xref: /linux/tools/power/x86/turbostat/turbostat.c (revision d53b8e36925256097a08d7cb749198d85cbf9b2b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * turbostat -- show CPU frequency and C-state residency
4  * on modern Intel and AMD processors.
5  *
6  * Copyright (c) 2024 Intel Corporation.
7  * Len Brown <len.brown@intel.com>
8  */
9 
10 #define _GNU_SOURCE
11 #include MSRHEADER
12 
13 // copied from arch/x86/include/asm/cpu_device_id.h
14 #define VFM_MODEL_BIT	0
15 #define VFM_FAMILY_BIT	8
16 #define VFM_VENDOR_BIT	16
17 #define VFM_RSVD_BIT	24
18 
19 #define	VFM_MODEL_MASK	GENMASK(VFM_FAMILY_BIT - 1, VFM_MODEL_BIT)
20 #define	VFM_FAMILY_MASK	GENMASK(VFM_VENDOR_BIT - 1, VFM_FAMILY_BIT)
21 #define	VFM_VENDOR_MASK	GENMASK(VFM_RSVD_BIT - 1, VFM_VENDOR_BIT)
22 
23 #define VFM_MODEL(vfm)	(((vfm) & VFM_MODEL_MASK) >> VFM_MODEL_BIT)
24 #define VFM_FAMILY(vfm)	(((vfm) & VFM_FAMILY_MASK) >> VFM_FAMILY_BIT)
25 #define VFM_VENDOR(vfm)	(((vfm) & VFM_VENDOR_MASK) >> VFM_VENDOR_BIT)
26 
27 #define	VFM_MAKE(_vendor, _family, _model) (	\
28 	((_model) << VFM_MODEL_BIT) |		\
29 	((_family) << VFM_FAMILY_BIT) |		\
30 	((_vendor) << VFM_VENDOR_BIT)		\
31 )
32 // end copied section
33 
34 #define X86_VENDOR_INTEL	0
35 
36 #include INTEL_FAMILY_HEADER
37 #include BUILD_BUG_HEADER
38 #include <stdarg.h>
39 #include <stdio.h>
40 #include <err.h>
41 #include <unistd.h>
42 #include <sys/types.h>
43 #include <sys/wait.h>
44 #include <sys/stat.h>
45 #include <sys/select.h>
46 #include <sys/resource.h>
47 #include <sys/mman.h>
48 #include <fcntl.h>
49 #include <signal.h>
50 #include <sys/time.h>
51 #include <stdlib.h>
52 #include <getopt.h>
53 #include <dirent.h>
54 #include <string.h>
55 #include <ctype.h>
56 #include <sched.h>
57 #include <time.h>
58 #include <cpuid.h>
59 #include <sys/capability.h>
60 #include <errno.h>
61 #include <math.h>
62 #include <linux/perf_event.h>
63 #include <asm/unistd.h>
64 #include <stdbool.h>
65 #include <assert.h>
66 #include <linux/kernel.h>
67 
68 #define UNUSED(x) (void)(x)
69 
70 /*
71  * This list matches the column headers, except
72  * 1. built-in only, the sysfs counters are not here -- we learn of those at run-time
73  * 2. Core and CPU are moved to the end, we can't have strings that contain them
74  *    matching on them for --show and --hide.
75  */
76 
77 /*
78  * buffer size used by sscanf() for added column names
79  * Usually truncated to 7 characters, but also handles 18 columns for raw 64-bit counters
80  */
81 #define	NAME_BYTES 20
82 #define PATH_BYTES 128
83 #define PERF_NAME_BYTES 128
84 
85 #define MAX_NOFILE 0x8000
86 
87 #define COUNTER_KIND_PERF_PREFIX "perf/"
88 #define COUNTER_KIND_PERF_PREFIX_LEN strlen(COUNTER_KIND_PERF_PREFIX)
89 #define PERF_DEV_NAME_BYTES 32
90 #define PERF_EVT_NAME_BYTES 32
91 
92 enum counter_scope { SCOPE_CPU, SCOPE_CORE, SCOPE_PACKAGE };
93 enum counter_type { COUNTER_ITEMS, COUNTER_CYCLES, COUNTER_SECONDS, COUNTER_USEC, COUNTER_K2M };
94 enum counter_format { FORMAT_RAW, FORMAT_DELTA, FORMAT_PERCENT, FORMAT_AVERAGE };
95 enum counter_source { COUNTER_SOURCE_NONE, COUNTER_SOURCE_PERF, COUNTER_SOURCE_MSR };
96 
97 struct perf_counter_info {
98 	struct perf_counter_info *next;
99 
100 	/* How to open the counter / What counter it is. */
101 	char device[PERF_DEV_NAME_BYTES];
102 	char event[PERF_EVT_NAME_BYTES];
103 
104 	/* How to show/format the counter. */
105 	char name[PERF_NAME_BYTES];
106 	unsigned int width;
107 	enum counter_scope scope;
108 	enum counter_type type;
109 	enum counter_format format;
110 	double scale;
111 
112 	/* For reading the counter. */
113 	int *fd_perf_per_domain;
114 	size_t num_domains;
115 };
116 
117 struct sysfs_path {
118 	char path[PATH_BYTES];
119 	int id;
120 	struct sysfs_path *next;
121 };
122 
123 struct msr_counter {
124 	unsigned int msr_num;
125 	char name[NAME_BYTES];
126 	struct sysfs_path *sp;
127 	unsigned int width;
128 	enum counter_type type;
129 	enum counter_format format;
130 	struct msr_counter *next;
131 	unsigned int flags;
132 #define	FLAGS_HIDE	(1 << 0)
133 #define	FLAGS_SHOW	(1 << 1)
134 #define	SYSFS_PERCPU	(1 << 1)
135 };
136 
137 struct msr_counter bic[] = {
138 	{ 0x0, "usec", NULL, 0, 0, 0, NULL, 0 },
139 	{ 0x0, "Time_Of_Day_Seconds", NULL, 0, 0, 0, NULL, 0 },
140 	{ 0x0, "Package", NULL, 0, 0, 0, NULL, 0 },
141 	{ 0x0, "Node", NULL, 0, 0, 0, NULL, 0 },
142 	{ 0x0, "Avg_MHz", NULL, 0, 0, 0, NULL, 0 },
143 	{ 0x0, "Busy%", NULL, 0, 0, 0, NULL, 0 },
144 	{ 0x0, "Bzy_MHz", NULL, 0, 0, 0, NULL, 0 },
145 	{ 0x0, "TSC_MHz", NULL, 0, 0, 0, NULL, 0 },
146 	{ 0x0, "IRQ", NULL, 0, 0, 0, NULL, 0 },
147 	{ 0x0, "SMI", NULL, 32, 0, FORMAT_DELTA, NULL, 0 },
148 	{ 0x0, "sysfs", NULL, 0, 0, 0, NULL, 0 },
149 	{ 0x0, "CPU%c1", NULL, 0, 0, 0, NULL, 0 },
150 	{ 0x0, "CPU%c3", NULL, 0, 0, 0, NULL, 0 },
151 	{ 0x0, "CPU%c6", NULL, 0, 0, 0, NULL, 0 },
152 	{ 0x0, "CPU%c7", NULL, 0, 0, 0, NULL, 0 },
153 	{ 0x0, "ThreadC", NULL, 0, 0, 0, NULL, 0 },
154 	{ 0x0, "CoreTmp", NULL, 0, 0, 0, NULL, 0 },
155 	{ 0x0, "CoreCnt", NULL, 0, 0, 0, NULL, 0 },
156 	{ 0x0, "PkgTmp", NULL, 0, 0, 0, NULL, 0 },
157 	{ 0x0, "GFX%rc6", NULL, 0, 0, 0, NULL, 0 },
158 	{ 0x0, "GFXMHz", NULL, 0, 0, 0, NULL, 0 },
159 	{ 0x0, "Pkg%pc2", NULL, 0, 0, 0, NULL, 0 },
160 	{ 0x0, "Pkg%pc3", NULL, 0, 0, 0, NULL, 0 },
161 	{ 0x0, "Pkg%pc6", NULL, 0, 0, 0, NULL, 0 },
162 	{ 0x0, "Pkg%pc7", NULL, 0, 0, 0, NULL, 0 },
163 	{ 0x0, "Pkg%pc8", NULL, 0, 0, 0, NULL, 0 },
164 	{ 0x0, "Pkg%pc9", NULL, 0, 0, 0, NULL, 0 },
165 	{ 0x0, "Pk%pc10", NULL, 0, 0, 0, NULL, 0 },
166 	{ 0x0, "CPU%LPI", NULL, 0, 0, 0, NULL, 0 },
167 	{ 0x0, "SYS%LPI", NULL, 0, 0, 0, NULL, 0 },
168 	{ 0x0, "PkgWatt", NULL, 0, 0, 0, NULL, 0 },
169 	{ 0x0, "CorWatt", NULL, 0, 0, 0, NULL, 0 },
170 	{ 0x0, "GFXWatt", NULL, 0, 0, 0, NULL, 0 },
171 	{ 0x0, "PkgCnt", NULL, 0, 0, 0, NULL, 0 },
172 	{ 0x0, "RAMWatt", NULL, 0, 0, 0, NULL, 0 },
173 	{ 0x0, "PKG_%", NULL, 0, 0, 0, NULL, 0 },
174 	{ 0x0, "RAM_%", NULL, 0, 0, 0, NULL, 0 },
175 	{ 0x0, "Pkg_J", NULL, 0, 0, 0, NULL, 0 },
176 	{ 0x0, "Cor_J", NULL, 0, 0, 0, NULL, 0 },
177 	{ 0x0, "GFX_J", NULL, 0, 0, 0, NULL, 0 },
178 	{ 0x0, "RAM_J", NULL, 0, 0, 0, NULL, 0 },
179 	{ 0x0, "Mod%c6", NULL, 0, 0, 0, NULL, 0 },
180 	{ 0x0, "Totl%C0", NULL, 0, 0, 0, NULL, 0 },
181 	{ 0x0, "Any%C0", NULL, 0, 0, 0, NULL, 0 },
182 	{ 0x0, "GFX%C0", NULL, 0, 0, 0, NULL, 0 },
183 	{ 0x0, "CPUGFX%", NULL, 0, 0, 0, NULL, 0 },
184 	{ 0x0, "Core", NULL, 0, 0, 0, NULL, 0 },
185 	{ 0x0, "CPU", NULL, 0, 0, 0, NULL, 0 },
186 	{ 0x0, "APIC", NULL, 0, 0, 0, NULL, 0 },
187 	{ 0x0, "X2APIC", NULL, 0, 0, 0, NULL, 0 },
188 	{ 0x0, "Die", NULL, 0, 0, 0, NULL, 0 },
189 	{ 0x0, "GFXAMHz", NULL, 0, 0, 0, NULL, 0 },
190 	{ 0x0, "IPC", NULL, 0, 0, 0, NULL, 0 },
191 	{ 0x0, "CoreThr", NULL, 0, 0, 0, NULL, 0 },
192 	{ 0x0, "UncMHz", NULL, 0, 0, 0, NULL, 0 },
193 	{ 0x0, "SAM%mc6", NULL, 0, 0, 0, NULL, 0 },
194 	{ 0x0, "SAMMHz", NULL, 0, 0, 0, NULL, 0 },
195 	{ 0x0, "SAMAMHz", NULL, 0, 0, 0, NULL, 0 },
196 	{ 0x0, "Die%c6", NULL, 0, 0, 0, NULL, 0 },
197 };
198 
199 #define MAX_BIC (sizeof(bic) / sizeof(struct msr_counter))
200 #define	BIC_USEC	(1ULL << 0)
201 #define	BIC_TOD		(1ULL << 1)
202 #define	BIC_Package	(1ULL << 2)
203 #define	BIC_Node	(1ULL << 3)
204 #define	BIC_Avg_MHz	(1ULL << 4)
205 #define	BIC_Busy	(1ULL << 5)
206 #define	BIC_Bzy_MHz	(1ULL << 6)
207 #define	BIC_TSC_MHz	(1ULL << 7)
208 #define	BIC_IRQ		(1ULL << 8)
209 #define	BIC_SMI		(1ULL << 9)
210 #define	BIC_sysfs	(1ULL << 10)
211 #define	BIC_CPU_c1	(1ULL << 11)
212 #define	BIC_CPU_c3	(1ULL << 12)
213 #define	BIC_CPU_c6	(1ULL << 13)
214 #define	BIC_CPU_c7	(1ULL << 14)
215 #define	BIC_ThreadC	(1ULL << 15)
216 #define	BIC_CoreTmp	(1ULL << 16)
217 #define	BIC_CoreCnt	(1ULL << 17)
218 #define	BIC_PkgTmp	(1ULL << 18)
219 #define	BIC_GFX_rc6	(1ULL << 19)
220 #define	BIC_GFXMHz	(1ULL << 20)
221 #define	BIC_Pkgpc2	(1ULL << 21)
222 #define	BIC_Pkgpc3	(1ULL << 22)
223 #define	BIC_Pkgpc6	(1ULL << 23)
224 #define	BIC_Pkgpc7	(1ULL << 24)
225 #define	BIC_Pkgpc8	(1ULL << 25)
226 #define	BIC_Pkgpc9	(1ULL << 26)
227 #define	BIC_Pkgpc10	(1ULL << 27)
228 #define BIC_CPU_LPI	(1ULL << 28)
229 #define BIC_SYS_LPI	(1ULL << 29)
230 #define	BIC_PkgWatt	(1ULL << 30)
231 #define	BIC_CorWatt	(1ULL << 31)
232 #define	BIC_GFXWatt	(1ULL << 32)
233 #define	BIC_PkgCnt	(1ULL << 33)
234 #define	BIC_RAMWatt	(1ULL << 34)
235 #define	BIC_PKG__	(1ULL << 35)
236 #define	BIC_RAM__	(1ULL << 36)
237 #define	BIC_Pkg_J	(1ULL << 37)
238 #define	BIC_Cor_J	(1ULL << 38)
239 #define	BIC_GFX_J	(1ULL << 39)
240 #define	BIC_RAM_J	(1ULL << 40)
241 #define	BIC_Mod_c6	(1ULL << 41)
242 #define	BIC_Totl_c0	(1ULL << 42)
243 #define	BIC_Any_c0	(1ULL << 43)
244 #define	BIC_GFX_c0	(1ULL << 44)
245 #define	BIC_CPUGFX	(1ULL << 45)
246 #define	BIC_Core	(1ULL << 46)
247 #define	BIC_CPU		(1ULL << 47)
248 #define	BIC_APIC	(1ULL << 48)
249 #define	BIC_X2APIC	(1ULL << 49)
250 #define	BIC_Die		(1ULL << 50)
251 #define	BIC_GFXACTMHz	(1ULL << 51)
252 #define	BIC_IPC		(1ULL << 52)
253 #define	BIC_CORE_THROT_CNT	(1ULL << 53)
254 #define	BIC_UNCORE_MHZ		(1ULL << 54)
255 #define	BIC_SAM_mc6		(1ULL << 55)
256 #define	BIC_SAMMHz		(1ULL << 56)
257 #define	BIC_SAMACTMHz		(1ULL << 57)
258 #define	BIC_Diec6		(1ULL << 58)
259 
260 #define BIC_TOPOLOGY (BIC_Package | BIC_Node | BIC_CoreCnt | BIC_PkgCnt | BIC_Core | BIC_CPU | BIC_Die )
261 #define BIC_THERMAL_PWR ( BIC_CoreTmp | BIC_PkgTmp | BIC_PkgWatt | BIC_CorWatt | BIC_GFXWatt | BIC_RAMWatt | BIC_PKG__ | BIC_RAM__)
262 #define BIC_FREQUENCY (BIC_Avg_MHz | BIC_Busy | BIC_Bzy_MHz | BIC_TSC_MHz | BIC_GFXMHz | BIC_GFXACTMHz | BIC_SAMMHz | BIC_SAMACTMHz | BIC_UNCORE_MHZ)
263 #define BIC_IDLE (BIC_sysfs | BIC_CPU_c1 | BIC_CPU_c3 | BIC_CPU_c6 | BIC_CPU_c7 | BIC_GFX_rc6 | BIC_Pkgpc2 | BIC_Pkgpc3 | BIC_Pkgpc6 | BIC_Pkgpc7 | BIC_Pkgpc8 | BIC_Pkgpc9 | BIC_Pkgpc10 | BIC_CPU_LPI | BIC_SYS_LPI | BIC_Mod_c6 | BIC_Totl_c0 | BIC_Any_c0 | BIC_GFX_c0 | BIC_CPUGFX | BIC_SAM_mc6 | BIC_Diec6)
264 #define BIC_OTHER ( BIC_IRQ | BIC_SMI | BIC_ThreadC | BIC_CoreTmp | BIC_IPC)
265 
266 #define BIC_DISABLED_BY_DEFAULT	(BIC_USEC | BIC_TOD | BIC_APIC | BIC_X2APIC)
267 
268 unsigned long long bic_enabled = (0xFFFFFFFFFFFFFFFFULL & ~BIC_DISABLED_BY_DEFAULT);
269 unsigned long long bic_present = BIC_USEC | BIC_TOD | BIC_sysfs | BIC_APIC | BIC_X2APIC;
270 
271 #define DO_BIC(COUNTER_NAME) (bic_enabled & bic_present & COUNTER_NAME)
272 #define DO_BIC_READ(COUNTER_NAME) (bic_present & COUNTER_NAME)
273 #define ENABLE_BIC(COUNTER_NAME) (bic_enabled |= COUNTER_NAME)
274 #define BIC_PRESENT(COUNTER_BIT) (bic_present |= COUNTER_BIT)
275 #define BIC_NOT_PRESENT(COUNTER_BIT) (bic_present &= ~COUNTER_BIT)
276 #define BIC_IS_ENABLED(COUNTER_BIT) (bic_enabled & COUNTER_BIT)
277 
278 /*
279  * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
280  * If you change the values, note they are used both in comparisons
281  * (>= PCL__7) and to index pkg_cstate_limit_strings[].
282  */
283 #define PCLUKN 0		/* Unknown */
284 #define PCLRSV 1		/* Reserved */
285 #define PCL__0 2		/* PC0 */
286 #define PCL__1 3		/* PC1 */
287 #define PCL__2 4		/* PC2 */
288 #define PCL__3 5		/* PC3 */
289 #define PCL__4 6		/* PC4 */
290 #define PCL__6 7		/* PC6 */
291 #define PCL_6N 8		/* PC6 No Retention */
292 #define PCL_6R 9		/* PC6 Retention */
293 #define PCL__7 10		/* PC7 */
294 #define PCL_7S 11		/* PC7 Shrink */
295 #define PCL__8 12		/* PC8 */
296 #define PCL__9 13		/* PC9 */
297 #define PCL_10 14		/* PC10 */
298 #define PCLUNL 15		/* Unlimited */
299 
300 struct amperf_group_fd;
301 
302 char *proc_stat = "/proc/stat";
303 FILE *outf;
304 int *fd_percpu;
305 int *fd_instr_count_percpu;
306 struct timeval interval_tv = { 5, 0 };
307 struct timespec interval_ts = { 5, 0 };
308 
309 unsigned int num_iterations;
310 unsigned int header_iterations;
311 unsigned int debug;
312 unsigned int quiet;
313 unsigned int shown;
314 unsigned int sums_need_wide_columns;
315 unsigned int rapl_joules;
316 unsigned int summary_only;
317 unsigned int list_header_only;
318 unsigned int dump_only;
319 unsigned int has_aperf;
320 unsigned int has_aperf_access;
321 unsigned int has_epb;
322 unsigned int has_turbo;
323 unsigned int is_hybrid;
324 unsigned int units = 1000000;	/* MHz etc */
325 unsigned int genuine_intel;
326 unsigned int authentic_amd;
327 unsigned int hygon_genuine;
328 unsigned int max_level, max_extended_level;
329 unsigned int has_invariant_tsc;
330 unsigned int aperf_mperf_multiplier = 1;
331 double bclk;
332 double base_hz;
333 unsigned int has_base_hz;
334 double tsc_tweak = 1.0;
335 unsigned int show_pkg_only;
336 unsigned int show_core_only;
337 char *output_buffer, *outp;
338 unsigned int do_dts;
339 unsigned int do_ptm;
340 unsigned int do_ipc;
341 unsigned long long cpuidle_cur_cpu_lpi_us;
342 unsigned long long cpuidle_cur_sys_lpi_us;
343 unsigned int tj_max;
344 unsigned int tj_max_override;
345 double rapl_power_units, rapl_time_units;
346 double rapl_dram_energy_units, rapl_energy_units;
347 double rapl_joule_counter_range;
348 unsigned int crystal_hz;
349 unsigned long long tsc_hz;
350 int base_cpu;
351 unsigned int has_hwp;		/* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
352 			/* IA32_HWP_REQUEST, IA32_HWP_STATUS */
353 unsigned int has_hwp_notify;	/* IA32_HWP_INTERRUPT */
354 unsigned int has_hwp_activity_window;	/* IA32_HWP_REQUEST[bits 41:32] */
355 unsigned int has_hwp_epp;	/* IA32_HWP_REQUEST[bits 31:24] */
356 unsigned int has_hwp_pkg;	/* IA32_HWP_REQUEST_PKG */
357 unsigned int first_counter_read = 1;
358 int ignore_stdin;
359 bool no_msr;
360 bool no_perf;
361 
362 enum gfx_sysfs_idx {
363 	GFX_rc6,
364 	GFX_MHz,
365 	GFX_ACTMHz,
366 	SAM_mc6,
367 	SAM_MHz,
368 	SAM_ACTMHz,
369 	GFX_MAX
370 };
371 
372 struct gfx_sysfs_info {
373 	const char *path;
374 	FILE *fp;
375 	unsigned int val;
376 	unsigned long long val_ull;
377 };
378 
379 static struct gfx_sysfs_info gfx_info[GFX_MAX];
380 
381 int get_msr(int cpu, off_t offset, unsigned long long *msr);
382 int add_counter(unsigned int msr_num, char *path, char *name,
383 		unsigned int width, enum counter_scope scope,
384 		enum counter_type type, enum counter_format format, int flags, int package_num);
385 
386 /* Model specific support Start */
387 
388 /* List of features that may diverge among different platforms */
389 struct platform_features {
390 	bool has_msr_misc_feature_control;	/* MSR_MISC_FEATURE_CONTROL */
391 	bool has_msr_misc_pwr_mgmt;	/* MSR_MISC_PWR_MGMT */
392 	bool has_nhm_msrs;	/* MSR_PLATFORM_INFO, MSR_IA32_TEMPERATURE_TARGET, MSR_SMI_COUNT, MSR_PKG_CST_CONFIG_CONTROL, MSR_IA32_POWER_CTL, TRL MSRs */
393 	bool has_config_tdp;	/* MSR_CONFIG_TDP_NOMINAL/LEVEL_1/LEVEL_2/CONTROL, MSR_TURBO_ACTIVATION_RATIO */
394 	int bclk_freq;		/* CPU base clock */
395 	int crystal_freq;	/* Crystal clock to use when not available from CPUID.15 */
396 	int supported_cstates;	/* Core cstates and Package cstates supported */
397 	int cst_limit;		/* MSR_PKG_CST_CONFIG_CONTROL */
398 	bool has_cst_auto_convension;	/* AUTOMATIC_CSTATE_CONVERSION bit in MSR_PKG_CST_CONFIG_CONTROL */
399 	bool has_irtl_msrs;	/* MSR_PKGC3/PKGC6/PKGC7/PKGC8/PKGC9/PKGC10_IRTL */
400 	bool has_msr_core_c1_res;	/* MSR_CORE_C1_RES */
401 	bool has_msr_module_c6_res_ms;	/* MSR_MODULE_C6_RES_MS */
402 	bool has_msr_c6_demotion_policy_config;	/* MSR_CC6_DEMOTION_POLICY_CONFIG/MSR_MC6_DEMOTION_POLICY_CONFIG */
403 	bool has_msr_atom_pkg_c6_residency;	/* MSR_ATOM_PKG_C6_RESIDENCY */
404 	bool has_msr_knl_core_c6_residency;	/* MSR_KNL_CORE_C6_RESIDENCY */
405 	bool has_ext_cst_msrs;	/* MSR_PKG_WEIGHTED_CORE_C0_RES/MSR_PKG_ANY_CORE_C0_RES/MSR_PKG_ANY_GFXE_C0_RES/MSR_PKG_BOTH_CORE_GFXE_C0_RES */
406 	bool has_cst_prewake_bit;	/* Cstate prewake bit in MSR_IA32_POWER_CTL */
407 	int trl_msrs;		/* MSR_TURBO_RATIO_LIMIT/LIMIT1/LIMIT2/SECONDARY, Atom TRL MSRs */
408 	int plr_msrs;		/* MSR_CORE/GFX/RING_PERF_LIMIT_REASONS */
409 	int rapl_msrs;		/* RAPL PKG/DRAM/CORE/GFX MSRs, AMD RAPL MSRs */
410 	bool has_per_core_rapl;	/* Indicates cores energy collection is per-core, not per-package. AMD specific for now */
411 	bool has_rapl_divisor;	/* Divisor for Energy unit raw value from MSR_RAPL_POWER_UNIT */
412 	bool has_fixed_rapl_unit;	/* Fixed Energy Unit used for DRAM RAPL Domain */
413 	int rapl_quirk_tdp;	/* Hardcoded TDP value when cannot be retrieved from hardware */
414 	int tcc_offset_bits;	/* TCC Offset bits in MSR_IA32_TEMPERATURE_TARGET */
415 	bool enable_tsc_tweak;	/* Use CPU Base freq instead of TSC freq for aperf/mperf counter */
416 	bool need_perf_multiplier;	/* mperf/aperf multiplier */
417 };
418 
419 struct platform_data {
420 	unsigned int vfm;
421 	const struct platform_features *features;
422 };
423 
424 /* For BCLK */
425 enum bclk_freq {
426 	BCLK_100MHZ = 1,
427 	BCLK_133MHZ,
428 	BCLK_SLV,
429 };
430 
431 #define SLM_BCLK_FREQS 5
432 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0 };
433 
434 double slm_bclk(void)
435 {
436 	unsigned long long msr = 3;
437 	unsigned int i;
438 	double freq;
439 
440 	if (get_msr(base_cpu, MSR_FSB_FREQ, &msr))
441 		fprintf(outf, "SLM BCLK: unknown\n");
442 
443 	i = msr & 0xf;
444 	if (i >= SLM_BCLK_FREQS) {
445 		fprintf(outf, "SLM BCLK[%d] invalid\n", i);
446 		i = 3;
447 	}
448 	freq = slm_freq_table[i];
449 
450 	if (!quiet)
451 		fprintf(outf, "SLM BCLK: %.1f Mhz\n", freq);
452 
453 	return freq;
454 }
455 
456 /* For Package cstate limit */
457 enum package_cstate_limit {
458 	CST_LIMIT_NHM = 1,
459 	CST_LIMIT_SNB,
460 	CST_LIMIT_HSW,
461 	CST_LIMIT_SKX,
462 	CST_LIMIT_ICX,
463 	CST_LIMIT_SLV,
464 	CST_LIMIT_AMT,
465 	CST_LIMIT_KNL,
466 	CST_LIMIT_GMT,
467 };
468 
469 /* For Turbo Ratio Limit MSRs */
470 enum turbo_ratio_limit_msrs {
471 	TRL_BASE = BIT(0),
472 	TRL_LIMIT1 = BIT(1),
473 	TRL_LIMIT2 = BIT(2),
474 	TRL_ATOM = BIT(3),
475 	TRL_KNL = BIT(4),
476 	TRL_CORECOUNT = BIT(5),
477 };
478 
479 /* For Perf Limit Reason MSRs */
480 enum perf_limit_reason_msrs {
481 	PLR_CORE = BIT(0),
482 	PLR_GFX = BIT(1),
483 	PLR_RING = BIT(2),
484 };
485 
486 /* For RAPL MSRs */
487 enum rapl_msrs {
488 	RAPL_PKG_POWER_LIMIT = BIT(0),	/* 0x610 MSR_PKG_POWER_LIMIT */
489 	RAPL_PKG_ENERGY_STATUS = BIT(1),	/* 0x611 MSR_PKG_ENERGY_STATUS */
490 	RAPL_PKG_PERF_STATUS = BIT(2),	/* 0x613 MSR_PKG_PERF_STATUS */
491 	RAPL_PKG_POWER_INFO = BIT(3),	/* 0x614 MSR_PKG_POWER_INFO */
492 	RAPL_DRAM_POWER_LIMIT = BIT(4),	/* 0x618 MSR_DRAM_POWER_LIMIT */
493 	RAPL_DRAM_ENERGY_STATUS = BIT(5),	/* 0x619 MSR_DRAM_ENERGY_STATUS */
494 	RAPL_DRAM_PERF_STATUS = BIT(6),	/* 0x61b MSR_DRAM_PERF_STATUS */
495 	RAPL_DRAM_POWER_INFO = BIT(7),	/* 0x61c MSR_DRAM_POWER_INFO */
496 	RAPL_CORE_POWER_LIMIT = BIT(8),	/* 0x638 MSR_PP0_POWER_LIMIT */
497 	RAPL_CORE_ENERGY_STATUS = BIT(9),	/* 0x639 MSR_PP0_ENERGY_STATUS */
498 	RAPL_CORE_POLICY = BIT(10),	/* 0x63a MSR_PP0_POLICY */
499 	RAPL_GFX_POWER_LIMIT = BIT(11),	/* 0x640 MSR_PP1_POWER_LIMIT */
500 	RAPL_GFX_ENERGY_STATUS = BIT(12),	/* 0x641 MSR_PP1_ENERGY_STATUS */
501 	RAPL_GFX_POLICY = BIT(13),	/* 0x642 MSR_PP1_POLICY */
502 	RAPL_AMD_PWR_UNIT = BIT(14),	/* 0xc0010299 MSR_AMD_RAPL_POWER_UNIT */
503 	RAPL_AMD_CORE_ENERGY_STAT = BIT(15),	/* 0xc001029a MSR_AMD_CORE_ENERGY_STATUS */
504 	RAPL_AMD_PKG_ENERGY_STAT = BIT(16),	/* 0xc001029b MSR_AMD_PKG_ENERGY_STATUS */
505 };
506 
507 #define RAPL_PKG	(RAPL_PKG_ENERGY_STATUS | RAPL_PKG_POWER_LIMIT)
508 #define RAPL_DRAM	(RAPL_DRAM_ENERGY_STATUS | RAPL_DRAM_POWER_LIMIT)
509 #define RAPL_CORE	(RAPL_CORE_ENERGY_STATUS | RAPL_CORE_POWER_LIMIT)
510 #define RAPL_GFX	(RAPL_GFX_POWER_LIMIT | RAPL_GFX_ENERGY_STATUS)
511 
512 #define RAPL_PKG_ALL	(RAPL_PKG | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO)
513 #define RAPL_DRAM_ALL	(RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_DRAM_POWER_INFO)
514 #define RAPL_CORE_ALL	(RAPL_CORE | RAPL_CORE_POLICY)
515 #define RAPL_GFX_ALL	(RAPL_GFX | RAPL_GFX_POLIGY)
516 
517 #define RAPL_AMD_F17H	(RAPL_AMD_PWR_UNIT | RAPL_AMD_CORE_ENERGY_STAT | RAPL_AMD_PKG_ENERGY_STAT)
518 
519 /* For Cstates */
520 enum cstates {
521 	CC1 = BIT(0),
522 	CC3 = BIT(1),
523 	CC6 = BIT(2),
524 	CC7 = BIT(3),
525 	PC2 = BIT(4),
526 	PC3 = BIT(5),
527 	PC6 = BIT(6),
528 	PC7 = BIT(7),
529 	PC8 = BIT(8),
530 	PC9 = BIT(9),
531 	PC10 = BIT(10),
532 };
533 
534 static const struct platform_features nhm_features = {
535 	.has_msr_misc_pwr_mgmt = 1,
536 	.has_nhm_msrs = 1,
537 	.bclk_freq = BCLK_133MHZ,
538 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
539 	.cst_limit = CST_LIMIT_NHM,
540 	.trl_msrs = TRL_BASE,
541 };
542 
543 static const struct platform_features nhx_features = {
544 	.has_msr_misc_pwr_mgmt = 1,
545 	.has_nhm_msrs = 1,
546 	.bclk_freq = BCLK_133MHZ,
547 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
548 	.cst_limit = CST_LIMIT_NHM,
549 };
550 
551 static const struct platform_features snb_features = {
552 	.has_msr_misc_feature_control = 1,
553 	.has_msr_misc_pwr_mgmt = 1,
554 	.has_nhm_msrs = 1,
555 	.bclk_freq = BCLK_100MHZ,
556 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
557 	.cst_limit = CST_LIMIT_SNB,
558 	.has_irtl_msrs = 1,
559 	.trl_msrs = TRL_BASE,
560 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
561 };
562 
563 static const struct platform_features snx_features = {
564 	.has_msr_misc_feature_control = 1,
565 	.has_msr_misc_pwr_mgmt = 1,
566 	.has_nhm_msrs = 1,
567 	.bclk_freq = BCLK_100MHZ,
568 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
569 	.cst_limit = CST_LIMIT_SNB,
570 	.has_irtl_msrs = 1,
571 	.trl_msrs = TRL_BASE,
572 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM_ALL,
573 };
574 
575 static const struct platform_features ivb_features = {
576 	.has_msr_misc_feature_control = 1,
577 	.has_msr_misc_pwr_mgmt = 1,
578 	.has_nhm_msrs = 1,
579 	.has_config_tdp = 1,
580 	.bclk_freq = BCLK_100MHZ,
581 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
582 	.cst_limit = CST_LIMIT_SNB,
583 	.has_irtl_msrs = 1,
584 	.trl_msrs = TRL_BASE,
585 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
586 };
587 
588 static const struct platform_features ivx_features = {
589 	.has_msr_misc_feature_control = 1,
590 	.has_msr_misc_pwr_mgmt = 1,
591 	.has_nhm_msrs = 1,
592 	.bclk_freq = BCLK_100MHZ,
593 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
594 	.cst_limit = CST_LIMIT_SNB,
595 	.has_irtl_msrs = 1,
596 	.trl_msrs = TRL_BASE | TRL_LIMIT1,
597 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM_ALL,
598 };
599 
600 static const struct platform_features hsw_features = {
601 	.has_msr_misc_feature_control = 1,
602 	.has_msr_misc_pwr_mgmt = 1,
603 	.has_nhm_msrs = 1,
604 	.has_config_tdp = 1,
605 	.bclk_freq = BCLK_100MHZ,
606 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
607 	.cst_limit = CST_LIMIT_HSW,
608 	.has_irtl_msrs = 1,
609 	.trl_msrs = TRL_BASE,
610 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
611 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
612 };
613 
614 static const struct platform_features hsx_features = {
615 	.has_msr_misc_feature_control = 1,
616 	.has_msr_misc_pwr_mgmt = 1,
617 	.has_nhm_msrs = 1,
618 	.has_config_tdp = 1,
619 	.bclk_freq = BCLK_100MHZ,
620 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
621 	.cst_limit = CST_LIMIT_HSW,
622 	.has_irtl_msrs = 1,
623 	.trl_msrs = TRL_BASE | TRL_LIMIT1 | TRL_LIMIT2,
624 	.plr_msrs = PLR_CORE | PLR_RING,
625 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
626 	.has_fixed_rapl_unit = 1,
627 };
628 
629 static const struct platform_features hswl_features = {
630 	.has_msr_misc_feature_control = 1,
631 	.has_msr_misc_pwr_mgmt = 1,
632 	.has_nhm_msrs = 1,
633 	.has_config_tdp = 1,
634 	.bclk_freq = BCLK_100MHZ,
635 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
636 	.cst_limit = CST_LIMIT_HSW,
637 	.has_irtl_msrs = 1,
638 	.trl_msrs = TRL_BASE,
639 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
640 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
641 };
642 
643 static const struct platform_features hswg_features = {
644 	.has_msr_misc_feature_control = 1,
645 	.has_msr_misc_pwr_mgmt = 1,
646 	.has_nhm_msrs = 1,
647 	.has_config_tdp = 1,
648 	.bclk_freq = BCLK_100MHZ,
649 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
650 	.cst_limit = CST_LIMIT_HSW,
651 	.has_irtl_msrs = 1,
652 	.trl_msrs = TRL_BASE,
653 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
654 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
655 };
656 
657 static const struct platform_features bdw_features = {
658 	.has_msr_misc_feature_control = 1,
659 	.has_msr_misc_pwr_mgmt = 1,
660 	.has_nhm_msrs = 1,
661 	.has_config_tdp = 1,
662 	.bclk_freq = BCLK_100MHZ,
663 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
664 	.cst_limit = CST_LIMIT_HSW,
665 	.has_irtl_msrs = 1,
666 	.trl_msrs = TRL_BASE,
667 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
668 };
669 
670 static const struct platform_features bdwg_features = {
671 	.has_msr_misc_feature_control = 1,
672 	.has_msr_misc_pwr_mgmt = 1,
673 	.has_nhm_msrs = 1,
674 	.has_config_tdp = 1,
675 	.bclk_freq = BCLK_100MHZ,
676 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
677 	.cst_limit = CST_LIMIT_HSW,
678 	.has_irtl_msrs = 1,
679 	.trl_msrs = TRL_BASE,
680 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
681 };
682 
683 static const struct platform_features bdx_features = {
684 	.has_msr_misc_feature_control = 1,
685 	.has_msr_misc_pwr_mgmt = 1,
686 	.has_nhm_msrs = 1,
687 	.has_config_tdp = 1,
688 	.bclk_freq = BCLK_100MHZ,
689 	.supported_cstates = CC1 | CC3 | CC6 | PC2 | PC3 | PC6,
690 	.cst_limit = CST_LIMIT_HSW,
691 	.has_irtl_msrs = 1,
692 	.has_cst_auto_convension = 1,
693 	.trl_msrs = TRL_BASE,
694 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
695 	.has_fixed_rapl_unit = 1,
696 };
697 
698 static const struct platform_features skl_features = {
699 	.has_msr_misc_feature_control = 1,
700 	.has_msr_misc_pwr_mgmt = 1,
701 	.has_nhm_msrs = 1,
702 	.has_config_tdp = 1,
703 	.bclk_freq = BCLK_100MHZ,
704 	.crystal_freq = 24000000,
705 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
706 	.cst_limit = CST_LIMIT_HSW,
707 	.has_irtl_msrs = 1,
708 	.has_ext_cst_msrs = 1,
709 	.trl_msrs = TRL_BASE,
710 	.tcc_offset_bits = 6,
711 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX,
712 	.enable_tsc_tweak = 1,
713 };
714 
715 static const struct platform_features cnl_features = {
716 	.has_msr_misc_feature_control = 1,
717 	.has_msr_misc_pwr_mgmt = 1,
718 	.has_nhm_msrs = 1,
719 	.has_config_tdp = 1,
720 	.bclk_freq = BCLK_100MHZ,
721 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
722 	.cst_limit = CST_LIMIT_HSW,
723 	.has_irtl_msrs = 1,
724 	.has_msr_core_c1_res = 1,
725 	.has_ext_cst_msrs = 1,
726 	.trl_msrs = TRL_BASE,
727 	.tcc_offset_bits = 6,
728 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX,
729 	.enable_tsc_tweak = 1,
730 };
731 
732 static const struct platform_features adl_features = {
733 	.has_msr_misc_feature_control = 1,
734 	.has_msr_misc_pwr_mgmt = 1,
735 	.has_nhm_msrs = 1,
736 	.has_config_tdp = 1,
737 	.bclk_freq = BCLK_100MHZ,
738 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC8 | PC10,
739 	.cst_limit = CST_LIMIT_HSW,
740 	.has_irtl_msrs = 1,
741 	.has_msr_core_c1_res = 1,
742 	.has_ext_cst_msrs = 1,
743 	.trl_msrs = TRL_BASE,
744 	.tcc_offset_bits = 6,
745 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX,
746 	.enable_tsc_tweak = 1,
747 };
748 
749 static const struct platform_features arl_features = {
750 	.has_msr_misc_feature_control = 1,
751 	.has_msr_misc_pwr_mgmt = 1,
752 	.has_nhm_msrs = 1,
753 	.has_config_tdp = 1,
754 	.bclk_freq = BCLK_100MHZ,
755 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC10,
756 	.cst_limit = CST_LIMIT_HSW,
757 	.has_irtl_msrs = 1,
758 	.has_msr_core_c1_res = 1,
759 	.has_ext_cst_msrs = 1,
760 	.trl_msrs = TRL_BASE,
761 	.tcc_offset_bits = 6,
762 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX,
763 	.enable_tsc_tweak = 1,
764 };
765 
766 static const struct platform_features skx_features = {
767 	.has_msr_misc_feature_control = 1,
768 	.has_msr_misc_pwr_mgmt = 1,
769 	.has_nhm_msrs = 1,
770 	.has_config_tdp = 1,
771 	.bclk_freq = BCLK_100MHZ,
772 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
773 	.cst_limit = CST_LIMIT_SKX,
774 	.has_irtl_msrs = 1,
775 	.has_cst_auto_convension = 1,
776 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
777 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
778 	.has_fixed_rapl_unit = 1,
779 };
780 
781 static const struct platform_features icx_features = {
782 	.has_msr_misc_feature_control = 1,
783 	.has_msr_misc_pwr_mgmt = 1,
784 	.has_nhm_msrs = 1,
785 	.has_config_tdp = 1,
786 	.bclk_freq = BCLK_100MHZ,
787 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
788 	.cst_limit = CST_LIMIT_ICX,
789 	.has_msr_core_c1_res = 1,
790 	.has_irtl_msrs = 1,
791 	.has_cst_prewake_bit = 1,
792 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
793 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
794 	.has_fixed_rapl_unit = 1,
795 };
796 
797 static const struct platform_features spr_features = {
798 	.has_msr_misc_feature_control = 1,
799 	.has_msr_misc_pwr_mgmt = 1,
800 	.has_nhm_msrs = 1,
801 	.has_config_tdp = 1,
802 	.bclk_freq = BCLK_100MHZ,
803 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
804 	.cst_limit = CST_LIMIT_SKX,
805 	.has_msr_core_c1_res = 1,
806 	.has_irtl_msrs = 1,
807 	.has_cst_prewake_bit = 1,
808 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
809 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
810 };
811 
812 static const struct platform_features srf_features = {
813 	.has_msr_misc_feature_control = 1,
814 	.has_msr_misc_pwr_mgmt = 1,
815 	.has_nhm_msrs = 1,
816 	.has_config_tdp = 1,
817 	.bclk_freq = BCLK_100MHZ,
818 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
819 	.cst_limit = CST_LIMIT_SKX,
820 	.has_msr_core_c1_res = 1,
821 	.has_msr_module_c6_res_ms = 1,
822 	.has_irtl_msrs = 1,
823 	.has_cst_prewake_bit = 1,
824 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
825 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
826 };
827 
828 static const struct platform_features grr_features = {
829 	.has_msr_misc_feature_control = 1,
830 	.has_msr_misc_pwr_mgmt = 1,
831 	.has_nhm_msrs = 1,
832 	.has_config_tdp = 1,
833 	.bclk_freq = BCLK_100MHZ,
834 	.supported_cstates = CC1 | CC6,
835 	.cst_limit = CST_LIMIT_SKX,
836 	.has_msr_core_c1_res = 1,
837 	.has_msr_module_c6_res_ms = 1,
838 	.has_irtl_msrs = 1,
839 	.has_cst_prewake_bit = 1,
840 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
841 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
842 };
843 
844 static const struct platform_features slv_features = {
845 	.has_nhm_msrs = 1,
846 	.bclk_freq = BCLK_SLV,
847 	.supported_cstates = CC1 | CC6 | PC6,
848 	.cst_limit = CST_LIMIT_SLV,
849 	.has_msr_core_c1_res = 1,
850 	.has_msr_module_c6_res_ms = 1,
851 	.has_msr_c6_demotion_policy_config = 1,
852 	.has_msr_atom_pkg_c6_residency = 1,
853 	.trl_msrs = TRL_ATOM,
854 	.rapl_msrs = RAPL_PKG | RAPL_CORE,
855 	.has_rapl_divisor = 1,
856 	.rapl_quirk_tdp = 30,
857 };
858 
859 static const struct platform_features slvd_features = {
860 	.has_msr_misc_pwr_mgmt = 1,
861 	.has_nhm_msrs = 1,
862 	.bclk_freq = BCLK_SLV,
863 	.supported_cstates = CC1 | CC6 | PC3 | PC6,
864 	.cst_limit = CST_LIMIT_SLV,
865 	.has_msr_atom_pkg_c6_residency = 1,
866 	.trl_msrs = TRL_BASE,
867 	.rapl_msrs = RAPL_PKG | RAPL_CORE,
868 	.rapl_quirk_tdp = 30,
869 };
870 
871 static const struct platform_features amt_features = {
872 	.has_nhm_msrs = 1,
873 	.bclk_freq = BCLK_133MHZ,
874 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
875 	.cst_limit = CST_LIMIT_AMT,
876 	.trl_msrs = TRL_BASE,
877 };
878 
879 static const struct platform_features gmt_features = {
880 	.has_msr_misc_pwr_mgmt = 1,
881 	.has_nhm_msrs = 1,
882 	.bclk_freq = BCLK_100MHZ,
883 	.crystal_freq = 19200000,
884 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
885 	.cst_limit = CST_LIMIT_GMT,
886 	.has_irtl_msrs = 1,
887 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
888 	.rapl_msrs = RAPL_PKG | RAPL_PKG_POWER_INFO,
889 };
890 
891 static const struct platform_features gmtd_features = {
892 	.has_msr_misc_pwr_mgmt = 1,
893 	.has_nhm_msrs = 1,
894 	.bclk_freq = BCLK_100MHZ,
895 	.crystal_freq = 25000000,
896 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
897 	.cst_limit = CST_LIMIT_GMT,
898 	.has_irtl_msrs = 1,
899 	.has_msr_core_c1_res = 1,
900 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
901 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_CORE_ENERGY_STATUS,
902 };
903 
904 static const struct platform_features gmtp_features = {
905 	.has_msr_misc_pwr_mgmt = 1,
906 	.has_nhm_msrs = 1,
907 	.bclk_freq = BCLK_100MHZ,
908 	.crystal_freq = 19200000,
909 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
910 	.cst_limit = CST_LIMIT_GMT,
911 	.has_irtl_msrs = 1,
912 	.trl_msrs = TRL_BASE,
913 	.rapl_msrs = RAPL_PKG | RAPL_PKG_POWER_INFO,
914 };
915 
916 static const struct platform_features tmt_features = {
917 	.has_msr_misc_pwr_mgmt = 1,
918 	.has_nhm_msrs = 1,
919 	.bclk_freq = BCLK_100MHZ,
920 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
921 	.cst_limit = CST_LIMIT_GMT,
922 	.has_irtl_msrs = 1,
923 	.trl_msrs = TRL_BASE,
924 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX,
925 	.enable_tsc_tweak = 1,
926 };
927 
928 static const struct platform_features tmtd_features = {
929 	.has_msr_misc_pwr_mgmt = 1,
930 	.has_nhm_msrs = 1,
931 	.bclk_freq = BCLK_100MHZ,
932 	.supported_cstates = CC1 | CC6,
933 	.cst_limit = CST_LIMIT_GMT,
934 	.has_irtl_msrs = 1,
935 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
936 	.rapl_msrs = RAPL_PKG_ALL,
937 };
938 
939 static const struct platform_features knl_features = {
940 	.has_msr_misc_pwr_mgmt = 1,
941 	.has_nhm_msrs = 1,
942 	.has_config_tdp = 1,
943 	.bclk_freq = BCLK_100MHZ,
944 	.supported_cstates = CC1 | CC6 | PC3 | PC6,
945 	.cst_limit = CST_LIMIT_KNL,
946 	.has_msr_knl_core_c6_residency = 1,
947 	.trl_msrs = TRL_KNL,
948 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
949 	.has_fixed_rapl_unit = 1,
950 	.need_perf_multiplier = 1,
951 };
952 
953 static const struct platform_features default_features = {
954 };
955 
956 static const struct platform_features amd_features_with_rapl = {
957 	.rapl_msrs = RAPL_AMD_F17H,
958 	.has_per_core_rapl = 1,
959 	.rapl_quirk_tdp = 280,	/* This is the max stock TDP of HEDT/Server Fam17h+ chips */
960 };
961 
962 static const struct platform_data turbostat_pdata[] = {
963 	{ INTEL_NEHALEM, &nhm_features },
964 	{ INTEL_NEHALEM_G, &nhm_features },
965 	{ INTEL_NEHALEM_EP, &nhm_features },
966 	{ INTEL_NEHALEM_EX, &nhx_features },
967 	{ INTEL_WESTMERE, &nhm_features },
968 	{ INTEL_WESTMERE_EP, &nhm_features },
969 	{ INTEL_WESTMERE_EX, &nhx_features },
970 	{ INTEL_SANDYBRIDGE, &snb_features },
971 	{ INTEL_SANDYBRIDGE_X, &snx_features },
972 	{ INTEL_IVYBRIDGE, &ivb_features },
973 	{ INTEL_IVYBRIDGE_X, &ivx_features },
974 	{ INTEL_HASWELL, &hsw_features },
975 	{ INTEL_HASWELL_X, &hsx_features },
976 	{ INTEL_HASWELL_L, &hswl_features },
977 	{ INTEL_HASWELL_G, &hswg_features },
978 	{ INTEL_BROADWELL, &bdw_features },
979 	{ INTEL_BROADWELL_G, &bdwg_features },
980 	{ INTEL_BROADWELL_X, &bdx_features },
981 	{ INTEL_BROADWELL_D, &bdx_features },
982 	{ INTEL_SKYLAKE_L, &skl_features },
983 	{ INTEL_SKYLAKE, &skl_features },
984 	{ INTEL_SKYLAKE_X, &skx_features },
985 	{ INTEL_KABYLAKE_L, &skl_features },
986 	{ INTEL_KABYLAKE, &skl_features },
987 	{ INTEL_COMETLAKE, &skl_features },
988 	{ INTEL_COMETLAKE_L, &skl_features },
989 	{ INTEL_CANNONLAKE_L, &cnl_features },
990 	{ INTEL_ICELAKE_X, &icx_features },
991 	{ INTEL_ICELAKE_D, &icx_features },
992 	{ INTEL_ICELAKE_L, &cnl_features },
993 	{ INTEL_ICELAKE_NNPI, &cnl_features },
994 	{ INTEL_ROCKETLAKE, &cnl_features },
995 	{ INTEL_TIGERLAKE_L, &cnl_features },
996 	{ INTEL_TIGERLAKE, &cnl_features },
997 	{ INTEL_SAPPHIRERAPIDS_X, &spr_features },
998 	{ INTEL_EMERALDRAPIDS_X, &spr_features },
999 	{ INTEL_GRANITERAPIDS_X, &spr_features },
1000 	{ INTEL_LAKEFIELD, &cnl_features },
1001 	{ INTEL_ALDERLAKE, &adl_features },
1002 	{ INTEL_ALDERLAKE_L, &adl_features },
1003 	{ INTEL_RAPTORLAKE, &adl_features },
1004 	{ INTEL_RAPTORLAKE_P, &adl_features },
1005 	{ INTEL_RAPTORLAKE_S, &adl_features },
1006 	{ INTEL_METEORLAKE, &cnl_features },
1007 	{ INTEL_METEORLAKE_L, &cnl_features },
1008 	{ INTEL_ARROWLAKE_H, &arl_features },
1009 	{ INTEL_ARROWLAKE_U, &arl_features },
1010 	{ INTEL_ARROWLAKE, &arl_features },
1011 	{ INTEL_LUNARLAKE_M, &arl_features },
1012 	{ INTEL_ATOM_SILVERMONT, &slv_features },
1013 	{ INTEL_ATOM_SILVERMONT_D, &slvd_features },
1014 	{ INTEL_ATOM_AIRMONT, &amt_features },
1015 	{ INTEL_ATOM_GOLDMONT, &gmt_features },
1016 	{ INTEL_ATOM_GOLDMONT_D, &gmtd_features },
1017 	{ INTEL_ATOM_GOLDMONT_PLUS, &gmtp_features },
1018 	{ INTEL_ATOM_TREMONT_D, &tmtd_features },
1019 	{ INTEL_ATOM_TREMONT, &tmt_features },
1020 	{ INTEL_ATOM_TREMONT_L, &tmt_features },
1021 	{ INTEL_ATOM_GRACEMONT, &adl_features },
1022 	{ INTEL_ATOM_CRESTMONT_X, &srf_features },
1023 	{ INTEL_ATOM_CRESTMONT, &grr_features },
1024 	{ INTEL_XEON_PHI_KNL, &knl_features },
1025 	{ INTEL_XEON_PHI_KNM, &knl_features },
1026 	/*
1027 	 * Missing support for
1028 	 * INTEL_ICELAKE
1029 	 * INTEL_ATOM_SILVERMONT_MID
1030 	 * INTEL_ATOM_AIRMONT_MID
1031 	 * INTEL_ATOM_AIRMONT_NP
1032 	 */
1033 	{ 0, NULL },
1034 };
1035 
1036 static const struct platform_features *platform;
1037 
1038 void probe_platform_features(unsigned int family, unsigned int model)
1039 {
1040 	int i;
1041 
1042 	platform = &default_features;
1043 
1044 	if (authentic_amd || hygon_genuine) {
1045 		if (max_extended_level >= 0x80000007) {
1046 			unsigned int eax, ebx, ecx, edx;
1047 
1048 			__cpuid(0x80000007, eax, ebx, ecx, edx);
1049 			/* RAPL (Fam 17h+) */
1050 			if ((edx & (1 << 14)) && family >= 0x17)
1051 				platform = &amd_features_with_rapl;
1052 		}
1053 		return;
1054 	}
1055 
1056 	if (!genuine_intel)
1057 		return;
1058 
1059 	for (i = 0; turbostat_pdata[i].features; i++) {
1060 		if (VFM_FAMILY(turbostat_pdata[i].vfm) == family && VFM_MODEL(turbostat_pdata[i].vfm) == model) {
1061 			platform = turbostat_pdata[i].features;
1062 			return;
1063 		}
1064 	}
1065 }
1066 
1067 /* Model specific support End */
1068 
1069 #define	TJMAX_DEFAULT	100
1070 
1071 /* MSRs that are not yet in the kernel-provided header. */
1072 #define MSR_RAPL_PWR_UNIT	0xc0010299
1073 #define MSR_CORE_ENERGY_STAT	0xc001029a
1074 #define MSR_PKG_ENERGY_STAT	0xc001029b
1075 
1076 #define MAX(a, b) ((a) > (b) ? (a) : (b))
1077 
1078 int backwards_count;
1079 char *progname;
1080 
1081 #define CPU_SUBSET_MAXCPUS	1024	/* need to use before probe... */
1082 cpu_set_t *cpu_present_set, *cpu_effective_set, *cpu_allowed_set, *cpu_affinity_set, *cpu_subset;
1083 size_t cpu_present_setsize, cpu_effective_setsize, cpu_allowed_setsize, cpu_affinity_setsize, cpu_subset_size;
1084 #define MAX_ADDED_THREAD_COUNTERS 24
1085 #define MAX_ADDED_CORE_COUNTERS 8
1086 #define MAX_ADDED_PACKAGE_COUNTERS 16
1087 #define PMT_MAX_ADDED_THREAD_COUNTERS 24
1088 #define PMT_MAX_ADDED_CORE_COUNTERS 8
1089 #define PMT_MAX_ADDED_PACKAGE_COUNTERS 16
1090 #define BITMASK_SIZE 32
1091 
1092 #define ZERO_ARRAY(arr) (memset(arr, 0, sizeof(arr)) + __must_be_array(arr))
1093 
1094 /* Indexes used to map data read from perf and MSRs into global variables */
1095 enum rapl_rci_index {
1096 	RAPL_RCI_INDEX_ENERGY_PKG = 0,
1097 	RAPL_RCI_INDEX_ENERGY_CORES = 1,
1098 	RAPL_RCI_INDEX_DRAM = 2,
1099 	RAPL_RCI_INDEX_GFX = 3,
1100 	RAPL_RCI_INDEX_PKG_PERF_STATUS = 4,
1101 	RAPL_RCI_INDEX_DRAM_PERF_STATUS = 5,
1102 	RAPL_RCI_INDEX_CORE_ENERGY = 6,
1103 	NUM_RAPL_COUNTERS,
1104 };
1105 
1106 enum rapl_unit {
1107 	RAPL_UNIT_INVALID,
1108 	RAPL_UNIT_JOULES,
1109 	RAPL_UNIT_WATTS,
1110 };
1111 
1112 struct rapl_counter_info_t {
1113 	unsigned long long data[NUM_RAPL_COUNTERS];
1114 	enum counter_source source[NUM_RAPL_COUNTERS];
1115 	unsigned long long flags[NUM_RAPL_COUNTERS];
1116 	double scale[NUM_RAPL_COUNTERS];
1117 	enum rapl_unit unit[NUM_RAPL_COUNTERS];
1118 	unsigned long long msr[NUM_RAPL_COUNTERS];
1119 	unsigned long long msr_mask[NUM_RAPL_COUNTERS];
1120 	int msr_shift[NUM_RAPL_COUNTERS];
1121 
1122 	int fd_perf;
1123 };
1124 
1125 /* struct rapl_counter_info_t for each RAPL domain */
1126 struct rapl_counter_info_t *rapl_counter_info_perdomain;
1127 unsigned int rapl_counter_info_perdomain_size;
1128 
1129 #define RAPL_COUNTER_FLAG_USE_MSR_SUM (1u << 1)
1130 
1131 struct rapl_counter_arch_info {
1132 	int feature_mask;	/* Mask for testing if the counter is supported on host */
1133 	const char *perf_subsys;
1134 	const char *perf_name;
1135 	unsigned long long msr;
1136 	unsigned long long msr_mask;
1137 	int msr_shift;		/* Positive mean shift right, negative mean shift left */
1138 	double *platform_rapl_msr_scale;	/* Scale applied to values read by MSR (platform dependent, filled at runtime) */
1139 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1140 	unsigned long long bic;
1141 	double compat_scale;	/* Some counters require constant scaling to be in the same range as other, similar ones */
1142 	unsigned long long flags;
1143 };
1144 
1145 static const struct rapl_counter_arch_info rapl_counter_arch_infos[] = {
1146 	{
1147 	 .feature_mask = RAPL_PKG,
1148 	 .perf_subsys = "power",
1149 	 .perf_name = "energy-pkg",
1150 	 .msr = MSR_PKG_ENERGY_STATUS,
1151 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1152 	 .msr_shift = 0,
1153 	 .platform_rapl_msr_scale = &rapl_energy_units,
1154 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1155 	 .bic = BIC_PkgWatt | BIC_Pkg_J,
1156 	 .compat_scale = 1.0,
1157 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1158 	  },
1159 	{
1160 	 .feature_mask = RAPL_AMD_F17H,
1161 	 .perf_subsys = "power",
1162 	 .perf_name = "energy-pkg",
1163 	 .msr = MSR_PKG_ENERGY_STAT,
1164 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1165 	 .msr_shift = 0,
1166 	 .platform_rapl_msr_scale = &rapl_energy_units,
1167 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1168 	 .bic = BIC_PkgWatt | BIC_Pkg_J,
1169 	 .compat_scale = 1.0,
1170 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1171 	  },
1172 	{
1173 	 .feature_mask = RAPL_CORE_ENERGY_STATUS,
1174 	 .perf_subsys = "power",
1175 	 .perf_name = "energy-cores",
1176 	 .msr = MSR_PP0_ENERGY_STATUS,
1177 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1178 	 .msr_shift = 0,
1179 	 .platform_rapl_msr_scale = &rapl_energy_units,
1180 	 .rci_index = RAPL_RCI_INDEX_ENERGY_CORES,
1181 	 .bic = BIC_CorWatt | BIC_Cor_J,
1182 	 .compat_scale = 1.0,
1183 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1184 	  },
1185 	{
1186 	 .feature_mask = RAPL_DRAM,
1187 	 .perf_subsys = "power",
1188 	 .perf_name = "energy-ram",
1189 	 .msr = MSR_DRAM_ENERGY_STATUS,
1190 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1191 	 .msr_shift = 0,
1192 	 .platform_rapl_msr_scale = &rapl_dram_energy_units,
1193 	 .rci_index = RAPL_RCI_INDEX_DRAM,
1194 	 .bic = BIC_RAMWatt | BIC_RAM_J,
1195 	 .compat_scale = 1.0,
1196 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1197 	  },
1198 	{
1199 	 .feature_mask = RAPL_GFX,
1200 	 .perf_subsys = "power",
1201 	 .perf_name = "energy-gpu",
1202 	 .msr = MSR_PP1_ENERGY_STATUS,
1203 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1204 	 .msr_shift = 0,
1205 	 .platform_rapl_msr_scale = &rapl_energy_units,
1206 	 .rci_index = RAPL_RCI_INDEX_GFX,
1207 	 .bic = BIC_GFXWatt | BIC_GFX_J,
1208 	 .compat_scale = 1.0,
1209 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1210 	  },
1211 	{
1212 	 .feature_mask = RAPL_PKG_PERF_STATUS,
1213 	 .perf_subsys = NULL,
1214 	 .perf_name = NULL,
1215 	 .msr = MSR_PKG_PERF_STATUS,
1216 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1217 	 .msr_shift = 0,
1218 	 .platform_rapl_msr_scale = &rapl_time_units,
1219 	 .rci_index = RAPL_RCI_INDEX_PKG_PERF_STATUS,
1220 	 .bic = BIC_PKG__,
1221 	 .compat_scale = 100.0,
1222 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1223 	  },
1224 	{
1225 	 .feature_mask = RAPL_DRAM_PERF_STATUS,
1226 	 .perf_subsys = NULL,
1227 	 .perf_name = NULL,
1228 	 .msr = MSR_DRAM_PERF_STATUS,
1229 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1230 	 .msr_shift = 0,
1231 	 .platform_rapl_msr_scale = &rapl_time_units,
1232 	 .rci_index = RAPL_RCI_INDEX_DRAM_PERF_STATUS,
1233 	 .bic = BIC_RAM__,
1234 	 .compat_scale = 100.0,
1235 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1236 	  },
1237 	{
1238 	 .feature_mask = RAPL_AMD_F17H,
1239 	 .perf_subsys = NULL,
1240 	 .perf_name = NULL,
1241 	 .msr = MSR_CORE_ENERGY_STAT,
1242 	 .msr_mask = 0xFFFFFFFF,
1243 	 .msr_shift = 0,
1244 	 .platform_rapl_msr_scale = &rapl_energy_units,
1245 	 .rci_index = RAPL_RCI_INDEX_CORE_ENERGY,
1246 	 .bic = BIC_CorWatt | BIC_Cor_J,
1247 	 .compat_scale = 1.0,
1248 	 .flags = 0,
1249 	  },
1250 };
1251 
1252 struct rapl_counter {
1253 	unsigned long long raw_value;
1254 	enum rapl_unit unit;
1255 	double scale;
1256 };
1257 
1258 /* Indexes used to map data read from perf and MSRs into global variables */
1259 enum ccstate_rci_index {
1260 	CCSTATE_RCI_INDEX_C1_RESIDENCY = 0,
1261 	CCSTATE_RCI_INDEX_C3_RESIDENCY = 1,
1262 	CCSTATE_RCI_INDEX_C6_RESIDENCY = 2,
1263 	CCSTATE_RCI_INDEX_C7_RESIDENCY = 3,
1264 	PCSTATE_RCI_INDEX_C2_RESIDENCY = 4,
1265 	PCSTATE_RCI_INDEX_C3_RESIDENCY = 5,
1266 	PCSTATE_RCI_INDEX_C6_RESIDENCY = 6,
1267 	PCSTATE_RCI_INDEX_C7_RESIDENCY = 7,
1268 	PCSTATE_RCI_INDEX_C8_RESIDENCY = 8,
1269 	PCSTATE_RCI_INDEX_C9_RESIDENCY = 9,
1270 	PCSTATE_RCI_INDEX_C10_RESIDENCY = 10,
1271 	NUM_CSTATE_COUNTERS,
1272 };
1273 
1274 struct cstate_counter_info_t {
1275 	unsigned long long data[NUM_CSTATE_COUNTERS];
1276 	enum counter_source source[NUM_CSTATE_COUNTERS];
1277 	unsigned long long msr[NUM_CSTATE_COUNTERS];
1278 	int fd_perf_core;
1279 	int fd_perf_pkg;
1280 };
1281 
1282 struct cstate_counter_info_t *ccstate_counter_info;
1283 unsigned int ccstate_counter_info_size;
1284 
1285 #define CSTATE_COUNTER_FLAG_COLLECT_PER_CORE   (1u << 0)
1286 #define CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD ((1u << 1) | CSTATE_COUNTER_FLAG_COLLECT_PER_CORE)
1287 #define CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY (1u << 2)
1288 
1289 struct cstate_counter_arch_info {
1290 	int feature_mask;	/* Mask for testing if the counter is supported on host */
1291 	const char *perf_subsys;
1292 	const char *perf_name;
1293 	unsigned long long msr;
1294 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1295 	unsigned long long bic;
1296 	unsigned long long flags;
1297 	int pkg_cstate_limit;
1298 };
1299 
1300 static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
1301 	{
1302 	 .feature_mask = CC1,
1303 	 .perf_subsys = "cstate_core",
1304 	 .perf_name = "c1-residency",
1305 	 .msr = MSR_CORE_C1_RES,
1306 	 .rci_index = CCSTATE_RCI_INDEX_C1_RESIDENCY,
1307 	 .bic = BIC_CPU_c1,
1308 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD,
1309 	 .pkg_cstate_limit = 0,
1310 	  },
1311 	{
1312 	 .feature_mask = CC3,
1313 	 .perf_subsys = "cstate_core",
1314 	 .perf_name = "c3-residency",
1315 	 .msr = MSR_CORE_C3_RESIDENCY,
1316 	 .rci_index = CCSTATE_RCI_INDEX_C3_RESIDENCY,
1317 	 .bic = BIC_CPU_c3,
1318 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1319 	 .pkg_cstate_limit = 0,
1320 	  },
1321 	{
1322 	 .feature_mask = CC6,
1323 	 .perf_subsys = "cstate_core",
1324 	 .perf_name = "c6-residency",
1325 	 .msr = MSR_CORE_C6_RESIDENCY,
1326 	 .rci_index = CCSTATE_RCI_INDEX_C6_RESIDENCY,
1327 	 .bic = BIC_CPU_c6,
1328 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1329 	 .pkg_cstate_limit = 0,
1330 	  },
1331 	{
1332 	 .feature_mask = CC7,
1333 	 .perf_subsys = "cstate_core",
1334 	 .perf_name = "c7-residency",
1335 	 .msr = MSR_CORE_C7_RESIDENCY,
1336 	 .rci_index = CCSTATE_RCI_INDEX_C7_RESIDENCY,
1337 	 .bic = BIC_CPU_c7,
1338 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1339 	 .pkg_cstate_limit = 0,
1340 	  },
1341 	{
1342 	 .feature_mask = PC2,
1343 	 .perf_subsys = "cstate_pkg",
1344 	 .perf_name = "c2-residency",
1345 	 .msr = MSR_PKG_C2_RESIDENCY,
1346 	 .rci_index = PCSTATE_RCI_INDEX_C2_RESIDENCY,
1347 	 .bic = BIC_Pkgpc2,
1348 	 .flags = 0,
1349 	 .pkg_cstate_limit = PCL__2,
1350 	  },
1351 	{
1352 	 .feature_mask = PC3,
1353 	 .perf_subsys = "cstate_pkg",
1354 	 .perf_name = "c3-residency",
1355 	 .msr = MSR_PKG_C3_RESIDENCY,
1356 	 .rci_index = PCSTATE_RCI_INDEX_C3_RESIDENCY,
1357 	 .bic = BIC_Pkgpc3,
1358 	 .flags = 0,
1359 	 .pkg_cstate_limit = PCL__3,
1360 	  },
1361 	{
1362 	 .feature_mask = PC6,
1363 	 .perf_subsys = "cstate_pkg",
1364 	 .perf_name = "c6-residency",
1365 	 .msr = MSR_PKG_C6_RESIDENCY,
1366 	 .rci_index = PCSTATE_RCI_INDEX_C6_RESIDENCY,
1367 	 .bic = BIC_Pkgpc6,
1368 	 .flags = 0,
1369 	 .pkg_cstate_limit = PCL__6,
1370 	  },
1371 	{
1372 	 .feature_mask = PC7,
1373 	 .perf_subsys = "cstate_pkg",
1374 	 .perf_name = "c7-residency",
1375 	 .msr = MSR_PKG_C7_RESIDENCY,
1376 	 .rci_index = PCSTATE_RCI_INDEX_C7_RESIDENCY,
1377 	 .bic = BIC_Pkgpc7,
1378 	 .flags = 0,
1379 	 .pkg_cstate_limit = PCL__7,
1380 	  },
1381 	{
1382 	 .feature_mask = PC8,
1383 	 .perf_subsys = "cstate_pkg",
1384 	 .perf_name = "c8-residency",
1385 	 .msr = MSR_PKG_C8_RESIDENCY,
1386 	 .rci_index = PCSTATE_RCI_INDEX_C8_RESIDENCY,
1387 	 .bic = BIC_Pkgpc8,
1388 	 .flags = 0,
1389 	 .pkg_cstate_limit = PCL__8,
1390 	  },
1391 	{
1392 	 .feature_mask = PC9,
1393 	 .perf_subsys = "cstate_pkg",
1394 	 .perf_name = "c9-residency",
1395 	 .msr = MSR_PKG_C9_RESIDENCY,
1396 	 .rci_index = PCSTATE_RCI_INDEX_C9_RESIDENCY,
1397 	 .bic = BIC_Pkgpc9,
1398 	 .flags = 0,
1399 	 .pkg_cstate_limit = PCL__9,
1400 	  },
1401 	{
1402 	 .feature_mask = PC10,
1403 	 .perf_subsys = "cstate_pkg",
1404 	 .perf_name = "c10-residency",
1405 	 .msr = MSR_PKG_C10_RESIDENCY,
1406 	 .rci_index = PCSTATE_RCI_INDEX_C10_RESIDENCY,
1407 	 .bic = BIC_Pkgpc10,
1408 	 .flags = 0,
1409 	 .pkg_cstate_limit = PCL_10,
1410 	  },
1411 };
1412 
1413 /* Indexes used to map data read from perf and MSRs into global variables */
1414 enum msr_rci_index {
1415 	MSR_RCI_INDEX_APERF = 0,
1416 	MSR_RCI_INDEX_MPERF = 1,
1417 	MSR_RCI_INDEX_SMI = 2,
1418 	NUM_MSR_COUNTERS,
1419 };
1420 
1421 struct msr_counter_info_t {
1422 	unsigned long long data[NUM_MSR_COUNTERS];
1423 	enum counter_source source[NUM_MSR_COUNTERS];
1424 	unsigned long long msr[NUM_MSR_COUNTERS];
1425 	unsigned long long msr_mask[NUM_MSR_COUNTERS];
1426 	int fd_perf;
1427 };
1428 
1429 struct msr_counter_info_t *msr_counter_info;
1430 unsigned int msr_counter_info_size;
1431 
1432 struct msr_counter_arch_info {
1433 	const char *perf_subsys;
1434 	const char *perf_name;
1435 	unsigned long long msr;
1436 	unsigned long long msr_mask;
1437 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1438 	bool needed;
1439 	bool present;
1440 };
1441 
1442 enum msr_arch_info_index {
1443 	MSR_ARCH_INFO_APERF_INDEX = 0,
1444 	MSR_ARCH_INFO_MPERF_INDEX = 1,
1445 	MSR_ARCH_INFO_SMI_INDEX = 2,
1446 };
1447 
1448 static struct msr_counter_arch_info msr_counter_arch_infos[] = {
1449 	[MSR_ARCH_INFO_APERF_INDEX] = {
1450 				       .perf_subsys = "msr",
1451 				       .perf_name = "aperf",
1452 				       .msr = MSR_IA32_APERF,
1453 				       .msr_mask = 0xFFFFFFFFFFFFFFFF,
1454 				       .rci_index = MSR_RCI_INDEX_APERF,
1455 				        },
1456 
1457 	[MSR_ARCH_INFO_MPERF_INDEX] = {
1458 				       .perf_subsys = "msr",
1459 				       .perf_name = "mperf",
1460 				       .msr = MSR_IA32_MPERF,
1461 				       .msr_mask = 0xFFFFFFFFFFFFFFFF,
1462 				       .rci_index = MSR_RCI_INDEX_MPERF,
1463 				        },
1464 
1465 	[MSR_ARCH_INFO_SMI_INDEX] = {
1466 				     .perf_subsys = "msr",
1467 				     .perf_name = "smi",
1468 				     .msr = MSR_SMI_COUNT,
1469 				     .msr_mask = 0xFFFFFFFF,
1470 				     .rci_index = MSR_RCI_INDEX_SMI,
1471 				      },
1472 };
1473 
1474 /* Can be redefined when compiling, useful for testing. */
1475 #ifndef SYSFS_TELEM_PATH
1476 #define SYSFS_TELEM_PATH "/sys/class/intel_pmt"
1477 #endif
1478 
1479 #define PMT_COUNTER_MTL_DC6_OFFSET 120
1480 #define PMT_COUNTER_MTL_DC6_LSB    0
1481 #define PMT_COUNTER_MTL_DC6_MSB    63
1482 #define PMT_MTL_DC6_GUID           0x1a067102
1483 
1484 #define PMT_COUNTER_NAME_SIZE_BYTES      16
1485 #define PMT_COUNTER_TYPE_NAME_SIZE_BYTES 32
1486 
1487 struct pmt_mmio {
1488 	struct pmt_mmio *next;
1489 
1490 	unsigned int guid;
1491 	unsigned int size;
1492 
1493 	/* Base pointer to the mmaped memory. */
1494 	void *mmio_base;
1495 
1496 	/*
1497 	 * Offset to be applied to the mmio_base
1498 	 * to get the beginning of the PMT counters for given GUID.
1499 	 */
1500 	unsigned long pmt_offset;
1501 } *pmt_mmios;
1502 
1503 enum pmt_datatype {
1504 	PMT_TYPE_RAW,
1505 	PMT_TYPE_XTAL_TIME,
1506 };
1507 
1508 struct pmt_domain_info {
1509 	/*
1510 	 * Pointer to the MMIO obtained by applying a counter offset
1511 	 * to the mmio_base of the mmaped region for the given GUID.
1512 	 *
1513 	 * This is where to read the raw value of the counter from.
1514 	 */
1515 	unsigned long *pcounter;
1516 };
1517 
1518 struct pmt_counter {
1519 	struct pmt_counter *next;
1520 
1521 	/* PMT metadata */
1522 	char name[PMT_COUNTER_NAME_SIZE_BYTES];
1523 	enum pmt_datatype type;
1524 	enum counter_scope scope;
1525 	unsigned int lsb;
1526 	unsigned int msb;
1527 
1528 	/* BIC-like metadata */
1529 	enum counter_format format;
1530 
1531 	unsigned int num_domains;
1532 	struct pmt_domain_info *domains;
1533 };
1534 
1535 unsigned int pmt_counter_get_width(const struct pmt_counter *p)
1536 {
1537 	return (p->msb - p->lsb) + 1;
1538 }
1539 
1540 void pmt_counter_resize_(struct pmt_counter *pcounter, unsigned int new_size)
1541 {
1542 	struct pmt_domain_info *new_mem;
1543 
1544 	new_mem = (struct pmt_domain_info *)reallocarray(pcounter->domains, new_size, sizeof(*pcounter->domains));
1545 	if (!new_mem) {
1546 		fprintf(stderr, "%s: failed to allocate memory for PMT counters\n", __func__);
1547 		exit(1);
1548 	}
1549 
1550 	/* Zero initialize just allocated memory. */
1551 	const size_t num_new_domains = new_size - pcounter->num_domains;
1552 
1553 	memset(&new_mem[pcounter->num_domains], 0, num_new_domains * sizeof(*pcounter->domains));
1554 
1555 	pcounter->num_domains = new_size;
1556 	pcounter->domains = new_mem;
1557 }
1558 
1559 void pmt_counter_resize(struct pmt_counter *pcounter, unsigned int new_size)
1560 {
1561 	/*
1562 	 * Allocate more memory ahead of time.
1563 	 *
1564 	 * Always allocate space for at least 8 elements
1565 	 * and double the size when growing.
1566 	 */
1567 	if (new_size < 8)
1568 		new_size = 8;
1569 	new_size = MAX(new_size, pcounter->num_domains * 2);
1570 
1571 	pmt_counter_resize_(pcounter, new_size);
1572 }
1573 
1574 struct thread_data {
1575 	struct timeval tv_begin;
1576 	struct timeval tv_end;
1577 	struct timeval tv_delta;
1578 	unsigned long long tsc;
1579 	unsigned long long aperf;
1580 	unsigned long long mperf;
1581 	unsigned long long c1;
1582 	unsigned long long instr_count;
1583 	unsigned long long irq_count;
1584 	unsigned int smi_count;
1585 	unsigned int cpu_id;
1586 	unsigned int apic_id;
1587 	unsigned int x2apic_id;
1588 	unsigned int flags;
1589 	bool is_atom;
1590 	unsigned long long counter[MAX_ADDED_THREAD_COUNTERS];
1591 	unsigned long long perf_counter[MAX_ADDED_THREAD_COUNTERS];
1592 	unsigned long long pmt_counter[PMT_MAX_ADDED_THREAD_COUNTERS];
1593 } *thread_even, *thread_odd;
1594 
1595 struct core_data {
1596 	int base_cpu;
1597 	unsigned long long c3;
1598 	unsigned long long c6;
1599 	unsigned long long c7;
1600 	unsigned long long mc6_us;	/* duplicate as per-core for now, even though per module */
1601 	unsigned int core_temp_c;
1602 	struct rapl_counter core_energy;	/* MSR_CORE_ENERGY_STAT */
1603 	unsigned int core_id;
1604 	unsigned long long core_throt_cnt;
1605 	unsigned long long counter[MAX_ADDED_CORE_COUNTERS];
1606 	unsigned long long perf_counter[MAX_ADDED_CORE_COUNTERS];
1607 	unsigned long long pmt_counter[PMT_MAX_ADDED_CORE_COUNTERS];
1608 } *core_even, *core_odd;
1609 
1610 struct pkg_data {
1611 	int base_cpu;
1612 	unsigned long long pc2;
1613 	unsigned long long pc3;
1614 	unsigned long long pc6;
1615 	unsigned long long pc7;
1616 	unsigned long long pc8;
1617 	unsigned long long pc9;
1618 	unsigned long long pc10;
1619 	long long cpu_lpi;
1620 	long long sys_lpi;
1621 	unsigned long long pkg_wtd_core_c0;
1622 	unsigned long long pkg_any_core_c0;
1623 	unsigned long long pkg_any_gfxe_c0;
1624 	unsigned long long pkg_both_core_gfxe_c0;
1625 	long long gfx_rc6_ms;
1626 	unsigned int gfx_mhz;
1627 	unsigned int gfx_act_mhz;
1628 	long long sam_mc6_ms;
1629 	unsigned int sam_mhz;
1630 	unsigned int sam_act_mhz;
1631 	unsigned int package_id;
1632 	struct rapl_counter energy_pkg;	/* MSR_PKG_ENERGY_STATUS */
1633 	struct rapl_counter energy_dram;	/* MSR_DRAM_ENERGY_STATUS */
1634 	struct rapl_counter energy_cores;	/* MSR_PP0_ENERGY_STATUS */
1635 	struct rapl_counter energy_gfx;	/* MSR_PP1_ENERGY_STATUS */
1636 	struct rapl_counter rapl_pkg_perf_status;	/* MSR_PKG_PERF_STATUS */
1637 	struct rapl_counter rapl_dram_perf_status;	/* MSR_DRAM_PERF_STATUS */
1638 	unsigned int pkg_temp_c;
1639 	unsigned int uncore_mhz;
1640 	unsigned long long die_c6;
1641 	unsigned long long counter[MAX_ADDED_PACKAGE_COUNTERS];
1642 	unsigned long long perf_counter[MAX_ADDED_PACKAGE_COUNTERS];
1643 	unsigned long long pmt_counter[PMT_MAX_ADDED_PACKAGE_COUNTERS];
1644 } *package_even, *package_odd;
1645 
1646 #define ODD_COUNTERS thread_odd, core_odd, package_odd
1647 #define EVEN_COUNTERS thread_even, core_even, package_even
1648 
1649 #define GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no)	      \
1650 	((thread_base) +						      \
1651 	 ((pkg_no) *							      \
1652 	  topo.nodes_per_pkg * topo.cores_per_node * topo.threads_per_core) + \
1653 	 ((node_no) * topo.cores_per_node * topo.threads_per_core) +	      \
1654 	 ((core_no) * topo.threads_per_core) +				      \
1655 	 (thread_no))
1656 
1657 #define GET_CORE(core_base, core_no, node_no, pkg_no)			\
1658 	((core_base) +							\
1659 	 ((pkg_no) *  topo.nodes_per_pkg * topo.cores_per_node) +	\
1660 	 ((node_no) * topo.cores_per_node) +				\
1661 	 (core_no))
1662 
1663 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
1664 
1665 /*
1666  * The accumulated sum of MSR is defined as a monotonic
1667  * increasing MSR, it will be accumulated periodically,
1668  * despite its register's bit width.
1669  */
1670 enum {
1671 	IDX_PKG_ENERGY,
1672 	IDX_DRAM_ENERGY,
1673 	IDX_PP0_ENERGY,
1674 	IDX_PP1_ENERGY,
1675 	IDX_PKG_PERF,
1676 	IDX_DRAM_PERF,
1677 	IDX_COUNT,
1678 };
1679 
1680 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr);
1681 
1682 struct msr_sum_array {
1683 	/* get_msr_sum() = sum + (get_msr() - last) */
1684 	struct {
1685 		/*The accumulated MSR value is updated by the timer */
1686 		unsigned long long sum;
1687 		/*The MSR footprint recorded in last timer */
1688 		unsigned long long last;
1689 	} entries[IDX_COUNT];
1690 };
1691 
1692 /* The percpu MSR sum array.*/
1693 struct msr_sum_array *per_cpu_msr_sum;
1694 
1695 off_t idx_to_offset(int idx)
1696 {
1697 	off_t offset;
1698 
1699 	switch (idx) {
1700 	case IDX_PKG_ENERGY:
1701 		if (platform->rapl_msrs & RAPL_AMD_F17H)
1702 			offset = MSR_PKG_ENERGY_STAT;
1703 		else
1704 			offset = MSR_PKG_ENERGY_STATUS;
1705 		break;
1706 	case IDX_DRAM_ENERGY:
1707 		offset = MSR_DRAM_ENERGY_STATUS;
1708 		break;
1709 	case IDX_PP0_ENERGY:
1710 		offset = MSR_PP0_ENERGY_STATUS;
1711 		break;
1712 	case IDX_PP1_ENERGY:
1713 		offset = MSR_PP1_ENERGY_STATUS;
1714 		break;
1715 	case IDX_PKG_PERF:
1716 		offset = MSR_PKG_PERF_STATUS;
1717 		break;
1718 	case IDX_DRAM_PERF:
1719 		offset = MSR_DRAM_PERF_STATUS;
1720 		break;
1721 	default:
1722 		offset = -1;
1723 	}
1724 	return offset;
1725 }
1726 
1727 int offset_to_idx(off_t offset)
1728 {
1729 	int idx;
1730 
1731 	switch (offset) {
1732 	case MSR_PKG_ENERGY_STATUS:
1733 	case MSR_PKG_ENERGY_STAT:
1734 		idx = IDX_PKG_ENERGY;
1735 		break;
1736 	case MSR_DRAM_ENERGY_STATUS:
1737 		idx = IDX_DRAM_ENERGY;
1738 		break;
1739 	case MSR_PP0_ENERGY_STATUS:
1740 		idx = IDX_PP0_ENERGY;
1741 		break;
1742 	case MSR_PP1_ENERGY_STATUS:
1743 		idx = IDX_PP1_ENERGY;
1744 		break;
1745 	case MSR_PKG_PERF_STATUS:
1746 		idx = IDX_PKG_PERF;
1747 		break;
1748 	case MSR_DRAM_PERF_STATUS:
1749 		idx = IDX_DRAM_PERF;
1750 		break;
1751 	default:
1752 		idx = -1;
1753 	}
1754 	return idx;
1755 }
1756 
1757 int idx_valid(int idx)
1758 {
1759 	switch (idx) {
1760 	case IDX_PKG_ENERGY:
1761 		return platform->rapl_msrs & (RAPL_PKG | RAPL_AMD_F17H);
1762 	case IDX_DRAM_ENERGY:
1763 		return platform->rapl_msrs & RAPL_DRAM;
1764 	case IDX_PP0_ENERGY:
1765 		return platform->rapl_msrs & RAPL_CORE_ENERGY_STATUS;
1766 	case IDX_PP1_ENERGY:
1767 		return platform->rapl_msrs & RAPL_GFX;
1768 	case IDX_PKG_PERF:
1769 		return platform->rapl_msrs & RAPL_PKG_PERF_STATUS;
1770 	case IDX_DRAM_PERF:
1771 		return platform->rapl_msrs & RAPL_DRAM_PERF_STATUS;
1772 	default:
1773 		return 0;
1774 	}
1775 }
1776 
1777 struct sys_counters {
1778 	/* MSR added counters */
1779 	unsigned int added_thread_counters;
1780 	unsigned int added_core_counters;
1781 	unsigned int added_package_counters;
1782 	struct msr_counter *tp;
1783 	struct msr_counter *cp;
1784 	struct msr_counter *pp;
1785 
1786 	/* perf added counters */
1787 	unsigned int added_thread_perf_counters;
1788 	unsigned int added_core_perf_counters;
1789 	unsigned int added_package_perf_counters;
1790 	struct perf_counter_info *perf_tp;
1791 	struct perf_counter_info *perf_cp;
1792 	struct perf_counter_info *perf_pp;
1793 
1794 	struct pmt_counter *pmt_tp;
1795 	struct pmt_counter *pmt_cp;
1796 	struct pmt_counter *pmt_pp;
1797 } sys;
1798 
1799 static size_t free_msr_counters_(struct msr_counter **pp)
1800 {
1801 	struct msr_counter *p = NULL;
1802 	size_t num_freed = 0;
1803 
1804 	while (*pp) {
1805 		p = *pp;
1806 
1807 		if (p->msr_num != 0) {
1808 			*pp = p->next;
1809 
1810 			free(p);
1811 			++num_freed;
1812 
1813 			continue;
1814 		}
1815 
1816 		pp = &p->next;
1817 	}
1818 
1819 	return num_freed;
1820 }
1821 
1822 /*
1823  * Free all added counters accessed via msr.
1824  */
1825 static void free_sys_msr_counters(void)
1826 {
1827 	/* Thread counters */
1828 	sys.added_thread_counters -= free_msr_counters_(&sys.tp);
1829 
1830 	/* Core counters */
1831 	sys.added_core_counters -= free_msr_counters_(&sys.cp);
1832 
1833 	/* Package counters */
1834 	sys.added_package_counters -= free_msr_counters_(&sys.pp);
1835 }
1836 
1837 struct system_summary {
1838 	struct thread_data threads;
1839 	struct core_data cores;
1840 	struct pkg_data packages;
1841 } average;
1842 
1843 struct cpu_topology {
1844 	int physical_package_id;
1845 	int die_id;
1846 	int logical_cpu_id;
1847 	int physical_node_id;
1848 	int logical_node_id;	/* 0-based count within the package */
1849 	int physical_core_id;
1850 	int thread_id;
1851 	cpu_set_t *put_ids;	/* Processing Unit/Thread IDs */
1852 } *cpus;
1853 
1854 struct topo_params {
1855 	int num_packages;
1856 	int num_die;
1857 	int num_cpus;
1858 	int num_cores;
1859 	int allowed_packages;
1860 	int allowed_cpus;
1861 	int allowed_cores;
1862 	int max_cpu_num;
1863 	int max_core_id;
1864 	int max_package_id;
1865 	int max_die_id;
1866 	int max_node_num;
1867 	int nodes_per_pkg;
1868 	int cores_per_node;
1869 	int threads_per_core;
1870 } topo;
1871 
1872 struct timeval tv_even, tv_odd, tv_delta;
1873 
1874 int *irq_column_2_cpu;		/* /proc/interrupts column numbers */
1875 int *irqs_per_cpu;		/* indexed by cpu_num */
1876 
1877 void setup_all_buffers(bool startup);
1878 
1879 char *sys_lpi_file;
1880 char *sys_lpi_file_sysfs = "/sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us";
1881 char *sys_lpi_file_debugfs = "/sys/kernel/debug/pmc_core/slp_s0_residency_usec";
1882 
1883 int cpu_is_not_present(int cpu)
1884 {
1885 	return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
1886 }
1887 
1888 int cpu_is_not_allowed(int cpu)
1889 {
1890 	return !CPU_ISSET_S(cpu, cpu_allowed_setsize, cpu_allowed_set);
1891 }
1892 
1893 /*
1894  * run func(thread, core, package) in topology order
1895  * skip non-present cpus
1896  */
1897 
1898 int for_all_cpus(int (func) (struct thread_data *, struct core_data *, struct pkg_data *),
1899 		 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
1900 {
1901 	int retval, pkg_no, core_no, thread_no, node_no;
1902 
1903 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1904 		for (node_no = 0; node_no < topo.nodes_per_pkg; node_no++) {
1905 			for (core_no = 0; core_no < topo.cores_per_node; ++core_no) {
1906 				for (thread_no = 0; thread_no < topo.threads_per_core; ++thread_no) {
1907 					struct thread_data *t;
1908 					struct core_data *c;
1909 					struct pkg_data *p;
1910 					t = GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no);
1911 
1912 					if (cpu_is_not_allowed(t->cpu_id))
1913 						continue;
1914 
1915 					c = GET_CORE(core_base, core_no, node_no, pkg_no);
1916 					p = GET_PKG(pkg_base, pkg_no);
1917 
1918 					retval = func(t, c, p);
1919 					if (retval)
1920 						return retval;
1921 				}
1922 			}
1923 		}
1924 	}
1925 	return 0;
1926 }
1927 
1928 int is_cpu_first_thread_in_core(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1929 {
1930 	UNUSED(p);
1931 
1932 	return ((int)t->cpu_id == c->base_cpu || c->base_cpu < 0);
1933 }
1934 
1935 int is_cpu_first_core_in_package(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1936 {
1937 	UNUSED(c);
1938 
1939 	return ((int)t->cpu_id == p->base_cpu || p->base_cpu < 0);
1940 }
1941 
1942 int is_cpu_first_thread_in_package(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1943 {
1944 	return is_cpu_first_thread_in_core(t, c, p) && is_cpu_first_core_in_package(t, c, p);
1945 }
1946 
1947 int cpu_migrate(int cpu)
1948 {
1949 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1950 	CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
1951 	if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
1952 		return -1;
1953 	else
1954 		return 0;
1955 }
1956 
1957 int get_msr_fd(int cpu)
1958 {
1959 	char pathname[32];
1960 	int fd;
1961 
1962 	fd = fd_percpu[cpu];
1963 
1964 	if (fd)
1965 		return fd;
1966 
1967 	sprintf(pathname, "/dev/cpu/%d/msr", cpu);
1968 	fd = open(pathname, O_RDONLY);
1969 	if (fd < 0)
1970 		err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, "
1971 		    "or run with --no-msr, or run as root", pathname);
1972 
1973 	fd_percpu[cpu] = fd;
1974 
1975 	return fd;
1976 }
1977 
1978 static void bic_disable_msr_access(void)
1979 {
1980 	const unsigned long bic_msrs = BIC_Mod_c6 | BIC_CoreTmp |
1981 	    BIC_Totl_c0 | BIC_Any_c0 | BIC_GFX_c0 | BIC_CPUGFX | BIC_PkgTmp;
1982 
1983 	bic_enabled &= ~bic_msrs;
1984 
1985 	free_sys_msr_counters();
1986 }
1987 
1988 static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid, int cpu, int group_fd, unsigned long flags)
1989 {
1990 	assert(!no_perf);
1991 
1992 	return syscall(__NR_perf_event_open, hw_event, pid, cpu, group_fd, flags);
1993 }
1994 
1995 static long open_perf_counter(int cpu, unsigned int type, unsigned int config, int group_fd, __u64 read_format)
1996 {
1997 	struct perf_event_attr attr;
1998 	const pid_t pid = -1;
1999 	const unsigned long flags = 0;
2000 
2001 	assert(!no_perf);
2002 
2003 	memset(&attr, 0, sizeof(struct perf_event_attr));
2004 
2005 	attr.type = type;
2006 	attr.size = sizeof(struct perf_event_attr);
2007 	attr.config = config;
2008 	attr.disabled = 0;
2009 	attr.sample_type = PERF_SAMPLE_IDENTIFIER;
2010 	attr.read_format = read_format;
2011 
2012 	const int fd = perf_event_open(&attr, pid, cpu, group_fd, flags);
2013 
2014 	return fd;
2015 }
2016 
2017 int get_instr_count_fd(int cpu)
2018 {
2019 	if (fd_instr_count_percpu[cpu])
2020 		return fd_instr_count_percpu[cpu];
2021 
2022 	fd_instr_count_percpu[cpu] = open_perf_counter(cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, -1, 0);
2023 
2024 	return fd_instr_count_percpu[cpu];
2025 }
2026 
2027 int get_msr(int cpu, off_t offset, unsigned long long *msr)
2028 {
2029 	ssize_t retval;
2030 
2031 	assert(!no_msr);
2032 
2033 	retval = pread(get_msr_fd(cpu), msr, sizeof(*msr), offset);
2034 
2035 	if (retval != sizeof *msr)
2036 		err(-1, "cpu%d: msr offset 0x%llx read failed", cpu, (unsigned long long)offset);
2037 
2038 	return 0;
2039 }
2040 
2041 int probe_msr(int cpu, off_t offset)
2042 {
2043 	ssize_t retval;
2044 	unsigned long long dummy;
2045 
2046 	assert(!no_msr);
2047 
2048 	retval = pread(get_msr_fd(cpu), &dummy, sizeof(dummy), offset);
2049 
2050 	if (retval != sizeof(dummy))
2051 		return 1;
2052 
2053 	return 0;
2054 }
2055 
2056 /* Convert CPU ID to domain ID for given added perf counter. */
2057 unsigned int cpu_to_domain(const struct perf_counter_info *pc, int cpu)
2058 {
2059 	switch (pc->scope) {
2060 	case SCOPE_CPU:
2061 		return cpu;
2062 
2063 	case SCOPE_CORE:
2064 		return cpus[cpu].physical_core_id;
2065 
2066 	case SCOPE_PACKAGE:
2067 		return cpus[cpu].physical_package_id;
2068 	}
2069 
2070 	__builtin_unreachable();
2071 }
2072 
2073 #define MAX_DEFERRED 16
2074 char *deferred_add_names[MAX_DEFERRED];
2075 char *deferred_skip_names[MAX_DEFERRED];
2076 int deferred_add_index;
2077 int deferred_skip_index;
2078 
2079 /*
2080  * HIDE_LIST - hide this list of counters, show the rest [default]
2081  * SHOW_LIST - show this list of counters, hide the rest
2082  */
2083 enum show_hide_mode { SHOW_LIST, HIDE_LIST } global_show_hide_mode = HIDE_LIST;
2084 
2085 void help(void)
2086 {
2087 	fprintf(outf,
2088 		"Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
2089 		"\n"
2090 		"Turbostat forks the specified COMMAND and prints statistics\n"
2091 		"when COMMAND completes.\n"
2092 		"If no COMMAND is specified, turbostat wakes every 5-seconds\n"
2093 		"to print statistics, until interrupted.\n"
2094 		"  -a, --add	add a counter\n"
2095 		"		  eg. --add msr0x10,u64,cpu,delta,MY_TSC\n"
2096 		"		  eg. --add perf/cstate_pkg/c2-residency,package,delta,percent,perfPC2\n"
2097 		"		  eg. --add pmt,name=XTAL,type=raw,domain=package0,offset=0,lsb=0,msb=63,guid=0x1a067102\n"
2098 		"  -c, --cpu	cpu-set	limit output to summary plus cpu-set:\n"
2099 		"		  {core | package | j,k,l..m,n-p }\n"
2100 		"  -d, --debug	displays usec, Time_Of_Day_Seconds and more debugging\n"
2101 		"		debug messages are printed to stderr\n"
2102 		"  -D, --Dump	displays the raw counter values\n"
2103 		"  -e, --enable	[all | column]\n"
2104 		"		shows all or the specified disabled column\n"
2105 		"  -H, --hide [column|column,column,...]\n"
2106 		"		hide the specified column(s)\n"
2107 		"  -i, --interval sec.subsec\n"
2108 		"		Override default 5-second measurement interval\n"
2109 		"  -J, --Joules	displays energy in Joules instead of Watts\n"
2110 		"  -l, --list	list column headers only\n"
2111 		"  -M, --no-msr Disable all uses of the MSR driver\n"
2112 		"  -P, --no-perf Disable all uses of the perf API\n"
2113 		"  -n, --num_iterations num\n"
2114 		"		number of the measurement iterations\n"
2115 		"  -N, --header_iterations num\n"
2116 		"		print header every num iterations\n"
2117 		"  -o, --out file\n"
2118 		"		create or truncate \"file\" for all output\n"
2119 		"  -q, --quiet	skip decoding system configuration header\n"
2120 		"  -s, --show [column|column,column,...]\n"
2121 		"		show only the specified column(s)\n"
2122 		"  -S, --Summary\n"
2123 		"		limits output to 1-line system summary per interval\n"
2124 		"  -T, --TCC temperature\n"
2125 		"		sets the Thermal Control Circuit temperature in\n"
2126 		"		  degrees Celsius\n"
2127 		"  -h, --help	print this help message\n"
2128 		"  -v, --version	print version information\n" "\n" "For more help, run \"man turbostat\"\n");
2129 }
2130 
2131 /*
2132  * bic_lookup
2133  * for all the strings in comma separate name_list,
2134  * set the approprate bit in return value.
2135  */
2136 unsigned long long bic_lookup(char *name_list, enum show_hide_mode mode)
2137 {
2138 	unsigned int i;
2139 	unsigned long long retval = 0;
2140 
2141 	while (name_list) {
2142 		char *comma;
2143 
2144 		comma = strchr(name_list, ',');
2145 
2146 		if (comma)
2147 			*comma = '\0';
2148 
2149 		for (i = 0; i < MAX_BIC; ++i) {
2150 			if (!strcmp(name_list, bic[i].name)) {
2151 				retval |= (1ULL << i);
2152 				break;
2153 			}
2154 			if (!strcmp(name_list, "all")) {
2155 				retval |= ~0;
2156 				break;
2157 			} else if (!strcmp(name_list, "topology")) {
2158 				retval |= BIC_TOPOLOGY;
2159 				break;
2160 			} else if (!strcmp(name_list, "power")) {
2161 				retval |= BIC_THERMAL_PWR;
2162 				break;
2163 			} else if (!strcmp(name_list, "idle")) {
2164 				retval |= BIC_IDLE;
2165 				break;
2166 			} else if (!strcmp(name_list, "frequency")) {
2167 				retval |= BIC_FREQUENCY;
2168 				break;
2169 			} else if (!strcmp(name_list, "other")) {
2170 				retval |= BIC_OTHER;
2171 				break;
2172 			}
2173 
2174 		}
2175 		if (i == MAX_BIC) {
2176 			if (mode == SHOW_LIST) {
2177 				deferred_add_names[deferred_add_index++] = name_list;
2178 				if (deferred_add_index >= MAX_DEFERRED) {
2179 					fprintf(stderr, "More than max %d un-recognized --add options '%s'\n",
2180 						MAX_DEFERRED, name_list);
2181 					help();
2182 					exit(1);
2183 				}
2184 			} else {
2185 				deferred_skip_names[deferred_skip_index++] = name_list;
2186 				if (debug)
2187 					fprintf(stderr, "deferred \"%s\"\n", name_list);
2188 				if (deferred_skip_index >= MAX_DEFERRED) {
2189 					fprintf(stderr, "More than max %d un-recognized --skip options '%s'\n",
2190 						MAX_DEFERRED, name_list);
2191 					help();
2192 					exit(1);
2193 				}
2194 			}
2195 		}
2196 
2197 		name_list = comma;
2198 		if (name_list)
2199 			name_list++;
2200 
2201 	}
2202 	return retval;
2203 }
2204 
2205 void print_header(char *delim)
2206 {
2207 	struct msr_counter *mp;
2208 	struct perf_counter_info *pp;
2209 	struct pmt_counter *ppmt;
2210 	int printed = 0;
2211 
2212 	if (DO_BIC(BIC_USEC))
2213 		outp += sprintf(outp, "%susec", (printed++ ? delim : ""));
2214 	if (DO_BIC(BIC_TOD))
2215 		outp += sprintf(outp, "%sTime_Of_Day_Seconds", (printed++ ? delim : ""));
2216 	if (DO_BIC(BIC_Package))
2217 		outp += sprintf(outp, "%sPackage", (printed++ ? delim : ""));
2218 	if (DO_BIC(BIC_Die))
2219 		outp += sprintf(outp, "%sDie", (printed++ ? delim : ""));
2220 	if (DO_BIC(BIC_Node))
2221 		outp += sprintf(outp, "%sNode", (printed++ ? delim : ""));
2222 	if (DO_BIC(BIC_Core))
2223 		outp += sprintf(outp, "%sCore", (printed++ ? delim : ""));
2224 	if (DO_BIC(BIC_CPU))
2225 		outp += sprintf(outp, "%sCPU", (printed++ ? delim : ""));
2226 	if (DO_BIC(BIC_APIC))
2227 		outp += sprintf(outp, "%sAPIC", (printed++ ? delim : ""));
2228 	if (DO_BIC(BIC_X2APIC))
2229 		outp += sprintf(outp, "%sX2APIC", (printed++ ? delim : ""));
2230 	if (DO_BIC(BIC_Avg_MHz))
2231 		outp += sprintf(outp, "%sAvg_MHz", (printed++ ? delim : ""));
2232 	if (DO_BIC(BIC_Busy))
2233 		outp += sprintf(outp, "%sBusy%%", (printed++ ? delim : ""));
2234 	if (DO_BIC(BIC_Bzy_MHz))
2235 		outp += sprintf(outp, "%sBzy_MHz", (printed++ ? delim : ""));
2236 	if (DO_BIC(BIC_TSC_MHz))
2237 		outp += sprintf(outp, "%sTSC_MHz", (printed++ ? delim : ""));
2238 
2239 	if (DO_BIC(BIC_IPC))
2240 		outp += sprintf(outp, "%sIPC", (printed++ ? delim : ""));
2241 
2242 	if (DO_BIC(BIC_IRQ)) {
2243 		if (sums_need_wide_columns)
2244 			outp += sprintf(outp, "%s     IRQ", (printed++ ? delim : ""));
2245 		else
2246 			outp += sprintf(outp, "%sIRQ", (printed++ ? delim : ""));
2247 	}
2248 
2249 	if (DO_BIC(BIC_SMI))
2250 		outp += sprintf(outp, "%sSMI", (printed++ ? delim : ""));
2251 
2252 	for (mp = sys.tp; mp; mp = mp->next) {
2253 
2254 		if (mp->format == FORMAT_RAW) {
2255 			if (mp->width == 64)
2256 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), mp->name);
2257 			else
2258 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), mp->name);
2259 		} else {
2260 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2261 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), mp->name);
2262 			else
2263 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), mp->name);
2264 		}
2265 	}
2266 
2267 	for (pp = sys.perf_tp; pp; pp = pp->next) {
2268 
2269 		if (pp->format == FORMAT_RAW) {
2270 			if (pp->width == 64)
2271 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), pp->name);
2272 			else
2273 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), pp->name);
2274 		} else {
2275 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2276 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), pp->name);
2277 			else
2278 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), pp->name);
2279 		}
2280 	}
2281 
2282 	ppmt = sys.pmt_tp;
2283 	while (ppmt) {
2284 		switch (ppmt->type) {
2285 		case PMT_TYPE_RAW:
2286 			if (pmt_counter_get_width(ppmt) <= 32)
2287 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), ppmt->name);
2288 			else
2289 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), ppmt->name);
2290 
2291 			break;
2292 
2293 		case PMT_TYPE_XTAL_TIME:
2294 			outp += sprintf(outp, "%s%s", delim, ppmt->name);
2295 			break;
2296 		}
2297 
2298 		ppmt = ppmt->next;
2299 	}
2300 
2301 	if (DO_BIC(BIC_CPU_c1))
2302 		outp += sprintf(outp, "%sCPU%%c1", (printed++ ? delim : ""));
2303 	if (DO_BIC(BIC_CPU_c3))
2304 		outp += sprintf(outp, "%sCPU%%c3", (printed++ ? delim : ""));
2305 	if (DO_BIC(BIC_CPU_c6))
2306 		outp += sprintf(outp, "%sCPU%%c6", (printed++ ? delim : ""));
2307 	if (DO_BIC(BIC_CPU_c7))
2308 		outp += sprintf(outp, "%sCPU%%c7", (printed++ ? delim : ""));
2309 
2310 	if (DO_BIC(BIC_Mod_c6))
2311 		outp += sprintf(outp, "%sMod%%c6", (printed++ ? delim : ""));
2312 
2313 	if (DO_BIC(BIC_CoreTmp))
2314 		outp += sprintf(outp, "%sCoreTmp", (printed++ ? delim : ""));
2315 
2316 	if (DO_BIC(BIC_CORE_THROT_CNT))
2317 		outp += sprintf(outp, "%sCoreThr", (printed++ ? delim : ""));
2318 
2319 	if (platform->rapl_msrs && !rapl_joules) {
2320 		if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
2321 			outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
2322 	} else if (platform->rapl_msrs && rapl_joules) {
2323 		if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
2324 			outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
2325 	}
2326 
2327 	for (mp = sys.cp; mp; mp = mp->next) {
2328 		if (mp->format == FORMAT_RAW) {
2329 			if (mp->width == 64)
2330 				outp += sprintf(outp, "%s%18.18s", delim, mp->name);
2331 			else
2332 				outp += sprintf(outp, "%s%10.10s", delim, mp->name);
2333 		} else {
2334 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2335 				outp += sprintf(outp, "%s%8s", delim, mp->name);
2336 			else
2337 				outp += sprintf(outp, "%s%s", delim, mp->name);
2338 		}
2339 	}
2340 
2341 	for (pp = sys.perf_cp; pp; pp = pp->next) {
2342 
2343 		if (pp->format == FORMAT_RAW) {
2344 			if (pp->width == 64)
2345 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), pp->name);
2346 			else
2347 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), pp->name);
2348 		} else {
2349 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2350 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), pp->name);
2351 			else
2352 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), pp->name);
2353 		}
2354 	}
2355 
2356 	ppmt = sys.pmt_cp;
2357 	while (ppmt) {
2358 		switch (ppmt->type) {
2359 		case PMT_TYPE_RAW:
2360 			if (pmt_counter_get_width(ppmt) <= 32)
2361 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), ppmt->name);
2362 			else
2363 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), ppmt->name);
2364 
2365 			break;
2366 
2367 		case PMT_TYPE_XTAL_TIME:
2368 			outp += sprintf(outp, "%s%s", delim, ppmt->name);
2369 			break;
2370 		}
2371 
2372 		ppmt = ppmt->next;
2373 	}
2374 
2375 	if (DO_BIC(BIC_PkgTmp))
2376 		outp += sprintf(outp, "%sPkgTmp", (printed++ ? delim : ""));
2377 
2378 	if (DO_BIC(BIC_GFX_rc6))
2379 		outp += sprintf(outp, "%sGFX%%rc6", (printed++ ? delim : ""));
2380 
2381 	if (DO_BIC(BIC_GFXMHz))
2382 		outp += sprintf(outp, "%sGFXMHz", (printed++ ? delim : ""));
2383 
2384 	if (DO_BIC(BIC_GFXACTMHz))
2385 		outp += sprintf(outp, "%sGFXAMHz", (printed++ ? delim : ""));
2386 
2387 	if (DO_BIC(BIC_SAM_mc6))
2388 		outp += sprintf(outp, "%sSAM%%mc6", (printed++ ? delim : ""));
2389 
2390 	if (DO_BIC(BIC_SAMMHz))
2391 		outp += sprintf(outp, "%sSAMMHz", (printed++ ? delim : ""));
2392 
2393 	if (DO_BIC(BIC_SAMACTMHz))
2394 		outp += sprintf(outp, "%sSAMAMHz", (printed++ ? delim : ""));
2395 
2396 	if (DO_BIC(BIC_Totl_c0))
2397 		outp += sprintf(outp, "%sTotl%%C0", (printed++ ? delim : ""));
2398 	if (DO_BIC(BIC_Any_c0))
2399 		outp += sprintf(outp, "%sAny%%C0", (printed++ ? delim : ""));
2400 	if (DO_BIC(BIC_GFX_c0))
2401 		outp += sprintf(outp, "%sGFX%%C0", (printed++ ? delim : ""));
2402 	if (DO_BIC(BIC_CPUGFX))
2403 		outp += sprintf(outp, "%sCPUGFX%%", (printed++ ? delim : ""));
2404 
2405 	if (DO_BIC(BIC_Pkgpc2))
2406 		outp += sprintf(outp, "%sPkg%%pc2", (printed++ ? delim : ""));
2407 	if (DO_BIC(BIC_Pkgpc3))
2408 		outp += sprintf(outp, "%sPkg%%pc3", (printed++ ? delim : ""));
2409 	if (DO_BIC(BIC_Pkgpc6))
2410 		outp += sprintf(outp, "%sPkg%%pc6", (printed++ ? delim : ""));
2411 	if (DO_BIC(BIC_Pkgpc7))
2412 		outp += sprintf(outp, "%sPkg%%pc7", (printed++ ? delim : ""));
2413 	if (DO_BIC(BIC_Pkgpc8))
2414 		outp += sprintf(outp, "%sPkg%%pc8", (printed++ ? delim : ""));
2415 	if (DO_BIC(BIC_Pkgpc9))
2416 		outp += sprintf(outp, "%sPkg%%pc9", (printed++ ? delim : ""));
2417 	if (DO_BIC(BIC_Pkgpc10))
2418 		outp += sprintf(outp, "%sPk%%pc10", (printed++ ? delim : ""));
2419 	if (DO_BIC(BIC_Diec6))
2420 		outp += sprintf(outp, "%sDie%%c6", (printed++ ? delim : ""));
2421 	if (DO_BIC(BIC_CPU_LPI))
2422 		outp += sprintf(outp, "%sCPU%%LPI", (printed++ ? delim : ""));
2423 	if (DO_BIC(BIC_SYS_LPI))
2424 		outp += sprintf(outp, "%sSYS%%LPI", (printed++ ? delim : ""));
2425 
2426 	if (platform->rapl_msrs && !rapl_joules) {
2427 		if (DO_BIC(BIC_PkgWatt))
2428 			outp += sprintf(outp, "%sPkgWatt", (printed++ ? delim : ""));
2429 		if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
2430 			outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
2431 		if (DO_BIC(BIC_GFXWatt))
2432 			outp += sprintf(outp, "%sGFXWatt", (printed++ ? delim : ""));
2433 		if (DO_BIC(BIC_RAMWatt))
2434 			outp += sprintf(outp, "%sRAMWatt", (printed++ ? delim : ""));
2435 		if (DO_BIC(BIC_PKG__))
2436 			outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
2437 		if (DO_BIC(BIC_RAM__))
2438 			outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
2439 	} else if (platform->rapl_msrs && rapl_joules) {
2440 		if (DO_BIC(BIC_Pkg_J))
2441 			outp += sprintf(outp, "%sPkg_J", (printed++ ? delim : ""));
2442 		if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
2443 			outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
2444 		if (DO_BIC(BIC_GFX_J))
2445 			outp += sprintf(outp, "%sGFX_J", (printed++ ? delim : ""));
2446 		if (DO_BIC(BIC_RAM_J))
2447 			outp += sprintf(outp, "%sRAM_J", (printed++ ? delim : ""));
2448 		if (DO_BIC(BIC_PKG__))
2449 			outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
2450 		if (DO_BIC(BIC_RAM__))
2451 			outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
2452 	}
2453 	if (DO_BIC(BIC_UNCORE_MHZ))
2454 		outp += sprintf(outp, "%sUncMHz", (printed++ ? delim : ""));
2455 
2456 	for (mp = sys.pp; mp; mp = mp->next) {
2457 		if (mp->format == FORMAT_RAW) {
2458 			if (mp->width == 64)
2459 				outp += sprintf(outp, "%s%18.18s", delim, mp->name);
2460 			else if (mp->width == 32)
2461 				outp += sprintf(outp, "%s%10.10s", delim, mp->name);
2462 			else
2463 				outp += sprintf(outp, "%s%7.7s", delim, mp->name);
2464 		} else {
2465 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2466 				outp += sprintf(outp, "%s%8s", delim, mp->name);
2467 			else
2468 				outp += sprintf(outp, "%s%7.7s", delim, mp->name);
2469 		}
2470 	}
2471 
2472 	for (pp = sys.perf_pp; pp; pp = pp->next) {
2473 
2474 		if (pp->format == FORMAT_RAW) {
2475 			if (pp->width == 64)
2476 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), pp->name);
2477 			else
2478 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), pp->name);
2479 		} else {
2480 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2481 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), pp->name);
2482 			else
2483 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), pp->name);
2484 		}
2485 	}
2486 
2487 	ppmt = sys.pmt_pp;
2488 	while (ppmt) {
2489 		switch (ppmt->type) {
2490 		case PMT_TYPE_RAW:
2491 			if (pmt_counter_get_width(ppmt) <= 32)
2492 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), ppmt->name);
2493 			else
2494 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), ppmt->name);
2495 
2496 			break;
2497 
2498 		case PMT_TYPE_XTAL_TIME:
2499 			outp += sprintf(outp, "%s%s", delim, ppmt->name);
2500 			break;
2501 		}
2502 
2503 		ppmt = ppmt->next;
2504 	}
2505 
2506 	outp += sprintf(outp, "\n");
2507 }
2508 
2509 int dump_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2510 {
2511 	int i;
2512 	struct msr_counter *mp;
2513 
2514 	outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
2515 
2516 	if (t) {
2517 		outp += sprintf(outp, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags);
2518 		outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
2519 		outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
2520 		outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
2521 		outp += sprintf(outp, "c1: %016llX\n", t->c1);
2522 
2523 		if (DO_BIC(BIC_IPC))
2524 			outp += sprintf(outp, "IPC: %lld\n", t->instr_count);
2525 
2526 		if (DO_BIC(BIC_IRQ))
2527 			outp += sprintf(outp, "IRQ: %lld\n", t->irq_count);
2528 		if (DO_BIC(BIC_SMI))
2529 			outp += sprintf(outp, "SMI: %d\n", t->smi_count);
2530 
2531 		for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
2532 			outp +=
2533 			    sprintf(outp, "tADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num,
2534 				    t->counter[i], mp->sp->path);
2535 		}
2536 	}
2537 
2538 	if (c && is_cpu_first_thread_in_core(t, c, p)) {
2539 		outp += sprintf(outp, "core: %d\n", c->core_id);
2540 		outp += sprintf(outp, "c3: %016llX\n", c->c3);
2541 		outp += sprintf(outp, "c6: %016llX\n", c->c6);
2542 		outp += sprintf(outp, "c7: %016llX\n", c->c7);
2543 		outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
2544 		outp += sprintf(outp, "cpu_throt_count: %016llX\n", c->core_throt_cnt);
2545 
2546 		const unsigned long long energy_value = c->core_energy.raw_value * c->core_energy.scale;
2547 		const double energy_scale = c->core_energy.scale;
2548 
2549 		if (c->core_energy.unit == RAPL_UNIT_JOULES)
2550 			outp += sprintf(outp, "Joules: %0llX (scale: %lf)\n", energy_value, energy_scale);
2551 
2552 		for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
2553 			outp +=
2554 			    sprintf(outp, "cADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num,
2555 				    c->counter[i], mp->sp->path);
2556 		}
2557 		outp += sprintf(outp, "mc6_us: %016llX\n", c->mc6_us);
2558 	}
2559 
2560 	if (p && is_cpu_first_core_in_package(t, c, p)) {
2561 		outp += sprintf(outp, "package: %d\n", p->package_id);
2562 
2563 		outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
2564 		outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
2565 		outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
2566 		outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
2567 
2568 		outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
2569 		if (DO_BIC(BIC_Pkgpc3))
2570 			outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
2571 		if (DO_BIC(BIC_Pkgpc6))
2572 			outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
2573 		if (DO_BIC(BIC_Pkgpc7))
2574 			outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
2575 		outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
2576 		outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
2577 		outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
2578 		outp += sprintf(outp, "cpu_lpi: %016llX\n", p->cpu_lpi);
2579 		outp += sprintf(outp, "sys_lpi: %016llX\n", p->sys_lpi);
2580 		outp += sprintf(outp, "Joules PKG: %0llX\n", p->energy_pkg.raw_value);
2581 		outp += sprintf(outp, "Joules COR: %0llX\n", p->energy_cores.raw_value);
2582 		outp += sprintf(outp, "Joules GFX: %0llX\n", p->energy_gfx.raw_value);
2583 		outp += sprintf(outp, "Joules RAM: %0llX\n", p->energy_dram.raw_value);
2584 		outp += sprintf(outp, "Throttle PKG: %0llX\n", p->rapl_pkg_perf_status.raw_value);
2585 		outp += sprintf(outp, "Throttle RAM: %0llX\n", p->rapl_dram_perf_status.raw_value);
2586 		outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
2587 
2588 		for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
2589 			outp +=
2590 			    sprintf(outp, "pADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num,
2591 				    p->counter[i], mp->sp->path);
2592 		}
2593 	}
2594 
2595 	outp += sprintf(outp, "\n");
2596 
2597 	return 0;
2598 }
2599 
2600 double rapl_counter_get_value(const struct rapl_counter *c, enum rapl_unit desired_unit, double interval)
2601 {
2602 	assert(desired_unit != RAPL_UNIT_INVALID);
2603 
2604 	/*
2605 	 * For now we don't expect anything other than joules,
2606 	 * so just simplify the logic.
2607 	 */
2608 	assert(c->unit == RAPL_UNIT_JOULES);
2609 
2610 	const double scaled = c->raw_value * c->scale;
2611 
2612 	if (desired_unit == RAPL_UNIT_WATTS)
2613 		return scaled / interval;
2614 	return scaled;
2615 }
2616 
2617 /*
2618  * column formatting convention & formats
2619  */
2620 int format_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2621 {
2622 	double interval_float, tsc;
2623 	char *fmt8;
2624 	int i;
2625 	struct msr_counter *mp;
2626 	struct perf_counter_info *pp;
2627 	struct pmt_counter *ppmt;
2628 	char *delim = "\t";
2629 	int printed = 0;
2630 
2631 	/* if showing only 1st thread in core and this isn't one, bail out */
2632 	if (show_core_only && !is_cpu_first_thread_in_core(t, c, p))
2633 		return 0;
2634 
2635 	/* if showing only 1st thread in pkg and this isn't one, bail out */
2636 	if (show_pkg_only && !is_cpu_first_core_in_package(t, c, p))
2637 		return 0;
2638 
2639 	/*if not summary line and --cpu is used */
2640 	if ((t != &average.threads) && (cpu_subset && !CPU_ISSET_S(t->cpu_id, cpu_subset_size, cpu_subset)))
2641 		return 0;
2642 
2643 	if (DO_BIC(BIC_USEC)) {
2644 		/* on each row, print how many usec each timestamp took to gather */
2645 		struct timeval tv;
2646 
2647 		timersub(&t->tv_end, &t->tv_begin, &tv);
2648 		outp += sprintf(outp, "%5ld\t", tv.tv_sec * 1000000 + tv.tv_usec);
2649 	}
2650 
2651 	/* Time_Of_Day_Seconds: on each row, print sec.usec last timestamp taken */
2652 	if (DO_BIC(BIC_TOD))
2653 		outp += sprintf(outp, "%10ld.%06ld\t", t->tv_end.tv_sec, t->tv_end.tv_usec);
2654 
2655 	interval_float = t->tv_delta.tv_sec + t->tv_delta.tv_usec / 1000000.0;
2656 
2657 	tsc = t->tsc * tsc_tweak;
2658 
2659 	/* topo columns, print blanks on 1st (average) line */
2660 	if (t == &average.threads) {
2661 		if (DO_BIC(BIC_Package))
2662 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2663 		if (DO_BIC(BIC_Die))
2664 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2665 		if (DO_BIC(BIC_Node))
2666 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2667 		if (DO_BIC(BIC_Core))
2668 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2669 		if (DO_BIC(BIC_CPU))
2670 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2671 		if (DO_BIC(BIC_APIC))
2672 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2673 		if (DO_BIC(BIC_X2APIC))
2674 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2675 	} else {
2676 		if (DO_BIC(BIC_Package)) {
2677 			if (p)
2678 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->package_id);
2679 			else
2680 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2681 		}
2682 		if (DO_BIC(BIC_Die)) {
2683 			if (c)
2684 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].die_id);
2685 			else
2686 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2687 		}
2688 		if (DO_BIC(BIC_Node)) {
2689 			if (t)
2690 				outp += sprintf(outp, "%s%d",
2691 						(printed++ ? delim : ""), cpus[t->cpu_id].physical_node_id);
2692 			else
2693 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2694 		}
2695 		if (DO_BIC(BIC_Core)) {
2696 			if (c)
2697 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_id);
2698 			else
2699 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2700 		}
2701 		if (DO_BIC(BIC_CPU))
2702 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->cpu_id);
2703 		if (DO_BIC(BIC_APIC))
2704 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->apic_id);
2705 		if (DO_BIC(BIC_X2APIC))
2706 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->x2apic_id);
2707 	}
2708 
2709 	if (DO_BIC(BIC_Avg_MHz))
2710 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 / units * t->aperf / interval_float);
2711 
2712 	if (DO_BIC(BIC_Busy))
2713 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->mperf / tsc);
2714 
2715 	if (DO_BIC(BIC_Bzy_MHz)) {
2716 		if (has_base_hz)
2717 			outp +=
2718 			    sprintf(outp, "%s%.0f", (printed++ ? delim : ""), base_hz / units * t->aperf / t->mperf);
2719 		else
2720 			outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""),
2721 					tsc / units * t->aperf / t->mperf / interval_float);
2722 	}
2723 
2724 	if (DO_BIC(BIC_TSC_MHz))
2725 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 * t->tsc / units / interval_float);
2726 
2727 	if (DO_BIC(BIC_IPC))
2728 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 1.0 * t->instr_count / t->aperf);
2729 
2730 	/* IRQ */
2731 	if (DO_BIC(BIC_IRQ)) {
2732 		if (sums_need_wide_columns)
2733 			outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->irq_count);
2734 		else
2735 			outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->irq_count);
2736 	}
2737 
2738 	/* SMI */
2739 	if (DO_BIC(BIC_SMI))
2740 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->smi_count);
2741 
2742 	/* Added counters */
2743 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
2744 		if (mp->format == FORMAT_RAW) {
2745 			if (mp->width == 32)
2746 				outp +=
2747 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int)t->counter[i]);
2748 			else
2749 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->counter[i]);
2750 		} else if (mp->format == FORMAT_DELTA) {
2751 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2752 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->counter[i]);
2753 			else
2754 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->counter[i]);
2755 		} else if (mp->format == FORMAT_PERCENT) {
2756 			if (mp->type == COUNTER_USEC)
2757 				outp +=
2758 				    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2759 					    t->counter[i] / interval_float / 10000);
2760 			else
2761 				outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->counter[i] / tsc);
2762 		}
2763 	}
2764 
2765 	/* Added perf counters */
2766 	for (i = 0, pp = sys.perf_tp; pp; ++i, pp = pp->next) {
2767 		if (pp->format == FORMAT_RAW) {
2768 			if (pp->width == 32)
2769 				outp +=
2770 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
2771 					    (unsigned int)t->perf_counter[i]);
2772 			else
2773 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->perf_counter[i]);
2774 		} else if (pp->format == FORMAT_DELTA) {
2775 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2776 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->perf_counter[i]);
2777 			else
2778 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->perf_counter[i]);
2779 		} else if (pp->format == FORMAT_PERCENT) {
2780 			if (pp->type == COUNTER_USEC)
2781 				outp +=
2782 				    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2783 					    t->perf_counter[i] / interval_float / 10000);
2784 			else
2785 				outp +=
2786 				    sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->perf_counter[i] / tsc);
2787 		}
2788 	}
2789 
2790 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
2791 		switch (ppmt->type) {
2792 		case PMT_TYPE_RAW:
2793 			if (pmt_counter_get_width(ppmt) <= 32)
2794 				outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
2795 						(unsigned int)t->pmt_counter[i]);
2796 			else
2797 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->pmt_counter[i]);
2798 
2799 			break;
2800 
2801 		case PMT_TYPE_XTAL_TIME:
2802 			const unsigned long value_raw = t->pmt_counter[i];
2803 			const double value_converted = 100.0 * value_raw / crystal_hz / interval_float;
2804 
2805 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
2806 			break;
2807 		}
2808 	}
2809 
2810 	/* C1 */
2811 	if (DO_BIC(BIC_CPU_c1))
2812 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->c1 / tsc);
2813 
2814 	/* print per-core data only for 1st thread in core */
2815 	if (!is_cpu_first_thread_in_core(t, c, p))
2816 		goto done;
2817 
2818 	if (DO_BIC(BIC_CPU_c3))
2819 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c3 / tsc);
2820 	if (DO_BIC(BIC_CPU_c6))
2821 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c6 / tsc);
2822 	if (DO_BIC(BIC_CPU_c7))
2823 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c7 / tsc);
2824 
2825 	/* Mod%c6 */
2826 	if (DO_BIC(BIC_Mod_c6))
2827 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->mc6_us / tsc);
2828 
2829 	if (DO_BIC(BIC_CoreTmp))
2830 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_temp_c);
2831 
2832 	/* Core throttle count */
2833 	if (DO_BIC(BIC_CORE_THROT_CNT))
2834 		outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->core_throt_cnt);
2835 
2836 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
2837 		if (mp->format == FORMAT_RAW) {
2838 			if (mp->width == 32)
2839 				outp +=
2840 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int)c->counter[i]);
2841 			else
2842 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->counter[i]);
2843 		} else if (mp->format == FORMAT_DELTA) {
2844 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2845 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), c->counter[i]);
2846 			else
2847 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->counter[i]);
2848 		} else if (mp->format == FORMAT_PERCENT) {
2849 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->counter[i] / tsc);
2850 		}
2851 	}
2852 
2853 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
2854 		if (pp->format == FORMAT_RAW) {
2855 			if (pp->width == 32)
2856 				outp +=
2857 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
2858 					    (unsigned int)c->perf_counter[i]);
2859 			else
2860 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->perf_counter[i]);
2861 		} else if (pp->format == FORMAT_DELTA) {
2862 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2863 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), c->perf_counter[i]);
2864 			else
2865 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->perf_counter[i]);
2866 		} else if (pp->format == FORMAT_PERCENT) {
2867 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->perf_counter[i] / tsc);
2868 		}
2869 	}
2870 
2871 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
2872 		switch (ppmt->type) {
2873 		case PMT_TYPE_RAW:
2874 			if (pmt_counter_get_width(ppmt) <= 32)
2875 				outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
2876 						(unsigned int)c->pmt_counter[i]);
2877 			else
2878 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->pmt_counter[i]);
2879 
2880 			break;
2881 
2882 		case PMT_TYPE_XTAL_TIME:
2883 			const unsigned long value_raw = c->pmt_counter[i];
2884 			const double value_converted = 100.0 * value_raw / crystal_hz / interval_float;
2885 
2886 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
2887 			break;
2888 		}
2889 	}
2890 
2891 	fmt8 = "%s%.2f";
2892 
2893 	if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
2894 		outp +=
2895 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
2896 			    rapl_counter_get_value(&c->core_energy, RAPL_UNIT_WATTS, interval_float));
2897 	if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
2898 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
2899 				rapl_counter_get_value(&c->core_energy, RAPL_UNIT_JOULES, interval_float));
2900 
2901 	/* print per-package data only for 1st core in package */
2902 	if (!is_cpu_first_core_in_package(t, c, p))
2903 		goto done;
2904 
2905 	/* PkgTmp */
2906 	if (DO_BIC(BIC_PkgTmp))
2907 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->pkg_temp_c);
2908 
2909 	/* GFXrc6 */
2910 	if (DO_BIC(BIC_GFX_rc6)) {
2911 		if (p->gfx_rc6_ms == -1) {	/* detect GFX counter reset */
2912 			outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
2913 		} else {
2914 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2915 					p->gfx_rc6_ms / 10.0 / interval_float);
2916 		}
2917 	}
2918 
2919 	/* GFXMHz */
2920 	if (DO_BIC(BIC_GFXMHz))
2921 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_mhz);
2922 
2923 	/* GFXACTMHz */
2924 	if (DO_BIC(BIC_GFXACTMHz))
2925 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_act_mhz);
2926 
2927 	/* SAMmc6 */
2928 	if (DO_BIC(BIC_SAM_mc6)) {
2929 		if (p->sam_mc6_ms == -1) {	/* detect GFX counter reset */
2930 			outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
2931 		} else {
2932 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2933 					p->sam_mc6_ms / 10.0 / interval_float);
2934 		}
2935 	}
2936 
2937 	/* SAMMHz */
2938 	if (DO_BIC(BIC_SAMMHz))
2939 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->sam_mhz);
2940 
2941 	/* SAMACTMHz */
2942 	if (DO_BIC(BIC_SAMACTMHz))
2943 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->sam_act_mhz);
2944 
2945 	/* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
2946 	if (DO_BIC(BIC_Totl_c0))
2947 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_wtd_core_c0 / tsc);
2948 	if (DO_BIC(BIC_Any_c0))
2949 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_core_c0 / tsc);
2950 	if (DO_BIC(BIC_GFX_c0))
2951 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_gfxe_c0 / tsc);
2952 	if (DO_BIC(BIC_CPUGFX))
2953 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_both_core_gfxe_c0 / tsc);
2954 
2955 	if (DO_BIC(BIC_Pkgpc2))
2956 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc2 / tsc);
2957 	if (DO_BIC(BIC_Pkgpc3))
2958 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc3 / tsc);
2959 	if (DO_BIC(BIC_Pkgpc6))
2960 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc6 / tsc);
2961 	if (DO_BIC(BIC_Pkgpc7))
2962 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc7 / tsc);
2963 	if (DO_BIC(BIC_Pkgpc8))
2964 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc8 / tsc);
2965 	if (DO_BIC(BIC_Pkgpc9))
2966 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc9 / tsc);
2967 	if (DO_BIC(BIC_Pkgpc10))
2968 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc10 / tsc);
2969 
2970 	if (DO_BIC(BIC_Diec6))
2971 		outp +=
2972 		    sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->die_c6 / crystal_hz / interval_float);
2973 
2974 	if (DO_BIC(BIC_CPU_LPI)) {
2975 		if (p->cpu_lpi >= 0)
2976 			outp +=
2977 			    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2978 				    100.0 * p->cpu_lpi / 1000000.0 / interval_float);
2979 		else
2980 			outp += sprintf(outp, "%s(neg)", (printed++ ? delim : ""));
2981 	}
2982 	if (DO_BIC(BIC_SYS_LPI)) {
2983 		if (p->sys_lpi >= 0)
2984 			outp +=
2985 			    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2986 				    100.0 * p->sys_lpi / 1000000.0 / interval_float);
2987 		else
2988 			outp += sprintf(outp, "%s(neg)", (printed++ ? delim : ""));
2989 	}
2990 
2991 	if (DO_BIC(BIC_PkgWatt))
2992 		outp +=
2993 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
2994 			    rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_WATTS, interval_float));
2995 	if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
2996 		outp +=
2997 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
2998 			    rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_WATTS, interval_float));
2999 	if (DO_BIC(BIC_GFXWatt))
3000 		outp +=
3001 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3002 			    rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_WATTS, interval_float));
3003 	if (DO_BIC(BIC_RAMWatt))
3004 		outp +=
3005 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3006 			    rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_WATTS, interval_float));
3007 	if (DO_BIC(BIC_Pkg_J))
3008 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3009 				rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_JOULES, interval_float));
3010 	if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
3011 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3012 				rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_JOULES, interval_float));
3013 	if (DO_BIC(BIC_GFX_J))
3014 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3015 				rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_JOULES, interval_float));
3016 	if (DO_BIC(BIC_RAM_J))
3017 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3018 				rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_JOULES, interval_float));
3019 	if (DO_BIC(BIC_PKG__))
3020 		outp +=
3021 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3022 			    rapl_counter_get_value(&p->rapl_pkg_perf_status, RAPL_UNIT_WATTS, interval_float));
3023 	if (DO_BIC(BIC_RAM__))
3024 		outp +=
3025 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3026 			    rapl_counter_get_value(&p->rapl_dram_perf_status, RAPL_UNIT_WATTS, interval_float));
3027 	/* UncMHz */
3028 	if (DO_BIC(BIC_UNCORE_MHZ))
3029 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->uncore_mhz);
3030 
3031 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3032 		if (mp->format == FORMAT_RAW) {
3033 			if (mp->width == 32)
3034 				outp +=
3035 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int)p->counter[i]);
3036 			else
3037 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->counter[i]);
3038 		} else if (mp->format == FORMAT_DELTA) {
3039 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
3040 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), p->counter[i]);
3041 			else
3042 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), p->counter[i]);
3043 		} else if (mp->format == FORMAT_PERCENT) {
3044 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->counter[i] / tsc);
3045 		} else if (mp->type == COUNTER_K2M)
3046 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), (unsigned int)p->counter[i] / 1000);
3047 	}
3048 
3049 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3050 		if (pp->format == FORMAT_RAW) {
3051 			if (pp->width == 32)
3052 				outp +=
3053 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
3054 					    (unsigned int)p->perf_counter[i]);
3055 			else
3056 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->perf_counter[i]);
3057 		} else if (pp->format == FORMAT_DELTA) {
3058 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
3059 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), p->perf_counter[i]);
3060 			else
3061 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), p->perf_counter[i]);
3062 		} else if (pp->format == FORMAT_PERCENT) {
3063 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->perf_counter[i] / tsc);
3064 		} else if (pp->type == COUNTER_K2M) {
3065 			outp +=
3066 			    sprintf(outp, "%s%d", (printed++ ? delim : ""), (unsigned int)p->perf_counter[i] / 1000);
3067 		}
3068 	}
3069 
3070 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3071 		switch (ppmt->type) {
3072 		case PMT_TYPE_RAW:
3073 			if (pmt_counter_get_width(ppmt) <= 32)
3074 				outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
3075 						(unsigned int)p->pmt_counter[i]);
3076 			else
3077 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->pmt_counter[i]);
3078 
3079 			break;
3080 
3081 		case PMT_TYPE_XTAL_TIME:
3082 			const unsigned long value_raw = p->pmt_counter[i];
3083 			const double value_converted = 100.0 * value_raw / crystal_hz / interval_float;
3084 
3085 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
3086 			break;
3087 		}
3088 	}
3089 
3090 done:
3091 	if (*(outp - 1) != '\n')
3092 		outp += sprintf(outp, "\n");
3093 
3094 	return 0;
3095 }
3096 
3097 void flush_output_stdout(void)
3098 {
3099 	FILE *filep;
3100 
3101 	if (outf == stderr)
3102 		filep = stdout;
3103 	else
3104 		filep = outf;
3105 
3106 	fputs(output_buffer, filep);
3107 	fflush(filep);
3108 
3109 	outp = output_buffer;
3110 }
3111 
3112 void flush_output_stderr(void)
3113 {
3114 	fputs(output_buffer, outf);
3115 	fflush(outf);
3116 	outp = output_buffer;
3117 }
3118 
3119 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3120 {
3121 	static int count;
3122 
3123 	if ((!count || (header_iterations && !(count % header_iterations))) || !summary_only)
3124 		print_header("\t");
3125 
3126 	format_counters(&average.threads, &average.cores, &average.packages);
3127 
3128 	count++;
3129 
3130 	if (summary_only)
3131 		return;
3132 
3133 	for_all_cpus(format_counters, t, c, p);
3134 }
3135 
3136 #define DELTA_WRAP32(new, old)			\
3137 	old = ((((unsigned long long)new << 32) - ((unsigned long long)old << 32)) >> 32);
3138 
3139 int delta_package(struct pkg_data *new, struct pkg_data *old)
3140 {
3141 	int i;
3142 	struct msr_counter *mp;
3143 	struct perf_counter_info *pp;
3144 	struct pmt_counter *ppmt;
3145 
3146 	if (DO_BIC(BIC_Totl_c0))
3147 		old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
3148 	if (DO_BIC(BIC_Any_c0))
3149 		old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
3150 	if (DO_BIC(BIC_GFX_c0))
3151 		old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
3152 	if (DO_BIC(BIC_CPUGFX))
3153 		old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
3154 
3155 	old->pc2 = new->pc2 - old->pc2;
3156 	if (DO_BIC(BIC_Pkgpc3))
3157 		old->pc3 = new->pc3 - old->pc3;
3158 	if (DO_BIC(BIC_Pkgpc6))
3159 		old->pc6 = new->pc6 - old->pc6;
3160 	if (DO_BIC(BIC_Pkgpc7))
3161 		old->pc7 = new->pc7 - old->pc7;
3162 	old->pc8 = new->pc8 - old->pc8;
3163 	old->pc9 = new->pc9 - old->pc9;
3164 	old->pc10 = new->pc10 - old->pc10;
3165 	old->die_c6 = new->die_c6 - old->die_c6;
3166 	old->cpu_lpi = new->cpu_lpi - old->cpu_lpi;
3167 	old->sys_lpi = new->sys_lpi - old->sys_lpi;
3168 	old->pkg_temp_c = new->pkg_temp_c;
3169 
3170 	/* flag an error when rc6 counter resets/wraps */
3171 	if (old->gfx_rc6_ms > new->gfx_rc6_ms)
3172 		old->gfx_rc6_ms = -1;
3173 	else
3174 		old->gfx_rc6_ms = new->gfx_rc6_ms - old->gfx_rc6_ms;
3175 
3176 	old->uncore_mhz = new->uncore_mhz;
3177 	old->gfx_mhz = new->gfx_mhz;
3178 	old->gfx_act_mhz = new->gfx_act_mhz;
3179 
3180 	/* flag an error when mc6 counter resets/wraps */
3181 	if (old->sam_mc6_ms > new->sam_mc6_ms)
3182 		old->sam_mc6_ms = -1;
3183 	else
3184 		old->sam_mc6_ms = new->sam_mc6_ms - old->sam_mc6_ms;
3185 
3186 	old->sam_mhz = new->sam_mhz;
3187 	old->sam_act_mhz = new->sam_act_mhz;
3188 
3189 	old->energy_pkg.raw_value = new->energy_pkg.raw_value - old->energy_pkg.raw_value;
3190 	old->energy_cores.raw_value = new->energy_cores.raw_value - old->energy_cores.raw_value;
3191 	old->energy_gfx.raw_value = new->energy_gfx.raw_value - old->energy_gfx.raw_value;
3192 	old->energy_dram.raw_value = new->energy_dram.raw_value - old->energy_dram.raw_value;
3193 	old->rapl_pkg_perf_status.raw_value = new->rapl_pkg_perf_status.raw_value - old->rapl_pkg_perf_status.raw_value;
3194 	old->rapl_dram_perf_status.raw_value =
3195 	    new->rapl_dram_perf_status.raw_value - old->rapl_dram_perf_status.raw_value;
3196 
3197 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3198 		if (mp->format == FORMAT_RAW)
3199 			old->counter[i] = new->counter[i];
3200 		else if (mp->format == FORMAT_AVERAGE)
3201 			old->counter[i] = new->counter[i];
3202 		else
3203 			old->counter[i] = new->counter[i] - old->counter[i];
3204 	}
3205 
3206 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3207 		if (pp->format == FORMAT_RAW)
3208 			old->perf_counter[i] = new->perf_counter[i];
3209 		else if (pp->format == FORMAT_AVERAGE)
3210 			old->perf_counter[i] = new->perf_counter[i];
3211 		else
3212 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3213 	}
3214 
3215 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3216 		if (ppmt->format == FORMAT_RAW)
3217 			old->pmt_counter[i] = new->pmt_counter[i];
3218 		else
3219 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3220 	}
3221 
3222 	return 0;
3223 }
3224 
3225 void delta_core(struct core_data *new, struct core_data *old)
3226 {
3227 	int i;
3228 	struct msr_counter *mp;
3229 	struct perf_counter_info *pp;
3230 	struct pmt_counter *ppmt;
3231 
3232 	old->c3 = new->c3 - old->c3;
3233 	old->c6 = new->c6 - old->c6;
3234 	old->c7 = new->c7 - old->c7;
3235 	old->core_temp_c = new->core_temp_c;
3236 	old->core_throt_cnt = new->core_throt_cnt;
3237 	old->mc6_us = new->mc6_us - old->mc6_us;
3238 
3239 	DELTA_WRAP32(new->core_energy.raw_value, old->core_energy.raw_value);
3240 
3241 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3242 		if (mp->format == FORMAT_RAW)
3243 			old->counter[i] = new->counter[i];
3244 		else
3245 			old->counter[i] = new->counter[i] - old->counter[i];
3246 	}
3247 
3248 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3249 		if (pp->format == FORMAT_RAW)
3250 			old->perf_counter[i] = new->perf_counter[i];
3251 		else
3252 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3253 	}
3254 
3255 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3256 		if (ppmt->format == FORMAT_RAW)
3257 			old->pmt_counter[i] = new->pmt_counter[i];
3258 		else
3259 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3260 	}
3261 }
3262 
3263 int soft_c1_residency_display(int bic)
3264 {
3265 	if (!DO_BIC(BIC_CPU_c1) || platform->has_msr_core_c1_res)
3266 		return 0;
3267 
3268 	return DO_BIC_READ(bic);
3269 }
3270 
3271 /*
3272  * old = new - old
3273  */
3274 int delta_thread(struct thread_data *new, struct thread_data *old, struct core_data *core_delta)
3275 {
3276 	int i;
3277 	struct msr_counter *mp;
3278 	struct perf_counter_info *pp;
3279 	struct pmt_counter *ppmt;
3280 
3281 	/* we run cpuid just the 1st time, copy the results */
3282 	if (DO_BIC(BIC_APIC))
3283 		new->apic_id = old->apic_id;
3284 	if (DO_BIC(BIC_X2APIC))
3285 		new->x2apic_id = old->x2apic_id;
3286 
3287 	/*
3288 	 * the timestamps from start of measurement interval are in "old"
3289 	 * the timestamp from end of measurement interval are in "new"
3290 	 * over-write old w/ new so we can print end of interval values
3291 	 */
3292 
3293 	timersub(&new->tv_begin, &old->tv_begin, &old->tv_delta);
3294 	old->tv_begin = new->tv_begin;
3295 	old->tv_end = new->tv_end;
3296 
3297 	old->tsc = new->tsc - old->tsc;
3298 
3299 	/* check for TSC < 1 Mcycles over interval */
3300 	if (old->tsc < (1000 * 1000))
3301 		errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
3302 		     "You can disable all c-states by booting with \"idle=poll\"\n"
3303 		     "or just the deep ones with \"processor.max_cstate=1\"");
3304 
3305 	old->c1 = new->c1 - old->c1;
3306 
3307 	if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz) || DO_BIC(BIC_IPC)
3308 	    || soft_c1_residency_display(BIC_Avg_MHz)) {
3309 		if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
3310 			old->aperf = new->aperf - old->aperf;
3311 			old->mperf = new->mperf - old->mperf;
3312 		} else {
3313 			return -1;
3314 		}
3315 	}
3316 
3317 	if (platform->has_msr_core_c1_res) {
3318 		/*
3319 		 * Some models have a dedicated C1 residency MSR,
3320 		 * which should be more accurate than the derivation below.
3321 		 */
3322 	} else {
3323 		/*
3324 		 * As counter collection is not atomic,
3325 		 * it is possible for mperf's non-halted cycles + idle states
3326 		 * to exceed TSC's all cycles: show c1 = 0% in that case.
3327 		 */
3328 		if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > (old->tsc * tsc_tweak))
3329 			old->c1 = 0;
3330 		else {
3331 			/* normal case, derive c1 */
3332 			old->c1 = (old->tsc * tsc_tweak) - old->mperf - core_delta->c3
3333 			    - core_delta->c6 - core_delta->c7;
3334 		}
3335 	}
3336 
3337 	if (old->mperf == 0) {
3338 		if (debug > 1)
3339 			fprintf(outf, "cpu%d MPERF 0!\n", old->cpu_id);
3340 		old->mperf = 1;	/* divide by 0 protection */
3341 	}
3342 
3343 	if (DO_BIC(BIC_IPC))
3344 		old->instr_count = new->instr_count - old->instr_count;
3345 
3346 	if (DO_BIC(BIC_IRQ))
3347 		old->irq_count = new->irq_count - old->irq_count;
3348 
3349 	if (DO_BIC(BIC_SMI))
3350 		old->smi_count = new->smi_count - old->smi_count;
3351 
3352 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3353 		if (mp->format == FORMAT_RAW)
3354 			old->counter[i] = new->counter[i];
3355 		else
3356 			old->counter[i] = new->counter[i] - old->counter[i];
3357 	}
3358 
3359 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
3360 		if (pp->format == FORMAT_RAW)
3361 			old->perf_counter[i] = new->perf_counter[i];
3362 		else
3363 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3364 	}
3365 
3366 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3367 		if (ppmt->format == FORMAT_RAW)
3368 			old->pmt_counter[i] = new->pmt_counter[i];
3369 		else
3370 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3371 	}
3372 
3373 	return 0;
3374 }
3375 
3376 int delta_cpu(struct thread_data *t, struct core_data *c,
3377 	      struct pkg_data *p, struct thread_data *t2, struct core_data *c2, struct pkg_data *p2)
3378 {
3379 	int retval = 0;
3380 
3381 	/* calculate core delta only for 1st thread in core */
3382 	if (is_cpu_first_thread_in_core(t, c, p))
3383 		delta_core(c, c2);
3384 
3385 	/* always calculate thread delta */
3386 	retval = delta_thread(t, t2, c2);	/* c2 is core delta */
3387 	if (retval)
3388 		return retval;
3389 
3390 	/* calculate package delta only for 1st core in package */
3391 	if (is_cpu_first_core_in_package(t, c, p))
3392 		retval = delta_package(p, p2);
3393 
3394 	return retval;
3395 }
3396 
3397 void rapl_counter_clear(struct rapl_counter *c)
3398 {
3399 	c->raw_value = 0;
3400 	c->scale = 0.0;
3401 	c->unit = RAPL_UNIT_INVALID;
3402 }
3403 
3404 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3405 {
3406 	int i;
3407 	struct msr_counter *mp;
3408 
3409 	t->tv_begin.tv_sec = 0;
3410 	t->tv_begin.tv_usec = 0;
3411 	t->tv_end.tv_sec = 0;
3412 	t->tv_end.tv_usec = 0;
3413 	t->tv_delta.tv_sec = 0;
3414 	t->tv_delta.tv_usec = 0;
3415 
3416 	t->tsc = 0;
3417 	t->aperf = 0;
3418 	t->mperf = 0;
3419 	t->c1 = 0;
3420 
3421 	t->instr_count = 0;
3422 
3423 	t->irq_count = 0;
3424 	t->smi_count = 0;
3425 
3426 	c->c3 = 0;
3427 	c->c6 = 0;
3428 	c->c7 = 0;
3429 	c->mc6_us = 0;
3430 	c->core_temp_c = 0;
3431 	rapl_counter_clear(&c->core_energy);
3432 	c->core_throt_cnt = 0;
3433 
3434 	p->pkg_wtd_core_c0 = 0;
3435 	p->pkg_any_core_c0 = 0;
3436 	p->pkg_any_gfxe_c0 = 0;
3437 	p->pkg_both_core_gfxe_c0 = 0;
3438 
3439 	p->pc2 = 0;
3440 	if (DO_BIC(BIC_Pkgpc3))
3441 		p->pc3 = 0;
3442 	if (DO_BIC(BIC_Pkgpc6))
3443 		p->pc6 = 0;
3444 	if (DO_BIC(BIC_Pkgpc7))
3445 		p->pc7 = 0;
3446 	p->pc8 = 0;
3447 	p->pc9 = 0;
3448 	p->pc10 = 0;
3449 	p->die_c6 = 0;
3450 	p->cpu_lpi = 0;
3451 	p->sys_lpi = 0;
3452 
3453 	rapl_counter_clear(&p->energy_pkg);
3454 	rapl_counter_clear(&p->energy_dram);
3455 	rapl_counter_clear(&p->energy_cores);
3456 	rapl_counter_clear(&p->energy_gfx);
3457 	rapl_counter_clear(&p->rapl_pkg_perf_status);
3458 	rapl_counter_clear(&p->rapl_dram_perf_status);
3459 	p->pkg_temp_c = 0;
3460 
3461 	p->gfx_rc6_ms = 0;
3462 	p->uncore_mhz = 0;
3463 	p->gfx_mhz = 0;
3464 	p->gfx_act_mhz = 0;
3465 	p->sam_mc6_ms = 0;
3466 	p->sam_mhz = 0;
3467 	p->sam_act_mhz = 0;
3468 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next)
3469 		t->counter[i] = 0;
3470 
3471 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next)
3472 		c->counter[i] = 0;
3473 
3474 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next)
3475 		p->counter[i] = 0;
3476 
3477 	memset(&t->perf_counter[0], 0, sizeof(t->perf_counter));
3478 	memset(&c->perf_counter[0], 0, sizeof(c->perf_counter));
3479 	memset(&p->perf_counter[0], 0, sizeof(p->perf_counter));
3480 
3481 	memset(&t->pmt_counter[0], 0, ARRAY_SIZE(t->pmt_counter));
3482 	memset(&c->pmt_counter[0], 0, ARRAY_SIZE(c->pmt_counter));
3483 	memset(&p->pmt_counter[0], 0, ARRAY_SIZE(p->pmt_counter));
3484 }
3485 
3486 void rapl_counter_accumulate(struct rapl_counter *dst, const struct rapl_counter *src)
3487 {
3488 	/* Copy unit and scale from src if dst is not initialized */
3489 	if (dst->unit == RAPL_UNIT_INVALID) {
3490 		dst->unit = src->unit;
3491 		dst->scale = src->scale;
3492 	}
3493 
3494 	assert(dst->unit == src->unit);
3495 	assert(dst->scale == src->scale);
3496 
3497 	dst->raw_value += src->raw_value;
3498 }
3499 
3500 int sum_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3501 {
3502 	int i;
3503 	struct msr_counter *mp;
3504 	struct perf_counter_info *pp;
3505 	struct pmt_counter *ppmt;
3506 
3507 	/* copy un-changing apic_id's */
3508 	if (DO_BIC(BIC_APIC))
3509 		average.threads.apic_id = t->apic_id;
3510 	if (DO_BIC(BIC_X2APIC))
3511 		average.threads.x2apic_id = t->x2apic_id;
3512 
3513 	/* remember first tv_begin */
3514 	if (average.threads.tv_begin.tv_sec == 0)
3515 		average.threads.tv_begin = t->tv_begin;
3516 
3517 	/* remember last tv_end */
3518 	average.threads.tv_end = t->tv_end;
3519 
3520 	average.threads.tsc += t->tsc;
3521 	average.threads.aperf += t->aperf;
3522 	average.threads.mperf += t->mperf;
3523 	average.threads.c1 += t->c1;
3524 
3525 	average.threads.instr_count += t->instr_count;
3526 
3527 	average.threads.irq_count += t->irq_count;
3528 	average.threads.smi_count += t->smi_count;
3529 
3530 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3531 		if (mp->format == FORMAT_RAW)
3532 			continue;
3533 		average.threads.counter[i] += t->counter[i];
3534 	}
3535 
3536 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
3537 		if (pp->format == FORMAT_RAW)
3538 			continue;
3539 		average.threads.perf_counter[i] += t->perf_counter[i];
3540 	}
3541 
3542 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3543 		average.threads.pmt_counter[i] += t->pmt_counter[i];
3544 	}
3545 
3546 	/* sum per-core values only for 1st thread in core */
3547 	if (!is_cpu_first_thread_in_core(t, c, p))
3548 		return 0;
3549 
3550 	average.cores.c3 += c->c3;
3551 	average.cores.c6 += c->c6;
3552 	average.cores.c7 += c->c7;
3553 	average.cores.mc6_us += c->mc6_us;
3554 
3555 	average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
3556 	average.cores.core_throt_cnt = MAX(average.cores.core_throt_cnt, c->core_throt_cnt);
3557 
3558 	rapl_counter_accumulate(&average.cores.core_energy, &c->core_energy);
3559 
3560 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3561 		if (mp->format == FORMAT_RAW)
3562 			continue;
3563 		average.cores.counter[i] += c->counter[i];
3564 	}
3565 
3566 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3567 		if (pp->format == FORMAT_RAW)
3568 			continue;
3569 		average.cores.perf_counter[i] += c->perf_counter[i];
3570 	}
3571 
3572 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3573 		average.cores.pmt_counter[i] += c->pmt_counter[i];
3574 	}
3575 
3576 	/* sum per-pkg values only for 1st core in pkg */
3577 	if (!is_cpu_first_core_in_package(t, c, p))
3578 		return 0;
3579 
3580 	if (DO_BIC(BIC_Totl_c0))
3581 		average.packages.pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
3582 	if (DO_BIC(BIC_Any_c0))
3583 		average.packages.pkg_any_core_c0 += p->pkg_any_core_c0;
3584 	if (DO_BIC(BIC_GFX_c0))
3585 		average.packages.pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
3586 	if (DO_BIC(BIC_CPUGFX))
3587 		average.packages.pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
3588 
3589 	average.packages.pc2 += p->pc2;
3590 	if (DO_BIC(BIC_Pkgpc3))
3591 		average.packages.pc3 += p->pc3;
3592 	if (DO_BIC(BIC_Pkgpc6))
3593 		average.packages.pc6 += p->pc6;
3594 	if (DO_BIC(BIC_Pkgpc7))
3595 		average.packages.pc7 += p->pc7;
3596 	average.packages.pc8 += p->pc8;
3597 	average.packages.pc9 += p->pc9;
3598 	average.packages.pc10 += p->pc10;
3599 	average.packages.die_c6 += p->die_c6;
3600 
3601 	average.packages.cpu_lpi = p->cpu_lpi;
3602 	average.packages.sys_lpi = p->sys_lpi;
3603 
3604 	rapl_counter_accumulate(&average.packages.energy_pkg, &p->energy_pkg);
3605 	rapl_counter_accumulate(&average.packages.energy_dram, &p->energy_dram);
3606 	rapl_counter_accumulate(&average.packages.energy_cores, &p->energy_cores);
3607 	rapl_counter_accumulate(&average.packages.energy_gfx, &p->energy_gfx);
3608 
3609 	average.packages.gfx_rc6_ms = p->gfx_rc6_ms;
3610 	average.packages.uncore_mhz = p->uncore_mhz;
3611 	average.packages.gfx_mhz = p->gfx_mhz;
3612 	average.packages.gfx_act_mhz = p->gfx_act_mhz;
3613 	average.packages.sam_mc6_ms = p->sam_mc6_ms;
3614 	average.packages.sam_mhz = p->sam_mhz;
3615 	average.packages.sam_act_mhz = p->sam_act_mhz;
3616 
3617 	average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
3618 
3619 	rapl_counter_accumulate(&average.packages.rapl_pkg_perf_status, &p->rapl_pkg_perf_status);
3620 	rapl_counter_accumulate(&average.packages.rapl_dram_perf_status, &p->rapl_dram_perf_status);
3621 
3622 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3623 		if ((mp->format == FORMAT_RAW) && (topo.num_packages == 0))
3624 			average.packages.counter[i] = p->counter[i];
3625 		else
3626 			average.packages.counter[i] += p->counter[i];
3627 	}
3628 
3629 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3630 		if ((pp->format == FORMAT_RAW) && (topo.num_packages == 0))
3631 			average.packages.perf_counter[i] = p->perf_counter[i];
3632 		else
3633 			average.packages.perf_counter[i] += p->perf_counter[i];
3634 	}
3635 
3636 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3637 		average.packages.pmt_counter[i] += p->pmt_counter[i];
3638 	}
3639 
3640 	return 0;
3641 }
3642 
3643 /*
3644  * sum the counters for all cpus in the system
3645  * compute the weighted average
3646  */
3647 void compute_average(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3648 {
3649 	int i;
3650 	struct msr_counter *mp;
3651 	struct perf_counter_info *pp;
3652 	struct pmt_counter *ppmt;
3653 
3654 	clear_counters(&average.threads, &average.cores, &average.packages);
3655 
3656 	for_all_cpus(sum_counters, t, c, p);
3657 
3658 	/* Use the global time delta for the average. */
3659 	average.threads.tv_delta = tv_delta;
3660 
3661 	average.threads.tsc /= topo.allowed_cpus;
3662 	average.threads.aperf /= topo.allowed_cpus;
3663 	average.threads.mperf /= topo.allowed_cpus;
3664 	average.threads.instr_count /= topo.allowed_cpus;
3665 	average.threads.c1 /= topo.allowed_cpus;
3666 
3667 	if (average.threads.irq_count > 9999999)
3668 		sums_need_wide_columns = 1;
3669 
3670 	average.cores.c3 /= topo.allowed_cores;
3671 	average.cores.c6 /= topo.allowed_cores;
3672 	average.cores.c7 /= topo.allowed_cores;
3673 	average.cores.mc6_us /= topo.allowed_cores;
3674 
3675 	if (DO_BIC(BIC_Totl_c0))
3676 		average.packages.pkg_wtd_core_c0 /= topo.allowed_packages;
3677 	if (DO_BIC(BIC_Any_c0))
3678 		average.packages.pkg_any_core_c0 /= topo.allowed_packages;
3679 	if (DO_BIC(BIC_GFX_c0))
3680 		average.packages.pkg_any_gfxe_c0 /= topo.allowed_packages;
3681 	if (DO_BIC(BIC_CPUGFX))
3682 		average.packages.pkg_both_core_gfxe_c0 /= topo.allowed_packages;
3683 
3684 	average.packages.pc2 /= topo.allowed_packages;
3685 	if (DO_BIC(BIC_Pkgpc3))
3686 		average.packages.pc3 /= topo.allowed_packages;
3687 	if (DO_BIC(BIC_Pkgpc6))
3688 		average.packages.pc6 /= topo.allowed_packages;
3689 	if (DO_BIC(BIC_Pkgpc7))
3690 		average.packages.pc7 /= topo.allowed_packages;
3691 
3692 	average.packages.pc8 /= topo.allowed_packages;
3693 	average.packages.pc9 /= topo.allowed_packages;
3694 	average.packages.pc10 /= topo.allowed_packages;
3695 	average.packages.die_c6 /= topo.allowed_packages;
3696 
3697 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3698 		if (mp->format == FORMAT_RAW)
3699 			continue;
3700 		if (mp->type == COUNTER_ITEMS) {
3701 			if (average.threads.counter[i] > 9999999)
3702 				sums_need_wide_columns = 1;
3703 			continue;
3704 		}
3705 		average.threads.counter[i] /= topo.allowed_cpus;
3706 	}
3707 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3708 		if (mp->format == FORMAT_RAW)
3709 			continue;
3710 		if (mp->type == COUNTER_ITEMS) {
3711 			if (average.cores.counter[i] > 9999999)
3712 				sums_need_wide_columns = 1;
3713 		}
3714 		average.cores.counter[i] /= topo.allowed_cores;
3715 	}
3716 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3717 		if (mp->format == FORMAT_RAW)
3718 			continue;
3719 		if (mp->type == COUNTER_ITEMS) {
3720 			if (average.packages.counter[i] > 9999999)
3721 				sums_need_wide_columns = 1;
3722 		}
3723 		average.packages.counter[i] /= topo.allowed_packages;
3724 	}
3725 
3726 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
3727 		if (pp->format == FORMAT_RAW)
3728 			continue;
3729 		if (pp->type == COUNTER_ITEMS) {
3730 			if (average.threads.perf_counter[i] > 9999999)
3731 				sums_need_wide_columns = 1;
3732 			continue;
3733 		}
3734 		average.threads.perf_counter[i] /= topo.allowed_cpus;
3735 	}
3736 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3737 		if (pp->format == FORMAT_RAW)
3738 			continue;
3739 		if (pp->type == COUNTER_ITEMS) {
3740 			if (average.cores.perf_counter[i] > 9999999)
3741 				sums_need_wide_columns = 1;
3742 		}
3743 		average.cores.perf_counter[i] /= topo.allowed_cores;
3744 	}
3745 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3746 		if (pp->format == FORMAT_RAW)
3747 			continue;
3748 		if (pp->type == COUNTER_ITEMS) {
3749 			if (average.packages.perf_counter[i] > 9999999)
3750 				sums_need_wide_columns = 1;
3751 		}
3752 		average.packages.perf_counter[i] /= topo.allowed_packages;
3753 	}
3754 
3755 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3756 		average.threads.pmt_counter[i] /= topo.allowed_cpus;
3757 	}
3758 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3759 		average.cores.pmt_counter[i] /= topo.allowed_cores;
3760 	}
3761 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3762 		average.packages.pmt_counter[i] /= topo.allowed_packages;
3763 	}
3764 }
3765 
3766 static unsigned long long rdtsc(void)
3767 {
3768 	unsigned int low, high;
3769 
3770 	asm volatile ("rdtsc":"=a" (low), "=d"(high));
3771 
3772 	return low | ((unsigned long long)high) << 32;
3773 }
3774 
3775 /*
3776  * Open a file, and exit on failure
3777  */
3778 FILE *fopen_or_die(const char *path, const char *mode)
3779 {
3780 	FILE *filep = fopen(path, mode);
3781 
3782 	if (!filep)
3783 		err(1, "%s: open failed", path);
3784 	return filep;
3785 }
3786 
3787 /*
3788  * snapshot_sysfs_counter()
3789  *
3790  * return snapshot of given counter
3791  */
3792 unsigned long long snapshot_sysfs_counter(char *path)
3793 {
3794 	FILE *fp;
3795 	int retval;
3796 	unsigned long long counter;
3797 
3798 	fp = fopen_or_die(path, "r");
3799 
3800 	retval = fscanf(fp, "%lld", &counter);
3801 	if (retval != 1)
3802 		err(1, "snapshot_sysfs_counter(%s)", path);
3803 
3804 	fclose(fp);
3805 
3806 	return counter;
3807 }
3808 
3809 int get_mp(int cpu, struct msr_counter *mp, unsigned long long *counterp, char *counter_path)
3810 {
3811 	if (mp->msr_num != 0) {
3812 		assert(!no_msr);
3813 		if (get_msr(cpu, mp->msr_num, counterp))
3814 			return -1;
3815 	} else {
3816 		char path[128 + PATH_BYTES];
3817 
3818 		if (mp->flags & SYSFS_PERCPU) {
3819 			sprintf(path, "/sys/devices/system/cpu/cpu%d/%s", cpu, mp->sp->path);
3820 
3821 			*counterp = snapshot_sysfs_counter(path);
3822 		} else {
3823 			*counterp = snapshot_sysfs_counter(counter_path);
3824 		}
3825 	}
3826 
3827 	return 0;
3828 }
3829 
3830 unsigned long long get_legacy_uncore_mhz(int package)
3831 {
3832 	char path[128];
3833 	int die;
3834 	static int warn_once;
3835 
3836 	/*
3837 	 * for this package, use the first die_id that exists
3838 	 */
3839 	for (die = 0; die <= topo.max_die_id; ++die) {
3840 
3841 		sprintf(path, "/sys/devices/system/cpu/intel_uncore_frequency/package_%02d_die_%02d/current_freq_khz",
3842 			package, die);
3843 
3844 		if (access(path, R_OK) == 0)
3845 			return (snapshot_sysfs_counter(path) / 1000);
3846 	}
3847 	if (!warn_once) {
3848 		warnx("BUG: %s: No %s", __func__, path);
3849 		warn_once = 1;
3850 	}
3851 
3852 	return 0;
3853 }
3854 
3855 int get_epb(int cpu)
3856 {
3857 	char path[128 + PATH_BYTES];
3858 	unsigned long long msr;
3859 	int ret, epb = -1;
3860 	FILE *fp;
3861 
3862 	sprintf(path, "/sys/devices/system/cpu/cpu%d/power/energy_perf_bias", cpu);
3863 
3864 	fp = fopen(path, "r");
3865 	if (!fp)
3866 		goto msr_fallback;
3867 
3868 	ret = fscanf(fp, "%d", &epb);
3869 	if (ret != 1)
3870 		err(1, "%s(%s)", __func__, path);
3871 
3872 	fclose(fp);
3873 
3874 	return epb;
3875 
3876 msr_fallback:
3877 	if (no_msr)
3878 		return -1;
3879 
3880 	get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr);
3881 
3882 	return msr & 0xf;
3883 }
3884 
3885 void get_apic_id(struct thread_data *t)
3886 {
3887 	unsigned int eax, ebx, ecx, edx;
3888 
3889 	if (DO_BIC(BIC_APIC)) {
3890 		eax = ebx = ecx = edx = 0;
3891 		__cpuid(1, eax, ebx, ecx, edx);
3892 
3893 		t->apic_id = (ebx >> 24) & 0xff;
3894 	}
3895 
3896 	if (!DO_BIC(BIC_X2APIC))
3897 		return;
3898 
3899 	if (authentic_amd || hygon_genuine) {
3900 		unsigned int topology_extensions;
3901 
3902 		if (max_extended_level < 0x8000001e)
3903 			return;
3904 
3905 		eax = ebx = ecx = edx = 0;
3906 		__cpuid(0x80000001, eax, ebx, ecx, edx);
3907 		topology_extensions = ecx & (1 << 22);
3908 
3909 		if (topology_extensions == 0)
3910 			return;
3911 
3912 		eax = ebx = ecx = edx = 0;
3913 		__cpuid(0x8000001e, eax, ebx, ecx, edx);
3914 
3915 		t->x2apic_id = eax;
3916 		return;
3917 	}
3918 
3919 	if (!genuine_intel)
3920 		return;
3921 
3922 	if (max_level < 0xb)
3923 		return;
3924 
3925 	ecx = 0;
3926 	__cpuid(0xb, eax, ebx, ecx, edx);
3927 	t->x2apic_id = edx;
3928 
3929 	if (debug && (t->apic_id != (t->x2apic_id & 0xff)))
3930 		fprintf(outf, "cpu%d: BIOS BUG: apic 0x%x x2apic 0x%x\n", t->cpu_id, t->apic_id, t->x2apic_id);
3931 }
3932 
3933 int get_core_throt_cnt(int cpu, unsigned long long *cnt)
3934 {
3935 	char path[128 + PATH_BYTES];
3936 	unsigned long long tmp;
3937 	FILE *fp;
3938 	int ret;
3939 
3940 	sprintf(path, "/sys/devices/system/cpu/cpu%d/thermal_throttle/core_throttle_count", cpu);
3941 	fp = fopen(path, "r");
3942 	if (!fp)
3943 		return -1;
3944 	ret = fscanf(fp, "%lld", &tmp);
3945 	fclose(fp);
3946 	if (ret != 1)
3947 		return -1;
3948 	*cnt = tmp;
3949 
3950 	return 0;
3951 }
3952 
3953 struct amperf_group_fd {
3954 	int aperf;		/* Also the group descriptor */
3955 	int mperf;
3956 };
3957 
3958 static int read_perf_counter_info(const char *const path, const char *const parse_format, void *value_ptr)
3959 {
3960 	int fdmt;
3961 	int bytes_read;
3962 	char buf[64];
3963 	int ret = -1;
3964 
3965 	fdmt = open(path, O_RDONLY, 0);
3966 	if (fdmt == -1) {
3967 		if (debug)
3968 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
3969 		ret = -1;
3970 		goto cleanup_and_exit;
3971 	}
3972 
3973 	bytes_read = read(fdmt, buf, sizeof(buf) - 1);
3974 	if (bytes_read <= 0 || bytes_read >= (int)sizeof(buf)) {
3975 		if (debug)
3976 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
3977 		ret = -1;
3978 		goto cleanup_and_exit;
3979 	}
3980 
3981 	buf[bytes_read] = '\0';
3982 
3983 	if (sscanf(buf, parse_format, value_ptr) != 1) {
3984 		if (debug)
3985 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
3986 		ret = -1;
3987 		goto cleanup_and_exit;
3988 	}
3989 
3990 	ret = 0;
3991 
3992 cleanup_and_exit:
3993 	close(fdmt);
3994 	return ret;
3995 }
3996 
3997 static unsigned int read_perf_counter_info_n(const char *const path, const char *const parse_format)
3998 {
3999 	unsigned int v;
4000 	int status;
4001 
4002 	status = read_perf_counter_info(path, parse_format, &v);
4003 	if (status)
4004 		v = -1;
4005 
4006 	return v;
4007 }
4008 
4009 static unsigned int read_perf_type(const char *subsys)
4010 {
4011 	const char *const path_format = "/sys/bus/event_source/devices/%s/type";
4012 	const char *const format = "%u";
4013 	char path[128];
4014 
4015 	snprintf(path, sizeof(path), path_format, subsys);
4016 
4017 	return read_perf_counter_info_n(path, format);
4018 }
4019 
4020 static unsigned int read_perf_config(const char *subsys, const char *event_name)
4021 {
4022 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s";
4023 	FILE *fconfig = NULL;
4024 	char path[128];
4025 	char config_str[64];
4026 	unsigned int config;
4027 	unsigned int umask;
4028 	bool has_config = false;
4029 	bool has_umask = false;
4030 	unsigned int ret = -1;
4031 
4032 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4033 
4034 	fconfig = fopen(path, "r");
4035 	if (!fconfig)
4036 		return -1;
4037 
4038 	if (fgets(config_str, ARRAY_SIZE(config_str), fconfig) != config_str)
4039 		goto cleanup_and_exit;
4040 
4041 	for (char *pconfig_str = &config_str[0]; pconfig_str;) {
4042 		if (sscanf(pconfig_str, "event=%x", &config) == 1) {
4043 			has_config = true;
4044 			goto next;
4045 		}
4046 
4047 		if (sscanf(pconfig_str, "umask=%x", &umask) == 1) {
4048 			has_umask = true;
4049 			goto next;
4050 		}
4051 
4052 next:
4053 		pconfig_str = strchr(pconfig_str, ',');
4054 		if (pconfig_str) {
4055 			*pconfig_str = '\0';
4056 			++pconfig_str;
4057 		}
4058 	}
4059 
4060 	if (!has_umask)
4061 		umask = 0;
4062 
4063 	if (has_config)
4064 		ret = (umask << 8) | config;
4065 
4066 cleanup_and_exit:
4067 	fclose(fconfig);
4068 	return ret;
4069 }
4070 
4071 static unsigned int read_perf_rapl_unit(const char *subsys, const char *event_name)
4072 {
4073 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.unit";
4074 	const char *const format = "%s";
4075 	char path[128];
4076 	char unit_buffer[16];
4077 
4078 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4079 
4080 	read_perf_counter_info(path, format, &unit_buffer);
4081 	if (strcmp("Joules", unit_buffer) == 0)
4082 		return RAPL_UNIT_JOULES;
4083 
4084 	return RAPL_UNIT_INVALID;
4085 }
4086 
4087 static double read_perf_scale(const char *subsys, const char *event_name)
4088 {
4089 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.scale";
4090 	const char *const format = "%lf";
4091 	char path[128];
4092 	double scale;
4093 
4094 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4095 
4096 	if (read_perf_counter_info(path, format, &scale))
4097 		return 0.0;
4098 
4099 	return scale;
4100 }
4101 
4102 size_t rapl_counter_info_count_perf(const struct rapl_counter_info_t *rci)
4103 {
4104 	size_t ret = 0;
4105 
4106 	for (int i = 0; i < NUM_RAPL_COUNTERS; ++i)
4107 		if (rci->source[i] == COUNTER_SOURCE_PERF)
4108 			++ret;
4109 
4110 	return ret;
4111 }
4112 
4113 static size_t cstate_counter_info_count_perf(const struct cstate_counter_info_t *cci)
4114 {
4115 	size_t ret = 0;
4116 
4117 	for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i)
4118 		if (cci->source[i] == COUNTER_SOURCE_PERF)
4119 			++ret;
4120 
4121 	return ret;
4122 }
4123 
4124 void write_rapl_counter(struct rapl_counter *rc, struct rapl_counter_info_t *rci, unsigned int idx)
4125 {
4126 	rc->raw_value = rci->data[idx];
4127 	rc->unit = rci->unit[idx];
4128 	rc->scale = rci->scale[idx];
4129 }
4130 
4131 int get_rapl_counters(int cpu, unsigned int domain, struct core_data *c, struct pkg_data *p)
4132 {
4133 	unsigned long long perf_data[NUM_RAPL_COUNTERS + 1];
4134 	struct rapl_counter_info_t *rci;
4135 
4136 	if (debug >= 2)
4137 		fprintf(stderr, "%s: cpu%d domain%d\n", __func__, cpu, domain);
4138 
4139 	assert(rapl_counter_info_perdomain);
4140 	assert(domain < rapl_counter_info_perdomain_size);
4141 
4142 	rci = &rapl_counter_info_perdomain[domain];
4143 
4144 	/*
4145 	 * If we have any perf counters to read, read them all now, in bulk
4146 	 */
4147 	if (rci->fd_perf != -1) {
4148 		size_t num_perf_counters = rapl_counter_info_count_perf(rci);
4149 		const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
4150 		const ssize_t actual_read_size = read(rci->fd_perf, &perf_data[0], sizeof(perf_data));
4151 
4152 		if (actual_read_size != expected_read_size)
4153 			err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size,
4154 			    actual_read_size);
4155 	}
4156 
4157 	for (unsigned int i = 0, pi = 1; i < NUM_RAPL_COUNTERS; ++i) {
4158 		switch (rci->source[i]) {
4159 		case COUNTER_SOURCE_NONE:
4160 			break;
4161 
4162 		case COUNTER_SOURCE_PERF:
4163 			assert(pi < ARRAY_SIZE(perf_data));
4164 			assert(rci->fd_perf != -1);
4165 
4166 			if (debug >= 2)
4167 				fprintf(stderr, "Reading rapl counter via perf at %u (%llu %e %lf)\n",
4168 					i, perf_data[pi], rci->scale[i], perf_data[pi] * rci->scale[i]);
4169 
4170 			rci->data[i] = perf_data[pi];
4171 
4172 			++pi;
4173 			break;
4174 
4175 		case COUNTER_SOURCE_MSR:
4176 			if (debug >= 2)
4177 				fprintf(stderr, "Reading rapl counter via msr at %u\n", i);
4178 
4179 			assert(!no_msr);
4180 			if (rci->flags[i] & RAPL_COUNTER_FLAG_USE_MSR_SUM) {
4181 				if (get_msr_sum(cpu, rci->msr[i], &rci->data[i]))
4182 					return -13 - i;
4183 			} else {
4184 				if (get_msr(cpu, rci->msr[i], &rci->data[i]))
4185 					return -13 - i;
4186 			}
4187 
4188 			rci->data[i] &= rci->msr_mask[i];
4189 			if (rci->msr_shift[i] >= 0)
4190 				rci->data[i] >>= abs(rci->msr_shift[i]);
4191 			else
4192 				rci->data[i] <<= abs(rci->msr_shift[i]);
4193 
4194 			break;
4195 		}
4196 	}
4197 
4198 	BUILD_BUG_ON(NUM_RAPL_COUNTERS != 7);
4199 	write_rapl_counter(&p->energy_pkg, rci, RAPL_RCI_INDEX_ENERGY_PKG);
4200 	write_rapl_counter(&p->energy_cores, rci, RAPL_RCI_INDEX_ENERGY_CORES);
4201 	write_rapl_counter(&p->energy_dram, rci, RAPL_RCI_INDEX_DRAM);
4202 	write_rapl_counter(&p->energy_gfx, rci, RAPL_RCI_INDEX_GFX);
4203 	write_rapl_counter(&p->rapl_pkg_perf_status, rci, RAPL_RCI_INDEX_PKG_PERF_STATUS);
4204 	write_rapl_counter(&p->rapl_dram_perf_status, rci, RAPL_RCI_INDEX_DRAM_PERF_STATUS);
4205 	write_rapl_counter(&c->core_energy, rci, RAPL_RCI_INDEX_CORE_ENERGY);
4206 
4207 	return 0;
4208 }
4209 
4210 char *find_sysfs_path_by_id(struct sysfs_path *sp, int id)
4211 {
4212 	while (sp) {
4213 		if (sp->id == id)
4214 			return (sp->path);
4215 		sp = sp->next;
4216 	}
4217 	if (debug)
4218 		warnx("%s: id%d not found", __func__, id);
4219 	return NULL;
4220 }
4221 
4222 int get_cstate_counters(unsigned int cpu, struct thread_data *t, struct core_data *c, struct pkg_data *p)
4223 {
4224 	/*
4225 	 * Overcommit memory a little bit here,
4226 	 * but skip calculating exact sizes for the buffers.
4227 	 */
4228 	unsigned long long perf_data[NUM_CSTATE_COUNTERS];
4229 	unsigned long long perf_data_core[NUM_CSTATE_COUNTERS + 1];
4230 	unsigned long long perf_data_pkg[NUM_CSTATE_COUNTERS + 1];
4231 
4232 	struct cstate_counter_info_t *cci;
4233 
4234 	if (debug >= 2)
4235 		fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
4236 
4237 	assert(ccstate_counter_info);
4238 	assert(cpu <= ccstate_counter_info_size);
4239 
4240 	ZERO_ARRAY(perf_data);
4241 	ZERO_ARRAY(perf_data_core);
4242 	ZERO_ARRAY(perf_data_pkg);
4243 
4244 	cci = &ccstate_counter_info[cpu];
4245 
4246 	/*
4247 	 * If we have any perf counters to read, read them all now, in bulk
4248 	 */
4249 	const size_t num_perf_counters = cstate_counter_info_count_perf(cci);
4250 	ssize_t expected_read_size = num_perf_counters * sizeof(unsigned long long);
4251 	ssize_t actual_read_size_core = 0, actual_read_size_pkg = 0;
4252 
4253 	if (cci->fd_perf_core != -1) {
4254 		/* Each descriptor read begins with number of counters read. */
4255 		expected_read_size += sizeof(unsigned long long);
4256 
4257 		actual_read_size_core = read(cci->fd_perf_core, &perf_data_core[0], sizeof(perf_data_core));
4258 
4259 		if (actual_read_size_core <= 0)
4260 			err(-1, "%s: read perf %s: %ld", __func__, "core", actual_read_size_core);
4261 	}
4262 
4263 	if (cci->fd_perf_pkg != -1) {
4264 		/* Each descriptor read begins with number of counters read. */
4265 		expected_read_size += sizeof(unsigned long long);
4266 
4267 		actual_read_size_pkg = read(cci->fd_perf_pkg, &perf_data_pkg[0], sizeof(perf_data_pkg));
4268 
4269 		if (actual_read_size_pkg <= 0)
4270 			err(-1, "%s: read perf %s: %ld", __func__, "pkg", actual_read_size_pkg);
4271 	}
4272 
4273 	const ssize_t actual_read_size_total = actual_read_size_core + actual_read_size_pkg;
4274 
4275 	if (actual_read_size_total != expected_read_size)
4276 		err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size, actual_read_size_total);
4277 
4278 	/*
4279 	 * Copy ccstate and pcstate data into unified buffer.
4280 	 *
4281 	 * Skip first element from core and pkg buffers.
4282 	 * Kernel puts there how many counters were read.
4283 	 */
4284 	const size_t num_core_counters = perf_data_core[0];
4285 	const size_t num_pkg_counters = perf_data_pkg[0];
4286 
4287 	assert(num_perf_counters == num_core_counters + num_pkg_counters);
4288 
4289 	/* Copy ccstate perf data */
4290 	memcpy(&perf_data[0], &perf_data_core[1], num_core_counters * sizeof(unsigned long long));
4291 
4292 	/* Copy pcstate perf data */
4293 	memcpy(&perf_data[num_core_counters], &perf_data_pkg[1], num_pkg_counters * sizeof(unsigned long long));
4294 
4295 	for (unsigned int i = 0, pi = 0; i < NUM_CSTATE_COUNTERS; ++i) {
4296 		switch (cci->source[i]) {
4297 		case COUNTER_SOURCE_NONE:
4298 			break;
4299 
4300 		case COUNTER_SOURCE_PERF:
4301 			assert(pi < ARRAY_SIZE(perf_data));
4302 			assert(cci->fd_perf_core != -1 || cci->fd_perf_pkg != -1);
4303 
4304 			if (debug >= 2)
4305 				fprintf(stderr, "cstate via %s %u: %llu\n", "perf", i, perf_data[pi]);
4306 
4307 			cci->data[i] = perf_data[pi];
4308 
4309 			++pi;
4310 			break;
4311 
4312 		case COUNTER_SOURCE_MSR:
4313 			assert(!no_msr);
4314 			if (get_msr(cpu, cci->msr[i], &cci->data[i]))
4315 				return -13 - i;
4316 
4317 			if (debug >= 2)
4318 				fprintf(stderr, "cstate via %s0x%llx %u: %llu\n", "msr", cci->msr[i], i, cci->data[i]);
4319 
4320 			break;
4321 		}
4322 	}
4323 
4324 	/*
4325 	 * Helper to write the data only if the source of
4326 	 * the counter for the current cpu is not none.
4327 	 *
4328 	 * Otherwise we would overwrite core data with 0 (default value),
4329 	 * when invoked for the thread sibling.
4330 	 */
4331 #define PERF_COUNTER_WRITE_DATA(out_counter, index) do {	\
4332 	if (cci->source[index] != COUNTER_SOURCE_NONE)		\
4333 		out_counter = cci->data[index];			\
4334 } while (0)
4335 
4336 	BUILD_BUG_ON(NUM_CSTATE_COUNTERS != 11);
4337 
4338 	PERF_COUNTER_WRITE_DATA(t->c1, CCSTATE_RCI_INDEX_C1_RESIDENCY);
4339 	PERF_COUNTER_WRITE_DATA(c->c3, CCSTATE_RCI_INDEX_C3_RESIDENCY);
4340 	PERF_COUNTER_WRITE_DATA(c->c6, CCSTATE_RCI_INDEX_C6_RESIDENCY);
4341 	PERF_COUNTER_WRITE_DATA(c->c7, CCSTATE_RCI_INDEX_C7_RESIDENCY);
4342 
4343 	PERF_COUNTER_WRITE_DATA(p->pc2, PCSTATE_RCI_INDEX_C2_RESIDENCY);
4344 	PERF_COUNTER_WRITE_DATA(p->pc3, PCSTATE_RCI_INDEX_C3_RESIDENCY);
4345 	PERF_COUNTER_WRITE_DATA(p->pc6, PCSTATE_RCI_INDEX_C6_RESIDENCY);
4346 	PERF_COUNTER_WRITE_DATA(p->pc7, PCSTATE_RCI_INDEX_C7_RESIDENCY);
4347 	PERF_COUNTER_WRITE_DATA(p->pc8, PCSTATE_RCI_INDEX_C8_RESIDENCY);
4348 	PERF_COUNTER_WRITE_DATA(p->pc9, PCSTATE_RCI_INDEX_C9_RESIDENCY);
4349 	PERF_COUNTER_WRITE_DATA(p->pc10, PCSTATE_RCI_INDEX_C10_RESIDENCY);
4350 
4351 #undef PERF_COUNTER_WRITE_DATA
4352 
4353 	return 0;
4354 }
4355 
4356 size_t msr_counter_info_count_perf(const struct msr_counter_info_t *mci)
4357 {
4358 	size_t ret = 0;
4359 
4360 	for (int i = 0; i < NUM_MSR_COUNTERS; ++i)
4361 		if (mci->source[i] == COUNTER_SOURCE_PERF)
4362 			++ret;
4363 
4364 	return ret;
4365 }
4366 
4367 int get_smi_aperf_mperf(unsigned int cpu, struct thread_data *t)
4368 {
4369 	unsigned long long perf_data[NUM_MSR_COUNTERS + 1];
4370 
4371 	struct msr_counter_info_t *mci;
4372 
4373 	if (debug >= 2)
4374 		fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
4375 
4376 	assert(msr_counter_info);
4377 	assert(cpu <= msr_counter_info_size);
4378 
4379 	mci = &msr_counter_info[cpu];
4380 
4381 	ZERO_ARRAY(perf_data);
4382 	ZERO_ARRAY(mci->data);
4383 
4384 	if (mci->fd_perf != -1) {
4385 		const size_t num_perf_counters = msr_counter_info_count_perf(mci);
4386 		const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
4387 		const ssize_t actual_read_size = read(mci->fd_perf, &perf_data[0], sizeof(perf_data));
4388 
4389 		if (actual_read_size != expected_read_size)
4390 			err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size,
4391 			    actual_read_size);
4392 	}
4393 
4394 	for (unsigned int i = 0, pi = 1; i < NUM_MSR_COUNTERS; ++i) {
4395 		switch (mci->source[i]) {
4396 		case COUNTER_SOURCE_NONE:
4397 			break;
4398 
4399 		case COUNTER_SOURCE_PERF:
4400 			assert(pi < ARRAY_SIZE(perf_data));
4401 			assert(mci->fd_perf != -1);
4402 
4403 			if (debug >= 2)
4404 				fprintf(stderr, "Reading msr counter via perf at %u: %llu\n", i, perf_data[pi]);
4405 
4406 			mci->data[i] = perf_data[pi];
4407 
4408 			++pi;
4409 			break;
4410 
4411 		case COUNTER_SOURCE_MSR:
4412 			assert(!no_msr);
4413 
4414 			if (get_msr(cpu, mci->msr[i], &mci->data[i]))
4415 				return -2 - i;
4416 
4417 			mci->data[i] &= mci->msr_mask[i];
4418 
4419 			if (debug >= 2)
4420 				fprintf(stderr, "Reading msr counter via msr at %u: %llu\n", i, mci->data[i]);
4421 
4422 			break;
4423 		}
4424 	}
4425 
4426 	BUILD_BUG_ON(NUM_MSR_COUNTERS != 3);
4427 	t->aperf = mci->data[MSR_RCI_INDEX_APERF];
4428 	t->mperf = mci->data[MSR_RCI_INDEX_MPERF];
4429 	t->smi_count = mci->data[MSR_RCI_INDEX_SMI];
4430 
4431 	return 0;
4432 }
4433 
4434 int perf_counter_info_read_values(struct perf_counter_info *pp, int cpu, unsigned long long *out, size_t out_size)
4435 {
4436 	unsigned int domain;
4437 	unsigned long long value;
4438 	int fd_counter;
4439 
4440 	for (size_t i = 0; pp; ++i, pp = pp->next) {
4441 		domain = cpu_to_domain(pp, cpu);
4442 		assert(domain < pp->num_domains);
4443 
4444 		fd_counter = pp->fd_perf_per_domain[domain];
4445 
4446 		if (fd_counter == -1)
4447 			continue;
4448 
4449 		if (read(fd_counter, &value, sizeof(value)) != sizeof(value))
4450 			return 1;
4451 
4452 		assert(i < out_size);
4453 		out[i] = value * pp->scale;
4454 	}
4455 
4456 	return 0;
4457 }
4458 
4459 unsigned long pmt_gen_value_mask(unsigned int lsb, unsigned int msb)
4460 {
4461 	unsigned long mask;
4462 
4463 	if (msb == 63)
4464 		mask = 0xffffffffffffffff;
4465 	else
4466 		mask = ((1 << (msb + 1)) - 1);
4467 
4468 	mask -= (1 << lsb) - 1;
4469 
4470 	return mask;
4471 }
4472 
4473 unsigned long pmt_read_counter(struct pmt_counter *ppmt, unsigned int domain_id)
4474 {
4475 	assert(domain_id < ppmt->num_domains);
4476 
4477 	const unsigned long *pmmio = ppmt->domains[domain_id].pcounter;
4478 	const unsigned long value = pmmio ? *pmmio : 0;
4479 	const unsigned long value_mask = pmt_gen_value_mask(ppmt->lsb, ppmt->msb);
4480 	const unsigned long value_shift = ppmt->lsb;
4481 
4482 	return (value & value_mask) >> value_shift;
4483 }
4484 
4485 /*
4486  * get_counters(...)
4487  * migrate to cpu
4488  * acquire and record local counters for that cpu
4489  */
4490 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4491 {
4492 	int cpu = t->cpu_id;
4493 	unsigned long long msr;
4494 	struct msr_counter *mp;
4495 	struct pmt_counter *pp;
4496 	int i;
4497 	int status;
4498 
4499 	if (cpu_migrate(cpu)) {
4500 		fprintf(outf, "%s: Could not migrate to CPU %d\n", __func__, cpu);
4501 		return -1;
4502 	}
4503 
4504 	gettimeofday(&t->tv_begin, (struct timezone *)NULL);
4505 
4506 	if (first_counter_read)
4507 		get_apic_id(t);
4508 
4509 	t->tsc = rdtsc();	/* we are running on local CPU of interest */
4510 
4511 	get_smi_aperf_mperf(cpu, t);
4512 
4513 	if (DO_BIC(BIC_IPC))
4514 		if (read(get_instr_count_fd(cpu), &t->instr_count, sizeof(long long)) != sizeof(long long))
4515 			return -4;
4516 
4517 	if (DO_BIC(BIC_IRQ))
4518 		t->irq_count = irqs_per_cpu[cpu];
4519 
4520 	get_cstate_counters(cpu, t, c, p);
4521 
4522 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
4523 		if (get_mp(cpu, mp, &t->counter[i], mp->sp->path))
4524 			return -10;
4525 	}
4526 
4527 	if (perf_counter_info_read_values(sys.perf_tp, cpu, t->perf_counter, MAX_ADDED_THREAD_COUNTERS))
4528 		return -10;
4529 
4530 	for (i = 0, pp = sys.pmt_tp; pp; i++, pp = pp->next)
4531 		t->pmt_counter[i] = pmt_read_counter(pp, t->cpu_id);
4532 
4533 	/* collect core counters only for 1st thread in core */
4534 	if (!is_cpu_first_thread_in_core(t, c, p))
4535 		goto done;
4536 
4537 	if (platform->has_per_core_rapl) {
4538 		status = get_rapl_counters(cpu, c->core_id, c, p);
4539 		if (status != 0)
4540 			return status;
4541 	}
4542 
4543 	if (DO_BIC(BIC_CPU_c7) && t->is_atom) {
4544 		/*
4545 		 * For Atom CPUs that has core cstate deeper than c6,
4546 		 * MSR_CORE_C6_RESIDENCY returns residency of cc6 and deeper.
4547 		 * Minus CC7 (and deeper cstates) residency to get
4548 		 * accturate cc6 residency.
4549 		 */
4550 		c->c6 -= c->c7;
4551 	}
4552 
4553 	if (DO_BIC(BIC_Mod_c6))
4554 		if (get_msr(cpu, MSR_MODULE_C6_RES_MS, &c->mc6_us))
4555 			return -8;
4556 
4557 	if (DO_BIC(BIC_CoreTmp)) {
4558 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
4559 			return -9;
4560 		c->core_temp_c = tj_max - ((msr >> 16) & 0x7F);
4561 	}
4562 
4563 	if (DO_BIC(BIC_CORE_THROT_CNT))
4564 		get_core_throt_cnt(cpu, &c->core_throt_cnt);
4565 
4566 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
4567 		if (get_mp(cpu, mp, &c->counter[i], mp->sp->path))
4568 			return -10;
4569 	}
4570 
4571 	if (perf_counter_info_read_values(sys.perf_cp, cpu, c->perf_counter, MAX_ADDED_CORE_COUNTERS))
4572 		return -10;
4573 
4574 	for (i = 0, pp = sys.pmt_cp; pp; i++, pp = pp->next)
4575 		c->pmt_counter[i] = pmt_read_counter(pp, c->core_id);
4576 
4577 	/* collect package counters only for 1st core in package */
4578 	if (!is_cpu_first_core_in_package(t, c, p))
4579 		goto done;
4580 
4581 	if (DO_BIC(BIC_Totl_c0)) {
4582 		if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
4583 			return -10;
4584 	}
4585 	if (DO_BIC(BIC_Any_c0)) {
4586 		if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
4587 			return -11;
4588 	}
4589 	if (DO_BIC(BIC_GFX_c0)) {
4590 		if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
4591 			return -12;
4592 	}
4593 	if (DO_BIC(BIC_CPUGFX)) {
4594 		if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
4595 			return -13;
4596 	}
4597 
4598 	if (DO_BIC(BIC_CPU_LPI))
4599 		p->cpu_lpi = cpuidle_cur_cpu_lpi_us;
4600 	if (DO_BIC(BIC_SYS_LPI))
4601 		p->sys_lpi = cpuidle_cur_sys_lpi_us;
4602 
4603 	if (!platform->has_per_core_rapl) {
4604 		status = get_rapl_counters(cpu, p->package_id, c, p);
4605 		if (status != 0)
4606 			return status;
4607 	}
4608 
4609 	if (DO_BIC(BIC_PkgTmp)) {
4610 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
4611 			return -17;
4612 		p->pkg_temp_c = tj_max - ((msr >> 16) & 0x7F);
4613 	}
4614 
4615 	if (DO_BIC(BIC_UNCORE_MHZ))
4616 		p->uncore_mhz = get_legacy_uncore_mhz(p->package_id);
4617 
4618 	if (DO_BIC(BIC_GFX_rc6))
4619 		p->gfx_rc6_ms = gfx_info[GFX_rc6].val_ull;
4620 
4621 	if (DO_BIC(BIC_GFXMHz))
4622 		p->gfx_mhz = gfx_info[GFX_MHz].val;
4623 
4624 	if (DO_BIC(BIC_GFXACTMHz))
4625 		p->gfx_act_mhz = gfx_info[GFX_ACTMHz].val;
4626 
4627 	if (DO_BIC(BIC_SAM_mc6))
4628 		p->sam_mc6_ms = gfx_info[SAM_mc6].val_ull;
4629 
4630 	if (DO_BIC(BIC_SAMMHz))
4631 		p->sam_mhz = gfx_info[SAM_MHz].val;
4632 
4633 	if (DO_BIC(BIC_SAMACTMHz))
4634 		p->sam_act_mhz = gfx_info[SAM_ACTMHz].val;
4635 
4636 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
4637 		char *path = NULL;
4638 
4639 		if (mp->msr_num == 0) {
4640 			path = find_sysfs_path_by_id(mp->sp, p->package_id);
4641 			if (path == NULL) {
4642 				warnx("%s: package_id %d not found", __func__, p->package_id);
4643 				return -10;
4644 			}
4645 		}
4646 		if (get_mp(cpu, mp, &p->counter[i], path))
4647 			return -10;
4648 	}
4649 
4650 	if (perf_counter_info_read_values(sys.perf_pp, cpu, p->perf_counter, MAX_ADDED_PACKAGE_COUNTERS))
4651 		return -10;
4652 
4653 	for (i = 0, pp = sys.pmt_pp; pp; i++, pp = pp->next)
4654 		p->pmt_counter[i] = pmt_read_counter(pp, p->package_id);
4655 
4656 done:
4657 	gettimeofday(&t->tv_end, (struct timezone *)NULL);
4658 
4659 	return 0;
4660 }
4661 
4662 int pkg_cstate_limit = PCLUKN;
4663 char *pkg_cstate_limit_strings[] = { "unknown", "reserved", "pc0", "pc1", "pc2",
4664 	"pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "pc10", "unlimited"
4665 };
4666 
4667 int nhm_pkg_cstate_limits[16] =
4668     { PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4669 	PCLRSV, PCLRSV
4670 };
4671 
4672 int snb_pkg_cstate_limits[16] =
4673     { PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4674 	PCLRSV, PCLRSV
4675 };
4676 
4677 int hsw_pkg_cstate_limits[16] =
4678     { PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4679 	PCLRSV, PCLRSV
4680 };
4681 
4682 int slv_pkg_cstate_limits[16] =
4683     { PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4684 	PCL__6, PCL__7
4685 };
4686 
4687 int amt_pkg_cstate_limits[16] =
4688     { PCLUNL, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4689 	PCLRSV, PCLRSV
4690 };
4691 
4692 int phi_pkg_cstate_limits[16] =
4693     { PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4694 	PCLRSV, PCLRSV
4695 };
4696 
4697 int glm_pkg_cstate_limits[16] =
4698     { PCLUNL, PCL__1, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCL_10, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4699 	PCLRSV, PCLRSV
4700 };
4701 
4702 int skx_pkg_cstate_limits[16] =
4703     { PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4704 	PCLRSV, PCLRSV
4705 };
4706 
4707 int icx_pkg_cstate_limits[16] =
4708     { PCL__0, PCL__2, PCL__6, PCL__6, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4709 	PCLRSV, PCLRSV
4710 };
4711 
4712 void probe_cst_limit(void)
4713 {
4714 	unsigned long long msr;
4715 	int *pkg_cstate_limits;
4716 
4717 	if (!platform->has_nhm_msrs || no_msr)
4718 		return;
4719 
4720 	switch (platform->cst_limit) {
4721 	case CST_LIMIT_NHM:
4722 		pkg_cstate_limits = nhm_pkg_cstate_limits;
4723 		break;
4724 	case CST_LIMIT_SNB:
4725 		pkg_cstate_limits = snb_pkg_cstate_limits;
4726 		break;
4727 	case CST_LIMIT_HSW:
4728 		pkg_cstate_limits = hsw_pkg_cstate_limits;
4729 		break;
4730 	case CST_LIMIT_SKX:
4731 		pkg_cstate_limits = skx_pkg_cstate_limits;
4732 		break;
4733 	case CST_LIMIT_ICX:
4734 		pkg_cstate_limits = icx_pkg_cstate_limits;
4735 		break;
4736 	case CST_LIMIT_SLV:
4737 		pkg_cstate_limits = slv_pkg_cstate_limits;
4738 		break;
4739 	case CST_LIMIT_AMT:
4740 		pkg_cstate_limits = amt_pkg_cstate_limits;
4741 		break;
4742 	case CST_LIMIT_KNL:
4743 		pkg_cstate_limits = phi_pkg_cstate_limits;
4744 		break;
4745 	case CST_LIMIT_GMT:
4746 		pkg_cstate_limits = glm_pkg_cstate_limits;
4747 		break;
4748 	default:
4749 		return;
4750 	}
4751 
4752 	get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
4753 	pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
4754 }
4755 
4756 static void dump_platform_info(void)
4757 {
4758 	unsigned long long msr;
4759 	unsigned int ratio;
4760 
4761 	if (!platform->has_nhm_msrs || no_msr)
4762 		return;
4763 
4764 	get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
4765 
4766 	fprintf(outf, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", base_cpu, msr);
4767 
4768 	ratio = (msr >> 40) & 0xFF;
4769 	fprintf(outf, "%d * %.1f = %.1f MHz max efficiency frequency\n", ratio, bclk, ratio * bclk);
4770 
4771 	ratio = (msr >> 8) & 0xFF;
4772 	fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n", ratio, bclk, ratio * bclk);
4773 }
4774 
4775 static void dump_power_ctl(void)
4776 {
4777 	unsigned long long msr;
4778 
4779 	if (!platform->has_nhm_msrs || no_msr)
4780 		return;
4781 
4782 	get_msr(base_cpu, MSR_IA32_POWER_CTL, &msr);
4783 	fprintf(outf, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
4784 		base_cpu, msr, msr & 0x2 ? "EN" : "DIS");
4785 
4786 	/* C-state Pre-wake Disable (CSTATE_PREWAKE_DISABLE) */
4787 	if (platform->has_cst_prewake_bit)
4788 		fprintf(outf, "C-state Pre-wake: %sabled\n", msr & 0x40000000 ? "DIS" : "EN");
4789 
4790 	return;
4791 }
4792 
4793 static void dump_turbo_ratio_limit2(void)
4794 {
4795 	unsigned long long msr;
4796 	unsigned int ratio;
4797 
4798 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
4799 
4800 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu, msr);
4801 
4802 	ratio = (msr >> 8) & 0xFF;
4803 	if (ratio)
4804 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 18 active cores\n", ratio, bclk, ratio * bclk);
4805 
4806 	ratio = (msr >> 0) & 0xFF;
4807 	if (ratio)
4808 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 17 active cores\n", ratio, bclk, ratio * bclk);
4809 	return;
4810 }
4811 
4812 static void dump_turbo_ratio_limit1(void)
4813 {
4814 	unsigned long long msr;
4815 	unsigned int ratio;
4816 
4817 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
4818 
4819 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, msr);
4820 
4821 	ratio = (msr >> 56) & 0xFF;
4822 	if (ratio)
4823 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 16 active cores\n", ratio, bclk, ratio * bclk);
4824 
4825 	ratio = (msr >> 48) & 0xFF;
4826 	if (ratio)
4827 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 15 active cores\n", ratio, bclk, ratio * bclk);
4828 
4829 	ratio = (msr >> 40) & 0xFF;
4830 	if (ratio)
4831 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 14 active cores\n", ratio, bclk, ratio * bclk);
4832 
4833 	ratio = (msr >> 32) & 0xFF;
4834 	if (ratio)
4835 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 13 active cores\n", ratio, bclk, ratio * bclk);
4836 
4837 	ratio = (msr >> 24) & 0xFF;
4838 	if (ratio)
4839 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 12 active cores\n", ratio, bclk, ratio * bclk);
4840 
4841 	ratio = (msr >> 16) & 0xFF;
4842 	if (ratio)
4843 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 11 active cores\n", ratio, bclk, ratio * bclk);
4844 
4845 	ratio = (msr >> 8) & 0xFF;
4846 	if (ratio)
4847 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 10 active cores\n", ratio, bclk, ratio * bclk);
4848 
4849 	ratio = (msr >> 0) & 0xFF;
4850 	if (ratio)
4851 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 9 active cores\n", ratio, bclk, ratio * bclk);
4852 	return;
4853 }
4854 
4855 static void dump_turbo_ratio_limits(int trl_msr_offset)
4856 {
4857 	unsigned long long msr, core_counts;
4858 	int shift;
4859 
4860 	get_msr(base_cpu, trl_msr_offset, &msr);
4861 	fprintf(outf, "cpu%d: MSR_%sTURBO_RATIO_LIMIT: 0x%08llx\n",
4862 		base_cpu, trl_msr_offset == MSR_SECONDARY_TURBO_RATIO_LIMIT ? "SECONDARY_" : "", msr);
4863 
4864 	if (platform->trl_msrs & TRL_CORECOUNT) {
4865 		get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &core_counts);
4866 		fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, core_counts);
4867 	} else {
4868 		core_counts = 0x0807060504030201;
4869 	}
4870 
4871 	for (shift = 56; shift >= 0; shift -= 8) {
4872 		unsigned int ratio, group_size;
4873 
4874 		ratio = (msr >> shift) & 0xFF;
4875 		group_size = (core_counts >> shift) & 0xFF;
4876 		if (ratio)
4877 			fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
4878 				ratio, bclk, ratio * bclk, group_size);
4879 	}
4880 
4881 	return;
4882 }
4883 
4884 static void dump_atom_turbo_ratio_limits(void)
4885 {
4886 	unsigned long long msr;
4887 	unsigned int ratio;
4888 
4889 	get_msr(base_cpu, MSR_ATOM_CORE_RATIOS, &msr);
4890 	fprintf(outf, "cpu%d: MSR_ATOM_CORE_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
4891 
4892 	ratio = (msr >> 0) & 0x3F;
4893 	if (ratio)
4894 		fprintf(outf, "%d * %.1f = %.1f MHz minimum operating frequency\n", ratio, bclk, ratio * bclk);
4895 
4896 	ratio = (msr >> 8) & 0x3F;
4897 	if (ratio)
4898 		fprintf(outf, "%d * %.1f = %.1f MHz low frequency mode (LFM)\n", ratio, bclk, ratio * bclk);
4899 
4900 	ratio = (msr >> 16) & 0x3F;
4901 	if (ratio)
4902 		fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n", ratio, bclk, ratio * bclk);
4903 
4904 	get_msr(base_cpu, MSR_ATOM_CORE_TURBO_RATIOS, &msr);
4905 	fprintf(outf, "cpu%d: MSR_ATOM_CORE_TURBO_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
4906 
4907 	ratio = (msr >> 24) & 0x3F;
4908 	if (ratio)
4909 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 4 active cores\n", ratio, bclk, ratio * bclk);
4910 
4911 	ratio = (msr >> 16) & 0x3F;
4912 	if (ratio)
4913 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 3 active cores\n", ratio, bclk, ratio * bclk);
4914 
4915 	ratio = (msr >> 8) & 0x3F;
4916 	if (ratio)
4917 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 2 active cores\n", ratio, bclk, ratio * bclk);
4918 
4919 	ratio = (msr >> 0) & 0x3F;
4920 	if (ratio)
4921 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 1 active core\n", ratio, bclk, ratio * bclk);
4922 }
4923 
4924 static void dump_knl_turbo_ratio_limits(void)
4925 {
4926 	const unsigned int buckets_no = 7;
4927 
4928 	unsigned long long msr;
4929 	int delta_cores, delta_ratio;
4930 	int i, b_nr;
4931 	unsigned int cores[buckets_no];
4932 	unsigned int ratio[buckets_no];
4933 
4934 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
4935 
4936 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu, msr);
4937 
4938 	/*
4939 	 * Turbo encoding in KNL is as follows:
4940 	 * [0] -- Reserved
4941 	 * [7:1] -- Base value of number of active cores of bucket 1.
4942 	 * [15:8] -- Base value of freq ratio of bucket 1.
4943 	 * [20:16] -- +ve delta of number of active cores of bucket 2.
4944 	 * i.e. active cores of bucket 2 =
4945 	 * active cores of bucket 1 + delta
4946 	 * [23:21] -- Negative delta of freq ratio of bucket 2.
4947 	 * i.e. freq ratio of bucket 2 =
4948 	 * freq ratio of bucket 1 - delta
4949 	 * [28:24]-- +ve delta of number of active cores of bucket 3.
4950 	 * [31:29]-- -ve delta of freq ratio of bucket 3.
4951 	 * [36:32]-- +ve delta of number of active cores of bucket 4.
4952 	 * [39:37]-- -ve delta of freq ratio of bucket 4.
4953 	 * [44:40]-- +ve delta of number of active cores of bucket 5.
4954 	 * [47:45]-- -ve delta of freq ratio of bucket 5.
4955 	 * [52:48]-- +ve delta of number of active cores of bucket 6.
4956 	 * [55:53]-- -ve delta of freq ratio of bucket 6.
4957 	 * [60:56]-- +ve delta of number of active cores of bucket 7.
4958 	 * [63:61]-- -ve delta of freq ratio of bucket 7.
4959 	 */
4960 
4961 	b_nr = 0;
4962 	cores[b_nr] = (msr & 0xFF) >> 1;
4963 	ratio[b_nr] = (msr >> 8) & 0xFF;
4964 
4965 	for (i = 16; i < 64; i += 8) {
4966 		delta_cores = (msr >> i) & 0x1F;
4967 		delta_ratio = (msr >> (i + 5)) & 0x7;
4968 
4969 		cores[b_nr + 1] = cores[b_nr] + delta_cores;
4970 		ratio[b_nr + 1] = ratio[b_nr] - delta_ratio;
4971 		b_nr++;
4972 	}
4973 
4974 	for (i = buckets_no - 1; i >= 0; i--)
4975 		if (i > 0 ? ratio[i] != ratio[i - 1] : 1)
4976 			fprintf(outf,
4977 				"%d * %.1f = %.1f MHz max turbo %d active cores\n",
4978 				ratio[i], bclk, ratio[i] * bclk, cores[i]);
4979 }
4980 
4981 static void dump_cst_cfg(void)
4982 {
4983 	unsigned long long msr;
4984 
4985 	if (!platform->has_nhm_msrs || no_msr)
4986 		return;
4987 
4988 	get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
4989 
4990 	fprintf(outf, "cpu%d: MSR_PKG_CST_CONFIG_CONTROL: 0x%08llx", base_cpu, msr);
4991 
4992 	fprintf(outf, " (%s%s%s%s%slocked, pkg-cstate-limit=%d (%s)",
4993 		(msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
4994 		(msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
4995 		(msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
4996 		(msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
4997 		(msr & (1 << 15)) ? "" : "UN", (unsigned int)msr & 0xF, pkg_cstate_limit_strings[pkg_cstate_limit]);
4998 
4999 #define AUTOMATIC_CSTATE_CONVERSION		(1UL << 16)
5000 	if (platform->has_cst_auto_convension) {
5001 		fprintf(outf, ", automatic c-state conversion=%s", (msr & AUTOMATIC_CSTATE_CONVERSION) ? "on" : "off");
5002 	}
5003 
5004 	fprintf(outf, ")\n");
5005 
5006 	return;
5007 }
5008 
5009 static void dump_config_tdp(void)
5010 {
5011 	unsigned long long msr;
5012 
5013 	get_msr(base_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
5014 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu, msr);
5015 	fprintf(outf, " (base_ratio=%d)\n", (unsigned int)msr & 0xFF);
5016 
5017 	get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
5018 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu, msr);
5019 	if (msr) {
5020 		fprintf(outf, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
5021 		fprintf(outf, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
5022 		fprintf(outf, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
5023 		fprintf(outf, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0x7FFF);
5024 	}
5025 	fprintf(outf, ")\n");
5026 
5027 	get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
5028 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu, msr);
5029 	if (msr) {
5030 		fprintf(outf, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
5031 		fprintf(outf, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
5032 		fprintf(outf, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
5033 		fprintf(outf, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0x7FFF);
5034 	}
5035 	fprintf(outf, ")\n");
5036 
5037 	get_msr(base_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
5038 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu, msr);
5039 	if ((msr) & 0x3)
5040 		fprintf(outf, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
5041 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
5042 	fprintf(outf, ")\n");
5043 
5044 	get_msr(base_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
5045 	fprintf(outf, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu, msr);
5046 	fprintf(outf, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xFF);
5047 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
5048 	fprintf(outf, ")\n");
5049 }
5050 
5051 unsigned int irtl_time_units[] = { 1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
5052 
5053 void print_irtl(void)
5054 {
5055 	unsigned long long msr;
5056 
5057 	if (!platform->has_irtl_msrs || no_msr)
5058 		return;
5059 
5060 	if (platform->supported_cstates & PC3) {
5061 		get_msr(base_cpu, MSR_PKGC3_IRTL, &msr);
5062 		fprintf(outf, "cpu%d: MSR_PKGC3_IRTL: 0x%08llx (", base_cpu, msr);
5063 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5064 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5065 	}
5066 
5067 	if (platform->supported_cstates & PC6) {
5068 		get_msr(base_cpu, MSR_PKGC6_IRTL, &msr);
5069 		fprintf(outf, "cpu%d: MSR_PKGC6_IRTL: 0x%08llx (", base_cpu, msr);
5070 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5071 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5072 	}
5073 
5074 	if (platform->supported_cstates & PC7) {
5075 		get_msr(base_cpu, MSR_PKGC7_IRTL, &msr);
5076 		fprintf(outf, "cpu%d: MSR_PKGC7_IRTL: 0x%08llx (", base_cpu, msr);
5077 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5078 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5079 	}
5080 
5081 	if (platform->supported_cstates & PC8) {
5082 		get_msr(base_cpu, MSR_PKGC8_IRTL, &msr);
5083 		fprintf(outf, "cpu%d: MSR_PKGC8_IRTL: 0x%08llx (", base_cpu, msr);
5084 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5085 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5086 	}
5087 
5088 	if (platform->supported_cstates & PC9) {
5089 		get_msr(base_cpu, MSR_PKGC9_IRTL, &msr);
5090 		fprintf(outf, "cpu%d: MSR_PKGC9_IRTL: 0x%08llx (", base_cpu, msr);
5091 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5092 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5093 	}
5094 
5095 	if (platform->supported_cstates & PC10) {
5096 		get_msr(base_cpu, MSR_PKGC10_IRTL, &msr);
5097 		fprintf(outf, "cpu%d: MSR_PKGC10_IRTL: 0x%08llx (", base_cpu, msr);
5098 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5099 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5100 	}
5101 }
5102 
5103 void free_fd_percpu(void)
5104 {
5105 	int i;
5106 
5107 	if (!fd_percpu)
5108 		return;
5109 
5110 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
5111 		if (fd_percpu[i] != 0)
5112 			close(fd_percpu[i]);
5113 	}
5114 
5115 	free(fd_percpu);
5116 	fd_percpu = NULL;
5117 }
5118 
5119 void free_fd_instr_count_percpu(void)
5120 {
5121 	if (!fd_instr_count_percpu)
5122 		return;
5123 
5124 	for (int i = 0; i < topo.max_cpu_num + 1; ++i) {
5125 		if (fd_instr_count_percpu[i] != 0)
5126 			close(fd_instr_count_percpu[i]);
5127 	}
5128 
5129 	free(fd_instr_count_percpu);
5130 	fd_instr_count_percpu = NULL;
5131 }
5132 
5133 void free_fd_cstate(void)
5134 {
5135 	if (!ccstate_counter_info)
5136 		return;
5137 
5138 	const int counter_info_num = ccstate_counter_info_size;
5139 
5140 	for (int counter_id = 0; counter_id < counter_info_num; ++counter_id) {
5141 		if (ccstate_counter_info[counter_id].fd_perf_core != -1)
5142 			close(ccstate_counter_info[counter_id].fd_perf_core);
5143 
5144 		if (ccstate_counter_info[counter_id].fd_perf_pkg != -1)
5145 			close(ccstate_counter_info[counter_id].fd_perf_pkg);
5146 	}
5147 
5148 	free(ccstate_counter_info);
5149 	ccstate_counter_info = NULL;
5150 	ccstate_counter_info_size = 0;
5151 }
5152 
5153 void free_fd_msr(void)
5154 {
5155 	if (!msr_counter_info)
5156 		return;
5157 
5158 	for (int cpu = 0; cpu < topo.max_cpu_num; ++cpu) {
5159 		if (msr_counter_info[cpu].fd_perf != -1)
5160 			close(msr_counter_info[cpu].fd_perf);
5161 	}
5162 
5163 	free(msr_counter_info);
5164 	msr_counter_info = NULL;
5165 	msr_counter_info_size = 0;
5166 }
5167 
5168 void free_fd_rapl_percpu(void)
5169 {
5170 	if (!rapl_counter_info_perdomain)
5171 		return;
5172 
5173 	const int num_domains = rapl_counter_info_perdomain_size;
5174 
5175 	for (int domain_id = 0; domain_id < num_domains; ++domain_id) {
5176 		if (rapl_counter_info_perdomain[domain_id].fd_perf != -1)
5177 			close(rapl_counter_info_perdomain[domain_id].fd_perf);
5178 	}
5179 
5180 	free(rapl_counter_info_perdomain);
5181 	rapl_counter_info_perdomain = NULL;
5182 	rapl_counter_info_perdomain_size = 0;
5183 }
5184 
5185 void free_fd_added_perf_counters_(struct perf_counter_info *pp)
5186 {
5187 	if (!pp)
5188 		return;
5189 
5190 	if (!pp->fd_perf_per_domain)
5191 		return;
5192 
5193 	while (pp) {
5194 		for (size_t domain = 0; domain < pp->num_domains; ++domain) {
5195 			if (pp->fd_perf_per_domain[domain] != -1) {
5196 				close(pp->fd_perf_per_domain[domain]);
5197 				pp->fd_perf_per_domain[domain] = -1;
5198 			}
5199 		}
5200 
5201 		free(pp->fd_perf_per_domain);
5202 		pp->fd_perf_per_domain = NULL;
5203 
5204 		pp = pp->next;
5205 	}
5206 }
5207 
5208 void free_fd_added_perf_counters(void)
5209 {
5210 	free_fd_added_perf_counters_(sys.perf_tp);
5211 	free_fd_added_perf_counters_(sys.perf_cp);
5212 	free_fd_added_perf_counters_(sys.perf_pp);
5213 }
5214 
5215 void free_all_buffers(void)
5216 {
5217 	int i;
5218 
5219 	CPU_FREE(cpu_present_set);
5220 	cpu_present_set = NULL;
5221 	cpu_present_setsize = 0;
5222 
5223 	CPU_FREE(cpu_effective_set);
5224 	cpu_effective_set = NULL;
5225 	cpu_effective_setsize = 0;
5226 
5227 	CPU_FREE(cpu_allowed_set);
5228 	cpu_allowed_set = NULL;
5229 	cpu_allowed_setsize = 0;
5230 
5231 	CPU_FREE(cpu_affinity_set);
5232 	cpu_affinity_set = NULL;
5233 	cpu_affinity_setsize = 0;
5234 
5235 	free(thread_even);
5236 	free(core_even);
5237 	free(package_even);
5238 
5239 	thread_even = NULL;
5240 	core_even = NULL;
5241 	package_even = NULL;
5242 
5243 	free(thread_odd);
5244 	free(core_odd);
5245 	free(package_odd);
5246 
5247 	thread_odd = NULL;
5248 	core_odd = NULL;
5249 	package_odd = NULL;
5250 
5251 	free(output_buffer);
5252 	output_buffer = NULL;
5253 	outp = NULL;
5254 
5255 	free_fd_percpu();
5256 	free_fd_instr_count_percpu();
5257 	free_fd_msr();
5258 	free_fd_rapl_percpu();
5259 	free_fd_cstate();
5260 	free_fd_added_perf_counters();
5261 
5262 	free(irq_column_2_cpu);
5263 	free(irqs_per_cpu);
5264 
5265 	for (i = 0; i <= topo.max_cpu_num; ++i) {
5266 		if (cpus[i].put_ids)
5267 			CPU_FREE(cpus[i].put_ids);
5268 	}
5269 	free(cpus);
5270 }
5271 
5272 /*
5273  * Parse a file containing a single int.
5274  * Return 0 if file can not be opened
5275  * Exit if file can be opened, but can not be parsed
5276  */
5277 int parse_int_file(const char *fmt, ...)
5278 {
5279 	va_list args;
5280 	char path[PATH_MAX];
5281 	FILE *filep;
5282 	int value;
5283 
5284 	va_start(args, fmt);
5285 	vsnprintf(path, sizeof(path), fmt, args);
5286 	va_end(args);
5287 	filep = fopen(path, "r");
5288 	if (!filep)
5289 		return 0;
5290 	if (fscanf(filep, "%d", &value) != 1)
5291 		err(1, "%s: failed to parse number from file", path);
5292 	fclose(filep);
5293 	return value;
5294 }
5295 
5296 /*
5297  * cpu_is_first_core_in_package(cpu)
5298  * return 1 if given CPU is 1st core in package
5299  */
5300 int cpu_is_first_core_in_package(int cpu)
5301 {
5302 	return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
5303 }
5304 
5305 int get_physical_package_id(int cpu)
5306 {
5307 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
5308 }
5309 
5310 int get_die_id(int cpu)
5311 {
5312 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/die_id", cpu);
5313 }
5314 
5315 int get_core_id(int cpu)
5316 {
5317 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
5318 }
5319 
5320 void set_node_data(void)
5321 {
5322 	int pkg, node, lnode, cpu, cpux;
5323 	int cpu_count;
5324 
5325 	/* initialize logical_node_id */
5326 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu)
5327 		cpus[cpu].logical_node_id = -1;
5328 
5329 	cpu_count = 0;
5330 	for (pkg = 0; pkg < topo.num_packages; pkg++) {
5331 		lnode = 0;
5332 		for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
5333 			if (cpus[cpu].physical_package_id != pkg)
5334 				continue;
5335 			/* find a cpu with an unset logical_node_id */
5336 			if (cpus[cpu].logical_node_id != -1)
5337 				continue;
5338 			cpus[cpu].logical_node_id = lnode;
5339 			node = cpus[cpu].physical_node_id;
5340 			cpu_count++;
5341 			/*
5342 			 * find all matching cpus on this pkg and set
5343 			 * the logical_node_id
5344 			 */
5345 			for (cpux = cpu; cpux <= topo.max_cpu_num; cpux++) {
5346 				if ((cpus[cpux].physical_package_id == pkg) && (cpus[cpux].physical_node_id == node)) {
5347 					cpus[cpux].logical_node_id = lnode;
5348 					cpu_count++;
5349 				}
5350 			}
5351 			lnode++;
5352 			if (lnode > topo.nodes_per_pkg)
5353 				topo.nodes_per_pkg = lnode;
5354 		}
5355 		if (cpu_count >= topo.max_cpu_num)
5356 			break;
5357 	}
5358 }
5359 
5360 int get_physical_node_id(struct cpu_topology *thiscpu)
5361 {
5362 	char path[80];
5363 	FILE *filep;
5364 	int i;
5365 	int cpu = thiscpu->logical_cpu_id;
5366 
5367 	for (i = 0; i <= topo.max_cpu_num; i++) {
5368 		sprintf(path, "/sys/devices/system/cpu/cpu%d/node%i/cpulist", cpu, i);
5369 		filep = fopen(path, "r");
5370 		if (!filep)
5371 			continue;
5372 		fclose(filep);
5373 		return i;
5374 	}
5375 	return -1;
5376 }
5377 
5378 static int parse_cpu_str(char *cpu_str, cpu_set_t *cpu_set, int cpu_set_size)
5379 {
5380 	unsigned int start, end;
5381 	char *next = cpu_str;
5382 
5383 	while (next && *next) {
5384 
5385 		if (*next == '-')	/* no negative cpu numbers */
5386 			return 1;
5387 
5388 		start = strtoul(next, &next, 10);
5389 
5390 		if (start >= CPU_SUBSET_MAXCPUS)
5391 			return 1;
5392 		CPU_SET_S(start, cpu_set_size, cpu_set);
5393 
5394 		if (*next == '\0' || *next == '\n')
5395 			break;
5396 
5397 		if (*next == ',') {
5398 			next += 1;
5399 			continue;
5400 		}
5401 
5402 		if (*next == '-') {
5403 			next += 1;	/* start range */
5404 		} else if (*next == '.') {
5405 			next += 1;
5406 			if (*next == '.')
5407 				next += 1;	/* start range */
5408 			else
5409 				return 1;
5410 		}
5411 
5412 		end = strtoul(next, &next, 10);
5413 		if (end <= start)
5414 			return 1;
5415 
5416 		while (++start <= end) {
5417 			if (start >= CPU_SUBSET_MAXCPUS)
5418 				return 1;
5419 			CPU_SET_S(start, cpu_set_size, cpu_set);
5420 		}
5421 
5422 		if (*next == ',')
5423 			next += 1;
5424 		else if (*next != '\0' && *next != '\n')
5425 			return 1;
5426 	}
5427 
5428 	return 0;
5429 }
5430 
5431 int get_thread_siblings(struct cpu_topology *thiscpu)
5432 {
5433 	char path[80], character;
5434 	FILE *filep;
5435 	unsigned long map;
5436 	int so, shift, sib_core;
5437 	int cpu = thiscpu->logical_cpu_id;
5438 	int offset = topo.max_cpu_num + 1;
5439 	size_t size;
5440 	int thread_id = 0;
5441 
5442 	thiscpu->put_ids = CPU_ALLOC((topo.max_cpu_num + 1));
5443 	if (thiscpu->thread_id < 0)
5444 		thiscpu->thread_id = thread_id++;
5445 	if (!thiscpu->put_ids)
5446 		return -1;
5447 
5448 	size = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
5449 	CPU_ZERO_S(size, thiscpu->put_ids);
5450 
5451 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", cpu);
5452 	filep = fopen(path, "r");
5453 
5454 	if (!filep) {
5455 		warnx("%s: open failed", path);
5456 		return -1;
5457 	}
5458 	do {
5459 		offset -= BITMASK_SIZE;
5460 		if (fscanf(filep, "%lx%c", &map, &character) != 2)
5461 			err(1, "%s: failed to parse file", path);
5462 		for (shift = 0; shift < BITMASK_SIZE; shift++) {
5463 			if ((map >> shift) & 0x1) {
5464 				so = shift + offset;
5465 				sib_core = get_core_id(so);
5466 				if (sib_core == thiscpu->physical_core_id) {
5467 					CPU_SET_S(so, size, thiscpu->put_ids);
5468 					if ((so != cpu) && (cpus[so].thread_id < 0))
5469 						cpus[so].thread_id = thread_id++;
5470 				}
5471 			}
5472 		}
5473 	} while (character == ',');
5474 	fclose(filep);
5475 
5476 	return CPU_COUNT_S(size, thiscpu->put_ids);
5477 }
5478 
5479 /*
5480  * run func(thread, core, package) in topology order
5481  * skip non-present cpus
5482  */
5483 
5484 int for_all_cpus_2(int (func) (struct thread_data *, struct core_data *,
5485 			       struct pkg_data *, struct thread_data *, struct core_data *,
5486 			       struct pkg_data *), struct thread_data *thread_base,
5487 		   struct core_data *core_base, struct pkg_data *pkg_base,
5488 		   struct thread_data *thread_base2, struct core_data *core_base2, struct pkg_data *pkg_base2)
5489 {
5490 	int retval, pkg_no, node_no, core_no, thread_no;
5491 
5492 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
5493 		for (node_no = 0; node_no < topo.nodes_per_pkg; ++node_no) {
5494 			for (core_no = 0; core_no < topo.cores_per_node; ++core_no) {
5495 				for (thread_no = 0; thread_no < topo.threads_per_core; ++thread_no) {
5496 					struct thread_data *t, *t2;
5497 					struct core_data *c, *c2;
5498 					struct pkg_data *p, *p2;
5499 
5500 					t = GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no);
5501 
5502 					if (cpu_is_not_allowed(t->cpu_id))
5503 						continue;
5504 
5505 					t2 = GET_THREAD(thread_base2, thread_no, core_no, node_no, pkg_no);
5506 
5507 					c = GET_CORE(core_base, core_no, node_no, pkg_no);
5508 					c2 = GET_CORE(core_base2, core_no, node_no, pkg_no);
5509 
5510 					p = GET_PKG(pkg_base, pkg_no);
5511 					p2 = GET_PKG(pkg_base2, pkg_no);
5512 
5513 					retval = func(t, c, p, t2, c2, p2);
5514 					if (retval)
5515 						return retval;
5516 				}
5517 			}
5518 		}
5519 	}
5520 	return 0;
5521 }
5522 
5523 /*
5524  * run func(cpu) on every cpu in /proc/stat
5525  * return max_cpu number
5526  */
5527 int for_all_proc_cpus(int (func) (int))
5528 {
5529 	FILE *fp;
5530 	int cpu_num;
5531 	int retval;
5532 
5533 	fp = fopen_or_die(proc_stat, "r");
5534 
5535 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
5536 	if (retval != 0)
5537 		err(1, "%s: failed to parse format", proc_stat);
5538 
5539 	while (1) {
5540 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
5541 		if (retval != 1)
5542 			break;
5543 
5544 		retval = func(cpu_num);
5545 		if (retval) {
5546 			fclose(fp);
5547 			return (retval);
5548 		}
5549 	}
5550 	fclose(fp);
5551 	return 0;
5552 }
5553 
5554 #define PATH_EFFECTIVE_CPUS	"/sys/fs/cgroup/cpuset.cpus.effective"
5555 
5556 static char cpu_effective_str[1024];
5557 
5558 static int update_effective_str(bool startup)
5559 {
5560 	FILE *fp;
5561 	char *pos;
5562 	char buf[1024];
5563 	int ret;
5564 
5565 	if (cpu_effective_str[0] == '\0' && !startup)
5566 		return 0;
5567 
5568 	fp = fopen(PATH_EFFECTIVE_CPUS, "r");
5569 	if (!fp)
5570 		return 0;
5571 
5572 	pos = fgets(buf, 1024, fp);
5573 	if (!pos)
5574 		err(1, "%s: file read failed\n", PATH_EFFECTIVE_CPUS);
5575 
5576 	fclose(fp);
5577 
5578 	ret = strncmp(cpu_effective_str, buf, 1024);
5579 	if (!ret)
5580 		return 0;
5581 
5582 	strncpy(cpu_effective_str, buf, 1024);
5583 	return 1;
5584 }
5585 
5586 static void update_effective_set(bool startup)
5587 {
5588 	update_effective_str(startup);
5589 
5590 	if (parse_cpu_str(cpu_effective_str, cpu_effective_set, cpu_effective_setsize))
5591 		err(1, "%s: cpu str malformat %s\n", PATH_EFFECTIVE_CPUS, cpu_effective_str);
5592 }
5593 
5594 void linux_perf_init(void);
5595 void msr_perf_init(void);
5596 void rapl_perf_init(void);
5597 void cstate_perf_init(void);
5598 void added_perf_counters_init(void);
5599 void pmt_init(void);
5600 
5601 void re_initialize(void)
5602 {
5603 	free_all_buffers();
5604 	setup_all_buffers(false);
5605 	linux_perf_init();
5606 	msr_perf_init();
5607 	rapl_perf_init();
5608 	cstate_perf_init();
5609 	added_perf_counters_init();
5610 	pmt_init();
5611 	fprintf(outf, "turbostat: re-initialized with num_cpus %d, allowed_cpus %d\n", topo.num_cpus,
5612 		topo.allowed_cpus);
5613 }
5614 
5615 void set_max_cpu_num(void)
5616 {
5617 	FILE *filep;
5618 	int base_cpu;
5619 	unsigned long dummy;
5620 	char pathname[64];
5621 
5622 	base_cpu = sched_getcpu();
5623 	if (base_cpu < 0)
5624 		err(1, "cannot find calling cpu ID");
5625 	sprintf(pathname, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", base_cpu);
5626 
5627 	filep = fopen_or_die(pathname, "r");
5628 	topo.max_cpu_num = 0;
5629 	while (fscanf(filep, "%lx,", &dummy) == 1)
5630 		topo.max_cpu_num += BITMASK_SIZE;
5631 	fclose(filep);
5632 	topo.max_cpu_num--;	/* 0 based */
5633 }
5634 
5635 /*
5636  * count_cpus()
5637  * remember the last one seen, it will be the max
5638  */
5639 int count_cpus(int cpu)
5640 {
5641 	UNUSED(cpu);
5642 
5643 	topo.num_cpus++;
5644 	return 0;
5645 }
5646 
5647 int mark_cpu_present(int cpu)
5648 {
5649 	CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
5650 	return 0;
5651 }
5652 
5653 int init_thread_id(int cpu)
5654 {
5655 	cpus[cpu].thread_id = -1;
5656 	return 0;
5657 }
5658 
5659 /*
5660  * snapshot_proc_interrupts()
5661  *
5662  * read and record summary of /proc/interrupts
5663  *
5664  * return 1 if config change requires a restart, else return 0
5665  */
5666 int snapshot_proc_interrupts(void)
5667 {
5668 	static FILE *fp;
5669 	int column, retval;
5670 
5671 	if (fp == NULL)
5672 		fp = fopen_or_die("/proc/interrupts", "r");
5673 	else
5674 		rewind(fp);
5675 
5676 	/* read 1st line of /proc/interrupts to get cpu* name for each column */
5677 	for (column = 0; column < topo.num_cpus; ++column) {
5678 		int cpu_number;
5679 
5680 		retval = fscanf(fp, " CPU%d", &cpu_number);
5681 		if (retval != 1)
5682 			break;
5683 
5684 		if (cpu_number > topo.max_cpu_num) {
5685 			warn("/proc/interrupts: cpu%d: > %d", cpu_number, topo.max_cpu_num);
5686 			return 1;
5687 		}
5688 
5689 		irq_column_2_cpu[column] = cpu_number;
5690 		irqs_per_cpu[cpu_number] = 0;
5691 	}
5692 
5693 	/* read /proc/interrupt count lines and sum up irqs per cpu */
5694 	while (1) {
5695 		int column;
5696 		char buf[64];
5697 
5698 		retval = fscanf(fp, " %s:", buf);	/* flush irq# "N:" */
5699 		if (retval != 1)
5700 			break;
5701 
5702 		/* read the count per cpu */
5703 		for (column = 0; column < topo.num_cpus; ++column) {
5704 
5705 			int cpu_number, irq_count;
5706 
5707 			retval = fscanf(fp, " %d", &irq_count);
5708 			if (retval != 1)
5709 				break;
5710 
5711 			cpu_number = irq_column_2_cpu[column];
5712 			irqs_per_cpu[cpu_number] += irq_count;
5713 
5714 		}
5715 
5716 		while (getc(fp) != '\n') ;	/* flush interrupt description */
5717 
5718 	}
5719 	return 0;
5720 }
5721 
5722 /*
5723  * snapshot_graphics()
5724  *
5725  * record snapshot of specified graphics sysfs knob
5726  *
5727  * return 1 if config change requires a restart, else return 0
5728  */
5729 int snapshot_graphics(int idx)
5730 {
5731 	FILE *fp;
5732 	int retval;
5733 
5734 	switch (idx) {
5735 	case GFX_rc6:
5736 	case SAM_mc6:
5737 		fp = fopen_or_die(gfx_info[idx].path, "r");
5738 		retval = fscanf(fp, "%lld", &gfx_info[idx].val_ull);
5739 		if (retval != 1)
5740 			err(1, "rc6");
5741 		fclose(fp);
5742 		return 0;
5743 	case GFX_MHz:
5744 	case GFX_ACTMHz:
5745 	case SAM_MHz:
5746 	case SAM_ACTMHz:
5747 		if (gfx_info[idx].fp == NULL) {
5748 			gfx_info[idx].fp = fopen_or_die(gfx_info[idx].path, "r");
5749 		} else {
5750 			rewind(gfx_info[idx].fp);
5751 			fflush(gfx_info[idx].fp);
5752 		}
5753 		retval = fscanf(gfx_info[idx].fp, "%d", &gfx_info[idx].val);
5754 		if (retval != 1)
5755 			err(1, "MHz");
5756 		return 0;
5757 	default:
5758 		return -EINVAL;
5759 	}
5760 }
5761 
5762 /*
5763  * snapshot_cpu_lpi()
5764  *
5765  * record snapshot of
5766  * /sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us
5767  */
5768 int snapshot_cpu_lpi_us(void)
5769 {
5770 	FILE *fp;
5771 	int retval;
5772 
5773 	fp = fopen_or_die("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", "r");
5774 
5775 	retval = fscanf(fp, "%lld", &cpuidle_cur_cpu_lpi_us);
5776 	if (retval != 1) {
5777 		fprintf(stderr, "Disabling Low Power Idle CPU output\n");
5778 		BIC_NOT_PRESENT(BIC_CPU_LPI);
5779 		fclose(fp);
5780 		return -1;
5781 	}
5782 
5783 	fclose(fp);
5784 
5785 	return 0;
5786 }
5787 
5788 /*
5789  * snapshot_sys_lpi()
5790  *
5791  * record snapshot of sys_lpi_file
5792  */
5793 int snapshot_sys_lpi_us(void)
5794 {
5795 	FILE *fp;
5796 	int retval;
5797 
5798 	fp = fopen_or_die(sys_lpi_file, "r");
5799 
5800 	retval = fscanf(fp, "%lld", &cpuidle_cur_sys_lpi_us);
5801 	if (retval != 1) {
5802 		fprintf(stderr, "Disabling Low Power Idle System output\n");
5803 		BIC_NOT_PRESENT(BIC_SYS_LPI);
5804 		fclose(fp);
5805 		return -1;
5806 	}
5807 	fclose(fp);
5808 
5809 	return 0;
5810 }
5811 
5812 /*
5813  * snapshot /proc and /sys files
5814  *
5815  * return 1 if configuration restart needed, else return 0
5816  */
5817 int snapshot_proc_sysfs_files(void)
5818 {
5819 	if (DO_BIC(BIC_IRQ))
5820 		if (snapshot_proc_interrupts())
5821 			return 1;
5822 
5823 	if (DO_BIC(BIC_GFX_rc6))
5824 		snapshot_graphics(GFX_rc6);
5825 
5826 	if (DO_BIC(BIC_GFXMHz))
5827 		snapshot_graphics(GFX_MHz);
5828 
5829 	if (DO_BIC(BIC_GFXACTMHz))
5830 		snapshot_graphics(GFX_ACTMHz);
5831 
5832 	if (DO_BIC(BIC_SAM_mc6))
5833 		snapshot_graphics(SAM_mc6);
5834 
5835 	if (DO_BIC(BIC_SAMMHz))
5836 		snapshot_graphics(SAM_MHz);
5837 
5838 	if (DO_BIC(BIC_SAMACTMHz))
5839 		snapshot_graphics(SAM_ACTMHz);
5840 
5841 	if (DO_BIC(BIC_CPU_LPI))
5842 		snapshot_cpu_lpi_us();
5843 
5844 	if (DO_BIC(BIC_SYS_LPI))
5845 		snapshot_sys_lpi_us();
5846 
5847 	return 0;
5848 }
5849 
5850 int exit_requested;
5851 
5852 static void signal_handler(int signal)
5853 {
5854 	switch (signal) {
5855 	case SIGINT:
5856 		exit_requested = 1;
5857 		if (debug)
5858 			fprintf(stderr, " SIGINT\n");
5859 		break;
5860 	case SIGUSR1:
5861 		if (debug > 1)
5862 			fprintf(stderr, "SIGUSR1\n");
5863 		break;
5864 	}
5865 }
5866 
5867 void setup_signal_handler(void)
5868 {
5869 	struct sigaction sa;
5870 
5871 	memset(&sa, 0, sizeof(sa));
5872 
5873 	sa.sa_handler = &signal_handler;
5874 
5875 	if (sigaction(SIGINT, &sa, NULL) < 0)
5876 		err(1, "sigaction SIGINT");
5877 	if (sigaction(SIGUSR1, &sa, NULL) < 0)
5878 		err(1, "sigaction SIGUSR1");
5879 }
5880 
5881 void do_sleep(void)
5882 {
5883 	struct timeval tout;
5884 	struct timespec rest;
5885 	fd_set readfds;
5886 	int retval;
5887 
5888 	FD_ZERO(&readfds);
5889 	FD_SET(0, &readfds);
5890 
5891 	if (ignore_stdin) {
5892 		nanosleep(&interval_ts, NULL);
5893 		return;
5894 	}
5895 
5896 	tout = interval_tv;
5897 	retval = select(1, &readfds, NULL, NULL, &tout);
5898 
5899 	if (retval == 1) {
5900 		switch (getc(stdin)) {
5901 		case 'q':
5902 			exit_requested = 1;
5903 			break;
5904 		case EOF:
5905 			/*
5906 			 * 'stdin' is a pipe closed on the other end. There
5907 			 * won't be any further input.
5908 			 */
5909 			ignore_stdin = 1;
5910 			/* Sleep the rest of the time */
5911 			rest.tv_sec = (tout.tv_sec + tout.tv_usec / 1000000);
5912 			rest.tv_nsec = (tout.tv_usec % 1000000) * 1000;
5913 			nanosleep(&rest, NULL);
5914 		}
5915 	}
5916 }
5917 
5918 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr)
5919 {
5920 	int ret, idx;
5921 	unsigned long long msr_cur, msr_last;
5922 
5923 	assert(!no_msr);
5924 
5925 	if (!per_cpu_msr_sum)
5926 		return 1;
5927 
5928 	idx = offset_to_idx(offset);
5929 	if (idx < 0)
5930 		return idx;
5931 	/* get_msr_sum() = sum + (get_msr() - last) */
5932 	ret = get_msr(cpu, offset, &msr_cur);
5933 	if (ret)
5934 		return ret;
5935 	msr_last = per_cpu_msr_sum[cpu].entries[idx].last;
5936 	DELTA_WRAP32(msr_cur, msr_last);
5937 	*msr = msr_last + per_cpu_msr_sum[cpu].entries[idx].sum;
5938 
5939 	return 0;
5940 }
5941 
5942 timer_t timerid;
5943 
5944 /* Timer callback, update the sum of MSRs periodically. */
5945 static int update_msr_sum(struct thread_data *t, struct core_data *c, struct pkg_data *p)
5946 {
5947 	int i, ret;
5948 	int cpu = t->cpu_id;
5949 
5950 	UNUSED(c);
5951 	UNUSED(p);
5952 
5953 	assert(!no_msr);
5954 
5955 	for (i = IDX_PKG_ENERGY; i < IDX_COUNT; i++) {
5956 		unsigned long long msr_cur, msr_last;
5957 		off_t offset;
5958 
5959 		if (!idx_valid(i))
5960 			continue;
5961 		offset = idx_to_offset(i);
5962 		if (offset < 0)
5963 			continue;
5964 		ret = get_msr(cpu, offset, &msr_cur);
5965 		if (ret) {
5966 			fprintf(outf, "Can not update msr(0x%llx)\n", (unsigned long long)offset);
5967 			continue;
5968 		}
5969 
5970 		msr_last = per_cpu_msr_sum[cpu].entries[i].last;
5971 		per_cpu_msr_sum[cpu].entries[i].last = msr_cur & 0xffffffff;
5972 
5973 		DELTA_WRAP32(msr_cur, msr_last);
5974 		per_cpu_msr_sum[cpu].entries[i].sum += msr_last;
5975 	}
5976 	return 0;
5977 }
5978 
5979 static void msr_record_handler(union sigval v)
5980 {
5981 	UNUSED(v);
5982 
5983 	for_all_cpus(update_msr_sum, EVEN_COUNTERS);
5984 }
5985 
5986 void msr_sum_record(void)
5987 {
5988 	struct itimerspec its;
5989 	struct sigevent sev;
5990 
5991 	per_cpu_msr_sum = calloc(topo.max_cpu_num + 1, sizeof(struct msr_sum_array));
5992 	if (!per_cpu_msr_sum) {
5993 		fprintf(outf, "Can not allocate memory for long time MSR.\n");
5994 		return;
5995 	}
5996 	/*
5997 	 * Signal handler might be restricted, so use thread notifier instead.
5998 	 */
5999 	memset(&sev, 0, sizeof(struct sigevent));
6000 	sev.sigev_notify = SIGEV_THREAD;
6001 	sev.sigev_notify_function = msr_record_handler;
6002 
6003 	sev.sigev_value.sival_ptr = &timerid;
6004 	if (timer_create(CLOCK_REALTIME, &sev, &timerid) == -1) {
6005 		fprintf(outf, "Can not create timer.\n");
6006 		goto release_msr;
6007 	}
6008 
6009 	its.it_value.tv_sec = 0;
6010 	its.it_value.tv_nsec = 1;
6011 	/*
6012 	 * A wraparound time has been calculated early.
6013 	 * Some sources state that the peak power for a
6014 	 * microprocessor is usually 1.5 times the TDP rating,
6015 	 * use 2 * TDP for safety.
6016 	 */
6017 	its.it_interval.tv_sec = rapl_joule_counter_range / 2;
6018 	its.it_interval.tv_nsec = 0;
6019 
6020 	if (timer_settime(timerid, 0, &its, NULL) == -1) {
6021 		fprintf(outf, "Can not set timer.\n");
6022 		goto release_timer;
6023 	}
6024 	return;
6025 
6026 release_timer:
6027 	timer_delete(timerid);
6028 release_msr:
6029 	free(per_cpu_msr_sum);
6030 }
6031 
6032 /*
6033  * set_my_sched_priority(pri)
6034  * return previous priority on success
6035  * return value < -20 on failure
6036  */
6037 int set_my_sched_priority(int priority)
6038 {
6039 	int retval;
6040 	int original_priority;
6041 
6042 	errno = 0;
6043 	original_priority = getpriority(PRIO_PROCESS, 0);
6044 	if (errno && (original_priority == -1))
6045 		return -21;
6046 
6047 	retval = setpriority(PRIO_PROCESS, 0, priority);
6048 	if (retval)
6049 		return -21;
6050 
6051 	errno = 0;
6052 	retval = getpriority(PRIO_PROCESS, 0);
6053 	if (retval != priority)
6054 		return -21;
6055 
6056 	return original_priority;
6057 }
6058 
6059 void turbostat_loop()
6060 {
6061 	int retval;
6062 	int restarted = 0;
6063 	unsigned int done_iters = 0;
6064 
6065 	setup_signal_handler();
6066 
6067 	/*
6068 	 * elevate own priority for interval mode
6069 	 *
6070 	 * ignore on error - we probably don't have permission to set it, but
6071 	 * it's not a big deal
6072 	 */
6073 	set_my_sched_priority(-20);
6074 
6075 restart:
6076 	restarted++;
6077 
6078 	snapshot_proc_sysfs_files();
6079 	retval = for_all_cpus(get_counters, EVEN_COUNTERS);
6080 	first_counter_read = 0;
6081 	if (retval < -1) {
6082 		exit(retval);
6083 	} else if (retval == -1) {
6084 		if (restarted > 10) {
6085 			exit(retval);
6086 		}
6087 		re_initialize();
6088 		goto restart;
6089 	}
6090 	restarted = 0;
6091 	done_iters = 0;
6092 	gettimeofday(&tv_even, (struct timezone *)NULL);
6093 
6094 	while (1) {
6095 		if (for_all_proc_cpus(cpu_is_not_present)) {
6096 			re_initialize();
6097 			goto restart;
6098 		}
6099 		if (update_effective_str(false)) {
6100 			re_initialize();
6101 			goto restart;
6102 		}
6103 		do_sleep();
6104 		if (snapshot_proc_sysfs_files())
6105 			goto restart;
6106 		retval = for_all_cpus(get_counters, ODD_COUNTERS);
6107 		if (retval < -1) {
6108 			exit(retval);
6109 		} else if (retval == -1) {
6110 			re_initialize();
6111 			goto restart;
6112 		}
6113 		gettimeofday(&tv_odd, (struct timezone *)NULL);
6114 		timersub(&tv_odd, &tv_even, &tv_delta);
6115 		if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) {
6116 			re_initialize();
6117 			goto restart;
6118 		}
6119 		compute_average(EVEN_COUNTERS);
6120 		format_all_counters(EVEN_COUNTERS);
6121 		flush_output_stdout();
6122 		if (exit_requested)
6123 			break;
6124 		if (num_iterations && ++done_iters >= num_iterations)
6125 			break;
6126 		do_sleep();
6127 		if (snapshot_proc_sysfs_files())
6128 			goto restart;
6129 		retval = for_all_cpus(get_counters, EVEN_COUNTERS);
6130 		if (retval < -1) {
6131 			exit(retval);
6132 		} else if (retval == -1) {
6133 			re_initialize();
6134 			goto restart;
6135 		}
6136 		gettimeofday(&tv_even, (struct timezone *)NULL);
6137 		timersub(&tv_even, &tv_odd, &tv_delta);
6138 		if (for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) {
6139 			re_initialize();
6140 			goto restart;
6141 		}
6142 		compute_average(ODD_COUNTERS);
6143 		format_all_counters(ODD_COUNTERS);
6144 		flush_output_stdout();
6145 		if (exit_requested)
6146 			break;
6147 		if (num_iterations && ++done_iters >= num_iterations)
6148 			break;
6149 	}
6150 }
6151 
6152 void check_dev_msr()
6153 {
6154 	struct stat sb;
6155 	char pathname[32];
6156 
6157 	if (no_msr)
6158 		return;
6159 
6160 	sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
6161 	if (stat(pathname, &sb))
6162 		if (system("/sbin/modprobe msr > /dev/null 2>&1"))
6163 			no_msr = 1;
6164 }
6165 
6166 /*
6167  * check for CAP_SYS_RAWIO
6168  * return 0 on success
6169  * return 1 on fail
6170  */
6171 int check_for_cap_sys_rawio(void)
6172 {
6173 	cap_t caps;
6174 	cap_flag_value_t cap_flag_value;
6175 	int ret = 0;
6176 
6177 	caps = cap_get_proc();
6178 	if (caps == NULL)
6179 		return 1;
6180 
6181 	if (cap_get_flag(caps, CAP_SYS_RAWIO, CAP_EFFECTIVE, &cap_flag_value)) {
6182 		ret = 1;
6183 		goto free_and_exit;
6184 	}
6185 
6186 	if (cap_flag_value != CAP_SET) {
6187 		ret = 1;
6188 		goto free_and_exit;
6189 	}
6190 
6191 free_and_exit:
6192 	if (cap_free(caps) == -1)
6193 		err(-6, "cap_free\n");
6194 
6195 	return ret;
6196 }
6197 
6198 void check_msr_permission(void)
6199 {
6200 	int failed = 0;
6201 	char pathname[32];
6202 
6203 	if (no_msr)
6204 		return;
6205 
6206 	/* check for CAP_SYS_RAWIO */
6207 	failed += check_for_cap_sys_rawio();
6208 
6209 	/* test file permissions */
6210 	sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
6211 	if (euidaccess(pathname, R_OK)) {
6212 		failed++;
6213 	}
6214 
6215 	if (failed) {
6216 		warnx("Failed to access %s. Some of the counters may not be available\n"
6217 		      "\tRun as root to enable them or use %s to disable the access explicitly", pathname, "--no-msr");
6218 		no_msr = 1;
6219 	}
6220 }
6221 
6222 void probe_bclk(void)
6223 {
6224 	unsigned long long msr;
6225 	unsigned int base_ratio;
6226 
6227 	if (!platform->has_nhm_msrs || no_msr)
6228 		return;
6229 
6230 	if (platform->bclk_freq == BCLK_100MHZ)
6231 		bclk = 100.00;
6232 	else if (platform->bclk_freq == BCLK_133MHZ)
6233 		bclk = 133.33;
6234 	else if (platform->bclk_freq == BCLK_SLV)
6235 		bclk = slm_bclk();
6236 	else
6237 		return;
6238 
6239 	get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
6240 	base_ratio = (msr >> 8) & 0xFF;
6241 
6242 	base_hz = base_ratio * bclk * 1000000;
6243 	has_base_hz = 1;
6244 
6245 	if (platform->enable_tsc_tweak)
6246 		tsc_tweak = base_hz / tsc_hz;
6247 }
6248 
6249 static void remove_underbar(char *s)
6250 {
6251 	char *to = s;
6252 
6253 	while (*s) {
6254 		if (*s != '_')
6255 			*to++ = *s;
6256 		s++;
6257 	}
6258 
6259 	*to = 0;
6260 }
6261 
6262 static void dump_turbo_ratio_info(void)
6263 {
6264 	if (!has_turbo)
6265 		return;
6266 
6267 	if (!platform->has_nhm_msrs || no_msr)
6268 		return;
6269 
6270 	if (platform->trl_msrs & TRL_LIMIT2)
6271 		dump_turbo_ratio_limit2();
6272 
6273 	if (platform->trl_msrs & TRL_LIMIT1)
6274 		dump_turbo_ratio_limit1();
6275 
6276 	if (platform->trl_msrs & TRL_BASE) {
6277 		dump_turbo_ratio_limits(MSR_TURBO_RATIO_LIMIT);
6278 
6279 		if (is_hybrid)
6280 			dump_turbo_ratio_limits(MSR_SECONDARY_TURBO_RATIO_LIMIT);
6281 	}
6282 
6283 	if (platform->trl_msrs & TRL_ATOM)
6284 		dump_atom_turbo_ratio_limits();
6285 
6286 	if (platform->trl_msrs & TRL_KNL)
6287 		dump_knl_turbo_ratio_limits();
6288 
6289 	if (platform->has_config_tdp)
6290 		dump_config_tdp();
6291 }
6292 
6293 static int read_sysfs_int(char *path)
6294 {
6295 	FILE *input;
6296 	int retval = -1;
6297 
6298 	input = fopen(path, "r");
6299 	if (input == NULL) {
6300 		if (debug)
6301 			fprintf(outf, "NSFOD %s\n", path);
6302 		return (-1);
6303 	}
6304 	if (fscanf(input, "%d", &retval) != 1)
6305 		err(1, "%s: failed to read int from file", path);
6306 	fclose(input);
6307 
6308 	return (retval);
6309 }
6310 
6311 static void dump_sysfs_file(char *path)
6312 {
6313 	FILE *input;
6314 	char cpuidle_buf[64];
6315 
6316 	input = fopen(path, "r");
6317 	if (input == NULL) {
6318 		if (debug)
6319 			fprintf(outf, "NSFOD %s\n", path);
6320 		return;
6321 	}
6322 	if (!fgets(cpuidle_buf, sizeof(cpuidle_buf), input))
6323 		err(1, "%s: failed to read file", path);
6324 	fclose(input);
6325 
6326 	fprintf(outf, "%s: %s", strrchr(path, '/') + 1, cpuidle_buf);
6327 }
6328 
6329 static void probe_intel_uncore_frequency_legacy(void)
6330 {
6331 	int i, j;
6332 	char path[256];
6333 
6334 	for (i = 0; i < topo.num_packages; ++i) {
6335 		for (j = 0; j <= topo.max_die_id; ++j) {
6336 			int k, l;
6337 			char path_base[128];
6338 
6339 			sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/package_%02d_die_%02d", i,
6340 				j);
6341 
6342 			if (access(path_base, R_OK))
6343 				continue;
6344 
6345 			BIC_PRESENT(BIC_UNCORE_MHZ);
6346 
6347 			if (quiet)
6348 				return;
6349 
6350 			sprintf(path, "%s/min_freq_khz", path_base);
6351 			k = read_sysfs_int(path);
6352 			sprintf(path, "%s/max_freq_khz", path_base);
6353 			l = read_sysfs_int(path);
6354 			fprintf(outf, "Uncore Frequency package%d die%d: %d - %d MHz ", i, j, k / 1000, l / 1000);
6355 
6356 			sprintf(path, "%s/initial_min_freq_khz", path_base);
6357 			k = read_sysfs_int(path);
6358 			sprintf(path, "%s/initial_max_freq_khz", path_base);
6359 			l = read_sysfs_int(path);
6360 			fprintf(outf, "(%d - %d MHz)", k / 1000, l / 1000);
6361 
6362 			sprintf(path, "%s/current_freq_khz", path_base);
6363 			k = read_sysfs_int(path);
6364 			fprintf(outf, " %d MHz\n", k / 1000);
6365 		}
6366 	}
6367 }
6368 
6369 static void probe_intel_uncore_frequency_cluster(void)
6370 {
6371 	int i, uncore_max_id;
6372 	char path[256];
6373 	char path_base[128];
6374 
6375 	if (access("/sys/devices/system/cpu/intel_uncore_frequency/uncore00/current_freq_khz", R_OK))
6376 		return;
6377 
6378 	for (uncore_max_id = 0;; ++uncore_max_id) {
6379 
6380 		sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/uncore%02d", uncore_max_id);
6381 
6382 		/* uncore## start at 00 and skips no numbers, so stop upon first missing */
6383 		if (access(path_base, R_OK)) {
6384 			uncore_max_id -= 1;
6385 			break;
6386 		}
6387 	}
6388 	for (i = uncore_max_id; i >= 0; --i) {
6389 		int k, l;
6390 		int package_id, domain_id, cluster_id;
6391 		char name_buf[16];
6392 
6393 		sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/uncore%02d", i);
6394 
6395 		if (access(path_base, R_OK))
6396 			err(1, "%s: %s\n", __func__, path_base);
6397 
6398 		sprintf(path, "%s/package_id", path_base);
6399 		package_id = read_sysfs_int(path);
6400 
6401 		sprintf(path, "%s/domain_id", path_base);
6402 		domain_id = read_sysfs_int(path);
6403 
6404 		sprintf(path, "%s/fabric_cluster_id", path_base);
6405 		cluster_id = read_sysfs_int(path);
6406 
6407 		sprintf(path, "%s/current_freq_khz", path_base);
6408 		sprintf(name_buf, "UMHz%d.%d", domain_id, cluster_id);
6409 
6410 		add_counter(0, path, name_buf, 0, SCOPE_PACKAGE, COUNTER_K2M, FORMAT_AVERAGE, 0, package_id);
6411 
6412 		if (quiet)
6413 			continue;
6414 
6415 		sprintf(path, "%s/min_freq_khz", path_base);
6416 		k = read_sysfs_int(path);
6417 		sprintf(path, "%s/max_freq_khz", path_base);
6418 		l = read_sysfs_int(path);
6419 		fprintf(outf, "Uncore Frequency package%d domain%d cluster%d: %d - %d MHz ", package_id, domain_id,
6420 			cluster_id, k / 1000, l / 1000);
6421 
6422 		sprintf(path, "%s/initial_min_freq_khz", path_base);
6423 		k = read_sysfs_int(path);
6424 		sprintf(path, "%s/initial_max_freq_khz", path_base);
6425 		l = read_sysfs_int(path);
6426 		fprintf(outf, "(%d - %d MHz)", k / 1000, l / 1000);
6427 
6428 		sprintf(path, "%s/current_freq_khz", path_base);
6429 		k = read_sysfs_int(path);
6430 		fprintf(outf, " %d MHz\n", k / 1000);
6431 	}
6432 }
6433 
6434 static void probe_intel_uncore_frequency(void)
6435 {
6436 	if (!genuine_intel)
6437 		return;
6438 
6439 	if (access("/sys/devices/system/cpu/intel_uncore_frequency/uncore00", R_OK) == 0)
6440 		probe_intel_uncore_frequency_cluster();
6441 	else
6442 		probe_intel_uncore_frequency_legacy();
6443 }
6444 
6445 static void probe_graphics(void)
6446 {
6447 	/* Xe graphics sysfs knobs */
6448 	if (!access("/sys/class/drm/card0/device/tile0/gt0/gtidle/idle_residency_ms", R_OK)) {
6449 		FILE *fp;
6450 		char buf[8];
6451 		bool gt0_is_gt;
6452 		int idx;
6453 
6454 		fp = fopen("/sys/class/drm/card0/device/tile0/gt0/gtidle/name", "r");
6455 		if (!fp)
6456 			goto next;
6457 
6458 		if (!fread(buf, sizeof(char), 7, fp)) {
6459 			fclose(fp);
6460 			goto next;
6461 		}
6462 		fclose(fp);
6463 
6464 		if (!strncmp(buf, "gt0-rc", strlen("gt0-rc")))
6465 			gt0_is_gt = true;
6466 		else if (!strncmp(buf, "gt0-mc", strlen("gt0-mc")))
6467 			gt0_is_gt = false;
6468 		else
6469 			goto next;
6470 
6471 		idx = gt0_is_gt ? GFX_rc6 : SAM_mc6;
6472 		gfx_info[idx].path = "/sys/class/drm/card0/device/tile0/gt0/gtidle/idle_residency_ms";
6473 
6474 		idx = gt0_is_gt ? GFX_MHz : SAM_MHz;
6475 		if (!access("/sys/class/drm/card0/device/tile0/gt0/freq0/cur_freq", R_OK))
6476 			gfx_info[idx].path = "/sys/class/drm/card0/device/tile0/gt0/freq0/cur_freq";
6477 
6478 		idx = gt0_is_gt ? GFX_ACTMHz : SAM_ACTMHz;
6479 		if (!access("/sys/class/drm/card0/device/tile0/gt0/freq0/act_freq", R_OK))
6480 			gfx_info[idx].path = "/sys/class/drm/card0/device/tile0/gt0/freq0/act_freq";
6481 
6482 		idx = gt0_is_gt ? SAM_mc6 : GFX_rc6;
6483 		if (!access("/sys/class/drm/card0/device/tile0/gt1/gtidle/idle_residency_ms", R_OK))
6484 			gfx_info[idx].path = "/sys/class/drm/card0/device/tile0/gt1/gtidle/idle_residency_ms";
6485 
6486 		idx = gt0_is_gt ? SAM_MHz : GFX_MHz;
6487 		if (!access("/sys/class/drm/card0/device/tile0/gt1/freq0/cur_freq", R_OK))
6488 			gfx_info[idx].path = "/sys/class/drm/card0/device/tile0/gt1/freq0/cur_freq";
6489 
6490 		idx = gt0_is_gt ? SAM_ACTMHz : GFX_ACTMHz;
6491 		if (!access("/sys/class/drm/card0/device/tile0/gt1/freq0/act_freq", R_OK))
6492 			gfx_info[idx].path = "/sys/class/drm/card0/device/tile0/gt1/freq0/act_freq";
6493 
6494 		goto end;
6495 	}
6496 
6497 next:
6498 	/* New i915 graphics sysfs knobs */
6499 	if (!access("/sys/class/drm/card0/gt/gt0/rc6_residency_ms", R_OK)) {
6500 		gfx_info[GFX_rc6].path = "/sys/class/drm/card0/gt/gt0/rc6_residency_ms";
6501 
6502 		if (!access("/sys/class/drm/card0/gt/gt0/rps_cur_freq_mhz", R_OK))
6503 			gfx_info[GFX_MHz].path = "/sys/class/drm/card0/gt/gt0/rps_cur_freq_mhz";
6504 
6505 		if (!access("/sys/class/drm/card0/gt/gt0/rps_act_freq_mhz", R_OK))
6506 			gfx_info[GFX_ACTMHz].path = "/sys/class/drm/card0/gt/gt0/rps_act_freq_mhz";
6507 
6508 		if (!access("/sys/class/drm/card0/gt/gt1/rc6_residency_ms", R_OK))
6509 			gfx_info[SAM_mc6].path = "/sys/class/drm/card0/gt/gt1/rc6_residency_ms";
6510 
6511 		if (!access("/sys/class/drm/card0/gt/gt1/rps_cur_freq_mhz", R_OK))
6512 			gfx_info[SAM_MHz].path = "/sys/class/drm/card0/gt/gt1/rps_cur_freq_mhz";
6513 
6514 		if (!access("/sys/class/drm/card0/gt/gt1/rps_act_freq_mhz", R_OK))
6515 			gfx_info[SAM_ACTMHz].path = "/sys/class/drm/card0/gt/gt1/rps_act_freq_mhz";
6516 
6517 		goto end;
6518 	}
6519 
6520 	/* Fall back to traditional i915 graphics sysfs knobs */
6521 	if (!access("/sys/class/drm/card0/power/rc6_residency_ms", R_OK))
6522 		gfx_info[GFX_rc6].path = "/sys/class/drm/card0/power/rc6_residency_ms";
6523 
6524 	if (!access("/sys/class/drm/card0/gt_cur_freq_mhz", R_OK))
6525 		gfx_info[GFX_MHz].path = "/sys/class/drm/card0/gt_cur_freq_mhz";
6526 	else if (!access("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", R_OK))
6527 		gfx_info[GFX_MHz].path = "/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz";
6528 
6529 	if (!access("/sys/class/drm/card0/gt_act_freq_mhz", R_OK))
6530 		gfx_info[GFX_ACTMHz].path = "/sys/class/drm/card0/gt_act_freq_mhz";
6531 	else if (!access("/sys/class/graphics/fb0/device/drm/card0/gt_act_freq_mhz", R_OK))
6532 		gfx_info[GFX_ACTMHz].path = "/sys/class/graphics/fb0/device/drm/card0/gt_act_freq_mhz";
6533 
6534 end:
6535 	if (gfx_info[GFX_rc6].path)
6536 		BIC_PRESENT(BIC_GFX_rc6);
6537 	if (gfx_info[GFX_MHz].path)
6538 		BIC_PRESENT(BIC_GFXMHz);
6539 	if (gfx_info[GFX_ACTMHz].path)
6540 		BIC_PRESENT(BIC_GFXACTMHz);
6541 	if (gfx_info[SAM_mc6].path)
6542 		BIC_PRESENT(BIC_SAM_mc6);
6543 	if (gfx_info[SAM_MHz].path)
6544 		BIC_PRESENT(BIC_SAMMHz);
6545 	if (gfx_info[SAM_ACTMHz].path)
6546 		BIC_PRESENT(BIC_SAMACTMHz);
6547 }
6548 
6549 static void dump_sysfs_cstate_config(void)
6550 {
6551 	char path[64];
6552 	char name_buf[16];
6553 	char desc[64];
6554 	FILE *input;
6555 	int state;
6556 	char *sp;
6557 
6558 	if (access("/sys/devices/system/cpu/cpuidle", R_OK)) {
6559 		fprintf(outf, "cpuidle not loaded\n");
6560 		return;
6561 	}
6562 
6563 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_driver");
6564 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor");
6565 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor_ro");
6566 
6567 	for (state = 0; state < 10; ++state) {
6568 
6569 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", base_cpu, state);
6570 		input = fopen(path, "r");
6571 		if (input == NULL)
6572 			continue;
6573 		if (!fgets(name_buf, sizeof(name_buf), input))
6574 			err(1, "%s: failed to read file", path);
6575 
6576 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
6577 		sp = strchr(name_buf, '-');
6578 		if (!sp)
6579 			sp = strchrnul(name_buf, '\n');
6580 		*sp = '\0';
6581 		fclose(input);
6582 
6583 		remove_underbar(name_buf);
6584 
6585 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/desc", base_cpu, state);
6586 		input = fopen(path, "r");
6587 		if (input == NULL)
6588 			continue;
6589 		if (!fgets(desc, sizeof(desc), input))
6590 			err(1, "%s: failed to read file", path);
6591 
6592 		fprintf(outf, "cpu%d: %s: %s", base_cpu, name_buf, desc);
6593 		fclose(input);
6594 	}
6595 }
6596 
6597 static void dump_sysfs_pstate_config(void)
6598 {
6599 	char path[64];
6600 	char driver_buf[64];
6601 	char governor_buf[64];
6602 	FILE *input;
6603 	int turbo;
6604 
6605 	sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_driver", base_cpu);
6606 	input = fopen(path, "r");
6607 	if (input == NULL) {
6608 		fprintf(outf, "NSFOD %s\n", path);
6609 		return;
6610 	}
6611 	if (!fgets(driver_buf, sizeof(driver_buf), input))
6612 		err(1, "%s: failed to read file", path);
6613 	fclose(input);
6614 
6615 	sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_governor", base_cpu);
6616 	input = fopen(path, "r");
6617 	if (input == NULL) {
6618 		fprintf(outf, "NSFOD %s\n", path);
6619 		return;
6620 	}
6621 	if (!fgets(governor_buf, sizeof(governor_buf), input))
6622 		err(1, "%s: failed to read file", path);
6623 	fclose(input);
6624 
6625 	fprintf(outf, "cpu%d: cpufreq driver: %s", base_cpu, driver_buf);
6626 	fprintf(outf, "cpu%d: cpufreq governor: %s", base_cpu, governor_buf);
6627 
6628 	sprintf(path, "/sys/devices/system/cpu/cpufreq/boost");
6629 	input = fopen(path, "r");
6630 	if (input != NULL) {
6631 		if (fscanf(input, "%d", &turbo) != 1)
6632 			err(1, "%s: failed to parse number from file", path);
6633 		fprintf(outf, "cpufreq boost: %d\n", turbo);
6634 		fclose(input);
6635 	}
6636 
6637 	sprintf(path, "/sys/devices/system/cpu/intel_pstate/no_turbo");
6638 	input = fopen(path, "r");
6639 	if (input != NULL) {
6640 		if (fscanf(input, "%d", &turbo) != 1)
6641 			err(1, "%s: failed to parse number from file", path);
6642 		fprintf(outf, "cpufreq intel_pstate no_turbo: %d\n", turbo);
6643 		fclose(input);
6644 	}
6645 }
6646 
6647 /*
6648  * print_epb()
6649  * Decode the ENERGY_PERF_BIAS MSR
6650  */
6651 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
6652 {
6653 	char *epb_string;
6654 	int cpu, epb;
6655 
6656 	UNUSED(c);
6657 	UNUSED(p);
6658 
6659 	if (!has_epb)
6660 		return 0;
6661 
6662 	cpu = t->cpu_id;
6663 
6664 	/* EPB is per-package */
6665 	if (!is_cpu_first_thread_in_package(t, c, p))
6666 		return 0;
6667 
6668 	if (cpu_migrate(cpu)) {
6669 		fprintf(outf, "print_epb: Could not migrate to CPU %d\n", cpu);
6670 		return -1;
6671 	}
6672 
6673 	epb = get_epb(cpu);
6674 	if (epb < 0)
6675 		return 0;
6676 
6677 	switch (epb) {
6678 	case ENERGY_PERF_BIAS_PERFORMANCE:
6679 		epb_string = "performance";
6680 		break;
6681 	case ENERGY_PERF_BIAS_NORMAL:
6682 		epb_string = "balanced";
6683 		break;
6684 	case ENERGY_PERF_BIAS_POWERSAVE:
6685 		epb_string = "powersave";
6686 		break;
6687 	default:
6688 		epb_string = "custom";
6689 		break;
6690 	}
6691 	fprintf(outf, "cpu%d: EPB: %d (%s)\n", cpu, epb, epb_string);
6692 
6693 	return 0;
6694 }
6695 
6696 /*
6697  * print_hwp()
6698  * Decode the MSR_HWP_CAPABILITIES
6699  */
6700 int print_hwp(struct thread_data *t, struct core_data *c, struct pkg_data *p)
6701 {
6702 	unsigned long long msr;
6703 	int cpu;
6704 
6705 	UNUSED(c);
6706 	UNUSED(p);
6707 
6708 	if (no_msr)
6709 		return 0;
6710 
6711 	if (!has_hwp)
6712 		return 0;
6713 
6714 	cpu = t->cpu_id;
6715 
6716 	/* MSR_HWP_CAPABILITIES is per-package */
6717 	if (!is_cpu_first_thread_in_package(t, c, p))
6718 		return 0;
6719 
6720 	if (cpu_migrate(cpu)) {
6721 		fprintf(outf, "print_hwp: Could not migrate to CPU %d\n", cpu);
6722 		return -1;
6723 	}
6724 
6725 	if (get_msr(cpu, MSR_PM_ENABLE, &msr))
6726 		return 0;
6727 
6728 	fprintf(outf, "cpu%d: MSR_PM_ENABLE: 0x%08llx (%sHWP)\n", cpu, msr, (msr & (1 << 0)) ? "" : "No-");
6729 
6730 	/* MSR_PM_ENABLE[1] == 1 if HWP is enabled and MSRs visible */
6731 	if ((msr & (1 << 0)) == 0)
6732 		return 0;
6733 
6734 	if (get_msr(cpu, MSR_HWP_CAPABILITIES, &msr))
6735 		return 0;
6736 
6737 	fprintf(outf, "cpu%d: MSR_HWP_CAPABILITIES: 0x%08llx "
6738 		"(high %d guar %d eff %d low %d)\n",
6739 		cpu, msr,
6740 		(unsigned int)HWP_HIGHEST_PERF(msr),
6741 		(unsigned int)HWP_GUARANTEED_PERF(msr),
6742 		(unsigned int)HWP_MOSTEFFICIENT_PERF(msr), (unsigned int)HWP_LOWEST_PERF(msr));
6743 
6744 	if (get_msr(cpu, MSR_HWP_REQUEST, &msr))
6745 		return 0;
6746 
6747 	fprintf(outf, "cpu%d: MSR_HWP_REQUEST: 0x%08llx "
6748 		"(min %d max %d des %d epp 0x%x window 0x%x pkg 0x%x)\n",
6749 		cpu, msr,
6750 		(unsigned int)(((msr) >> 0) & 0xff),
6751 		(unsigned int)(((msr) >> 8) & 0xff),
6752 		(unsigned int)(((msr) >> 16) & 0xff),
6753 		(unsigned int)(((msr) >> 24) & 0xff),
6754 		(unsigned int)(((msr) >> 32) & 0xff3), (unsigned int)(((msr) >> 42) & 0x1));
6755 
6756 	if (has_hwp_pkg) {
6757 		if (get_msr(cpu, MSR_HWP_REQUEST_PKG, &msr))
6758 			return 0;
6759 
6760 		fprintf(outf, "cpu%d: MSR_HWP_REQUEST_PKG: 0x%08llx "
6761 			"(min %d max %d des %d epp 0x%x window 0x%x)\n",
6762 			cpu, msr,
6763 			(unsigned int)(((msr) >> 0) & 0xff),
6764 			(unsigned int)(((msr) >> 8) & 0xff),
6765 			(unsigned int)(((msr) >> 16) & 0xff),
6766 			(unsigned int)(((msr) >> 24) & 0xff), (unsigned int)(((msr) >> 32) & 0xff3));
6767 	}
6768 	if (has_hwp_notify) {
6769 		if (get_msr(cpu, MSR_HWP_INTERRUPT, &msr))
6770 			return 0;
6771 
6772 		fprintf(outf, "cpu%d: MSR_HWP_INTERRUPT: 0x%08llx "
6773 			"(%s_Guaranteed_Perf_Change, %s_Excursion_Min)\n",
6774 			cpu, msr, ((msr) & 0x1) ? "EN" : "Dis", ((msr) & 0x2) ? "EN" : "Dis");
6775 	}
6776 	if (get_msr(cpu, MSR_HWP_STATUS, &msr))
6777 		return 0;
6778 
6779 	fprintf(outf, "cpu%d: MSR_HWP_STATUS: 0x%08llx "
6780 		"(%sGuaranteed_Perf_Change, %sExcursion_Min)\n",
6781 		cpu, msr, ((msr) & 0x1) ? "" : "No-", ((msr) & 0x4) ? "" : "No-");
6782 
6783 	return 0;
6784 }
6785 
6786 /*
6787  * print_perf_limit()
6788  */
6789 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
6790 {
6791 	unsigned long long msr;
6792 	int cpu;
6793 
6794 	UNUSED(c);
6795 	UNUSED(p);
6796 
6797 	if (no_msr)
6798 		return 0;
6799 
6800 	cpu = t->cpu_id;
6801 
6802 	/* per-package */
6803 	if (!is_cpu_first_thread_in_package(t, c, p))
6804 		return 0;
6805 
6806 	if (cpu_migrate(cpu)) {
6807 		fprintf(outf, "print_perf_limit: Could not migrate to CPU %d\n", cpu);
6808 		return -1;
6809 	}
6810 
6811 	if (platform->plr_msrs & PLR_CORE) {
6812 		get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
6813 		fprintf(outf, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
6814 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
6815 			(msr & 1 << 15) ? "bit15, " : "",
6816 			(msr & 1 << 14) ? "bit14, " : "",
6817 			(msr & 1 << 13) ? "Transitions, " : "",
6818 			(msr & 1 << 12) ? "MultiCoreTurbo, " : "",
6819 			(msr & 1 << 11) ? "PkgPwrL2, " : "",
6820 			(msr & 1 << 10) ? "PkgPwrL1, " : "",
6821 			(msr & 1 << 9) ? "CorePwr, " : "",
6822 			(msr & 1 << 8) ? "Amps, " : "",
6823 			(msr & 1 << 6) ? "VR-Therm, " : "",
6824 			(msr & 1 << 5) ? "Auto-HWP, " : "",
6825 			(msr & 1 << 4) ? "Graphics, " : "",
6826 			(msr & 1 << 2) ? "bit2, " : "",
6827 			(msr & 1 << 1) ? "ThermStatus, " : "", (msr & 1 << 0) ? "PROCHOT, " : "");
6828 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
6829 			(msr & 1 << 31) ? "bit31, " : "",
6830 			(msr & 1 << 30) ? "bit30, " : "",
6831 			(msr & 1 << 29) ? "Transitions, " : "",
6832 			(msr & 1 << 28) ? "MultiCoreTurbo, " : "",
6833 			(msr & 1 << 27) ? "PkgPwrL2, " : "",
6834 			(msr & 1 << 26) ? "PkgPwrL1, " : "",
6835 			(msr & 1 << 25) ? "CorePwr, " : "",
6836 			(msr & 1 << 24) ? "Amps, " : "",
6837 			(msr & 1 << 22) ? "VR-Therm, " : "",
6838 			(msr & 1 << 21) ? "Auto-HWP, " : "",
6839 			(msr & 1 << 20) ? "Graphics, " : "",
6840 			(msr & 1 << 18) ? "bit18, " : "",
6841 			(msr & 1 << 17) ? "ThermStatus, " : "", (msr & 1 << 16) ? "PROCHOT, " : "");
6842 
6843 	}
6844 	if (platform->plr_msrs & PLR_GFX) {
6845 		get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
6846 		fprintf(outf, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
6847 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s)",
6848 			(msr & 1 << 0) ? "PROCHOT, " : "",
6849 			(msr & 1 << 1) ? "ThermStatus, " : "",
6850 			(msr & 1 << 4) ? "Graphics, " : "",
6851 			(msr & 1 << 6) ? "VR-Therm, " : "",
6852 			(msr & 1 << 8) ? "Amps, " : "",
6853 			(msr & 1 << 9) ? "GFXPwr, " : "",
6854 			(msr & 1 << 10) ? "PkgPwrL1, " : "", (msr & 1 << 11) ? "PkgPwrL2, " : "");
6855 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s)\n",
6856 			(msr & 1 << 16) ? "PROCHOT, " : "",
6857 			(msr & 1 << 17) ? "ThermStatus, " : "",
6858 			(msr & 1 << 20) ? "Graphics, " : "",
6859 			(msr & 1 << 22) ? "VR-Therm, " : "",
6860 			(msr & 1 << 24) ? "Amps, " : "",
6861 			(msr & 1 << 25) ? "GFXPwr, " : "",
6862 			(msr & 1 << 26) ? "PkgPwrL1, " : "", (msr & 1 << 27) ? "PkgPwrL2, " : "");
6863 	}
6864 	if (platform->plr_msrs & PLR_RING) {
6865 		get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
6866 		fprintf(outf, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
6867 		fprintf(outf, " (Active: %s%s%s%s%s%s)",
6868 			(msr & 1 << 0) ? "PROCHOT, " : "",
6869 			(msr & 1 << 1) ? "ThermStatus, " : "",
6870 			(msr & 1 << 6) ? "VR-Therm, " : "",
6871 			(msr & 1 << 8) ? "Amps, " : "",
6872 			(msr & 1 << 10) ? "PkgPwrL1, " : "", (msr & 1 << 11) ? "PkgPwrL2, " : "");
6873 		fprintf(outf, " (Logged: %s%s%s%s%s%s)\n",
6874 			(msr & 1 << 16) ? "PROCHOT, " : "",
6875 			(msr & 1 << 17) ? "ThermStatus, " : "",
6876 			(msr & 1 << 22) ? "VR-Therm, " : "",
6877 			(msr & 1 << 24) ? "Amps, " : "",
6878 			(msr & 1 << 26) ? "PkgPwrL1, " : "", (msr & 1 << 27) ? "PkgPwrL2, " : "");
6879 	}
6880 	return 0;
6881 }
6882 
6883 #define	RAPL_POWER_GRANULARITY	0x7FFF	/* 15 bit power granularity */
6884 #define	RAPL_TIME_GRANULARITY	0x3F	/* 6 bit time granularity */
6885 
6886 double get_quirk_tdp(void)
6887 {
6888 	if (platform->rapl_quirk_tdp)
6889 		return platform->rapl_quirk_tdp;
6890 
6891 	return 135.0;
6892 }
6893 
6894 double get_tdp_intel(void)
6895 {
6896 	unsigned long long msr;
6897 
6898 	if (platform->rapl_msrs & RAPL_PKG_POWER_INFO)
6899 		if (!get_msr(base_cpu, MSR_PKG_POWER_INFO, &msr))
6900 			return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
6901 	return get_quirk_tdp();
6902 }
6903 
6904 double get_tdp_amd(void)
6905 {
6906 	return get_quirk_tdp();
6907 }
6908 
6909 void rapl_probe_intel(void)
6910 {
6911 	unsigned long long msr;
6912 	unsigned int time_unit;
6913 	double tdp;
6914 	const unsigned long long bic_watt_bits = BIC_PkgWatt | BIC_CorWatt | BIC_RAMWatt | BIC_GFXWatt;
6915 	const unsigned long long bic_joules_bits = BIC_Pkg_J | BIC_Cor_J | BIC_RAM_J | BIC_GFX_J;
6916 
6917 	if (rapl_joules)
6918 		bic_enabled &= ~bic_watt_bits;
6919 	else
6920 		bic_enabled &= ~bic_joules_bits;
6921 
6922 	if (!(platform->rapl_msrs & RAPL_PKG_PERF_STATUS))
6923 		bic_enabled &= ~BIC_PKG__;
6924 	if (!(platform->rapl_msrs & RAPL_DRAM_PERF_STATUS))
6925 		bic_enabled &= ~BIC_RAM__;
6926 
6927 	/* units on package 0, verify later other packages match */
6928 	if (get_msr(base_cpu, MSR_RAPL_POWER_UNIT, &msr))
6929 		return;
6930 
6931 	rapl_power_units = 1.0 / (1 << (msr & 0xF));
6932 	if (platform->has_rapl_divisor)
6933 		rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
6934 	else
6935 		rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
6936 
6937 	if (platform->has_fixed_rapl_unit)
6938 		rapl_dram_energy_units = (15.3 / 1000000);
6939 	else
6940 		rapl_dram_energy_units = rapl_energy_units;
6941 
6942 	time_unit = msr >> 16 & 0xF;
6943 	if (time_unit == 0)
6944 		time_unit = 0xA;
6945 
6946 	rapl_time_units = 1.0 / (1 << (time_unit));
6947 
6948 	tdp = get_tdp_intel();
6949 
6950 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
6951 	if (!quiet)
6952 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
6953 }
6954 
6955 void rapl_probe_amd(void)
6956 {
6957 	unsigned long long msr;
6958 	double tdp;
6959 	const unsigned long long bic_watt_bits = BIC_PkgWatt | BIC_CorWatt;
6960 	const unsigned long long bic_joules_bits = BIC_Pkg_J | BIC_Cor_J;
6961 
6962 	if (rapl_joules)
6963 		bic_enabled &= ~bic_watt_bits;
6964 	else
6965 		bic_enabled &= ~bic_joules_bits;
6966 
6967 	if (get_msr(base_cpu, MSR_RAPL_PWR_UNIT, &msr))
6968 		return;
6969 
6970 	rapl_time_units = ldexp(1.0, -(msr >> 16 & 0xf));
6971 	rapl_energy_units = ldexp(1.0, -(msr >> 8 & 0x1f));
6972 	rapl_power_units = ldexp(1.0, -(msr & 0xf));
6973 
6974 	tdp = get_tdp_amd();
6975 
6976 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
6977 	if (!quiet)
6978 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
6979 }
6980 
6981 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
6982 {
6983 	fprintf(outf, "cpu%d: %s: %sabled (%0.3f Watts, %f sec, clamp %sabled)\n",
6984 		cpu, label,
6985 		((msr >> 15) & 1) ? "EN" : "DIS",
6986 		((msr >> 0) & 0x7FFF) * rapl_power_units,
6987 		(1.0 + (((msr >> 22) & 0x3) / 4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
6988 		(((msr >> 16) & 1) ? "EN" : "DIS"));
6989 
6990 	return;
6991 }
6992 
6993 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
6994 {
6995 	unsigned long long msr;
6996 	const char *msr_name;
6997 	int cpu;
6998 
6999 	UNUSED(c);
7000 	UNUSED(p);
7001 
7002 	if (!platform->rapl_msrs)
7003 		return 0;
7004 
7005 	/* RAPL counters are per package, so print only for 1st thread/package */
7006 	if (!is_cpu_first_thread_in_package(t, c, p))
7007 		return 0;
7008 
7009 	cpu = t->cpu_id;
7010 	if (cpu_migrate(cpu)) {
7011 		fprintf(outf, "print_rapl: Could not migrate to CPU %d\n", cpu);
7012 		return -1;
7013 	}
7014 
7015 	if (platform->rapl_msrs & RAPL_AMD_F17H) {
7016 		msr_name = "MSR_RAPL_PWR_UNIT";
7017 		if (get_msr(cpu, MSR_RAPL_PWR_UNIT, &msr))
7018 			return -1;
7019 	} else {
7020 		msr_name = "MSR_RAPL_POWER_UNIT";
7021 		if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
7022 			return -1;
7023 	}
7024 
7025 	fprintf(outf, "cpu%d: %s: 0x%08llx (%f Watts, %f Joules, %f sec.)\n", cpu, msr_name, msr,
7026 		rapl_power_units, rapl_energy_units, rapl_time_units);
7027 
7028 	if (platform->rapl_msrs & RAPL_PKG_POWER_INFO) {
7029 
7030 		if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
7031 			return -5;
7032 
7033 		fprintf(outf, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
7034 			cpu, msr,
7035 			((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7036 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7037 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7038 			((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
7039 
7040 	}
7041 	if (platform->rapl_msrs & RAPL_PKG) {
7042 
7043 		if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
7044 			return -9;
7045 
7046 		fprintf(outf, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
7047 			cpu, msr, (msr >> 63) & 1 ? "" : "UN");
7048 
7049 		print_power_limit_msr(cpu, msr, "PKG Limit #1");
7050 		fprintf(outf, "cpu%d: PKG Limit #2: %sabled (%0.3f Watts, %f* sec, clamp %sabled)\n",
7051 			cpu,
7052 			((msr >> 47) & 1) ? "EN" : "DIS",
7053 			((msr >> 32) & 0x7FFF) * rapl_power_units,
7054 			(1.0 + (((msr >> 54) & 0x3) / 4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
7055 			((msr >> 48) & 1) ? "EN" : "DIS");
7056 
7057 		if (get_msr(cpu, MSR_VR_CURRENT_CONFIG, &msr))
7058 			return -9;
7059 
7060 		fprintf(outf, "cpu%d: MSR_VR_CURRENT_CONFIG: 0x%08llx\n", cpu, msr);
7061 		fprintf(outf, "cpu%d: PKG Limit #4: %f Watts (%slocked)\n",
7062 			cpu, ((msr >> 0) & 0x1FFF) * rapl_power_units, (msr >> 31) & 1 ? "" : "UN");
7063 	}
7064 
7065 	if (platform->rapl_msrs & RAPL_DRAM_POWER_INFO) {
7066 		if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
7067 			return -6;
7068 
7069 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
7070 			cpu, msr,
7071 			((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7072 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7073 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7074 			((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
7075 	}
7076 	if (platform->rapl_msrs & RAPL_DRAM) {
7077 		if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
7078 			return -9;
7079 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
7080 			cpu, msr, (msr >> 31) & 1 ? "" : "UN");
7081 
7082 		print_power_limit_msr(cpu, msr, "DRAM Limit");
7083 	}
7084 	if (platform->rapl_msrs & RAPL_CORE_POLICY) {
7085 		if (get_msr(cpu, MSR_PP0_POLICY, &msr))
7086 			return -7;
7087 
7088 		fprintf(outf, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
7089 	}
7090 	if (platform->rapl_msrs & RAPL_CORE_POWER_LIMIT) {
7091 		if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
7092 			return -9;
7093 		fprintf(outf, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
7094 			cpu, msr, (msr >> 31) & 1 ? "" : "UN");
7095 		print_power_limit_msr(cpu, msr, "Cores Limit");
7096 	}
7097 	if (platform->rapl_msrs & RAPL_GFX) {
7098 		if (get_msr(cpu, MSR_PP1_POLICY, &msr))
7099 			return -8;
7100 
7101 		fprintf(outf, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
7102 
7103 		if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
7104 			return -9;
7105 		fprintf(outf, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
7106 			cpu, msr, (msr >> 31) & 1 ? "" : "UN");
7107 		print_power_limit_msr(cpu, msr, "GFX Limit");
7108 	}
7109 	return 0;
7110 }
7111 
7112 /*
7113  * probe_rapl()
7114  *
7115  * sets rapl_power_units, rapl_energy_units, rapl_time_units
7116  */
7117 void probe_rapl(void)
7118 {
7119 	if (!platform->rapl_msrs || no_msr)
7120 		return;
7121 
7122 	if (genuine_intel)
7123 		rapl_probe_intel();
7124 	if (authentic_amd || hygon_genuine)
7125 		rapl_probe_amd();
7126 
7127 	if (quiet)
7128 		return;
7129 
7130 	for_all_cpus(print_rapl, ODD_COUNTERS);
7131 }
7132 
7133 /*
7134  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
7135  * the Thermal Control Circuit (TCC) activates.
7136  * This is usually equal to tjMax.
7137  *
7138  * Older processors do not have this MSR, so there we guess,
7139  * but also allow cmdline over-ride with -T.
7140  *
7141  * Several MSR temperature values are in units of degrees-C
7142  * below this value, including the Digital Thermal Sensor (DTS),
7143  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
7144  */
7145 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7146 {
7147 	unsigned long long msr;
7148 	unsigned int tcc_default, tcc_offset;
7149 	int cpu;
7150 
7151 	UNUSED(c);
7152 	UNUSED(p);
7153 
7154 	/* tj_max is used only for dts or ptm */
7155 	if (!(do_dts || do_ptm))
7156 		return 0;
7157 
7158 	/* this is a per-package concept */
7159 	if (!is_cpu_first_thread_in_package(t, c, p))
7160 		return 0;
7161 
7162 	cpu = t->cpu_id;
7163 	if (cpu_migrate(cpu)) {
7164 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
7165 		return -1;
7166 	}
7167 
7168 	if (tj_max_override != 0) {
7169 		tj_max = tj_max_override;
7170 		fprintf(outf, "cpu%d: Using cmdline TCC Target (%d C)\n", cpu, tj_max);
7171 		return 0;
7172 	}
7173 
7174 	/* Temperature Target MSR is Nehalem and newer only */
7175 	if (!platform->has_nhm_msrs || no_msr)
7176 		goto guess;
7177 
7178 	if (get_msr(base_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
7179 		goto guess;
7180 
7181 	tcc_default = (msr >> 16) & 0xFF;
7182 
7183 	if (!quiet) {
7184 		int bits = platform->tcc_offset_bits;
7185 		unsigned long long enabled = 0;
7186 
7187 		if (bits && !get_msr(base_cpu, MSR_PLATFORM_INFO, &enabled))
7188 			enabled = (enabled >> 30) & 1;
7189 
7190 		if (bits && enabled) {
7191 			tcc_offset = (msr >> 24) & GENMASK(bits - 1, 0);
7192 			fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C) (%d default - %d offset)\n",
7193 				cpu, msr, tcc_default - tcc_offset, tcc_default, tcc_offset);
7194 		} else {
7195 			fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n", cpu, msr, tcc_default);
7196 		}
7197 	}
7198 
7199 	if (!tcc_default)
7200 		goto guess;
7201 
7202 	tj_max = tcc_default;
7203 
7204 	return 0;
7205 
7206 guess:
7207 	tj_max = TJMAX_DEFAULT;
7208 	fprintf(outf, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n", cpu, tj_max);
7209 
7210 	return 0;
7211 }
7212 
7213 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7214 {
7215 	unsigned long long msr;
7216 	unsigned int dts, dts2;
7217 	int cpu;
7218 
7219 	UNUSED(c);
7220 	UNUSED(p);
7221 
7222 	if (no_msr)
7223 		return 0;
7224 
7225 	if (!(do_dts || do_ptm))
7226 		return 0;
7227 
7228 	cpu = t->cpu_id;
7229 
7230 	/* DTS is per-core, no need to print for each thread */
7231 	if (!is_cpu_first_thread_in_core(t, c, p))
7232 		return 0;
7233 
7234 	if (cpu_migrate(cpu)) {
7235 		fprintf(outf, "print_thermal: Could not migrate to CPU %d\n", cpu);
7236 		return -1;
7237 	}
7238 
7239 	if (do_ptm && is_cpu_first_core_in_package(t, c, p)) {
7240 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
7241 			return 0;
7242 
7243 		dts = (msr >> 16) & 0x7F;
7244 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n", cpu, msr, tj_max - dts);
7245 
7246 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
7247 			return 0;
7248 
7249 		dts = (msr >> 16) & 0x7F;
7250 		dts2 = (msr >> 8) & 0x7F;
7251 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
7252 			cpu, msr, tj_max - dts, tj_max - dts2);
7253 	}
7254 
7255 	if (do_dts && debug) {
7256 		unsigned int resolution;
7257 
7258 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
7259 			return 0;
7260 
7261 		dts = (msr >> 16) & 0x7F;
7262 		resolution = (msr >> 27) & 0xF;
7263 		fprintf(outf, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
7264 			cpu, msr, tj_max - dts, resolution);
7265 
7266 		if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
7267 			return 0;
7268 
7269 		dts = (msr >> 16) & 0x7F;
7270 		dts2 = (msr >> 8) & 0x7F;
7271 		fprintf(outf, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
7272 			cpu, msr, tj_max - dts, tj_max - dts2);
7273 	}
7274 
7275 	return 0;
7276 }
7277 
7278 void probe_thermal(void)
7279 {
7280 	if (!access("/sys/devices/system/cpu/cpu0/thermal_throttle/core_throttle_count", R_OK))
7281 		BIC_PRESENT(BIC_CORE_THROT_CNT);
7282 	else
7283 		BIC_NOT_PRESENT(BIC_CORE_THROT_CNT);
7284 
7285 	for_all_cpus(set_temperature_target, ODD_COUNTERS);
7286 
7287 	if (quiet)
7288 		return;
7289 
7290 	for_all_cpus(print_thermal, ODD_COUNTERS);
7291 }
7292 
7293 int get_cpu_type(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7294 {
7295 	unsigned int eax, ebx, ecx, edx;
7296 
7297 	UNUSED(c);
7298 	UNUSED(p);
7299 
7300 	if (!genuine_intel)
7301 		return 0;
7302 
7303 	if (cpu_migrate(t->cpu_id)) {
7304 		fprintf(outf, "Could not migrate to CPU %d\n", t->cpu_id);
7305 		return -1;
7306 	}
7307 
7308 	if (max_level < 0x1a)
7309 		return 0;
7310 
7311 	__cpuid(0x1a, eax, ebx, ecx, edx);
7312 	eax = (eax >> 24) & 0xFF;
7313 	if (eax == 0x20)
7314 		t->is_atom = true;
7315 	return 0;
7316 }
7317 
7318 void decode_feature_control_msr(void)
7319 {
7320 	unsigned long long msr;
7321 
7322 	if (no_msr)
7323 		return;
7324 
7325 	if (!get_msr(base_cpu, MSR_IA32_FEAT_CTL, &msr))
7326 		fprintf(outf, "cpu%d: MSR_IA32_FEATURE_CONTROL: 0x%08llx (%sLocked %s)\n",
7327 			base_cpu, msr, msr & FEAT_CTL_LOCKED ? "" : "UN-", msr & (1 << 18) ? "SGX" : "");
7328 }
7329 
7330 void decode_misc_enable_msr(void)
7331 {
7332 	unsigned long long msr;
7333 
7334 	if (no_msr)
7335 		return;
7336 
7337 	if (!genuine_intel)
7338 		return;
7339 
7340 	if (!get_msr(base_cpu, MSR_IA32_MISC_ENABLE, &msr))
7341 		fprintf(outf, "cpu%d: MSR_IA32_MISC_ENABLE: 0x%08llx (%sTCC %sEIST %sMWAIT %sPREFETCH %sTURBO)\n",
7342 			base_cpu, msr,
7343 			msr & MSR_IA32_MISC_ENABLE_TM1 ? "" : "No-",
7344 			msr & MSR_IA32_MISC_ENABLE_ENHANCED_SPEEDSTEP ? "" : "No-",
7345 			msr & MSR_IA32_MISC_ENABLE_MWAIT ? "" : "No-",
7346 			msr & MSR_IA32_MISC_ENABLE_PREFETCH_DISABLE ? "No-" : "",
7347 			msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ? "No-" : "");
7348 }
7349 
7350 void decode_misc_feature_control(void)
7351 {
7352 	unsigned long long msr;
7353 
7354 	if (no_msr)
7355 		return;
7356 
7357 	if (!platform->has_msr_misc_feature_control)
7358 		return;
7359 
7360 	if (!get_msr(base_cpu, MSR_MISC_FEATURE_CONTROL, &msr))
7361 		fprintf(outf,
7362 			"cpu%d: MSR_MISC_FEATURE_CONTROL: 0x%08llx (%sL2-Prefetch %sL2-Prefetch-pair %sL1-Prefetch %sL1-IP-Prefetch)\n",
7363 			base_cpu, msr, msr & (0 << 0) ? "No-" : "", msr & (1 << 0) ? "No-" : "",
7364 			msr & (2 << 0) ? "No-" : "", msr & (3 << 0) ? "No-" : "");
7365 }
7366 
7367 /*
7368  * Decode MSR_MISC_PWR_MGMT
7369  *
7370  * Decode the bits according to the Nehalem documentation
7371  * bit[0] seems to continue to have same meaning going forward
7372  * bit[1] less so...
7373  */
7374 void decode_misc_pwr_mgmt_msr(void)
7375 {
7376 	unsigned long long msr;
7377 
7378 	if (no_msr)
7379 		return;
7380 
7381 	if (!platform->has_msr_misc_pwr_mgmt)
7382 		return;
7383 
7384 	if (!get_msr(base_cpu, MSR_MISC_PWR_MGMT, &msr))
7385 		fprintf(outf, "cpu%d: MSR_MISC_PWR_MGMT: 0x%08llx (%sable-EIST_Coordination %sable-EPB %sable-OOB)\n",
7386 			base_cpu, msr,
7387 			msr & (1 << 0) ? "DIS" : "EN", msr & (1 << 1) ? "EN" : "DIS", msr & (1 << 8) ? "EN" : "DIS");
7388 }
7389 
7390 /*
7391  * Decode MSR_CC6_DEMOTION_POLICY_CONFIG, MSR_MC6_DEMOTION_POLICY_CONFIG
7392  *
7393  * This MSRs are present on Silvermont processors,
7394  * Intel Atom processor E3000 series (Baytrail), and friends.
7395  */
7396 void decode_c6_demotion_policy_msr(void)
7397 {
7398 	unsigned long long msr;
7399 
7400 	if (no_msr)
7401 		return;
7402 
7403 	if (!platform->has_msr_c6_demotion_policy_config)
7404 		return;
7405 
7406 	if (!get_msr(base_cpu, MSR_CC6_DEMOTION_POLICY_CONFIG, &msr))
7407 		fprintf(outf, "cpu%d: MSR_CC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-CC6-Demotion)\n",
7408 			base_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
7409 
7410 	if (!get_msr(base_cpu, MSR_MC6_DEMOTION_POLICY_CONFIG, &msr))
7411 		fprintf(outf, "cpu%d: MSR_MC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-MC6-Demotion)\n",
7412 			base_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
7413 }
7414 
7415 void print_dev_latency(void)
7416 {
7417 	char *path = "/dev/cpu_dma_latency";
7418 	int fd;
7419 	int value;
7420 	int retval;
7421 
7422 	fd = open(path, O_RDONLY);
7423 	if (fd < 0) {
7424 		if (debug)
7425 			warnx("Read %s failed", path);
7426 		return;
7427 	}
7428 
7429 	retval = read(fd, (void *)&value, sizeof(int));
7430 	if (retval != sizeof(int)) {
7431 		warn("read failed %s", path);
7432 		close(fd);
7433 		return;
7434 	}
7435 	fprintf(outf, "/dev/cpu_dma_latency: %d usec (%s)\n", value, value == 2000000000 ? "default" : "constrained");
7436 
7437 	close(fd);
7438 }
7439 
7440 static int has_instr_count_access(void)
7441 {
7442 	int fd;
7443 	int has_access;
7444 
7445 	if (no_perf)
7446 		return 0;
7447 
7448 	fd = open_perf_counter(base_cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, -1, 0);
7449 	has_access = fd != -1;
7450 
7451 	if (fd != -1)
7452 		close(fd);
7453 
7454 	if (!has_access)
7455 		warnx("Failed to access %s. Some of the counters may not be available\n"
7456 		      "\tRun as root to enable them or use %s to disable the access explicitly",
7457 		      "instructions retired perf counter", "--no-perf");
7458 
7459 	return has_access;
7460 }
7461 
7462 int add_rapl_perf_counter_(int cpu, struct rapl_counter_info_t *rci, const struct rapl_counter_arch_info *cai,
7463 			   double *scale_, enum rapl_unit *unit_)
7464 {
7465 	if (no_perf)
7466 		return -1;
7467 
7468 	const double scale = read_perf_scale(cai->perf_subsys, cai->perf_name);
7469 
7470 	if (scale == 0.0)
7471 		return -1;
7472 
7473 	const enum rapl_unit unit = read_perf_rapl_unit(cai->perf_subsys, cai->perf_name);
7474 
7475 	if (unit == RAPL_UNIT_INVALID)
7476 		return -1;
7477 
7478 	const unsigned int rapl_type = read_perf_type(cai->perf_subsys);
7479 	const unsigned int rapl_energy_pkg_config = read_perf_config(cai->perf_subsys, cai->perf_name);
7480 
7481 	const int fd_counter =
7482 	    open_perf_counter(cpu, rapl_type, rapl_energy_pkg_config, rci->fd_perf, PERF_FORMAT_GROUP);
7483 	if (fd_counter == -1)
7484 		return -1;
7485 
7486 	/* If it's the first counter opened, make it a group descriptor */
7487 	if (rci->fd_perf == -1)
7488 		rci->fd_perf = fd_counter;
7489 
7490 	*scale_ = scale;
7491 	*unit_ = unit;
7492 	return fd_counter;
7493 }
7494 
7495 int add_rapl_perf_counter(int cpu, struct rapl_counter_info_t *rci, const struct rapl_counter_arch_info *cai,
7496 			  double *scale, enum rapl_unit *unit)
7497 {
7498 	int ret = add_rapl_perf_counter_(cpu, rci, cai, scale, unit);
7499 
7500 	if (debug >= 2)
7501 		fprintf(stderr, "%s: %d (cpu: %d)\n", __func__, ret, cpu);
7502 
7503 	return ret;
7504 }
7505 
7506 /*
7507  * Linux-perf manages the HW instructions-retired counter
7508  * by enabling when requested, and hiding rollover
7509  */
7510 void linux_perf_init(void)
7511 {
7512 	if (access("/proc/sys/kernel/perf_event_paranoid", F_OK))
7513 		return;
7514 
7515 	if (BIC_IS_ENABLED(BIC_IPC) && has_aperf) {
7516 		fd_instr_count_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
7517 		if (fd_instr_count_percpu == NULL)
7518 			err(-1, "calloc fd_instr_count_percpu");
7519 	}
7520 }
7521 
7522 void rapl_perf_init(void)
7523 {
7524 	const unsigned int num_domains = (platform->has_per_core_rapl ? topo.max_core_id : topo.max_package_id) + 1;
7525 	bool *domain_visited = calloc(num_domains, sizeof(bool));
7526 
7527 	rapl_counter_info_perdomain = calloc(num_domains, sizeof(*rapl_counter_info_perdomain));
7528 	if (rapl_counter_info_perdomain == NULL)
7529 		err(-1, "calloc rapl_counter_info_percpu");
7530 	rapl_counter_info_perdomain_size = num_domains;
7531 
7532 	/*
7533 	 * Initialize rapl_counter_info_percpu
7534 	 */
7535 	for (unsigned int domain_id = 0; domain_id < num_domains; ++domain_id) {
7536 		struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[domain_id];
7537 
7538 		rci->fd_perf = -1;
7539 		for (size_t i = 0; i < NUM_RAPL_COUNTERS; ++i) {
7540 			rci->data[i] = 0;
7541 			rci->source[i] = COUNTER_SOURCE_NONE;
7542 		}
7543 	}
7544 
7545 	/*
7546 	 * Open/probe the counters
7547 	 * If can't get it via perf, fallback to MSR
7548 	 */
7549 	for (size_t i = 0; i < ARRAY_SIZE(rapl_counter_arch_infos); ++i) {
7550 
7551 		const struct rapl_counter_arch_info *const cai = &rapl_counter_arch_infos[i];
7552 		bool has_counter = 0;
7553 		double scale;
7554 		enum rapl_unit unit;
7555 		unsigned int next_domain;
7556 
7557 		memset(domain_visited, 0, num_domains * sizeof(*domain_visited));
7558 
7559 		for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
7560 
7561 			if (cpu_is_not_allowed(cpu))
7562 				continue;
7563 
7564 			/* Skip already seen and handled RAPL domains */
7565 			next_domain =
7566 			    platform->has_per_core_rapl ? cpus[cpu].physical_core_id : cpus[cpu].physical_package_id;
7567 
7568 			assert(next_domain < num_domains);
7569 
7570 			if (domain_visited[next_domain])
7571 				continue;
7572 
7573 			domain_visited[next_domain] = 1;
7574 
7575 			struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[next_domain];
7576 
7577 			/* Check if the counter is enabled and accessible */
7578 			if (BIC_IS_ENABLED(cai->bic) && (platform->rapl_msrs & cai->feature_mask)) {
7579 
7580 				/* Use perf API for this counter */
7581 				if (!no_perf && cai->perf_name
7582 				    && add_rapl_perf_counter(cpu, rci, cai, &scale, &unit) != -1) {
7583 					rci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
7584 					rci->scale[cai->rci_index] = scale * cai->compat_scale;
7585 					rci->unit[cai->rci_index] = unit;
7586 					rci->flags[cai->rci_index] = cai->flags;
7587 
7588 					/* Use MSR for this counter */
7589 				} else if (!no_msr && cai->msr && probe_msr(cpu, cai->msr) == 0) {
7590 					rci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
7591 					rci->msr[cai->rci_index] = cai->msr;
7592 					rci->msr_mask[cai->rci_index] = cai->msr_mask;
7593 					rci->msr_shift[cai->rci_index] = cai->msr_shift;
7594 					rci->unit[cai->rci_index] = RAPL_UNIT_JOULES;
7595 					rci->scale[cai->rci_index] = *cai->platform_rapl_msr_scale * cai->compat_scale;
7596 					rci->flags[cai->rci_index] = cai->flags;
7597 				}
7598 			}
7599 
7600 			if (rci->source[cai->rci_index] != COUNTER_SOURCE_NONE)
7601 				has_counter = 1;
7602 		}
7603 
7604 		/* If any CPU has access to the counter, make it present */
7605 		if (has_counter)
7606 			BIC_PRESENT(cai->bic);
7607 	}
7608 
7609 	free(domain_visited);
7610 }
7611 
7612 /* Assumes msr_counter_info is populated */
7613 static int has_amperf_access(void)
7614 {
7615 	return msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].present &&
7616 	    msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].present;
7617 }
7618 
7619 int *get_cstate_perf_group_fd(struct cstate_counter_info_t *cci, const char *group_name)
7620 {
7621 	if (strcmp(group_name, "cstate_core") == 0)
7622 		return &cci->fd_perf_core;
7623 
7624 	if (strcmp(group_name, "cstate_pkg") == 0)
7625 		return &cci->fd_perf_pkg;
7626 
7627 	return NULL;
7628 }
7629 
7630 int add_cstate_perf_counter_(int cpu, struct cstate_counter_info_t *cci, const struct cstate_counter_arch_info *cai)
7631 {
7632 	if (no_perf)
7633 		return -1;
7634 
7635 	int *pfd_group = get_cstate_perf_group_fd(cci, cai->perf_subsys);
7636 
7637 	if (pfd_group == NULL)
7638 		return -1;
7639 
7640 	const unsigned int type = read_perf_type(cai->perf_subsys);
7641 	const unsigned int config = read_perf_config(cai->perf_subsys, cai->perf_name);
7642 
7643 	const int fd_counter = open_perf_counter(cpu, type, config, *pfd_group, PERF_FORMAT_GROUP);
7644 
7645 	if (fd_counter == -1)
7646 		return -1;
7647 
7648 	/* If it's the first counter opened, make it a group descriptor */
7649 	if (*pfd_group == -1)
7650 		*pfd_group = fd_counter;
7651 
7652 	return fd_counter;
7653 }
7654 
7655 int add_cstate_perf_counter(int cpu, struct cstate_counter_info_t *cci, const struct cstate_counter_arch_info *cai)
7656 {
7657 	int ret = add_cstate_perf_counter_(cpu, cci, cai);
7658 
7659 	if (debug >= 2)
7660 		fprintf(stderr, "%s: %d (cpu: %d)\n", __func__, ret, cpu);
7661 
7662 	return ret;
7663 }
7664 
7665 int add_msr_perf_counter_(int cpu, struct msr_counter_info_t *cci, const struct msr_counter_arch_info *cai)
7666 {
7667 	if (no_perf)
7668 		return -1;
7669 
7670 	const unsigned int type = read_perf_type(cai->perf_subsys);
7671 	const unsigned int config = read_perf_config(cai->perf_subsys, cai->perf_name);
7672 
7673 	const int fd_counter = open_perf_counter(cpu, type, config, cci->fd_perf, PERF_FORMAT_GROUP);
7674 
7675 	if (fd_counter == -1)
7676 		return -1;
7677 
7678 	/* If it's the first counter opened, make it a group descriptor */
7679 	if (cci->fd_perf == -1)
7680 		cci->fd_perf = fd_counter;
7681 
7682 	return fd_counter;
7683 }
7684 
7685 int add_msr_perf_counter(int cpu, struct msr_counter_info_t *cci, const struct msr_counter_arch_info *cai)
7686 {
7687 	int ret = add_msr_perf_counter_(cpu, cci, cai);
7688 
7689 	if (debug)
7690 		fprintf(stderr, "%s: %s/%s: %d (cpu: %d)\n", __func__, cai->perf_subsys, cai->perf_name, ret, cpu);
7691 
7692 	return ret;
7693 }
7694 
7695 void msr_perf_init_(void)
7696 {
7697 	const int mci_num = topo.max_cpu_num + 1;
7698 
7699 	msr_counter_info = calloc(mci_num, sizeof(*msr_counter_info));
7700 	if (!msr_counter_info)
7701 		err(1, "calloc msr_counter_info");
7702 	msr_counter_info_size = mci_num;
7703 
7704 	for (int cpu = 0; cpu < mci_num; ++cpu)
7705 		msr_counter_info[cpu].fd_perf = -1;
7706 
7707 	for (int cidx = 0; cidx < NUM_MSR_COUNTERS; ++cidx) {
7708 
7709 		struct msr_counter_arch_info *cai = &msr_counter_arch_infos[cidx];
7710 
7711 		cai->present = false;
7712 
7713 		for (int cpu = 0; cpu < mci_num; ++cpu) {
7714 
7715 			struct msr_counter_info_t *const cci = &msr_counter_info[cpu];
7716 
7717 			if (cpu_is_not_allowed(cpu))
7718 				continue;
7719 
7720 			if (cai->needed) {
7721 				/* Use perf API for this counter */
7722 				if (!no_perf && cai->perf_name && add_msr_perf_counter(cpu, cci, cai) != -1) {
7723 					cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
7724 					cai->present = true;
7725 
7726 					/* User MSR for this counter */
7727 				} else if (!no_msr && cai->msr && probe_msr(cpu, cai->msr) == 0) {
7728 					cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
7729 					cci->msr[cai->rci_index] = cai->msr;
7730 					cci->msr_mask[cai->rci_index] = cai->msr_mask;
7731 					cai->present = true;
7732 				}
7733 			}
7734 		}
7735 	}
7736 }
7737 
7738 /* Initialize data for reading perf counters from the MSR group. */
7739 void msr_perf_init(void)
7740 {
7741 	bool need_amperf = false, need_smi = false;
7742 	const bool need_soft_c1 = (!platform->has_msr_core_c1_res) && (platform->supported_cstates & CC1);
7743 
7744 	need_amperf = BIC_IS_ENABLED(BIC_Avg_MHz) || BIC_IS_ENABLED(BIC_Busy) || BIC_IS_ENABLED(BIC_Bzy_MHz)
7745 	    || BIC_IS_ENABLED(BIC_IPC) || need_soft_c1;
7746 
7747 	if (BIC_IS_ENABLED(BIC_SMI))
7748 		need_smi = true;
7749 
7750 	/* Enable needed counters */
7751 	msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].needed = need_amperf;
7752 	msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].needed = need_amperf;
7753 	msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].needed = need_smi;
7754 
7755 	msr_perf_init_();
7756 
7757 	const bool has_amperf = has_amperf_access();
7758 	const bool has_smi = msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].present;
7759 
7760 	has_aperf_access = has_amperf;
7761 
7762 	if (has_amperf) {
7763 		BIC_PRESENT(BIC_Avg_MHz);
7764 		BIC_PRESENT(BIC_Busy);
7765 		BIC_PRESENT(BIC_Bzy_MHz);
7766 		BIC_PRESENT(BIC_SMI);
7767 	}
7768 
7769 	if (has_smi)
7770 		BIC_PRESENT(BIC_SMI);
7771 }
7772 
7773 void cstate_perf_init_(bool soft_c1)
7774 {
7775 	bool has_counter;
7776 	bool *cores_visited = NULL, *pkg_visited = NULL;
7777 	const int cores_visited_elems = topo.max_core_id + 1;
7778 	const int pkg_visited_elems = topo.max_package_id + 1;
7779 	const int cci_num = topo.max_cpu_num + 1;
7780 
7781 	ccstate_counter_info = calloc(cci_num, sizeof(*ccstate_counter_info));
7782 	if (!ccstate_counter_info)
7783 		err(1, "calloc ccstate_counter_arch_info");
7784 	ccstate_counter_info_size = cci_num;
7785 
7786 	cores_visited = calloc(cores_visited_elems, sizeof(*cores_visited));
7787 	if (!cores_visited)
7788 		err(1, "calloc cores_visited");
7789 
7790 	pkg_visited = calloc(pkg_visited_elems, sizeof(*pkg_visited));
7791 	if (!pkg_visited)
7792 		err(1, "calloc pkg_visited");
7793 
7794 	/* Initialize cstate_counter_info_percpu */
7795 	for (int cpu = 0; cpu < cci_num; ++cpu) {
7796 		ccstate_counter_info[cpu].fd_perf_core = -1;
7797 		ccstate_counter_info[cpu].fd_perf_pkg = -1;
7798 	}
7799 
7800 	for (int cidx = 0; cidx < NUM_CSTATE_COUNTERS; ++cidx) {
7801 		has_counter = false;
7802 		memset(cores_visited, 0, cores_visited_elems * sizeof(*cores_visited));
7803 		memset(pkg_visited, 0, pkg_visited_elems * sizeof(*pkg_visited));
7804 
7805 		const struct cstate_counter_arch_info *cai = &ccstate_counter_arch_infos[cidx];
7806 
7807 		for (int cpu = 0; cpu < cci_num; ++cpu) {
7808 
7809 			struct cstate_counter_info_t *const cci = &ccstate_counter_info[cpu];
7810 
7811 			if (cpu_is_not_allowed(cpu))
7812 				continue;
7813 
7814 			const int core_id = cpus[cpu].physical_core_id;
7815 			const int pkg_id = cpus[cpu].physical_package_id;
7816 
7817 			assert(core_id < cores_visited_elems);
7818 			assert(pkg_id < pkg_visited_elems);
7819 
7820 			const bool per_thread = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD;
7821 			const bool per_core = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_CORE;
7822 
7823 			if (!per_thread && cores_visited[core_id])
7824 				continue;
7825 
7826 			if (!per_core && pkg_visited[pkg_id])
7827 				continue;
7828 
7829 			const bool counter_needed = BIC_IS_ENABLED(cai->bic) ||
7830 			    (soft_c1 && (cai->flags & CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY));
7831 			const bool counter_supported = (platform->supported_cstates & cai->feature_mask);
7832 
7833 			if (counter_needed && counter_supported) {
7834 				/* Use perf API for this counter */
7835 				if (!no_perf && cai->perf_name && add_cstate_perf_counter(cpu, cci, cai) != -1) {
7836 
7837 					cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
7838 
7839 					/* User MSR for this counter */
7840 				} else if (!no_msr && cai->msr && pkg_cstate_limit >= cai->pkg_cstate_limit
7841 					   && probe_msr(cpu, cai->msr) == 0) {
7842 					cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
7843 					cci->msr[cai->rci_index] = cai->msr;
7844 				}
7845 			}
7846 
7847 			if (cci->source[cai->rci_index] != COUNTER_SOURCE_NONE) {
7848 				has_counter = true;
7849 				cores_visited[core_id] = true;
7850 				pkg_visited[pkg_id] = true;
7851 			}
7852 		}
7853 
7854 		/* If any CPU has access to the counter, make it present */
7855 		if (has_counter)
7856 			BIC_PRESENT(cai->bic);
7857 	}
7858 
7859 	free(cores_visited);
7860 	free(pkg_visited);
7861 }
7862 
7863 void cstate_perf_init(void)
7864 {
7865 	/*
7866 	 * If we don't have a C1 residency MSR, we calculate it "in software",
7867 	 * but we need APERF, MPERF too.
7868 	 */
7869 	const bool soft_c1 = !platform->has_msr_core_c1_res && has_amperf_access()
7870 	    && platform->supported_cstates & CC1;
7871 
7872 	if (soft_c1)
7873 		BIC_PRESENT(BIC_CPU_c1);
7874 
7875 	cstate_perf_init_(soft_c1);
7876 }
7877 
7878 void probe_cstates(void)
7879 {
7880 	probe_cst_limit();
7881 
7882 	if (platform->has_msr_module_c6_res_ms)
7883 		BIC_PRESENT(BIC_Mod_c6);
7884 
7885 	if (platform->has_ext_cst_msrs && !no_msr) {
7886 		BIC_PRESENT(BIC_Totl_c0);
7887 		BIC_PRESENT(BIC_Any_c0);
7888 		BIC_PRESENT(BIC_GFX_c0);
7889 		BIC_PRESENT(BIC_CPUGFX);
7890 	}
7891 
7892 	if (quiet)
7893 		return;
7894 
7895 	dump_power_ctl();
7896 	dump_cst_cfg();
7897 	decode_c6_demotion_policy_msr();
7898 	print_dev_latency();
7899 	dump_sysfs_cstate_config();
7900 	print_irtl();
7901 }
7902 
7903 void probe_lpi(void)
7904 {
7905 	if (!access("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", R_OK))
7906 		BIC_PRESENT(BIC_CPU_LPI);
7907 	else
7908 		BIC_NOT_PRESENT(BIC_CPU_LPI);
7909 
7910 	if (!access(sys_lpi_file_sysfs, R_OK)) {
7911 		sys_lpi_file = sys_lpi_file_sysfs;
7912 		BIC_PRESENT(BIC_SYS_LPI);
7913 	} else if (!access(sys_lpi_file_debugfs, R_OK)) {
7914 		sys_lpi_file = sys_lpi_file_debugfs;
7915 		BIC_PRESENT(BIC_SYS_LPI);
7916 	} else {
7917 		sys_lpi_file_sysfs = NULL;
7918 		BIC_NOT_PRESENT(BIC_SYS_LPI);
7919 	}
7920 
7921 }
7922 
7923 void probe_pstates(void)
7924 {
7925 	probe_bclk();
7926 
7927 	if (quiet)
7928 		return;
7929 
7930 	dump_platform_info();
7931 	dump_turbo_ratio_info();
7932 	dump_sysfs_pstate_config();
7933 	decode_misc_pwr_mgmt_msr();
7934 
7935 	for_all_cpus(print_hwp, ODD_COUNTERS);
7936 	for_all_cpus(print_epb, ODD_COUNTERS);
7937 	for_all_cpus(print_perf_limit, ODD_COUNTERS);
7938 }
7939 
7940 void process_cpuid()
7941 {
7942 	unsigned int eax, ebx, ecx, edx;
7943 	unsigned int fms, family, model, stepping, ecx_flags, edx_flags;
7944 	unsigned long long ucode_patch = 0;
7945 	bool ucode_patch_valid = false;
7946 
7947 	eax = ebx = ecx = edx = 0;
7948 
7949 	__cpuid(0, max_level, ebx, ecx, edx);
7950 
7951 	if (ebx == 0x756e6547 && ecx == 0x6c65746e && edx == 0x49656e69)
7952 		genuine_intel = 1;
7953 	else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65)
7954 		authentic_amd = 1;
7955 	else if (ebx == 0x6f677948 && ecx == 0x656e6975 && edx == 0x6e65476e)
7956 		hygon_genuine = 1;
7957 
7958 	if (!quiet)
7959 		fprintf(outf, "CPUID(0): %.4s%.4s%.4s 0x%x CPUID levels\n",
7960 			(char *)&ebx, (char *)&edx, (char *)&ecx, max_level);
7961 
7962 	__cpuid(1, fms, ebx, ecx, edx);
7963 	family = (fms >> 8) & 0xf;
7964 	model = (fms >> 4) & 0xf;
7965 	stepping = fms & 0xf;
7966 	if (family == 0xf)
7967 		family += (fms >> 20) & 0xff;
7968 	if (family >= 6)
7969 		model += ((fms >> 16) & 0xf) << 4;
7970 	ecx_flags = ecx;
7971 	edx_flags = edx;
7972 
7973 	if (!no_msr) {
7974 		if (get_msr(sched_getcpu(), MSR_IA32_UCODE_REV, &ucode_patch))
7975 			warnx("get_msr(UCODE)");
7976 		else
7977 			ucode_patch_valid = true;
7978 	}
7979 
7980 	/*
7981 	 * check max extended function levels of CPUID.
7982 	 * This is needed to check for invariant TSC.
7983 	 * This check is valid for both Intel and AMD.
7984 	 */
7985 	ebx = ecx = edx = 0;
7986 	__cpuid(0x80000000, max_extended_level, ebx, ecx, edx);
7987 
7988 	if (!quiet) {
7989 		fprintf(outf, "CPUID(1): family:model:stepping 0x%x:%x:%x (%d:%d:%d)",
7990 			family, model, stepping, family, model, stepping);
7991 		if (ucode_patch_valid)
7992 			fprintf(outf, " microcode 0x%x", (unsigned int)((ucode_patch >> 32) & 0xFFFFFFFF));
7993 		fputc('\n', outf);
7994 
7995 		fprintf(outf, "CPUID(0x80000000): max_extended_levels: 0x%x\n", max_extended_level);
7996 		fprintf(outf, "CPUID(1): %s %s %s %s %s %s %s %s %s %s\n",
7997 			ecx_flags & (1 << 0) ? "SSE3" : "-",
7998 			ecx_flags & (1 << 3) ? "MONITOR" : "-",
7999 			ecx_flags & (1 << 6) ? "SMX" : "-",
8000 			ecx_flags & (1 << 7) ? "EIST" : "-",
8001 			ecx_flags & (1 << 8) ? "TM2" : "-",
8002 			edx_flags & (1 << 4) ? "TSC" : "-",
8003 			edx_flags & (1 << 5) ? "MSR" : "-",
8004 			edx_flags & (1 << 22) ? "ACPI-TM" : "-",
8005 			edx_flags & (1 << 28) ? "HT" : "-", edx_flags & (1 << 29) ? "TM" : "-");
8006 	}
8007 
8008 	probe_platform_features(family, model);
8009 
8010 	if (!(edx_flags & (1 << 5)))
8011 		errx(1, "CPUID: no MSR");
8012 
8013 	if (max_extended_level >= 0x80000007) {
8014 
8015 		/*
8016 		 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
8017 		 * this check is valid for both Intel and AMD
8018 		 */
8019 		__cpuid(0x80000007, eax, ebx, ecx, edx);
8020 		has_invariant_tsc = edx & (1 << 8);
8021 	}
8022 
8023 	/*
8024 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
8025 	 * this check is valid for both Intel and AMD
8026 	 */
8027 
8028 	__cpuid(0x6, eax, ebx, ecx, edx);
8029 	has_aperf = ecx & (1 << 0);
8030 	do_dts = eax & (1 << 0);
8031 	if (do_dts)
8032 		BIC_PRESENT(BIC_CoreTmp);
8033 	has_turbo = eax & (1 << 1);
8034 	do_ptm = eax & (1 << 6);
8035 	if (do_ptm)
8036 		BIC_PRESENT(BIC_PkgTmp);
8037 	has_hwp = eax & (1 << 7);
8038 	has_hwp_notify = eax & (1 << 8);
8039 	has_hwp_activity_window = eax & (1 << 9);
8040 	has_hwp_epp = eax & (1 << 10);
8041 	has_hwp_pkg = eax & (1 << 11);
8042 	has_epb = ecx & (1 << 3);
8043 
8044 	if (!quiet)
8045 		fprintf(outf, "CPUID(6): %sAPERF, %sTURBO, %sDTS, %sPTM, %sHWP, "
8046 			"%sHWPnotify, %sHWPwindow, %sHWPepp, %sHWPpkg, %sEPB\n",
8047 			has_aperf ? "" : "No-",
8048 			has_turbo ? "" : "No-",
8049 			do_dts ? "" : "No-",
8050 			do_ptm ? "" : "No-",
8051 			has_hwp ? "" : "No-",
8052 			has_hwp_notify ? "" : "No-",
8053 			has_hwp_activity_window ? "" : "No-",
8054 			has_hwp_epp ? "" : "No-", has_hwp_pkg ? "" : "No-", has_epb ? "" : "No-");
8055 
8056 	if (!quiet)
8057 		decode_misc_enable_msr();
8058 
8059 	if (max_level >= 0x7 && !quiet) {
8060 		int has_sgx;
8061 
8062 		ecx = 0;
8063 
8064 		__cpuid_count(0x7, 0, eax, ebx, ecx, edx);
8065 
8066 		has_sgx = ebx & (1 << 2);
8067 
8068 		is_hybrid = edx & (1 << 15);
8069 
8070 		fprintf(outf, "CPUID(7): %sSGX %sHybrid\n", has_sgx ? "" : "No-", is_hybrid ? "" : "No-");
8071 
8072 		if (has_sgx)
8073 			decode_feature_control_msr();
8074 	}
8075 
8076 	if (max_level >= 0x15) {
8077 		unsigned int eax_crystal;
8078 		unsigned int ebx_tsc;
8079 
8080 		/*
8081 		 * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
8082 		 */
8083 		eax_crystal = ebx_tsc = crystal_hz = edx = 0;
8084 		__cpuid(0x15, eax_crystal, ebx_tsc, crystal_hz, edx);
8085 
8086 		if (ebx_tsc != 0) {
8087 			if (!quiet && (ebx != 0))
8088 				fprintf(outf, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n",
8089 					eax_crystal, ebx_tsc, crystal_hz);
8090 
8091 			if (crystal_hz == 0)
8092 				crystal_hz = platform->crystal_freq;
8093 
8094 			if (crystal_hz) {
8095 				tsc_hz = (unsigned long long)crystal_hz *ebx_tsc / eax_crystal;
8096 				if (!quiet)
8097 					fprintf(outf, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n",
8098 						tsc_hz / 1000000, crystal_hz, ebx_tsc, eax_crystal);
8099 			}
8100 		}
8101 	}
8102 	if (max_level >= 0x16) {
8103 		unsigned int base_mhz, max_mhz, bus_mhz, edx;
8104 
8105 		/*
8106 		 * CPUID 16H Base MHz, Max MHz, Bus MHz
8107 		 */
8108 		base_mhz = max_mhz = bus_mhz = edx = 0;
8109 
8110 		__cpuid(0x16, base_mhz, max_mhz, bus_mhz, edx);
8111 
8112 		bclk = bus_mhz;
8113 
8114 		base_hz = base_mhz * 1000000;
8115 		has_base_hz = 1;
8116 
8117 		if (platform->enable_tsc_tweak)
8118 			tsc_tweak = base_hz / tsc_hz;
8119 
8120 		if (!quiet)
8121 			fprintf(outf, "CPUID(0x16): base_mhz: %d max_mhz: %d bus_mhz: %d\n",
8122 				base_mhz, max_mhz, bus_mhz);
8123 	}
8124 
8125 	if (has_aperf)
8126 		aperf_mperf_multiplier = platform->need_perf_multiplier ? 1024 : 1;
8127 
8128 	BIC_PRESENT(BIC_IRQ);
8129 	BIC_PRESENT(BIC_TSC_MHz);
8130 }
8131 
8132 static void counter_info_init(void)
8133 {
8134 	for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i) {
8135 		struct cstate_counter_arch_info *const cai = &ccstate_counter_arch_infos[i];
8136 
8137 		if (platform->has_msr_knl_core_c6_residency && cai->msr == MSR_CORE_C6_RESIDENCY)
8138 			cai->msr = MSR_KNL_CORE_C6_RESIDENCY;
8139 
8140 		if (!platform->has_msr_core_c1_res && cai->msr == MSR_CORE_C1_RES)
8141 			cai->msr = 0;
8142 
8143 		if (platform->has_msr_atom_pkg_c6_residency && cai->msr == MSR_PKG_C6_RESIDENCY)
8144 			cai->msr = MSR_ATOM_PKG_C6_RESIDENCY;
8145 	}
8146 
8147 	for (int i = 0; i < NUM_MSR_COUNTERS; ++i) {
8148 		msr_counter_arch_infos[i].present = false;
8149 		msr_counter_arch_infos[i].needed = false;
8150 	}
8151 }
8152 
8153 void probe_pm_features(void)
8154 {
8155 	probe_pstates();
8156 
8157 	probe_cstates();
8158 
8159 	probe_lpi();
8160 
8161 	probe_intel_uncore_frequency();
8162 
8163 	probe_graphics();
8164 
8165 	probe_rapl();
8166 
8167 	probe_thermal();
8168 
8169 	if (platform->has_nhm_msrs && !no_msr)
8170 		BIC_PRESENT(BIC_SMI);
8171 
8172 	if (!quiet)
8173 		decode_misc_feature_control();
8174 }
8175 
8176 /*
8177  * in /dev/cpu/ return success for names that are numbers
8178  * ie. filter out ".", "..", "microcode".
8179  */
8180 int dir_filter(const struct dirent *dirp)
8181 {
8182 	if (isdigit(dirp->d_name[0]))
8183 		return 1;
8184 	else
8185 		return 0;
8186 }
8187 
8188 void topology_probe(bool startup)
8189 {
8190 	int i;
8191 	int max_core_id = 0;
8192 	int max_package_id = 0;
8193 	int max_siblings = 0;
8194 
8195 	/* Initialize num_cpus, max_cpu_num */
8196 	set_max_cpu_num();
8197 	topo.num_cpus = 0;
8198 	for_all_proc_cpus(count_cpus);
8199 	if (!summary_only && topo.num_cpus > 1)
8200 		BIC_PRESENT(BIC_CPU);
8201 
8202 	if (debug > 1)
8203 		fprintf(outf, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
8204 
8205 	cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
8206 	if (cpus == NULL)
8207 		err(1, "calloc cpus");
8208 
8209 	/*
8210 	 * Allocate and initialize cpu_present_set
8211 	 */
8212 	cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
8213 	if (cpu_present_set == NULL)
8214 		err(3, "CPU_ALLOC");
8215 	cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8216 	CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
8217 	for_all_proc_cpus(mark_cpu_present);
8218 
8219 	/*
8220 	 * Allocate and initialize cpu_effective_set
8221 	 */
8222 	cpu_effective_set = CPU_ALLOC((topo.max_cpu_num + 1));
8223 	if (cpu_effective_set == NULL)
8224 		err(3, "CPU_ALLOC");
8225 	cpu_effective_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8226 	CPU_ZERO_S(cpu_effective_setsize, cpu_effective_set);
8227 	update_effective_set(startup);
8228 
8229 	/*
8230 	 * Allocate and initialize cpu_allowed_set
8231 	 */
8232 	cpu_allowed_set = CPU_ALLOC((topo.max_cpu_num + 1));
8233 	if (cpu_allowed_set == NULL)
8234 		err(3, "CPU_ALLOC");
8235 	cpu_allowed_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8236 	CPU_ZERO_S(cpu_allowed_setsize, cpu_allowed_set);
8237 
8238 	/*
8239 	 * Validate and update cpu_allowed_set.
8240 	 *
8241 	 * Make sure all cpus in cpu_subset are also in cpu_present_set during startup.
8242 	 * Give a warning when cpus in cpu_subset become unavailable at runtime.
8243 	 * Give a warning when cpus are not effective because of cgroup setting.
8244 	 *
8245 	 * cpu_allowed_set is the intersection of cpu_present_set/cpu_effective_set/cpu_subset.
8246 	 */
8247 	for (i = 0; i < CPU_SUBSET_MAXCPUS; ++i) {
8248 		if (cpu_subset && !CPU_ISSET_S(i, cpu_subset_size, cpu_subset))
8249 			continue;
8250 
8251 		if (!CPU_ISSET_S(i, cpu_present_setsize, cpu_present_set)) {
8252 			if (cpu_subset) {
8253 				/* cpus in cpu_subset must be in cpu_present_set during startup */
8254 				if (startup)
8255 					err(1, "cpu%d not present", i);
8256 				else
8257 					fprintf(stderr, "cpu%d not present\n", i);
8258 			}
8259 			continue;
8260 		}
8261 
8262 		if (CPU_COUNT_S(cpu_effective_setsize, cpu_effective_set)) {
8263 			if (!CPU_ISSET_S(i, cpu_effective_setsize, cpu_effective_set)) {
8264 				fprintf(stderr, "cpu%d not effective\n", i);
8265 				continue;
8266 			}
8267 		}
8268 
8269 		CPU_SET_S(i, cpu_allowed_setsize, cpu_allowed_set);
8270 	}
8271 
8272 	if (!CPU_COUNT_S(cpu_allowed_setsize, cpu_allowed_set))
8273 		err(-ENODEV, "No valid cpus found");
8274 	sched_setaffinity(0, cpu_allowed_setsize, cpu_allowed_set);
8275 
8276 	/*
8277 	 * Allocate and initialize cpu_affinity_set
8278 	 */
8279 	cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
8280 	if (cpu_affinity_set == NULL)
8281 		err(3, "CPU_ALLOC");
8282 	cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8283 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
8284 
8285 	for_all_proc_cpus(init_thread_id);
8286 
8287 	/*
8288 	 * For online cpus
8289 	 * find max_core_id, max_package_id
8290 	 */
8291 	for (i = 0; i <= topo.max_cpu_num; ++i) {
8292 		int siblings;
8293 
8294 		if (cpu_is_not_present(i)) {
8295 			if (debug > 1)
8296 				fprintf(outf, "cpu%d NOT PRESENT\n", i);
8297 			continue;
8298 		}
8299 
8300 		cpus[i].logical_cpu_id = i;
8301 
8302 		/* get package information */
8303 		cpus[i].physical_package_id = get_physical_package_id(i);
8304 		if (cpus[i].physical_package_id > max_package_id)
8305 			max_package_id = cpus[i].physical_package_id;
8306 
8307 		/* get die information */
8308 		cpus[i].die_id = get_die_id(i);
8309 		if (cpus[i].die_id > topo.max_die_id)
8310 			topo.max_die_id = cpus[i].die_id;
8311 
8312 		/* get numa node information */
8313 		cpus[i].physical_node_id = get_physical_node_id(&cpus[i]);
8314 		if (cpus[i].physical_node_id > topo.max_node_num)
8315 			topo.max_node_num = cpus[i].physical_node_id;
8316 
8317 		/* get core information */
8318 		cpus[i].physical_core_id = get_core_id(i);
8319 		if (cpus[i].physical_core_id > max_core_id)
8320 			max_core_id = cpus[i].physical_core_id;
8321 
8322 		/* get thread information */
8323 		siblings = get_thread_siblings(&cpus[i]);
8324 		if (siblings > max_siblings)
8325 			max_siblings = siblings;
8326 		if (cpus[i].thread_id == 0)
8327 			topo.num_cores++;
8328 	}
8329 	topo.max_core_id = max_core_id;
8330 	topo.max_package_id = max_package_id;
8331 
8332 	topo.cores_per_node = max_core_id + 1;
8333 	if (debug > 1)
8334 		fprintf(outf, "max_core_id %d, sizing for %d cores per package\n", max_core_id, topo.cores_per_node);
8335 	if (!summary_only && topo.cores_per_node > 1)
8336 		BIC_PRESENT(BIC_Core);
8337 
8338 	topo.num_die = topo.max_die_id + 1;
8339 	if (debug > 1)
8340 		fprintf(outf, "max_die_id %d, sizing for %d die\n", topo.max_die_id, topo.num_die);
8341 	if (!summary_only && topo.num_die > 1)
8342 		BIC_PRESENT(BIC_Die);
8343 
8344 	topo.num_packages = max_package_id + 1;
8345 	if (debug > 1)
8346 		fprintf(outf, "max_package_id %d, sizing for %d packages\n", max_package_id, topo.num_packages);
8347 	if (!summary_only && topo.num_packages > 1)
8348 		BIC_PRESENT(BIC_Package);
8349 
8350 	set_node_data();
8351 	if (debug > 1)
8352 		fprintf(outf, "nodes_per_pkg %d\n", topo.nodes_per_pkg);
8353 	if (!summary_only && topo.nodes_per_pkg > 1)
8354 		BIC_PRESENT(BIC_Node);
8355 
8356 	topo.threads_per_core = max_siblings;
8357 	if (debug > 1)
8358 		fprintf(outf, "max_siblings %d\n", max_siblings);
8359 
8360 	if (debug < 1)
8361 		return;
8362 
8363 	for (i = 0; i <= topo.max_cpu_num; ++i) {
8364 		if (cpu_is_not_present(i))
8365 			continue;
8366 		fprintf(outf,
8367 			"cpu %d pkg %d die %d node %d lnode %d core %d thread %d\n",
8368 			i, cpus[i].physical_package_id, cpus[i].die_id,
8369 			cpus[i].physical_node_id, cpus[i].logical_node_id, cpus[i].physical_core_id, cpus[i].thread_id);
8370 	}
8371 
8372 }
8373 
8374 void allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
8375 {
8376 	int i;
8377 	int num_cores = topo.cores_per_node * topo.nodes_per_pkg * topo.num_packages;
8378 	int num_threads = topo.threads_per_core * num_cores;
8379 
8380 	*t = calloc(num_threads, sizeof(struct thread_data));
8381 	if (*t == NULL)
8382 		goto error;
8383 
8384 	for (i = 0; i < num_threads; i++)
8385 		(*t)[i].cpu_id = -1;
8386 
8387 	*c = calloc(num_cores, sizeof(struct core_data));
8388 	if (*c == NULL)
8389 		goto error;
8390 
8391 	for (i = 0; i < num_cores; i++) {
8392 		(*c)[i].core_id = -1;
8393 		(*c)[i].base_cpu = -1;
8394 	}
8395 
8396 	*p = calloc(topo.num_packages, sizeof(struct pkg_data));
8397 	if (*p == NULL)
8398 		goto error;
8399 
8400 	for (i = 0; i < topo.num_packages; i++) {
8401 		(*p)[i].package_id = i;
8402 		(*p)[i].base_cpu = -1;
8403 	}
8404 
8405 	return;
8406 error:
8407 	err(1, "calloc counters");
8408 }
8409 
8410 /*
8411  * init_counter()
8412  *
8413  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
8414  */
8415 void init_counter(struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base, int cpu_id)
8416 {
8417 	int pkg_id = cpus[cpu_id].physical_package_id;
8418 	int node_id = cpus[cpu_id].logical_node_id;
8419 	int core_id = cpus[cpu_id].physical_core_id;
8420 	int thread_id = cpus[cpu_id].thread_id;
8421 	struct thread_data *t;
8422 	struct core_data *c;
8423 	struct pkg_data *p;
8424 
8425 	/* Workaround for systems where physical_node_id==-1
8426 	 * and logical_node_id==(-1 - topo.num_cpus)
8427 	 */
8428 	if (node_id < 0)
8429 		node_id = 0;
8430 
8431 	t = GET_THREAD(thread_base, thread_id, core_id, node_id, pkg_id);
8432 	c = GET_CORE(core_base, core_id, node_id, pkg_id);
8433 	p = GET_PKG(pkg_base, pkg_id);
8434 
8435 	t->cpu_id = cpu_id;
8436 	if (!cpu_is_not_allowed(cpu_id)) {
8437 		if (c->base_cpu < 0)
8438 			c->base_cpu = t->cpu_id;
8439 		if (p->base_cpu < 0)
8440 			p->base_cpu = t->cpu_id;
8441 	}
8442 
8443 	c->core_id = core_id;
8444 	p->package_id = pkg_id;
8445 }
8446 
8447 int initialize_counters(int cpu_id)
8448 {
8449 	init_counter(EVEN_COUNTERS, cpu_id);
8450 	init_counter(ODD_COUNTERS, cpu_id);
8451 	return 0;
8452 }
8453 
8454 void allocate_output_buffer()
8455 {
8456 	output_buffer = calloc(1, (1 + topo.num_cpus) * 2048);
8457 	outp = output_buffer;
8458 	if (outp == NULL)
8459 		err(-1, "calloc output buffer");
8460 }
8461 
8462 void allocate_fd_percpu(void)
8463 {
8464 	fd_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8465 	if (fd_percpu == NULL)
8466 		err(-1, "calloc fd_percpu");
8467 }
8468 
8469 void allocate_irq_buffers(void)
8470 {
8471 	irq_column_2_cpu = calloc(topo.num_cpus, sizeof(int));
8472 	if (irq_column_2_cpu == NULL)
8473 		err(-1, "calloc %d", topo.num_cpus);
8474 
8475 	irqs_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8476 	if (irqs_per_cpu == NULL)
8477 		err(-1, "calloc %d", topo.max_cpu_num + 1);
8478 }
8479 
8480 int update_topo(struct thread_data *t, struct core_data *c, struct pkg_data *p)
8481 {
8482 	topo.allowed_cpus++;
8483 	if ((int)t->cpu_id == c->base_cpu)
8484 		topo.allowed_cores++;
8485 	if ((int)t->cpu_id == p->base_cpu)
8486 		topo.allowed_packages++;
8487 
8488 	return 0;
8489 }
8490 
8491 void topology_update(void)
8492 {
8493 	topo.allowed_cpus = 0;
8494 	topo.allowed_cores = 0;
8495 	topo.allowed_packages = 0;
8496 	for_all_cpus(update_topo, ODD_COUNTERS);
8497 }
8498 
8499 void setup_all_buffers(bool startup)
8500 {
8501 	topology_probe(startup);
8502 	allocate_irq_buffers();
8503 	allocate_fd_percpu();
8504 	allocate_counters(&thread_even, &core_even, &package_even);
8505 	allocate_counters(&thread_odd, &core_odd, &package_odd);
8506 	allocate_output_buffer();
8507 	for_all_proc_cpus(initialize_counters);
8508 	topology_update();
8509 }
8510 
8511 void set_base_cpu(void)
8512 {
8513 	int i;
8514 
8515 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
8516 		if (cpu_is_not_allowed(i))
8517 			continue;
8518 		base_cpu = i;
8519 		if (debug > 1)
8520 			fprintf(outf, "base_cpu = %d\n", base_cpu);
8521 		return;
8522 	}
8523 	err(-ENODEV, "No valid cpus found");
8524 }
8525 
8526 bool has_added_counters(void)
8527 {
8528 	/*
8529 	 * It only makes sense to call this after the command line is parsed,
8530 	 * otherwise sys structure is not populated.
8531 	 */
8532 
8533 	return sys.added_core_counters | sys.added_thread_counters | sys.added_package_counters;
8534 }
8535 
8536 void check_msr_access(void)
8537 {
8538 	check_dev_msr();
8539 	check_msr_permission();
8540 
8541 	if (no_msr)
8542 		bic_disable_msr_access();
8543 }
8544 
8545 void check_perf_access(void)
8546 {
8547 	if (no_perf || !BIC_IS_ENABLED(BIC_IPC) || !has_instr_count_access())
8548 		bic_enabled &= ~BIC_IPC;
8549 }
8550 
8551 int added_perf_counters_init_(struct perf_counter_info *pinfo)
8552 {
8553 	size_t num_domains = 0;
8554 	unsigned int next_domain;
8555 	bool *domain_visited;
8556 	unsigned int perf_type, perf_config;
8557 	double perf_scale;
8558 	int fd_perf;
8559 
8560 	if (!pinfo)
8561 		return 0;
8562 
8563 	const size_t max_num_domains = MAX(topo.max_cpu_num + 1, MAX(topo.max_core_id + 1, topo.max_package_id + 1));
8564 
8565 	domain_visited = calloc(max_num_domains, sizeof(*domain_visited));
8566 
8567 	while (pinfo) {
8568 		switch (pinfo->scope) {
8569 		case SCOPE_CPU:
8570 			num_domains = topo.max_cpu_num + 1;
8571 			break;
8572 
8573 		case SCOPE_CORE:
8574 			num_domains = topo.max_core_id + 1;
8575 			break;
8576 
8577 		case SCOPE_PACKAGE:
8578 			num_domains = topo.max_package_id + 1;
8579 			break;
8580 		}
8581 
8582 		/* Allocate buffer for file descriptor for each domain. */
8583 		pinfo->fd_perf_per_domain = calloc(num_domains, sizeof(*pinfo->fd_perf_per_domain));
8584 		if (!pinfo->fd_perf_per_domain)
8585 			errx(1, "%s: alloc %s", __func__, "fd_perf_per_domain");
8586 
8587 		for (size_t i = 0; i < num_domains; ++i)
8588 			pinfo->fd_perf_per_domain[i] = -1;
8589 
8590 		pinfo->num_domains = num_domains;
8591 		pinfo->scale = 1.0;
8592 
8593 		memset(domain_visited, 0, max_num_domains * sizeof(*domain_visited));
8594 
8595 		for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
8596 
8597 			next_domain = cpu_to_domain(pinfo, cpu);
8598 
8599 			assert(next_domain < num_domains);
8600 
8601 			if (cpu_is_not_allowed(cpu))
8602 				continue;
8603 
8604 			if (domain_visited[next_domain])
8605 				continue;
8606 
8607 			perf_type = read_perf_type(pinfo->device);
8608 			if (perf_type == (unsigned int)-1) {
8609 				warnx("%s: perf/%s/%s: failed to read %s",
8610 				      __func__, pinfo->device, pinfo->event, "type");
8611 				continue;
8612 			}
8613 
8614 			perf_config = read_perf_config(pinfo->device, pinfo->event);
8615 			if (perf_config == (unsigned int)-1) {
8616 				warnx("%s: perf/%s/%s: failed to read %s",
8617 				      __func__, pinfo->device, pinfo->event, "config");
8618 				continue;
8619 			}
8620 
8621 			/* Scale is not required, some counters just don't have it. */
8622 			perf_scale = read_perf_scale(pinfo->device, pinfo->event);
8623 			if (perf_scale == 0.0)
8624 				perf_scale = 1.0;
8625 
8626 			fd_perf = open_perf_counter(cpu, perf_type, perf_config, -1, 0);
8627 			if (fd_perf == -1) {
8628 				warnx("%s: perf/%s/%s: failed to open counter on cpu%d",
8629 				      __func__, pinfo->device, pinfo->event, cpu);
8630 				continue;
8631 			}
8632 
8633 			domain_visited[next_domain] = 1;
8634 			pinfo->fd_perf_per_domain[next_domain] = fd_perf;
8635 			pinfo->scale = perf_scale;
8636 
8637 			if (debug)
8638 				fprintf(stderr, "Add perf/%s/%s cpu%d: %d\n",
8639 					pinfo->device, pinfo->event, cpu, pinfo->fd_perf_per_domain[next_domain]);
8640 		}
8641 
8642 		pinfo = pinfo->next;
8643 	}
8644 
8645 	free(domain_visited);
8646 
8647 	return 0;
8648 }
8649 
8650 void added_perf_counters_init(void)
8651 {
8652 	if (added_perf_counters_init_(sys.perf_tp))
8653 		errx(1, "%s: %s", __func__, "thread");
8654 
8655 	if (added_perf_counters_init_(sys.perf_cp))
8656 		errx(1, "%s: %s", __func__, "core");
8657 
8658 	if (added_perf_counters_init_(sys.perf_pp))
8659 		errx(1, "%s: %s", __func__, "package");
8660 }
8661 
8662 int parse_telem_info_file(int fd_dir, const char *info_filename, const char *format, unsigned long *output)
8663 {
8664 	int fd_telem_info;
8665 	FILE *file_telem_info;
8666 	unsigned long value;
8667 
8668 	fd_telem_info = openat(fd_dir, info_filename, O_RDONLY);
8669 	if (fd_telem_info == -1)
8670 		return -1;
8671 
8672 	file_telem_info = fdopen(fd_telem_info, "r");
8673 	if (file_telem_info == NULL) {
8674 		close(fd_telem_info);
8675 		return -1;
8676 	}
8677 
8678 	if (fscanf(file_telem_info, format, &value) != 1) {
8679 		fclose(file_telem_info);
8680 		return -1;
8681 	}
8682 
8683 	fclose(file_telem_info);
8684 
8685 	*output = value;
8686 
8687 	return 0;
8688 }
8689 
8690 struct pmt_mmio *pmt_mmio_open(unsigned int target_guid)
8691 {
8692 	DIR *dirp;
8693 	struct dirent *entry;
8694 	struct stat st;
8695 	unsigned int telem_idx;
8696 	int fd_telem_dir, fd_pmt;
8697 	unsigned long guid, size, offset;
8698 	size_t mmap_size;
8699 	void *mmio;
8700 	struct pmt_mmio *ret = NULL;
8701 
8702 	if (stat(SYSFS_TELEM_PATH, &st) == -1)
8703 		return NULL;
8704 
8705 	dirp = opendir(SYSFS_TELEM_PATH);
8706 	if (dirp == NULL)
8707 		return NULL;
8708 
8709 	for (;;) {
8710 		entry = readdir(dirp);
8711 
8712 		if (entry == NULL)
8713 			break;
8714 
8715 		if (strcmp(entry->d_name, ".") == 0)
8716 			continue;
8717 
8718 		if (strcmp(entry->d_name, "..") == 0)
8719 			continue;
8720 
8721 		if (sscanf(entry->d_name, "telem%u", &telem_idx) != 1)
8722 			continue;
8723 
8724 		if (fstatat(dirfd(dirp), entry->d_name, &st, 0) == -1) {
8725 			break;
8726 		}
8727 
8728 		if (!S_ISDIR(st.st_mode))
8729 			continue;
8730 
8731 		fd_telem_dir = openat(dirfd(dirp), entry->d_name, O_RDONLY);
8732 		if (fd_telem_dir == -1) {
8733 			break;
8734 		}
8735 
8736 		if (parse_telem_info_file(fd_telem_dir, "guid", "%lx", &guid)) {
8737 			close(fd_telem_dir);
8738 			break;
8739 		}
8740 
8741 		if (parse_telem_info_file(fd_telem_dir, "size", "%lu", &size)) {
8742 			close(fd_telem_dir);
8743 			break;
8744 		}
8745 
8746 		if (guid != target_guid) {
8747 			close(fd_telem_dir);
8748 			continue;
8749 		}
8750 
8751 		if (parse_telem_info_file(fd_telem_dir, "offset", "%lu", &offset)) {
8752 			close(fd_telem_dir);
8753 			break;
8754 		}
8755 
8756 		assert(offset == 0);
8757 
8758 		fd_pmt = openat(fd_telem_dir, "telem", O_RDONLY);
8759 		if (fd_pmt == -1)
8760 			goto loop_cleanup_and_break;
8761 
8762 		mmap_size = (size + 0x1000UL) & (~0x1000UL);
8763 		mmio = mmap(0, mmap_size, PROT_READ, MAP_SHARED, fd_pmt, 0);
8764 		if (mmio != MAP_FAILED) {
8765 
8766 			if (debug)
8767 				fprintf(stderr, "%s: 0x%lx mmaped at: %p\n", __func__, guid, mmio);
8768 
8769 			ret = calloc(1, sizeof(*ret));
8770 
8771 			if (!ret) {
8772 				fprintf(stderr, "%s: Failed to allocate pmt_mmio\n", __func__);
8773 				exit(1);
8774 			}
8775 
8776 			ret->guid = guid;
8777 			ret->mmio_base = mmio;
8778 			ret->pmt_offset = offset;
8779 			ret->size = size;
8780 
8781 			ret->next = pmt_mmios;
8782 			pmt_mmios = ret;
8783 		}
8784 
8785 loop_cleanup_and_break:
8786 		close(fd_pmt);
8787 		close(fd_telem_dir);
8788 		break;
8789 	}
8790 
8791 	closedir(dirp);
8792 
8793 	return ret;
8794 }
8795 
8796 struct pmt_mmio *pmt_mmio_find(unsigned int guid)
8797 {
8798 	struct pmt_mmio *pmmio = pmt_mmios;
8799 
8800 	while (pmmio) {
8801 		if (pmmio->guid == guid)
8802 			return pmmio;
8803 
8804 		pmmio = pmmio->next;
8805 	}
8806 
8807 	return NULL;
8808 }
8809 
8810 void *pmt_get_counter_pointer(struct pmt_mmio *pmmio, unsigned long counter_offset)
8811 {
8812 	char *ret;
8813 
8814 	/* Get base of mmaped PMT file. */
8815 	ret = (char *)pmmio->mmio_base;
8816 
8817 	/*
8818 	 * Apply PMT MMIO offset to obtain beginning of the mmaped telemetry data.
8819 	 * It's not guaranteed that the mmaped memory begins with the telemetry data
8820 	 *      - we might have to apply the offset first.
8821 	 */
8822 	ret += pmmio->pmt_offset;
8823 
8824 	/* Apply the counter offset to get the address to the mmaped counter. */
8825 	ret += counter_offset;
8826 
8827 	return ret;
8828 }
8829 
8830 struct pmt_mmio *pmt_add_guid(unsigned int guid)
8831 {
8832 	struct pmt_mmio *ret;
8833 
8834 	ret = pmt_mmio_find(guid);
8835 	if (!ret)
8836 		ret = pmt_mmio_open(guid);
8837 
8838 	return ret;
8839 }
8840 
8841 enum pmt_open_mode {
8842 	PMT_OPEN_TRY,		/* Open failure is not an error. */
8843 	PMT_OPEN_REQUIRED,	/* Open failure is a fatal error. */
8844 };
8845 
8846 struct pmt_counter *pmt_find_counter(struct pmt_counter *pcounter, const char *name)
8847 {
8848 	while (pcounter) {
8849 		if (strcmp(pcounter->name, name) == 0)
8850 			break;
8851 
8852 		pcounter = pcounter->next;
8853 	}
8854 
8855 	return pcounter;
8856 }
8857 
8858 struct pmt_counter **pmt_get_scope_root(enum counter_scope scope)
8859 {
8860 	switch (scope) {
8861 	case SCOPE_CPU:
8862 		return &sys.pmt_tp;
8863 	case SCOPE_CORE:
8864 		return &sys.pmt_cp;
8865 	case SCOPE_PACKAGE:
8866 		return &sys.pmt_pp;
8867 	}
8868 
8869 	__builtin_unreachable();
8870 }
8871 
8872 void pmt_counter_add_domain(struct pmt_counter *pcounter, unsigned long *pmmio, unsigned int domain_id)
8873 {
8874 	/* Make sure the new domain fits. */
8875 	if (domain_id >= pcounter->num_domains)
8876 		pmt_counter_resize(pcounter, domain_id + 1);
8877 
8878 	assert(pcounter->domains);
8879 	assert(domain_id < pcounter->num_domains);
8880 
8881 	pcounter->domains[domain_id].pcounter = pmmio;
8882 }
8883 
8884 int pmt_add_counter(unsigned int guid, const char *name, enum pmt_datatype type,
8885 		    unsigned int lsb, unsigned int msb, unsigned int offset, enum counter_scope scope,
8886 		    enum counter_format format, unsigned int domain_id, enum pmt_open_mode mode)
8887 {
8888 	struct pmt_mmio *mmio;
8889 	struct pmt_counter *pcounter;
8890 	struct pmt_counter **const pmt_root = pmt_get_scope_root(scope);
8891 	bool new_counter = false;
8892 	int conflict = 0;
8893 
8894 	if (lsb > msb) {
8895 		fprintf(stderr, "%s: %s: `%s` must be satisfied\n", __func__, "lsb <= msb", name);
8896 		exit(1);
8897 	}
8898 
8899 	if (msb >= 64) {
8900 		fprintf(stderr, "%s: %s: `%s` must be satisfied\n", __func__, "msb < 64", name);
8901 		exit(1);
8902 	}
8903 
8904 	mmio = pmt_add_guid(guid);
8905 	if (!mmio) {
8906 		if (mode != PMT_OPEN_TRY) {
8907 			fprintf(stderr, "%s: failed to map PMT MMIO for guid %x\n", __func__, guid);
8908 			exit(1);
8909 		}
8910 
8911 		return 1;
8912 	}
8913 
8914 	if (offset >= mmio->size) {
8915 		if (mode != PMT_OPEN_TRY) {
8916 			fprintf(stderr, "%s: offset %u outside of PMT MMIO size %u\n", __func__, offset, mmio->size);
8917 			exit(1);
8918 		}
8919 
8920 		return 1;
8921 	}
8922 
8923 	pcounter = pmt_find_counter(*pmt_root, name);
8924 	if (!pcounter) {
8925 		pcounter = calloc(1, sizeof(*pcounter));
8926 		new_counter = true;
8927 	}
8928 
8929 	if (new_counter) {
8930 		strncpy(pcounter->name, name, ARRAY_SIZE(pcounter->name) - 1);
8931 		pcounter->type = type;
8932 		pcounter->scope = scope;
8933 		pcounter->lsb = lsb;
8934 		pcounter->msb = msb;
8935 		pcounter->format = format;
8936 	} else {
8937 		conflict += pcounter->type != type;
8938 		conflict += pcounter->scope != scope;
8939 		conflict += pcounter->lsb != lsb;
8940 		conflict += pcounter->msb != msb;
8941 		conflict += pcounter->format != format;
8942 	}
8943 
8944 	if (conflict) {
8945 		fprintf(stderr, "%s: conflicting parameters for the PMT counter with the same name %s\n",
8946 			__func__, name);
8947 		exit(1);
8948 	}
8949 
8950 	pmt_counter_add_domain(pcounter, pmt_get_counter_pointer(mmio, offset), domain_id);
8951 
8952 	if (new_counter) {
8953 		pcounter->next = *pmt_root;
8954 		*pmt_root = pcounter;
8955 	}
8956 
8957 	return 0;
8958 }
8959 
8960 void pmt_init(void)
8961 {
8962 	if (BIC_IS_ENABLED(BIC_Diec6)) {
8963 		pmt_add_counter(PMT_MTL_DC6_GUID, "Die%c6", PMT_TYPE_XTAL_TIME, PMT_COUNTER_MTL_DC6_LSB,
8964 				PMT_COUNTER_MTL_DC6_MSB, PMT_COUNTER_MTL_DC6_OFFSET, SCOPE_PACKAGE, FORMAT_DELTA,
8965 				0, PMT_OPEN_TRY);
8966 	}
8967 }
8968 
8969 void turbostat_init()
8970 {
8971 	setup_all_buffers(true);
8972 	set_base_cpu();
8973 	check_msr_access();
8974 	check_perf_access();
8975 	process_cpuid();
8976 	counter_info_init();
8977 	probe_pm_features();
8978 	msr_perf_init();
8979 	linux_perf_init();
8980 	rapl_perf_init();
8981 	cstate_perf_init();
8982 	added_perf_counters_init();
8983 	pmt_init();
8984 
8985 	for_all_cpus(get_cpu_type, ODD_COUNTERS);
8986 	for_all_cpus(get_cpu_type, EVEN_COUNTERS);
8987 
8988 	if (BIC_IS_ENABLED(BIC_IPC) && has_aperf_access && get_instr_count_fd(base_cpu) != -1)
8989 		BIC_PRESENT(BIC_IPC);
8990 
8991 	/*
8992 	 * If TSC tweak is needed, but couldn't get it,
8993 	 * disable more BICs, since it can't be reported accurately.
8994 	 */
8995 	if (platform->enable_tsc_tweak && !has_base_hz) {
8996 		bic_enabled &= ~BIC_Busy;
8997 		bic_enabled &= ~BIC_Bzy_MHz;
8998 	}
8999 }
9000 
9001 int fork_it(char **argv)
9002 {
9003 	pid_t child_pid;
9004 	int status;
9005 
9006 	snapshot_proc_sysfs_files();
9007 	status = for_all_cpus(get_counters, EVEN_COUNTERS);
9008 	first_counter_read = 0;
9009 	if (status)
9010 		exit(status);
9011 	gettimeofday(&tv_even, (struct timezone *)NULL);
9012 
9013 	child_pid = fork();
9014 	if (!child_pid) {
9015 		/* child */
9016 		execvp(argv[0], argv);
9017 		err(errno, "exec %s", argv[0]);
9018 	} else {
9019 
9020 		/* parent */
9021 		if (child_pid == -1)
9022 			err(1, "fork");
9023 
9024 		signal(SIGINT, SIG_IGN);
9025 		signal(SIGQUIT, SIG_IGN);
9026 		if (waitpid(child_pid, &status, 0) == -1)
9027 			err(status, "waitpid");
9028 
9029 		if (WIFEXITED(status))
9030 			status = WEXITSTATUS(status);
9031 	}
9032 	/*
9033 	 * n.b. fork_it() does not check for errors from for_all_cpus()
9034 	 * because re-starting is problematic when forking
9035 	 */
9036 	snapshot_proc_sysfs_files();
9037 	for_all_cpus(get_counters, ODD_COUNTERS);
9038 	gettimeofday(&tv_odd, (struct timezone *)NULL);
9039 	timersub(&tv_odd, &tv_even, &tv_delta);
9040 	if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS))
9041 		fprintf(outf, "%s: Counter reset detected\n", progname);
9042 	else {
9043 		compute_average(EVEN_COUNTERS);
9044 		format_all_counters(EVEN_COUNTERS);
9045 	}
9046 
9047 	fprintf(outf, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec / 1000000.0);
9048 
9049 	flush_output_stderr();
9050 
9051 	return status;
9052 }
9053 
9054 int get_and_dump_counters(void)
9055 {
9056 	int status;
9057 
9058 	snapshot_proc_sysfs_files();
9059 	status = for_all_cpus(get_counters, ODD_COUNTERS);
9060 	if (status)
9061 		return status;
9062 
9063 	status = for_all_cpus(dump_counters, ODD_COUNTERS);
9064 	if (status)
9065 		return status;
9066 
9067 	flush_output_stdout();
9068 
9069 	return status;
9070 }
9071 
9072 void print_version()
9073 {
9074 	fprintf(outf, "turbostat version 2024.07.26 - Len Brown <lenb@kernel.org>\n");
9075 }
9076 
9077 #define COMMAND_LINE_SIZE 2048
9078 
9079 void print_bootcmd(void)
9080 {
9081 	char bootcmd[COMMAND_LINE_SIZE];
9082 	FILE *fp;
9083 	int ret;
9084 
9085 	memset(bootcmd, 0, COMMAND_LINE_SIZE);
9086 	fp = fopen("/proc/cmdline", "r");
9087 	if (!fp)
9088 		return;
9089 
9090 	ret = fread(bootcmd, sizeof(char), COMMAND_LINE_SIZE - 1, fp);
9091 	if (ret) {
9092 		bootcmd[ret] = '\0';
9093 		/* the last character is already '\n' */
9094 		fprintf(outf, "Kernel command line: %s", bootcmd);
9095 	}
9096 
9097 	fclose(fp);
9098 }
9099 
9100 struct msr_counter *find_msrp_by_name(struct msr_counter *head, char *name)
9101 {
9102 	struct msr_counter *mp;
9103 
9104 	for (mp = head; mp; mp = mp->next) {
9105 		if (debug)
9106 			fprintf(stderr, "%s: %s %s\n", __func__, name, mp->name);
9107 		if (!strncmp(name, mp->name, strlen(mp->name)))
9108 			return mp;
9109 	}
9110 	return NULL;
9111 }
9112 
9113 int add_counter(unsigned int msr_num, char *path, char *name,
9114 		unsigned int width, enum counter_scope scope,
9115 		enum counter_type type, enum counter_format format, int flags, int id)
9116 {
9117 	struct msr_counter *msrp;
9118 
9119 	if (no_msr && msr_num)
9120 		errx(1, "Requested MSR counter 0x%x, but in --no-msr mode", msr_num);
9121 
9122 	if (debug)
9123 		fprintf(stderr, "%s(msr%d, %s, %s, width%d, scope%d, type%d, format%d, flags%x, id%d)\n",
9124 			__func__, msr_num, path, name, width, scope, type, format, flags, id);
9125 
9126 	switch (scope) {
9127 
9128 	case SCOPE_CPU:
9129 		msrp = find_msrp_by_name(sys.tp, name);
9130 		if (msrp) {
9131 			if (debug)
9132 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
9133 			break;
9134 		}
9135 		if (sys.added_thread_counters++ >= MAX_ADDED_THREAD_COUNTERS) {
9136 			warnx("ignoring thread counter %s", name);
9137 			return -1;
9138 		}
9139 		break;
9140 	case SCOPE_CORE:
9141 		msrp = find_msrp_by_name(sys.cp, name);
9142 		if (msrp) {
9143 			if (debug)
9144 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
9145 			break;
9146 		}
9147 		if (sys.added_core_counters++ >= MAX_ADDED_CORE_COUNTERS) {
9148 			warnx("ignoring core counter %s", name);
9149 			return -1;
9150 		}
9151 		break;
9152 	case SCOPE_PACKAGE:
9153 		msrp = find_msrp_by_name(sys.pp, name);
9154 		if (msrp) {
9155 			if (debug)
9156 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
9157 			break;
9158 		}
9159 		if (sys.added_package_counters++ >= MAX_ADDED_PACKAGE_COUNTERS) {
9160 			warnx("ignoring package counter %s", name);
9161 			return -1;
9162 		}
9163 		break;
9164 	default:
9165 		warnx("ignoring counter %s with unknown scope", name);
9166 		return -1;
9167 	}
9168 
9169 	if (msrp == NULL) {
9170 		msrp = calloc(1, sizeof(struct msr_counter));
9171 		if (msrp == NULL)
9172 			err(-1, "calloc msr_counter");
9173 
9174 		msrp->msr_num = msr_num;
9175 		strncpy(msrp->name, name, NAME_BYTES - 1);
9176 		msrp->width = width;
9177 		msrp->type = type;
9178 		msrp->format = format;
9179 		msrp->flags = flags;
9180 
9181 		switch (scope) {
9182 		case SCOPE_CPU:
9183 			msrp->next = sys.tp;
9184 			sys.tp = msrp;
9185 			break;
9186 		case SCOPE_CORE:
9187 			msrp->next = sys.cp;
9188 			sys.cp = msrp;
9189 			break;
9190 		case SCOPE_PACKAGE:
9191 			msrp->next = sys.pp;
9192 			sys.pp = msrp;
9193 			break;
9194 		}
9195 	}
9196 
9197 	if (path) {
9198 		struct sysfs_path *sp;
9199 
9200 		sp = calloc(1, sizeof(struct sysfs_path));
9201 		if (sp == NULL) {
9202 			perror("calloc");
9203 			exit(1);
9204 		}
9205 		strncpy(sp->path, path, PATH_BYTES - 1);
9206 		sp->id = id;
9207 		sp->next = msrp->sp;
9208 		msrp->sp = sp;
9209 	}
9210 
9211 	return 0;
9212 }
9213 
9214 /*
9215  * Initialize the fields used for identifying and opening the counter.
9216  *
9217  * Defer the initialization of any runtime buffers for actually reading
9218  * the counters for when we initialize all perf counters, so we can later
9219  * easily call re_initialize().
9220  */
9221 struct perf_counter_info *make_perf_counter_info(const char *perf_device,
9222 						 const char *perf_event,
9223 						 const char *name,
9224 						 unsigned int width,
9225 						 enum counter_scope scope,
9226 						 enum counter_type type, enum counter_format format)
9227 {
9228 	struct perf_counter_info *pinfo;
9229 
9230 	pinfo = calloc(1, sizeof(*pinfo));
9231 	if (!pinfo)
9232 		errx(1, "%s: Failed to allocate %s/%s\n", __func__, perf_device, perf_event);
9233 
9234 	strncpy(pinfo->device, perf_device, ARRAY_SIZE(pinfo->device) - 1);
9235 	strncpy(pinfo->event, perf_event, ARRAY_SIZE(pinfo->event) - 1);
9236 
9237 	strncpy(pinfo->name, name, ARRAY_SIZE(pinfo->name) - 1);
9238 	pinfo->width = width;
9239 	pinfo->scope = scope;
9240 	pinfo->type = type;
9241 	pinfo->format = format;
9242 
9243 	return pinfo;
9244 }
9245 
9246 int add_perf_counter(const char *perf_device, const char *perf_event, const char *name_buffer, unsigned int width,
9247 		     enum counter_scope scope, enum counter_type type, enum counter_format format)
9248 {
9249 	struct perf_counter_info *pinfo;
9250 
9251 	switch (scope) {
9252 	case SCOPE_CPU:
9253 		if (sys.added_thread_perf_counters >= MAX_ADDED_THREAD_COUNTERS) {
9254 			warnx("ignoring thread counter perf/%s/%s", perf_device, perf_event);
9255 			return -1;
9256 		}
9257 		break;
9258 
9259 	case SCOPE_CORE:
9260 		if (sys.added_core_perf_counters >= MAX_ADDED_CORE_COUNTERS) {
9261 			warnx("ignoring core counter perf/%s/%s", perf_device, perf_event);
9262 			return -1;
9263 		}
9264 		break;
9265 
9266 	case SCOPE_PACKAGE:
9267 		if (sys.added_package_perf_counters >= MAX_ADDED_PACKAGE_COUNTERS) {
9268 			warnx("ignoring package counter perf/%s/%s", perf_device, perf_event);
9269 			return -1;
9270 		}
9271 		break;
9272 	}
9273 
9274 	pinfo = make_perf_counter_info(perf_device, perf_event, name_buffer, width, scope, type, format);
9275 
9276 	if (!pinfo)
9277 		return -1;
9278 
9279 	switch (scope) {
9280 	case SCOPE_CPU:
9281 		pinfo->next = sys.perf_tp;
9282 		sys.perf_tp = pinfo;
9283 		++sys.added_thread_perf_counters;
9284 		break;
9285 
9286 	case SCOPE_CORE:
9287 		pinfo->next = sys.perf_cp;
9288 		sys.perf_cp = pinfo;
9289 		++sys.added_core_perf_counters;
9290 		break;
9291 
9292 	case SCOPE_PACKAGE:
9293 		pinfo->next = sys.perf_pp;
9294 		sys.perf_pp = pinfo;
9295 		++sys.added_package_perf_counters;
9296 		break;
9297 	}
9298 
9299 	// FIXME: we might not have debug here yet
9300 	if (debug)
9301 		fprintf(stderr, "%s: %s/%s, name: %s, scope%d\n",
9302 			__func__, pinfo->device, pinfo->event, pinfo->name, pinfo->scope);
9303 
9304 	return 0;
9305 }
9306 
9307 void parse_add_command_msr(char *add_command)
9308 {
9309 	int msr_num = 0;
9310 	char *path = NULL;
9311 	char perf_device[PERF_DEV_NAME_BYTES] = "";
9312 	char perf_event[PERF_EVT_NAME_BYTES] = "";
9313 	char name_buffer[PERF_NAME_BYTES] = "";
9314 	int width = 64;
9315 	int fail = 0;
9316 	enum counter_scope scope = SCOPE_CPU;
9317 	enum counter_type type = COUNTER_CYCLES;
9318 	enum counter_format format = FORMAT_DELTA;
9319 
9320 	while (add_command) {
9321 
9322 		if (sscanf(add_command, "msr0x%x", &msr_num) == 1)
9323 			goto next;
9324 
9325 		if (sscanf(add_command, "msr%d", &msr_num) == 1)
9326 			goto next;
9327 
9328 		BUILD_BUG_ON(ARRAY_SIZE(perf_device) <= 31);
9329 		BUILD_BUG_ON(ARRAY_SIZE(perf_event) <= 31);
9330 		if (sscanf(add_command, "perf/%31[^/]/%31[^,]", &perf_device[0], &perf_event[0]) == 2)
9331 			goto next;
9332 
9333 		if (*add_command == '/') {
9334 			path = add_command;
9335 			goto next;
9336 		}
9337 
9338 		if (sscanf(add_command, "u%d", &width) == 1) {
9339 			if ((width == 32) || (width == 64))
9340 				goto next;
9341 			width = 64;
9342 		}
9343 		if (!strncmp(add_command, "cpu", strlen("cpu"))) {
9344 			scope = SCOPE_CPU;
9345 			goto next;
9346 		}
9347 		if (!strncmp(add_command, "core", strlen("core"))) {
9348 			scope = SCOPE_CORE;
9349 			goto next;
9350 		}
9351 		if (!strncmp(add_command, "package", strlen("package"))) {
9352 			scope = SCOPE_PACKAGE;
9353 			goto next;
9354 		}
9355 		if (!strncmp(add_command, "cycles", strlen("cycles"))) {
9356 			type = COUNTER_CYCLES;
9357 			goto next;
9358 		}
9359 		if (!strncmp(add_command, "seconds", strlen("seconds"))) {
9360 			type = COUNTER_SECONDS;
9361 			goto next;
9362 		}
9363 		if (!strncmp(add_command, "usec", strlen("usec"))) {
9364 			type = COUNTER_USEC;
9365 			goto next;
9366 		}
9367 		if (!strncmp(add_command, "raw", strlen("raw"))) {
9368 			format = FORMAT_RAW;
9369 			goto next;
9370 		}
9371 		if (!strncmp(add_command, "delta", strlen("delta"))) {
9372 			format = FORMAT_DELTA;
9373 			goto next;
9374 		}
9375 		if (!strncmp(add_command, "percent", strlen("percent"))) {
9376 			format = FORMAT_PERCENT;
9377 			goto next;
9378 		}
9379 
9380 		BUILD_BUG_ON(ARRAY_SIZE(name_buffer) <= 18);
9381 		if (sscanf(add_command, "%18s,%*s", name_buffer) == 1) {
9382 			char *eos;
9383 
9384 			eos = strchr(name_buffer, ',');
9385 			if (eos)
9386 				*eos = '\0';
9387 			goto next;
9388 		}
9389 
9390 next:
9391 		add_command = strchr(add_command, ',');
9392 		if (add_command) {
9393 			*add_command = '\0';
9394 			add_command++;
9395 		}
9396 
9397 	}
9398 	if ((msr_num == 0) && (path == NULL) && (perf_device[0] == '\0' || perf_event[0] == '\0')) {
9399 		fprintf(stderr, "--add: (msrDDD | msr0xXXX | /path_to_counter | perf/device/event ) required\n");
9400 		fail++;
9401 	}
9402 
9403 	/* Test for non-empty perf_device and perf_event */
9404 	const bool is_perf_counter = perf_device[0] && perf_event[0];
9405 
9406 	/* generate default column header */
9407 	if (*name_buffer == '\0') {
9408 		if (is_perf_counter) {
9409 			snprintf(name_buffer, ARRAY_SIZE(name_buffer), "perf/%s", perf_event);
9410 		} else {
9411 			if (width == 32)
9412 				sprintf(name_buffer, "M0x%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
9413 			else
9414 				sprintf(name_buffer, "M0X%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
9415 		}
9416 	}
9417 
9418 	if (is_perf_counter) {
9419 		if (add_perf_counter(perf_device, perf_event, name_buffer, width, scope, type, format))
9420 			fail++;
9421 	} else {
9422 		if (add_counter(msr_num, path, name_buffer, width, scope, type, format, 0, 0))
9423 			fail++;
9424 	}
9425 
9426 	if (fail) {
9427 		help();
9428 		exit(1);
9429 	}
9430 }
9431 
9432 bool starts_with(const char *str, const char *prefix)
9433 {
9434 	return strncmp(prefix, str, strlen(prefix)) == 0;
9435 }
9436 
9437 void parse_add_command_pmt(char *add_command)
9438 {
9439 	char *name = NULL;
9440 	char *type_name = NULL;
9441 	char *format_name = NULL;
9442 	unsigned int offset;
9443 	unsigned int lsb;
9444 	unsigned int msb;
9445 	unsigned int guid;
9446 	unsigned int domain_id;
9447 	enum counter_scope scope = 0;
9448 	enum pmt_datatype type = PMT_TYPE_RAW;
9449 	enum counter_format format = FORMAT_RAW;
9450 	bool has_offset = false;
9451 	bool has_lsb = false;
9452 	bool has_msb = false;
9453 	bool has_format = true;	/* Format has a default value. */
9454 	bool has_guid = false;
9455 	bool has_scope = false;
9456 	bool has_type = true;	/* Type has a default value. */
9457 
9458 	/* Consume the "pmt," prefix. */
9459 	add_command = strchr(add_command, ',');
9460 	if (!add_command) {
9461 		help();
9462 		exit(1);
9463 	}
9464 	++add_command;
9465 
9466 	while (add_command) {
9467 		if (starts_with(add_command, "name=")) {
9468 			name = add_command + strlen("name=");
9469 			goto next;
9470 		}
9471 
9472 		if (starts_with(add_command, "type=")) {
9473 			type_name = add_command + strlen("type=");
9474 			goto next;
9475 		}
9476 
9477 		if (starts_with(add_command, "domain=")) {
9478 			const size_t prefix_len = strlen("domain=");
9479 
9480 			if (sscanf(add_command + prefix_len, "cpu%u", &domain_id) == 1) {
9481 				scope = SCOPE_CPU;
9482 				has_scope = true;
9483 			} else if (sscanf(add_command + prefix_len, "core%u", &domain_id) == 1) {
9484 				scope = SCOPE_CORE;
9485 				has_scope = true;
9486 			} else if (sscanf(add_command + prefix_len, "package%u", &domain_id) == 1) {
9487 				scope = SCOPE_PACKAGE;
9488 				has_scope = true;
9489 			}
9490 
9491 			if (!has_scope) {
9492 				printf("%s: invalid value for scope. Expected cpu%%u, core%%u or package%%u.\n",
9493 				       __func__);
9494 				exit(1);
9495 			}
9496 
9497 			goto next;
9498 		}
9499 
9500 		if (starts_with(add_command, "format=")) {
9501 			format_name = add_command + strlen("format=");
9502 			goto next;
9503 		}
9504 
9505 		if (sscanf(add_command, "offset=%u", &offset) == 1) {
9506 			has_offset = true;
9507 			goto next;
9508 		}
9509 
9510 		if (sscanf(add_command, "lsb=%u", &lsb) == 1) {
9511 			has_lsb = true;
9512 			goto next;
9513 		}
9514 
9515 		if (sscanf(add_command, "msb=%u", &msb) == 1) {
9516 			has_msb = true;
9517 			goto next;
9518 		}
9519 
9520 		if (sscanf(add_command, "guid=%x", &guid) == 1) {
9521 			has_guid = true;
9522 			goto next;
9523 		}
9524 
9525 next:
9526 		add_command = strchr(add_command, ',');
9527 		if (add_command) {
9528 			*add_command = '\0';
9529 			add_command++;
9530 		}
9531 	}
9532 
9533 	if (!name) {
9534 		printf("%s: missing %s\n", __func__, "name");
9535 		exit(1);
9536 	}
9537 
9538 	if (strlen(name) >= PMT_COUNTER_NAME_SIZE_BYTES) {
9539 		printf("%s: name has to be at most %d characters long\n", __func__, PMT_COUNTER_NAME_SIZE_BYTES);
9540 		exit(1);
9541 	}
9542 
9543 	if (format_name) {
9544 		has_format = false;
9545 
9546 		if (strcmp("raw", format_name) == 0) {
9547 			format = FORMAT_RAW;
9548 			has_format = true;
9549 		}
9550 
9551 		if (strcmp("delta", format_name) == 0) {
9552 			format = FORMAT_DELTA;
9553 			has_format = true;
9554 		}
9555 
9556 		if (!has_format) {
9557 			fprintf(stderr, "%s: Invalid format %s. Expected raw or delta\n", __func__, format_name);
9558 			exit(1);
9559 		}
9560 	}
9561 
9562 	if (type_name) {
9563 		has_type = false;
9564 
9565 		if (strcmp("raw", type_name) == 0) {
9566 			type = PMT_TYPE_RAW;
9567 			has_type = true;
9568 		}
9569 
9570 		if (strcmp("txtal_time", type_name) == 0) {
9571 			type = PMT_TYPE_XTAL_TIME;
9572 			has_type = true;
9573 		}
9574 
9575 		if (!has_type) {
9576 			printf("%s: invalid %s: %s\n", __func__, "type", type_name);
9577 			exit(1);
9578 		}
9579 	}
9580 
9581 	if (!has_offset) {
9582 		printf("%s : missing %s\n", __func__, "offset");
9583 		exit(1);
9584 	}
9585 
9586 	if (!has_lsb) {
9587 		printf("%s: missing %s\n", __func__, "lsb");
9588 		exit(1);
9589 	}
9590 
9591 	if (!has_msb) {
9592 		printf("%s: missing %s\n", __func__, "msb");
9593 		exit(1);
9594 	}
9595 
9596 	if (!has_guid) {
9597 		printf("%s: missing %s\n", __func__, "guid");
9598 		exit(1);
9599 	}
9600 
9601 	if (!has_scope) {
9602 		printf("%s: missing %s\n", __func__, "scope");
9603 		exit(1);
9604 	}
9605 
9606 	if (lsb > msb) {
9607 		printf("%s: lsb > msb doesn't make sense\n", __func__);
9608 		exit(1);
9609 	}
9610 
9611 	pmt_add_counter(guid, name, type, lsb, msb, offset, scope, format, domain_id, PMT_OPEN_REQUIRED);
9612 }
9613 
9614 void parse_add_command(char *add_command)
9615 {
9616 	if (strncmp(add_command, "pmt", strlen("pmt")) == 0)
9617 		return parse_add_command_pmt(add_command);
9618 	return parse_add_command_msr(add_command);
9619 }
9620 
9621 int is_deferred_add(char *name)
9622 {
9623 	int i;
9624 
9625 	for (i = 0; i < deferred_add_index; ++i)
9626 		if (!strcmp(name, deferred_add_names[i]))
9627 			return 1;
9628 	return 0;
9629 }
9630 
9631 int is_deferred_skip(char *name)
9632 {
9633 	int i;
9634 
9635 	for (i = 0; i < deferred_skip_index; ++i)
9636 		if (!strcmp(name, deferred_skip_names[i]))
9637 			return 1;
9638 	return 0;
9639 }
9640 
9641 void probe_sysfs(void)
9642 {
9643 	char path[64];
9644 	char name_buf[16];
9645 	FILE *input;
9646 	int state;
9647 	char *sp;
9648 
9649 	for (state = 10; state >= 0; --state) {
9650 
9651 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", base_cpu, state);
9652 		input = fopen(path, "r");
9653 		if (input == NULL)
9654 			continue;
9655 		if (!fgets(name_buf, sizeof(name_buf), input))
9656 			err(1, "%s: failed to read file", path);
9657 
9658 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
9659 		sp = strchr(name_buf, '-');
9660 		if (!sp)
9661 			sp = strchrnul(name_buf, '\n');
9662 		*sp = '%';
9663 		*(sp + 1) = '\0';
9664 
9665 		remove_underbar(name_buf);
9666 
9667 		fclose(input);
9668 
9669 		sprintf(path, "cpuidle/state%d/time", state);
9670 
9671 		if (!DO_BIC(BIC_sysfs) && !is_deferred_add(name_buf))
9672 			continue;
9673 
9674 		if (is_deferred_skip(name_buf))
9675 			continue;
9676 
9677 		add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_USEC, FORMAT_PERCENT, SYSFS_PERCPU, 0);
9678 	}
9679 
9680 	for (state = 10; state >= 0; --state) {
9681 
9682 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", base_cpu, state);
9683 		input = fopen(path, "r");
9684 		if (input == NULL)
9685 			continue;
9686 		if (!fgets(name_buf, sizeof(name_buf), input))
9687 			err(1, "%s: failed to read file", path);
9688 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
9689 		sp = strchr(name_buf, '-');
9690 		if (!sp)
9691 			sp = strchrnul(name_buf, '\n');
9692 		*sp = '\0';
9693 		fclose(input);
9694 
9695 		remove_underbar(name_buf);
9696 
9697 		sprintf(path, "cpuidle/state%d/usage", state);
9698 
9699 		if (!DO_BIC(BIC_sysfs) && !is_deferred_add(name_buf))
9700 			continue;
9701 
9702 		if (is_deferred_skip(name_buf))
9703 			continue;
9704 
9705 		add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS, FORMAT_DELTA, SYSFS_PERCPU, 0);
9706 	}
9707 
9708 }
9709 
9710 /*
9711  * parse cpuset with following syntax
9712  * 1,2,4..6,8-10 and set bits in cpu_subset
9713  */
9714 void parse_cpu_command(char *optarg)
9715 {
9716 	if (!strcmp(optarg, "core")) {
9717 		if (cpu_subset)
9718 			goto error;
9719 		show_core_only++;
9720 		return;
9721 	}
9722 	if (!strcmp(optarg, "package")) {
9723 		if (cpu_subset)
9724 			goto error;
9725 		show_pkg_only++;
9726 		return;
9727 	}
9728 	if (show_core_only || show_pkg_only)
9729 		goto error;
9730 
9731 	cpu_subset = CPU_ALLOC(CPU_SUBSET_MAXCPUS);
9732 	if (cpu_subset == NULL)
9733 		err(3, "CPU_ALLOC");
9734 	cpu_subset_size = CPU_ALLOC_SIZE(CPU_SUBSET_MAXCPUS);
9735 
9736 	CPU_ZERO_S(cpu_subset_size, cpu_subset);
9737 
9738 	if (parse_cpu_str(optarg, cpu_subset, cpu_subset_size))
9739 		goto error;
9740 
9741 	return;
9742 
9743 error:
9744 	fprintf(stderr, "\"--cpu %s\" malformed\n", optarg);
9745 	help();
9746 	exit(-1);
9747 }
9748 
9749 void cmdline(int argc, char **argv)
9750 {
9751 	int opt;
9752 	int option_index = 0;
9753 	static struct option long_options[] = {
9754 		{ "add", required_argument, 0, 'a' },
9755 		{ "cpu", required_argument, 0, 'c' },
9756 		{ "Dump", no_argument, 0, 'D' },
9757 		{ "debug", no_argument, 0, 'd' },	/* internal, not documented */
9758 		{ "enable", required_argument, 0, 'e' },
9759 		{ "interval", required_argument, 0, 'i' },
9760 		{ "IPC", no_argument, 0, 'I' },
9761 		{ "num_iterations", required_argument, 0, 'n' },
9762 		{ "header_iterations", required_argument, 0, 'N' },
9763 		{ "help", no_argument, 0, 'h' },
9764 		{ "hide", required_argument, 0, 'H' },	// meh, -h taken by --help
9765 		{ "Joules", no_argument, 0, 'J' },
9766 		{ "list", no_argument, 0, 'l' },
9767 		{ "out", required_argument, 0, 'o' },
9768 		{ "quiet", no_argument, 0, 'q' },
9769 		{ "no-msr", no_argument, 0, 'M' },
9770 		{ "no-perf", no_argument, 0, 'P' },
9771 		{ "show", required_argument, 0, 's' },
9772 		{ "Summary", no_argument, 0, 'S' },
9773 		{ "TCC", required_argument, 0, 'T' },
9774 		{ "version", no_argument, 0, 'v' },
9775 		{ 0, 0, 0, 0 }
9776 	};
9777 
9778 	progname = argv[0];
9779 
9780 	/*
9781 	 * Parse some options early, because they may make other options invalid,
9782 	 * like adding the MSR counter with --add and at the same time using --no-msr.
9783 	 */
9784 	while ((opt = getopt_long_only(argc, argv, "MPn:", long_options, &option_index)) != -1) {
9785 		switch (opt) {
9786 		case 'M':
9787 			no_msr = 1;
9788 			break;
9789 		case 'P':
9790 			no_perf = 1;
9791 			break;
9792 		default:
9793 			break;
9794 		}
9795 	}
9796 	optind = 0;
9797 
9798 	while ((opt = getopt_long_only(argc, argv, "+C:c:Dde:hi:Jn:o:qMST:v", long_options, &option_index)) != -1) {
9799 		switch (opt) {
9800 		case 'a':
9801 			parse_add_command(optarg);
9802 			break;
9803 		case 'c':
9804 			parse_cpu_command(optarg);
9805 			break;
9806 		case 'D':
9807 			dump_only++;
9808 			break;
9809 		case 'e':
9810 			/* --enable specified counter */
9811 			bic_enabled = bic_enabled | bic_lookup(optarg, SHOW_LIST);
9812 			break;
9813 		case 'd':
9814 			debug++;
9815 			ENABLE_BIC(BIC_DISABLED_BY_DEFAULT);
9816 			break;
9817 		case 'H':
9818 			/*
9819 			 * --hide: do not show those specified
9820 			 *  multiple invocations simply clear more bits in enabled mask
9821 			 */
9822 			bic_enabled &= ~bic_lookup(optarg, HIDE_LIST);
9823 			break;
9824 		case 'h':
9825 		default:
9826 			help();
9827 			exit(1);
9828 		case 'i':
9829 			{
9830 				double interval = strtod(optarg, NULL);
9831 
9832 				if (interval < 0.001) {
9833 					fprintf(outf, "interval %f seconds is too small\n", interval);
9834 					exit(2);
9835 				}
9836 
9837 				interval_tv.tv_sec = interval_ts.tv_sec = interval;
9838 				interval_tv.tv_usec = (interval - interval_tv.tv_sec) * 1000000;
9839 				interval_ts.tv_nsec = (interval - interval_ts.tv_sec) * 1000000000;
9840 			}
9841 			break;
9842 		case 'J':
9843 			rapl_joules++;
9844 			break;
9845 		case 'l':
9846 			ENABLE_BIC(BIC_DISABLED_BY_DEFAULT);
9847 			list_header_only++;
9848 			quiet++;
9849 			break;
9850 		case 'o':
9851 			outf = fopen_or_die(optarg, "w");
9852 			break;
9853 		case 'q':
9854 			quiet = 1;
9855 			break;
9856 		case 'M':
9857 		case 'P':
9858 			/* Parsed earlier */
9859 			break;
9860 		case 'n':
9861 			num_iterations = strtod(optarg, NULL);
9862 
9863 			if (num_iterations <= 0) {
9864 				fprintf(outf, "iterations %d should be positive number\n", num_iterations);
9865 				exit(2);
9866 			}
9867 			break;
9868 		case 'N':
9869 			header_iterations = strtod(optarg, NULL);
9870 
9871 			if (header_iterations <= 0) {
9872 				fprintf(outf, "iterations %d should be positive number\n", header_iterations);
9873 				exit(2);
9874 			}
9875 			break;
9876 		case 's':
9877 			/*
9878 			 * --show: show only those specified
9879 			 *  The 1st invocation will clear and replace the enabled mask
9880 			 *  subsequent invocations can add to it.
9881 			 */
9882 			if (shown == 0)
9883 				bic_enabled = bic_lookup(optarg, SHOW_LIST);
9884 			else
9885 				bic_enabled |= bic_lookup(optarg, SHOW_LIST);
9886 			shown = 1;
9887 			break;
9888 		case 'S':
9889 			summary_only++;
9890 			break;
9891 		case 'T':
9892 			tj_max_override = atoi(optarg);
9893 			break;
9894 		case 'v':
9895 			print_version();
9896 			exit(0);
9897 			break;
9898 		}
9899 	}
9900 }
9901 
9902 void set_rlimit(void)
9903 {
9904 	struct rlimit limit;
9905 
9906 	if (getrlimit(RLIMIT_NOFILE, &limit) < 0)
9907 		err(1, "Failed to get rlimit");
9908 
9909 	if (limit.rlim_max < MAX_NOFILE)
9910 		limit.rlim_max = MAX_NOFILE;
9911 	if (limit.rlim_cur < MAX_NOFILE)
9912 		limit.rlim_cur = MAX_NOFILE;
9913 
9914 	if (setrlimit(RLIMIT_NOFILE, &limit) < 0)
9915 		err(1, "Failed to set rlimit");
9916 }
9917 
9918 int main(int argc, char **argv)
9919 {
9920 	int fd, ret;
9921 
9922 	fd = open("/sys/fs/cgroup/cgroup.procs", O_WRONLY);
9923 	if (fd < 0)
9924 		goto skip_cgroup_setting;
9925 
9926 	ret = write(fd, "0\n", 2);
9927 	if (ret == -1)
9928 		perror("Can't update cgroup\n");
9929 
9930 	close(fd);
9931 
9932 skip_cgroup_setting:
9933 	outf = stderr;
9934 	cmdline(argc, argv);
9935 
9936 	if (!quiet) {
9937 		print_version();
9938 		print_bootcmd();
9939 	}
9940 
9941 	probe_sysfs();
9942 
9943 	if (!getuid())
9944 		set_rlimit();
9945 
9946 	turbostat_init();
9947 
9948 	if (!no_msr)
9949 		msr_sum_record();
9950 
9951 	/* dump counters and exit */
9952 	if (dump_only)
9953 		return get_and_dump_counters();
9954 
9955 	/* list header and exit */
9956 	if (list_header_only) {
9957 		print_header(",");
9958 		flush_output_stdout();
9959 		return 0;
9960 	}
9961 
9962 	/*
9963 	 * if any params left, it must be a command to fork
9964 	 */
9965 	if (argc - optind)
9966 		return fork_it(argv + optind);
9967 	else
9968 		turbostat_loop();
9969 
9970 	return 0;
9971 }
9972