xref: /linux/tools/power/x86/turbostat/turbostat.c (revision ba6ec09911b805778a2fed6d626bfe77b011a717)
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) 2025 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 CPUID_LEAF_MODEL_ID			0x1A
35 #define CPUID_LEAF_MODEL_ID_CORE_TYPE_SHIFT	24
36 
37 #define X86_VENDOR_INTEL	0
38 
39 #include INTEL_FAMILY_HEADER
40 #include BUILD_BUG_HEADER
41 #include <stdarg.h>
42 #include <stdio.h>
43 #include <err.h>
44 #include <unistd.h>
45 #include <sys/types.h>
46 #include <sys/wait.h>
47 #include <sys/stat.h>
48 #include <sys/select.h>
49 #include <sys/resource.h>
50 #include <sys/mman.h>
51 #include <fcntl.h>
52 #include <signal.h>
53 #include <sys/time.h>
54 #include <stdlib.h>
55 #include <getopt.h>
56 #include <dirent.h>
57 #include <string.h>
58 #include <ctype.h>
59 #include <sched.h>
60 #include <time.h>
61 #include <cpuid.h>
62 #include <sys/capability.h>
63 #include <errno.h>
64 #include <math.h>
65 #include <linux/perf_event.h>
66 #include <asm/unistd.h>
67 #include <stdbool.h>
68 #include <assert.h>
69 #include <linux/kernel.h>
70 
71 #define UNUSED(x) (void)(x)
72 
73 /*
74  * This list matches the column headers, except
75  * 1. built-in only, the sysfs counters are not here -- we learn of those at run-time
76  * 2. Core and CPU are moved to the end, we can't have strings that contain them
77  *    matching on them for --show and --hide.
78  */
79 
80 /*
81  * buffer size used by sscanf() for added column names
82  * Usually truncated to 7 characters, but also handles 18 columns for raw 64-bit counters
83  */
84 #define	NAME_BYTES 20
85 #define PATH_BYTES 128
86 #define PERF_NAME_BYTES 128
87 
88 #define MAX_NOFILE 0x8000
89 
90 #define COUNTER_KIND_PERF_PREFIX "perf/"
91 #define COUNTER_KIND_PERF_PREFIX_LEN strlen(COUNTER_KIND_PERF_PREFIX)
92 #define PERF_DEV_NAME_BYTES 32
93 #define PERF_EVT_NAME_BYTES 32
94 
95 #define INTEL_ECORE_TYPE	0x20
96 #define INTEL_PCORE_TYPE	0x40
97 
98 #define ROUND_UP_TO_PAGE_SIZE(n) (((n) + 0x1000UL-1UL) & ~(0x1000UL-1UL))
99 
100 enum counter_scope { SCOPE_CPU, SCOPE_CORE, SCOPE_PACKAGE };
101 enum counter_type { COUNTER_ITEMS, COUNTER_CYCLES, COUNTER_SECONDS, COUNTER_USEC, COUNTER_K2M };
102 enum counter_format { FORMAT_RAW, FORMAT_DELTA, FORMAT_PERCENT, FORMAT_AVERAGE };
103 enum counter_source { COUNTER_SOURCE_NONE, COUNTER_SOURCE_PERF, COUNTER_SOURCE_MSR };
104 
105 struct perf_counter_info {
106 	struct perf_counter_info *next;
107 
108 	/* How to open the counter / What counter it is. */
109 	char device[PERF_DEV_NAME_BYTES];
110 	char event[PERF_EVT_NAME_BYTES];
111 
112 	/* How to show/format the counter. */
113 	char name[PERF_NAME_BYTES];
114 	unsigned int width;
115 	enum counter_scope scope;
116 	enum counter_type type;
117 	enum counter_format format;
118 	double scale;
119 
120 	/* For reading the counter. */
121 	int *fd_perf_per_domain;
122 	size_t num_domains;
123 };
124 
125 struct sysfs_path {
126 	char path[PATH_BYTES];
127 	int id;
128 	struct sysfs_path *next;
129 };
130 
131 struct msr_counter {
132 	unsigned int msr_num;
133 	char name[NAME_BYTES];
134 	struct sysfs_path *sp;
135 	unsigned int width;
136 	enum counter_type type;
137 	enum counter_format format;
138 	struct msr_counter *next;
139 	unsigned int flags;
140 #define	FLAGS_HIDE	(1 << 0)
141 #define	FLAGS_SHOW	(1 << 1)
142 #define	SYSFS_PERCPU	(1 << 1)
143 };
144 
145 struct msr_counter bic[] = {
146 	{ 0x0, "usec", NULL, 0, 0, 0, NULL, 0 },
147 	{ 0x0, "Time_Of_Day_Seconds", NULL, 0, 0, 0, NULL, 0 },
148 	{ 0x0, "Package", NULL, 0, 0, 0, NULL, 0 },
149 	{ 0x0, "Node", NULL, 0, 0, 0, NULL, 0 },
150 	{ 0x0, "Avg_MHz", NULL, 0, 0, 0, NULL, 0 },
151 	{ 0x0, "Busy%", NULL, 0, 0, 0, NULL, 0 },
152 	{ 0x0, "Bzy_MHz", NULL, 0, 0, 0, NULL, 0 },
153 	{ 0x0, "TSC_MHz", NULL, 0, 0, 0, NULL, 0 },
154 	{ 0x0, "IRQ", NULL, 0, 0, 0, NULL, 0 },
155 	{ 0x0, "SMI", NULL, 32, 0, FORMAT_DELTA, NULL, 0 },
156 	{ 0x0, "sysfs", NULL, 0, 0, 0, NULL, 0 },
157 	{ 0x0, "CPU%c1", NULL, 0, 0, 0, NULL, 0 },
158 	{ 0x0, "CPU%c3", NULL, 0, 0, 0, NULL, 0 },
159 	{ 0x0, "CPU%c6", NULL, 0, 0, 0, NULL, 0 },
160 	{ 0x0, "CPU%c7", NULL, 0, 0, 0, NULL, 0 },
161 	{ 0x0, "ThreadC", NULL, 0, 0, 0, NULL, 0 },
162 	{ 0x0, "CoreTmp", NULL, 0, 0, 0, NULL, 0 },
163 	{ 0x0, "CoreCnt", NULL, 0, 0, 0, NULL, 0 },
164 	{ 0x0, "PkgTmp", NULL, 0, 0, 0, NULL, 0 },
165 	{ 0x0, "GFX%rc6", NULL, 0, 0, 0, NULL, 0 },
166 	{ 0x0, "GFXMHz", NULL, 0, 0, 0, NULL, 0 },
167 	{ 0x0, "Pkg%pc2", NULL, 0, 0, 0, NULL, 0 },
168 	{ 0x0, "Pkg%pc3", NULL, 0, 0, 0, NULL, 0 },
169 	{ 0x0, "Pkg%pc6", NULL, 0, 0, 0, NULL, 0 },
170 	{ 0x0, "Pkg%pc7", NULL, 0, 0, 0, NULL, 0 },
171 	{ 0x0, "Pkg%pc8", NULL, 0, 0, 0, NULL, 0 },
172 	{ 0x0, "Pkg%pc9", NULL, 0, 0, 0, NULL, 0 },
173 	{ 0x0, "Pk%pc10", NULL, 0, 0, 0, NULL, 0 },
174 	{ 0x0, "CPU%LPI", NULL, 0, 0, 0, NULL, 0 },
175 	{ 0x0, "SYS%LPI", NULL, 0, 0, 0, NULL, 0 },
176 	{ 0x0, "PkgWatt", NULL, 0, 0, 0, NULL, 0 },
177 	{ 0x0, "CorWatt", NULL, 0, 0, 0, NULL, 0 },
178 	{ 0x0, "GFXWatt", NULL, 0, 0, 0, NULL, 0 },
179 	{ 0x0, "PkgCnt", NULL, 0, 0, 0, NULL, 0 },
180 	{ 0x0, "RAMWatt", NULL, 0, 0, 0, NULL, 0 },
181 	{ 0x0, "PKG_%", NULL, 0, 0, 0, NULL, 0 },
182 	{ 0x0, "RAM_%", NULL, 0, 0, 0, NULL, 0 },
183 	{ 0x0, "Pkg_J", NULL, 0, 0, 0, NULL, 0 },
184 	{ 0x0, "Cor_J", NULL, 0, 0, 0, NULL, 0 },
185 	{ 0x0, "GFX_J", NULL, 0, 0, 0, NULL, 0 },
186 	{ 0x0, "RAM_J", NULL, 0, 0, 0, NULL, 0 },
187 	{ 0x0, "Mod%c6", NULL, 0, 0, 0, NULL, 0 },
188 	{ 0x0, "Totl%C0", NULL, 0, 0, 0, NULL, 0 },
189 	{ 0x0, "Any%C0", NULL, 0, 0, 0, NULL, 0 },
190 	{ 0x0, "GFX%C0", NULL, 0, 0, 0, NULL, 0 },
191 	{ 0x0, "CPUGFX%", NULL, 0, 0, 0, NULL, 0 },
192 	{ 0x0, "Core", NULL, 0, 0, 0, NULL, 0 },
193 	{ 0x0, "CPU", NULL, 0, 0, 0, NULL, 0 },
194 	{ 0x0, "APIC", NULL, 0, 0, 0, NULL, 0 },
195 	{ 0x0, "X2APIC", NULL, 0, 0, 0, NULL, 0 },
196 	{ 0x0, "Die", NULL, 0, 0, 0, NULL, 0 },
197 	{ 0x0, "GFXAMHz", NULL, 0, 0, 0, NULL, 0 },
198 	{ 0x0, "IPC", NULL, 0, 0, 0, NULL, 0 },
199 	{ 0x0, "CoreThr", NULL, 0, 0, 0, NULL, 0 },
200 	{ 0x0, "UncMHz", NULL, 0, 0, 0, NULL, 0 },
201 	{ 0x0, "SAM%mc6", NULL, 0, 0, 0, NULL, 0 },
202 	{ 0x0, "SAMMHz", NULL, 0, 0, 0, NULL, 0 },
203 	{ 0x0, "SAMAMHz", NULL, 0, 0, 0, NULL, 0 },
204 	{ 0x0, "Die%c6", NULL, 0, 0, 0, NULL, 0 },
205 	{ 0x0, "SysWatt", NULL, 0, 0, 0, NULL, 0 },
206 	{ 0x0, "Sys_J", NULL, 0, 0, 0, NULL, 0 },
207 	{ 0x0, "NMI", NULL, 0, 0, 0, NULL, 0 },
208 	{ 0x0, "CPU%c1e", NULL, 0, 0, 0, NULL, 0 },
209 };
210 
211 #define MAX_BIC (sizeof(bic) / sizeof(struct msr_counter))
212 #define	BIC_USEC	(1ULL << 0)
213 #define	BIC_TOD		(1ULL << 1)
214 #define	BIC_Package	(1ULL << 2)
215 #define	BIC_Node	(1ULL << 3)
216 #define	BIC_Avg_MHz	(1ULL << 4)
217 #define	BIC_Busy	(1ULL << 5)
218 #define	BIC_Bzy_MHz	(1ULL << 6)
219 #define	BIC_TSC_MHz	(1ULL << 7)
220 #define	BIC_IRQ		(1ULL << 8)
221 #define	BIC_SMI		(1ULL << 9)
222 #define	BIC_sysfs	(1ULL << 10)
223 #define	BIC_CPU_c1	(1ULL << 11)
224 #define	BIC_CPU_c3	(1ULL << 12)
225 #define	BIC_CPU_c6	(1ULL << 13)
226 #define	BIC_CPU_c7	(1ULL << 14)
227 #define	BIC_ThreadC	(1ULL << 15)
228 #define	BIC_CoreTmp	(1ULL << 16)
229 #define	BIC_CoreCnt	(1ULL << 17)
230 #define	BIC_PkgTmp	(1ULL << 18)
231 #define	BIC_GFX_rc6	(1ULL << 19)
232 #define	BIC_GFXMHz	(1ULL << 20)
233 #define	BIC_Pkgpc2	(1ULL << 21)
234 #define	BIC_Pkgpc3	(1ULL << 22)
235 #define	BIC_Pkgpc6	(1ULL << 23)
236 #define	BIC_Pkgpc7	(1ULL << 24)
237 #define	BIC_Pkgpc8	(1ULL << 25)
238 #define	BIC_Pkgpc9	(1ULL << 26)
239 #define	BIC_Pkgpc10	(1ULL << 27)
240 #define BIC_CPU_LPI	(1ULL << 28)
241 #define BIC_SYS_LPI	(1ULL << 29)
242 #define	BIC_PkgWatt	(1ULL << 30)
243 #define	BIC_CorWatt	(1ULL << 31)
244 #define	BIC_GFXWatt	(1ULL << 32)
245 #define	BIC_PkgCnt	(1ULL << 33)
246 #define	BIC_RAMWatt	(1ULL << 34)
247 #define	BIC_PKG__	(1ULL << 35)
248 #define	BIC_RAM__	(1ULL << 36)
249 #define	BIC_Pkg_J	(1ULL << 37)
250 #define	BIC_Cor_J	(1ULL << 38)
251 #define	BIC_GFX_J	(1ULL << 39)
252 #define	BIC_RAM_J	(1ULL << 40)
253 #define	BIC_Mod_c6	(1ULL << 41)
254 #define	BIC_Totl_c0	(1ULL << 42)
255 #define	BIC_Any_c0	(1ULL << 43)
256 #define	BIC_GFX_c0	(1ULL << 44)
257 #define	BIC_CPUGFX	(1ULL << 45)
258 #define	BIC_Core	(1ULL << 46)
259 #define	BIC_CPU		(1ULL << 47)
260 #define	BIC_APIC	(1ULL << 48)
261 #define	BIC_X2APIC	(1ULL << 49)
262 #define	BIC_Die		(1ULL << 50)
263 #define	BIC_GFXACTMHz	(1ULL << 51)
264 #define	BIC_IPC		(1ULL << 52)
265 #define	BIC_CORE_THROT_CNT	(1ULL << 53)
266 #define	BIC_UNCORE_MHZ		(1ULL << 54)
267 #define	BIC_SAM_mc6		(1ULL << 55)
268 #define	BIC_SAMMHz		(1ULL << 56)
269 #define	BIC_SAMACTMHz		(1ULL << 57)
270 #define	BIC_Diec6		(1ULL << 58)
271 #define	BIC_SysWatt		(1ULL << 59)
272 #define	BIC_Sys_J		(1ULL << 60)
273 #define	BIC_NMI			(1ULL << 61)
274 #define	BIC_CPU_c1e		(1ULL << 62)
275 
276 #define BIC_TOPOLOGY (BIC_Package | BIC_Node | BIC_CoreCnt | BIC_PkgCnt | BIC_Core | BIC_CPU | BIC_Die)
277 #define BIC_THERMAL_PWR (BIC_CoreTmp | BIC_PkgTmp | BIC_PkgWatt | BIC_CorWatt | BIC_GFXWatt | BIC_RAMWatt | BIC_PKG__ | BIC_RAM__ | BIC_SysWatt)
278 #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)
279 #define BIC_IDLE (BIC_Busy | 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)
280 #define BIC_OTHER (BIC_IRQ | BIC_NMI | BIC_SMI | BIC_ThreadC | BIC_CoreTmp | BIC_IPC)
281 
282 #define BIC_DISABLED_BY_DEFAULT	(BIC_USEC | BIC_TOD | BIC_APIC | BIC_X2APIC)
283 
284 unsigned long long bic_enabled = (0xFFFFFFFFFFFFFFFFULL & ~BIC_DISABLED_BY_DEFAULT);
285 unsigned long long bic_present = BIC_USEC | BIC_TOD | BIC_sysfs | BIC_APIC | BIC_X2APIC;
286 
287 #define DO_BIC(COUNTER_NAME) (bic_enabled & bic_present & COUNTER_NAME)
288 #define DO_BIC_READ(COUNTER_NAME) (bic_present & COUNTER_NAME)
289 #define ENABLE_BIC(COUNTER_NAME) (bic_enabled |= COUNTER_NAME)
290 #define BIC_PRESENT(COUNTER_BIT) (bic_present |= COUNTER_BIT)
291 #define BIC_NOT_PRESENT(COUNTER_BIT) (bic_present &= ~COUNTER_BIT)
292 #define BIC_IS_ENABLED(COUNTER_BIT) (bic_enabled & COUNTER_BIT)
293 
294 /*
295  * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
296  * If you change the values, note they are used both in comparisons
297  * (>= PCL__7) and to index pkg_cstate_limit_strings[].
298  */
299 #define PCLUKN 0		/* Unknown */
300 #define PCLRSV 1		/* Reserved */
301 #define PCL__0 2		/* PC0 */
302 #define PCL__1 3		/* PC1 */
303 #define PCL__2 4		/* PC2 */
304 #define PCL__3 5		/* PC3 */
305 #define PCL__4 6		/* PC4 */
306 #define PCL__6 7		/* PC6 */
307 #define PCL_6N 8		/* PC6 No Retention */
308 #define PCL_6R 9		/* PC6 Retention */
309 #define PCL__7 10		/* PC7 */
310 #define PCL_7S 11		/* PC7 Shrink */
311 #define PCL__8 12		/* PC8 */
312 #define PCL__9 13		/* PC9 */
313 #define PCL_10 14		/* PC10 */
314 #define PCLUNL 15		/* Unlimited */
315 
316 struct amperf_group_fd;
317 
318 char *proc_stat = "/proc/stat";
319 FILE *outf;
320 int *fd_percpu;
321 int *fd_instr_count_percpu;
322 struct timeval interval_tv = { 5, 0 };
323 struct timespec interval_ts = { 5, 0 };
324 
325 unsigned int num_iterations;
326 unsigned int header_iterations;
327 unsigned int debug;
328 unsigned int quiet;
329 unsigned int shown;
330 unsigned int sums_need_wide_columns;
331 unsigned int rapl_joules;
332 unsigned int summary_only;
333 unsigned int list_header_only;
334 unsigned int dump_only;
335 unsigned int force_load;
336 unsigned int has_aperf;
337 unsigned int has_aperf_access;
338 unsigned int has_epb;
339 unsigned int has_turbo;
340 unsigned int is_hybrid;
341 unsigned int units = 1000000;	/* MHz etc */
342 unsigned int genuine_intel;
343 unsigned int authentic_amd;
344 unsigned int hygon_genuine;
345 unsigned int max_level, max_extended_level;
346 unsigned int has_invariant_tsc;
347 unsigned int aperf_mperf_multiplier = 1;
348 double bclk;
349 double base_hz;
350 unsigned int has_base_hz;
351 double tsc_tweak = 1.0;
352 unsigned int show_pkg_only;
353 unsigned int show_core_only;
354 char *output_buffer, *outp;
355 unsigned int do_dts;
356 unsigned int do_ptm;
357 unsigned int do_ipc;
358 unsigned long long cpuidle_cur_cpu_lpi_us;
359 unsigned long long cpuidle_cur_sys_lpi_us;
360 unsigned int tj_max;
361 unsigned int tj_max_override;
362 double rapl_power_units, rapl_time_units;
363 double rapl_dram_energy_units, rapl_energy_units, rapl_psys_energy_units;
364 double rapl_joule_counter_range;
365 unsigned int crystal_hz;
366 unsigned long long tsc_hz;
367 int base_cpu;
368 unsigned int has_hwp;		/* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
369 			/* IA32_HWP_REQUEST, IA32_HWP_STATUS */
370 unsigned int has_hwp_notify;	/* IA32_HWP_INTERRUPT */
371 unsigned int has_hwp_activity_window;	/* IA32_HWP_REQUEST[bits 41:32] */
372 unsigned int has_hwp_epp;	/* IA32_HWP_REQUEST[bits 31:24] */
373 unsigned int has_hwp_pkg;	/* IA32_HWP_REQUEST_PKG */
374 unsigned int first_counter_read = 1;
375 
376 static struct timeval procsysfs_tv_begin;
377 
378 int ignore_stdin;
379 bool no_msr;
380 bool no_perf;
381 
382 enum gfx_sysfs_idx {
383 	GFX_rc6,
384 	GFX_MHz,
385 	GFX_ACTMHz,
386 	SAM_mc6,
387 	SAM_MHz,
388 	SAM_ACTMHz,
389 	GFX_MAX
390 };
391 
392 struct gfx_sysfs_info {
393 	FILE *fp;
394 	unsigned int val;
395 	unsigned long long val_ull;
396 };
397 
398 static struct gfx_sysfs_info gfx_info[GFX_MAX];
399 
400 int get_msr(int cpu, off_t offset, unsigned long long *msr);
401 int add_counter(unsigned int msr_num, char *path, char *name,
402 		unsigned int width, enum counter_scope scope,
403 		enum counter_type type, enum counter_format format, int flags, int package_num);
404 
405 /* Model specific support Start */
406 
407 /* List of features that may diverge among different platforms */
408 struct platform_features {
409 	bool has_msr_misc_feature_control;	/* MSR_MISC_FEATURE_CONTROL */
410 	bool has_msr_misc_pwr_mgmt;	/* MSR_MISC_PWR_MGMT */
411 	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 */
412 	bool has_config_tdp;	/* MSR_CONFIG_TDP_NOMINAL/LEVEL_1/LEVEL_2/CONTROL, MSR_TURBO_ACTIVATION_RATIO */
413 	int bclk_freq;		/* CPU base clock */
414 	int crystal_freq;	/* Crystal clock to use when not available from CPUID.15 */
415 	int supported_cstates;	/* Core cstates and Package cstates supported */
416 	int cst_limit;		/* MSR_PKG_CST_CONFIG_CONTROL */
417 	bool has_cst_auto_convension;	/* AUTOMATIC_CSTATE_CONVERSION bit in MSR_PKG_CST_CONFIG_CONTROL */
418 	bool has_irtl_msrs;	/* MSR_PKGC3/PKGC6/PKGC7/PKGC8/PKGC9/PKGC10_IRTL */
419 	bool has_msr_core_c1_res;	/* MSR_CORE_C1_RES */
420 	bool has_msr_module_c6_res_ms;	/* MSR_MODULE_C6_RES_MS */
421 	bool has_msr_c6_demotion_policy_config;	/* MSR_CC6_DEMOTION_POLICY_CONFIG/MSR_MC6_DEMOTION_POLICY_CONFIG */
422 	bool has_msr_atom_pkg_c6_residency;	/* MSR_ATOM_PKG_C6_RESIDENCY */
423 	bool has_msr_knl_core_c6_residency;	/* MSR_KNL_CORE_C6_RESIDENCY */
424 	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 */
425 	bool has_cst_prewake_bit;	/* Cstate prewake bit in MSR_IA32_POWER_CTL */
426 	int trl_msrs;		/* MSR_TURBO_RATIO_LIMIT/LIMIT1/LIMIT2/SECONDARY, Atom TRL MSRs */
427 	int plr_msrs;		/* MSR_CORE/GFX/RING_PERF_LIMIT_REASONS */
428 	int rapl_msrs;		/* RAPL PKG/DRAM/CORE/GFX MSRs, AMD RAPL MSRs */
429 	bool has_per_core_rapl;	/* Indicates cores energy collection is per-core, not per-package. AMD specific for now */
430 	bool has_rapl_divisor;	/* Divisor for Energy unit raw value from MSR_RAPL_POWER_UNIT */
431 	bool has_fixed_rapl_unit;	/* Fixed Energy Unit used for DRAM RAPL Domain */
432 	bool has_fixed_rapl_psys_unit;	/* Fixed Energy Unit used for PSYS RAPL Domain */
433 	int rapl_quirk_tdp;	/* Hardcoded TDP value when cannot be retrieved from hardware */
434 	int tcc_offset_bits;	/* TCC Offset bits in MSR_IA32_TEMPERATURE_TARGET */
435 	bool enable_tsc_tweak;	/* Use CPU Base freq instead of TSC freq for aperf/mperf counter */
436 	bool need_perf_multiplier;	/* mperf/aperf multiplier */
437 };
438 
439 struct platform_data {
440 	unsigned int vfm;
441 	const struct platform_features *features;
442 };
443 
444 /* For BCLK */
445 enum bclk_freq {
446 	BCLK_100MHZ = 1,
447 	BCLK_133MHZ,
448 	BCLK_SLV,
449 };
450 
451 #define SLM_BCLK_FREQS 5
452 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0 };
453 
slm_bclk(void)454 double slm_bclk(void)
455 {
456 	unsigned long long msr = 3;
457 	unsigned int i;
458 	double freq;
459 
460 	if (get_msr(base_cpu, MSR_FSB_FREQ, &msr))
461 		fprintf(outf, "SLM BCLK: unknown\n");
462 
463 	i = msr & 0xf;
464 	if (i >= SLM_BCLK_FREQS) {
465 		fprintf(outf, "SLM BCLK[%d] invalid\n", i);
466 		i = 3;
467 	}
468 	freq = slm_freq_table[i];
469 
470 	if (!quiet)
471 		fprintf(outf, "SLM BCLK: %.1f Mhz\n", freq);
472 
473 	return freq;
474 }
475 
476 /* For Package cstate limit */
477 enum package_cstate_limit {
478 	CST_LIMIT_NHM = 1,
479 	CST_LIMIT_SNB,
480 	CST_LIMIT_HSW,
481 	CST_LIMIT_SKX,
482 	CST_LIMIT_ICX,
483 	CST_LIMIT_SLV,
484 	CST_LIMIT_AMT,
485 	CST_LIMIT_KNL,
486 	CST_LIMIT_GMT,
487 };
488 
489 /* For Turbo Ratio Limit MSRs */
490 enum turbo_ratio_limit_msrs {
491 	TRL_BASE = BIT(0),
492 	TRL_LIMIT1 = BIT(1),
493 	TRL_LIMIT2 = BIT(2),
494 	TRL_ATOM = BIT(3),
495 	TRL_KNL = BIT(4),
496 	TRL_CORECOUNT = BIT(5),
497 };
498 
499 /* For Perf Limit Reason MSRs */
500 enum perf_limit_reason_msrs {
501 	PLR_CORE = BIT(0),
502 	PLR_GFX = BIT(1),
503 	PLR_RING = BIT(2),
504 };
505 
506 /* For RAPL MSRs */
507 enum rapl_msrs {
508 	RAPL_PKG_POWER_LIMIT = BIT(0),	/* 0x610 MSR_PKG_POWER_LIMIT */
509 	RAPL_PKG_ENERGY_STATUS = BIT(1),	/* 0x611 MSR_PKG_ENERGY_STATUS */
510 	RAPL_PKG_PERF_STATUS = BIT(2),	/* 0x613 MSR_PKG_PERF_STATUS */
511 	RAPL_PKG_POWER_INFO = BIT(3),	/* 0x614 MSR_PKG_POWER_INFO */
512 	RAPL_DRAM_POWER_LIMIT = BIT(4),	/* 0x618 MSR_DRAM_POWER_LIMIT */
513 	RAPL_DRAM_ENERGY_STATUS = BIT(5),	/* 0x619 MSR_DRAM_ENERGY_STATUS */
514 	RAPL_DRAM_PERF_STATUS = BIT(6),	/* 0x61b MSR_DRAM_PERF_STATUS */
515 	RAPL_DRAM_POWER_INFO = BIT(7),	/* 0x61c MSR_DRAM_POWER_INFO */
516 	RAPL_CORE_POWER_LIMIT = BIT(8),	/* 0x638 MSR_PP0_POWER_LIMIT */
517 	RAPL_CORE_ENERGY_STATUS = BIT(9),	/* 0x639 MSR_PP0_ENERGY_STATUS */
518 	RAPL_CORE_POLICY = BIT(10),	/* 0x63a MSR_PP0_POLICY */
519 	RAPL_GFX_POWER_LIMIT = BIT(11),	/* 0x640 MSR_PP1_POWER_LIMIT */
520 	RAPL_GFX_ENERGY_STATUS = BIT(12),	/* 0x641 MSR_PP1_ENERGY_STATUS */
521 	RAPL_GFX_POLICY = BIT(13),	/* 0x642 MSR_PP1_POLICY */
522 	RAPL_AMD_PWR_UNIT = BIT(14),	/* 0xc0010299 MSR_AMD_RAPL_POWER_UNIT */
523 	RAPL_AMD_CORE_ENERGY_STAT = BIT(15),	/* 0xc001029a MSR_AMD_CORE_ENERGY_STATUS */
524 	RAPL_AMD_PKG_ENERGY_STAT = BIT(16),	/* 0xc001029b MSR_AMD_PKG_ENERGY_STATUS */
525 	RAPL_PLATFORM_ENERGY_LIMIT = BIT(17),	/* 0x64c MSR_PLATFORM_ENERGY_LIMIT */
526 	RAPL_PLATFORM_ENERGY_STATUS = BIT(18),	/* 0x64d MSR_PLATFORM_ENERGY_STATUS */
527 };
528 
529 #define RAPL_PKG	(RAPL_PKG_ENERGY_STATUS | RAPL_PKG_POWER_LIMIT)
530 #define RAPL_DRAM	(RAPL_DRAM_ENERGY_STATUS | RAPL_DRAM_POWER_LIMIT)
531 #define RAPL_CORE	(RAPL_CORE_ENERGY_STATUS | RAPL_CORE_POWER_LIMIT)
532 #define RAPL_GFX	(RAPL_GFX_POWER_LIMIT | RAPL_GFX_ENERGY_STATUS)
533 #define RAPL_PSYS	(RAPL_PLATFORM_ENERGY_STATUS | RAPL_PLATFORM_ENERGY_LIMIT)
534 
535 #define RAPL_PKG_ALL	(RAPL_PKG | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO)
536 #define RAPL_DRAM_ALL	(RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_DRAM_POWER_INFO)
537 #define RAPL_CORE_ALL	(RAPL_CORE | RAPL_CORE_POLICY)
538 #define RAPL_GFX_ALL	(RAPL_GFX | RAPL_GFX_POLIGY)
539 
540 #define RAPL_AMD_F17H	(RAPL_AMD_PWR_UNIT | RAPL_AMD_CORE_ENERGY_STAT | RAPL_AMD_PKG_ENERGY_STAT)
541 
542 /* For Cstates */
543 enum cstates {
544 	CC1 = BIT(0),
545 	CC3 = BIT(1),
546 	CC6 = BIT(2),
547 	CC7 = BIT(3),
548 	PC2 = BIT(4),
549 	PC3 = BIT(5),
550 	PC6 = BIT(6),
551 	PC7 = BIT(7),
552 	PC8 = BIT(8),
553 	PC9 = BIT(9),
554 	PC10 = BIT(10),
555 };
556 
557 static const struct platform_features nhm_features = {
558 	.has_msr_misc_pwr_mgmt = 1,
559 	.has_nhm_msrs = 1,
560 	.bclk_freq = BCLK_133MHZ,
561 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
562 	.cst_limit = CST_LIMIT_NHM,
563 	.trl_msrs = TRL_BASE,
564 };
565 
566 static const struct platform_features nhx_features = {
567 	.has_msr_misc_pwr_mgmt = 1,
568 	.has_nhm_msrs = 1,
569 	.bclk_freq = BCLK_133MHZ,
570 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
571 	.cst_limit = CST_LIMIT_NHM,
572 };
573 
574 static const struct platform_features snb_features = {
575 	.has_msr_misc_feature_control = 1,
576 	.has_msr_misc_pwr_mgmt = 1,
577 	.has_nhm_msrs = 1,
578 	.bclk_freq = BCLK_100MHZ,
579 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
580 	.cst_limit = CST_LIMIT_SNB,
581 	.has_irtl_msrs = 1,
582 	.trl_msrs = TRL_BASE,
583 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
584 };
585 
586 static const struct platform_features snx_features = {
587 	.has_msr_misc_feature_control = 1,
588 	.has_msr_misc_pwr_mgmt = 1,
589 	.has_nhm_msrs = 1,
590 	.bclk_freq = BCLK_100MHZ,
591 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
592 	.cst_limit = CST_LIMIT_SNB,
593 	.has_irtl_msrs = 1,
594 	.trl_msrs = TRL_BASE,
595 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM_ALL,
596 };
597 
598 static const struct platform_features ivb_features = {
599 	.has_msr_misc_feature_control = 1,
600 	.has_msr_misc_pwr_mgmt = 1,
601 	.has_nhm_msrs = 1,
602 	.has_config_tdp = 1,
603 	.bclk_freq = BCLK_100MHZ,
604 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
605 	.cst_limit = CST_LIMIT_SNB,
606 	.has_irtl_msrs = 1,
607 	.trl_msrs = TRL_BASE,
608 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
609 };
610 
611 static const struct platform_features ivx_features = {
612 	.has_msr_misc_feature_control = 1,
613 	.has_msr_misc_pwr_mgmt = 1,
614 	.has_nhm_msrs = 1,
615 	.bclk_freq = BCLK_100MHZ,
616 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
617 	.cst_limit = CST_LIMIT_SNB,
618 	.has_irtl_msrs = 1,
619 	.trl_msrs = TRL_BASE | TRL_LIMIT1,
620 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM_ALL,
621 };
622 
623 static const struct platform_features hsw_features = {
624 	.has_msr_misc_feature_control = 1,
625 	.has_msr_misc_pwr_mgmt = 1,
626 	.has_nhm_msrs = 1,
627 	.has_config_tdp = 1,
628 	.bclk_freq = BCLK_100MHZ,
629 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
630 	.cst_limit = CST_LIMIT_HSW,
631 	.has_irtl_msrs = 1,
632 	.trl_msrs = TRL_BASE,
633 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
634 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
635 };
636 
637 static const struct platform_features hsx_features = {
638 	.has_msr_misc_feature_control = 1,
639 	.has_msr_misc_pwr_mgmt = 1,
640 	.has_nhm_msrs = 1,
641 	.has_config_tdp = 1,
642 	.bclk_freq = BCLK_100MHZ,
643 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
644 	.cst_limit = CST_LIMIT_HSW,
645 	.has_irtl_msrs = 1,
646 	.trl_msrs = TRL_BASE | TRL_LIMIT1 | TRL_LIMIT2,
647 	.plr_msrs = PLR_CORE | PLR_RING,
648 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
649 	.has_fixed_rapl_unit = 1,
650 };
651 
652 static const struct platform_features hswl_features = {
653 	.has_msr_misc_feature_control = 1,
654 	.has_msr_misc_pwr_mgmt = 1,
655 	.has_nhm_msrs = 1,
656 	.has_config_tdp = 1,
657 	.bclk_freq = BCLK_100MHZ,
658 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
659 	.cst_limit = CST_LIMIT_HSW,
660 	.has_irtl_msrs = 1,
661 	.trl_msrs = TRL_BASE,
662 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
663 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
664 };
665 
666 static const struct platform_features hswg_features = {
667 	.has_msr_misc_feature_control = 1,
668 	.has_msr_misc_pwr_mgmt = 1,
669 	.has_nhm_msrs = 1,
670 	.has_config_tdp = 1,
671 	.bclk_freq = BCLK_100MHZ,
672 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
673 	.cst_limit = CST_LIMIT_HSW,
674 	.has_irtl_msrs = 1,
675 	.trl_msrs = TRL_BASE,
676 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
677 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
678 };
679 
680 static const struct platform_features bdw_features = {
681 	.has_msr_misc_feature_control = 1,
682 	.has_msr_misc_pwr_mgmt = 1,
683 	.has_nhm_msrs = 1,
684 	.has_config_tdp = 1,
685 	.bclk_freq = BCLK_100MHZ,
686 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
687 	.cst_limit = CST_LIMIT_HSW,
688 	.has_irtl_msrs = 1,
689 	.trl_msrs = TRL_BASE,
690 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
691 };
692 
693 static const struct platform_features bdwg_features = {
694 	.has_msr_misc_feature_control = 1,
695 	.has_msr_misc_pwr_mgmt = 1,
696 	.has_nhm_msrs = 1,
697 	.has_config_tdp = 1,
698 	.bclk_freq = BCLK_100MHZ,
699 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
700 	.cst_limit = CST_LIMIT_HSW,
701 	.has_irtl_msrs = 1,
702 	.trl_msrs = TRL_BASE,
703 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
704 };
705 
706 static const struct platform_features bdx_features = {
707 	.has_msr_misc_feature_control = 1,
708 	.has_msr_misc_pwr_mgmt = 1,
709 	.has_nhm_msrs = 1,
710 	.has_config_tdp = 1,
711 	.bclk_freq = BCLK_100MHZ,
712 	.supported_cstates = CC1 | CC3 | CC6 | PC2 | PC3 | PC6,
713 	.cst_limit = CST_LIMIT_HSW,
714 	.has_irtl_msrs = 1,
715 	.has_cst_auto_convension = 1,
716 	.trl_msrs = TRL_BASE,
717 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
718 	.has_fixed_rapl_unit = 1,
719 };
720 
721 static const struct platform_features skl_features = {
722 	.has_msr_misc_feature_control = 1,
723 	.has_msr_misc_pwr_mgmt = 1,
724 	.has_nhm_msrs = 1,
725 	.has_config_tdp = 1,
726 	.bclk_freq = BCLK_100MHZ,
727 	.crystal_freq = 24000000,
728 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
729 	.cst_limit = CST_LIMIT_HSW,
730 	.has_irtl_msrs = 1,
731 	.has_ext_cst_msrs = 1,
732 	.trl_msrs = TRL_BASE,
733 	.tcc_offset_bits = 6,
734 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX | RAPL_PSYS,
735 	.enable_tsc_tweak = 1,
736 };
737 
738 static const struct platform_features cnl_features = {
739 	.has_msr_misc_feature_control = 1,
740 	.has_msr_misc_pwr_mgmt = 1,
741 	.has_nhm_msrs = 1,
742 	.has_config_tdp = 1,
743 	.bclk_freq = BCLK_100MHZ,
744 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
745 	.cst_limit = CST_LIMIT_HSW,
746 	.has_irtl_msrs = 1,
747 	.has_msr_core_c1_res = 1,
748 	.has_ext_cst_msrs = 1,
749 	.trl_msrs = TRL_BASE,
750 	.tcc_offset_bits = 6,
751 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX | RAPL_PSYS,
752 	.enable_tsc_tweak = 1,
753 };
754 
755 /* Copied from cnl_features, with PC7/PC9 removed */
756 static const struct platform_features adl_features = {
757 	.has_msr_misc_feature_control	= cnl_features.has_msr_misc_feature_control,
758 	.has_msr_misc_pwr_mgmt		= cnl_features.has_msr_misc_pwr_mgmt,
759 	.has_nhm_msrs			= cnl_features.has_nhm_msrs,
760 	.has_config_tdp			= cnl_features.has_config_tdp,
761 	.bclk_freq			= cnl_features.bclk_freq,
762 	.supported_cstates		= CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC8 | PC10,
763 	.cst_limit			= cnl_features.cst_limit,
764 	.has_irtl_msrs			= cnl_features.has_irtl_msrs,
765 	.has_msr_core_c1_res		= cnl_features.has_msr_core_c1_res,
766 	.has_ext_cst_msrs		= cnl_features.has_ext_cst_msrs,
767 	.trl_msrs			= cnl_features.trl_msrs,
768 	.tcc_offset_bits		= cnl_features.tcc_offset_bits,
769 	.rapl_msrs			= cnl_features.rapl_msrs,
770 	.enable_tsc_tweak		= cnl_features.enable_tsc_tweak,
771 };
772 
773 /* Copied from adl_features, with PC3/PC8 removed */
774 static const struct platform_features lnl_features = {
775 	.has_msr_misc_feature_control	= adl_features.has_msr_misc_feature_control,
776 	.has_msr_misc_pwr_mgmt		= adl_features.has_msr_misc_pwr_mgmt,
777 	.has_nhm_msrs			= adl_features.has_nhm_msrs,
778 	.has_config_tdp			= adl_features.has_config_tdp,
779 	.bclk_freq			= adl_features.bclk_freq,
780 	.supported_cstates		= CC1 | CC6 | CC7 | PC2 | PC6 | PC10,
781 	.cst_limit			= adl_features.cst_limit,
782 	.has_irtl_msrs			= adl_features.has_irtl_msrs,
783 	.has_msr_core_c1_res		= adl_features.has_msr_core_c1_res,
784 	.has_ext_cst_msrs		= adl_features.has_ext_cst_msrs,
785 	.trl_msrs			= adl_features.trl_msrs,
786 	.tcc_offset_bits		= adl_features.tcc_offset_bits,
787 	.rapl_msrs			= adl_features.rapl_msrs,
788 	.enable_tsc_tweak		= adl_features.enable_tsc_tweak,
789 };
790 
791 static const struct platform_features skx_features = {
792 	.has_msr_misc_feature_control = 1,
793 	.has_msr_misc_pwr_mgmt = 1,
794 	.has_nhm_msrs = 1,
795 	.has_config_tdp = 1,
796 	.bclk_freq = BCLK_100MHZ,
797 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
798 	.cst_limit = CST_LIMIT_SKX,
799 	.has_irtl_msrs = 1,
800 	.has_cst_auto_convension = 1,
801 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
802 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
803 	.has_fixed_rapl_unit = 1,
804 };
805 
806 static const struct platform_features icx_features = {
807 	.has_msr_misc_feature_control = 1,
808 	.has_msr_misc_pwr_mgmt = 1,
809 	.has_nhm_msrs = 1,
810 	.has_config_tdp = 1,
811 	.bclk_freq = BCLK_100MHZ,
812 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
813 	.cst_limit = CST_LIMIT_ICX,
814 	.has_msr_core_c1_res = 1,
815 	.has_irtl_msrs = 1,
816 	.has_cst_prewake_bit = 1,
817 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
818 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
819 	.has_fixed_rapl_unit = 1,
820 };
821 
822 static const struct platform_features spr_features = {
823 	.has_msr_misc_feature_control = 1,
824 	.has_msr_misc_pwr_mgmt = 1,
825 	.has_nhm_msrs = 1,
826 	.has_config_tdp = 1,
827 	.bclk_freq = BCLK_100MHZ,
828 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
829 	.cst_limit = CST_LIMIT_SKX,
830 	.has_msr_core_c1_res = 1,
831 	.has_irtl_msrs = 1,
832 	.has_cst_prewake_bit = 1,
833 	.has_fixed_rapl_psys_unit = 1,
834 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
835 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
836 };
837 
838 static const struct platform_features srf_features = {
839 	.has_msr_misc_feature_control = 1,
840 	.has_msr_misc_pwr_mgmt = 1,
841 	.has_nhm_msrs = 1,
842 	.has_config_tdp = 1,
843 	.bclk_freq = BCLK_100MHZ,
844 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
845 	.cst_limit = CST_LIMIT_SKX,
846 	.has_msr_core_c1_res = 1,
847 	.has_msr_module_c6_res_ms = 1,
848 	.has_irtl_msrs = 1,
849 	.has_cst_prewake_bit = 1,
850 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
851 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
852 };
853 
854 static const struct platform_features grr_features = {
855 	.has_msr_misc_feature_control = 1,
856 	.has_msr_misc_pwr_mgmt = 1,
857 	.has_nhm_msrs = 1,
858 	.has_config_tdp = 1,
859 	.bclk_freq = BCLK_100MHZ,
860 	.supported_cstates = CC1 | CC6,
861 	.cst_limit = CST_LIMIT_SKX,
862 	.has_msr_core_c1_res = 1,
863 	.has_msr_module_c6_res_ms = 1,
864 	.has_irtl_msrs = 1,
865 	.has_cst_prewake_bit = 1,
866 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
867 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
868 };
869 
870 static const struct platform_features slv_features = {
871 	.has_nhm_msrs = 1,
872 	.bclk_freq = BCLK_SLV,
873 	.supported_cstates = CC1 | CC6 | PC6,
874 	.cst_limit = CST_LIMIT_SLV,
875 	.has_msr_core_c1_res = 1,
876 	.has_msr_module_c6_res_ms = 1,
877 	.has_msr_c6_demotion_policy_config = 1,
878 	.has_msr_atom_pkg_c6_residency = 1,
879 	.trl_msrs = TRL_ATOM,
880 	.rapl_msrs = RAPL_PKG | RAPL_CORE,
881 	.has_rapl_divisor = 1,
882 	.rapl_quirk_tdp = 30,
883 };
884 
885 static const struct platform_features slvd_features = {
886 	.has_msr_misc_pwr_mgmt = 1,
887 	.has_nhm_msrs = 1,
888 	.bclk_freq = BCLK_SLV,
889 	.supported_cstates = CC1 | CC6 | PC3 | PC6,
890 	.cst_limit = CST_LIMIT_SLV,
891 	.has_msr_atom_pkg_c6_residency = 1,
892 	.trl_msrs = TRL_BASE,
893 	.rapl_msrs = RAPL_PKG | RAPL_CORE,
894 	.rapl_quirk_tdp = 30,
895 };
896 
897 static const struct platform_features amt_features = {
898 	.has_nhm_msrs = 1,
899 	.bclk_freq = BCLK_133MHZ,
900 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
901 	.cst_limit = CST_LIMIT_AMT,
902 	.trl_msrs = TRL_BASE,
903 };
904 
905 static const struct platform_features gmt_features = {
906 	.has_msr_misc_pwr_mgmt = 1,
907 	.has_nhm_msrs = 1,
908 	.bclk_freq = BCLK_100MHZ,
909 	.crystal_freq = 19200000,
910 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
911 	.cst_limit = CST_LIMIT_GMT,
912 	.has_irtl_msrs = 1,
913 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
914 	.rapl_msrs = RAPL_PKG | RAPL_PKG_POWER_INFO,
915 };
916 
917 static const struct platform_features gmtd_features = {
918 	.has_msr_misc_pwr_mgmt = 1,
919 	.has_nhm_msrs = 1,
920 	.bclk_freq = BCLK_100MHZ,
921 	.crystal_freq = 25000000,
922 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
923 	.cst_limit = CST_LIMIT_GMT,
924 	.has_irtl_msrs = 1,
925 	.has_msr_core_c1_res = 1,
926 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
927 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_CORE_ENERGY_STATUS,
928 };
929 
930 static const struct platform_features gmtp_features = {
931 	.has_msr_misc_pwr_mgmt = 1,
932 	.has_nhm_msrs = 1,
933 	.bclk_freq = BCLK_100MHZ,
934 	.crystal_freq = 19200000,
935 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
936 	.cst_limit = CST_LIMIT_GMT,
937 	.has_irtl_msrs = 1,
938 	.trl_msrs = TRL_BASE,
939 	.rapl_msrs = RAPL_PKG | RAPL_PKG_POWER_INFO,
940 };
941 
942 static const struct platform_features tmt_features = {
943 	.has_msr_misc_pwr_mgmt = 1,
944 	.has_nhm_msrs = 1,
945 	.bclk_freq = BCLK_100MHZ,
946 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
947 	.cst_limit = CST_LIMIT_GMT,
948 	.has_irtl_msrs = 1,
949 	.trl_msrs = TRL_BASE,
950 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX,
951 	.enable_tsc_tweak = 1,
952 };
953 
954 static const struct platform_features tmtd_features = {
955 	.has_msr_misc_pwr_mgmt = 1,
956 	.has_nhm_msrs = 1,
957 	.bclk_freq = BCLK_100MHZ,
958 	.supported_cstates = CC1 | CC6,
959 	.cst_limit = CST_LIMIT_GMT,
960 	.has_irtl_msrs = 1,
961 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
962 	.rapl_msrs = RAPL_PKG_ALL,
963 };
964 
965 static const struct platform_features knl_features = {
966 	.has_msr_misc_pwr_mgmt = 1,
967 	.has_nhm_msrs = 1,
968 	.has_config_tdp = 1,
969 	.bclk_freq = BCLK_100MHZ,
970 	.supported_cstates = CC1 | CC6 | PC3 | PC6,
971 	.cst_limit = CST_LIMIT_KNL,
972 	.has_msr_knl_core_c6_residency = 1,
973 	.trl_msrs = TRL_KNL,
974 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
975 	.has_fixed_rapl_unit = 1,
976 	.need_perf_multiplier = 1,
977 };
978 
979 static const struct platform_features default_features = {
980 };
981 
982 static const struct platform_features amd_features_with_rapl = {
983 	.rapl_msrs = RAPL_AMD_F17H,
984 	.has_per_core_rapl = 1,
985 	.rapl_quirk_tdp = 280,	/* This is the max stock TDP of HEDT/Server Fam17h+ chips */
986 };
987 
988 static const struct platform_data turbostat_pdata[] = {
989 	{ INTEL_NEHALEM, &nhm_features },
990 	{ INTEL_NEHALEM_G, &nhm_features },
991 	{ INTEL_NEHALEM_EP, &nhm_features },
992 	{ INTEL_NEHALEM_EX, &nhx_features },
993 	{ INTEL_WESTMERE, &nhm_features },
994 	{ INTEL_WESTMERE_EP, &nhm_features },
995 	{ INTEL_WESTMERE_EX, &nhx_features },
996 	{ INTEL_SANDYBRIDGE, &snb_features },
997 	{ INTEL_SANDYBRIDGE_X, &snx_features },
998 	{ INTEL_IVYBRIDGE, &ivb_features },
999 	{ INTEL_IVYBRIDGE_X, &ivx_features },
1000 	{ INTEL_HASWELL, &hsw_features },
1001 	{ INTEL_HASWELL_X, &hsx_features },
1002 	{ INTEL_HASWELL_L, &hswl_features },
1003 	{ INTEL_HASWELL_G, &hswg_features },
1004 	{ INTEL_BROADWELL, &bdw_features },
1005 	{ INTEL_BROADWELL_G, &bdwg_features },
1006 	{ INTEL_BROADWELL_X, &bdx_features },
1007 	{ INTEL_BROADWELL_D, &bdx_features },
1008 	{ INTEL_SKYLAKE_L, &skl_features },
1009 	{ INTEL_SKYLAKE, &skl_features },
1010 	{ INTEL_SKYLAKE_X, &skx_features },
1011 	{ INTEL_KABYLAKE_L, &skl_features },
1012 	{ INTEL_KABYLAKE, &skl_features },
1013 	{ INTEL_COMETLAKE, &skl_features },
1014 	{ INTEL_COMETLAKE_L, &skl_features },
1015 	{ INTEL_CANNONLAKE_L, &cnl_features },
1016 	{ INTEL_ICELAKE_X, &icx_features },
1017 	{ INTEL_ICELAKE_D, &icx_features },
1018 	{ INTEL_ICELAKE_L, &cnl_features },
1019 	{ INTEL_ICELAKE_NNPI, &cnl_features },
1020 	{ INTEL_ROCKETLAKE, &cnl_features },
1021 	{ INTEL_TIGERLAKE_L, &cnl_features },
1022 	{ INTEL_TIGERLAKE, &cnl_features },
1023 	{ INTEL_SAPPHIRERAPIDS_X, &spr_features },
1024 	{ INTEL_EMERALDRAPIDS_X, &spr_features },
1025 	{ INTEL_GRANITERAPIDS_X, &spr_features },
1026 	{ INTEL_GRANITERAPIDS_D, &spr_features },
1027 	{ INTEL_LAKEFIELD, &cnl_features },
1028 	{ INTEL_ALDERLAKE, &adl_features },
1029 	{ INTEL_ALDERLAKE_L, &adl_features },
1030 	{ INTEL_RAPTORLAKE, &adl_features },
1031 	{ INTEL_RAPTORLAKE_P, &adl_features },
1032 	{ INTEL_RAPTORLAKE_S, &adl_features },
1033 	{ INTEL_METEORLAKE, &adl_features },
1034 	{ INTEL_METEORLAKE_L, &adl_features },
1035 	{ INTEL_ARROWLAKE_H, &adl_features },
1036 	{ INTEL_ARROWLAKE_U, &adl_features },
1037 	{ INTEL_ARROWLAKE, &adl_features },
1038 	{ INTEL_LUNARLAKE_M, &lnl_features },
1039 	{ INTEL_PANTHERLAKE_L, &lnl_features },
1040 	{ INTEL_ATOM_SILVERMONT, &slv_features },
1041 	{ INTEL_ATOM_SILVERMONT_D, &slvd_features },
1042 	{ INTEL_ATOM_AIRMONT, &amt_features },
1043 	{ INTEL_ATOM_GOLDMONT, &gmt_features },
1044 	{ INTEL_ATOM_GOLDMONT_D, &gmtd_features },
1045 	{ INTEL_ATOM_GOLDMONT_PLUS, &gmtp_features },
1046 	{ INTEL_ATOM_TREMONT_D, &tmtd_features },
1047 	{ INTEL_ATOM_TREMONT, &tmt_features },
1048 	{ INTEL_ATOM_TREMONT_L, &tmt_features },
1049 	{ INTEL_ATOM_GRACEMONT, &adl_features },
1050 	{ INTEL_ATOM_CRESTMONT_X, &srf_features },
1051 	{ INTEL_ATOM_CRESTMONT, &grr_features },
1052 	{ INTEL_ATOM_DARKMONT_X, &srf_features },
1053 	{ INTEL_XEON_PHI_KNL, &knl_features },
1054 	{ INTEL_XEON_PHI_KNM, &knl_features },
1055 	/*
1056 	 * Missing support for
1057 	 * INTEL_ICELAKE
1058 	 * INTEL_ATOM_SILVERMONT_MID
1059 	 * INTEL_ATOM_SILVERMONT_MID2
1060 	 * INTEL_ATOM_AIRMONT_NP
1061 	 */
1062 	{ 0, NULL },
1063 };
1064 
1065 static const struct platform_features *platform;
1066 
probe_platform_features(unsigned int family,unsigned int model)1067 void probe_platform_features(unsigned int family, unsigned int model)
1068 {
1069 	int i;
1070 
1071 
1072 	if (authentic_amd || hygon_genuine) {
1073 		/* fallback to default features on unsupported models */
1074 		force_load++;
1075 		if (max_extended_level >= 0x80000007) {
1076 			unsigned int eax, ebx, ecx, edx;
1077 
1078 			__cpuid(0x80000007, eax, ebx, ecx, edx);
1079 			/* RAPL (Fam 17h+) */
1080 			if ((edx & (1 << 14)) && family >= 0x17)
1081 				platform = &amd_features_with_rapl;
1082 		}
1083 		goto end;
1084 	}
1085 
1086 	if (!genuine_intel)
1087 		goto end;
1088 
1089 	for (i = 0; turbostat_pdata[i].features; i++) {
1090 		if (VFM_FAMILY(turbostat_pdata[i].vfm) == family && VFM_MODEL(turbostat_pdata[i].vfm) == model) {
1091 			platform = turbostat_pdata[i].features;
1092 			return;
1093 		}
1094 	}
1095 
1096 end:
1097 	if (force_load && !platform) {
1098 		fprintf(outf, "Forced to run on unsupported platform!\n");
1099 		platform = &default_features;
1100 	}
1101 
1102 	if (platform)
1103 		return;
1104 
1105 	fprintf(stderr, "Unsupported platform detected.\n"
1106 		"\tSee RUN THE LATEST VERSION on turbostat(8)\n");
1107 	exit(1);
1108 }
1109 
1110 /* Model specific support End */
1111 
1112 #define	TJMAX_DEFAULT	100
1113 
1114 /* MSRs that are not yet in the kernel-provided header. */
1115 #define MSR_RAPL_PWR_UNIT	0xc0010299
1116 #define MSR_CORE_ENERGY_STAT	0xc001029a
1117 #define MSR_PKG_ENERGY_STAT	0xc001029b
1118 
1119 #define MAX(a, b) ((a) > (b) ? (a) : (b))
1120 
1121 int backwards_count;
1122 char *progname;
1123 
1124 #define CPU_SUBSET_MAXCPUS	1024	/* need to use before probe... */
1125 cpu_set_t *cpu_present_set, *cpu_possible_set, *cpu_effective_set, *cpu_allowed_set, *cpu_affinity_set, *cpu_subset;
1126 size_t cpu_present_setsize, cpu_possible_setsize, cpu_effective_setsize, cpu_allowed_setsize, cpu_affinity_setsize, cpu_subset_size;
1127 #define MAX_ADDED_THREAD_COUNTERS 24
1128 #define MAX_ADDED_CORE_COUNTERS 8
1129 #define MAX_ADDED_PACKAGE_COUNTERS 16
1130 #define PMT_MAX_ADDED_THREAD_COUNTERS 24
1131 #define PMT_MAX_ADDED_CORE_COUNTERS 8
1132 #define PMT_MAX_ADDED_PACKAGE_COUNTERS 16
1133 #define BITMASK_SIZE 32
1134 
1135 #define ZERO_ARRAY(arr) (memset(arr, 0, sizeof(arr)) + __must_be_array(arr))
1136 
1137 /* Indexes used to map data read from perf and MSRs into global variables */
1138 enum rapl_rci_index {
1139 	RAPL_RCI_INDEX_ENERGY_PKG = 0,
1140 	RAPL_RCI_INDEX_ENERGY_CORES = 1,
1141 	RAPL_RCI_INDEX_DRAM = 2,
1142 	RAPL_RCI_INDEX_GFX = 3,
1143 	RAPL_RCI_INDEX_PKG_PERF_STATUS = 4,
1144 	RAPL_RCI_INDEX_DRAM_PERF_STATUS = 5,
1145 	RAPL_RCI_INDEX_CORE_ENERGY = 6,
1146 	RAPL_RCI_INDEX_ENERGY_PLATFORM = 7,
1147 	NUM_RAPL_COUNTERS,
1148 };
1149 
1150 enum rapl_unit {
1151 	RAPL_UNIT_INVALID,
1152 	RAPL_UNIT_JOULES,
1153 	RAPL_UNIT_WATTS,
1154 };
1155 
1156 struct rapl_counter_info_t {
1157 	unsigned long long data[NUM_RAPL_COUNTERS];
1158 	enum counter_source source[NUM_RAPL_COUNTERS];
1159 	unsigned long long flags[NUM_RAPL_COUNTERS];
1160 	double scale[NUM_RAPL_COUNTERS];
1161 	enum rapl_unit unit[NUM_RAPL_COUNTERS];
1162 	unsigned long long msr[NUM_RAPL_COUNTERS];
1163 	unsigned long long msr_mask[NUM_RAPL_COUNTERS];
1164 	int msr_shift[NUM_RAPL_COUNTERS];
1165 
1166 	int fd_perf;
1167 };
1168 
1169 /* struct rapl_counter_info_t for each RAPL domain */
1170 struct rapl_counter_info_t *rapl_counter_info_perdomain;
1171 unsigned int rapl_counter_info_perdomain_size;
1172 
1173 #define RAPL_COUNTER_FLAG_PLATFORM_COUNTER (1u << 0)
1174 #define RAPL_COUNTER_FLAG_USE_MSR_SUM (1u << 1)
1175 
1176 struct rapl_counter_arch_info {
1177 	int feature_mask;	/* Mask for testing if the counter is supported on host */
1178 	const char *perf_subsys;
1179 	const char *perf_name;
1180 	unsigned long long msr;
1181 	unsigned long long msr_mask;
1182 	int msr_shift;		/* Positive mean shift right, negative mean shift left */
1183 	double *platform_rapl_msr_scale;	/* Scale applied to values read by MSR (platform dependent, filled at runtime) */
1184 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1185 	unsigned long long bic;
1186 	double compat_scale;	/* Some counters require constant scaling to be in the same range as other, similar ones */
1187 	unsigned long long flags;
1188 };
1189 
1190 static const struct rapl_counter_arch_info rapl_counter_arch_infos[] = {
1191 	{
1192 	 .feature_mask = RAPL_PKG,
1193 	 .perf_subsys = "power",
1194 	 .perf_name = "energy-pkg",
1195 	 .msr = MSR_PKG_ENERGY_STATUS,
1196 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1197 	 .msr_shift = 0,
1198 	 .platform_rapl_msr_scale = &rapl_energy_units,
1199 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1200 	 .bic = BIC_PkgWatt | BIC_Pkg_J,
1201 	 .compat_scale = 1.0,
1202 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1203 	  },
1204 	{
1205 	 .feature_mask = RAPL_AMD_F17H,
1206 	 .perf_subsys = "power",
1207 	 .perf_name = "energy-pkg",
1208 	 .msr = MSR_PKG_ENERGY_STAT,
1209 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1210 	 .msr_shift = 0,
1211 	 .platform_rapl_msr_scale = &rapl_energy_units,
1212 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1213 	 .bic = BIC_PkgWatt | BIC_Pkg_J,
1214 	 .compat_scale = 1.0,
1215 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1216 	  },
1217 	{
1218 	 .feature_mask = RAPL_CORE_ENERGY_STATUS,
1219 	 .perf_subsys = "power",
1220 	 .perf_name = "energy-cores",
1221 	 .msr = MSR_PP0_ENERGY_STATUS,
1222 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1223 	 .msr_shift = 0,
1224 	 .platform_rapl_msr_scale = &rapl_energy_units,
1225 	 .rci_index = RAPL_RCI_INDEX_ENERGY_CORES,
1226 	 .bic = BIC_CorWatt | BIC_Cor_J,
1227 	 .compat_scale = 1.0,
1228 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1229 	  },
1230 	{
1231 	 .feature_mask = RAPL_DRAM,
1232 	 .perf_subsys = "power",
1233 	 .perf_name = "energy-ram",
1234 	 .msr = MSR_DRAM_ENERGY_STATUS,
1235 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1236 	 .msr_shift = 0,
1237 	 .platform_rapl_msr_scale = &rapl_dram_energy_units,
1238 	 .rci_index = RAPL_RCI_INDEX_DRAM,
1239 	 .bic = BIC_RAMWatt | BIC_RAM_J,
1240 	 .compat_scale = 1.0,
1241 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1242 	  },
1243 	{
1244 	 .feature_mask = RAPL_GFX,
1245 	 .perf_subsys = "power",
1246 	 .perf_name = "energy-gpu",
1247 	 .msr = MSR_PP1_ENERGY_STATUS,
1248 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1249 	 .msr_shift = 0,
1250 	 .platform_rapl_msr_scale = &rapl_energy_units,
1251 	 .rci_index = RAPL_RCI_INDEX_GFX,
1252 	 .bic = BIC_GFXWatt | BIC_GFX_J,
1253 	 .compat_scale = 1.0,
1254 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1255 	  },
1256 	{
1257 	 .feature_mask = RAPL_PKG_PERF_STATUS,
1258 	 .perf_subsys = NULL,
1259 	 .perf_name = NULL,
1260 	 .msr = MSR_PKG_PERF_STATUS,
1261 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1262 	 .msr_shift = 0,
1263 	 .platform_rapl_msr_scale = &rapl_time_units,
1264 	 .rci_index = RAPL_RCI_INDEX_PKG_PERF_STATUS,
1265 	 .bic = BIC_PKG__,
1266 	 .compat_scale = 100.0,
1267 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1268 	  },
1269 	{
1270 	 .feature_mask = RAPL_DRAM_PERF_STATUS,
1271 	 .perf_subsys = NULL,
1272 	 .perf_name = NULL,
1273 	 .msr = MSR_DRAM_PERF_STATUS,
1274 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1275 	 .msr_shift = 0,
1276 	 .platform_rapl_msr_scale = &rapl_time_units,
1277 	 .rci_index = RAPL_RCI_INDEX_DRAM_PERF_STATUS,
1278 	 .bic = BIC_RAM__,
1279 	 .compat_scale = 100.0,
1280 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1281 	  },
1282 	{
1283 	 .feature_mask = RAPL_AMD_F17H,
1284 	 .perf_subsys = NULL,
1285 	 .perf_name = NULL,
1286 	 .msr = MSR_CORE_ENERGY_STAT,
1287 	 .msr_mask = 0xFFFFFFFF,
1288 	 .msr_shift = 0,
1289 	 .platform_rapl_msr_scale = &rapl_energy_units,
1290 	 .rci_index = RAPL_RCI_INDEX_CORE_ENERGY,
1291 	 .bic = BIC_CorWatt | BIC_Cor_J,
1292 	 .compat_scale = 1.0,
1293 	 .flags = 0,
1294 	  },
1295 	{
1296 	 .feature_mask = RAPL_PSYS,
1297 	 .perf_subsys = "power",
1298 	 .perf_name = "energy-psys",
1299 	 .msr = MSR_PLATFORM_ENERGY_STATUS,
1300 	 .msr_mask = 0x00000000FFFFFFFF,
1301 	 .msr_shift = 0,
1302 	 .platform_rapl_msr_scale = &rapl_psys_energy_units,
1303 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PLATFORM,
1304 	 .bic = BIC_SysWatt | BIC_Sys_J,
1305 	 .compat_scale = 1.0,
1306 	 .flags = RAPL_COUNTER_FLAG_PLATFORM_COUNTER | RAPL_COUNTER_FLAG_USE_MSR_SUM,
1307 	  },
1308 };
1309 
1310 struct rapl_counter {
1311 	unsigned long long raw_value;
1312 	enum rapl_unit unit;
1313 	double scale;
1314 };
1315 
1316 /* Indexes used to map data read from perf and MSRs into global variables */
1317 enum ccstate_rci_index {
1318 	CCSTATE_RCI_INDEX_C1_RESIDENCY = 0,
1319 	CCSTATE_RCI_INDEX_C3_RESIDENCY = 1,
1320 	CCSTATE_RCI_INDEX_C6_RESIDENCY = 2,
1321 	CCSTATE_RCI_INDEX_C7_RESIDENCY = 3,
1322 	PCSTATE_RCI_INDEX_C2_RESIDENCY = 4,
1323 	PCSTATE_RCI_INDEX_C3_RESIDENCY = 5,
1324 	PCSTATE_RCI_INDEX_C6_RESIDENCY = 6,
1325 	PCSTATE_RCI_INDEX_C7_RESIDENCY = 7,
1326 	PCSTATE_RCI_INDEX_C8_RESIDENCY = 8,
1327 	PCSTATE_RCI_INDEX_C9_RESIDENCY = 9,
1328 	PCSTATE_RCI_INDEX_C10_RESIDENCY = 10,
1329 	NUM_CSTATE_COUNTERS,
1330 };
1331 
1332 struct cstate_counter_info_t {
1333 	unsigned long long data[NUM_CSTATE_COUNTERS];
1334 	enum counter_source source[NUM_CSTATE_COUNTERS];
1335 	unsigned long long msr[NUM_CSTATE_COUNTERS];
1336 	int fd_perf_core;
1337 	int fd_perf_pkg;
1338 };
1339 
1340 struct cstate_counter_info_t *ccstate_counter_info;
1341 unsigned int ccstate_counter_info_size;
1342 
1343 #define CSTATE_COUNTER_FLAG_COLLECT_PER_CORE   (1u << 0)
1344 #define CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD ((1u << 1) | CSTATE_COUNTER_FLAG_COLLECT_PER_CORE)
1345 #define CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY (1u << 2)
1346 
1347 struct cstate_counter_arch_info {
1348 	int feature_mask;	/* Mask for testing if the counter is supported on host */
1349 	const char *perf_subsys;
1350 	const char *perf_name;
1351 	unsigned long long msr;
1352 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1353 	unsigned long long bic;
1354 	unsigned long long flags;
1355 	int pkg_cstate_limit;
1356 };
1357 
1358 static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
1359 	{
1360 	 .feature_mask = CC1,
1361 	 .perf_subsys = "cstate_core",
1362 	 .perf_name = "c1-residency",
1363 	 .msr = MSR_CORE_C1_RES,
1364 	 .rci_index = CCSTATE_RCI_INDEX_C1_RESIDENCY,
1365 	 .bic = BIC_CPU_c1,
1366 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD,
1367 	 .pkg_cstate_limit = 0,
1368 	  },
1369 	{
1370 	 .feature_mask = CC3,
1371 	 .perf_subsys = "cstate_core",
1372 	 .perf_name = "c3-residency",
1373 	 .msr = MSR_CORE_C3_RESIDENCY,
1374 	 .rci_index = CCSTATE_RCI_INDEX_C3_RESIDENCY,
1375 	 .bic = BIC_CPU_c3,
1376 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1377 	 .pkg_cstate_limit = 0,
1378 	  },
1379 	{
1380 	 .feature_mask = CC6,
1381 	 .perf_subsys = "cstate_core",
1382 	 .perf_name = "c6-residency",
1383 	 .msr = MSR_CORE_C6_RESIDENCY,
1384 	 .rci_index = CCSTATE_RCI_INDEX_C6_RESIDENCY,
1385 	 .bic = BIC_CPU_c6,
1386 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1387 	 .pkg_cstate_limit = 0,
1388 	  },
1389 	{
1390 	 .feature_mask = CC7,
1391 	 .perf_subsys = "cstate_core",
1392 	 .perf_name = "c7-residency",
1393 	 .msr = MSR_CORE_C7_RESIDENCY,
1394 	 .rci_index = CCSTATE_RCI_INDEX_C7_RESIDENCY,
1395 	 .bic = BIC_CPU_c7,
1396 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1397 	 .pkg_cstate_limit = 0,
1398 	  },
1399 	{
1400 	 .feature_mask = PC2,
1401 	 .perf_subsys = "cstate_pkg",
1402 	 .perf_name = "c2-residency",
1403 	 .msr = MSR_PKG_C2_RESIDENCY,
1404 	 .rci_index = PCSTATE_RCI_INDEX_C2_RESIDENCY,
1405 	 .bic = BIC_Pkgpc2,
1406 	 .flags = 0,
1407 	 .pkg_cstate_limit = PCL__2,
1408 	  },
1409 	{
1410 	 .feature_mask = PC3,
1411 	 .perf_subsys = "cstate_pkg",
1412 	 .perf_name = "c3-residency",
1413 	 .msr = MSR_PKG_C3_RESIDENCY,
1414 	 .rci_index = PCSTATE_RCI_INDEX_C3_RESIDENCY,
1415 	 .bic = BIC_Pkgpc3,
1416 	 .flags = 0,
1417 	 .pkg_cstate_limit = PCL__3,
1418 	  },
1419 	{
1420 	 .feature_mask = PC6,
1421 	 .perf_subsys = "cstate_pkg",
1422 	 .perf_name = "c6-residency",
1423 	 .msr = MSR_PKG_C6_RESIDENCY,
1424 	 .rci_index = PCSTATE_RCI_INDEX_C6_RESIDENCY,
1425 	 .bic = BIC_Pkgpc6,
1426 	 .flags = 0,
1427 	 .pkg_cstate_limit = PCL__6,
1428 	  },
1429 	{
1430 	 .feature_mask = PC7,
1431 	 .perf_subsys = "cstate_pkg",
1432 	 .perf_name = "c7-residency",
1433 	 .msr = MSR_PKG_C7_RESIDENCY,
1434 	 .rci_index = PCSTATE_RCI_INDEX_C7_RESIDENCY,
1435 	 .bic = BIC_Pkgpc7,
1436 	 .flags = 0,
1437 	 .pkg_cstate_limit = PCL__7,
1438 	  },
1439 	{
1440 	 .feature_mask = PC8,
1441 	 .perf_subsys = "cstate_pkg",
1442 	 .perf_name = "c8-residency",
1443 	 .msr = MSR_PKG_C8_RESIDENCY,
1444 	 .rci_index = PCSTATE_RCI_INDEX_C8_RESIDENCY,
1445 	 .bic = BIC_Pkgpc8,
1446 	 .flags = 0,
1447 	 .pkg_cstate_limit = PCL__8,
1448 	  },
1449 	{
1450 	 .feature_mask = PC9,
1451 	 .perf_subsys = "cstate_pkg",
1452 	 .perf_name = "c9-residency",
1453 	 .msr = MSR_PKG_C9_RESIDENCY,
1454 	 .rci_index = PCSTATE_RCI_INDEX_C9_RESIDENCY,
1455 	 .bic = BIC_Pkgpc9,
1456 	 .flags = 0,
1457 	 .pkg_cstate_limit = PCL__9,
1458 	  },
1459 	{
1460 	 .feature_mask = PC10,
1461 	 .perf_subsys = "cstate_pkg",
1462 	 .perf_name = "c10-residency",
1463 	 .msr = MSR_PKG_C10_RESIDENCY,
1464 	 .rci_index = PCSTATE_RCI_INDEX_C10_RESIDENCY,
1465 	 .bic = BIC_Pkgpc10,
1466 	 .flags = 0,
1467 	 .pkg_cstate_limit = PCL_10,
1468 	  },
1469 };
1470 
1471 /* Indexes used to map data read from perf and MSRs into global variables */
1472 enum msr_rci_index {
1473 	MSR_RCI_INDEX_APERF = 0,
1474 	MSR_RCI_INDEX_MPERF = 1,
1475 	MSR_RCI_INDEX_SMI = 2,
1476 	NUM_MSR_COUNTERS,
1477 };
1478 
1479 struct msr_counter_info_t {
1480 	unsigned long long data[NUM_MSR_COUNTERS];
1481 	enum counter_source source[NUM_MSR_COUNTERS];
1482 	unsigned long long msr[NUM_MSR_COUNTERS];
1483 	unsigned long long msr_mask[NUM_MSR_COUNTERS];
1484 	int fd_perf;
1485 };
1486 
1487 struct msr_counter_info_t *msr_counter_info;
1488 unsigned int msr_counter_info_size;
1489 
1490 struct msr_counter_arch_info {
1491 	const char *perf_subsys;
1492 	const char *perf_name;
1493 	unsigned long long msr;
1494 	unsigned long long msr_mask;
1495 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1496 	bool needed;
1497 	bool present;
1498 };
1499 
1500 enum msr_arch_info_index {
1501 	MSR_ARCH_INFO_APERF_INDEX = 0,
1502 	MSR_ARCH_INFO_MPERF_INDEX = 1,
1503 	MSR_ARCH_INFO_SMI_INDEX = 2,
1504 };
1505 
1506 static struct msr_counter_arch_info msr_counter_arch_infos[] = {
1507 	[MSR_ARCH_INFO_APERF_INDEX] = {
1508 				       .perf_subsys = "msr",
1509 				       .perf_name = "aperf",
1510 				       .msr = MSR_IA32_APERF,
1511 				       .msr_mask = 0xFFFFFFFFFFFFFFFF,
1512 				       .rci_index = MSR_RCI_INDEX_APERF,
1513 				        },
1514 
1515 	[MSR_ARCH_INFO_MPERF_INDEX] = {
1516 				       .perf_subsys = "msr",
1517 				       .perf_name = "mperf",
1518 				       .msr = MSR_IA32_MPERF,
1519 				       .msr_mask = 0xFFFFFFFFFFFFFFFF,
1520 				       .rci_index = MSR_RCI_INDEX_MPERF,
1521 				        },
1522 
1523 	[MSR_ARCH_INFO_SMI_INDEX] = {
1524 				     .perf_subsys = "msr",
1525 				     .perf_name = "smi",
1526 				     .msr = MSR_SMI_COUNT,
1527 				     .msr_mask = 0xFFFFFFFF,
1528 				     .rci_index = MSR_RCI_INDEX_SMI,
1529 				      },
1530 };
1531 
1532 /* Can be redefined when compiling, useful for testing. */
1533 #ifndef SYSFS_TELEM_PATH
1534 #define SYSFS_TELEM_PATH "/sys/class/intel_pmt"
1535 #endif
1536 
1537 #define PMT_COUNTER_MTL_DC6_OFFSET 120
1538 #define PMT_COUNTER_MTL_DC6_LSB    0
1539 #define PMT_COUNTER_MTL_DC6_MSB    63
1540 #define PMT_MTL_DC6_GUID           0x1a067102
1541 #define PMT_MTL_DC6_SEQ            0
1542 
1543 #define PMT_COUNTER_CWF_MC1E_OFFSET_BASE          20936
1544 #define PMT_COUNTER_CWF_MC1E_OFFSET_INCREMENT     24
1545 #define PMT_COUNTER_CWF_MC1E_NUM_MODULES_PER_FILE 12
1546 #define PMT_COUNTER_CWF_CPUS_PER_MODULE           4
1547 #define PMT_COUNTER_CWF_MC1E_LSB                  0
1548 #define PMT_COUNTER_CWF_MC1E_MSB                  63
1549 #define PMT_CWF_MC1E_GUID                         0x14421519
1550 
1551 unsigned long long tcore_clock_freq_hz = 800000000;
1552 
1553 #define PMT_COUNTER_NAME_SIZE_BYTES      16
1554 #define PMT_COUNTER_TYPE_NAME_SIZE_BYTES 32
1555 
1556 struct pmt_mmio {
1557 	struct pmt_mmio *next;
1558 
1559 	unsigned int guid;
1560 	unsigned int size;
1561 
1562 	/* Base pointer to the mmaped memory. */
1563 	void *mmio_base;
1564 
1565 	/*
1566 	 * Offset to be applied to the mmio_base
1567 	 * to get the beginning of the PMT counters for given GUID.
1568 	 */
1569 	unsigned long pmt_offset;
1570 } *pmt_mmios;
1571 
1572 enum pmt_datatype {
1573 	PMT_TYPE_RAW,
1574 	PMT_TYPE_XTAL_TIME,
1575 	PMT_TYPE_TCORE_CLOCK,
1576 };
1577 
1578 struct pmt_domain_info {
1579 	/*
1580 	 * Pointer to the MMIO obtained by applying a counter offset
1581 	 * to the mmio_base of the mmaped region for the given GUID.
1582 	 *
1583 	 * This is where to read the raw value of the counter from.
1584 	 */
1585 	unsigned long *pcounter;
1586 };
1587 
1588 struct pmt_counter {
1589 	struct pmt_counter *next;
1590 
1591 	/* PMT metadata */
1592 	char name[PMT_COUNTER_NAME_SIZE_BYTES];
1593 	enum pmt_datatype type;
1594 	enum counter_scope scope;
1595 	unsigned int lsb;
1596 	unsigned int msb;
1597 
1598 	/* BIC-like metadata */
1599 	enum counter_format format;
1600 
1601 	unsigned int num_domains;
1602 	struct pmt_domain_info *domains;
1603 };
1604 
1605 /*
1606  * PMT telemetry directory iterator.
1607  * Used to iterate telemetry files in sysfs in correct order.
1608  */
1609 struct pmt_diriter_t {
1610 	DIR *dir;
1611 	struct dirent **namelist;
1612 	unsigned int num_names;
1613 	unsigned int current_name_idx;
1614 };
1615 
pmt_telemdir_filter(const struct dirent * e)1616 int pmt_telemdir_filter(const struct dirent *e)
1617 {
1618 	unsigned int dummy;
1619 
1620 	return sscanf(e->d_name, "telem%u", &dummy);
1621 }
1622 
pmt_telemdir_sort(const struct dirent ** a,const struct dirent ** b)1623 int pmt_telemdir_sort(const struct dirent **a, const struct dirent **b)
1624 {
1625 	unsigned int aidx = 0, bidx = 0;
1626 
1627 	sscanf((*a)->d_name, "telem%u", &aidx);
1628 	sscanf((*b)->d_name, "telem%u", &bidx);
1629 
1630 	return aidx >= bidx;
1631 }
1632 
pmt_diriter_next(struct pmt_diriter_t * iter)1633 const struct dirent *pmt_diriter_next(struct pmt_diriter_t *iter)
1634 {
1635 	const struct dirent *ret = NULL;
1636 
1637 	if (!iter->dir)
1638 		return NULL;
1639 
1640 	if (iter->current_name_idx >= iter->num_names)
1641 		return NULL;
1642 
1643 	ret = iter->namelist[iter->current_name_idx];
1644 	++iter->current_name_idx;
1645 
1646 	return ret;
1647 }
1648 
pmt_diriter_begin(struct pmt_diriter_t * iter,const char * pmt_root_path)1649 const struct dirent *pmt_diriter_begin(struct pmt_diriter_t *iter, const char *pmt_root_path)
1650 {
1651 	int num_names = iter->num_names;
1652 
1653 	if (!iter->dir) {
1654 		iter->dir = opendir(pmt_root_path);
1655 		if (iter->dir == NULL)
1656 			return NULL;
1657 
1658 		num_names = scandir(pmt_root_path, &iter->namelist, pmt_telemdir_filter, pmt_telemdir_sort);
1659 		if (num_names == -1)
1660 			return NULL;
1661 	}
1662 
1663 	iter->current_name_idx = 0;
1664 	iter->num_names = num_names;
1665 
1666 	return pmt_diriter_next(iter);
1667 }
1668 
pmt_diriter_init(struct pmt_diriter_t * iter)1669 void pmt_diriter_init(struct pmt_diriter_t *iter)
1670 {
1671 	memset(iter, 0, sizeof(*iter));
1672 }
1673 
pmt_diriter_remove(struct pmt_diriter_t * iter)1674 void pmt_diriter_remove(struct pmt_diriter_t *iter)
1675 {
1676 	if (iter->namelist) {
1677 		for (unsigned int i = 0; i < iter->num_names; i++) {
1678 			free(iter->namelist[i]);
1679 			iter->namelist[i] = NULL;
1680 		}
1681 	}
1682 
1683 	free(iter->namelist);
1684 	iter->namelist = NULL;
1685 	iter->num_names = 0;
1686 	iter->current_name_idx = 0;
1687 
1688 	closedir(iter->dir);
1689 	iter->dir = NULL;
1690 }
1691 
pmt_counter_get_width(const struct pmt_counter * p)1692 unsigned int pmt_counter_get_width(const struct pmt_counter *p)
1693 {
1694 	return (p->msb - p->lsb) + 1;
1695 }
1696 
pmt_counter_resize_(struct pmt_counter * pcounter,unsigned int new_size)1697 void pmt_counter_resize_(struct pmt_counter *pcounter, unsigned int new_size)
1698 {
1699 	struct pmt_domain_info *new_mem;
1700 
1701 	new_mem = (struct pmt_domain_info *)reallocarray(pcounter->domains, new_size, sizeof(*pcounter->domains));
1702 	if (!new_mem) {
1703 		fprintf(stderr, "%s: failed to allocate memory for PMT counters\n", __func__);
1704 		exit(1);
1705 	}
1706 
1707 	/* Zero initialize just allocated memory. */
1708 	const size_t num_new_domains = new_size - pcounter->num_domains;
1709 
1710 	memset(&new_mem[pcounter->num_domains], 0, num_new_domains * sizeof(*pcounter->domains));
1711 
1712 	pcounter->num_domains = new_size;
1713 	pcounter->domains = new_mem;
1714 }
1715 
pmt_counter_resize(struct pmt_counter * pcounter,unsigned int new_size)1716 void pmt_counter_resize(struct pmt_counter *pcounter, unsigned int new_size)
1717 {
1718 	/*
1719 	 * Allocate more memory ahead of time.
1720 	 *
1721 	 * Always allocate space for at least 8 elements
1722 	 * and double the size when growing.
1723 	 */
1724 	if (new_size < 8)
1725 		new_size = 8;
1726 	new_size = MAX(new_size, pcounter->num_domains * 2);
1727 
1728 	pmt_counter_resize_(pcounter, new_size);
1729 }
1730 
1731 struct thread_data {
1732 	struct timeval tv_begin;
1733 	struct timeval tv_end;
1734 	struct timeval tv_delta;
1735 	unsigned long long tsc;
1736 	unsigned long long aperf;
1737 	unsigned long long mperf;
1738 	unsigned long long c1;
1739 	unsigned long long instr_count;
1740 	unsigned long long irq_count;
1741 	unsigned long long nmi_count;
1742 	unsigned int smi_count;
1743 	unsigned int cpu_id;
1744 	unsigned int apic_id;
1745 	unsigned int x2apic_id;
1746 	unsigned int flags;
1747 	bool is_atom;
1748 	unsigned long long counter[MAX_ADDED_THREAD_COUNTERS];
1749 	unsigned long long perf_counter[MAX_ADDED_THREAD_COUNTERS];
1750 	unsigned long long pmt_counter[PMT_MAX_ADDED_THREAD_COUNTERS];
1751 } *thread_even, *thread_odd;
1752 
1753 struct core_data {
1754 	int base_cpu;
1755 	unsigned long long c3;
1756 	unsigned long long c6;
1757 	unsigned long long c7;
1758 	unsigned long long mc6_us;	/* duplicate as per-core for now, even though per module */
1759 	unsigned int core_temp_c;
1760 	struct rapl_counter core_energy;	/* MSR_CORE_ENERGY_STAT */
1761 	unsigned int core_id;
1762 	unsigned long long core_throt_cnt;
1763 	unsigned long long counter[MAX_ADDED_CORE_COUNTERS];
1764 	unsigned long long perf_counter[MAX_ADDED_CORE_COUNTERS];
1765 	unsigned long long pmt_counter[PMT_MAX_ADDED_CORE_COUNTERS];
1766 } *core_even, *core_odd;
1767 
1768 struct pkg_data {
1769 	int base_cpu;
1770 	unsigned long long pc2;
1771 	unsigned long long pc3;
1772 	unsigned long long pc6;
1773 	unsigned long long pc7;
1774 	unsigned long long pc8;
1775 	unsigned long long pc9;
1776 	unsigned long long pc10;
1777 	long long cpu_lpi;
1778 	long long sys_lpi;
1779 	unsigned long long pkg_wtd_core_c0;
1780 	unsigned long long pkg_any_core_c0;
1781 	unsigned long long pkg_any_gfxe_c0;
1782 	unsigned long long pkg_both_core_gfxe_c0;
1783 	long long gfx_rc6_ms;
1784 	unsigned int gfx_mhz;
1785 	unsigned int gfx_act_mhz;
1786 	long long sam_mc6_ms;
1787 	unsigned int sam_mhz;
1788 	unsigned int sam_act_mhz;
1789 	unsigned int package_id;
1790 	struct rapl_counter energy_pkg;	/* MSR_PKG_ENERGY_STATUS */
1791 	struct rapl_counter energy_dram;	/* MSR_DRAM_ENERGY_STATUS */
1792 	struct rapl_counter energy_cores;	/* MSR_PP0_ENERGY_STATUS */
1793 	struct rapl_counter energy_gfx;	/* MSR_PP1_ENERGY_STATUS */
1794 	struct rapl_counter rapl_pkg_perf_status;	/* MSR_PKG_PERF_STATUS */
1795 	struct rapl_counter rapl_dram_perf_status;	/* MSR_DRAM_PERF_STATUS */
1796 	unsigned int pkg_temp_c;
1797 	unsigned int uncore_mhz;
1798 	unsigned long long die_c6;
1799 	unsigned long long counter[MAX_ADDED_PACKAGE_COUNTERS];
1800 	unsigned long long perf_counter[MAX_ADDED_PACKAGE_COUNTERS];
1801 	unsigned long long pmt_counter[PMT_MAX_ADDED_PACKAGE_COUNTERS];
1802 } *package_even, *package_odd;
1803 
1804 #define ODD_COUNTERS thread_odd, core_odd, package_odd
1805 #define EVEN_COUNTERS thread_even, core_even, package_even
1806 
1807 #define GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no)	      \
1808 	((thread_base) +						      \
1809 	 ((pkg_no) *							      \
1810 	  topo.nodes_per_pkg * topo.cores_per_node * topo.threads_per_core) + \
1811 	 ((node_no) * topo.cores_per_node * topo.threads_per_core) +	      \
1812 	 ((core_no) * topo.threads_per_core) +				      \
1813 	 (thread_no))
1814 
1815 #define GET_CORE(core_base, core_no, node_no, pkg_no)			\
1816 	((core_base) +							\
1817 	 ((pkg_no) *  topo.nodes_per_pkg * topo.cores_per_node) +	\
1818 	 ((node_no) * topo.cores_per_node) +				\
1819 	 (core_no))
1820 
1821 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
1822 
1823 /*
1824  * The accumulated sum of MSR is defined as a monotonic
1825  * increasing MSR, it will be accumulated periodically,
1826  * despite its register's bit width.
1827  */
1828 enum {
1829 	IDX_PKG_ENERGY,
1830 	IDX_DRAM_ENERGY,
1831 	IDX_PP0_ENERGY,
1832 	IDX_PP1_ENERGY,
1833 	IDX_PKG_PERF,
1834 	IDX_DRAM_PERF,
1835 	IDX_PSYS_ENERGY,
1836 	IDX_COUNT,
1837 };
1838 
1839 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr);
1840 
1841 struct msr_sum_array {
1842 	/* get_msr_sum() = sum + (get_msr() - last) */
1843 	struct {
1844 		/*The accumulated MSR value is updated by the timer */
1845 		unsigned long long sum;
1846 		/*The MSR footprint recorded in last timer */
1847 		unsigned long long last;
1848 	} entries[IDX_COUNT];
1849 };
1850 
1851 /* The percpu MSR sum array.*/
1852 struct msr_sum_array *per_cpu_msr_sum;
1853 
idx_to_offset(int idx)1854 off_t idx_to_offset(int idx)
1855 {
1856 	off_t offset;
1857 
1858 	switch (idx) {
1859 	case IDX_PKG_ENERGY:
1860 		if (platform->rapl_msrs & RAPL_AMD_F17H)
1861 			offset = MSR_PKG_ENERGY_STAT;
1862 		else
1863 			offset = MSR_PKG_ENERGY_STATUS;
1864 		break;
1865 	case IDX_DRAM_ENERGY:
1866 		offset = MSR_DRAM_ENERGY_STATUS;
1867 		break;
1868 	case IDX_PP0_ENERGY:
1869 		offset = MSR_PP0_ENERGY_STATUS;
1870 		break;
1871 	case IDX_PP1_ENERGY:
1872 		offset = MSR_PP1_ENERGY_STATUS;
1873 		break;
1874 	case IDX_PKG_PERF:
1875 		offset = MSR_PKG_PERF_STATUS;
1876 		break;
1877 	case IDX_DRAM_PERF:
1878 		offset = MSR_DRAM_PERF_STATUS;
1879 		break;
1880 	case IDX_PSYS_ENERGY:
1881 		offset = MSR_PLATFORM_ENERGY_STATUS;
1882 		break;
1883 	default:
1884 		offset = -1;
1885 	}
1886 	return offset;
1887 }
1888 
offset_to_idx(off_t offset)1889 int offset_to_idx(off_t offset)
1890 {
1891 	int idx;
1892 
1893 	switch (offset) {
1894 	case MSR_PKG_ENERGY_STATUS:
1895 	case MSR_PKG_ENERGY_STAT:
1896 		idx = IDX_PKG_ENERGY;
1897 		break;
1898 	case MSR_DRAM_ENERGY_STATUS:
1899 		idx = IDX_DRAM_ENERGY;
1900 		break;
1901 	case MSR_PP0_ENERGY_STATUS:
1902 		idx = IDX_PP0_ENERGY;
1903 		break;
1904 	case MSR_PP1_ENERGY_STATUS:
1905 		idx = IDX_PP1_ENERGY;
1906 		break;
1907 	case MSR_PKG_PERF_STATUS:
1908 		idx = IDX_PKG_PERF;
1909 		break;
1910 	case MSR_DRAM_PERF_STATUS:
1911 		idx = IDX_DRAM_PERF;
1912 		break;
1913 	case MSR_PLATFORM_ENERGY_STATUS:
1914 		idx = IDX_PSYS_ENERGY;
1915 		break;
1916 	default:
1917 		idx = -1;
1918 	}
1919 	return idx;
1920 }
1921 
idx_valid(int idx)1922 int idx_valid(int idx)
1923 {
1924 	switch (idx) {
1925 	case IDX_PKG_ENERGY:
1926 		return platform->rapl_msrs & (RAPL_PKG | RAPL_AMD_F17H);
1927 	case IDX_DRAM_ENERGY:
1928 		return platform->rapl_msrs & RAPL_DRAM;
1929 	case IDX_PP0_ENERGY:
1930 		return platform->rapl_msrs & RAPL_CORE_ENERGY_STATUS;
1931 	case IDX_PP1_ENERGY:
1932 		return platform->rapl_msrs & RAPL_GFX;
1933 	case IDX_PKG_PERF:
1934 		return platform->rapl_msrs & RAPL_PKG_PERF_STATUS;
1935 	case IDX_DRAM_PERF:
1936 		return platform->rapl_msrs & RAPL_DRAM_PERF_STATUS;
1937 	case IDX_PSYS_ENERGY:
1938 		return platform->rapl_msrs & RAPL_PSYS;
1939 	default:
1940 		return 0;
1941 	}
1942 }
1943 
1944 struct sys_counters {
1945 	/* MSR added counters */
1946 	unsigned int added_thread_counters;
1947 	unsigned int added_core_counters;
1948 	unsigned int added_package_counters;
1949 	struct msr_counter *tp;
1950 	struct msr_counter *cp;
1951 	struct msr_counter *pp;
1952 
1953 	/* perf added counters */
1954 	unsigned int added_thread_perf_counters;
1955 	unsigned int added_core_perf_counters;
1956 	unsigned int added_package_perf_counters;
1957 	struct perf_counter_info *perf_tp;
1958 	struct perf_counter_info *perf_cp;
1959 	struct perf_counter_info *perf_pp;
1960 
1961 	struct pmt_counter *pmt_tp;
1962 	struct pmt_counter *pmt_cp;
1963 	struct pmt_counter *pmt_pp;
1964 } sys;
1965 
free_msr_counters_(struct msr_counter ** pp)1966 static size_t free_msr_counters_(struct msr_counter **pp)
1967 {
1968 	struct msr_counter *p = NULL;
1969 	size_t num_freed = 0;
1970 
1971 	while (*pp) {
1972 		p = *pp;
1973 
1974 		if (p->msr_num != 0) {
1975 			*pp = p->next;
1976 
1977 			free(p);
1978 			++num_freed;
1979 
1980 			continue;
1981 		}
1982 
1983 		pp = &p->next;
1984 	}
1985 
1986 	return num_freed;
1987 }
1988 
1989 /*
1990  * Free all added counters accessed via msr.
1991  */
free_sys_msr_counters(void)1992 static void free_sys_msr_counters(void)
1993 {
1994 	/* Thread counters */
1995 	sys.added_thread_counters -= free_msr_counters_(&sys.tp);
1996 
1997 	/* Core counters */
1998 	sys.added_core_counters -= free_msr_counters_(&sys.cp);
1999 
2000 	/* Package counters */
2001 	sys.added_package_counters -= free_msr_counters_(&sys.pp);
2002 }
2003 
2004 struct system_summary {
2005 	struct thread_data threads;
2006 	struct core_data cores;
2007 	struct pkg_data packages;
2008 } average;
2009 
2010 struct platform_counters {
2011 	struct rapl_counter energy_psys;	/* MSR_PLATFORM_ENERGY_STATUS */
2012 } platform_counters_odd, platform_counters_even;
2013 
2014 struct cpu_topology {
2015 	int physical_package_id;
2016 	int die_id;
2017 	int logical_cpu_id;
2018 	int physical_node_id;
2019 	int logical_node_id;	/* 0-based count within the package */
2020 	int physical_core_id;
2021 	int thread_id;
2022 	int type;
2023 	cpu_set_t *put_ids;	/* Processing Unit/Thread IDs */
2024 } *cpus;
2025 
2026 struct topo_params {
2027 	int num_packages;
2028 	int num_die;
2029 	int num_cpus;
2030 	int num_cores;
2031 	int allowed_packages;
2032 	int allowed_cpus;
2033 	int allowed_cores;
2034 	int max_cpu_num;
2035 	int max_core_id;
2036 	int max_package_id;
2037 	int max_die_id;
2038 	int max_node_num;
2039 	int nodes_per_pkg;
2040 	int cores_per_node;
2041 	int threads_per_core;
2042 } topo;
2043 
2044 struct timeval tv_even, tv_odd, tv_delta;
2045 
2046 int *irq_column_2_cpu;		/* /proc/interrupts column numbers */
2047 int *irqs_per_cpu;		/* indexed by cpu_num */
2048 int *nmi_per_cpu;		/* indexed by cpu_num */
2049 
2050 void setup_all_buffers(bool startup);
2051 
2052 char *sys_lpi_file;
2053 char *sys_lpi_file_sysfs = "/sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us";
2054 char *sys_lpi_file_debugfs = "/sys/kernel/debug/pmc_core/slp_s0_residency_usec";
2055 
cpu_is_not_present(int cpu)2056 int cpu_is_not_present(int cpu)
2057 {
2058 	return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
2059 }
2060 
cpu_is_not_allowed(int cpu)2061 int cpu_is_not_allowed(int cpu)
2062 {
2063 	return !CPU_ISSET_S(cpu, cpu_allowed_setsize, cpu_allowed_set);
2064 }
2065 
2066 /*
2067  * run func(thread, core, package) in topology order
2068  * skip non-present cpus
2069  */
2070 
for_all_cpus(int (func)(struct thread_data *,struct core_data *,struct pkg_data *),struct thread_data * thread_base,struct core_data * core_base,struct pkg_data * pkg_base)2071 int for_all_cpus(int (func) (struct thread_data *, struct core_data *, struct pkg_data *),
2072 		 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
2073 {
2074 	int retval, pkg_no, core_no, thread_no, node_no;
2075 
2076 	retval = 0;
2077 
2078 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
2079 		for (node_no = 0; node_no < topo.nodes_per_pkg; node_no++) {
2080 			for (core_no = 0; core_no < topo.cores_per_node; ++core_no) {
2081 				for (thread_no = 0; thread_no < topo.threads_per_core; ++thread_no) {
2082 					struct thread_data *t;
2083 					struct core_data *c;
2084 					struct pkg_data *p;
2085 					t = GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no);
2086 
2087 					if (cpu_is_not_allowed(t->cpu_id))
2088 						continue;
2089 
2090 					c = GET_CORE(core_base, core_no, node_no, pkg_no);
2091 					p = GET_PKG(pkg_base, pkg_no);
2092 
2093 					retval |= func(t, c, p);
2094 				}
2095 			}
2096 		}
2097 	}
2098 	return retval;
2099 }
2100 
is_cpu_first_thread_in_core(struct thread_data * t,struct core_data * c,struct pkg_data * p)2101 int is_cpu_first_thread_in_core(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2102 {
2103 	UNUSED(p);
2104 
2105 	return ((int)t->cpu_id == c->base_cpu || c->base_cpu < 0);
2106 }
2107 
is_cpu_first_core_in_package(struct thread_data * t,struct core_data * c,struct pkg_data * p)2108 int is_cpu_first_core_in_package(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2109 {
2110 	UNUSED(c);
2111 
2112 	return ((int)t->cpu_id == p->base_cpu || p->base_cpu < 0);
2113 }
2114 
is_cpu_first_thread_in_package(struct thread_data * t,struct core_data * c,struct pkg_data * p)2115 int is_cpu_first_thread_in_package(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2116 {
2117 	return is_cpu_first_thread_in_core(t, c, p) && is_cpu_first_core_in_package(t, c, p);
2118 }
2119 
cpu_migrate(int cpu)2120 int cpu_migrate(int cpu)
2121 {
2122 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2123 	CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
2124 	if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
2125 		return -1;
2126 	else
2127 		return 0;
2128 }
2129 
get_msr_fd(int cpu)2130 int get_msr_fd(int cpu)
2131 {
2132 	char pathname[32];
2133 	int fd;
2134 
2135 	fd = fd_percpu[cpu];
2136 
2137 	if (fd)
2138 		return fd;
2139 
2140 	sprintf(pathname, "/dev/cpu/%d/msr", cpu);
2141 	fd = open(pathname, O_RDONLY);
2142 	if (fd < 0)
2143 		err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, "
2144 		    "or run with --no-msr, or run as root", pathname);
2145 
2146 	fd_percpu[cpu] = fd;
2147 
2148 	return fd;
2149 }
2150 
bic_disable_msr_access(void)2151 static void bic_disable_msr_access(void)
2152 {
2153 	const unsigned long bic_msrs = BIC_Mod_c6 | BIC_CoreTmp |
2154 	    BIC_Totl_c0 | BIC_Any_c0 | BIC_GFX_c0 | BIC_CPUGFX | BIC_PkgTmp;
2155 
2156 	bic_enabled &= ~bic_msrs;
2157 
2158 	free_sys_msr_counters();
2159 }
2160 
perf_event_open(struct perf_event_attr * hw_event,pid_t pid,int cpu,int group_fd,unsigned long flags)2161 static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid, int cpu, int group_fd, unsigned long flags)
2162 {
2163 	assert(!no_perf);
2164 
2165 	return syscall(__NR_perf_event_open, hw_event, pid, cpu, group_fd, flags);
2166 }
2167 
open_perf_counter(int cpu,unsigned int type,unsigned int config,int group_fd,__u64 read_format)2168 static long open_perf_counter(int cpu, unsigned int type, unsigned int config, int group_fd, __u64 read_format)
2169 {
2170 	struct perf_event_attr attr;
2171 	const pid_t pid = -1;
2172 	const unsigned long flags = 0;
2173 
2174 	assert(!no_perf);
2175 
2176 	memset(&attr, 0, sizeof(struct perf_event_attr));
2177 
2178 	attr.type = type;
2179 	attr.size = sizeof(struct perf_event_attr);
2180 	attr.config = config;
2181 	attr.disabled = 0;
2182 	attr.sample_type = PERF_SAMPLE_IDENTIFIER;
2183 	attr.read_format = read_format;
2184 
2185 	const int fd = perf_event_open(&attr, pid, cpu, group_fd, flags);
2186 
2187 	return fd;
2188 }
2189 
get_instr_count_fd(int cpu)2190 int get_instr_count_fd(int cpu)
2191 {
2192 	if (fd_instr_count_percpu[cpu])
2193 		return fd_instr_count_percpu[cpu];
2194 
2195 	fd_instr_count_percpu[cpu] = open_perf_counter(cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, -1, 0);
2196 
2197 	return fd_instr_count_percpu[cpu];
2198 }
2199 
get_msr(int cpu,off_t offset,unsigned long long * msr)2200 int get_msr(int cpu, off_t offset, unsigned long long *msr)
2201 {
2202 	ssize_t retval;
2203 
2204 	assert(!no_msr);
2205 
2206 	retval = pread(get_msr_fd(cpu), msr, sizeof(*msr), offset);
2207 
2208 	if (retval != sizeof *msr)
2209 		err(-1, "cpu%d: msr offset 0x%llx read failed", cpu, (unsigned long long)offset);
2210 
2211 	return 0;
2212 }
2213 
probe_msr(int cpu,off_t offset)2214 int probe_msr(int cpu, off_t offset)
2215 {
2216 	ssize_t retval;
2217 	unsigned long long value;
2218 
2219 	assert(!no_msr);
2220 
2221 	retval = pread(get_msr_fd(cpu), &value, sizeof(value), offset);
2222 
2223 	/*
2224 	 * Expect MSRs to accumulate some non-zero value since the system was powered on.
2225 	 * Treat zero as a read failure.
2226 	 */
2227 	if (retval != sizeof(value) || value == 0)
2228 		return 1;
2229 
2230 	return 0;
2231 }
2232 
2233 /* Convert CPU ID to domain ID for given added perf counter. */
cpu_to_domain(const struct perf_counter_info * pc,int cpu)2234 unsigned int cpu_to_domain(const struct perf_counter_info *pc, int cpu)
2235 {
2236 	switch (pc->scope) {
2237 	case SCOPE_CPU:
2238 		return cpu;
2239 
2240 	case SCOPE_CORE:
2241 		return cpus[cpu].physical_core_id;
2242 
2243 	case SCOPE_PACKAGE:
2244 		return cpus[cpu].physical_package_id;
2245 	}
2246 
2247 	__builtin_unreachable();
2248 }
2249 
2250 #define MAX_DEFERRED 16
2251 char *deferred_add_names[MAX_DEFERRED];
2252 char *deferred_skip_names[MAX_DEFERRED];
2253 int deferred_add_index;
2254 int deferred_skip_index;
2255 
2256 /*
2257  * HIDE_LIST - hide this list of counters, show the rest [default]
2258  * SHOW_LIST - show this list of counters, hide the rest
2259  */
2260 enum show_hide_mode { SHOW_LIST, HIDE_LIST } global_show_hide_mode = HIDE_LIST;
2261 
help(void)2262 void help(void)
2263 {
2264 	fprintf(outf,
2265 		"Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
2266 		"\n"
2267 		"Turbostat forks the specified COMMAND and prints statistics\n"
2268 		"when COMMAND completes.\n"
2269 		"If no COMMAND is specified, turbostat wakes every 5-seconds\n"
2270 		"to print statistics, until interrupted.\n"
2271 		"  -a, --add counter\n"
2272 		"		add a counter\n"
2273 		"		  eg. --add msr0x10,u64,cpu,delta,MY_TSC\n"
2274 		"		  eg. --add perf/cstate_pkg/c2-residency,package,delta,percent,perfPC2\n"
2275 		"		  eg. --add pmt,name=XTAL,type=raw,domain=package0,offset=0,lsb=0,msb=63,guid=0x1a067102\n"
2276 		"  -c, --cpu cpu-set\n"
2277 		"		limit output to summary plus cpu-set:\n"
2278 		"		  {core | package | j,k,l..m,n-p }\n"
2279 		"  -d, --debug\n"
2280 		"		displays usec, Time_Of_Day_Seconds and more debugging\n"
2281 		"		debug messages are printed to stderr\n"
2282 		"  -D, --Dump\n"
2283 		"		displays the raw counter values\n"
2284 		"  -e, --enable [all | column]\n"
2285 		"		shows all or the specified disabled column\n"
2286 		"  -f, --force\n"
2287 		"		force load turbostat with minimum default features on unsupported platforms.\n"
2288 		"  -H, --hide [column | column,column,...]\n"
2289 		"		hide the specified column(s)\n"
2290 		"  -i, --interval sec.subsec\n"
2291 		"		override default 5-second measurement interval\n"
2292 		"  -J, --Joules\n"
2293 		"		displays energy in Joules instead of Watts\n"
2294 		"  -l, --list\n"
2295 		"		list column headers only\n"
2296 		"  -M, --no-msr\n"
2297 		"		disable all uses of the MSR driver\n"
2298 		"  -P, --no-perf\n"
2299 		"		disable all uses of the perf API\n"
2300 		"  -n, --num_iterations num\n"
2301 		"		number of the measurement iterations\n"
2302 		"  -N, --header_iterations num\n"
2303 		"		print header every num iterations\n"
2304 		"  -o, --out file\n"
2305 		"		create or truncate \"file\" for all output\n"
2306 		"  -q, --quiet\n"
2307 		"		skip decoding system configuration header\n"
2308 		"  -s, --show [column | column,column,...]\n"
2309 		"		show only the specified column(s)\n"
2310 		"  -S, --Summary\n"
2311 		"		limits output to 1-line system summary per interval\n"
2312 		"  -T, --TCC temperature\n"
2313 		"		sets the Thermal Control Circuit temperature in\n"
2314 		"		  degrees Celsius\n"
2315 		"  -h, --help\n"
2316 		"		print this help message\n"
2317 		"  -v, --version\n"
2318 		"		print version information\n\nFor more help, run \"man turbostat\"\n");
2319 }
2320 
2321 /*
2322  * bic_lookup
2323  * for all the strings in comma separate name_list,
2324  * set the approprate bit in return value.
2325  */
bic_lookup(char * name_list,enum show_hide_mode mode)2326 unsigned long long bic_lookup(char *name_list, enum show_hide_mode mode)
2327 {
2328 	unsigned int i;
2329 	unsigned long long retval = 0;
2330 
2331 	while (name_list) {
2332 		char *comma;
2333 
2334 		comma = strchr(name_list, ',');
2335 
2336 		if (comma)
2337 			*comma = '\0';
2338 
2339 		for (i = 0; i < MAX_BIC; ++i) {
2340 			if (!strcmp(name_list, bic[i].name)) {
2341 				retval |= (1ULL << i);
2342 				break;
2343 			}
2344 			if (!strcmp(name_list, "all")) {
2345 				retval |= ~0;
2346 				break;
2347 			} else if (!strcmp(name_list, "topology")) {
2348 				retval |= BIC_TOPOLOGY;
2349 				break;
2350 			} else if (!strcmp(name_list, "power")) {
2351 				retval |= BIC_THERMAL_PWR;
2352 				break;
2353 			} else if (!strcmp(name_list, "idle")) {
2354 				retval |= BIC_IDLE;
2355 				break;
2356 			} else if (!strcmp(name_list, "frequency")) {
2357 				retval |= BIC_FREQUENCY;
2358 				break;
2359 			} else if (!strcmp(name_list, "other")) {
2360 				retval |= BIC_OTHER;
2361 				break;
2362 			}
2363 
2364 		}
2365 		if (i == MAX_BIC) {
2366 			if (mode == SHOW_LIST) {
2367 				deferred_add_names[deferred_add_index++] = name_list;
2368 				if (deferred_add_index >= MAX_DEFERRED) {
2369 					fprintf(stderr, "More than max %d un-recognized --add options '%s'\n",
2370 						MAX_DEFERRED, name_list);
2371 					help();
2372 					exit(1);
2373 				}
2374 			} else {
2375 				deferred_skip_names[deferred_skip_index++] = name_list;
2376 				if (debug)
2377 					fprintf(stderr, "deferred \"%s\"\n", name_list);
2378 				if (deferred_skip_index >= MAX_DEFERRED) {
2379 					fprintf(stderr, "More than max %d un-recognized --skip options '%s'\n",
2380 						MAX_DEFERRED, name_list);
2381 					help();
2382 					exit(1);
2383 				}
2384 			}
2385 		}
2386 
2387 		name_list = comma;
2388 		if (name_list)
2389 			name_list++;
2390 
2391 	}
2392 	return retval;
2393 }
2394 
print_header(char * delim)2395 void print_header(char *delim)
2396 {
2397 	struct msr_counter *mp;
2398 	struct perf_counter_info *pp;
2399 	struct pmt_counter *ppmt;
2400 	int printed = 0;
2401 
2402 	if (DO_BIC(BIC_USEC))
2403 		outp += sprintf(outp, "%susec", (printed++ ? delim : ""));
2404 	if (DO_BIC(BIC_TOD))
2405 		outp += sprintf(outp, "%sTime_Of_Day_Seconds", (printed++ ? delim : ""));
2406 	if (DO_BIC(BIC_Package))
2407 		outp += sprintf(outp, "%sPackage", (printed++ ? delim : ""));
2408 	if (DO_BIC(BIC_Die))
2409 		outp += sprintf(outp, "%sDie", (printed++ ? delim : ""));
2410 	if (DO_BIC(BIC_Node))
2411 		outp += sprintf(outp, "%sNode", (printed++ ? delim : ""));
2412 	if (DO_BIC(BIC_Core))
2413 		outp += sprintf(outp, "%sCore", (printed++ ? delim : ""));
2414 	if (DO_BIC(BIC_CPU))
2415 		outp += sprintf(outp, "%sCPU", (printed++ ? delim : ""));
2416 	if (DO_BIC(BIC_APIC))
2417 		outp += sprintf(outp, "%sAPIC", (printed++ ? delim : ""));
2418 	if (DO_BIC(BIC_X2APIC))
2419 		outp += sprintf(outp, "%sX2APIC", (printed++ ? delim : ""));
2420 	if (DO_BIC(BIC_Avg_MHz))
2421 		outp += sprintf(outp, "%sAvg_MHz", (printed++ ? delim : ""));
2422 	if (DO_BIC(BIC_Busy))
2423 		outp += sprintf(outp, "%sBusy%%", (printed++ ? delim : ""));
2424 	if (DO_BIC(BIC_Bzy_MHz))
2425 		outp += sprintf(outp, "%sBzy_MHz", (printed++ ? delim : ""));
2426 	if (DO_BIC(BIC_TSC_MHz))
2427 		outp += sprintf(outp, "%sTSC_MHz", (printed++ ? delim : ""));
2428 
2429 	if (DO_BIC(BIC_IPC))
2430 		outp += sprintf(outp, "%sIPC", (printed++ ? delim : ""));
2431 
2432 	if (DO_BIC(BIC_IRQ)) {
2433 		if (sums_need_wide_columns)
2434 			outp += sprintf(outp, "%s     IRQ", (printed++ ? delim : ""));
2435 		else
2436 			outp += sprintf(outp, "%sIRQ", (printed++ ? delim : ""));
2437 	}
2438 	if (DO_BIC(BIC_NMI)) {
2439 		if (sums_need_wide_columns)
2440 			outp += sprintf(outp, "%s     NMI", (printed++ ? delim : ""));
2441 		else
2442 			outp += sprintf(outp, "%sNMI", (printed++ ? delim : ""));
2443 	}
2444 
2445 	if (DO_BIC(BIC_SMI))
2446 		outp += sprintf(outp, "%sSMI", (printed++ ? delim : ""));
2447 
2448 	for (mp = sys.tp; mp; mp = mp->next) {
2449 
2450 		if (mp->format == FORMAT_RAW) {
2451 			if (mp->width == 64)
2452 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), mp->name);
2453 			else
2454 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), mp->name);
2455 		} else {
2456 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2457 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), mp->name);
2458 			else
2459 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), mp->name);
2460 		}
2461 	}
2462 
2463 	for (pp = sys.perf_tp; pp; pp = pp->next) {
2464 
2465 		if (pp->format == FORMAT_RAW) {
2466 			if (pp->width == 64)
2467 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), pp->name);
2468 			else
2469 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), pp->name);
2470 		} else {
2471 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2472 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), pp->name);
2473 			else
2474 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), pp->name);
2475 		}
2476 	}
2477 
2478 	ppmt = sys.pmt_tp;
2479 	while (ppmt) {
2480 		switch (ppmt->type) {
2481 		case PMT_TYPE_RAW:
2482 			if (pmt_counter_get_width(ppmt) <= 32)
2483 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), ppmt->name);
2484 			else
2485 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), ppmt->name);
2486 
2487 			break;
2488 
2489 		case PMT_TYPE_XTAL_TIME:
2490 		case PMT_TYPE_TCORE_CLOCK:
2491 			outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), ppmt->name);
2492 			break;
2493 		}
2494 
2495 		ppmt = ppmt->next;
2496 	}
2497 
2498 	if (DO_BIC(BIC_CPU_c1))
2499 		outp += sprintf(outp, "%sCPU%%c1", (printed++ ? delim : ""));
2500 	if (DO_BIC(BIC_CPU_c3))
2501 		outp += sprintf(outp, "%sCPU%%c3", (printed++ ? delim : ""));
2502 	if (DO_BIC(BIC_CPU_c6))
2503 		outp += sprintf(outp, "%sCPU%%c6", (printed++ ? delim : ""));
2504 	if (DO_BIC(BIC_CPU_c7))
2505 		outp += sprintf(outp, "%sCPU%%c7", (printed++ ? delim : ""));
2506 
2507 	if (DO_BIC(BIC_Mod_c6))
2508 		outp += sprintf(outp, "%sMod%%c6", (printed++ ? delim : ""));
2509 
2510 	if (DO_BIC(BIC_CoreTmp))
2511 		outp += sprintf(outp, "%sCoreTmp", (printed++ ? delim : ""));
2512 
2513 	if (DO_BIC(BIC_CORE_THROT_CNT))
2514 		outp += sprintf(outp, "%sCoreThr", (printed++ ? delim : ""));
2515 
2516 	if (platform->rapl_msrs && !rapl_joules) {
2517 		if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
2518 			outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
2519 	} else if (platform->rapl_msrs && rapl_joules) {
2520 		if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
2521 			outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
2522 	}
2523 
2524 	for (mp = sys.cp; mp; mp = mp->next) {
2525 		if (mp->format == FORMAT_RAW) {
2526 			if (mp->width == 64)
2527 				outp += sprintf(outp, "%s%18.18s", delim, mp->name);
2528 			else
2529 				outp += sprintf(outp, "%s%10.10s", delim, mp->name);
2530 		} else {
2531 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2532 				outp += sprintf(outp, "%s%8s", delim, mp->name);
2533 			else
2534 				outp += sprintf(outp, "%s%s", delim, mp->name);
2535 		}
2536 	}
2537 
2538 	for (pp = sys.perf_cp; pp; pp = pp->next) {
2539 
2540 		if (pp->format == FORMAT_RAW) {
2541 			if (pp->width == 64)
2542 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), pp->name);
2543 			else
2544 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), pp->name);
2545 		} else {
2546 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2547 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), pp->name);
2548 			else
2549 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), pp->name);
2550 		}
2551 	}
2552 
2553 	ppmt = sys.pmt_cp;
2554 	while (ppmt) {
2555 		switch (ppmt->type) {
2556 		case PMT_TYPE_RAW:
2557 			if (pmt_counter_get_width(ppmt) <= 32)
2558 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), ppmt->name);
2559 			else
2560 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), ppmt->name);
2561 
2562 			break;
2563 
2564 		case PMT_TYPE_XTAL_TIME:
2565 		case PMT_TYPE_TCORE_CLOCK:
2566 			outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), ppmt->name);
2567 			break;
2568 		}
2569 
2570 		ppmt = ppmt->next;
2571 	}
2572 
2573 	if (DO_BIC(BIC_PkgTmp))
2574 		outp += sprintf(outp, "%sPkgTmp", (printed++ ? delim : ""));
2575 
2576 	if (DO_BIC(BIC_GFX_rc6))
2577 		outp += sprintf(outp, "%sGFX%%rc6", (printed++ ? delim : ""));
2578 
2579 	if (DO_BIC(BIC_GFXMHz))
2580 		outp += sprintf(outp, "%sGFXMHz", (printed++ ? delim : ""));
2581 
2582 	if (DO_BIC(BIC_GFXACTMHz))
2583 		outp += sprintf(outp, "%sGFXAMHz", (printed++ ? delim : ""));
2584 
2585 	if (DO_BIC(BIC_SAM_mc6))
2586 		outp += sprintf(outp, "%sSAM%%mc6", (printed++ ? delim : ""));
2587 
2588 	if (DO_BIC(BIC_SAMMHz))
2589 		outp += sprintf(outp, "%sSAMMHz", (printed++ ? delim : ""));
2590 
2591 	if (DO_BIC(BIC_SAMACTMHz))
2592 		outp += sprintf(outp, "%sSAMAMHz", (printed++ ? delim : ""));
2593 
2594 	if (DO_BIC(BIC_Totl_c0))
2595 		outp += sprintf(outp, "%sTotl%%C0", (printed++ ? delim : ""));
2596 	if (DO_BIC(BIC_Any_c0))
2597 		outp += sprintf(outp, "%sAny%%C0", (printed++ ? delim : ""));
2598 	if (DO_BIC(BIC_GFX_c0))
2599 		outp += sprintf(outp, "%sGFX%%C0", (printed++ ? delim : ""));
2600 	if (DO_BIC(BIC_CPUGFX))
2601 		outp += sprintf(outp, "%sCPUGFX%%", (printed++ ? delim : ""));
2602 
2603 	if (DO_BIC(BIC_Pkgpc2))
2604 		outp += sprintf(outp, "%sPkg%%pc2", (printed++ ? delim : ""));
2605 	if (DO_BIC(BIC_Pkgpc3))
2606 		outp += sprintf(outp, "%sPkg%%pc3", (printed++ ? delim : ""));
2607 	if (DO_BIC(BIC_Pkgpc6))
2608 		outp += sprintf(outp, "%sPkg%%pc6", (printed++ ? delim : ""));
2609 	if (DO_BIC(BIC_Pkgpc7))
2610 		outp += sprintf(outp, "%sPkg%%pc7", (printed++ ? delim : ""));
2611 	if (DO_BIC(BIC_Pkgpc8))
2612 		outp += sprintf(outp, "%sPkg%%pc8", (printed++ ? delim : ""));
2613 	if (DO_BIC(BIC_Pkgpc9))
2614 		outp += sprintf(outp, "%sPkg%%pc9", (printed++ ? delim : ""));
2615 	if (DO_BIC(BIC_Pkgpc10))
2616 		outp += sprintf(outp, "%sPk%%pc10", (printed++ ? delim : ""));
2617 	if (DO_BIC(BIC_Diec6))
2618 		outp += sprintf(outp, "%sDie%%c6", (printed++ ? delim : ""));
2619 	if (DO_BIC(BIC_CPU_LPI))
2620 		outp += sprintf(outp, "%sCPU%%LPI", (printed++ ? delim : ""));
2621 	if (DO_BIC(BIC_SYS_LPI))
2622 		outp += sprintf(outp, "%sSYS%%LPI", (printed++ ? delim : ""));
2623 
2624 	if (platform->rapl_msrs && !rapl_joules) {
2625 		if (DO_BIC(BIC_PkgWatt))
2626 			outp += sprintf(outp, "%sPkgWatt", (printed++ ? delim : ""));
2627 		if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
2628 			outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
2629 		if (DO_BIC(BIC_GFXWatt))
2630 			outp += sprintf(outp, "%sGFXWatt", (printed++ ? delim : ""));
2631 		if (DO_BIC(BIC_RAMWatt))
2632 			outp += sprintf(outp, "%sRAMWatt", (printed++ ? delim : ""));
2633 		if (DO_BIC(BIC_PKG__))
2634 			outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
2635 		if (DO_BIC(BIC_RAM__))
2636 			outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
2637 	} else if (platform->rapl_msrs && rapl_joules) {
2638 		if (DO_BIC(BIC_Pkg_J))
2639 			outp += sprintf(outp, "%sPkg_J", (printed++ ? delim : ""));
2640 		if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
2641 			outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
2642 		if (DO_BIC(BIC_GFX_J))
2643 			outp += sprintf(outp, "%sGFX_J", (printed++ ? delim : ""));
2644 		if (DO_BIC(BIC_RAM_J))
2645 			outp += sprintf(outp, "%sRAM_J", (printed++ ? delim : ""));
2646 		if (DO_BIC(BIC_PKG__))
2647 			outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
2648 		if (DO_BIC(BIC_RAM__))
2649 			outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
2650 	}
2651 	if (DO_BIC(BIC_UNCORE_MHZ))
2652 		outp += sprintf(outp, "%sUncMHz", (printed++ ? delim : ""));
2653 
2654 	for (mp = sys.pp; mp; mp = mp->next) {
2655 		if (mp->format == FORMAT_RAW) {
2656 			if (mp->width == 64)
2657 				outp += sprintf(outp, "%s%18.18s", delim, mp->name);
2658 			else if (mp->width == 32)
2659 				outp += sprintf(outp, "%s%10.10s", delim, mp->name);
2660 			else
2661 				outp += sprintf(outp, "%s%7.7s", delim, mp->name);
2662 		} else {
2663 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2664 				outp += sprintf(outp, "%s%8s", delim, mp->name);
2665 			else
2666 				outp += sprintf(outp, "%s%7.7s", delim, mp->name);
2667 		}
2668 	}
2669 
2670 	for (pp = sys.perf_pp; pp; pp = pp->next) {
2671 
2672 		if (pp->format == FORMAT_RAW) {
2673 			if (pp->width == 64)
2674 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), pp->name);
2675 			else
2676 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), pp->name);
2677 		} else {
2678 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2679 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), pp->name);
2680 			else
2681 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), pp->name);
2682 		}
2683 	}
2684 
2685 	ppmt = sys.pmt_pp;
2686 	while (ppmt) {
2687 		switch (ppmt->type) {
2688 		case PMT_TYPE_RAW:
2689 			if (pmt_counter_get_width(ppmt) <= 32)
2690 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), ppmt->name);
2691 			else
2692 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), ppmt->name);
2693 
2694 			break;
2695 
2696 		case PMT_TYPE_XTAL_TIME:
2697 		case PMT_TYPE_TCORE_CLOCK:
2698 			outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), ppmt->name);
2699 			break;
2700 		}
2701 
2702 		ppmt = ppmt->next;
2703 	}
2704 
2705 	if (DO_BIC(BIC_SysWatt))
2706 		outp += sprintf(outp, "%sSysWatt", (printed++ ? delim : ""));
2707 	if (DO_BIC(BIC_Sys_J))
2708 		outp += sprintf(outp, "%sSys_J", (printed++ ? delim : ""));
2709 
2710 	outp += sprintf(outp, "\n");
2711 }
2712 
dump_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)2713 int dump_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2714 {
2715 	int i;
2716 	struct msr_counter *mp;
2717 	struct platform_counters *pplat_cnt = p == package_odd ? &platform_counters_odd : &platform_counters_even;
2718 
2719 	outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
2720 
2721 	if (t) {
2722 		outp += sprintf(outp, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags);
2723 		outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
2724 		outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
2725 		outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
2726 		outp += sprintf(outp, "c1: %016llX\n", t->c1);
2727 
2728 		if (DO_BIC(BIC_IPC))
2729 			outp += sprintf(outp, "IPC: %lld\n", t->instr_count);
2730 
2731 		if (DO_BIC(BIC_IRQ))
2732 			outp += sprintf(outp, "IRQ: %lld\n", t->irq_count);
2733 		if (DO_BIC(BIC_NMI))
2734 			outp += sprintf(outp, "IRQ: %lld\n", t->nmi_count);
2735 		if (DO_BIC(BIC_SMI))
2736 			outp += sprintf(outp, "SMI: %d\n", t->smi_count);
2737 
2738 		for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
2739 			outp +=
2740 			    sprintf(outp, "tADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num,
2741 				    t->counter[i], mp->sp->path);
2742 		}
2743 	}
2744 
2745 	if (c && is_cpu_first_thread_in_core(t, c, p)) {
2746 		outp += sprintf(outp, "core: %d\n", c->core_id);
2747 		outp += sprintf(outp, "c3: %016llX\n", c->c3);
2748 		outp += sprintf(outp, "c6: %016llX\n", c->c6);
2749 		outp += sprintf(outp, "c7: %016llX\n", c->c7);
2750 		outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
2751 		outp += sprintf(outp, "cpu_throt_count: %016llX\n", c->core_throt_cnt);
2752 
2753 		const unsigned long long energy_value = c->core_energy.raw_value * c->core_energy.scale;
2754 		const double energy_scale = c->core_energy.scale;
2755 
2756 		if (c->core_energy.unit == RAPL_UNIT_JOULES)
2757 			outp += sprintf(outp, "Joules: %0llX (scale: %lf)\n", energy_value, energy_scale);
2758 
2759 		for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
2760 			outp +=
2761 			    sprintf(outp, "cADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num,
2762 				    c->counter[i], mp->sp->path);
2763 		}
2764 		outp += sprintf(outp, "mc6_us: %016llX\n", c->mc6_us);
2765 	}
2766 
2767 	if (p && is_cpu_first_core_in_package(t, c, p)) {
2768 		outp += sprintf(outp, "package: %d\n", p->package_id);
2769 
2770 		outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
2771 		outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
2772 		outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
2773 		outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
2774 
2775 		outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
2776 		if (DO_BIC(BIC_Pkgpc3))
2777 			outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
2778 		if (DO_BIC(BIC_Pkgpc6))
2779 			outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
2780 		if (DO_BIC(BIC_Pkgpc7))
2781 			outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
2782 		outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
2783 		outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
2784 		outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
2785 		outp += sprintf(outp, "cpu_lpi: %016llX\n", p->cpu_lpi);
2786 		outp += sprintf(outp, "sys_lpi: %016llX\n", p->sys_lpi);
2787 		outp += sprintf(outp, "Joules PKG: %0llX\n", p->energy_pkg.raw_value);
2788 		outp += sprintf(outp, "Joules COR: %0llX\n", p->energy_cores.raw_value);
2789 		outp += sprintf(outp, "Joules GFX: %0llX\n", p->energy_gfx.raw_value);
2790 		outp += sprintf(outp, "Joules RAM: %0llX\n", p->energy_dram.raw_value);
2791 		outp += sprintf(outp, "Joules PSYS: %0llX\n", pplat_cnt->energy_psys.raw_value);
2792 		outp += sprintf(outp, "Throttle PKG: %0llX\n", p->rapl_pkg_perf_status.raw_value);
2793 		outp += sprintf(outp, "Throttle RAM: %0llX\n", p->rapl_dram_perf_status.raw_value);
2794 		outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
2795 
2796 		for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
2797 			outp +=
2798 			    sprintf(outp, "pADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num,
2799 				    p->counter[i], mp->sp->path);
2800 		}
2801 	}
2802 
2803 	outp += sprintf(outp, "\n");
2804 
2805 	return 0;
2806 }
2807 
rapl_counter_get_value(const struct rapl_counter * c,enum rapl_unit desired_unit,double interval)2808 double rapl_counter_get_value(const struct rapl_counter *c, enum rapl_unit desired_unit, double interval)
2809 {
2810 	assert(desired_unit != RAPL_UNIT_INVALID);
2811 
2812 	/*
2813 	 * For now we don't expect anything other than joules,
2814 	 * so just simplify the logic.
2815 	 */
2816 	assert(c->unit == RAPL_UNIT_JOULES);
2817 
2818 	const double scaled = c->raw_value * c->scale;
2819 
2820 	if (desired_unit == RAPL_UNIT_WATTS)
2821 		return scaled / interval;
2822 	return scaled;
2823 }
2824 
2825 /*
2826  * column formatting convention & formats
2827  */
format_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)2828 int format_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2829 {
2830 	static int count;
2831 
2832 	struct platform_counters *pplat_cnt = NULL;
2833 	double interval_float, tsc;
2834 	char *fmt8;
2835 	int i;
2836 	struct msr_counter *mp;
2837 	struct perf_counter_info *pp;
2838 	struct pmt_counter *ppmt;
2839 	char *delim = "\t";
2840 	int printed = 0;
2841 
2842 	if (t == &average.threads) {
2843 		pplat_cnt = count & 1 ? &platform_counters_odd : &platform_counters_even;
2844 		++count;
2845 	}
2846 
2847 	/* if showing only 1st thread in core and this isn't one, bail out */
2848 	if (show_core_only && !is_cpu_first_thread_in_core(t, c, p))
2849 		return 0;
2850 
2851 	/* if showing only 1st thread in pkg and this isn't one, bail out */
2852 	if (show_pkg_only && !is_cpu_first_core_in_package(t, c, p))
2853 		return 0;
2854 
2855 	/*if not summary line and --cpu is used */
2856 	if ((t != &average.threads) && (cpu_subset && !CPU_ISSET_S(t->cpu_id, cpu_subset_size, cpu_subset)))
2857 		return 0;
2858 
2859 	if (DO_BIC(BIC_USEC)) {
2860 		/* on each row, print how many usec each timestamp took to gather */
2861 		struct timeval tv;
2862 
2863 		timersub(&t->tv_end, &t->tv_begin, &tv);
2864 		outp += sprintf(outp, "%5ld\t", tv.tv_sec * 1000000 + tv.tv_usec);
2865 	}
2866 
2867 	/* Time_Of_Day_Seconds: on each row, print sec.usec last timestamp taken */
2868 	if (DO_BIC(BIC_TOD))
2869 		outp += sprintf(outp, "%10ld.%06ld\t", t->tv_end.tv_sec, t->tv_end.tv_usec);
2870 
2871 	interval_float = t->tv_delta.tv_sec + t->tv_delta.tv_usec / 1000000.0;
2872 
2873 	tsc = t->tsc * tsc_tweak;
2874 
2875 	/* topo columns, print blanks on 1st (average) line */
2876 	if (t == &average.threads) {
2877 		if (DO_BIC(BIC_Package))
2878 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2879 		if (DO_BIC(BIC_Die))
2880 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2881 		if (DO_BIC(BIC_Node))
2882 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2883 		if (DO_BIC(BIC_Core))
2884 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2885 		if (DO_BIC(BIC_CPU))
2886 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2887 		if (DO_BIC(BIC_APIC))
2888 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2889 		if (DO_BIC(BIC_X2APIC))
2890 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2891 	} else {
2892 		if (DO_BIC(BIC_Package)) {
2893 			if (p)
2894 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->package_id);
2895 			else
2896 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2897 		}
2898 		if (DO_BIC(BIC_Die)) {
2899 			if (c)
2900 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].die_id);
2901 			else
2902 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2903 		}
2904 		if (DO_BIC(BIC_Node)) {
2905 			if (t)
2906 				outp += sprintf(outp, "%s%d",
2907 						(printed++ ? delim : ""), cpus[t->cpu_id].physical_node_id);
2908 			else
2909 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2910 		}
2911 		if (DO_BIC(BIC_Core)) {
2912 			if (c)
2913 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_id);
2914 			else
2915 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2916 		}
2917 		if (DO_BIC(BIC_CPU))
2918 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->cpu_id);
2919 		if (DO_BIC(BIC_APIC))
2920 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->apic_id);
2921 		if (DO_BIC(BIC_X2APIC))
2922 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->x2apic_id);
2923 	}
2924 
2925 	if (DO_BIC(BIC_Avg_MHz))
2926 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 / units * t->aperf / interval_float);
2927 
2928 	if (DO_BIC(BIC_Busy))
2929 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->mperf / tsc);
2930 
2931 	if (DO_BIC(BIC_Bzy_MHz)) {
2932 		if (has_base_hz)
2933 			outp +=
2934 			    sprintf(outp, "%s%.0f", (printed++ ? delim : ""), base_hz / units * t->aperf / t->mperf);
2935 		else
2936 			outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""),
2937 					tsc / units * t->aperf / t->mperf / interval_float);
2938 	}
2939 
2940 	if (DO_BIC(BIC_TSC_MHz))
2941 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 * t->tsc / units / interval_float);
2942 
2943 	if (DO_BIC(BIC_IPC))
2944 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 1.0 * t->instr_count / t->aperf);
2945 
2946 	/* IRQ */
2947 	if (DO_BIC(BIC_IRQ)) {
2948 		if (sums_need_wide_columns)
2949 			outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->irq_count);
2950 		else
2951 			outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->irq_count);
2952 	}
2953 
2954 	/* NMI */
2955 	if (DO_BIC(BIC_NMI)) {
2956 		if (sums_need_wide_columns)
2957 			outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->nmi_count);
2958 		else
2959 			outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->nmi_count);
2960 	}
2961 
2962 	/* SMI */
2963 	if (DO_BIC(BIC_SMI))
2964 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->smi_count);
2965 
2966 	/* Added counters */
2967 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
2968 		if (mp->format == FORMAT_RAW) {
2969 			if (mp->width == 32)
2970 				outp +=
2971 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int)t->counter[i]);
2972 			else
2973 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->counter[i]);
2974 		} else if (mp->format == FORMAT_DELTA) {
2975 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2976 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->counter[i]);
2977 			else
2978 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->counter[i]);
2979 		} else if (mp->format == FORMAT_PERCENT) {
2980 			if (mp->type == COUNTER_USEC)
2981 				outp +=
2982 				    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2983 					    t->counter[i] / interval_float / 10000);
2984 			else
2985 				outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->counter[i] / tsc);
2986 		}
2987 	}
2988 
2989 	/* Added perf counters */
2990 	for (i = 0, pp = sys.perf_tp; pp; ++i, pp = pp->next) {
2991 		if (pp->format == FORMAT_RAW) {
2992 			if (pp->width == 32)
2993 				outp +=
2994 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
2995 					    (unsigned int)t->perf_counter[i]);
2996 			else
2997 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->perf_counter[i]);
2998 		} else if (pp->format == FORMAT_DELTA) {
2999 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
3000 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->perf_counter[i]);
3001 			else
3002 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->perf_counter[i]);
3003 		} else if (pp->format == FORMAT_PERCENT) {
3004 			if (pp->type == COUNTER_USEC)
3005 				outp +=
3006 				    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
3007 					    t->perf_counter[i] / interval_float / 10000);
3008 			else
3009 				outp +=
3010 				    sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->perf_counter[i] / tsc);
3011 		}
3012 	}
3013 
3014 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3015 		const unsigned long value_raw = t->pmt_counter[i];
3016 		double value_converted;
3017 		switch (ppmt->type) {
3018 		case PMT_TYPE_RAW:
3019 			if (pmt_counter_get_width(ppmt) <= 32)
3020 				outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
3021 						(unsigned int)t->pmt_counter[i]);
3022 			else
3023 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->pmt_counter[i]);
3024 
3025 			break;
3026 
3027 		case PMT_TYPE_XTAL_TIME:
3028 			value_converted = 100.0 * value_raw / crystal_hz / interval_float;
3029 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
3030 			break;
3031 
3032 		case PMT_TYPE_TCORE_CLOCK:
3033 			value_converted = 100.0 * value_raw / tcore_clock_freq_hz / interval_float;
3034 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
3035 		}
3036 	}
3037 
3038 	/* C1 */
3039 	if (DO_BIC(BIC_CPU_c1))
3040 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->c1 / tsc);
3041 
3042 	/* print per-core data only for 1st thread in core */
3043 	if (!is_cpu_first_thread_in_core(t, c, p))
3044 		goto done;
3045 
3046 	if (DO_BIC(BIC_CPU_c3))
3047 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c3 / tsc);
3048 	if (DO_BIC(BIC_CPU_c6))
3049 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c6 / tsc);
3050 	if (DO_BIC(BIC_CPU_c7))
3051 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c7 / tsc);
3052 
3053 	/* Mod%c6 */
3054 	if (DO_BIC(BIC_Mod_c6))
3055 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->mc6_us / tsc);
3056 
3057 	if (DO_BIC(BIC_CoreTmp))
3058 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_temp_c);
3059 
3060 	/* Core throttle count */
3061 	if (DO_BIC(BIC_CORE_THROT_CNT))
3062 		outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->core_throt_cnt);
3063 
3064 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3065 		if (mp->format == FORMAT_RAW) {
3066 			if (mp->width == 32)
3067 				outp +=
3068 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int)c->counter[i]);
3069 			else
3070 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->counter[i]);
3071 		} else if (mp->format == FORMAT_DELTA) {
3072 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
3073 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), c->counter[i]);
3074 			else
3075 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->counter[i]);
3076 		} else if (mp->format == FORMAT_PERCENT) {
3077 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->counter[i] / tsc);
3078 		}
3079 	}
3080 
3081 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3082 		if (pp->format == FORMAT_RAW) {
3083 			if (pp->width == 32)
3084 				outp +=
3085 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
3086 					    (unsigned int)c->perf_counter[i]);
3087 			else
3088 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->perf_counter[i]);
3089 		} else if (pp->format == FORMAT_DELTA) {
3090 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
3091 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), c->perf_counter[i]);
3092 			else
3093 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->perf_counter[i]);
3094 		} else if (pp->format == FORMAT_PERCENT) {
3095 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->perf_counter[i] / tsc);
3096 		}
3097 	}
3098 
3099 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3100 		const unsigned long value_raw = c->pmt_counter[i];
3101 		double value_converted;
3102 		switch (ppmt->type) {
3103 		case PMT_TYPE_RAW:
3104 			if (pmt_counter_get_width(ppmt) <= 32)
3105 				outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
3106 						(unsigned int)c->pmt_counter[i]);
3107 			else
3108 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->pmt_counter[i]);
3109 
3110 			break;
3111 
3112 		case PMT_TYPE_XTAL_TIME:
3113 			value_converted = 100.0 * value_raw / crystal_hz / interval_float;
3114 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
3115 			break;
3116 
3117 		case PMT_TYPE_TCORE_CLOCK:
3118 			value_converted = 100.0 * value_raw / tcore_clock_freq_hz / interval_float;
3119 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
3120 		}
3121 	}
3122 
3123 	fmt8 = "%s%.2f";
3124 
3125 	if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
3126 		outp +=
3127 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3128 			    rapl_counter_get_value(&c->core_energy, RAPL_UNIT_WATTS, interval_float));
3129 	if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
3130 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3131 				rapl_counter_get_value(&c->core_energy, RAPL_UNIT_JOULES, interval_float));
3132 
3133 	/* print per-package data only for 1st core in package */
3134 	if (!is_cpu_first_core_in_package(t, c, p))
3135 		goto done;
3136 
3137 	/* PkgTmp */
3138 	if (DO_BIC(BIC_PkgTmp))
3139 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->pkg_temp_c);
3140 
3141 	/* GFXrc6 */
3142 	if (DO_BIC(BIC_GFX_rc6)) {
3143 		if (p->gfx_rc6_ms == -1) {	/* detect GFX counter reset */
3144 			outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
3145 		} else {
3146 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
3147 					p->gfx_rc6_ms / 10.0 / interval_float);
3148 		}
3149 	}
3150 
3151 	/* GFXMHz */
3152 	if (DO_BIC(BIC_GFXMHz))
3153 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_mhz);
3154 
3155 	/* GFXACTMHz */
3156 	if (DO_BIC(BIC_GFXACTMHz))
3157 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_act_mhz);
3158 
3159 	/* SAMmc6 */
3160 	if (DO_BIC(BIC_SAM_mc6)) {
3161 		if (p->sam_mc6_ms == -1) {	/* detect GFX counter reset */
3162 			outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
3163 		} else {
3164 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
3165 					p->sam_mc6_ms / 10.0 / interval_float);
3166 		}
3167 	}
3168 
3169 	/* SAMMHz */
3170 	if (DO_BIC(BIC_SAMMHz))
3171 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->sam_mhz);
3172 
3173 	/* SAMACTMHz */
3174 	if (DO_BIC(BIC_SAMACTMHz))
3175 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->sam_act_mhz);
3176 
3177 	/* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
3178 	if (DO_BIC(BIC_Totl_c0))
3179 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_wtd_core_c0 / tsc);
3180 	if (DO_BIC(BIC_Any_c0))
3181 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_core_c0 / tsc);
3182 	if (DO_BIC(BIC_GFX_c0))
3183 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_gfxe_c0 / tsc);
3184 	if (DO_BIC(BIC_CPUGFX))
3185 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_both_core_gfxe_c0 / tsc);
3186 
3187 	if (DO_BIC(BIC_Pkgpc2))
3188 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc2 / tsc);
3189 	if (DO_BIC(BIC_Pkgpc3))
3190 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc3 / tsc);
3191 	if (DO_BIC(BIC_Pkgpc6))
3192 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc6 / tsc);
3193 	if (DO_BIC(BIC_Pkgpc7))
3194 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc7 / tsc);
3195 	if (DO_BIC(BIC_Pkgpc8))
3196 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc8 / tsc);
3197 	if (DO_BIC(BIC_Pkgpc9))
3198 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc9 / tsc);
3199 	if (DO_BIC(BIC_Pkgpc10))
3200 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc10 / tsc);
3201 
3202 	if (DO_BIC(BIC_Diec6))
3203 		outp +=
3204 		    sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->die_c6 / crystal_hz / interval_float);
3205 
3206 	if (DO_BIC(BIC_CPU_LPI)) {
3207 		if (p->cpu_lpi >= 0)
3208 			outp +=
3209 			    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
3210 				    100.0 * p->cpu_lpi / 1000000.0 / interval_float);
3211 		else
3212 			outp += sprintf(outp, "%s(neg)", (printed++ ? delim : ""));
3213 	}
3214 	if (DO_BIC(BIC_SYS_LPI)) {
3215 		if (p->sys_lpi >= 0)
3216 			outp +=
3217 			    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
3218 				    100.0 * p->sys_lpi / 1000000.0 / interval_float);
3219 		else
3220 			outp += sprintf(outp, "%s(neg)", (printed++ ? delim : ""));
3221 	}
3222 
3223 	if (DO_BIC(BIC_PkgWatt))
3224 		outp +=
3225 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3226 			    rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_WATTS, interval_float));
3227 	if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
3228 		outp +=
3229 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3230 			    rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_WATTS, interval_float));
3231 	if (DO_BIC(BIC_GFXWatt))
3232 		outp +=
3233 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3234 			    rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_WATTS, interval_float));
3235 	if (DO_BIC(BIC_RAMWatt))
3236 		outp +=
3237 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3238 			    rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_WATTS, interval_float));
3239 	if (DO_BIC(BIC_Pkg_J))
3240 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3241 				rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_JOULES, interval_float));
3242 	if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
3243 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3244 				rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_JOULES, interval_float));
3245 	if (DO_BIC(BIC_GFX_J))
3246 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3247 				rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_JOULES, interval_float));
3248 	if (DO_BIC(BIC_RAM_J))
3249 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3250 				rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_JOULES, interval_float));
3251 	if (DO_BIC(BIC_PKG__))
3252 		outp +=
3253 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3254 			    rapl_counter_get_value(&p->rapl_pkg_perf_status, RAPL_UNIT_WATTS, interval_float));
3255 	if (DO_BIC(BIC_RAM__))
3256 		outp +=
3257 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3258 			    rapl_counter_get_value(&p->rapl_dram_perf_status, RAPL_UNIT_WATTS, interval_float));
3259 	/* UncMHz */
3260 	if (DO_BIC(BIC_UNCORE_MHZ))
3261 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->uncore_mhz);
3262 
3263 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3264 		if (mp->format == FORMAT_RAW) {
3265 			if (mp->width == 32)
3266 				outp +=
3267 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int)p->counter[i]);
3268 			else
3269 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->counter[i]);
3270 		} else if (mp->format == FORMAT_DELTA) {
3271 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
3272 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), p->counter[i]);
3273 			else
3274 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), p->counter[i]);
3275 		} else if (mp->format == FORMAT_PERCENT) {
3276 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->counter[i] / tsc);
3277 		} else if (mp->type == COUNTER_K2M)
3278 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), (unsigned int)p->counter[i] / 1000);
3279 	}
3280 
3281 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3282 		if (pp->format == FORMAT_RAW) {
3283 			if (pp->width == 32)
3284 				outp +=
3285 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
3286 					    (unsigned int)p->perf_counter[i]);
3287 			else
3288 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->perf_counter[i]);
3289 		} else if (pp->format == FORMAT_DELTA) {
3290 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
3291 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), p->perf_counter[i]);
3292 			else
3293 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), p->perf_counter[i]);
3294 		} else if (pp->format == FORMAT_PERCENT) {
3295 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->perf_counter[i] / tsc);
3296 		} else if (pp->type == COUNTER_K2M) {
3297 			outp +=
3298 			    sprintf(outp, "%s%d", (printed++ ? delim : ""), (unsigned int)p->perf_counter[i] / 1000);
3299 		}
3300 	}
3301 
3302 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3303 		const unsigned long value_raw = p->pmt_counter[i];
3304 		double value_converted;
3305 		switch (ppmt->type) {
3306 		case PMT_TYPE_RAW:
3307 			if (pmt_counter_get_width(ppmt) <= 32)
3308 				outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
3309 						(unsigned int)p->pmt_counter[i]);
3310 			else
3311 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->pmt_counter[i]);
3312 
3313 			break;
3314 
3315 		case PMT_TYPE_XTAL_TIME:
3316 			value_converted = 100.0 * value_raw / crystal_hz / interval_float;
3317 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
3318 			break;
3319 
3320 		case PMT_TYPE_TCORE_CLOCK:
3321 			value_converted = 100.0 * value_raw / tcore_clock_freq_hz / interval_float;
3322 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
3323 		}
3324 	}
3325 
3326 	if (DO_BIC(BIC_SysWatt) && (t == &average.threads))
3327 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3328 				rapl_counter_get_value(&pplat_cnt->energy_psys, RAPL_UNIT_WATTS, interval_float));
3329 	if (DO_BIC(BIC_Sys_J) && (t == &average.threads))
3330 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3331 				rapl_counter_get_value(&pplat_cnt->energy_psys, RAPL_UNIT_JOULES, interval_float));
3332 
3333 done:
3334 	if (*(outp - 1) != '\n')
3335 		outp += sprintf(outp, "\n");
3336 
3337 	return 0;
3338 }
3339 
flush_output_stdout(void)3340 void flush_output_stdout(void)
3341 {
3342 	FILE *filep;
3343 
3344 	if (outf == stderr)
3345 		filep = stdout;
3346 	else
3347 		filep = outf;
3348 
3349 	fputs(output_buffer, filep);
3350 	fflush(filep);
3351 
3352 	outp = output_buffer;
3353 }
3354 
flush_output_stderr(void)3355 void flush_output_stderr(void)
3356 {
3357 	fputs(output_buffer, outf);
3358 	fflush(outf);
3359 	outp = output_buffer;
3360 }
3361 
format_all_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)3362 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3363 {
3364 	static int count;
3365 
3366 	if ((!count || (header_iterations && !(count % header_iterations))) || !summary_only)
3367 		print_header("\t");
3368 
3369 	format_counters(&average.threads, &average.cores, &average.packages);
3370 
3371 	count++;
3372 
3373 	if (summary_only)
3374 		return;
3375 
3376 	for_all_cpus(format_counters, t, c, p);
3377 }
3378 
3379 #define DELTA_WRAP32(new, old)			\
3380 	old = ((((unsigned long long)new << 32) - ((unsigned long long)old << 32)) >> 32);
3381 
delta_package(struct pkg_data * new,struct pkg_data * old)3382 int delta_package(struct pkg_data *new, struct pkg_data *old)
3383 {
3384 	int i;
3385 	struct msr_counter *mp;
3386 	struct perf_counter_info *pp;
3387 	struct pmt_counter *ppmt;
3388 
3389 	if (DO_BIC(BIC_Totl_c0))
3390 		old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
3391 	if (DO_BIC(BIC_Any_c0))
3392 		old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
3393 	if (DO_BIC(BIC_GFX_c0))
3394 		old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
3395 	if (DO_BIC(BIC_CPUGFX))
3396 		old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
3397 
3398 	old->pc2 = new->pc2 - old->pc2;
3399 	if (DO_BIC(BIC_Pkgpc3))
3400 		old->pc3 = new->pc3 - old->pc3;
3401 	if (DO_BIC(BIC_Pkgpc6))
3402 		old->pc6 = new->pc6 - old->pc6;
3403 	if (DO_BIC(BIC_Pkgpc7))
3404 		old->pc7 = new->pc7 - old->pc7;
3405 	old->pc8 = new->pc8 - old->pc8;
3406 	old->pc9 = new->pc9 - old->pc9;
3407 	old->pc10 = new->pc10 - old->pc10;
3408 	old->die_c6 = new->die_c6 - old->die_c6;
3409 	old->cpu_lpi = new->cpu_lpi - old->cpu_lpi;
3410 	old->sys_lpi = new->sys_lpi - old->sys_lpi;
3411 	old->pkg_temp_c = new->pkg_temp_c;
3412 
3413 	/* flag an error when rc6 counter resets/wraps */
3414 	if (old->gfx_rc6_ms > new->gfx_rc6_ms)
3415 		old->gfx_rc6_ms = -1;
3416 	else
3417 		old->gfx_rc6_ms = new->gfx_rc6_ms - old->gfx_rc6_ms;
3418 
3419 	old->uncore_mhz = new->uncore_mhz;
3420 	old->gfx_mhz = new->gfx_mhz;
3421 	old->gfx_act_mhz = new->gfx_act_mhz;
3422 
3423 	/* flag an error when mc6 counter resets/wraps */
3424 	if (old->sam_mc6_ms > new->sam_mc6_ms)
3425 		old->sam_mc6_ms = -1;
3426 	else
3427 		old->sam_mc6_ms = new->sam_mc6_ms - old->sam_mc6_ms;
3428 
3429 	old->sam_mhz = new->sam_mhz;
3430 	old->sam_act_mhz = new->sam_act_mhz;
3431 
3432 	old->energy_pkg.raw_value = new->energy_pkg.raw_value - old->energy_pkg.raw_value;
3433 	old->energy_cores.raw_value = new->energy_cores.raw_value - old->energy_cores.raw_value;
3434 	old->energy_gfx.raw_value = new->energy_gfx.raw_value - old->energy_gfx.raw_value;
3435 	old->energy_dram.raw_value = new->energy_dram.raw_value - old->energy_dram.raw_value;
3436 	old->rapl_pkg_perf_status.raw_value = new->rapl_pkg_perf_status.raw_value - old->rapl_pkg_perf_status.raw_value;
3437 	old->rapl_dram_perf_status.raw_value =
3438 	    new->rapl_dram_perf_status.raw_value - old->rapl_dram_perf_status.raw_value;
3439 
3440 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3441 		if (mp->format == FORMAT_RAW)
3442 			old->counter[i] = new->counter[i];
3443 		else if (mp->format == FORMAT_AVERAGE)
3444 			old->counter[i] = new->counter[i];
3445 		else
3446 			old->counter[i] = new->counter[i] - old->counter[i];
3447 	}
3448 
3449 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3450 		if (pp->format == FORMAT_RAW)
3451 			old->perf_counter[i] = new->perf_counter[i];
3452 		else if (pp->format == FORMAT_AVERAGE)
3453 			old->perf_counter[i] = new->perf_counter[i];
3454 		else
3455 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3456 	}
3457 
3458 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3459 		if (ppmt->format == FORMAT_RAW)
3460 			old->pmt_counter[i] = new->pmt_counter[i];
3461 		else
3462 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3463 	}
3464 
3465 	return 0;
3466 }
3467 
delta_core(struct core_data * new,struct core_data * old)3468 void delta_core(struct core_data *new, struct core_data *old)
3469 {
3470 	int i;
3471 	struct msr_counter *mp;
3472 	struct perf_counter_info *pp;
3473 	struct pmt_counter *ppmt;
3474 
3475 	old->c3 = new->c3 - old->c3;
3476 	old->c6 = new->c6 - old->c6;
3477 	old->c7 = new->c7 - old->c7;
3478 	old->core_temp_c = new->core_temp_c;
3479 	old->core_throt_cnt = new->core_throt_cnt;
3480 	old->mc6_us = new->mc6_us - old->mc6_us;
3481 
3482 	DELTA_WRAP32(new->core_energy.raw_value, old->core_energy.raw_value);
3483 
3484 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3485 		if (mp->format == FORMAT_RAW)
3486 			old->counter[i] = new->counter[i];
3487 		else
3488 			old->counter[i] = new->counter[i] - old->counter[i];
3489 	}
3490 
3491 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3492 		if (pp->format == FORMAT_RAW)
3493 			old->perf_counter[i] = new->perf_counter[i];
3494 		else
3495 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3496 	}
3497 
3498 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3499 		if (ppmt->format == FORMAT_RAW)
3500 			old->pmt_counter[i] = new->pmt_counter[i];
3501 		else
3502 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3503 	}
3504 }
3505 
soft_c1_residency_display(int bic)3506 int soft_c1_residency_display(int bic)
3507 {
3508 	if (!DO_BIC(BIC_CPU_c1) || platform->has_msr_core_c1_res)
3509 		return 0;
3510 
3511 	return DO_BIC_READ(bic);
3512 }
3513 
3514 /*
3515  * old = new - old
3516  */
delta_thread(struct thread_data * new,struct thread_data * old,struct core_data * core_delta)3517 int delta_thread(struct thread_data *new, struct thread_data *old, struct core_data *core_delta)
3518 {
3519 	int i;
3520 	struct msr_counter *mp;
3521 	struct perf_counter_info *pp;
3522 	struct pmt_counter *ppmt;
3523 
3524 	/* we run cpuid just the 1st time, copy the results */
3525 	if (DO_BIC(BIC_APIC))
3526 		new->apic_id = old->apic_id;
3527 	if (DO_BIC(BIC_X2APIC))
3528 		new->x2apic_id = old->x2apic_id;
3529 
3530 	/*
3531 	 * the timestamps from start of measurement interval are in "old"
3532 	 * the timestamp from end of measurement interval are in "new"
3533 	 * over-write old w/ new so we can print end of interval values
3534 	 */
3535 
3536 	timersub(&new->tv_begin, &old->tv_begin, &old->tv_delta);
3537 	old->tv_begin = new->tv_begin;
3538 	old->tv_end = new->tv_end;
3539 
3540 	old->tsc = new->tsc - old->tsc;
3541 
3542 	/* check for TSC < 1 Mcycles over interval */
3543 	if (old->tsc < (1000 * 1000))
3544 		errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
3545 		     "You can disable all c-states by booting with \"idle=poll\"\n"
3546 		     "or just the deep ones with \"processor.max_cstate=1\"");
3547 
3548 	old->c1 = new->c1 - old->c1;
3549 
3550 	if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz) || DO_BIC(BIC_IPC)
3551 	    || soft_c1_residency_display(BIC_Avg_MHz)) {
3552 		if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
3553 			old->aperf = new->aperf - old->aperf;
3554 			old->mperf = new->mperf - old->mperf;
3555 		} else {
3556 			return -1;
3557 		}
3558 	}
3559 
3560 	if (platform->has_msr_core_c1_res) {
3561 		/*
3562 		 * Some models have a dedicated C1 residency MSR,
3563 		 * which should be more accurate than the derivation below.
3564 		 */
3565 	} else {
3566 		/*
3567 		 * As counter collection is not atomic,
3568 		 * it is possible for mperf's non-halted cycles + idle states
3569 		 * to exceed TSC's all cycles: show c1 = 0% in that case.
3570 		 */
3571 		if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > (old->tsc * tsc_tweak))
3572 			old->c1 = 0;
3573 		else {
3574 			/* normal case, derive c1 */
3575 			old->c1 = (old->tsc * tsc_tweak) - old->mperf - core_delta->c3
3576 			    - core_delta->c6 - core_delta->c7;
3577 		}
3578 	}
3579 
3580 	if (old->mperf == 0) {
3581 		if (debug > 1)
3582 			fprintf(outf, "cpu%d MPERF 0!\n", old->cpu_id);
3583 		old->mperf = 1;	/* divide by 0 protection */
3584 	}
3585 
3586 	if (DO_BIC(BIC_IPC))
3587 		old->instr_count = new->instr_count - old->instr_count;
3588 
3589 	if (DO_BIC(BIC_IRQ))
3590 		old->irq_count = new->irq_count - old->irq_count;
3591 
3592 	if (DO_BIC(BIC_NMI))
3593 		old->nmi_count = new->nmi_count - old->nmi_count;
3594 
3595 	if (DO_BIC(BIC_SMI))
3596 		old->smi_count = new->smi_count - old->smi_count;
3597 
3598 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3599 		if (mp->format == FORMAT_RAW)
3600 			old->counter[i] = new->counter[i];
3601 		else
3602 			old->counter[i] = new->counter[i] - old->counter[i];
3603 	}
3604 
3605 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
3606 		if (pp->format == FORMAT_RAW)
3607 			old->perf_counter[i] = new->perf_counter[i];
3608 		else
3609 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3610 	}
3611 
3612 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3613 		if (ppmt->format == FORMAT_RAW)
3614 			old->pmt_counter[i] = new->pmt_counter[i];
3615 		else
3616 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3617 	}
3618 
3619 	return 0;
3620 }
3621 
delta_cpu(struct thread_data * t,struct core_data * c,struct pkg_data * p,struct thread_data * t2,struct core_data * c2,struct pkg_data * p2)3622 int delta_cpu(struct thread_data *t, struct core_data *c,
3623 	      struct pkg_data *p, struct thread_data *t2, struct core_data *c2, struct pkg_data *p2)
3624 {
3625 	int retval = 0;
3626 
3627 	/* calculate core delta only for 1st thread in core */
3628 	if (is_cpu_first_thread_in_core(t, c, p))
3629 		delta_core(c, c2);
3630 
3631 	/* always calculate thread delta */
3632 	retval = delta_thread(t, t2, c2);	/* c2 is core delta */
3633 
3634 	/* calculate package delta only for 1st core in package */
3635 	if (is_cpu_first_core_in_package(t, c, p))
3636 		retval |= delta_package(p, p2);
3637 
3638 	return retval;
3639 }
3640 
delta_platform(struct platform_counters * new,struct platform_counters * old)3641 void delta_platform(struct platform_counters *new, struct platform_counters *old)
3642 {
3643 	old->energy_psys.raw_value = new->energy_psys.raw_value - old->energy_psys.raw_value;
3644 }
3645 
rapl_counter_clear(struct rapl_counter * c)3646 void rapl_counter_clear(struct rapl_counter *c)
3647 {
3648 	c->raw_value = 0;
3649 	c->scale = 0.0;
3650 	c->unit = RAPL_UNIT_INVALID;
3651 }
3652 
clear_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)3653 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3654 {
3655 	int i;
3656 	struct msr_counter *mp;
3657 
3658 	t->tv_begin.tv_sec = 0;
3659 	t->tv_begin.tv_usec = 0;
3660 	t->tv_end.tv_sec = 0;
3661 	t->tv_end.tv_usec = 0;
3662 	t->tv_delta.tv_sec = 0;
3663 	t->tv_delta.tv_usec = 0;
3664 
3665 	t->tsc = 0;
3666 	t->aperf = 0;
3667 	t->mperf = 0;
3668 	t->c1 = 0;
3669 
3670 	t->instr_count = 0;
3671 
3672 	t->irq_count = 0;
3673 	t->nmi_count = 0;
3674 	t->smi_count = 0;
3675 
3676 	c->c3 = 0;
3677 	c->c6 = 0;
3678 	c->c7 = 0;
3679 	c->mc6_us = 0;
3680 	c->core_temp_c = 0;
3681 	rapl_counter_clear(&c->core_energy);
3682 	c->core_throt_cnt = 0;
3683 
3684 	p->pkg_wtd_core_c0 = 0;
3685 	p->pkg_any_core_c0 = 0;
3686 	p->pkg_any_gfxe_c0 = 0;
3687 	p->pkg_both_core_gfxe_c0 = 0;
3688 
3689 	p->pc2 = 0;
3690 	if (DO_BIC(BIC_Pkgpc3))
3691 		p->pc3 = 0;
3692 	if (DO_BIC(BIC_Pkgpc6))
3693 		p->pc6 = 0;
3694 	if (DO_BIC(BIC_Pkgpc7))
3695 		p->pc7 = 0;
3696 	p->pc8 = 0;
3697 	p->pc9 = 0;
3698 	p->pc10 = 0;
3699 	p->die_c6 = 0;
3700 	p->cpu_lpi = 0;
3701 	p->sys_lpi = 0;
3702 
3703 	rapl_counter_clear(&p->energy_pkg);
3704 	rapl_counter_clear(&p->energy_dram);
3705 	rapl_counter_clear(&p->energy_cores);
3706 	rapl_counter_clear(&p->energy_gfx);
3707 	rapl_counter_clear(&p->rapl_pkg_perf_status);
3708 	rapl_counter_clear(&p->rapl_dram_perf_status);
3709 	p->pkg_temp_c = 0;
3710 
3711 	p->gfx_rc6_ms = 0;
3712 	p->uncore_mhz = 0;
3713 	p->gfx_mhz = 0;
3714 	p->gfx_act_mhz = 0;
3715 	p->sam_mc6_ms = 0;
3716 	p->sam_mhz = 0;
3717 	p->sam_act_mhz = 0;
3718 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next)
3719 		t->counter[i] = 0;
3720 
3721 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next)
3722 		c->counter[i] = 0;
3723 
3724 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next)
3725 		p->counter[i] = 0;
3726 
3727 	memset(&t->perf_counter[0], 0, sizeof(t->perf_counter));
3728 	memset(&c->perf_counter[0], 0, sizeof(c->perf_counter));
3729 	memset(&p->perf_counter[0], 0, sizeof(p->perf_counter));
3730 
3731 	memset(&t->pmt_counter[0], 0, ARRAY_SIZE(t->pmt_counter));
3732 	memset(&c->pmt_counter[0], 0, ARRAY_SIZE(c->pmt_counter));
3733 	memset(&p->pmt_counter[0], 0, ARRAY_SIZE(p->pmt_counter));
3734 }
3735 
rapl_counter_accumulate(struct rapl_counter * dst,const struct rapl_counter * src)3736 void rapl_counter_accumulate(struct rapl_counter *dst, const struct rapl_counter *src)
3737 {
3738 	/* Copy unit and scale from src if dst is not initialized */
3739 	if (dst->unit == RAPL_UNIT_INVALID) {
3740 		dst->unit = src->unit;
3741 		dst->scale = src->scale;
3742 	}
3743 
3744 	assert(dst->unit == src->unit);
3745 	assert(dst->scale == src->scale);
3746 
3747 	dst->raw_value += src->raw_value;
3748 }
3749 
sum_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)3750 int sum_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3751 {
3752 	int i;
3753 	struct msr_counter *mp;
3754 	struct perf_counter_info *pp;
3755 	struct pmt_counter *ppmt;
3756 
3757 	/* copy un-changing apic_id's */
3758 	if (DO_BIC(BIC_APIC))
3759 		average.threads.apic_id = t->apic_id;
3760 	if (DO_BIC(BIC_X2APIC))
3761 		average.threads.x2apic_id = t->x2apic_id;
3762 
3763 	/* remember first tv_begin */
3764 	if (average.threads.tv_begin.tv_sec == 0)
3765 		average.threads.tv_begin = procsysfs_tv_begin;
3766 
3767 	/* remember last tv_end */
3768 	average.threads.tv_end = t->tv_end;
3769 
3770 	average.threads.tsc += t->tsc;
3771 	average.threads.aperf += t->aperf;
3772 	average.threads.mperf += t->mperf;
3773 	average.threads.c1 += t->c1;
3774 
3775 	average.threads.instr_count += t->instr_count;
3776 
3777 	average.threads.irq_count += t->irq_count;
3778 	average.threads.nmi_count += t->nmi_count;
3779 	average.threads.smi_count += t->smi_count;
3780 
3781 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3782 		if (mp->format == FORMAT_RAW)
3783 			continue;
3784 		average.threads.counter[i] += t->counter[i];
3785 	}
3786 
3787 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
3788 		if (pp->format == FORMAT_RAW)
3789 			continue;
3790 		average.threads.perf_counter[i] += t->perf_counter[i];
3791 	}
3792 
3793 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3794 		average.threads.pmt_counter[i] += t->pmt_counter[i];
3795 	}
3796 
3797 	/* sum per-core values only for 1st thread in core */
3798 	if (!is_cpu_first_thread_in_core(t, c, p))
3799 		return 0;
3800 
3801 	average.cores.c3 += c->c3;
3802 	average.cores.c6 += c->c6;
3803 	average.cores.c7 += c->c7;
3804 	average.cores.mc6_us += c->mc6_us;
3805 
3806 	average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
3807 	average.cores.core_throt_cnt = MAX(average.cores.core_throt_cnt, c->core_throt_cnt);
3808 
3809 	rapl_counter_accumulate(&average.cores.core_energy, &c->core_energy);
3810 
3811 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3812 		if (mp->format == FORMAT_RAW)
3813 			continue;
3814 		average.cores.counter[i] += c->counter[i];
3815 	}
3816 
3817 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3818 		if (pp->format == FORMAT_RAW)
3819 			continue;
3820 		average.cores.perf_counter[i] += c->perf_counter[i];
3821 	}
3822 
3823 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3824 		average.cores.pmt_counter[i] += c->pmt_counter[i];
3825 	}
3826 
3827 	/* sum per-pkg values only for 1st core in pkg */
3828 	if (!is_cpu_first_core_in_package(t, c, p))
3829 		return 0;
3830 
3831 	if (DO_BIC(BIC_Totl_c0))
3832 		average.packages.pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
3833 	if (DO_BIC(BIC_Any_c0))
3834 		average.packages.pkg_any_core_c0 += p->pkg_any_core_c0;
3835 	if (DO_BIC(BIC_GFX_c0))
3836 		average.packages.pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
3837 	if (DO_BIC(BIC_CPUGFX))
3838 		average.packages.pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
3839 
3840 	average.packages.pc2 += p->pc2;
3841 	if (DO_BIC(BIC_Pkgpc3))
3842 		average.packages.pc3 += p->pc3;
3843 	if (DO_BIC(BIC_Pkgpc6))
3844 		average.packages.pc6 += p->pc6;
3845 	if (DO_BIC(BIC_Pkgpc7))
3846 		average.packages.pc7 += p->pc7;
3847 	average.packages.pc8 += p->pc8;
3848 	average.packages.pc9 += p->pc9;
3849 	average.packages.pc10 += p->pc10;
3850 	average.packages.die_c6 += p->die_c6;
3851 
3852 	average.packages.cpu_lpi = p->cpu_lpi;
3853 	average.packages.sys_lpi = p->sys_lpi;
3854 
3855 	rapl_counter_accumulate(&average.packages.energy_pkg, &p->energy_pkg);
3856 	rapl_counter_accumulate(&average.packages.energy_dram, &p->energy_dram);
3857 	rapl_counter_accumulate(&average.packages.energy_cores, &p->energy_cores);
3858 	rapl_counter_accumulate(&average.packages.energy_gfx, &p->energy_gfx);
3859 
3860 	average.packages.gfx_rc6_ms = p->gfx_rc6_ms;
3861 	average.packages.uncore_mhz = p->uncore_mhz;
3862 	average.packages.gfx_mhz = p->gfx_mhz;
3863 	average.packages.gfx_act_mhz = p->gfx_act_mhz;
3864 	average.packages.sam_mc6_ms = p->sam_mc6_ms;
3865 	average.packages.sam_mhz = p->sam_mhz;
3866 	average.packages.sam_act_mhz = p->sam_act_mhz;
3867 
3868 	average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
3869 
3870 	rapl_counter_accumulate(&average.packages.rapl_pkg_perf_status, &p->rapl_pkg_perf_status);
3871 	rapl_counter_accumulate(&average.packages.rapl_dram_perf_status, &p->rapl_dram_perf_status);
3872 
3873 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3874 		if ((mp->format == FORMAT_RAW) && (topo.num_packages == 0))
3875 			average.packages.counter[i] = p->counter[i];
3876 		else
3877 			average.packages.counter[i] += p->counter[i];
3878 	}
3879 
3880 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3881 		if ((pp->format == FORMAT_RAW) && (topo.num_packages == 0))
3882 			average.packages.perf_counter[i] = p->perf_counter[i];
3883 		else
3884 			average.packages.perf_counter[i] += p->perf_counter[i];
3885 	}
3886 
3887 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3888 		average.packages.pmt_counter[i] += p->pmt_counter[i];
3889 	}
3890 
3891 	return 0;
3892 }
3893 
3894 /*
3895  * sum the counters for all cpus in the system
3896  * compute the weighted average
3897  */
compute_average(struct thread_data * t,struct core_data * c,struct pkg_data * p)3898 void compute_average(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3899 {
3900 	int i;
3901 	struct msr_counter *mp;
3902 	struct perf_counter_info *pp;
3903 	struct pmt_counter *ppmt;
3904 
3905 	clear_counters(&average.threads, &average.cores, &average.packages);
3906 
3907 	for_all_cpus(sum_counters, t, c, p);
3908 
3909 	/* Use the global time delta for the average. */
3910 	average.threads.tv_delta = tv_delta;
3911 
3912 	average.threads.tsc /= topo.allowed_cpus;
3913 	average.threads.aperf /= topo.allowed_cpus;
3914 	average.threads.mperf /= topo.allowed_cpus;
3915 	average.threads.instr_count /= topo.allowed_cpus;
3916 	average.threads.c1 /= topo.allowed_cpus;
3917 
3918 	if (average.threads.irq_count > 9999999)
3919 		sums_need_wide_columns = 1;
3920 	if (average.threads.nmi_count > 9999999)
3921 		sums_need_wide_columns = 1;
3922 
3923 
3924 	average.cores.c3 /= topo.allowed_cores;
3925 	average.cores.c6 /= topo.allowed_cores;
3926 	average.cores.c7 /= topo.allowed_cores;
3927 	average.cores.mc6_us /= topo.allowed_cores;
3928 
3929 	if (DO_BIC(BIC_Totl_c0))
3930 		average.packages.pkg_wtd_core_c0 /= topo.allowed_packages;
3931 	if (DO_BIC(BIC_Any_c0))
3932 		average.packages.pkg_any_core_c0 /= topo.allowed_packages;
3933 	if (DO_BIC(BIC_GFX_c0))
3934 		average.packages.pkg_any_gfxe_c0 /= topo.allowed_packages;
3935 	if (DO_BIC(BIC_CPUGFX))
3936 		average.packages.pkg_both_core_gfxe_c0 /= topo.allowed_packages;
3937 
3938 	average.packages.pc2 /= topo.allowed_packages;
3939 	if (DO_BIC(BIC_Pkgpc3))
3940 		average.packages.pc3 /= topo.allowed_packages;
3941 	if (DO_BIC(BIC_Pkgpc6))
3942 		average.packages.pc6 /= topo.allowed_packages;
3943 	if (DO_BIC(BIC_Pkgpc7))
3944 		average.packages.pc7 /= topo.allowed_packages;
3945 
3946 	average.packages.pc8 /= topo.allowed_packages;
3947 	average.packages.pc9 /= topo.allowed_packages;
3948 	average.packages.pc10 /= topo.allowed_packages;
3949 	average.packages.die_c6 /= topo.allowed_packages;
3950 
3951 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3952 		if (mp->format == FORMAT_RAW)
3953 			continue;
3954 		if (mp->type == COUNTER_ITEMS) {
3955 			if (average.threads.counter[i] > 9999999)
3956 				sums_need_wide_columns = 1;
3957 			continue;
3958 		}
3959 		average.threads.counter[i] /= topo.allowed_cpus;
3960 	}
3961 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3962 		if (mp->format == FORMAT_RAW)
3963 			continue;
3964 		if (mp->type == COUNTER_ITEMS) {
3965 			if (average.cores.counter[i] > 9999999)
3966 				sums_need_wide_columns = 1;
3967 		}
3968 		average.cores.counter[i] /= topo.allowed_cores;
3969 	}
3970 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3971 		if (mp->format == FORMAT_RAW)
3972 			continue;
3973 		if (mp->type == COUNTER_ITEMS) {
3974 			if (average.packages.counter[i] > 9999999)
3975 				sums_need_wide_columns = 1;
3976 		}
3977 		average.packages.counter[i] /= topo.allowed_packages;
3978 	}
3979 
3980 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
3981 		if (pp->format == FORMAT_RAW)
3982 			continue;
3983 		if (pp->type == COUNTER_ITEMS) {
3984 			if (average.threads.perf_counter[i] > 9999999)
3985 				sums_need_wide_columns = 1;
3986 			continue;
3987 		}
3988 		average.threads.perf_counter[i] /= topo.allowed_cpus;
3989 	}
3990 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3991 		if (pp->format == FORMAT_RAW)
3992 			continue;
3993 		if (pp->type == COUNTER_ITEMS) {
3994 			if (average.cores.perf_counter[i] > 9999999)
3995 				sums_need_wide_columns = 1;
3996 		}
3997 		average.cores.perf_counter[i] /= topo.allowed_cores;
3998 	}
3999 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
4000 		if (pp->format == FORMAT_RAW)
4001 			continue;
4002 		if (pp->type == COUNTER_ITEMS) {
4003 			if (average.packages.perf_counter[i] > 9999999)
4004 				sums_need_wide_columns = 1;
4005 		}
4006 		average.packages.perf_counter[i] /= topo.allowed_packages;
4007 	}
4008 
4009 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
4010 		average.threads.pmt_counter[i] /= topo.allowed_cpus;
4011 	}
4012 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
4013 		average.cores.pmt_counter[i] /= topo.allowed_cores;
4014 	}
4015 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
4016 		average.packages.pmt_counter[i] /= topo.allowed_packages;
4017 	}
4018 }
4019 
rdtsc(void)4020 static unsigned long long rdtsc(void)
4021 {
4022 	unsigned int low, high;
4023 
4024 	asm volatile ("rdtsc":"=a" (low), "=d"(high));
4025 
4026 	return low | ((unsigned long long)high) << 32;
4027 }
4028 
4029 /*
4030  * Open a file, and exit on failure
4031  */
fopen_or_die(const char * path,const char * mode)4032 FILE *fopen_or_die(const char *path, const char *mode)
4033 {
4034 	FILE *filep = fopen(path, mode);
4035 
4036 	if (!filep)
4037 		err(1, "%s: open failed", path);
4038 	return filep;
4039 }
4040 
4041 /*
4042  * snapshot_sysfs_counter()
4043  *
4044  * return snapshot of given counter
4045  */
snapshot_sysfs_counter(char * path)4046 unsigned long long snapshot_sysfs_counter(char *path)
4047 {
4048 	FILE *fp;
4049 	int retval;
4050 	unsigned long long counter;
4051 
4052 	fp = fopen_or_die(path, "r");
4053 
4054 	retval = fscanf(fp, "%lld", &counter);
4055 	if (retval != 1)
4056 		err(1, "snapshot_sysfs_counter(%s)", path);
4057 
4058 	fclose(fp);
4059 
4060 	return counter;
4061 }
4062 
get_mp(int cpu,struct msr_counter * mp,unsigned long long * counterp,char * counter_path)4063 int get_mp(int cpu, struct msr_counter *mp, unsigned long long *counterp, char *counter_path)
4064 {
4065 	if (mp->msr_num != 0) {
4066 		assert(!no_msr);
4067 		if (get_msr(cpu, mp->msr_num, counterp))
4068 			return -1;
4069 	} else {
4070 		char path[128 + PATH_BYTES];
4071 
4072 		if (mp->flags & SYSFS_PERCPU) {
4073 			sprintf(path, "/sys/devices/system/cpu/cpu%d/%s", cpu, mp->sp->path);
4074 
4075 			*counterp = snapshot_sysfs_counter(path);
4076 		} else {
4077 			*counterp = snapshot_sysfs_counter(counter_path);
4078 		}
4079 	}
4080 
4081 	return 0;
4082 }
4083 
get_legacy_uncore_mhz(int package)4084 unsigned long long get_legacy_uncore_mhz(int package)
4085 {
4086 	char path[128];
4087 	int die;
4088 	static int warn_once;
4089 
4090 	/*
4091 	 * for this package, use the first die_id that exists
4092 	 */
4093 	for (die = 0; die <= topo.max_die_id; ++die) {
4094 
4095 		sprintf(path, "/sys/devices/system/cpu/intel_uncore_frequency/package_%02d_die_%02d/current_freq_khz",
4096 			package, die);
4097 
4098 		if (access(path, R_OK) == 0)
4099 			return (snapshot_sysfs_counter(path) / 1000);
4100 	}
4101 	if (!warn_once) {
4102 		warnx("BUG: %s: No %s", __func__, path);
4103 		warn_once = 1;
4104 	}
4105 
4106 	return 0;
4107 }
4108 
get_epb(int cpu)4109 int get_epb(int cpu)
4110 {
4111 	char path[128 + PATH_BYTES];
4112 	unsigned long long msr;
4113 	int ret, epb = -1;
4114 	FILE *fp;
4115 
4116 	sprintf(path, "/sys/devices/system/cpu/cpu%d/power/energy_perf_bias", cpu);
4117 
4118 	fp = fopen(path, "r");
4119 	if (!fp)
4120 		goto msr_fallback;
4121 
4122 	ret = fscanf(fp, "%d", &epb);
4123 	if (ret != 1)
4124 		err(1, "%s(%s)", __func__, path);
4125 
4126 	fclose(fp);
4127 
4128 	return epb;
4129 
4130 msr_fallback:
4131 	if (no_msr)
4132 		return -1;
4133 
4134 	get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr);
4135 
4136 	return msr & 0xf;
4137 }
4138 
get_apic_id(struct thread_data * t)4139 void get_apic_id(struct thread_data *t)
4140 {
4141 	unsigned int eax, ebx, ecx, edx;
4142 
4143 	if (DO_BIC(BIC_APIC)) {
4144 		eax = ebx = ecx = edx = 0;
4145 		__cpuid(1, eax, ebx, ecx, edx);
4146 
4147 		t->apic_id = (ebx >> 24) & 0xff;
4148 	}
4149 
4150 	if (!DO_BIC(BIC_X2APIC))
4151 		return;
4152 
4153 	if (authentic_amd || hygon_genuine) {
4154 		unsigned int topology_extensions;
4155 
4156 		if (max_extended_level < 0x8000001e)
4157 			return;
4158 
4159 		eax = ebx = ecx = edx = 0;
4160 		__cpuid(0x80000001, eax, ebx, ecx, edx);
4161 		topology_extensions = ecx & (1 << 22);
4162 
4163 		if (topology_extensions == 0)
4164 			return;
4165 
4166 		eax = ebx = ecx = edx = 0;
4167 		__cpuid(0x8000001e, eax, ebx, ecx, edx);
4168 
4169 		t->x2apic_id = eax;
4170 		return;
4171 	}
4172 
4173 	if (!genuine_intel)
4174 		return;
4175 
4176 	if (max_level < 0xb)
4177 		return;
4178 
4179 	ecx = 0;
4180 	__cpuid(0xb, eax, ebx, ecx, edx);
4181 	t->x2apic_id = edx;
4182 
4183 	if (debug && (t->apic_id != (t->x2apic_id & 0xff)))
4184 		fprintf(outf, "cpu%d: BIOS BUG: apic 0x%x x2apic 0x%x\n", t->cpu_id, t->apic_id, t->x2apic_id);
4185 }
4186 
get_core_throt_cnt(int cpu,unsigned long long * cnt)4187 int get_core_throt_cnt(int cpu, unsigned long long *cnt)
4188 {
4189 	char path[128 + PATH_BYTES];
4190 	unsigned long long tmp;
4191 	FILE *fp;
4192 	int ret;
4193 
4194 	sprintf(path, "/sys/devices/system/cpu/cpu%d/thermal_throttle/core_throttle_count", cpu);
4195 	fp = fopen(path, "r");
4196 	if (!fp)
4197 		return -1;
4198 	ret = fscanf(fp, "%lld", &tmp);
4199 	fclose(fp);
4200 	if (ret != 1)
4201 		return -1;
4202 	*cnt = tmp;
4203 
4204 	return 0;
4205 }
4206 
4207 struct amperf_group_fd {
4208 	int aperf;		/* Also the group descriptor */
4209 	int mperf;
4210 };
4211 
read_perf_counter_info(const char * const path,const char * const parse_format,void * value_ptr)4212 static int read_perf_counter_info(const char *const path, const char *const parse_format, void *value_ptr)
4213 {
4214 	int fdmt;
4215 	int bytes_read;
4216 	char buf[64];
4217 	int ret = -1;
4218 
4219 	fdmt = open(path, O_RDONLY, 0);
4220 	if (fdmt == -1) {
4221 		if (debug)
4222 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4223 		ret = -1;
4224 		goto cleanup_and_exit;
4225 	}
4226 
4227 	bytes_read = read(fdmt, buf, sizeof(buf) - 1);
4228 	if (bytes_read <= 0 || bytes_read >= (int)sizeof(buf)) {
4229 		if (debug)
4230 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4231 		ret = -1;
4232 		goto cleanup_and_exit;
4233 	}
4234 
4235 	buf[bytes_read] = '\0';
4236 
4237 	if (sscanf(buf, parse_format, value_ptr) != 1) {
4238 		if (debug)
4239 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4240 		ret = -1;
4241 		goto cleanup_and_exit;
4242 	}
4243 
4244 	ret = 0;
4245 
4246 cleanup_and_exit:
4247 	close(fdmt);
4248 	return ret;
4249 }
4250 
read_perf_counter_info_n(const char * const path,const char * const parse_format)4251 static unsigned int read_perf_counter_info_n(const char *const path, const char *const parse_format)
4252 {
4253 	unsigned int v;
4254 	int status;
4255 
4256 	status = read_perf_counter_info(path, parse_format, &v);
4257 	if (status)
4258 		v = -1;
4259 
4260 	return v;
4261 }
4262 
read_perf_type(const char * subsys)4263 static unsigned int read_perf_type(const char *subsys)
4264 {
4265 	const char *const path_format = "/sys/bus/event_source/devices/%s/type";
4266 	const char *const format = "%u";
4267 	char path[128];
4268 
4269 	snprintf(path, sizeof(path), path_format, subsys);
4270 
4271 	return read_perf_counter_info_n(path, format);
4272 }
4273 
read_perf_config(const char * subsys,const char * event_name)4274 static unsigned int read_perf_config(const char *subsys, const char *event_name)
4275 {
4276 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s";
4277 	FILE *fconfig = NULL;
4278 	char path[128];
4279 	char config_str[64];
4280 	unsigned int config;
4281 	unsigned int umask;
4282 	bool has_config = false;
4283 	bool has_umask = false;
4284 	unsigned int ret = -1;
4285 
4286 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4287 
4288 	fconfig = fopen(path, "r");
4289 	if (!fconfig)
4290 		return -1;
4291 
4292 	if (fgets(config_str, ARRAY_SIZE(config_str), fconfig) != config_str)
4293 		goto cleanup_and_exit;
4294 
4295 	for (char *pconfig_str = &config_str[0]; pconfig_str;) {
4296 		if (sscanf(pconfig_str, "event=%x", &config) == 1) {
4297 			has_config = true;
4298 			goto next;
4299 		}
4300 
4301 		if (sscanf(pconfig_str, "umask=%x", &umask) == 1) {
4302 			has_umask = true;
4303 			goto next;
4304 		}
4305 
4306 next:
4307 		pconfig_str = strchr(pconfig_str, ',');
4308 		if (pconfig_str) {
4309 			*pconfig_str = '\0';
4310 			++pconfig_str;
4311 		}
4312 	}
4313 
4314 	if (!has_umask)
4315 		umask = 0;
4316 
4317 	if (has_config)
4318 		ret = (umask << 8) | config;
4319 
4320 cleanup_and_exit:
4321 	fclose(fconfig);
4322 	return ret;
4323 }
4324 
read_perf_rapl_unit(const char * subsys,const char * event_name)4325 static unsigned int read_perf_rapl_unit(const char *subsys, const char *event_name)
4326 {
4327 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.unit";
4328 	const char *const format = "%s";
4329 	char path[128];
4330 	char unit_buffer[16];
4331 
4332 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4333 
4334 	read_perf_counter_info(path, format, &unit_buffer);
4335 	if (strcmp("Joules", unit_buffer) == 0)
4336 		return RAPL_UNIT_JOULES;
4337 
4338 	return RAPL_UNIT_INVALID;
4339 }
4340 
read_perf_scale(const char * subsys,const char * event_name)4341 static double read_perf_scale(const char *subsys, const char *event_name)
4342 {
4343 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.scale";
4344 	const char *const format = "%lf";
4345 	char path[128];
4346 	double scale;
4347 
4348 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4349 
4350 	if (read_perf_counter_info(path, format, &scale))
4351 		return 0.0;
4352 
4353 	return scale;
4354 }
4355 
rapl_counter_info_count_perf(const struct rapl_counter_info_t * rci)4356 size_t rapl_counter_info_count_perf(const struct rapl_counter_info_t *rci)
4357 {
4358 	size_t ret = 0;
4359 
4360 	for (int i = 0; i < NUM_RAPL_COUNTERS; ++i)
4361 		if (rci->source[i] == COUNTER_SOURCE_PERF)
4362 			++ret;
4363 
4364 	return ret;
4365 }
4366 
cstate_counter_info_count_perf(const struct cstate_counter_info_t * cci)4367 static size_t cstate_counter_info_count_perf(const struct cstate_counter_info_t *cci)
4368 {
4369 	size_t ret = 0;
4370 
4371 	for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i)
4372 		if (cci->source[i] == COUNTER_SOURCE_PERF)
4373 			++ret;
4374 
4375 	return ret;
4376 }
4377 
write_rapl_counter(struct rapl_counter * rc,struct rapl_counter_info_t * rci,unsigned int idx)4378 void write_rapl_counter(struct rapl_counter *rc, struct rapl_counter_info_t *rci, unsigned int idx)
4379 {
4380 	if (rci->source[idx] == COUNTER_SOURCE_NONE)
4381 		return;
4382 
4383 	rc->raw_value = rci->data[idx];
4384 	rc->unit = rci->unit[idx];
4385 	rc->scale = rci->scale[idx];
4386 }
4387 
get_rapl_counters(int cpu,unsigned int domain,struct core_data * c,struct pkg_data * p)4388 int get_rapl_counters(int cpu, unsigned int domain, struct core_data *c, struct pkg_data *p)
4389 {
4390 	struct platform_counters *pplat_cnt = p == package_odd ? &platform_counters_odd : &platform_counters_even;
4391 	unsigned long long perf_data[NUM_RAPL_COUNTERS + 1];
4392 	struct rapl_counter_info_t *rci;
4393 
4394 	if (debug >= 2)
4395 		fprintf(stderr, "%s: cpu%d domain%d\n", __func__, cpu, domain);
4396 
4397 	assert(rapl_counter_info_perdomain);
4398 	assert(domain < rapl_counter_info_perdomain_size);
4399 
4400 	rci = &rapl_counter_info_perdomain[domain];
4401 
4402 	/*
4403 	 * If we have any perf counters to read, read them all now, in bulk
4404 	 */
4405 	if (rci->fd_perf != -1) {
4406 		size_t num_perf_counters = rapl_counter_info_count_perf(rci);
4407 		const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
4408 		const ssize_t actual_read_size = read(rci->fd_perf, &perf_data[0], sizeof(perf_data));
4409 
4410 		if (actual_read_size != expected_read_size)
4411 			err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size,
4412 			    actual_read_size);
4413 	}
4414 
4415 	for (unsigned int i = 0, pi = 1; i < NUM_RAPL_COUNTERS; ++i) {
4416 		switch (rci->source[i]) {
4417 		case COUNTER_SOURCE_NONE:
4418 			rci->data[i] = 0;
4419 			break;
4420 
4421 		case COUNTER_SOURCE_PERF:
4422 			assert(pi < ARRAY_SIZE(perf_data));
4423 			assert(rci->fd_perf != -1);
4424 
4425 			if (debug >= 2)
4426 				fprintf(stderr, "Reading rapl counter via perf at %u (%llu %e %lf)\n",
4427 					i, perf_data[pi], rci->scale[i], perf_data[pi] * rci->scale[i]);
4428 
4429 			rci->data[i] = perf_data[pi];
4430 
4431 			++pi;
4432 			break;
4433 
4434 		case COUNTER_SOURCE_MSR:
4435 			if (debug >= 2)
4436 				fprintf(stderr, "Reading rapl counter via msr at %u\n", i);
4437 
4438 			assert(!no_msr);
4439 			if (rci->flags[i] & RAPL_COUNTER_FLAG_USE_MSR_SUM) {
4440 				if (get_msr_sum(cpu, rci->msr[i], &rci->data[i]))
4441 					return -13 - i;
4442 			} else {
4443 				if (get_msr(cpu, rci->msr[i], &rci->data[i]))
4444 					return -13 - i;
4445 			}
4446 
4447 			rci->data[i] &= rci->msr_mask[i];
4448 			if (rci->msr_shift[i] >= 0)
4449 				rci->data[i] >>= abs(rci->msr_shift[i]);
4450 			else
4451 				rci->data[i] <<= abs(rci->msr_shift[i]);
4452 
4453 			break;
4454 		}
4455 	}
4456 
4457 	BUILD_BUG_ON(NUM_RAPL_COUNTERS != 8);
4458 	write_rapl_counter(&p->energy_pkg, rci, RAPL_RCI_INDEX_ENERGY_PKG);
4459 	write_rapl_counter(&p->energy_cores, rci, RAPL_RCI_INDEX_ENERGY_CORES);
4460 	write_rapl_counter(&p->energy_dram, rci, RAPL_RCI_INDEX_DRAM);
4461 	write_rapl_counter(&p->energy_gfx, rci, RAPL_RCI_INDEX_GFX);
4462 	write_rapl_counter(&p->rapl_pkg_perf_status, rci, RAPL_RCI_INDEX_PKG_PERF_STATUS);
4463 	write_rapl_counter(&p->rapl_dram_perf_status, rci, RAPL_RCI_INDEX_DRAM_PERF_STATUS);
4464 	write_rapl_counter(&c->core_energy, rci, RAPL_RCI_INDEX_CORE_ENERGY);
4465 	write_rapl_counter(&pplat_cnt->energy_psys, rci, RAPL_RCI_INDEX_ENERGY_PLATFORM);
4466 
4467 	return 0;
4468 }
4469 
find_sysfs_path_by_id(struct sysfs_path * sp,int id)4470 char *find_sysfs_path_by_id(struct sysfs_path *sp, int id)
4471 {
4472 	while (sp) {
4473 		if (sp->id == id)
4474 			return (sp->path);
4475 		sp = sp->next;
4476 	}
4477 	if (debug)
4478 		warnx("%s: id%d not found", __func__, id);
4479 	return NULL;
4480 }
4481 
get_cstate_counters(unsigned int cpu,struct thread_data * t,struct core_data * c,struct pkg_data * p)4482 int get_cstate_counters(unsigned int cpu, struct thread_data *t, struct core_data *c, struct pkg_data *p)
4483 {
4484 	/*
4485 	 * Overcommit memory a little bit here,
4486 	 * but skip calculating exact sizes for the buffers.
4487 	 */
4488 	unsigned long long perf_data[NUM_CSTATE_COUNTERS];
4489 	unsigned long long perf_data_core[NUM_CSTATE_COUNTERS + 1];
4490 	unsigned long long perf_data_pkg[NUM_CSTATE_COUNTERS + 1];
4491 
4492 	struct cstate_counter_info_t *cci;
4493 
4494 	if (debug >= 2)
4495 		fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
4496 
4497 	assert(ccstate_counter_info);
4498 	assert(cpu <= ccstate_counter_info_size);
4499 
4500 	ZERO_ARRAY(perf_data);
4501 	ZERO_ARRAY(perf_data_core);
4502 	ZERO_ARRAY(perf_data_pkg);
4503 
4504 	cci = &ccstate_counter_info[cpu];
4505 
4506 	/*
4507 	 * If we have any perf counters to read, read them all now, in bulk
4508 	 */
4509 	const size_t num_perf_counters = cstate_counter_info_count_perf(cci);
4510 	ssize_t expected_read_size = num_perf_counters * sizeof(unsigned long long);
4511 	ssize_t actual_read_size_core = 0, actual_read_size_pkg = 0;
4512 
4513 	if (cci->fd_perf_core != -1) {
4514 		/* Each descriptor read begins with number of counters read. */
4515 		expected_read_size += sizeof(unsigned long long);
4516 
4517 		actual_read_size_core = read(cci->fd_perf_core, &perf_data_core[0], sizeof(perf_data_core));
4518 
4519 		if (actual_read_size_core <= 0)
4520 			err(-1, "%s: read perf %s: %ld", __func__, "core", actual_read_size_core);
4521 	}
4522 
4523 	if (cci->fd_perf_pkg != -1) {
4524 		/* Each descriptor read begins with number of counters read. */
4525 		expected_read_size += sizeof(unsigned long long);
4526 
4527 		actual_read_size_pkg = read(cci->fd_perf_pkg, &perf_data_pkg[0], sizeof(perf_data_pkg));
4528 
4529 		if (actual_read_size_pkg <= 0)
4530 			err(-1, "%s: read perf %s: %ld", __func__, "pkg", actual_read_size_pkg);
4531 	}
4532 
4533 	const ssize_t actual_read_size_total = actual_read_size_core + actual_read_size_pkg;
4534 
4535 	if (actual_read_size_total != expected_read_size)
4536 		err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size, actual_read_size_total);
4537 
4538 	/*
4539 	 * Copy ccstate and pcstate data into unified buffer.
4540 	 *
4541 	 * Skip first element from core and pkg buffers.
4542 	 * Kernel puts there how many counters were read.
4543 	 */
4544 	const size_t num_core_counters = perf_data_core[0];
4545 	const size_t num_pkg_counters = perf_data_pkg[0];
4546 
4547 	assert(num_perf_counters == num_core_counters + num_pkg_counters);
4548 
4549 	/* Copy ccstate perf data */
4550 	memcpy(&perf_data[0], &perf_data_core[1], num_core_counters * sizeof(unsigned long long));
4551 
4552 	/* Copy pcstate perf data */
4553 	memcpy(&perf_data[num_core_counters], &perf_data_pkg[1], num_pkg_counters * sizeof(unsigned long long));
4554 
4555 	for (unsigned int i = 0, pi = 0; i < NUM_CSTATE_COUNTERS; ++i) {
4556 		switch (cci->source[i]) {
4557 		case COUNTER_SOURCE_NONE:
4558 			break;
4559 
4560 		case COUNTER_SOURCE_PERF:
4561 			assert(pi < ARRAY_SIZE(perf_data));
4562 			assert(cci->fd_perf_core != -1 || cci->fd_perf_pkg != -1);
4563 
4564 			if (debug >= 2)
4565 				fprintf(stderr, "cstate via %s %u: %llu\n", "perf", i, perf_data[pi]);
4566 
4567 			cci->data[i] = perf_data[pi];
4568 
4569 			++pi;
4570 			break;
4571 
4572 		case COUNTER_SOURCE_MSR:
4573 			assert(!no_msr);
4574 			if (get_msr(cpu, cci->msr[i], &cci->data[i]))
4575 				return -13 - i;
4576 
4577 			if (debug >= 2)
4578 				fprintf(stderr, "cstate via %s0x%llx %u: %llu\n", "msr", cci->msr[i], i, cci->data[i]);
4579 
4580 			break;
4581 		}
4582 	}
4583 
4584 	/*
4585 	 * Helper to write the data only if the source of
4586 	 * the counter for the current cpu is not none.
4587 	 *
4588 	 * Otherwise we would overwrite core data with 0 (default value),
4589 	 * when invoked for the thread sibling.
4590 	 */
4591 #define PERF_COUNTER_WRITE_DATA(out_counter, index) do {	\
4592 	if (cci->source[index] != COUNTER_SOURCE_NONE)		\
4593 		out_counter = cci->data[index];			\
4594 } while (0)
4595 
4596 	BUILD_BUG_ON(NUM_CSTATE_COUNTERS != 11);
4597 
4598 	PERF_COUNTER_WRITE_DATA(t->c1, CCSTATE_RCI_INDEX_C1_RESIDENCY);
4599 	PERF_COUNTER_WRITE_DATA(c->c3, CCSTATE_RCI_INDEX_C3_RESIDENCY);
4600 	PERF_COUNTER_WRITE_DATA(c->c6, CCSTATE_RCI_INDEX_C6_RESIDENCY);
4601 	PERF_COUNTER_WRITE_DATA(c->c7, CCSTATE_RCI_INDEX_C7_RESIDENCY);
4602 
4603 	PERF_COUNTER_WRITE_DATA(p->pc2, PCSTATE_RCI_INDEX_C2_RESIDENCY);
4604 	PERF_COUNTER_WRITE_DATA(p->pc3, PCSTATE_RCI_INDEX_C3_RESIDENCY);
4605 	PERF_COUNTER_WRITE_DATA(p->pc6, PCSTATE_RCI_INDEX_C6_RESIDENCY);
4606 	PERF_COUNTER_WRITE_DATA(p->pc7, PCSTATE_RCI_INDEX_C7_RESIDENCY);
4607 	PERF_COUNTER_WRITE_DATA(p->pc8, PCSTATE_RCI_INDEX_C8_RESIDENCY);
4608 	PERF_COUNTER_WRITE_DATA(p->pc9, PCSTATE_RCI_INDEX_C9_RESIDENCY);
4609 	PERF_COUNTER_WRITE_DATA(p->pc10, PCSTATE_RCI_INDEX_C10_RESIDENCY);
4610 
4611 #undef PERF_COUNTER_WRITE_DATA
4612 
4613 	return 0;
4614 }
4615 
msr_counter_info_count_perf(const struct msr_counter_info_t * mci)4616 size_t msr_counter_info_count_perf(const struct msr_counter_info_t *mci)
4617 {
4618 	size_t ret = 0;
4619 
4620 	for (int i = 0; i < NUM_MSR_COUNTERS; ++i)
4621 		if (mci->source[i] == COUNTER_SOURCE_PERF)
4622 			++ret;
4623 
4624 	return ret;
4625 }
4626 
get_smi_aperf_mperf(unsigned int cpu,struct thread_data * t)4627 int get_smi_aperf_mperf(unsigned int cpu, struct thread_data *t)
4628 {
4629 	unsigned long long perf_data[NUM_MSR_COUNTERS + 1];
4630 
4631 	struct msr_counter_info_t *mci;
4632 
4633 	if (debug >= 2)
4634 		fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
4635 
4636 	assert(msr_counter_info);
4637 	assert(cpu <= msr_counter_info_size);
4638 
4639 	mci = &msr_counter_info[cpu];
4640 
4641 	ZERO_ARRAY(perf_data);
4642 	ZERO_ARRAY(mci->data);
4643 
4644 	if (mci->fd_perf != -1) {
4645 		const size_t num_perf_counters = msr_counter_info_count_perf(mci);
4646 		const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
4647 		const ssize_t actual_read_size = read(mci->fd_perf, &perf_data[0], sizeof(perf_data));
4648 
4649 		if (actual_read_size != expected_read_size)
4650 			err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size,
4651 			    actual_read_size);
4652 	}
4653 
4654 	for (unsigned int i = 0, pi = 1; i < NUM_MSR_COUNTERS; ++i) {
4655 		switch (mci->source[i]) {
4656 		case COUNTER_SOURCE_NONE:
4657 			break;
4658 
4659 		case COUNTER_SOURCE_PERF:
4660 			assert(pi < ARRAY_SIZE(perf_data));
4661 			assert(mci->fd_perf != -1);
4662 
4663 			if (debug >= 2)
4664 				fprintf(stderr, "Reading msr counter via perf at %u: %llu\n", i, perf_data[pi]);
4665 
4666 			mci->data[i] = perf_data[pi];
4667 
4668 			++pi;
4669 			break;
4670 
4671 		case COUNTER_SOURCE_MSR:
4672 			assert(!no_msr);
4673 
4674 			if (get_msr(cpu, mci->msr[i], &mci->data[i]))
4675 				return -2 - i;
4676 
4677 			mci->data[i] &= mci->msr_mask[i];
4678 
4679 			if (debug >= 2)
4680 				fprintf(stderr, "Reading msr counter via msr at %u: %llu\n", i, mci->data[i]);
4681 
4682 			break;
4683 		}
4684 	}
4685 
4686 	BUILD_BUG_ON(NUM_MSR_COUNTERS != 3);
4687 	t->aperf = mci->data[MSR_RCI_INDEX_APERF];
4688 	t->mperf = mci->data[MSR_RCI_INDEX_MPERF];
4689 	t->smi_count = mci->data[MSR_RCI_INDEX_SMI];
4690 
4691 	return 0;
4692 }
4693 
perf_counter_info_read_values(struct perf_counter_info * pp,int cpu,unsigned long long * out,size_t out_size)4694 int perf_counter_info_read_values(struct perf_counter_info *pp, int cpu, unsigned long long *out, size_t out_size)
4695 {
4696 	unsigned int domain;
4697 	unsigned long long value;
4698 	int fd_counter;
4699 
4700 	for (size_t i = 0; pp; ++i, pp = pp->next) {
4701 		domain = cpu_to_domain(pp, cpu);
4702 		assert(domain < pp->num_domains);
4703 
4704 		fd_counter = pp->fd_perf_per_domain[domain];
4705 
4706 		if (fd_counter == -1)
4707 			continue;
4708 
4709 		if (read(fd_counter, &value, sizeof(value)) != sizeof(value))
4710 			return 1;
4711 
4712 		assert(i < out_size);
4713 		out[i] = value * pp->scale;
4714 	}
4715 
4716 	return 0;
4717 }
4718 
pmt_gen_value_mask(unsigned int lsb,unsigned int msb)4719 unsigned long pmt_gen_value_mask(unsigned int lsb, unsigned int msb)
4720 {
4721 	unsigned long mask;
4722 
4723 	if (msb == 63)
4724 		mask = 0xffffffffffffffff;
4725 	else
4726 		mask = ((1 << (msb + 1)) - 1);
4727 
4728 	mask -= (1 << lsb) - 1;
4729 
4730 	return mask;
4731 }
4732 
pmt_read_counter(struct pmt_counter * ppmt,unsigned int domain_id)4733 unsigned long pmt_read_counter(struct pmt_counter *ppmt, unsigned int domain_id)
4734 {
4735 	if (domain_id >= ppmt->num_domains)
4736 		return 0;
4737 
4738 	const unsigned long *pmmio = ppmt->domains[domain_id].pcounter;
4739 	const unsigned long value = pmmio ? *pmmio : 0;
4740 	const unsigned long value_mask = pmt_gen_value_mask(ppmt->lsb, ppmt->msb);
4741 	const unsigned long value_shift = ppmt->lsb;
4742 
4743 	return (value & value_mask) >> value_shift;
4744 }
4745 
4746 /*
4747  * get_counters(...)
4748  * migrate to cpu
4749  * acquire and record local counters for that cpu
4750  */
get_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)4751 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4752 {
4753 	int cpu = t->cpu_id;
4754 	unsigned long long msr;
4755 	struct msr_counter *mp;
4756 	struct pmt_counter *pp;
4757 	int i;
4758 	int status;
4759 
4760 	if (cpu_migrate(cpu)) {
4761 		fprintf(outf, "%s: Could not migrate to CPU %d\n", __func__, cpu);
4762 		return -1;
4763 	}
4764 
4765 	gettimeofday(&t->tv_begin, (struct timezone *)NULL);
4766 
4767 	if (first_counter_read)
4768 		get_apic_id(t);
4769 
4770 	t->tsc = rdtsc();	/* we are running on local CPU of interest */
4771 
4772 	get_smi_aperf_mperf(cpu, t);
4773 
4774 	if (DO_BIC(BIC_IPC))
4775 		if (read(get_instr_count_fd(cpu), &t->instr_count, sizeof(long long)) != sizeof(long long))
4776 			return -4;
4777 
4778 	if (DO_BIC(BIC_IRQ))
4779 		t->irq_count = irqs_per_cpu[cpu];
4780 	if (DO_BIC(BIC_NMI))
4781 		t->nmi_count = nmi_per_cpu[cpu];
4782 
4783 	get_cstate_counters(cpu, t, c, p);
4784 
4785 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
4786 		if (get_mp(cpu, mp, &t->counter[i], mp->sp->path))
4787 			return -10;
4788 	}
4789 
4790 	if (perf_counter_info_read_values(sys.perf_tp, cpu, t->perf_counter, MAX_ADDED_THREAD_COUNTERS))
4791 		return -10;
4792 
4793 	for (i = 0, pp = sys.pmt_tp; pp; i++, pp = pp->next)
4794 		t->pmt_counter[i] = pmt_read_counter(pp, t->cpu_id);
4795 
4796 	/* collect core counters only for 1st thread in core */
4797 	if (!is_cpu_first_thread_in_core(t, c, p))
4798 		goto done;
4799 
4800 	if (platform->has_per_core_rapl) {
4801 		status = get_rapl_counters(cpu, c->core_id, c, p);
4802 		if (status != 0)
4803 			return status;
4804 	}
4805 
4806 	if (DO_BIC(BIC_CPU_c7) && t->is_atom) {
4807 		/*
4808 		 * For Atom CPUs that has core cstate deeper than c6,
4809 		 * MSR_CORE_C6_RESIDENCY returns residency of cc6 and deeper.
4810 		 * Minus CC7 (and deeper cstates) residency to get
4811 		 * accturate cc6 residency.
4812 		 */
4813 		c->c6 -= c->c7;
4814 	}
4815 
4816 	if (DO_BIC(BIC_Mod_c6))
4817 		if (get_msr(cpu, MSR_MODULE_C6_RES_MS, &c->mc6_us))
4818 			return -8;
4819 
4820 	if (DO_BIC(BIC_CoreTmp)) {
4821 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
4822 			return -9;
4823 		c->core_temp_c = tj_max - ((msr >> 16) & 0x7F);
4824 	}
4825 
4826 	if (DO_BIC(BIC_CORE_THROT_CNT))
4827 		get_core_throt_cnt(cpu, &c->core_throt_cnt);
4828 
4829 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
4830 		if (get_mp(cpu, mp, &c->counter[i], mp->sp->path))
4831 			return -10;
4832 	}
4833 
4834 	if (perf_counter_info_read_values(sys.perf_cp, cpu, c->perf_counter, MAX_ADDED_CORE_COUNTERS))
4835 		return -10;
4836 
4837 	for (i = 0, pp = sys.pmt_cp; pp; i++, pp = pp->next)
4838 		c->pmt_counter[i] = pmt_read_counter(pp, c->core_id);
4839 
4840 	/* collect package counters only for 1st core in package */
4841 	if (!is_cpu_first_core_in_package(t, c, p))
4842 		goto done;
4843 
4844 	if (DO_BIC(BIC_Totl_c0)) {
4845 		if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
4846 			return -10;
4847 	}
4848 	if (DO_BIC(BIC_Any_c0)) {
4849 		if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
4850 			return -11;
4851 	}
4852 	if (DO_BIC(BIC_GFX_c0)) {
4853 		if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
4854 			return -12;
4855 	}
4856 	if (DO_BIC(BIC_CPUGFX)) {
4857 		if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
4858 			return -13;
4859 	}
4860 
4861 	if (DO_BIC(BIC_CPU_LPI))
4862 		p->cpu_lpi = cpuidle_cur_cpu_lpi_us;
4863 	if (DO_BIC(BIC_SYS_LPI))
4864 		p->sys_lpi = cpuidle_cur_sys_lpi_us;
4865 
4866 	if (!platform->has_per_core_rapl) {
4867 		status = get_rapl_counters(cpu, p->package_id, c, p);
4868 		if (status != 0)
4869 			return status;
4870 	}
4871 
4872 	if (DO_BIC(BIC_PkgTmp)) {
4873 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
4874 			return -17;
4875 		p->pkg_temp_c = tj_max - ((msr >> 16) & 0x7F);
4876 	}
4877 
4878 	if (DO_BIC(BIC_UNCORE_MHZ))
4879 		p->uncore_mhz = get_legacy_uncore_mhz(p->package_id);
4880 
4881 	if (DO_BIC(BIC_GFX_rc6))
4882 		p->gfx_rc6_ms = gfx_info[GFX_rc6].val_ull;
4883 
4884 	if (DO_BIC(BIC_GFXMHz))
4885 		p->gfx_mhz = gfx_info[GFX_MHz].val;
4886 
4887 	if (DO_BIC(BIC_GFXACTMHz))
4888 		p->gfx_act_mhz = gfx_info[GFX_ACTMHz].val;
4889 
4890 	if (DO_BIC(BIC_SAM_mc6))
4891 		p->sam_mc6_ms = gfx_info[SAM_mc6].val_ull;
4892 
4893 	if (DO_BIC(BIC_SAMMHz))
4894 		p->sam_mhz = gfx_info[SAM_MHz].val;
4895 
4896 	if (DO_BIC(BIC_SAMACTMHz))
4897 		p->sam_act_mhz = gfx_info[SAM_ACTMHz].val;
4898 
4899 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
4900 		char *path = NULL;
4901 
4902 		if (mp->msr_num == 0) {
4903 			path = find_sysfs_path_by_id(mp->sp, p->package_id);
4904 			if (path == NULL) {
4905 				warnx("%s: package_id %d not found", __func__, p->package_id);
4906 				return -10;
4907 			}
4908 		}
4909 		if (get_mp(cpu, mp, &p->counter[i], path))
4910 			return -10;
4911 	}
4912 
4913 	if (perf_counter_info_read_values(sys.perf_pp, cpu, p->perf_counter, MAX_ADDED_PACKAGE_COUNTERS))
4914 		return -10;
4915 
4916 	for (i = 0, pp = sys.pmt_pp; pp; i++, pp = pp->next)
4917 		p->pmt_counter[i] = pmt_read_counter(pp, p->package_id);
4918 
4919 done:
4920 	gettimeofday(&t->tv_end, (struct timezone *)NULL);
4921 
4922 	return 0;
4923 }
4924 
4925 int pkg_cstate_limit = PCLUKN;
4926 char *pkg_cstate_limit_strings[] = { "unknown", "reserved", "pc0", "pc1", "pc2",
4927 	"pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "pc10", "unlimited"
4928 };
4929 
4930 int nhm_pkg_cstate_limits[16] =
4931     { PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4932 	PCLRSV, PCLRSV
4933 };
4934 
4935 int snb_pkg_cstate_limits[16] =
4936     { PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4937 	PCLRSV, PCLRSV
4938 };
4939 
4940 int hsw_pkg_cstate_limits[16] =
4941     { PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4942 	PCLRSV, PCLRSV
4943 };
4944 
4945 int slv_pkg_cstate_limits[16] =
4946     { PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4947 	PCL__6, PCL__7
4948 };
4949 
4950 int amt_pkg_cstate_limits[16] =
4951     { PCLUNL, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4952 	PCLRSV, PCLRSV
4953 };
4954 
4955 int phi_pkg_cstate_limits[16] =
4956     { PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4957 	PCLRSV, PCLRSV
4958 };
4959 
4960 int glm_pkg_cstate_limits[16] =
4961     { PCLUNL, PCL__1, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCL_10, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4962 	PCLRSV, PCLRSV
4963 };
4964 
4965 int skx_pkg_cstate_limits[16] =
4966     { PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4967 	PCLRSV, PCLRSV
4968 };
4969 
4970 int icx_pkg_cstate_limits[16] =
4971     { PCL__0, PCL__2, PCL__6, PCL__6, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4972 	PCLRSV, PCLRSV
4973 };
4974 
probe_cst_limit(void)4975 void probe_cst_limit(void)
4976 {
4977 	unsigned long long msr;
4978 	int *pkg_cstate_limits;
4979 
4980 	if (!platform->has_nhm_msrs || no_msr)
4981 		return;
4982 
4983 	switch (platform->cst_limit) {
4984 	case CST_LIMIT_NHM:
4985 		pkg_cstate_limits = nhm_pkg_cstate_limits;
4986 		break;
4987 	case CST_LIMIT_SNB:
4988 		pkg_cstate_limits = snb_pkg_cstate_limits;
4989 		break;
4990 	case CST_LIMIT_HSW:
4991 		pkg_cstate_limits = hsw_pkg_cstate_limits;
4992 		break;
4993 	case CST_LIMIT_SKX:
4994 		pkg_cstate_limits = skx_pkg_cstate_limits;
4995 		break;
4996 	case CST_LIMIT_ICX:
4997 		pkg_cstate_limits = icx_pkg_cstate_limits;
4998 		break;
4999 	case CST_LIMIT_SLV:
5000 		pkg_cstate_limits = slv_pkg_cstate_limits;
5001 		break;
5002 	case CST_LIMIT_AMT:
5003 		pkg_cstate_limits = amt_pkg_cstate_limits;
5004 		break;
5005 	case CST_LIMIT_KNL:
5006 		pkg_cstate_limits = phi_pkg_cstate_limits;
5007 		break;
5008 	case CST_LIMIT_GMT:
5009 		pkg_cstate_limits = glm_pkg_cstate_limits;
5010 		break;
5011 	default:
5012 		return;
5013 	}
5014 
5015 	get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
5016 	pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
5017 }
5018 
dump_platform_info(void)5019 static void dump_platform_info(void)
5020 {
5021 	unsigned long long msr;
5022 	unsigned int ratio;
5023 
5024 	if (!platform->has_nhm_msrs || no_msr)
5025 		return;
5026 
5027 	get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
5028 
5029 	fprintf(outf, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", base_cpu, msr);
5030 
5031 	ratio = (msr >> 40) & 0xFF;
5032 	fprintf(outf, "%d * %.1f = %.1f MHz max efficiency frequency\n", ratio, bclk, ratio * bclk);
5033 
5034 	ratio = (msr >> 8) & 0xFF;
5035 	fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n", ratio, bclk, ratio * bclk);
5036 }
5037 
dump_power_ctl(void)5038 static void dump_power_ctl(void)
5039 {
5040 	unsigned long long msr;
5041 
5042 	if (!platform->has_nhm_msrs || no_msr)
5043 		return;
5044 
5045 	get_msr(base_cpu, MSR_IA32_POWER_CTL, &msr);
5046 	fprintf(outf, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
5047 		base_cpu, msr, msr & 0x2 ? "EN" : "DIS");
5048 
5049 	/* C-state Pre-wake Disable (CSTATE_PREWAKE_DISABLE) */
5050 	if (platform->has_cst_prewake_bit)
5051 		fprintf(outf, "C-state Pre-wake: %sabled\n", msr & 0x40000000 ? "DIS" : "EN");
5052 
5053 	return;
5054 }
5055 
dump_turbo_ratio_limit2(void)5056 static void dump_turbo_ratio_limit2(void)
5057 {
5058 	unsigned long long msr;
5059 	unsigned int ratio;
5060 
5061 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
5062 
5063 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu, msr);
5064 
5065 	ratio = (msr >> 8) & 0xFF;
5066 	if (ratio)
5067 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 18 active cores\n", ratio, bclk, ratio * bclk);
5068 
5069 	ratio = (msr >> 0) & 0xFF;
5070 	if (ratio)
5071 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 17 active cores\n", ratio, bclk, ratio * bclk);
5072 	return;
5073 }
5074 
dump_turbo_ratio_limit1(void)5075 static void dump_turbo_ratio_limit1(void)
5076 {
5077 	unsigned long long msr;
5078 	unsigned int ratio;
5079 
5080 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
5081 
5082 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, msr);
5083 
5084 	ratio = (msr >> 56) & 0xFF;
5085 	if (ratio)
5086 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 16 active cores\n", ratio, bclk, ratio * bclk);
5087 
5088 	ratio = (msr >> 48) & 0xFF;
5089 	if (ratio)
5090 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 15 active cores\n", ratio, bclk, ratio * bclk);
5091 
5092 	ratio = (msr >> 40) & 0xFF;
5093 	if (ratio)
5094 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 14 active cores\n", ratio, bclk, ratio * bclk);
5095 
5096 	ratio = (msr >> 32) & 0xFF;
5097 	if (ratio)
5098 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 13 active cores\n", ratio, bclk, ratio * bclk);
5099 
5100 	ratio = (msr >> 24) & 0xFF;
5101 	if (ratio)
5102 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 12 active cores\n", ratio, bclk, ratio * bclk);
5103 
5104 	ratio = (msr >> 16) & 0xFF;
5105 	if (ratio)
5106 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 11 active cores\n", ratio, bclk, ratio * bclk);
5107 
5108 	ratio = (msr >> 8) & 0xFF;
5109 	if (ratio)
5110 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 10 active cores\n", ratio, bclk, ratio * bclk);
5111 
5112 	ratio = (msr >> 0) & 0xFF;
5113 	if (ratio)
5114 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 9 active cores\n", ratio, bclk, ratio * bclk);
5115 	return;
5116 }
5117 
dump_turbo_ratio_limits(int trl_msr_offset)5118 static void dump_turbo_ratio_limits(int trl_msr_offset)
5119 {
5120 	unsigned long long msr, core_counts;
5121 	int shift;
5122 
5123 	get_msr(base_cpu, trl_msr_offset, &msr);
5124 	fprintf(outf, "cpu%d: MSR_%sTURBO_RATIO_LIMIT: 0x%08llx\n",
5125 		base_cpu, trl_msr_offset == MSR_SECONDARY_TURBO_RATIO_LIMIT ? "SECONDARY_" : "", msr);
5126 
5127 	if (platform->trl_msrs & TRL_CORECOUNT) {
5128 		get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &core_counts);
5129 		fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, core_counts);
5130 	} else {
5131 		core_counts = 0x0807060504030201;
5132 	}
5133 
5134 	for (shift = 56; shift >= 0; shift -= 8) {
5135 		unsigned int ratio, group_size;
5136 
5137 		ratio = (msr >> shift) & 0xFF;
5138 		group_size = (core_counts >> shift) & 0xFF;
5139 		if (ratio)
5140 			fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
5141 				ratio, bclk, ratio * bclk, group_size);
5142 	}
5143 
5144 	return;
5145 }
5146 
dump_atom_turbo_ratio_limits(void)5147 static void dump_atom_turbo_ratio_limits(void)
5148 {
5149 	unsigned long long msr;
5150 	unsigned int ratio;
5151 
5152 	get_msr(base_cpu, MSR_ATOM_CORE_RATIOS, &msr);
5153 	fprintf(outf, "cpu%d: MSR_ATOM_CORE_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
5154 
5155 	ratio = (msr >> 0) & 0x3F;
5156 	if (ratio)
5157 		fprintf(outf, "%d * %.1f = %.1f MHz minimum operating frequency\n", ratio, bclk, ratio * bclk);
5158 
5159 	ratio = (msr >> 8) & 0x3F;
5160 	if (ratio)
5161 		fprintf(outf, "%d * %.1f = %.1f MHz low frequency mode (LFM)\n", ratio, bclk, ratio * bclk);
5162 
5163 	ratio = (msr >> 16) & 0x3F;
5164 	if (ratio)
5165 		fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n", ratio, bclk, ratio * bclk);
5166 
5167 	get_msr(base_cpu, MSR_ATOM_CORE_TURBO_RATIOS, &msr);
5168 	fprintf(outf, "cpu%d: MSR_ATOM_CORE_TURBO_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
5169 
5170 	ratio = (msr >> 24) & 0x3F;
5171 	if (ratio)
5172 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 4 active cores\n", ratio, bclk, ratio * bclk);
5173 
5174 	ratio = (msr >> 16) & 0x3F;
5175 	if (ratio)
5176 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 3 active cores\n", ratio, bclk, ratio * bclk);
5177 
5178 	ratio = (msr >> 8) & 0x3F;
5179 	if (ratio)
5180 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 2 active cores\n", ratio, bclk, ratio * bclk);
5181 
5182 	ratio = (msr >> 0) & 0x3F;
5183 	if (ratio)
5184 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 1 active core\n", ratio, bclk, ratio * bclk);
5185 }
5186 
dump_knl_turbo_ratio_limits(void)5187 static void dump_knl_turbo_ratio_limits(void)
5188 {
5189 	const unsigned int buckets_no = 7;
5190 
5191 	unsigned long long msr;
5192 	int delta_cores, delta_ratio;
5193 	int i, b_nr;
5194 	unsigned int cores[buckets_no];
5195 	unsigned int ratio[buckets_no];
5196 
5197 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
5198 
5199 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu, msr);
5200 
5201 	/*
5202 	 * Turbo encoding in KNL is as follows:
5203 	 * [0] -- Reserved
5204 	 * [7:1] -- Base value of number of active cores of bucket 1.
5205 	 * [15:8] -- Base value of freq ratio of bucket 1.
5206 	 * [20:16] -- +ve delta of number of active cores of bucket 2.
5207 	 * i.e. active cores of bucket 2 =
5208 	 * active cores of bucket 1 + delta
5209 	 * [23:21] -- Negative delta of freq ratio of bucket 2.
5210 	 * i.e. freq ratio of bucket 2 =
5211 	 * freq ratio of bucket 1 - delta
5212 	 * [28:24]-- +ve delta of number of active cores of bucket 3.
5213 	 * [31:29]-- -ve delta of freq ratio of bucket 3.
5214 	 * [36:32]-- +ve delta of number of active cores of bucket 4.
5215 	 * [39:37]-- -ve delta of freq ratio of bucket 4.
5216 	 * [44:40]-- +ve delta of number of active cores of bucket 5.
5217 	 * [47:45]-- -ve delta of freq ratio of bucket 5.
5218 	 * [52:48]-- +ve delta of number of active cores of bucket 6.
5219 	 * [55:53]-- -ve delta of freq ratio of bucket 6.
5220 	 * [60:56]-- +ve delta of number of active cores of bucket 7.
5221 	 * [63:61]-- -ve delta of freq ratio of bucket 7.
5222 	 */
5223 
5224 	b_nr = 0;
5225 	cores[b_nr] = (msr & 0xFF) >> 1;
5226 	ratio[b_nr] = (msr >> 8) & 0xFF;
5227 
5228 	for (i = 16; i < 64; i += 8) {
5229 		delta_cores = (msr >> i) & 0x1F;
5230 		delta_ratio = (msr >> (i + 5)) & 0x7;
5231 
5232 		cores[b_nr + 1] = cores[b_nr] + delta_cores;
5233 		ratio[b_nr + 1] = ratio[b_nr] - delta_ratio;
5234 		b_nr++;
5235 	}
5236 
5237 	for (i = buckets_no - 1; i >= 0; i--)
5238 		if (i > 0 ? ratio[i] != ratio[i - 1] : 1)
5239 			fprintf(outf,
5240 				"%d * %.1f = %.1f MHz max turbo %d active cores\n",
5241 				ratio[i], bclk, ratio[i] * bclk, cores[i]);
5242 }
5243 
dump_cst_cfg(void)5244 static void dump_cst_cfg(void)
5245 {
5246 	unsigned long long msr;
5247 
5248 	if (!platform->has_nhm_msrs || no_msr)
5249 		return;
5250 
5251 	get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
5252 
5253 	fprintf(outf, "cpu%d: MSR_PKG_CST_CONFIG_CONTROL: 0x%08llx", base_cpu, msr);
5254 
5255 	fprintf(outf, " (%s%s%s%s%slocked, pkg-cstate-limit=%d (%s)",
5256 		(msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
5257 		(msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
5258 		(msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
5259 		(msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
5260 		(msr & (1 << 15)) ? "" : "UN", (unsigned int)msr & 0xF, pkg_cstate_limit_strings[pkg_cstate_limit]);
5261 
5262 #define AUTOMATIC_CSTATE_CONVERSION		(1UL << 16)
5263 	if (platform->has_cst_auto_convension) {
5264 		fprintf(outf, ", automatic c-state conversion=%s", (msr & AUTOMATIC_CSTATE_CONVERSION) ? "on" : "off");
5265 	}
5266 
5267 	fprintf(outf, ")\n");
5268 
5269 	return;
5270 }
5271 
dump_config_tdp(void)5272 static void dump_config_tdp(void)
5273 {
5274 	unsigned long long msr;
5275 
5276 	get_msr(base_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
5277 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu, msr);
5278 	fprintf(outf, " (base_ratio=%d)\n", (unsigned int)msr & 0xFF);
5279 
5280 	get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
5281 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu, msr);
5282 	if (msr) {
5283 		fprintf(outf, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
5284 		fprintf(outf, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
5285 		fprintf(outf, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
5286 		fprintf(outf, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0x7FFF);
5287 	}
5288 	fprintf(outf, ")\n");
5289 
5290 	get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
5291 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu, msr);
5292 	if (msr) {
5293 		fprintf(outf, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
5294 		fprintf(outf, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
5295 		fprintf(outf, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
5296 		fprintf(outf, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0x7FFF);
5297 	}
5298 	fprintf(outf, ")\n");
5299 
5300 	get_msr(base_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
5301 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu, msr);
5302 	if ((msr) & 0x3)
5303 		fprintf(outf, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
5304 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
5305 	fprintf(outf, ")\n");
5306 
5307 	get_msr(base_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
5308 	fprintf(outf, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu, msr);
5309 	fprintf(outf, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xFF);
5310 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
5311 	fprintf(outf, ")\n");
5312 }
5313 
5314 unsigned int irtl_time_units[] = { 1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
5315 
print_irtl(void)5316 void print_irtl(void)
5317 {
5318 	unsigned long long msr;
5319 
5320 	if (!platform->has_irtl_msrs || no_msr)
5321 		return;
5322 
5323 	if (platform->supported_cstates & PC3) {
5324 		get_msr(base_cpu, MSR_PKGC3_IRTL, &msr);
5325 		fprintf(outf, "cpu%d: MSR_PKGC3_IRTL: 0x%08llx (", base_cpu, msr);
5326 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5327 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5328 	}
5329 
5330 	if (platform->supported_cstates & PC6) {
5331 		get_msr(base_cpu, MSR_PKGC6_IRTL, &msr);
5332 		fprintf(outf, "cpu%d: MSR_PKGC6_IRTL: 0x%08llx (", base_cpu, msr);
5333 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5334 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5335 	}
5336 
5337 	if (platform->supported_cstates & PC7) {
5338 		get_msr(base_cpu, MSR_PKGC7_IRTL, &msr);
5339 		fprintf(outf, "cpu%d: MSR_PKGC7_IRTL: 0x%08llx (", base_cpu, msr);
5340 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5341 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5342 	}
5343 
5344 	if (platform->supported_cstates & PC8) {
5345 		get_msr(base_cpu, MSR_PKGC8_IRTL, &msr);
5346 		fprintf(outf, "cpu%d: MSR_PKGC8_IRTL: 0x%08llx (", base_cpu, msr);
5347 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5348 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5349 	}
5350 
5351 	if (platform->supported_cstates & PC9) {
5352 		get_msr(base_cpu, MSR_PKGC9_IRTL, &msr);
5353 		fprintf(outf, "cpu%d: MSR_PKGC9_IRTL: 0x%08llx (", base_cpu, msr);
5354 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5355 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5356 	}
5357 
5358 	if (platform->supported_cstates & PC10) {
5359 		get_msr(base_cpu, MSR_PKGC10_IRTL, &msr);
5360 		fprintf(outf, "cpu%d: MSR_PKGC10_IRTL: 0x%08llx (", base_cpu, msr);
5361 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5362 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5363 	}
5364 }
5365 
free_fd_percpu(void)5366 void free_fd_percpu(void)
5367 {
5368 	int i;
5369 
5370 	if (!fd_percpu)
5371 		return;
5372 
5373 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
5374 		if (fd_percpu[i] != 0)
5375 			close(fd_percpu[i]);
5376 	}
5377 
5378 	free(fd_percpu);
5379 	fd_percpu = NULL;
5380 }
5381 
free_fd_instr_count_percpu(void)5382 void free_fd_instr_count_percpu(void)
5383 {
5384 	if (!fd_instr_count_percpu)
5385 		return;
5386 
5387 	for (int i = 0; i < topo.max_cpu_num + 1; ++i) {
5388 		if (fd_instr_count_percpu[i] != 0)
5389 			close(fd_instr_count_percpu[i]);
5390 	}
5391 
5392 	free(fd_instr_count_percpu);
5393 	fd_instr_count_percpu = NULL;
5394 }
5395 
free_fd_cstate(void)5396 void free_fd_cstate(void)
5397 {
5398 	if (!ccstate_counter_info)
5399 		return;
5400 
5401 	const int counter_info_num = ccstate_counter_info_size;
5402 
5403 	for (int counter_id = 0; counter_id < counter_info_num; ++counter_id) {
5404 		if (ccstate_counter_info[counter_id].fd_perf_core != -1)
5405 			close(ccstate_counter_info[counter_id].fd_perf_core);
5406 
5407 		if (ccstate_counter_info[counter_id].fd_perf_pkg != -1)
5408 			close(ccstate_counter_info[counter_id].fd_perf_pkg);
5409 	}
5410 
5411 	free(ccstate_counter_info);
5412 	ccstate_counter_info = NULL;
5413 	ccstate_counter_info_size = 0;
5414 }
5415 
free_fd_msr(void)5416 void free_fd_msr(void)
5417 {
5418 	if (!msr_counter_info)
5419 		return;
5420 
5421 	for (int cpu = 0; cpu < topo.max_cpu_num; ++cpu) {
5422 		if (msr_counter_info[cpu].fd_perf != -1)
5423 			close(msr_counter_info[cpu].fd_perf);
5424 	}
5425 
5426 	free(msr_counter_info);
5427 	msr_counter_info = NULL;
5428 	msr_counter_info_size = 0;
5429 }
5430 
free_fd_rapl_percpu(void)5431 void free_fd_rapl_percpu(void)
5432 {
5433 	if (!rapl_counter_info_perdomain)
5434 		return;
5435 
5436 	const int num_domains = rapl_counter_info_perdomain_size;
5437 
5438 	for (int domain_id = 0; domain_id < num_domains; ++domain_id) {
5439 		if (rapl_counter_info_perdomain[domain_id].fd_perf != -1)
5440 			close(rapl_counter_info_perdomain[domain_id].fd_perf);
5441 	}
5442 
5443 	free(rapl_counter_info_perdomain);
5444 	rapl_counter_info_perdomain = NULL;
5445 	rapl_counter_info_perdomain_size = 0;
5446 }
5447 
free_fd_added_perf_counters_(struct perf_counter_info * pp)5448 void free_fd_added_perf_counters_(struct perf_counter_info *pp)
5449 {
5450 	if (!pp)
5451 		return;
5452 
5453 	if (!pp->fd_perf_per_domain)
5454 		return;
5455 
5456 	while (pp) {
5457 		for (size_t domain = 0; domain < pp->num_domains; ++domain) {
5458 			if (pp->fd_perf_per_domain[domain] != -1) {
5459 				close(pp->fd_perf_per_domain[domain]);
5460 				pp->fd_perf_per_domain[domain] = -1;
5461 			}
5462 		}
5463 
5464 		free(pp->fd_perf_per_domain);
5465 		pp->fd_perf_per_domain = NULL;
5466 
5467 		pp = pp->next;
5468 	}
5469 }
5470 
free_fd_added_perf_counters(void)5471 void free_fd_added_perf_counters(void)
5472 {
5473 	free_fd_added_perf_counters_(sys.perf_tp);
5474 	free_fd_added_perf_counters_(sys.perf_cp);
5475 	free_fd_added_perf_counters_(sys.perf_pp);
5476 }
5477 
free_all_buffers(void)5478 void free_all_buffers(void)
5479 {
5480 	int i;
5481 
5482 	CPU_FREE(cpu_present_set);
5483 	cpu_present_set = NULL;
5484 	cpu_present_setsize = 0;
5485 
5486 	CPU_FREE(cpu_effective_set);
5487 	cpu_effective_set = NULL;
5488 	cpu_effective_setsize = 0;
5489 
5490 	CPU_FREE(cpu_allowed_set);
5491 	cpu_allowed_set = NULL;
5492 	cpu_allowed_setsize = 0;
5493 
5494 	CPU_FREE(cpu_affinity_set);
5495 	cpu_affinity_set = NULL;
5496 	cpu_affinity_setsize = 0;
5497 
5498 	free(thread_even);
5499 	free(core_even);
5500 	free(package_even);
5501 
5502 	thread_even = NULL;
5503 	core_even = NULL;
5504 	package_even = NULL;
5505 
5506 	free(thread_odd);
5507 	free(core_odd);
5508 	free(package_odd);
5509 
5510 	thread_odd = NULL;
5511 	core_odd = NULL;
5512 	package_odd = NULL;
5513 
5514 	free(output_buffer);
5515 	output_buffer = NULL;
5516 	outp = NULL;
5517 
5518 	free_fd_percpu();
5519 	free_fd_instr_count_percpu();
5520 	free_fd_msr();
5521 	free_fd_rapl_percpu();
5522 	free_fd_cstate();
5523 	free_fd_added_perf_counters();
5524 
5525 	free(irq_column_2_cpu);
5526 	free(irqs_per_cpu);
5527 	free(nmi_per_cpu);
5528 
5529 	for (i = 0; i <= topo.max_cpu_num; ++i) {
5530 		if (cpus[i].put_ids)
5531 			CPU_FREE(cpus[i].put_ids);
5532 	}
5533 	free(cpus);
5534 }
5535 
5536 /*
5537  * Parse a file containing a single int.
5538  * Return 0 if file can not be opened
5539  * Exit if file can be opened, but can not be parsed
5540  */
parse_int_file(const char * fmt,...)5541 int parse_int_file(const char *fmt, ...)
5542 {
5543 	va_list args;
5544 	char path[PATH_MAX];
5545 	FILE *filep;
5546 	int value;
5547 
5548 	va_start(args, fmt);
5549 	vsnprintf(path, sizeof(path), fmt, args);
5550 	va_end(args);
5551 	filep = fopen(path, "r");
5552 	if (!filep)
5553 		return 0;
5554 	if (fscanf(filep, "%d", &value) != 1)
5555 		err(1, "%s: failed to parse number from file", path);
5556 	fclose(filep);
5557 	return value;
5558 }
5559 
5560 /*
5561  * cpu_is_first_core_in_package(cpu)
5562  * return 1 if given CPU is 1st core in package
5563  */
cpu_is_first_core_in_package(int cpu)5564 int cpu_is_first_core_in_package(int cpu)
5565 {
5566 	return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
5567 }
5568 
get_physical_package_id(int cpu)5569 int get_physical_package_id(int cpu)
5570 {
5571 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
5572 }
5573 
get_die_id(int cpu)5574 int get_die_id(int cpu)
5575 {
5576 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/die_id", cpu);
5577 }
5578 
get_core_id(int cpu)5579 int get_core_id(int cpu)
5580 {
5581 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
5582 }
5583 
set_node_data(void)5584 void set_node_data(void)
5585 {
5586 	int pkg, node, lnode, cpu, cpux;
5587 	int cpu_count;
5588 
5589 	/* initialize logical_node_id */
5590 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu)
5591 		cpus[cpu].logical_node_id = -1;
5592 
5593 	cpu_count = 0;
5594 	for (pkg = 0; pkg < topo.num_packages; pkg++) {
5595 		lnode = 0;
5596 		for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
5597 			if (cpus[cpu].physical_package_id != pkg)
5598 				continue;
5599 			/* find a cpu with an unset logical_node_id */
5600 			if (cpus[cpu].logical_node_id != -1)
5601 				continue;
5602 			cpus[cpu].logical_node_id = lnode;
5603 			node = cpus[cpu].physical_node_id;
5604 			cpu_count++;
5605 			/*
5606 			 * find all matching cpus on this pkg and set
5607 			 * the logical_node_id
5608 			 */
5609 			for (cpux = cpu; cpux <= topo.max_cpu_num; cpux++) {
5610 				if ((cpus[cpux].physical_package_id == pkg) && (cpus[cpux].physical_node_id == node)) {
5611 					cpus[cpux].logical_node_id = lnode;
5612 					cpu_count++;
5613 				}
5614 			}
5615 			lnode++;
5616 			if (lnode > topo.nodes_per_pkg)
5617 				topo.nodes_per_pkg = lnode;
5618 		}
5619 		if (cpu_count >= topo.max_cpu_num)
5620 			break;
5621 	}
5622 }
5623 
get_physical_node_id(struct cpu_topology * thiscpu)5624 int get_physical_node_id(struct cpu_topology *thiscpu)
5625 {
5626 	char path[80];
5627 	FILE *filep;
5628 	int i;
5629 	int cpu = thiscpu->logical_cpu_id;
5630 
5631 	for (i = 0; i <= topo.max_cpu_num; i++) {
5632 		sprintf(path, "/sys/devices/system/cpu/cpu%d/node%i/cpulist", cpu, i);
5633 		filep = fopen(path, "r");
5634 		if (!filep)
5635 			continue;
5636 		fclose(filep);
5637 		return i;
5638 	}
5639 	return -1;
5640 }
5641 
parse_cpu_str(char * cpu_str,cpu_set_t * cpu_set,int cpu_set_size)5642 static int parse_cpu_str(char *cpu_str, cpu_set_t *cpu_set, int cpu_set_size)
5643 {
5644 	unsigned int start, end;
5645 	char *next = cpu_str;
5646 
5647 	while (next && *next) {
5648 
5649 		if (*next == '-')	/* no negative cpu numbers */
5650 			return 1;
5651 
5652 		if (*next == '\0' || *next == '\n')
5653 			break;
5654 
5655 		start = strtoul(next, &next, 10);
5656 
5657 		if (start >= CPU_SUBSET_MAXCPUS)
5658 			return 1;
5659 		CPU_SET_S(start, cpu_set_size, cpu_set);
5660 
5661 		if (*next == '\0' || *next == '\n')
5662 			break;
5663 
5664 		if (*next == ',') {
5665 			next += 1;
5666 			continue;
5667 		}
5668 
5669 		if (*next == '-') {
5670 			next += 1;	/* start range */
5671 		} else if (*next == '.') {
5672 			next += 1;
5673 			if (*next == '.')
5674 				next += 1;	/* start range */
5675 			else
5676 				return 1;
5677 		}
5678 
5679 		end = strtoul(next, &next, 10);
5680 		if (end <= start)
5681 			return 1;
5682 
5683 		while (++start <= end) {
5684 			if (start >= CPU_SUBSET_MAXCPUS)
5685 				return 1;
5686 			CPU_SET_S(start, cpu_set_size, cpu_set);
5687 		}
5688 
5689 		if (*next == ',')
5690 			next += 1;
5691 		else if (*next != '\0' && *next != '\n')
5692 			return 1;
5693 	}
5694 
5695 	return 0;
5696 }
5697 
get_thread_siblings(struct cpu_topology * thiscpu)5698 int get_thread_siblings(struct cpu_topology *thiscpu)
5699 {
5700 	char path[80], character;
5701 	FILE *filep;
5702 	unsigned long map;
5703 	int so, shift, sib_core;
5704 	int cpu = thiscpu->logical_cpu_id;
5705 	int offset = topo.max_cpu_num + 1;
5706 	size_t size;
5707 	int thread_id = 0;
5708 
5709 	thiscpu->put_ids = CPU_ALLOC((topo.max_cpu_num + 1));
5710 	if (thiscpu->thread_id < 0)
5711 		thiscpu->thread_id = thread_id++;
5712 	if (!thiscpu->put_ids)
5713 		return -1;
5714 
5715 	size = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
5716 	CPU_ZERO_S(size, thiscpu->put_ids);
5717 
5718 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", cpu);
5719 	filep = fopen(path, "r");
5720 
5721 	if (!filep) {
5722 		warnx("%s: open failed", path);
5723 		return -1;
5724 	}
5725 	do {
5726 		offset -= BITMASK_SIZE;
5727 		if (fscanf(filep, "%lx%c", &map, &character) != 2)
5728 			err(1, "%s: failed to parse file", path);
5729 		for (shift = 0; shift < BITMASK_SIZE; shift++) {
5730 			if ((map >> shift) & 0x1) {
5731 				so = shift + offset;
5732 				sib_core = get_core_id(so);
5733 				if (sib_core == thiscpu->physical_core_id) {
5734 					CPU_SET_S(so, size, thiscpu->put_ids);
5735 					if ((so != cpu) && (cpus[so].thread_id < 0))
5736 						cpus[so].thread_id = thread_id++;
5737 				}
5738 			}
5739 		}
5740 	} while (character == ',');
5741 	fclose(filep);
5742 
5743 	return CPU_COUNT_S(size, thiscpu->put_ids);
5744 }
5745 
5746 /*
5747  * run func(thread, core, package) in topology order
5748  * skip non-present cpus
5749  */
5750 
for_all_cpus_2(int (func)(struct thread_data *,struct core_data *,struct pkg_data *,struct thread_data *,struct core_data *,struct pkg_data *),struct thread_data * thread_base,struct core_data * core_base,struct pkg_data * pkg_base,struct thread_data * thread_base2,struct core_data * core_base2,struct pkg_data * pkg_base2)5751 int for_all_cpus_2(int (func) (struct thread_data *, struct core_data *,
5752 			       struct pkg_data *, struct thread_data *, struct core_data *,
5753 			       struct pkg_data *), struct thread_data *thread_base,
5754 		   struct core_data *core_base, struct pkg_data *pkg_base,
5755 		   struct thread_data *thread_base2, struct core_data *core_base2, struct pkg_data *pkg_base2)
5756 {
5757 	int retval, pkg_no, node_no, core_no, thread_no;
5758 
5759 	retval = 0;
5760 
5761 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
5762 		for (node_no = 0; node_no < topo.nodes_per_pkg; ++node_no) {
5763 			for (core_no = 0; core_no < topo.cores_per_node; ++core_no) {
5764 				for (thread_no = 0; thread_no < topo.threads_per_core; ++thread_no) {
5765 					struct thread_data *t, *t2;
5766 					struct core_data *c, *c2;
5767 					struct pkg_data *p, *p2;
5768 
5769 					t = GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no);
5770 
5771 					if (cpu_is_not_allowed(t->cpu_id))
5772 						continue;
5773 
5774 					t2 = GET_THREAD(thread_base2, thread_no, core_no, node_no, pkg_no);
5775 
5776 					c = GET_CORE(core_base, core_no, node_no, pkg_no);
5777 					c2 = GET_CORE(core_base2, core_no, node_no, pkg_no);
5778 
5779 					p = GET_PKG(pkg_base, pkg_no);
5780 					p2 = GET_PKG(pkg_base2, pkg_no);
5781 
5782 					retval |= func(t, c, p, t2, c2, p2);
5783 				}
5784 			}
5785 		}
5786 	}
5787 	return retval;
5788 }
5789 
5790 /*
5791  * run func(cpu) on every cpu in /proc/stat
5792  * return max_cpu number
5793  */
for_all_proc_cpus(int (func)(int))5794 int for_all_proc_cpus(int (func) (int))
5795 {
5796 	FILE *fp;
5797 	int cpu_num;
5798 	int retval;
5799 
5800 	fp = fopen_or_die(proc_stat, "r");
5801 
5802 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
5803 	if (retval != 0)
5804 		err(1, "%s: failed to parse format", proc_stat);
5805 
5806 	while (1) {
5807 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
5808 		if (retval != 1)
5809 			break;
5810 
5811 		retval = func(cpu_num);
5812 		if (retval) {
5813 			fclose(fp);
5814 			return (retval);
5815 		}
5816 	}
5817 	fclose(fp);
5818 	return 0;
5819 }
5820 
5821 #define PATH_EFFECTIVE_CPUS	"/sys/fs/cgroup/cpuset.cpus.effective"
5822 
5823 static char cpu_effective_str[1024];
5824 
update_effective_str(bool startup)5825 static int update_effective_str(bool startup)
5826 {
5827 	FILE *fp;
5828 	char *pos;
5829 	char buf[1024];
5830 	int ret;
5831 
5832 	if (cpu_effective_str[0] == '\0' && !startup)
5833 		return 0;
5834 
5835 	fp = fopen(PATH_EFFECTIVE_CPUS, "r");
5836 	if (!fp)
5837 		return 0;
5838 
5839 	pos = fgets(buf, 1024, fp);
5840 	if (!pos)
5841 		err(1, "%s: file read failed\n", PATH_EFFECTIVE_CPUS);
5842 
5843 	fclose(fp);
5844 
5845 	ret = strncmp(cpu_effective_str, buf, 1024);
5846 	if (!ret)
5847 		return 0;
5848 
5849 	strncpy(cpu_effective_str, buf, 1024);
5850 	return 1;
5851 }
5852 
update_effective_set(bool startup)5853 static void update_effective_set(bool startup)
5854 {
5855 	update_effective_str(startup);
5856 
5857 	if (parse_cpu_str(cpu_effective_str, cpu_effective_set, cpu_effective_setsize))
5858 		err(1, "%s: cpu str malformat %s\n", PATH_EFFECTIVE_CPUS, cpu_effective_str);
5859 }
5860 
5861 void linux_perf_init(void);
5862 void msr_perf_init(void);
5863 void rapl_perf_init(void);
5864 void cstate_perf_init(void);
5865 void added_perf_counters_init(void);
5866 void pmt_init(void);
5867 
re_initialize(void)5868 void re_initialize(void)
5869 {
5870 	free_all_buffers();
5871 	setup_all_buffers(false);
5872 	linux_perf_init();
5873 	msr_perf_init();
5874 	rapl_perf_init();
5875 	cstate_perf_init();
5876 	added_perf_counters_init();
5877 	pmt_init();
5878 	fprintf(outf, "turbostat: re-initialized with num_cpus %d, allowed_cpus %d\n", topo.num_cpus,
5879 		topo.allowed_cpus);
5880 }
5881 
set_max_cpu_num(void)5882 void set_max_cpu_num(void)
5883 {
5884 	FILE *filep;
5885 	int base_cpu;
5886 	unsigned long dummy;
5887 	char pathname[64];
5888 
5889 	base_cpu = sched_getcpu();
5890 	if (base_cpu < 0)
5891 		err(1, "cannot find calling cpu ID");
5892 	sprintf(pathname, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", base_cpu);
5893 
5894 	filep = fopen_or_die(pathname, "r");
5895 	topo.max_cpu_num = 0;
5896 	while (fscanf(filep, "%lx,", &dummy) == 1)
5897 		topo.max_cpu_num += BITMASK_SIZE;
5898 	fclose(filep);
5899 	topo.max_cpu_num--;	/* 0 based */
5900 }
5901 
5902 /*
5903  * count_cpus()
5904  * remember the last one seen, it will be the max
5905  */
count_cpus(int cpu)5906 int count_cpus(int cpu)
5907 {
5908 	UNUSED(cpu);
5909 
5910 	topo.num_cpus++;
5911 	return 0;
5912 }
5913 
mark_cpu_present(int cpu)5914 int mark_cpu_present(int cpu)
5915 {
5916 	CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
5917 	return 0;
5918 }
5919 
init_thread_id(int cpu)5920 int init_thread_id(int cpu)
5921 {
5922 	cpus[cpu].thread_id = -1;
5923 	return 0;
5924 }
5925 
set_my_cpu_type(void)5926 int set_my_cpu_type(void)
5927 {
5928 	unsigned int eax, ebx, ecx, edx;
5929 	unsigned int max_level;
5930 
5931 	__cpuid(0, max_level, ebx, ecx, edx);
5932 
5933 	if (max_level < CPUID_LEAF_MODEL_ID)
5934 		return 0;
5935 
5936 	__cpuid(CPUID_LEAF_MODEL_ID, eax, ebx, ecx, edx);
5937 
5938 	return (eax >> CPUID_LEAF_MODEL_ID_CORE_TYPE_SHIFT);
5939 }
5940 
set_cpu_hybrid_type(int cpu)5941 int set_cpu_hybrid_type(int cpu)
5942 {
5943 	if (cpu_migrate(cpu))
5944 		return -1;
5945 
5946 	int type = set_my_cpu_type();
5947 
5948 	cpus[cpu].type = type;
5949 	return 0;
5950 }
5951 
5952 /*
5953  * snapshot_proc_interrupts()
5954  *
5955  * read and record summary of /proc/interrupts
5956  *
5957  * return 1 if config change requires a restart, else return 0
5958  */
snapshot_proc_interrupts(void)5959 int snapshot_proc_interrupts(void)
5960 {
5961 	static FILE *fp;
5962 	int column, retval;
5963 
5964 	if (fp == NULL)
5965 		fp = fopen_or_die("/proc/interrupts", "r");
5966 	else
5967 		rewind(fp);
5968 
5969 	/* read 1st line of /proc/interrupts to get cpu* name for each column */
5970 	for (column = 0; column < topo.num_cpus; ++column) {
5971 		int cpu_number;
5972 
5973 		retval = fscanf(fp, " CPU%d", &cpu_number);
5974 		if (retval != 1)
5975 			break;
5976 
5977 		if (cpu_number > topo.max_cpu_num) {
5978 			warn("/proc/interrupts: cpu%d: > %d", cpu_number, topo.max_cpu_num);
5979 			return 1;
5980 		}
5981 
5982 		irq_column_2_cpu[column] = cpu_number;
5983 		irqs_per_cpu[cpu_number] = 0;
5984 		nmi_per_cpu[cpu_number] = 0;
5985 	}
5986 
5987 	/* read /proc/interrupt count lines and sum up irqs per cpu */
5988 	while (1) {
5989 		int column;
5990 		char buf[64];
5991 		int this_row_is_nmi = 0;
5992 
5993 		retval = fscanf(fp, " %s:", buf);	/* irq# "N:" */
5994 		if (retval != 1)
5995 			break;
5996 
5997 		if (strncmp(buf, "NMI", strlen("NMI")) == 0)
5998 			this_row_is_nmi = 1;
5999 
6000 		/* read the count per cpu */
6001 		for (column = 0; column < topo.num_cpus; ++column) {
6002 
6003 			int cpu_number, irq_count;
6004 
6005 			retval = fscanf(fp, " %d", &irq_count);
6006 
6007 			if (retval != 1)
6008 				break;
6009 
6010 			cpu_number = irq_column_2_cpu[column];
6011 			irqs_per_cpu[cpu_number] += irq_count;
6012 			if (this_row_is_nmi)
6013 				nmi_per_cpu[cpu_number] += irq_count;
6014 		}
6015 		while (getc(fp) != '\n') ;	/* flush interrupt description */
6016 
6017 	}
6018 	return 0;
6019 }
6020 
6021 /*
6022  * snapshot_graphics()
6023  *
6024  * record snapshot of specified graphics sysfs knob
6025  *
6026  * return 1 if config change requires a restart, else return 0
6027  */
snapshot_graphics(int idx)6028 int snapshot_graphics(int idx)
6029 {
6030 	int retval;
6031 
6032 	rewind(gfx_info[idx].fp);
6033 
6034 	switch (idx) {
6035 	case GFX_rc6:
6036 	case SAM_mc6:
6037 		retval = fscanf(gfx_info[idx].fp, "%lld", &gfx_info[idx].val_ull);
6038 		if (retval != 1)
6039 			err(1, "rc6");
6040 		return 0;
6041 	case GFX_MHz:
6042 	case GFX_ACTMHz:
6043 	case SAM_MHz:
6044 	case SAM_ACTMHz:
6045 		retval = fscanf(gfx_info[idx].fp, "%d", &gfx_info[idx].val);
6046 		if (retval != 1)
6047 			err(1, "MHz");
6048 		return 0;
6049 	default:
6050 		return -EINVAL;
6051 	}
6052 }
6053 
6054 /*
6055  * snapshot_cpu_lpi()
6056  *
6057  * record snapshot of
6058  * /sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us
6059  */
snapshot_cpu_lpi_us(void)6060 int snapshot_cpu_lpi_us(void)
6061 {
6062 	FILE *fp;
6063 	int retval;
6064 
6065 	fp = fopen_or_die("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", "r");
6066 
6067 	retval = fscanf(fp, "%lld", &cpuidle_cur_cpu_lpi_us);
6068 	if (retval != 1) {
6069 		fprintf(stderr, "Disabling Low Power Idle CPU output\n");
6070 		BIC_NOT_PRESENT(BIC_CPU_LPI);
6071 		fclose(fp);
6072 		return -1;
6073 	}
6074 
6075 	fclose(fp);
6076 
6077 	return 0;
6078 }
6079 
6080 /*
6081  * snapshot_sys_lpi()
6082  *
6083  * record snapshot of sys_lpi_file
6084  */
snapshot_sys_lpi_us(void)6085 int snapshot_sys_lpi_us(void)
6086 {
6087 	FILE *fp;
6088 	int retval;
6089 
6090 	fp = fopen_or_die(sys_lpi_file, "r");
6091 
6092 	retval = fscanf(fp, "%lld", &cpuidle_cur_sys_lpi_us);
6093 	if (retval != 1) {
6094 		fprintf(stderr, "Disabling Low Power Idle System output\n");
6095 		BIC_NOT_PRESENT(BIC_SYS_LPI);
6096 		fclose(fp);
6097 		return -1;
6098 	}
6099 	fclose(fp);
6100 
6101 	return 0;
6102 }
6103 
6104 /*
6105  * snapshot /proc and /sys files
6106  *
6107  * return 1 if configuration restart needed, else return 0
6108  */
snapshot_proc_sysfs_files(void)6109 int snapshot_proc_sysfs_files(void)
6110 {
6111 	gettimeofday(&procsysfs_tv_begin, (struct timezone *)NULL);
6112 
6113 	if (DO_BIC(BIC_IRQ) || DO_BIC(BIC_NMI))
6114 		if (snapshot_proc_interrupts())
6115 			return 1;
6116 
6117 	if (DO_BIC(BIC_GFX_rc6))
6118 		snapshot_graphics(GFX_rc6);
6119 
6120 	if (DO_BIC(BIC_GFXMHz))
6121 		snapshot_graphics(GFX_MHz);
6122 
6123 	if (DO_BIC(BIC_GFXACTMHz))
6124 		snapshot_graphics(GFX_ACTMHz);
6125 
6126 	if (DO_BIC(BIC_SAM_mc6))
6127 		snapshot_graphics(SAM_mc6);
6128 
6129 	if (DO_BIC(BIC_SAMMHz))
6130 		snapshot_graphics(SAM_MHz);
6131 
6132 	if (DO_BIC(BIC_SAMACTMHz))
6133 		snapshot_graphics(SAM_ACTMHz);
6134 
6135 	if (DO_BIC(BIC_CPU_LPI))
6136 		snapshot_cpu_lpi_us();
6137 
6138 	if (DO_BIC(BIC_SYS_LPI))
6139 		snapshot_sys_lpi_us();
6140 
6141 	return 0;
6142 }
6143 
6144 int exit_requested;
6145 
signal_handler(int signal)6146 static void signal_handler(int signal)
6147 {
6148 	switch (signal) {
6149 	case SIGINT:
6150 		exit_requested = 1;
6151 		if (debug)
6152 			fprintf(stderr, " SIGINT\n");
6153 		break;
6154 	case SIGUSR1:
6155 		if (debug > 1)
6156 			fprintf(stderr, "SIGUSR1\n");
6157 		break;
6158 	}
6159 }
6160 
setup_signal_handler(void)6161 void setup_signal_handler(void)
6162 {
6163 	struct sigaction sa;
6164 
6165 	memset(&sa, 0, sizeof(sa));
6166 
6167 	sa.sa_handler = &signal_handler;
6168 
6169 	if (sigaction(SIGINT, &sa, NULL) < 0)
6170 		err(1, "sigaction SIGINT");
6171 	if (sigaction(SIGUSR1, &sa, NULL) < 0)
6172 		err(1, "sigaction SIGUSR1");
6173 }
6174 
do_sleep(void)6175 void do_sleep(void)
6176 {
6177 	struct timeval tout;
6178 	struct timespec rest;
6179 	fd_set readfds;
6180 	int retval;
6181 
6182 	FD_ZERO(&readfds);
6183 	FD_SET(0, &readfds);
6184 
6185 	if (ignore_stdin) {
6186 		nanosleep(&interval_ts, NULL);
6187 		return;
6188 	}
6189 
6190 	tout = interval_tv;
6191 	retval = select(1, &readfds, NULL, NULL, &tout);
6192 
6193 	if (retval == 1) {
6194 		switch (getc(stdin)) {
6195 		case 'q':
6196 			exit_requested = 1;
6197 			break;
6198 		case EOF:
6199 			/*
6200 			 * 'stdin' is a pipe closed on the other end. There
6201 			 * won't be any further input.
6202 			 */
6203 			ignore_stdin = 1;
6204 			/* Sleep the rest of the time */
6205 			rest.tv_sec = (tout.tv_sec + tout.tv_usec / 1000000);
6206 			rest.tv_nsec = (tout.tv_usec % 1000000) * 1000;
6207 			nanosleep(&rest, NULL);
6208 		}
6209 	}
6210 }
6211 
get_msr_sum(int cpu,off_t offset,unsigned long long * msr)6212 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr)
6213 {
6214 	int ret, idx;
6215 	unsigned long long msr_cur, msr_last;
6216 
6217 	assert(!no_msr);
6218 
6219 	if (!per_cpu_msr_sum)
6220 		return 1;
6221 
6222 	idx = offset_to_idx(offset);
6223 	if (idx < 0)
6224 		return idx;
6225 	/* get_msr_sum() = sum + (get_msr() - last) */
6226 	ret = get_msr(cpu, offset, &msr_cur);
6227 	if (ret)
6228 		return ret;
6229 	msr_last = per_cpu_msr_sum[cpu].entries[idx].last;
6230 	DELTA_WRAP32(msr_cur, msr_last);
6231 	*msr = msr_last + per_cpu_msr_sum[cpu].entries[idx].sum;
6232 
6233 	return 0;
6234 }
6235 
6236 timer_t timerid;
6237 
6238 /* Timer callback, update the sum of MSRs periodically. */
update_msr_sum(struct thread_data * t,struct core_data * c,struct pkg_data * p)6239 static int update_msr_sum(struct thread_data *t, struct core_data *c, struct pkg_data *p)
6240 {
6241 	int i, ret;
6242 	int cpu = t->cpu_id;
6243 
6244 	UNUSED(c);
6245 	UNUSED(p);
6246 
6247 	assert(!no_msr);
6248 
6249 	for (i = IDX_PKG_ENERGY; i < IDX_COUNT; i++) {
6250 		unsigned long long msr_cur, msr_last;
6251 		off_t offset;
6252 
6253 		if (!idx_valid(i))
6254 			continue;
6255 		offset = idx_to_offset(i);
6256 		if (offset < 0)
6257 			continue;
6258 		ret = get_msr(cpu, offset, &msr_cur);
6259 		if (ret) {
6260 			fprintf(outf, "Can not update msr(0x%llx)\n", (unsigned long long)offset);
6261 			continue;
6262 		}
6263 
6264 		msr_last = per_cpu_msr_sum[cpu].entries[i].last;
6265 		per_cpu_msr_sum[cpu].entries[i].last = msr_cur & 0xffffffff;
6266 
6267 		DELTA_WRAP32(msr_cur, msr_last);
6268 		per_cpu_msr_sum[cpu].entries[i].sum += msr_last;
6269 	}
6270 	return 0;
6271 }
6272 
msr_record_handler(union sigval v)6273 static void msr_record_handler(union sigval v)
6274 {
6275 	UNUSED(v);
6276 
6277 	for_all_cpus(update_msr_sum, EVEN_COUNTERS);
6278 }
6279 
msr_sum_record(void)6280 void msr_sum_record(void)
6281 {
6282 	struct itimerspec its;
6283 	struct sigevent sev;
6284 
6285 	per_cpu_msr_sum = calloc(topo.max_cpu_num + 1, sizeof(struct msr_sum_array));
6286 	if (!per_cpu_msr_sum) {
6287 		fprintf(outf, "Can not allocate memory for long time MSR.\n");
6288 		return;
6289 	}
6290 	/*
6291 	 * Signal handler might be restricted, so use thread notifier instead.
6292 	 */
6293 	memset(&sev, 0, sizeof(struct sigevent));
6294 	sev.sigev_notify = SIGEV_THREAD;
6295 	sev.sigev_notify_function = msr_record_handler;
6296 
6297 	sev.sigev_value.sival_ptr = &timerid;
6298 	if (timer_create(CLOCK_REALTIME, &sev, &timerid) == -1) {
6299 		fprintf(outf, "Can not create timer.\n");
6300 		goto release_msr;
6301 	}
6302 
6303 	its.it_value.tv_sec = 0;
6304 	its.it_value.tv_nsec = 1;
6305 	/*
6306 	 * A wraparound time has been calculated early.
6307 	 * Some sources state that the peak power for a
6308 	 * microprocessor is usually 1.5 times the TDP rating,
6309 	 * use 2 * TDP for safety.
6310 	 */
6311 	its.it_interval.tv_sec = rapl_joule_counter_range / 2;
6312 	its.it_interval.tv_nsec = 0;
6313 
6314 	if (timer_settime(timerid, 0, &its, NULL) == -1) {
6315 		fprintf(outf, "Can not set timer.\n");
6316 		goto release_timer;
6317 	}
6318 	return;
6319 
6320 release_timer:
6321 	timer_delete(timerid);
6322 release_msr:
6323 	free(per_cpu_msr_sum);
6324 }
6325 
6326 /*
6327  * set_my_sched_priority(pri)
6328  * return previous priority on success
6329  * return value < -20 on failure
6330  */
set_my_sched_priority(int priority)6331 int set_my_sched_priority(int priority)
6332 {
6333 	int retval;
6334 	int original_priority;
6335 
6336 	errno = 0;
6337 	original_priority = getpriority(PRIO_PROCESS, 0);
6338 	if (errno && (original_priority == -1))
6339 		return -21;
6340 
6341 	retval = setpriority(PRIO_PROCESS, 0, priority);
6342 	if (retval)
6343 		return -21;
6344 
6345 	errno = 0;
6346 	retval = getpriority(PRIO_PROCESS, 0);
6347 	if (retval != priority)
6348 		return -21;
6349 
6350 	return original_priority;
6351 }
6352 
turbostat_loop()6353 void turbostat_loop()
6354 {
6355 	int retval;
6356 	int restarted = 0;
6357 	unsigned int done_iters = 0;
6358 
6359 	setup_signal_handler();
6360 
6361 	/*
6362 	 * elevate own priority for interval mode
6363 	 *
6364 	 * ignore on error - we probably don't have permission to set it, but
6365 	 * it's not a big deal
6366 	 */
6367 	set_my_sched_priority(-20);
6368 
6369 restart:
6370 	restarted++;
6371 
6372 	snapshot_proc_sysfs_files();
6373 	retval = for_all_cpus(get_counters, EVEN_COUNTERS);
6374 	first_counter_read = 0;
6375 	if (retval < -1) {
6376 		exit(retval);
6377 	} else if (retval == -1) {
6378 		if (restarted > 10) {
6379 			exit(retval);
6380 		}
6381 		re_initialize();
6382 		goto restart;
6383 	}
6384 	restarted = 0;
6385 	done_iters = 0;
6386 	gettimeofday(&tv_even, (struct timezone *)NULL);
6387 
6388 	while (1) {
6389 		if (for_all_proc_cpus(cpu_is_not_present)) {
6390 			re_initialize();
6391 			goto restart;
6392 		}
6393 		if (update_effective_str(false)) {
6394 			re_initialize();
6395 			goto restart;
6396 		}
6397 		do_sleep();
6398 		if (snapshot_proc_sysfs_files())
6399 			goto restart;
6400 		retval = for_all_cpus(get_counters, ODD_COUNTERS);
6401 		if (retval < -1) {
6402 			exit(retval);
6403 		} else if (retval == -1) {
6404 			re_initialize();
6405 			goto restart;
6406 		}
6407 		gettimeofday(&tv_odd, (struct timezone *)NULL);
6408 		timersub(&tv_odd, &tv_even, &tv_delta);
6409 		if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) {
6410 			re_initialize();
6411 			goto restart;
6412 		}
6413 		delta_platform(&platform_counters_odd, &platform_counters_even);
6414 		compute_average(EVEN_COUNTERS);
6415 		format_all_counters(EVEN_COUNTERS);
6416 		flush_output_stdout();
6417 		if (exit_requested)
6418 			break;
6419 		if (num_iterations && ++done_iters >= num_iterations)
6420 			break;
6421 		do_sleep();
6422 		if (snapshot_proc_sysfs_files())
6423 			goto restart;
6424 		retval = for_all_cpus(get_counters, EVEN_COUNTERS);
6425 		if (retval < -1) {
6426 			exit(retval);
6427 		} else if (retval == -1) {
6428 			re_initialize();
6429 			goto restart;
6430 		}
6431 		gettimeofday(&tv_even, (struct timezone *)NULL);
6432 		timersub(&tv_even, &tv_odd, &tv_delta);
6433 		if (for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) {
6434 			re_initialize();
6435 			goto restart;
6436 		}
6437 		delta_platform(&platform_counters_even, &platform_counters_odd);
6438 		compute_average(ODD_COUNTERS);
6439 		format_all_counters(ODD_COUNTERS);
6440 		flush_output_stdout();
6441 		if (exit_requested)
6442 			break;
6443 		if (num_iterations && ++done_iters >= num_iterations)
6444 			break;
6445 	}
6446 }
6447 
check_dev_msr()6448 void check_dev_msr()
6449 {
6450 	struct stat sb;
6451 	char pathname[32];
6452 
6453 	if (no_msr)
6454 		return;
6455 
6456 	sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
6457 	if (stat(pathname, &sb))
6458 		if (system("/sbin/modprobe msr > /dev/null 2>&1"))
6459 			no_msr = 1;
6460 }
6461 
6462 /*
6463  * check for CAP_SYS_RAWIO
6464  * return 0 on success
6465  * return 1 on fail
6466  */
check_for_cap_sys_rawio(void)6467 int check_for_cap_sys_rawio(void)
6468 {
6469 	cap_t caps;
6470 	cap_flag_value_t cap_flag_value;
6471 	int ret = 0;
6472 
6473 	caps = cap_get_proc();
6474 	if (caps == NULL)
6475 		return 1;
6476 
6477 	if (cap_get_flag(caps, CAP_SYS_RAWIO, CAP_EFFECTIVE, &cap_flag_value)) {
6478 		ret = 1;
6479 		goto free_and_exit;
6480 	}
6481 
6482 	if (cap_flag_value != CAP_SET) {
6483 		ret = 1;
6484 		goto free_and_exit;
6485 	}
6486 
6487 free_and_exit:
6488 	if (cap_free(caps) == -1)
6489 		err(-6, "cap_free\n");
6490 
6491 	return ret;
6492 }
6493 
check_msr_permission(void)6494 void check_msr_permission(void)
6495 {
6496 	int failed = 0;
6497 	char pathname[32];
6498 
6499 	if (no_msr)
6500 		return;
6501 
6502 	/* check for CAP_SYS_RAWIO */
6503 	failed += check_for_cap_sys_rawio();
6504 
6505 	/* test file permissions */
6506 	sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
6507 	if (euidaccess(pathname, R_OK)) {
6508 		failed++;
6509 	}
6510 
6511 	if (failed) {
6512 		warnx("Failed to access %s. Some of the counters may not be available\n"
6513 		      "\tRun as root to enable them or use %s to disable the access explicitly", pathname, "--no-msr");
6514 		no_msr = 1;
6515 	}
6516 }
6517 
probe_bclk(void)6518 void probe_bclk(void)
6519 {
6520 	unsigned long long msr;
6521 	unsigned int base_ratio;
6522 
6523 	if (!platform->has_nhm_msrs || no_msr)
6524 		return;
6525 
6526 	if (platform->bclk_freq == BCLK_100MHZ)
6527 		bclk = 100.00;
6528 	else if (platform->bclk_freq == BCLK_133MHZ)
6529 		bclk = 133.33;
6530 	else if (platform->bclk_freq == BCLK_SLV)
6531 		bclk = slm_bclk();
6532 	else
6533 		return;
6534 
6535 	get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
6536 	base_ratio = (msr >> 8) & 0xFF;
6537 
6538 	base_hz = base_ratio * bclk * 1000000;
6539 	has_base_hz = 1;
6540 
6541 	if (platform->enable_tsc_tweak)
6542 		tsc_tweak = base_hz / tsc_hz;
6543 }
6544 
remove_underbar(char * s)6545 static void remove_underbar(char *s)
6546 {
6547 	char *to = s;
6548 
6549 	while (*s) {
6550 		if (*s != '_')
6551 			*to++ = *s;
6552 		s++;
6553 	}
6554 
6555 	*to = 0;
6556 }
6557 
dump_turbo_ratio_info(void)6558 static void dump_turbo_ratio_info(void)
6559 {
6560 	if (!has_turbo)
6561 		return;
6562 
6563 	if (!platform->has_nhm_msrs || no_msr)
6564 		return;
6565 
6566 	if (platform->trl_msrs & TRL_LIMIT2)
6567 		dump_turbo_ratio_limit2();
6568 
6569 	if (platform->trl_msrs & TRL_LIMIT1)
6570 		dump_turbo_ratio_limit1();
6571 
6572 	if (platform->trl_msrs & TRL_BASE) {
6573 		dump_turbo_ratio_limits(MSR_TURBO_RATIO_LIMIT);
6574 
6575 		if (is_hybrid)
6576 			dump_turbo_ratio_limits(MSR_SECONDARY_TURBO_RATIO_LIMIT);
6577 	}
6578 
6579 	if (platform->trl_msrs & TRL_ATOM)
6580 		dump_atom_turbo_ratio_limits();
6581 
6582 	if (platform->trl_msrs & TRL_KNL)
6583 		dump_knl_turbo_ratio_limits();
6584 
6585 	if (platform->has_config_tdp)
6586 		dump_config_tdp();
6587 }
6588 
read_sysfs_int(char * path)6589 static int read_sysfs_int(char *path)
6590 {
6591 	FILE *input;
6592 	int retval = -1;
6593 
6594 	input = fopen(path, "r");
6595 	if (input == NULL) {
6596 		if (debug)
6597 			fprintf(outf, "NSFOD %s\n", path);
6598 		return (-1);
6599 	}
6600 	if (fscanf(input, "%d", &retval) != 1)
6601 		err(1, "%s: failed to read int from file", path);
6602 	fclose(input);
6603 
6604 	return (retval);
6605 }
6606 
dump_sysfs_file(char * path)6607 static void dump_sysfs_file(char *path)
6608 {
6609 	FILE *input;
6610 	char cpuidle_buf[64];
6611 
6612 	input = fopen(path, "r");
6613 	if (input == NULL) {
6614 		if (debug)
6615 			fprintf(outf, "NSFOD %s\n", path);
6616 		return;
6617 	}
6618 	if (!fgets(cpuidle_buf, sizeof(cpuidle_buf), input))
6619 		err(1, "%s: failed to read file", path);
6620 	fclose(input);
6621 
6622 	fprintf(outf, "%s: %s", strrchr(path, '/') + 1, cpuidle_buf);
6623 }
6624 
probe_intel_uncore_frequency_legacy(void)6625 static void probe_intel_uncore_frequency_legacy(void)
6626 {
6627 	int i, j;
6628 	char path[256];
6629 
6630 	for (i = 0; i < topo.num_packages; ++i) {
6631 		for (j = 0; j <= topo.max_die_id; ++j) {
6632 			int k, l;
6633 			char path_base[128];
6634 
6635 			sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/package_%02d_die_%02d", i,
6636 				j);
6637 
6638 			if (access(path_base, R_OK))
6639 				continue;
6640 
6641 			BIC_PRESENT(BIC_UNCORE_MHZ);
6642 
6643 			if (quiet)
6644 				return;
6645 
6646 			sprintf(path, "%s/min_freq_khz", path_base);
6647 			k = read_sysfs_int(path);
6648 			sprintf(path, "%s/max_freq_khz", path_base);
6649 			l = read_sysfs_int(path);
6650 			fprintf(outf, "Uncore Frequency package%d die%d: %d - %d MHz ", i, j, k / 1000, l / 1000);
6651 
6652 			sprintf(path, "%s/initial_min_freq_khz", path_base);
6653 			k = read_sysfs_int(path);
6654 			sprintf(path, "%s/initial_max_freq_khz", path_base);
6655 			l = read_sysfs_int(path);
6656 			fprintf(outf, "(%d - %d MHz)", k / 1000, l / 1000);
6657 
6658 			sprintf(path, "%s/current_freq_khz", path_base);
6659 			k = read_sysfs_int(path);
6660 			fprintf(outf, " %d MHz\n", k / 1000);
6661 		}
6662 	}
6663 }
6664 
probe_intel_uncore_frequency_cluster(void)6665 static void probe_intel_uncore_frequency_cluster(void)
6666 {
6667 	int i, uncore_max_id;
6668 	char path[256];
6669 	char path_base[128];
6670 
6671 	if (access("/sys/devices/system/cpu/intel_uncore_frequency/uncore00/current_freq_khz", R_OK))
6672 		return;
6673 
6674 	for (uncore_max_id = 0;; ++uncore_max_id) {
6675 
6676 		sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/uncore%02d", uncore_max_id);
6677 
6678 		/* uncore## start at 00 and skips no numbers, so stop upon first missing */
6679 		if (access(path_base, R_OK)) {
6680 			uncore_max_id -= 1;
6681 			break;
6682 		}
6683 	}
6684 	for (i = uncore_max_id; i >= 0; --i) {
6685 		int k, l;
6686 		int package_id, domain_id, cluster_id;
6687 		char name_buf[16];
6688 
6689 		sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/uncore%02d", i);
6690 
6691 		if (access(path_base, R_OK))
6692 			err(1, "%s: %s\n", __func__, path_base);
6693 
6694 		sprintf(path, "%s/package_id", path_base);
6695 		package_id = read_sysfs_int(path);
6696 
6697 		sprintf(path, "%s/domain_id", path_base);
6698 		domain_id = read_sysfs_int(path);
6699 
6700 		sprintf(path, "%s/fabric_cluster_id", path_base);
6701 		cluster_id = read_sysfs_int(path);
6702 
6703 		sprintf(path, "%s/current_freq_khz", path_base);
6704 		sprintf(name_buf, "UMHz%d.%d", domain_id, cluster_id);
6705 
6706 		add_counter(0, path, name_buf, 0, SCOPE_PACKAGE, COUNTER_K2M, FORMAT_AVERAGE, 0, package_id);
6707 
6708 		if (quiet)
6709 			continue;
6710 
6711 		sprintf(path, "%s/min_freq_khz", path_base);
6712 		k = read_sysfs_int(path);
6713 		sprintf(path, "%s/max_freq_khz", path_base);
6714 		l = read_sysfs_int(path);
6715 		fprintf(outf, "Uncore Frequency package%d domain%d cluster%d: %d - %d MHz ", package_id, domain_id,
6716 			cluster_id, k / 1000, l / 1000);
6717 
6718 		sprintf(path, "%s/initial_min_freq_khz", path_base);
6719 		k = read_sysfs_int(path);
6720 		sprintf(path, "%s/initial_max_freq_khz", path_base);
6721 		l = read_sysfs_int(path);
6722 		fprintf(outf, "(%d - %d MHz)", k / 1000, l / 1000);
6723 
6724 		sprintf(path, "%s/current_freq_khz", path_base);
6725 		k = read_sysfs_int(path);
6726 		fprintf(outf, " %d MHz\n", k / 1000);
6727 	}
6728 }
6729 
probe_intel_uncore_frequency(void)6730 static void probe_intel_uncore_frequency(void)
6731 {
6732 	if (!genuine_intel)
6733 		return;
6734 
6735 	if (access("/sys/devices/system/cpu/intel_uncore_frequency/uncore00", R_OK) == 0)
6736 		probe_intel_uncore_frequency_cluster();
6737 	else
6738 		probe_intel_uncore_frequency_legacy();
6739 }
6740 
set_graphics_fp(char * path,int idx)6741 static void set_graphics_fp(char *path, int idx)
6742 {
6743 	if (!access(path, R_OK))
6744 		gfx_info[idx].fp = fopen_or_die(path, "r");
6745 }
6746 
6747 /* Enlarge this if there are /sys/class/drm/card2 ... */
6748 #define GFX_MAX_CARDS	2
6749 
probe_graphics(void)6750 static void probe_graphics(void)
6751 {
6752 	char path[PATH_MAX];
6753 	int i;
6754 
6755 	/* Xe graphics sysfs knobs */
6756 	if (!access("/sys/class/drm/card0/device/tile0/gt0/gtidle/idle_residency_ms", R_OK)) {
6757 		FILE *fp;
6758 		char buf[8];
6759 		bool gt0_is_gt;
6760 
6761 		fp = fopen("/sys/class/drm/card0/device/tile0/gt0/gtidle/name", "r");
6762 		if (!fp)
6763 			goto next;
6764 
6765 		if (!fread(buf, sizeof(char), 7, fp)) {
6766 			fclose(fp);
6767 			goto next;
6768 		}
6769 		fclose(fp);
6770 
6771 		if (!strncmp(buf, "gt0-rc", strlen("gt0-rc")))
6772 			gt0_is_gt = true;
6773 		else if (!strncmp(buf, "gt0-mc", strlen("gt0-mc")))
6774 			gt0_is_gt = false;
6775 		else
6776 			goto next;
6777 
6778 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/gtidle/idle_residency_ms", gt0_is_gt ? GFX_rc6 : SAM_mc6);
6779 
6780 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/freq0/cur_freq", gt0_is_gt ? GFX_MHz : SAM_MHz);
6781 
6782 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/freq0/act_freq", gt0_is_gt ? GFX_ACTMHz : SAM_ACTMHz);
6783 
6784 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/gtidle/idle_residency_ms", gt0_is_gt ? SAM_mc6 : GFX_rc6);
6785 
6786 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/freq0/cur_freq", gt0_is_gt ? SAM_MHz : GFX_MHz);
6787 
6788 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/freq0/act_freq", gt0_is_gt ? SAM_ACTMHz : GFX_ACTMHz);
6789 
6790 		goto end;
6791 	}
6792 
6793 next:
6794 	/* New i915 graphics sysfs knobs */
6795 	for (i = 0; i < GFX_MAX_CARDS; i++) {
6796 		snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rc6_residency_ms", i);
6797 		if (!access(path, R_OK))
6798 			break;
6799 	}
6800 
6801 	if (i == GFX_MAX_CARDS)
6802 		goto legacy_i915;
6803 
6804 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rc6_residency_ms", i);
6805 	set_graphics_fp(path, GFX_rc6);
6806 
6807 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rps_cur_freq_mhz", i);
6808 	set_graphics_fp(path, GFX_MHz);
6809 
6810 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rps_act_freq_mhz", i);
6811 	set_graphics_fp(path, GFX_ACTMHz);
6812 
6813 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rc6_residency_ms", i);
6814 	set_graphics_fp(path, SAM_mc6);
6815 
6816 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rps_cur_freq_mhz", i);
6817 	set_graphics_fp(path, SAM_MHz);
6818 
6819 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rps_act_freq_mhz", i);
6820 	set_graphics_fp(path, SAM_ACTMHz);
6821 
6822 	goto end;
6823 
6824 legacy_i915:
6825 	/* Fall back to traditional i915 graphics sysfs knobs */
6826 	set_graphics_fp("/sys/class/drm/card0/power/rc6_residency_ms", GFX_rc6);
6827 
6828 	set_graphics_fp("/sys/class/drm/card0/gt_cur_freq_mhz", GFX_MHz);
6829 	if (!gfx_info[GFX_MHz].fp)
6830 		set_graphics_fp("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", GFX_MHz);
6831 
6832 	set_graphics_fp("/sys/class/drm/card0/gt_act_freq_mhz", GFX_ACTMHz);
6833 	if (!gfx_info[GFX_ACTMHz].fp)
6834 		set_graphics_fp("/sys/class/graphics/fb0/device/drm/card0/gt_act_freq_mhz", GFX_ACTMHz);
6835 
6836 end:
6837 	if (gfx_info[GFX_rc6].fp)
6838 		BIC_PRESENT(BIC_GFX_rc6);
6839 	if (gfx_info[GFX_MHz].fp)
6840 		BIC_PRESENT(BIC_GFXMHz);
6841 	if (gfx_info[GFX_ACTMHz].fp)
6842 		BIC_PRESENT(BIC_GFXACTMHz);
6843 	if (gfx_info[SAM_mc6].fp)
6844 		BIC_PRESENT(BIC_SAM_mc6);
6845 	if (gfx_info[SAM_MHz].fp)
6846 		BIC_PRESENT(BIC_SAMMHz);
6847 	if (gfx_info[SAM_ACTMHz].fp)
6848 		BIC_PRESENT(BIC_SAMACTMHz);
6849 }
6850 
dump_sysfs_cstate_config(void)6851 static void dump_sysfs_cstate_config(void)
6852 {
6853 	char path[64];
6854 	char name_buf[16];
6855 	char desc[64];
6856 	FILE *input;
6857 	int state;
6858 	char *sp;
6859 
6860 	if (access("/sys/devices/system/cpu/cpuidle", R_OK)) {
6861 		fprintf(outf, "cpuidle not loaded\n");
6862 		return;
6863 	}
6864 
6865 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_driver");
6866 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor");
6867 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor_ro");
6868 
6869 	for (state = 0; state < 10; ++state) {
6870 
6871 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", base_cpu, state);
6872 		input = fopen(path, "r");
6873 		if (input == NULL)
6874 			continue;
6875 		if (!fgets(name_buf, sizeof(name_buf), input))
6876 			err(1, "%s: failed to read file", path);
6877 
6878 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
6879 		sp = strchr(name_buf, '-');
6880 		if (!sp)
6881 			sp = strchrnul(name_buf, '\n');
6882 		*sp = '\0';
6883 		fclose(input);
6884 
6885 		remove_underbar(name_buf);
6886 
6887 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/desc", base_cpu, state);
6888 		input = fopen(path, "r");
6889 		if (input == NULL)
6890 			continue;
6891 		if (!fgets(desc, sizeof(desc), input))
6892 			err(1, "%s: failed to read file", path);
6893 
6894 		fprintf(outf, "cpu%d: %s: %s", base_cpu, name_buf, desc);
6895 		fclose(input);
6896 	}
6897 }
6898 
dump_sysfs_pstate_config(void)6899 static void dump_sysfs_pstate_config(void)
6900 {
6901 	char path[64];
6902 	char driver_buf[64];
6903 	char governor_buf[64];
6904 	FILE *input;
6905 	int turbo;
6906 
6907 	sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_driver", base_cpu);
6908 	input = fopen(path, "r");
6909 	if (input == NULL) {
6910 		fprintf(outf, "NSFOD %s\n", path);
6911 		return;
6912 	}
6913 	if (!fgets(driver_buf, sizeof(driver_buf), input))
6914 		err(1, "%s: failed to read file", path);
6915 	fclose(input);
6916 
6917 	sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_governor", base_cpu);
6918 	input = fopen(path, "r");
6919 	if (input == NULL) {
6920 		fprintf(outf, "NSFOD %s\n", path);
6921 		return;
6922 	}
6923 	if (!fgets(governor_buf, sizeof(governor_buf), input))
6924 		err(1, "%s: failed to read file", path);
6925 	fclose(input);
6926 
6927 	fprintf(outf, "cpu%d: cpufreq driver: %s", base_cpu, driver_buf);
6928 	fprintf(outf, "cpu%d: cpufreq governor: %s", base_cpu, governor_buf);
6929 
6930 	sprintf(path, "/sys/devices/system/cpu/cpufreq/boost");
6931 	input = fopen(path, "r");
6932 	if (input != NULL) {
6933 		if (fscanf(input, "%d", &turbo) != 1)
6934 			err(1, "%s: failed to parse number from file", path);
6935 		fprintf(outf, "cpufreq boost: %d\n", turbo);
6936 		fclose(input);
6937 	}
6938 
6939 	sprintf(path, "/sys/devices/system/cpu/intel_pstate/no_turbo");
6940 	input = fopen(path, "r");
6941 	if (input != NULL) {
6942 		if (fscanf(input, "%d", &turbo) != 1)
6943 			err(1, "%s: failed to parse number from file", path);
6944 		fprintf(outf, "cpufreq intel_pstate no_turbo: %d\n", turbo);
6945 		fclose(input);
6946 	}
6947 }
6948 
6949 /*
6950  * print_epb()
6951  * Decode the ENERGY_PERF_BIAS MSR
6952  */
print_epb(struct thread_data * t,struct core_data * c,struct pkg_data * p)6953 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
6954 {
6955 	char *epb_string;
6956 	int cpu, epb;
6957 
6958 	UNUSED(c);
6959 	UNUSED(p);
6960 
6961 	if (!has_epb)
6962 		return 0;
6963 
6964 	cpu = t->cpu_id;
6965 
6966 	/* EPB is per-package */
6967 	if (!is_cpu_first_thread_in_package(t, c, p))
6968 		return 0;
6969 
6970 	if (cpu_migrate(cpu)) {
6971 		fprintf(outf, "print_epb: Could not migrate to CPU %d\n", cpu);
6972 		return -1;
6973 	}
6974 
6975 	epb = get_epb(cpu);
6976 	if (epb < 0)
6977 		return 0;
6978 
6979 	switch (epb) {
6980 	case ENERGY_PERF_BIAS_PERFORMANCE:
6981 		epb_string = "performance";
6982 		break;
6983 	case ENERGY_PERF_BIAS_NORMAL:
6984 		epb_string = "balanced";
6985 		break;
6986 	case ENERGY_PERF_BIAS_POWERSAVE:
6987 		epb_string = "powersave";
6988 		break;
6989 	default:
6990 		epb_string = "custom";
6991 		break;
6992 	}
6993 	fprintf(outf, "cpu%d: EPB: %d (%s)\n", cpu, epb, epb_string);
6994 
6995 	return 0;
6996 }
6997 
6998 /*
6999  * print_hwp()
7000  * Decode the MSR_HWP_CAPABILITIES
7001  */
print_hwp(struct thread_data * t,struct core_data * c,struct pkg_data * p)7002 int print_hwp(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7003 {
7004 	unsigned long long msr;
7005 	int cpu;
7006 
7007 	UNUSED(c);
7008 	UNUSED(p);
7009 
7010 	if (no_msr)
7011 		return 0;
7012 
7013 	if (!has_hwp)
7014 		return 0;
7015 
7016 	cpu = t->cpu_id;
7017 
7018 	/* MSR_HWP_CAPABILITIES is per-package */
7019 	if (!is_cpu_first_thread_in_package(t, c, p))
7020 		return 0;
7021 
7022 	if (cpu_migrate(cpu)) {
7023 		fprintf(outf, "print_hwp: Could not migrate to CPU %d\n", cpu);
7024 		return -1;
7025 	}
7026 
7027 	if (get_msr(cpu, MSR_PM_ENABLE, &msr))
7028 		return 0;
7029 
7030 	fprintf(outf, "cpu%d: MSR_PM_ENABLE: 0x%08llx (%sHWP)\n", cpu, msr, (msr & (1 << 0)) ? "" : "No-");
7031 
7032 	/* MSR_PM_ENABLE[1] == 1 if HWP is enabled and MSRs visible */
7033 	if ((msr & (1 << 0)) == 0)
7034 		return 0;
7035 
7036 	if (get_msr(cpu, MSR_HWP_CAPABILITIES, &msr))
7037 		return 0;
7038 
7039 	fprintf(outf, "cpu%d: MSR_HWP_CAPABILITIES: 0x%08llx "
7040 		"(high %d guar %d eff %d low %d)\n",
7041 		cpu, msr,
7042 		(unsigned int)HWP_HIGHEST_PERF(msr),
7043 		(unsigned int)HWP_GUARANTEED_PERF(msr),
7044 		(unsigned int)HWP_MOSTEFFICIENT_PERF(msr), (unsigned int)HWP_LOWEST_PERF(msr));
7045 
7046 	if (get_msr(cpu, MSR_HWP_REQUEST, &msr))
7047 		return 0;
7048 
7049 	fprintf(outf, "cpu%d: MSR_HWP_REQUEST: 0x%08llx "
7050 		"(min %d max %d des %d epp 0x%x window 0x%x pkg 0x%x)\n",
7051 		cpu, msr,
7052 		(unsigned int)(((msr) >> 0) & 0xff),
7053 		(unsigned int)(((msr) >> 8) & 0xff),
7054 		(unsigned int)(((msr) >> 16) & 0xff),
7055 		(unsigned int)(((msr) >> 24) & 0xff),
7056 		(unsigned int)(((msr) >> 32) & 0xff3), (unsigned int)(((msr) >> 42) & 0x1));
7057 
7058 	if (has_hwp_pkg) {
7059 		if (get_msr(cpu, MSR_HWP_REQUEST_PKG, &msr))
7060 			return 0;
7061 
7062 		fprintf(outf, "cpu%d: MSR_HWP_REQUEST_PKG: 0x%08llx "
7063 			"(min %d max %d des %d epp 0x%x window 0x%x)\n",
7064 			cpu, msr,
7065 			(unsigned int)(((msr) >> 0) & 0xff),
7066 			(unsigned int)(((msr) >> 8) & 0xff),
7067 			(unsigned int)(((msr) >> 16) & 0xff),
7068 			(unsigned int)(((msr) >> 24) & 0xff), (unsigned int)(((msr) >> 32) & 0xff3));
7069 	}
7070 	if (has_hwp_notify) {
7071 		if (get_msr(cpu, MSR_HWP_INTERRUPT, &msr))
7072 			return 0;
7073 
7074 		fprintf(outf, "cpu%d: MSR_HWP_INTERRUPT: 0x%08llx "
7075 			"(%s_Guaranteed_Perf_Change, %s_Excursion_Min)\n",
7076 			cpu, msr, ((msr) & 0x1) ? "EN" : "Dis", ((msr) & 0x2) ? "EN" : "Dis");
7077 	}
7078 	if (get_msr(cpu, MSR_HWP_STATUS, &msr))
7079 		return 0;
7080 
7081 	fprintf(outf, "cpu%d: MSR_HWP_STATUS: 0x%08llx "
7082 		"(%sGuaranteed_Perf_Change, %sExcursion_Min)\n",
7083 		cpu, msr, ((msr) & 0x1) ? "" : "No-", ((msr) & 0x4) ? "" : "No-");
7084 
7085 	return 0;
7086 }
7087 
7088 /*
7089  * print_perf_limit()
7090  */
print_perf_limit(struct thread_data * t,struct core_data * c,struct pkg_data * p)7091 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7092 {
7093 	unsigned long long msr;
7094 	int cpu;
7095 
7096 	UNUSED(c);
7097 	UNUSED(p);
7098 
7099 	if (no_msr)
7100 		return 0;
7101 
7102 	cpu = t->cpu_id;
7103 
7104 	/* per-package */
7105 	if (!is_cpu_first_thread_in_package(t, c, p))
7106 		return 0;
7107 
7108 	if (cpu_migrate(cpu)) {
7109 		fprintf(outf, "print_perf_limit: Could not migrate to CPU %d\n", cpu);
7110 		return -1;
7111 	}
7112 
7113 	if (platform->plr_msrs & PLR_CORE) {
7114 		get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
7115 		fprintf(outf, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
7116 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
7117 			(msr & 1 << 15) ? "bit15, " : "",
7118 			(msr & 1 << 14) ? "bit14, " : "",
7119 			(msr & 1 << 13) ? "Transitions, " : "",
7120 			(msr & 1 << 12) ? "MultiCoreTurbo, " : "",
7121 			(msr & 1 << 11) ? "PkgPwrL2, " : "",
7122 			(msr & 1 << 10) ? "PkgPwrL1, " : "",
7123 			(msr & 1 << 9) ? "CorePwr, " : "",
7124 			(msr & 1 << 8) ? "Amps, " : "",
7125 			(msr & 1 << 6) ? "VR-Therm, " : "",
7126 			(msr & 1 << 5) ? "Auto-HWP, " : "",
7127 			(msr & 1 << 4) ? "Graphics, " : "",
7128 			(msr & 1 << 2) ? "bit2, " : "",
7129 			(msr & 1 << 1) ? "ThermStatus, " : "", (msr & 1 << 0) ? "PROCHOT, " : "");
7130 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
7131 			(msr & 1 << 31) ? "bit31, " : "",
7132 			(msr & 1 << 30) ? "bit30, " : "",
7133 			(msr & 1 << 29) ? "Transitions, " : "",
7134 			(msr & 1 << 28) ? "MultiCoreTurbo, " : "",
7135 			(msr & 1 << 27) ? "PkgPwrL2, " : "",
7136 			(msr & 1 << 26) ? "PkgPwrL1, " : "",
7137 			(msr & 1 << 25) ? "CorePwr, " : "",
7138 			(msr & 1 << 24) ? "Amps, " : "",
7139 			(msr & 1 << 22) ? "VR-Therm, " : "",
7140 			(msr & 1 << 21) ? "Auto-HWP, " : "",
7141 			(msr & 1 << 20) ? "Graphics, " : "",
7142 			(msr & 1 << 18) ? "bit18, " : "",
7143 			(msr & 1 << 17) ? "ThermStatus, " : "", (msr & 1 << 16) ? "PROCHOT, " : "");
7144 
7145 	}
7146 	if (platform->plr_msrs & PLR_GFX) {
7147 		get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
7148 		fprintf(outf, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
7149 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s)",
7150 			(msr & 1 << 0) ? "PROCHOT, " : "",
7151 			(msr & 1 << 1) ? "ThermStatus, " : "",
7152 			(msr & 1 << 4) ? "Graphics, " : "",
7153 			(msr & 1 << 6) ? "VR-Therm, " : "",
7154 			(msr & 1 << 8) ? "Amps, " : "",
7155 			(msr & 1 << 9) ? "GFXPwr, " : "",
7156 			(msr & 1 << 10) ? "PkgPwrL1, " : "", (msr & 1 << 11) ? "PkgPwrL2, " : "");
7157 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s)\n",
7158 			(msr & 1 << 16) ? "PROCHOT, " : "",
7159 			(msr & 1 << 17) ? "ThermStatus, " : "",
7160 			(msr & 1 << 20) ? "Graphics, " : "",
7161 			(msr & 1 << 22) ? "VR-Therm, " : "",
7162 			(msr & 1 << 24) ? "Amps, " : "",
7163 			(msr & 1 << 25) ? "GFXPwr, " : "",
7164 			(msr & 1 << 26) ? "PkgPwrL1, " : "", (msr & 1 << 27) ? "PkgPwrL2, " : "");
7165 	}
7166 	if (platform->plr_msrs & PLR_RING) {
7167 		get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
7168 		fprintf(outf, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
7169 		fprintf(outf, " (Active: %s%s%s%s%s%s)",
7170 			(msr & 1 << 0) ? "PROCHOT, " : "",
7171 			(msr & 1 << 1) ? "ThermStatus, " : "",
7172 			(msr & 1 << 6) ? "VR-Therm, " : "",
7173 			(msr & 1 << 8) ? "Amps, " : "",
7174 			(msr & 1 << 10) ? "PkgPwrL1, " : "", (msr & 1 << 11) ? "PkgPwrL2, " : "");
7175 		fprintf(outf, " (Logged: %s%s%s%s%s%s)\n",
7176 			(msr & 1 << 16) ? "PROCHOT, " : "",
7177 			(msr & 1 << 17) ? "ThermStatus, " : "",
7178 			(msr & 1 << 22) ? "VR-Therm, " : "",
7179 			(msr & 1 << 24) ? "Amps, " : "",
7180 			(msr & 1 << 26) ? "PkgPwrL1, " : "", (msr & 1 << 27) ? "PkgPwrL2, " : "");
7181 	}
7182 	return 0;
7183 }
7184 
7185 #define	RAPL_POWER_GRANULARITY	0x7FFF	/* 15 bit power granularity */
7186 #define	RAPL_TIME_GRANULARITY	0x3F	/* 6 bit time granularity */
7187 
get_quirk_tdp(void)7188 double get_quirk_tdp(void)
7189 {
7190 	if (platform->rapl_quirk_tdp)
7191 		return platform->rapl_quirk_tdp;
7192 
7193 	return 135.0;
7194 }
7195 
get_tdp_intel(void)7196 double get_tdp_intel(void)
7197 {
7198 	unsigned long long msr;
7199 
7200 	if (platform->rapl_msrs & RAPL_PKG_POWER_INFO)
7201 		if (!get_msr(base_cpu, MSR_PKG_POWER_INFO, &msr))
7202 			return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
7203 	return get_quirk_tdp();
7204 }
7205 
get_tdp_amd(void)7206 double get_tdp_amd(void)
7207 {
7208 	return get_quirk_tdp();
7209 }
7210 
rapl_probe_intel(void)7211 void rapl_probe_intel(void)
7212 {
7213 	unsigned long long msr;
7214 	unsigned int time_unit;
7215 	double tdp;
7216 	const unsigned long long bic_watt_bits = BIC_SysWatt | BIC_PkgWatt | BIC_CorWatt | BIC_RAMWatt | BIC_GFXWatt;
7217 	const unsigned long long bic_joules_bits = BIC_Sys_J | BIC_Pkg_J | BIC_Cor_J | BIC_RAM_J | BIC_GFX_J;
7218 
7219 	if (rapl_joules)
7220 		bic_enabled &= ~bic_watt_bits;
7221 	else
7222 		bic_enabled &= ~bic_joules_bits;
7223 
7224 	if (!(platform->rapl_msrs & RAPL_PKG_PERF_STATUS))
7225 		bic_enabled &= ~BIC_PKG__;
7226 	if (!(platform->rapl_msrs & RAPL_DRAM_PERF_STATUS))
7227 		bic_enabled &= ~BIC_RAM__;
7228 
7229 	/* units on package 0, verify later other packages match */
7230 	if (get_msr(base_cpu, MSR_RAPL_POWER_UNIT, &msr))
7231 		return;
7232 
7233 	rapl_power_units = 1.0 / (1 << (msr & 0xF));
7234 	if (platform->has_rapl_divisor)
7235 		rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
7236 	else
7237 		rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
7238 
7239 	if (platform->has_fixed_rapl_unit)
7240 		rapl_dram_energy_units = (15.3 / 1000000);
7241 	else
7242 		rapl_dram_energy_units = rapl_energy_units;
7243 
7244 	if (platform->has_fixed_rapl_psys_unit)
7245 		rapl_psys_energy_units = 1.0;
7246 	else
7247 		rapl_psys_energy_units = rapl_energy_units;
7248 
7249 	time_unit = msr >> 16 & 0xF;
7250 	if (time_unit == 0)
7251 		time_unit = 0xA;
7252 
7253 	rapl_time_units = 1.0 / (1 << (time_unit));
7254 
7255 	tdp = get_tdp_intel();
7256 
7257 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
7258 	if (!quiet)
7259 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
7260 }
7261 
rapl_probe_amd(void)7262 void rapl_probe_amd(void)
7263 {
7264 	unsigned long long msr;
7265 	double tdp;
7266 	const unsigned long long bic_watt_bits = BIC_PkgWatt | BIC_CorWatt;
7267 	const unsigned long long bic_joules_bits = BIC_Pkg_J | BIC_Cor_J;
7268 
7269 	if (rapl_joules)
7270 		bic_enabled &= ~bic_watt_bits;
7271 	else
7272 		bic_enabled &= ~bic_joules_bits;
7273 
7274 	if (get_msr(base_cpu, MSR_RAPL_PWR_UNIT, &msr))
7275 		return;
7276 
7277 	rapl_time_units = ldexp(1.0, -(msr >> 16 & 0xf));
7278 	rapl_energy_units = ldexp(1.0, -(msr >> 8 & 0x1f));
7279 	rapl_power_units = ldexp(1.0, -(msr & 0xf));
7280 
7281 	tdp = get_tdp_amd();
7282 
7283 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
7284 	if (!quiet)
7285 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
7286 }
7287 
print_power_limit_msr(int cpu,unsigned long long msr,char * label)7288 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
7289 {
7290 	fprintf(outf, "cpu%d: %s: %sabled (%0.3f Watts, %f sec, clamp %sabled)\n",
7291 		cpu, label,
7292 		((msr >> 15) & 1) ? "EN" : "DIS",
7293 		((msr >> 0) & 0x7FFF) * rapl_power_units,
7294 		(1.0 + (((msr >> 22) & 0x3) / 4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
7295 		(((msr >> 16) & 1) ? "EN" : "DIS"));
7296 
7297 	return;
7298 }
7299 
print_rapl(struct thread_data * t,struct core_data * c,struct pkg_data * p)7300 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7301 {
7302 	unsigned long long msr;
7303 	const char *msr_name;
7304 	int cpu;
7305 
7306 	UNUSED(c);
7307 	UNUSED(p);
7308 
7309 	if (!platform->rapl_msrs)
7310 		return 0;
7311 
7312 	/* RAPL counters are per package, so print only for 1st thread/package */
7313 	if (!is_cpu_first_thread_in_package(t, c, p))
7314 		return 0;
7315 
7316 	cpu = t->cpu_id;
7317 	if (cpu_migrate(cpu)) {
7318 		fprintf(outf, "print_rapl: Could not migrate to CPU %d\n", cpu);
7319 		return -1;
7320 	}
7321 
7322 	if (platform->rapl_msrs & RAPL_AMD_F17H) {
7323 		msr_name = "MSR_RAPL_PWR_UNIT";
7324 		if (get_msr(cpu, MSR_RAPL_PWR_UNIT, &msr))
7325 			return -1;
7326 	} else {
7327 		msr_name = "MSR_RAPL_POWER_UNIT";
7328 		if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
7329 			return -1;
7330 	}
7331 
7332 	fprintf(outf, "cpu%d: %s: 0x%08llx (%f Watts, %f Joules, %f sec.)\n", cpu, msr_name, msr,
7333 		rapl_power_units, rapl_energy_units, rapl_time_units);
7334 
7335 	if (platform->rapl_msrs & RAPL_PKG_POWER_INFO) {
7336 
7337 		if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
7338 			return -5;
7339 
7340 		fprintf(outf, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
7341 			cpu, msr,
7342 			((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7343 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7344 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7345 			((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
7346 
7347 	}
7348 	if (platform->rapl_msrs & RAPL_PKG) {
7349 
7350 		if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
7351 			return -9;
7352 
7353 		fprintf(outf, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
7354 			cpu, msr, (msr >> 63) & 1 ? "" : "UN");
7355 
7356 		print_power_limit_msr(cpu, msr, "PKG Limit #1");
7357 		fprintf(outf, "cpu%d: PKG Limit #2: %sabled (%0.3f Watts, %f* sec, clamp %sabled)\n",
7358 			cpu,
7359 			((msr >> 47) & 1) ? "EN" : "DIS",
7360 			((msr >> 32) & 0x7FFF) * rapl_power_units,
7361 			(1.0 + (((msr >> 54) & 0x3) / 4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
7362 			((msr >> 48) & 1) ? "EN" : "DIS");
7363 
7364 		if (get_msr(cpu, MSR_VR_CURRENT_CONFIG, &msr))
7365 			return -9;
7366 
7367 		fprintf(outf, "cpu%d: MSR_VR_CURRENT_CONFIG: 0x%08llx\n", cpu, msr);
7368 		fprintf(outf, "cpu%d: PKG Limit #4: %f Watts (%slocked)\n",
7369 			cpu, ((msr >> 0) & 0x1FFF) * rapl_power_units, (msr >> 31) & 1 ? "" : "UN");
7370 	}
7371 
7372 	if (platform->rapl_msrs & RAPL_DRAM_POWER_INFO) {
7373 		if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
7374 			return -6;
7375 
7376 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
7377 			cpu, msr,
7378 			((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7379 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7380 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7381 			((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
7382 	}
7383 	if (platform->rapl_msrs & RAPL_DRAM) {
7384 		if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
7385 			return -9;
7386 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
7387 			cpu, msr, (msr >> 31) & 1 ? "" : "UN");
7388 
7389 		print_power_limit_msr(cpu, msr, "DRAM Limit");
7390 	}
7391 	if (platform->rapl_msrs & RAPL_CORE_POLICY) {
7392 		if (get_msr(cpu, MSR_PP0_POLICY, &msr))
7393 			return -7;
7394 
7395 		fprintf(outf, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
7396 	}
7397 	if (platform->rapl_msrs & RAPL_CORE_POWER_LIMIT) {
7398 		if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
7399 			return -9;
7400 		fprintf(outf, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
7401 			cpu, msr, (msr >> 31) & 1 ? "" : "UN");
7402 		print_power_limit_msr(cpu, msr, "Cores Limit");
7403 	}
7404 	if (platform->rapl_msrs & RAPL_GFX) {
7405 		if (get_msr(cpu, MSR_PP1_POLICY, &msr))
7406 			return -8;
7407 
7408 		fprintf(outf, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
7409 
7410 		if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
7411 			return -9;
7412 		fprintf(outf, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
7413 			cpu, msr, (msr >> 31) & 1 ? "" : "UN");
7414 		print_power_limit_msr(cpu, msr, "GFX Limit");
7415 	}
7416 	return 0;
7417 }
7418 
7419 /*
7420  * probe_rapl()
7421  *
7422  * sets rapl_power_units, rapl_energy_units, rapl_time_units
7423  */
probe_rapl(void)7424 void probe_rapl(void)
7425 {
7426 	if (!platform->rapl_msrs || no_msr)
7427 		return;
7428 
7429 	if (genuine_intel)
7430 		rapl_probe_intel();
7431 	if (authentic_amd || hygon_genuine)
7432 		rapl_probe_amd();
7433 
7434 	if (quiet)
7435 		return;
7436 
7437 	for_all_cpus(print_rapl, ODD_COUNTERS);
7438 }
7439 
7440 /*
7441  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
7442  * the Thermal Control Circuit (TCC) activates.
7443  * This is usually equal to tjMax.
7444  *
7445  * Older processors do not have this MSR, so there we guess,
7446  * but also allow cmdline over-ride with -T.
7447  *
7448  * Several MSR temperature values are in units of degrees-C
7449  * below this value, including the Digital Thermal Sensor (DTS),
7450  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
7451  */
set_temperature_target(struct thread_data * t,struct core_data * c,struct pkg_data * p)7452 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7453 {
7454 	unsigned long long msr;
7455 	unsigned int tcc_default, tcc_offset;
7456 	int cpu;
7457 
7458 	UNUSED(c);
7459 	UNUSED(p);
7460 
7461 	/* tj_max is used only for dts or ptm */
7462 	if (!(do_dts || do_ptm))
7463 		return 0;
7464 
7465 	/* this is a per-package concept */
7466 	if (!is_cpu_first_thread_in_package(t, c, p))
7467 		return 0;
7468 
7469 	cpu = t->cpu_id;
7470 	if (cpu_migrate(cpu)) {
7471 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
7472 		return -1;
7473 	}
7474 
7475 	if (tj_max_override != 0) {
7476 		tj_max = tj_max_override;
7477 		fprintf(outf, "cpu%d: Using cmdline TCC Target (%d C)\n", cpu, tj_max);
7478 		return 0;
7479 	}
7480 
7481 	/* Temperature Target MSR is Nehalem and newer only */
7482 	if (!platform->has_nhm_msrs || no_msr)
7483 		goto guess;
7484 
7485 	if (get_msr(base_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
7486 		goto guess;
7487 
7488 	tcc_default = (msr >> 16) & 0xFF;
7489 
7490 	if (!quiet) {
7491 		int bits = platform->tcc_offset_bits;
7492 		unsigned long long enabled = 0;
7493 
7494 		if (bits && !get_msr(base_cpu, MSR_PLATFORM_INFO, &enabled))
7495 			enabled = (enabled >> 30) & 1;
7496 
7497 		if (bits && enabled) {
7498 			tcc_offset = (msr >> 24) & GENMASK(bits - 1, 0);
7499 			fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C) (%d default - %d offset)\n",
7500 				cpu, msr, tcc_default - tcc_offset, tcc_default, tcc_offset);
7501 		} else {
7502 			fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n", cpu, msr, tcc_default);
7503 		}
7504 	}
7505 
7506 	if (!tcc_default)
7507 		goto guess;
7508 
7509 	tj_max = tcc_default;
7510 
7511 	return 0;
7512 
7513 guess:
7514 	tj_max = TJMAX_DEFAULT;
7515 	fprintf(outf, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n", cpu, tj_max);
7516 
7517 	return 0;
7518 }
7519 
print_thermal(struct thread_data * t,struct core_data * c,struct pkg_data * p)7520 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7521 {
7522 	unsigned long long msr;
7523 	unsigned int dts, dts2;
7524 	int cpu;
7525 
7526 	UNUSED(c);
7527 	UNUSED(p);
7528 
7529 	if (no_msr)
7530 		return 0;
7531 
7532 	if (!(do_dts || do_ptm))
7533 		return 0;
7534 
7535 	cpu = t->cpu_id;
7536 
7537 	/* DTS is per-core, no need to print for each thread */
7538 	if (!is_cpu_first_thread_in_core(t, c, p))
7539 		return 0;
7540 
7541 	if (cpu_migrate(cpu)) {
7542 		fprintf(outf, "print_thermal: Could not migrate to CPU %d\n", cpu);
7543 		return -1;
7544 	}
7545 
7546 	if (do_ptm && is_cpu_first_core_in_package(t, c, p)) {
7547 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
7548 			return 0;
7549 
7550 		dts = (msr >> 16) & 0x7F;
7551 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n", cpu, msr, tj_max - dts);
7552 
7553 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
7554 			return 0;
7555 
7556 		dts = (msr >> 16) & 0x7F;
7557 		dts2 = (msr >> 8) & 0x7F;
7558 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
7559 			cpu, msr, tj_max - dts, tj_max - dts2);
7560 	}
7561 
7562 	if (do_dts && debug) {
7563 		unsigned int resolution;
7564 
7565 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
7566 			return 0;
7567 
7568 		dts = (msr >> 16) & 0x7F;
7569 		resolution = (msr >> 27) & 0xF;
7570 		fprintf(outf, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
7571 			cpu, msr, tj_max - dts, resolution);
7572 
7573 		if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
7574 			return 0;
7575 
7576 		dts = (msr >> 16) & 0x7F;
7577 		dts2 = (msr >> 8) & 0x7F;
7578 		fprintf(outf, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
7579 			cpu, msr, tj_max - dts, tj_max - dts2);
7580 	}
7581 
7582 	return 0;
7583 }
7584 
probe_thermal(void)7585 void probe_thermal(void)
7586 {
7587 	if (!access("/sys/devices/system/cpu/cpu0/thermal_throttle/core_throttle_count", R_OK))
7588 		BIC_PRESENT(BIC_CORE_THROT_CNT);
7589 	else
7590 		BIC_NOT_PRESENT(BIC_CORE_THROT_CNT);
7591 
7592 	for_all_cpus(set_temperature_target, ODD_COUNTERS);
7593 
7594 	if (quiet)
7595 		return;
7596 
7597 	for_all_cpus(print_thermal, ODD_COUNTERS);
7598 }
7599 
get_cpu_type(struct thread_data * t,struct core_data * c,struct pkg_data * p)7600 int get_cpu_type(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7601 {
7602 	unsigned int eax, ebx, ecx, edx;
7603 
7604 	UNUSED(c);
7605 	UNUSED(p);
7606 
7607 	if (!genuine_intel)
7608 		return 0;
7609 
7610 	if (cpu_migrate(t->cpu_id)) {
7611 		fprintf(outf, "Could not migrate to CPU %d\n", t->cpu_id);
7612 		return -1;
7613 	}
7614 
7615 	if (max_level < 0x1a)
7616 		return 0;
7617 
7618 	__cpuid(0x1a, eax, ebx, ecx, edx);
7619 	eax = (eax >> 24) & 0xFF;
7620 	if (eax == 0x20)
7621 		t->is_atom = true;
7622 	return 0;
7623 }
7624 
decode_feature_control_msr(void)7625 void decode_feature_control_msr(void)
7626 {
7627 	unsigned long long msr;
7628 
7629 	if (no_msr)
7630 		return;
7631 
7632 	if (!get_msr(base_cpu, MSR_IA32_FEAT_CTL, &msr))
7633 		fprintf(outf, "cpu%d: MSR_IA32_FEATURE_CONTROL: 0x%08llx (%sLocked %s)\n",
7634 			base_cpu, msr, msr & FEAT_CTL_LOCKED ? "" : "UN-", msr & (1 << 18) ? "SGX" : "");
7635 }
7636 
decode_misc_enable_msr(void)7637 void decode_misc_enable_msr(void)
7638 {
7639 	unsigned long long msr;
7640 
7641 	if (no_msr)
7642 		return;
7643 
7644 	if (!genuine_intel)
7645 		return;
7646 
7647 	if (!get_msr(base_cpu, MSR_IA32_MISC_ENABLE, &msr))
7648 		fprintf(outf, "cpu%d: MSR_IA32_MISC_ENABLE: 0x%08llx (%sTCC %sEIST %sMWAIT %sPREFETCH %sTURBO)\n",
7649 			base_cpu, msr,
7650 			msr & MSR_IA32_MISC_ENABLE_TM1 ? "" : "No-",
7651 			msr & MSR_IA32_MISC_ENABLE_ENHANCED_SPEEDSTEP ? "" : "No-",
7652 			msr & MSR_IA32_MISC_ENABLE_MWAIT ? "" : "No-",
7653 			msr & MSR_IA32_MISC_ENABLE_PREFETCH_DISABLE ? "No-" : "",
7654 			msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ? "No-" : "");
7655 }
7656 
decode_misc_feature_control(void)7657 void decode_misc_feature_control(void)
7658 {
7659 	unsigned long long msr;
7660 
7661 	if (no_msr)
7662 		return;
7663 
7664 	if (!platform->has_msr_misc_feature_control)
7665 		return;
7666 
7667 	if (!get_msr(base_cpu, MSR_MISC_FEATURE_CONTROL, &msr))
7668 		fprintf(outf,
7669 			"cpu%d: MSR_MISC_FEATURE_CONTROL: 0x%08llx (%sL2-Prefetch %sL2-Prefetch-pair %sL1-Prefetch %sL1-IP-Prefetch)\n",
7670 			base_cpu, msr, msr & (0 << 0) ? "No-" : "", msr & (1 << 0) ? "No-" : "",
7671 			msr & (2 << 0) ? "No-" : "", msr & (3 << 0) ? "No-" : "");
7672 }
7673 
7674 /*
7675  * Decode MSR_MISC_PWR_MGMT
7676  *
7677  * Decode the bits according to the Nehalem documentation
7678  * bit[0] seems to continue to have same meaning going forward
7679  * bit[1] less so...
7680  */
decode_misc_pwr_mgmt_msr(void)7681 void decode_misc_pwr_mgmt_msr(void)
7682 {
7683 	unsigned long long msr;
7684 
7685 	if (no_msr)
7686 		return;
7687 
7688 	if (!platform->has_msr_misc_pwr_mgmt)
7689 		return;
7690 
7691 	if (!get_msr(base_cpu, MSR_MISC_PWR_MGMT, &msr))
7692 		fprintf(outf, "cpu%d: MSR_MISC_PWR_MGMT: 0x%08llx (%sable-EIST_Coordination %sable-EPB %sable-OOB)\n",
7693 			base_cpu, msr,
7694 			msr & (1 << 0) ? "DIS" : "EN", msr & (1 << 1) ? "EN" : "DIS", msr & (1 << 8) ? "EN" : "DIS");
7695 }
7696 
7697 /*
7698  * Decode MSR_CC6_DEMOTION_POLICY_CONFIG, MSR_MC6_DEMOTION_POLICY_CONFIG
7699  *
7700  * This MSRs are present on Silvermont processors,
7701  * Intel Atom processor E3000 series (Baytrail), and friends.
7702  */
decode_c6_demotion_policy_msr(void)7703 void decode_c6_demotion_policy_msr(void)
7704 {
7705 	unsigned long long msr;
7706 
7707 	if (no_msr)
7708 		return;
7709 
7710 	if (!platform->has_msr_c6_demotion_policy_config)
7711 		return;
7712 
7713 	if (!get_msr(base_cpu, MSR_CC6_DEMOTION_POLICY_CONFIG, &msr))
7714 		fprintf(outf, "cpu%d: MSR_CC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-CC6-Demotion)\n",
7715 			base_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
7716 
7717 	if (!get_msr(base_cpu, MSR_MC6_DEMOTION_POLICY_CONFIG, &msr))
7718 		fprintf(outf, "cpu%d: MSR_MC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-MC6-Demotion)\n",
7719 			base_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
7720 }
7721 
print_dev_latency(void)7722 void print_dev_latency(void)
7723 {
7724 	char *path = "/dev/cpu_dma_latency";
7725 	int fd;
7726 	int value;
7727 	int retval;
7728 
7729 	fd = open(path, O_RDONLY);
7730 	if (fd < 0) {
7731 		if (debug)
7732 			warnx("Read %s failed", path);
7733 		return;
7734 	}
7735 
7736 	retval = read(fd, (void *)&value, sizeof(int));
7737 	if (retval != sizeof(int)) {
7738 		warn("read failed %s", path);
7739 		close(fd);
7740 		return;
7741 	}
7742 	fprintf(outf, "/dev/cpu_dma_latency: %d usec (%s)\n", value, value == 2000000000 ? "default" : "constrained");
7743 
7744 	close(fd);
7745 }
7746 
has_instr_count_access(void)7747 static int has_instr_count_access(void)
7748 {
7749 	int fd;
7750 	int has_access;
7751 
7752 	if (no_perf)
7753 		return 0;
7754 
7755 	fd = open_perf_counter(base_cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, -1, 0);
7756 	has_access = fd != -1;
7757 
7758 	if (fd != -1)
7759 		close(fd);
7760 
7761 	if (!has_access)
7762 		warnx("Failed to access %s. Some of the counters may not be available\n"
7763 		      "\tRun as root to enable them or use %s to disable the access explicitly",
7764 		      "instructions retired perf counter", "--no-perf");
7765 
7766 	return has_access;
7767 }
7768 
add_rapl_perf_counter_(int cpu,struct rapl_counter_info_t * rci,const struct rapl_counter_arch_info * cai,double * scale_,enum rapl_unit * unit_)7769 int add_rapl_perf_counter_(int cpu, struct rapl_counter_info_t *rci, const struct rapl_counter_arch_info *cai,
7770 			   double *scale_, enum rapl_unit *unit_)
7771 {
7772 	if (no_perf)
7773 		return -1;
7774 
7775 	const double scale = read_perf_scale(cai->perf_subsys, cai->perf_name);
7776 
7777 	if (scale == 0.0)
7778 		return -1;
7779 
7780 	const enum rapl_unit unit = read_perf_rapl_unit(cai->perf_subsys, cai->perf_name);
7781 
7782 	if (unit == RAPL_UNIT_INVALID)
7783 		return -1;
7784 
7785 	const unsigned int rapl_type = read_perf_type(cai->perf_subsys);
7786 	const unsigned int rapl_energy_pkg_config = read_perf_config(cai->perf_subsys, cai->perf_name);
7787 
7788 	const int fd_counter =
7789 	    open_perf_counter(cpu, rapl_type, rapl_energy_pkg_config, rci->fd_perf, PERF_FORMAT_GROUP);
7790 	if (fd_counter == -1)
7791 		return -1;
7792 
7793 	/* If it's the first counter opened, make it a group descriptor */
7794 	if (rci->fd_perf == -1)
7795 		rci->fd_perf = fd_counter;
7796 
7797 	*scale_ = scale;
7798 	*unit_ = unit;
7799 	return fd_counter;
7800 }
7801 
add_rapl_perf_counter(int cpu,struct rapl_counter_info_t * rci,const struct rapl_counter_arch_info * cai,double * scale,enum rapl_unit * unit)7802 int add_rapl_perf_counter(int cpu, struct rapl_counter_info_t *rci, const struct rapl_counter_arch_info *cai,
7803 			  double *scale, enum rapl_unit *unit)
7804 {
7805 	int ret = add_rapl_perf_counter_(cpu, rci, cai, scale, unit);
7806 
7807 	if (debug >= 2)
7808 		fprintf(stderr, "%s: %d (cpu: %d)\n", __func__, ret, cpu);
7809 
7810 	return ret;
7811 }
7812 
7813 /*
7814  * Linux-perf manages the HW instructions-retired counter
7815  * by enabling when requested, and hiding rollover
7816  */
linux_perf_init(void)7817 void linux_perf_init(void)
7818 {
7819 	if (access("/proc/sys/kernel/perf_event_paranoid", F_OK))
7820 		return;
7821 
7822 	if (BIC_IS_ENABLED(BIC_IPC) && has_aperf) {
7823 		fd_instr_count_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
7824 		if (fd_instr_count_percpu == NULL)
7825 			err(-1, "calloc fd_instr_count_percpu");
7826 	}
7827 }
7828 
rapl_perf_init(void)7829 void rapl_perf_init(void)
7830 {
7831 	const unsigned int num_domains = (platform->has_per_core_rapl ? topo.max_core_id : topo.max_package_id) + 1;
7832 	bool *domain_visited = calloc(num_domains, sizeof(bool));
7833 
7834 	rapl_counter_info_perdomain = calloc(num_domains, sizeof(*rapl_counter_info_perdomain));
7835 	if (rapl_counter_info_perdomain == NULL)
7836 		err(-1, "calloc rapl_counter_info_percpu");
7837 	rapl_counter_info_perdomain_size = num_domains;
7838 
7839 	/*
7840 	 * Initialize rapl_counter_info_percpu
7841 	 */
7842 	for (unsigned int domain_id = 0; domain_id < num_domains; ++domain_id) {
7843 		struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[domain_id];
7844 
7845 		rci->fd_perf = -1;
7846 		for (size_t i = 0; i < NUM_RAPL_COUNTERS; ++i) {
7847 			rci->data[i] = 0;
7848 			rci->source[i] = COUNTER_SOURCE_NONE;
7849 		}
7850 	}
7851 
7852 	/*
7853 	 * Open/probe the counters
7854 	 * If can't get it via perf, fallback to MSR
7855 	 */
7856 	for (size_t i = 0; i < ARRAY_SIZE(rapl_counter_arch_infos); ++i) {
7857 
7858 		const struct rapl_counter_arch_info *const cai = &rapl_counter_arch_infos[i];
7859 		bool has_counter = 0;
7860 		double scale;
7861 		enum rapl_unit unit;
7862 		unsigned int next_domain;
7863 
7864 		memset(domain_visited, 0, num_domains * sizeof(*domain_visited));
7865 
7866 		for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
7867 
7868 			if (cpu_is_not_allowed(cpu))
7869 				continue;
7870 
7871 			/* Skip already seen and handled RAPL domains */
7872 			next_domain =
7873 			    platform->has_per_core_rapl ? cpus[cpu].physical_core_id : cpus[cpu].physical_package_id;
7874 
7875 			assert(next_domain < num_domains);
7876 
7877 			if (domain_visited[next_domain])
7878 				continue;
7879 
7880 			domain_visited[next_domain] = 1;
7881 
7882 			if ((cai->flags & RAPL_COUNTER_FLAG_PLATFORM_COUNTER) && (cpu != base_cpu))
7883 				continue;
7884 
7885 			struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[next_domain];
7886 
7887 			/* Check if the counter is enabled and accessible */
7888 			if (BIC_IS_ENABLED(cai->bic) && (platform->rapl_msrs & cai->feature_mask)) {
7889 
7890 				/* Use perf API for this counter */
7891 				if (!no_perf && cai->perf_name
7892 				    && add_rapl_perf_counter(cpu, rci, cai, &scale, &unit) != -1) {
7893 					rci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
7894 					rci->scale[cai->rci_index] = scale * cai->compat_scale;
7895 					rci->unit[cai->rci_index] = unit;
7896 					rci->flags[cai->rci_index] = cai->flags;
7897 
7898 					/* Use MSR for this counter */
7899 				} else if (!no_msr && cai->msr && probe_msr(cpu, cai->msr) == 0) {
7900 					rci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
7901 					rci->msr[cai->rci_index] = cai->msr;
7902 					rci->msr_mask[cai->rci_index] = cai->msr_mask;
7903 					rci->msr_shift[cai->rci_index] = cai->msr_shift;
7904 					rci->unit[cai->rci_index] = RAPL_UNIT_JOULES;
7905 					rci->scale[cai->rci_index] = *cai->platform_rapl_msr_scale * cai->compat_scale;
7906 					rci->flags[cai->rci_index] = cai->flags;
7907 				}
7908 			}
7909 
7910 			if (rci->source[cai->rci_index] != COUNTER_SOURCE_NONE)
7911 				has_counter = 1;
7912 		}
7913 
7914 		/* If any CPU has access to the counter, make it present */
7915 		if (has_counter)
7916 			BIC_PRESENT(cai->bic);
7917 	}
7918 
7919 	free(domain_visited);
7920 }
7921 
7922 /* Assumes msr_counter_info is populated */
has_amperf_access(void)7923 static int has_amperf_access(void)
7924 {
7925 	return msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].present &&
7926 	    msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].present;
7927 }
7928 
get_cstate_perf_group_fd(struct cstate_counter_info_t * cci,const char * group_name)7929 int *get_cstate_perf_group_fd(struct cstate_counter_info_t *cci, const char *group_name)
7930 {
7931 	if (strcmp(group_name, "cstate_core") == 0)
7932 		return &cci->fd_perf_core;
7933 
7934 	if (strcmp(group_name, "cstate_pkg") == 0)
7935 		return &cci->fd_perf_pkg;
7936 
7937 	return NULL;
7938 }
7939 
add_cstate_perf_counter_(int cpu,struct cstate_counter_info_t * cci,const struct cstate_counter_arch_info * cai)7940 int add_cstate_perf_counter_(int cpu, struct cstate_counter_info_t *cci, const struct cstate_counter_arch_info *cai)
7941 {
7942 	if (no_perf)
7943 		return -1;
7944 
7945 	int *pfd_group = get_cstate_perf_group_fd(cci, cai->perf_subsys);
7946 
7947 	if (pfd_group == NULL)
7948 		return -1;
7949 
7950 	const unsigned int type = read_perf_type(cai->perf_subsys);
7951 	const unsigned int config = read_perf_config(cai->perf_subsys, cai->perf_name);
7952 
7953 	const int fd_counter = open_perf_counter(cpu, type, config, *pfd_group, PERF_FORMAT_GROUP);
7954 
7955 	if (fd_counter == -1)
7956 		return -1;
7957 
7958 	/* If it's the first counter opened, make it a group descriptor */
7959 	if (*pfd_group == -1)
7960 		*pfd_group = fd_counter;
7961 
7962 	return fd_counter;
7963 }
7964 
add_cstate_perf_counter(int cpu,struct cstate_counter_info_t * cci,const struct cstate_counter_arch_info * cai)7965 int add_cstate_perf_counter(int cpu, struct cstate_counter_info_t *cci, const struct cstate_counter_arch_info *cai)
7966 {
7967 	int ret = add_cstate_perf_counter_(cpu, cci, cai);
7968 
7969 	if (debug >= 2)
7970 		fprintf(stderr, "%s: %d (cpu: %d)\n", __func__, ret, cpu);
7971 
7972 	return ret;
7973 }
7974 
add_msr_perf_counter_(int cpu,struct msr_counter_info_t * cci,const struct msr_counter_arch_info * cai)7975 int add_msr_perf_counter_(int cpu, struct msr_counter_info_t *cci, const struct msr_counter_arch_info *cai)
7976 {
7977 	if (no_perf)
7978 		return -1;
7979 
7980 	const unsigned int type = read_perf_type(cai->perf_subsys);
7981 	const unsigned int config = read_perf_config(cai->perf_subsys, cai->perf_name);
7982 
7983 	const int fd_counter = open_perf_counter(cpu, type, config, cci->fd_perf, PERF_FORMAT_GROUP);
7984 
7985 	if (fd_counter == -1)
7986 		return -1;
7987 
7988 	/* If it's the first counter opened, make it a group descriptor */
7989 	if (cci->fd_perf == -1)
7990 		cci->fd_perf = fd_counter;
7991 
7992 	return fd_counter;
7993 }
7994 
add_msr_perf_counter(int cpu,struct msr_counter_info_t * cci,const struct msr_counter_arch_info * cai)7995 int add_msr_perf_counter(int cpu, struct msr_counter_info_t *cci, const struct msr_counter_arch_info *cai)
7996 {
7997 	int ret = add_msr_perf_counter_(cpu, cci, cai);
7998 
7999 	if (debug)
8000 		fprintf(stderr, "%s: %s/%s: %d (cpu: %d)\n", __func__, cai->perf_subsys, cai->perf_name, ret, cpu);
8001 
8002 	return ret;
8003 }
8004 
msr_perf_init_(void)8005 void msr_perf_init_(void)
8006 {
8007 	const int mci_num = topo.max_cpu_num + 1;
8008 
8009 	msr_counter_info = calloc(mci_num, sizeof(*msr_counter_info));
8010 	if (!msr_counter_info)
8011 		err(1, "calloc msr_counter_info");
8012 	msr_counter_info_size = mci_num;
8013 
8014 	for (int cpu = 0; cpu < mci_num; ++cpu)
8015 		msr_counter_info[cpu].fd_perf = -1;
8016 
8017 	for (int cidx = 0; cidx < NUM_MSR_COUNTERS; ++cidx) {
8018 
8019 		struct msr_counter_arch_info *cai = &msr_counter_arch_infos[cidx];
8020 
8021 		cai->present = false;
8022 
8023 		for (int cpu = 0; cpu < mci_num; ++cpu) {
8024 
8025 			struct msr_counter_info_t *const cci = &msr_counter_info[cpu];
8026 
8027 			if (cpu_is_not_allowed(cpu))
8028 				continue;
8029 
8030 			if (cai->needed) {
8031 				/* Use perf API for this counter */
8032 				if (!no_perf && cai->perf_name && add_msr_perf_counter(cpu, cci, cai) != -1) {
8033 					cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
8034 					cai->present = true;
8035 
8036 					/* User MSR for this counter */
8037 				} else if (!no_msr && cai->msr && probe_msr(cpu, cai->msr) == 0) {
8038 					cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
8039 					cci->msr[cai->rci_index] = cai->msr;
8040 					cci->msr_mask[cai->rci_index] = cai->msr_mask;
8041 					cai->present = true;
8042 				}
8043 			}
8044 		}
8045 	}
8046 }
8047 
8048 /* Initialize data for reading perf counters from the MSR group. */
msr_perf_init(void)8049 void msr_perf_init(void)
8050 {
8051 	bool need_amperf = false, need_smi = false;
8052 	const bool need_soft_c1 = (!platform->has_msr_core_c1_res) && (platform->supported_cstates & CC1);
8053 
8054 	need_amperf = BIC_IS_ENABLED(BIC_Avg_MHz) || BIC_IS_ENABLED(BIC_Busy) || BIC_IS_ENABLED(BIC_Bzy_MHz)
8055 	    || BIC_IS_ENABLED(BIC_IPC) || need_soft_c1;
8056 
8057 	if (BIC_IS_ENABLED(BIC_SMI))
8058 		need_smi = true;
8059 
8060 	/* Enable needed counters */
8061 	msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].needed = need_amperf;
8062 	msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].needed = need_amperf;
8063 	msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].needed = need_smi;
8064 
8065 	msr_perf_init_();
8066 
8067 	const bool has_amperf = has_amperf_access();
8068 	const bool has_smi = msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].present;
8069 
8070 	has_aperf_access = has_amperf;
8071 
8072 	if (has_amperf) {
8073 		BIC_PRESENT(BIC_Avg_MHz);
8074 		BIC_PRESENT(BIC_Busy);
8075 		BIC_PRESENT(BIC_Bzy_MHz);
8076 		BIC_PRESENT(BIC_SMI);
8077 	}
8078 
8079 	if (has_smi)
8080 		BIC_PRESENT(BIC_SMI);
8081 }
8082 
cstate_perf_init_(bool soft_c1)8083 void cstate_perf_init_(bool soft_c1)
8084 {
8085 	bool has_counter;
8086 	bool *cores_visited = NULL, *pkg_visited = NULL;
8087 	const int cores_visited_elems = topo.max_core_id + 1;
8088 	const int pkg_visited_elems = topo.max_package_id + 1;
8089 	const int cci_num = topo.max_cpu_num + 1;
8090 
8091 	ccstate_counter_info = calloc(cci_num, sizeof(*ccstate_counter_info));
8092 	if (!ccstate_counter_info)
8093 		err(1, "calloc ccstate_counter_arch_info");
8094 	ccstate_counter_info_size = cci_num;
8095 
8096 	cores_visited = calloc(cores_visited_elems, sizeof(*cores_visited));
8097 	if (!cores_visited)
8098 		err(1, "calloc cores_visited");
8099 
8100 	pkg_visited = calloc(pkg_visited_elems, sizeof(*pkg_visited));
8101 	if (!pkg_visited)
8102 		err(1, "calloc pkg_visited");
8103 
8104 	/* Initialize cstate_counter_info_percpu */
8105 	for (int cpu = 0; cpu < cci_num; ++cpu) {
8106 		ccstate_counter_info[cpu].fd_perf_core = -1;
8107 		ccstate_counter_info[cpu].fd_perf_pkg = -1;
8108 	}
8109 
8110 	for (int cidx = 0; cidx < NUM_CSTATE_COUNTERS; ++cidx) {
8111 		has_counter = false;
8112 		memset(cores_visited, 0, cores_visited_elems * sizeof(*cores_visited));
8113 		memset(pkg_visited, 0, pkg_visited_elems * sizeof(*pkg_visited));
8114 
8115 		const struct cstate_counter_arch_info *cai = &ccstate_counter_arch_infos[cidx];
8116 
8117 		for (int cpu = 0; cpu < cci_num; ++cpu) {
8118 
8119 			struct cstate_counter_info_t *const cci = &ccstate_counter_info[cpu];
8120 
8121 			if (cpu_is_not_allowed(cpu))
8122 				continue;
8123 
8124 			const int core_id = cpus[cpu].physical_core_id;
8125 			const int pkg_id = cpus[cpu].physical_package_id;
8126 
8127 			assert(core_id < cores_visited_elems);
8128 			assert(pkg_id < pkg_visited_elems);
8129 
8130 			const bool per_thread = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD;
8131 			const bool per_core = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_CORE;
8132 
8133 			if (!per_thread && cores_visited[core_id])
8134 				continue;
8135 
8136 			if (!per_core && pkg_visited[pkg_id])
8137 				continue;
8138 
8139 			const bool counter_needed = BIC_IS_ENABLED(cai->bic) ||
8140 			    (soft_c1 && (cai->flags & CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY));
8141 			const bool counter_supported = (platform->supported_cstates & cai->feature_mask);
8142 
8143 			if (counter_needed && counter_supported) {
8144 				/* Use perf API for this counter */
8145 				if (!no_perf && cai->perf_name && add_cstate_perf_counter(cpu, cci, cai) != -1) {
8146 
8147 					cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
8148 
8149 					/* User MSR for this counter */
8150 				} else if (!no_msr && cai->msr && pkg_cstate_limit >= cai->pkg_cstate_limit
8151 					   && probe_msr(cpu, cai->msr) == 0) {
8152 					cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
8153 					cci->msr[cai->rci_index] = cai->msr;
8154 				}
8155 			}
8156 
8157 			if (cci->source[cai->rci_index] != COUNTER_SOURCE_NONE) {
8158 				has_counter = true;
8159 				cores_visited[core_id] = true;
8160 				pkg_visited[pkg_id] = true;
8161 			}
8162 		}
8163 
8164 		/* If any CPU has access to the counter, make it present */
8165 		if (has_counter)
8166 			BIC_PRESENT(cai->bic);
8167 	}
8168 
8169 	free(cores_visited);
8170 	free(pkg_visited);
8171 }
8172 
cstate_perf_init(void)8173 void cstate_perf_init(void)
8174 {
8175 	/*
8176 	 * If we don't have a C1 residency MSR, we calculate it "in software",
8177 	 * but we need APERF, MPERF too.
8178 	 */
8179 	const bool soft_c1 = !platform->has_msr_core_c1_res && has_amperf_access()
8180 	    && platform->supported_cstates & CC1;
8181 
8182 	if (soft_c1)
8183 		BIC_PRESENT(BIC_CPU_c1);
8184 
8185 	cstate_perf_init_(soft_c1);
8186 }
8187 
probe_cstates(void)8188 void probe_cstates(void)
8189 {
8190 	probe_cst_limit();
8191 
8192 	if (platform->has_msr_module_c6_res_ms)
8193 		BIC_PRESENT(BIC_Mod_c6);
8194 
8195 	if (platform->has_ext_cst_msrs && !no_msr) {
8196 		BIC_PRESENT(BIC_Totl_c0);
8197 		BIC_PRESENT(BIC_Any_c0);
8198 		BIC_PRESENT(BIC_GFX_c0);
8199 		BIC_PRESENT(BIC_CPUGFX);
8200 	}
8201 
8202 	if (quiet)
8203 		return;
8204 
8205 	dump_power_ctl();
8206 	dump_cst_cfg();
8207 	decode_c6_demotion_policy_msr();
8208 	print_dev_latency();
8209 	dump_sysfs_cstate_config();
8210 	print_irtl();
8211 }
8212 
probe_lpi(void)8213 void probe_lpi(void)
8214 {
8215 	if (!access("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", R_OK))
8216 		BIC_PRESENT(BIC_CPU_LPI);
8217 	else
8218 		BIC_NOT_PRESENT(BIC_CPU_LPI);
8219 
8220 	if (!access(sys_lpi_file_sysfs, R_OK)) {
8221 		sys_lpi_file = sys_lpi_file_sysfs;
8222 		BIC_PRESENT(BIC_SYS_LPI);
8223 	} else if (!access(sys_lpi_file_debugfs, R_OK)) {
8224 		sys_lpi_file = sys_lpi_file_debugfs;
8225 		BIC_PRESENT(BIC_SYS_LPI);
8226 	} else {
8227 		sys_lpi_file_sysfs = NULL;
8228 		BIC_NOT_PRESENT(BIC_SYS_LPI);
8229 	}
8230 
8231 }
8232 
probe_pstates(void)8233 void probe_pstates(void)
8234 {
8235 	probe_bclk();
8236 
8237 	if (quiet)
8238 		return;
8239 
8240 	dump_platform_info();
8241 	dump_turbo_ratio_info();
8242 	dump_sysfs_pstate_config();
8243 	decode_misc_pwr_mgmt_msr();
8244 
8245 	for_all_cpus(print_hwp, ODD_COUNTERS);
8246 	for_all_cpus(print_epb, ODD_COUNTERS);
8247 	for_all_cpus(print_perf_limit, ODD_COUNTERS);
8248 }
8249 
process_cpuid()8250 void process_cpuid()
8251 {
8252 	unsigned int eax, ebx, ecx, edx;
8253 	unsigned int fms, family, model, stepping, ecx_flags, edx_flags;
8254 	unsigned long long ucode_patch = 0;
8255 	bool ucode_patch_valid = false;
8256 
8257 	eax = ebx = ecx = edx = 0;
8258 
8259 	__cpuid(0, max_level, ebx, ecx, edx);
8260 
8261 	if (ebx == 0x756e6547 && ecx == 0x6c65746e && edx == 0x49656e69)
8262 		genuine_intel = 1;
8263 	else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65)
8264 		authentic_amd = 1;
8265 	else if (ebx == 0x6f677948 && ecx == 0x656e6975 && edx == 0x6e65476e)
8266 		hygon_genuine = 1;
8267 
8268 	if (!quiet)
8269 		fprintf(outf, "CPUID(0): %.4s%.4s%.4s 0x%x CPUID levels\n",
8270 			(char *)&ebx, (char *)&edx, (char *)&ecx, max_level);
8271 
8272 	__cpuid(1, fms, ebx, ecx, edx);
8273 	family = (fms >> 8) & 0xf;
8274 	model = (fms >> 4) & 0xf;
8275 	stepping = fms & 0xf;
8276 	if (family == 0xf)
8277 		family += (fms >> 20) & 0xff;
8278 	if (family >= 6)
8279 		model += ((fms >> 16) & 0xf) << 4;
8280 	ecx_flags = ecx;
8281 	edx_flags = edx;
8282 
8283 	if (!no_msr) {
8284 		if (get_msr(sched_getcpu(), MSR_IA32_UCODE_REV, &ucode_patch))
8285 			warnx("get_msr(UCODE)");
8286 		else
8287 			ucode_patch_valid = true;
8288 	}
8289 
8290 	/*
8291 	 * check max extended function levels of CPUID.
8292 	 * This is needed to check for invariant TSC.
8293 	 * This check is valid for both Intel and AMD.
8294 	 */
8295 	ebx = ecx = edx = 0;
8296 	__cpuid(0x80000000, max_extended_level, ebx, ecx, edx);
8297 
8298 	if (!quiet) {
8299 		fprintf(outf, "CPUID(1): family:model:stepping 0x%x:%x:%x (%d:%d:%d)",
8300 			family, model, stepping, family, model, stepping);
8301 		if (ucode_patch_valid)
8302 			fprintf(outf, " microcode 0x%x", (unsigned int)((ucode_patch >> 32) & 0xFFFFFFFF));
8303 		fputc('\n', outf);
8304 
8305 		fprintf(outf, "CPUID(0x80000000): max_extended_levels: 0x%x\n", max_extended_level);
8306 		fprintf(outf, "CPUID(1): %s %s %s %s %s %s %s %s %s %s\n",
8307 			ecx_flags & (1 << 0) ? "SSE3" : "-",
8308 			ecx_flags & (1 << 3) ? "MONITOR" : "-",
8309 			ecx_flags & (1 << 6) ? "SMX" : "-",
8310 			ecx_flags & (1 << 7) ? "EIST" : "-",
8311 			ecx_flags & (1 << 8) ? "TM2" : "-",
8312 			edx_flags & (1 << 4) ? "TSC" : "-",
8313 			edx_flags & (1 << 5) ? "MSR" : "-",
8314 			edx_flags & (1 << 22) ? "ACPI-TM" : "-",
8315 			edx_flags & (1 << 28) ? "HT" : "-", edx_flags & (1 << 29) ? "TM" : "-");
8316 	}
8317 
8318 	probe_platform_features(family, model);
8319 
8320 	if (!(edx_flags & (1 << 5)))
8321 		errx(1, "CPUID: no MSR");
8322 
8323 	if (max_extended_level >= 0x80000007) {
8324 
8325 		/*
8326 		 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
8327 		 * this check is valid for both Intel and AMD
8328 		 */
8329 		__cpuid(0x80000007, eax, ebx, ecx, edx);
8330 		has_invariant_tsc = edx & (1 << 8);
8331 	}
8332 
8333 	/*
8334 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
8335 	 * this check is valid for both Intel and AMD
8336 	 */
8337 
8338 	__cpuid(0x6, eax, ebx, ecx, edx);
8339 	has_aperf = ecx & (1 << 0);
8340 	do_dts = eax & (1 << 0);
8341 	if (do_dts)
8342 		BIC_PRESENT(BIC_CoreTmp);
8343 	has_turbo = eax & (1 << 1);
8344 	do_ptm = eax & (1 << 6);
8345 	if (do_ptm)
8346 		BIC_PRESENT(BIC_PkgTmp);
8347 	has_hwp = eax & (1 << 7);
8348 	has_hwp_notify = eax & (1 << 8);
8349 	has_hwp_activity_window = eax & (1 << 9);
8350 	has_hwp_epp = eax & (1 << 10);
8351 	has_hwp_pkg = eax & (1 << 11);
8352 	has_epb = ecx & (1 << 3);
8353 
8354 	if (!quiet)
8355 		fprintf(outf, "CPUID(6): %sAPERF, %sTURBO, %sDTS, %sPTM, %sHWP, "
8356 			"%sHWPnotify, %sHWPwindow, %sHWPepp, %sHWPpkg, %sEPB\n",
8357 			has_aperf ? "" : "No-",
8358 			has_turbo ? "" : "No-",
8359 			do_dts ? "" : "No-",
8360 			do_ptm ? "" : "No-",
8361 			has_hwp ? "" : "No-",
8362 			has_hwp_notify ? "" : "No-",
8363 			has_hwp_activity_window ? "" : "No-",
8364 			has_hwp_epp ? "" : "No-", has_hwp_pkg ? "" : "No-", has_epb ? "" : "No-");
8365 
8366 	if (!quiet)
8367 		decode_misc_enable_msr();
8368 
8369 	if (max_level >= 0x7 && !quiet) {
8370 		int has_sgx;
8371 
8372 		ecx = 0;
8373 
8374 		__cpuid_count(0x7, 0, eax, ebx, ecx, edx);
8375 
8376 		has_sgx = ebx & (1 << 2);
8377 
8378 		is_hybrid = edx & (1 << 15);
8379 
8380 		fprintf(outf, "CPUID(7): %sSGX %sHybrid\n", has_sgx ? "" : "No-", is_hybrid ? "" : "No-");
8381 
8382 		if (has_sgx)
8383 			decode_feature_control_msr();
8384 	}
8385 
8386 	if (max_level >= 0x15) {
8387 		unsigned int eax_crystal;
8388 		unsigned int ebx_tsc;
8389 
8390 		/*
8391 		 * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
8392 		 */
8393 		eax_crystal = ebx_tsc = crystal_hz = edx = 0;
8394 		__cpuid(0x15, eax_crystal, ebx_tsc, crystal_hz, edx);
8395 
8396 		if (ebx_tsc != 0) {
8397 			if (!quiet && (ebx != 0))
8398 				fprintf(outf, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n",
8399 					eax_crystal, ebx_tsc, crystal_hz);
8400 
8401 			if (crystal_hz == 0)
8402 				crystal_hz = platform->crystal_freq;
8403 
8404 			if (crystal_hz) {
8405 				tsc_hz = (unsigned long long)crystal_hz *ebx_tsc / eax_crystal;
8406 				if (!quiet)
8407 					fprintf(outf, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n",
8408 						tsc_hz / 1000000, crystal_hz, ebx_tsc, eax_crystal);
8409 			}
8410 		}
8411 	}
8412 	if (max_level >= 0x16) {
8413 		unsigned int base_mhz, max_mhz, bus_mhz, edx;
8414 
8415 		/*
8416 		 * CPUID 16H Base MHz, Max MHz, Bus MHz
8417 		 */
8418 		base_mhz = max_mhz = bus_mhz = edx = 0;
8419 
8420 		__cpuid(0x16, base_mhz, max_mhz, bus_mhz, edx);
8421 
8422 		bclk = bus_mhz;
8423 
8424 		base_hz = base_mhz * 1000000;
8425 		has_base_hz = 1;
8426 
8427 		if (platform->enable_tsc_tweak)
8428 			tsc_tweak = base_hz / tsc_hz;
8429 
8430 		if (!quiet)
8431 			fprintf(outf, "CPUID(0x16): base_mhz: %d max_mhz: %d bus_mhz: %d\n",
8432 				base_mhz, max_mhz, bus_mhz);
8433 	}
8434 
8435 	if (has_aperf)
8436 		aperf_mperf_multiplier = platform->need_perf_multiplier ? 1024 : 1;
8437 
8438 	BIC_PRESENT(BIC_IRQ);
8439 	BIC_PRESENT(BIC_NMI);
8440 	BIC_PRESENT(BIC_TSC_MHz);
8441 }
8442 
counter_info_init(void)8443 static void counter_info_init(void)
8444 {
8445 	for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i) {
8446 		struct cstate_counter_arch_info *const cai = &ccstate_counter_arch_infos[i];
8447 
8448 		if (platform->has_msr_knl_core_c6_residency && cai->msr == MSR_CORE_C6_RESIDENCY)
8449 			cai->msr = MSR_KNL_CORE_C6_RESIDENCY;
8450 
8451 		if (!platform->has_msr_core_c1_res && cai->msr == MSR_CORE_C1_RES)
8452 			cai->msr = 0;
8453 
8454 		if (platform->has_msr_atom_pkg_c6_residency && cai->msr == MSR_PKG_C6_RESIDENCY)
8455 			cai->msr = MSR_ATOM_PKG_C6_RESIDENCY;
8456 	}
8457 
8458 	for (int i = 0; i < NUM_MSR_COUNTERS; ++i) {
8459 		msr_counter_arch_infos[i].present = false;
8460 		msr_counter_arch_infos[i].needed = false;
8461 	}
8462 }
8463 
probe_pm_features(void)8464 void probe_pm_features(void)
8465 {
8466 	probe_pstates();
8467 
8468 	probe_cstates();
8469 
8470 	probe_lpi();
8471 
8472 	probe_intel_uncore_frequency();
8473 
8474 	probe_graphics();
8475 
8476 	probe_rapl();
8477 
8478 	probe_thermal();
8479 
8480 	if (platform->has_nhm_msrs && !no_msr)
8481 		BIC_PRESENT(BIC_SMI);
8482 
8483 	if (!quiet)
8484 		decode_misc_feature_control();
8485 }
8486 
8487 /*
8488  * in /dev/cpu/ return success for names that are numbers
8489  * ie. filter out ".", "..", "microcode".
8490  */
dir_filter(const struct dirent * dirp)8491 int dir_filter(const struct dirent *dirp)
8492 {
8493 	if (isdigit(dirp->d_name[0]))
8494 		return 1;
8495 	else
8496 		return 0;
8497 }
8498 
8499 char *possible_file = "/sys/devices/system/cpu/possible";
8500 char possible_buf[1024];
8501 
initialize_cpu_possible_set(void)8502 int initialize_cpu_possible_set(void)
8503 {
8504 	FILE *fp;
8505 
8506 	fp = fopen(possible_file, "r");
8507 	if (!fp) {
8508 		warn("open %s", possible_file);
8509 		return -1;
8510 	}
8511 	if (fread(possible_buf, sizeof(char), 1024, fp) == 0) {
8512 		warn("read %s", possible_file);
8513 		goto err;
8514 	}
8515 	if (parse_cpu_str(possible_buf, cpu_possible_set, cpu_possible_setsize)) {
8516 		warnx("%s: cpu str malformat %s\n", possible_file, cpu_effective_str);
8517 		goto err;
8518 	}
8519 	return 0;
8520 
8521 err:
8522 	fclose(fp);
8523 	return -1;
8524 }
8525 
topology_probe(bool startup)8526 void topology_probe(bool startup)
8527 {
8528 	int i;
8529 	int max_core_id = 0;
8530 	int max_package_id = 0;
8531 	int max_siblings = 0;
8532 
8533 	/* Initialize num_cpus, max_cpu_num */
8534 	set_max_cpu_num();
8535 	topo.num_cpus = 0;
8536 	for_all_proc_cpus(count_cpus);
8537 	if (!summary_only)
8538 		BIC_PRESENT(BIC_CPU);
8539 
8540 	if (debug > 1)
8541 		fprintf(outf, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
8542 
8543 	cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
8544 	if (cpus == NULL)
8545 		err(1, "calloc cpus");
8546 
8547 	/*
8548 	 * Allocate and initialize cpu_present_set
8549 	 */
8550 	cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
8551 	if (cpu_present_set == NULL)
8552 		err(3, "CPU_ALLOC");
8553 	cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8554 	CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
8555 	for_all_proc_cpus(mark_cpu_present);
8556 
8557 	/*
8558 	 * Allocate and initialize cpu_possible_set
8559 	 */
8560 	cpu_possible_set = CPU_ALLOC((topo.max_cpu_num + 1));
8561 	if (cpu_possible_set == NULL)
8562 		err(3, "CPU_ALLOC");
8563 	cpu_possible_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8564 	CPU_ZERO_S(cpu_possible_setsize, cpu_possible_set);
8565 	initialize_cpu_possible_set();
8566 
8567 	/*
8568 	 * Allocate and initialize cpu_effective_set
8569 	 */
8570 	cpu_effective_set = CPU_ALLOC((topo.max_cpu_num + 1));
8571 	if (cpu_effective_set == NULL)
8572 		err(3, "CPU_ALLOC");
8573 	cpu_effective_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8574 	CPU_ZERO_S(cpu_effective_setsize, cpu_effective_set);
8575 	update_effective_set(startup);
8576 
8577 	/*
8578 	 * Allocate and initialize cpu_allowed_set
8579 	 */
8580 	cpu_allowed_set = CPU_ALLOC((topo.max_cpu_num + 1));
8581 	if (cpu_allowed_set == NULL)
8582 		err(3, "CPU_ALLOC");
8583 	cpu_allowed_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8584 	CPU_ZERO_S(cpu_allowed_setsize, cpu_allowed_set);
8585 
8586 	/*
8587 	 * Validate and update cpu_allowed_set.
8588 	 *
8589 	 * Make sure all cpus in cpu_subset are also in cpu_present_set during startup.
8590 	 * Give a warning when cpus in cpu_subset become unavailable at runtime.
8591 	 * Give a warning when cpus are not effective because of cgroup setting.
8592 	 *
8593 	 * cpu_allowed_set is the intersection of cpu_present_set/cpu_effective_set/cpu_subset.
8594 	 */
8595 	for (i = 0; i < CPU_SUBSET_MAXCPUS; ++i) {
8596 		if (cpu_subset && !CPU_ISSET_S(i, cpu_subset_size, cpu_subset))
8597 			continue;
8598 
8599 		if (!CPU_ISSET_S(i, cpu_present_setsize, cpu_present_set)) {
8600 			if (cpu_subset) {
8601 				/* cpus in cpu_subset must be in cpu_present_set during startup */
8602 				if (startup)
8603 					err(1, "cpu%d not present", i);
8604 				else
8605 					fprintf(stderr, "cpu%d not present\n", i);
8606 			}
8607 			continue;
8608 		}
8609 
8610 		if (CPU_COUNT_S(cpu_effective_setsize, cpu_effective_set)) {
8611 			if (!CPU_ISSET_S(i, cpu_effective_setsize, cpu_effective_set)) {
8612 				fprintf(stderr, "cpu%d not effective\n", i);
8613 				continue;
8614 			}
8615 		}
8616 
8617 		CPU_SET_S(i, cpu_allowed_setsize, cpu_allowed_set);
8618 	}
8619 
8620 	if (!CPU_COUNT_S(cpu_allowed_setsize, cpu_allowed_set))
8621 		err(-ENODEV, "No valid cpus found");
8622 	sched_setaffinity(0, cpu_allowed_setsize, cpu_allowed_set);
8623 
8624 	/*
8625 	 * Allocate and initialize cpu_affinity_set
8626 	 */
8627 	cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
8628 	if (cpu_affinity_set == NULL)
8629 		err(3, "CPU_ALLOC");
8630 	cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8631 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
8632 
8633 	for_all_proc_cpus(init_thread_id);
8634 
8635 	for_all_proc_cpus(set_cpu_hybrid_type);
8636 
8637 	/*
8638 	 * For online cpus
8639 	 * find max_core_id, max_package_id
8640 	 */
8641 	for (i = 0; i <= topo.max_cpu_num; ++i) {
8642 		int siblings;
8643 
8644 		if (cpu_is_not_present(i)) {
8645 			if (debug > 1)
8646 				fprintf(outf, "cpu%d NOT PRESENT\n", i);
8647 			continue;
8648 		}
8649 
8650 		cpus[i].logical_cpu_id = i;
8651 
8652 		/* get package information */
8653 		cpus[i].physical_package_id = get_physical_package_id(i);
8654 		if (cpus[i].physical_package_id > max_package_id)
8655 			max_package_id = cpus[i].physical_package_id;
8656 
8657 		/* get die information */
8658 		cpus[i].die_id = get_die_id(i);
8659 		if (cpus[i].die_id > topo.max_die_id)
8660 			topo.max_die_id = cpus[i].die_id;
8661 
8662 		/* get numa node information */
8663 		cpus[i].physical_node_id = get_physical_node_id(&cpus[i]);
8664 		if (cpus[i].physical_node_id > topo.max_node_num)
8665 			topo.max_node_num = cpus[i].physical_node_id;
8666 
8667 		/* get core information */
8668 		cpus[i].physical_core_id = get_core_id(i);
8669 		if (cpus[i].physical_core_id > max_core_id)
8670 			max_core_id = cpus[i].physical_core_id;
8671 
8672 		/* get thread information */
8673 		siblings = get_thread_siblings(&cpus[i]);
8674 		if (siblings > max_siblings)
8675 			max_siblings = siblings;
8676 		if (cpus[i].thread_id == 0)
8677 			topo.num_cores++;
8678 	}
8679 	topo.max_core_id = max_core_id;
8680 	topo.max_package_id = max_package_id;
8681 
8682 	topo.cores_per_node = max_core_id + 1;
8683 	if (debug > 1)
8684 		fprintf(outf, "max_core_id %d, sizing for %d cores per package\n", max_core_id, topo.cores_per_node);
8685 	if (!summary_only)
8686 		BIC_PRESENT(BIC_Core);
8687 
8688 	topo.num_die = topo.max_die_id + 1;
8689 	if (debug > 1)
8690 		fprintf(outf, "max_die_id %d, sizing for %d die\n", topo.max_die_id, topo.num_die);
8691 	if (!summary_only && topo.num_die > 1)
8692 		BIC_PRESENT(BIC_Die);
8693 
8694 	topo.num_packages = max_package_id + 1;
8695 	if (debug > 1)
8696 		fprintf(outf, "max_package_id %d, sizing for %d packages\n", max_package_id, topo.num_packages);
8697 	if (!summary_only && topo.num_packages > 1)
8698 		BIC_PRESENT(BIC_Package);
8699 
8700 	set_node_data();
8701 	if (debug > 1)
8702 		fprintf(outf, "nodes_per_pkg %d\n", topo.nodes_per_pkg);
8703 	if (!summary_only && topo.nodes_per_pkg > 1)
8704 		BIC_PRESENT(BIC_Node);
8705 
8706 	topo.threads_per_core = max_siblings;
8707 	if (debug > 1)
8708 		fprintf(outf, "max_siblings %d\n", max_siblings);
8709 
8710 	if (debug < 1)
8711 		return;
8712 
8713 	for (i = 0; i <= topo.max_cpu_num; ++i) {
8714 		if (cpu_is_not_present(i))
8715 			continue;
8716 		fprintf(outf,
8717 			"cpu %d pkg %d die %d node %d lnode %d core %d thread %d\n",
8718 			i, cpus[i].physical_package_id, cpus[i].die_id,
8719 			cpus[i].physical_node_id, cpus[i].logical_node_id, cpus[i].physical_core_id, cpus[i].thread_id);
8720 	}
8721 
8722 }
8723 
allocate_counters(struct thread_data ** t,struct core_data ** c,struct pkg_data ** p)8724 void allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
8725 {
8726 	int i;
8727 	int num_cores = topo.cores_per_node * topo.nodes_per_pkg * topo.num_packages;
8728 	int num_threads = topo.threads_per_core * num_cores;
8729 
8730 	*t = calloc(num_threads, sizeof(struct thread_data));
8731 	if (*t == NULL)
8732 		goto error;
8733 
8734 	for (i = 0; i < num_threads; i++)
8735 		(*t)[i].cpu_id = -1;
8736 
8737 	*c = calloc(num_cores, sizeof(struct core_data));
8738 	if (*c == NULL)
8739 		goto error;
8740 
8741 	for (i = 0; i < num_cores; i++) {
8742 		(*c)[i].core_id = -1;
8743 		(*c)[i].base_cpu = -1;
8744 	}
8745 
8746 	*p = calloc(topo.num_packages, sizeof(struct pkg_data));
8747 	if (*p == NULL)
8748 		goto error;
8749 
8750 	for (i = 0; i < topo.num_packages; i++) {
8751 		(*p)[i].package_id = i;
8752 		(*p)[i].base_cpu = -1;
8753 	}
8754 
8755 	return;
8756 error:
8757 	err(1, "calloc counters");
8758 }
8759 
8760 /*
8761  * init_counter()
8762  *
8763  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
8764  */
init_counter(struct thread_data * thread_base,struct core_data * core_base,struct pkg_data * pkg_base,int cpu_id)8765 void init_counter(struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base, int cpu_id)
8766 {
8767 	int pkg_id = cpus[cpu_id].physical_package_id;
8768 	int node_id = cpus[cpu_id].logical_node_id;
8769 	int core_id = cpus[cpu_id].physical_core_id;
8770 	int thread_id = cpus[cpu_id].thread_id;
8771 	struct thread_data *t;
8772 	struct core_data *c;
8773 	struct pkg_data *p;
8774 
8775 	/* Workaround for systems where physical_node_id==-1
8776 	 * and logical_node_id==(-1 - topo.num_cpus)
8777 	 */
8778 	if (node_id < 0)
8779 		node_id = 0;
8780 
8781 	t = GET_THREAD(thread_base, thread_id, core_id, node_id, pkg_id);
8782 	c = GET_CORE(core_base, core_id, node_id, pkg_id);
8783 	p = GET_PKG(pkg_base, pkg_id);
8784 
8785 	t->cpu_id = cpu_id;
8786 	if (!cpu_is_not_allowed(cpu_id)) {
8787 		if (c->base_cpu < 0)
8788 			c->base_cpu = t->cpu_id;
8789 		if (p->base_cpu < 0)
8790 			p->base_cpu = t->cpu_id;
8791 	}
8792 
8793 	c->core_id = core_id;
8794 	p->package_id = pkg_id;
8795 }
8796 
initialize_counters(int cpu_id)8797 int initialize_counters(int cpu_id)
8798 {
8799 	init_counter(EVEN_COUNTERS, cpu_id);
8800 	init_counter(ODD_COUNTERS, cpu_id);
8801 	return 0;
8802 }
8803 
allocate_output_buffer()8804 void allocate_output_buffer()
8805 {
8806 	output_buffer = calloc(1, (1 + topo.num_cpus) * 2048);
8807 	outp = output_buffer;
8808 	if (outp == NULL)
8809 		err(-1, "calloc output buffer");
8810 }
8811 
allocate_fd_percpu(void)8812 void allocate_fd_percpu(void)
8813 {
8814 	fd_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8815 	if (fd_percpu == NULL)
8816 		err(-1, "calloc fd_percpu");
8817 }
8818 
allocate_irq_buffers(void)8819 void allocate_irq_buffers(void)
8820 {
8821 	irq_column_2_cpu = calloc(topo.num_cpus, sizeof(int));
8822 	if (irq_column_2_cpu == NULL)
8823 		err(-1, "calloc %d", topo.num_cpus);
8824 
8825 	irqs_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8826 	if (irqs_per_cpu == NULL)
8827 		err(-1, "calloc %d IRQ", topo.max_cpu_num + 1);
8828 
8829 	nmi_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8830 	if (nmi_per_cpu == NULL)
8831 		err(-1, "calloc %d NMI", topo.max_cpu_num + 1);
8832 }
8833 
update_topo(struct thread_data * t,struct core_data * c,struct pkg_data * p)8834 int update_topo(struct thread_data *t, struct core_data *c, struct pkg_data *p)
8835 {
8836 	topo.allowed_cpus++;
8837 	if ((int)t->cpu_id == c->base_cpu)
8838 		topo.allowed_cores++;
8839 	if ((int)t->cpu_id == p->base_cpu)
8840 		topo.allowed_packages++;
8841 
8842 	return 0;
8843 }
8844 
topology_update(void)8845 void topology_update(void)
8846 {
8847 	topo.allowed_cpus = 0;
8848 	topo.allowed_cores = 0;
8849 	topo.allowed_packages = 0;
8850 	for_all_cpus(update_topo, ODD_COUNTERS);
8851 }
8852 
setup_all_buffers(bool startup)8853 void setup_all_buffers(bool startup)
8854 {
8855 	topology_probe(startup);
8856 	allocate_irq_buffers();
8857 	allocate_fd_percpu();
8858 	allocate_counters(&thread_even, &core_even, &package_even);
8859 	allocate_counters(&thread_odd, &core_odd, &package_odd);
8860 	allocate_output_buffer();
8861 	for_all_proc_cpus(initialize_counters);
8862 	topology_update();
8863 }
8864 
set_base_cpu(void)8865 void set_base_cpu(void)
8866 {
8867 	int i;
8868 
8869 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
8870 		if (cpu_is_not_allowed(i))
8871 			continue;
8872 		base_cpu = i;
8873 		if (debug > 1)
8874 			fprintf(outf, "base_cpu = %d\n", base_cpu);
8875 		return;
8876 	}
8877 	err(-ENODEV, "No valid cpus found");
8878 }
8879 
has_added_counters(void)8880 bool has_added_counters(void)
8881 {
8882 	/*
8883 	 * It only makes sense to call this after the command line is parsed,
8884 	 * otherwise sys structure is not populated.
8885 	 */
8886 
8887 	return sys.added_core_counters | sys.added_thread_counters | sys.added_package_counters;
8888 }
8889 
check_msr_access(void)8890 void check_msr_access(void)
8891 {
8892 	check_dev_msr();
8893 	check_msr_permission();
8894 
8895 	if (no_msr)
8896 		bic_disable_msr_access();
8897 }
8898 
check_perf_access(void)8899 void check_perf_access(void)
8900 {
8901 	if (no_perf || !BIC_IS_ENABLED(BIC_IPC) || !has_instr_count_access())
8902 		bic_enabled &= ~BIC_IPC;
8903 }
8904 
perf_has_hybrid_devices(void)8905 bool perf_has_hybrid_devices(void)
8906 {
8907 	/*
8908 	 *  0: unknown
8909 	 *  1: has separate perf device for p and e core
8910 	 * -1: doesn't have separate perf device for p and e core
8911 	 */
8912 	static int cached;
8913 
8914 	if (cached > 0)
8915 		return true;
8916 
8917 	if (cached < 0)
8918 		return false;
8919 
8920 	if (access("/sys/bus/event_source/devices/cpu_core", F_OK)) {
8921 		cached = -1;
8922 		return false;
8923 	}
8924 
8925 	if (access("/sys/bus/event_source/devices/cpu_atom", F_OK)) {
8926 		cached = -1;
8927 		return false;
8928 	}
8929 
8930 	cached = 1;
8931 	return true;
8932 }
8933 
added_perf_counters_init_(struct perf_counter_info * pinfo)8934 int added_perf_counters_init_(struct perf_counter_info *pinfo)
8935 {
8936 	size_t num_domains = 0;
8937 	unsigned int next_domain;
8938 	bool *domain_visited;
8939 	unsigned int perf_type, perf_config;
8940 	double perf_scale;
8941 	int fd_perf;
8942 
8943 	if (!pinfo)
8944 		return 0;
8945 
8946 	const size_t max_num_domains = MAX(topo.max_cpu_num + 1, MAX(topo.max_core_id + 1, topo.max_package_id + 1));
8947 
8948 	domain_visited = calloc(max_num_domains, sizeof(*domain_visited));
8949 
8950 	while (pinfo) {
8951 		switch (pinfo->scope) {
8952 		case SCOPE_CPU:
8953 			num_domains = topo.max_cpu_num + 1;
8954 			break;
8955 
8956 		case SCOPE_CORE:
8957 			num_domains = topo.max_core_id + 1;
8958 			break;
8959 
8960 		case SCOPE_PACKAGE:
8961 			num_domains = topo.max_package_id + 1;
8962 			break;
8963 		}
8964 
8965 		/* Allocate buffer for file descriptor for each domain. */
8966 		pinfo->fd_perf_per_domain = calloc(num_domains, sizeof(*pinfo->fd_perf_per_domain));
8967 		if (!pinfo->fd_perf_per_domain)
8968 			errx(1, "%s: alloc %s", __func__, "fd_perf_per_domain");
8969 
8970 		for (size_t i = 0; i < num_domains; ++i)
8971 			pinfo->fd_perf_per_domain[i] = -1;
8972 
8973 		pinfo->num_domains = num_domains;
8974 		pinfo->scale = 1.0;
8975 
8976 		memset(domain_visited, 0, max_num_domains * sizeof(*domain_visited));
8977 
8978 		for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
8979 
8980 			next_domain = cpu_to_domain(pinfo, cpu);
8981 
8982 			assert(next_domain < num_domains);
8983 
8984 			if (cpu_is_not_allowed(cpu))
8985 				continue;
8986 
8987 			if (domain_visited[next_domain])
8988 				continue;
8989 
8990 			/*
8991 			 * Intel hybrid platforms expose different perf devices for P and E cores.
8992 			 * Instead of one, "/sys/bus/event_source/devices/cpu" device, there are
8993 			 * "/sys/bus/event_source/devices/{cpu_core,cpu_atom}".
8994 			 *
8995 			 * This makes it more complicated to the user, because most of the counters
8996 			 * are available on both and have to be handled manually, otherwise.
8997 			 *
8998 			 * Code below, allow user to use the old "cpu" name, which is translated accordingly.
8999 			 */
9000 			const char *perf_device = pinfo->device;
9001 
9002 			if (strcmp(perf_device, "cpu") == 0 && perf_has_hybrid_devices()) {
9003 				switch (cpus[cpu].type) {
9004 				case INTEL_PCORE_TYPE:
9005 					perf_device = "cpu_core";
9006 					break;
9007 
9008 				case INTEL_ECORE_TYPE:
9009 					perf_device = "cpu_atom";
9010 					break;
9011 
9012 				default: /* Don't change, we will probably fail and report a problem soon. */
9013 					break;
9014 				}
9015 			}
9016 
9017 			perf_type = read_perf_type(perf_device);
9018 			if (perf_type == (unsigned int)-1) {
9019 				warnx("%s: perf/%s/%s: failed to read %s",
9020 				      __func__, perf_device, pinfo->event, "type");
9021 				continue;
9022 			}
9023 
9024 			perf_config = read_perf_config(perf_device, pinfo->event);
9025 			if (perf_config == (unsigned int)-1) {
9026 				warnx("%s: perf/%s/%s: failed to read %s",
9027 				      __func__, perf_device, pinfo->event, "config");
9028 				continue;
9029 			}
9030 
9031 			/* Scale is not required, some counters just don't have it. */
9032 			perf_scale = read_perf_scale(perf_device, pinfo->event);
9033 			if (perf_scale == 0.0)
9034 				perf_scale = 1.0;
9035 
9036 			fd_perf = open_perf_counter(cpu, perf_type, perf_config, -1, 0);
9037 			if (fd_perf == -1) {
9038 				warnx("%s: perf/%s/%s: failed to open counter on cpu%d",
9039 				      __func__, perf_device, pinfo->event, cpu);
9040 				continue;
9041 			}
9042 
9043 			domain_visited[next_domain] = 1;
9044 			pinfo->fd_perf_per_domain[next_domain] = fd_perf;
9045 			pinfo->scale = perf_scale;
9046 
9047 			if (debug)
9048 				fprintf(stderr, "Add perf/%s/%s cpu%d: %d\n",
9049 					perf_device, pinfo->event, cpu, pinfo->fd_perf_per_domain[next_domain]);
9050 		}
9051 
9052 		pinfo = pinfo->next;
9053 	}
9054 
9055 	free(domain_visited);
9056 
9057 	return 0;
9058 }
9059 
added_perf_counters_init(void)9060 void added_perf_counters_init(void)
9061 {
9062 	if (added_perf_counters_init_(sys.perf_tp))
9063 		errx(1, "%s: %s", __func__, "thread");
9064 
9065 	if (added_perf_counters_init_(sys.perf_cp))
9066 		errx(1, "%s: %s", __func__, "core");
9067 
9068 	if (added_perf_counters_init_(sys.perf_pp))
9069 		errx(1, "%s: %s", __func__, "package");
9070 }
9071 
parse_telem_info_file(int fd_dir,const char * info_filename,const char * format,unsigned long * output)9072 int parse_telem_info_file(int fd_dir, const char *info_filename, const char *format, unsigned long *output)
9073 {
9074 	int fd_telem_info;
9075 	FILE *file_telem_info;
9076 	unsigned long value;
9077 
9078 	fd_telem_info = openat(fd_dir, info_filename, O_RDONLY);
9079 	if (fd_telem_info == -1)
9080 		return -1;
9081 
9082 	file_telem_info = fdopen(fd_telem_info, "r");
9083 	if (file_telem_info == NULL) {
9084 		close(fd_telem_info);
9085 		return -1;
9086 	}
9087 
9088 	if (fscanf(file_telem_info, format, &value) != 1) {
9089 		fclose(file_telem_info);
9090 		return -1;
9091 	}
9092 
9093 	fclose(file_telem_info);
9094 
9095 	*output = value;
9096 
9097 	return 0;
9098 }
9099 
pmt_mmio_open(unsigned int target_guid)9100 struct pmt_mmio *pmt_mmio_open(unsigned int target_guid)
9101 {
9102 	struct pmt_diriter_t pmt_iter;
9103 	const struct dirent *entry;
9104 	struct stat st;
9105 	int fd_telem_dir, fd_pmt;
9106 	unsigned long guid, size, offset;
9107 	size_t mmap_size;
9108 	void *mmio;
9109 	struct pmt_mmio *head = NULL, *last = NULL;
9110 	struct pmt_mmio *new_pmt = NULL;
9111 
9112 	if (stat(SYSFS_TELEM_PATH, &st) == -1)
9113 		return NULL;
9114 
9115 	pmt_diriter_init(&pmt_iter);
9116 	entry = pmt_diriter_begin(&pmt_iter, SYSFS_TELEM_PATH);
9117 	if (!entry) {
9118 		pmt_diriter_remove(&pmt_iter);
9119 		return NULL;
9120 	}
9121 
9122 	for ( ; entry != NULL; entry = pmt_diriter_next(&pmt_iter)) {
9123 		if (fstatat(dirfd(pmt_iter.dir), entry->d_name, &st, 0) == -1)
9124 			break;
9125 
9126 		if (!S_ISDIR(st.st_mode))
9127 			continue;
9128 
9129 		fd_telem_dir = openat(dirfd(pmt_iter.dir), entry->d_name, O_RDONLY);
9130 		if (fd_telem_dir == -1)
9131 			break;
9132 
9133 		if (parse_telem_info_file(fd_telem_dir, "guid", "%lx", &guid)) {
9134 			close(fd_telem_dir);
9135 			break;
9136 		}
9137 
9138 		if (parse_telem_info_file(fd_telem_dir, "size", "%lu", &size)) {
9139 			close(fd_telem_dir);
9140 			break;
9141 		}
9142 
9143 		if (guid != target_guid) {
9144 			close(fd_telem_dir);
9145 			continue;
9146 		}
9147 
9148 		if (parse_telem_info_file(fd_telem_dir, "offset", "%lu", &offset)) {
9149 			close(fd_telem_dir);
9150 			break;
9151 		}
9152 
9153 		assert(offset == 0);
9154 
9155 		fd_pmt = openat(fd_telem_dir, "telem", O_RDONLY);
9156 		if (fd_pmt == -1)
9157 			goto loop_cleanup_and_break;
9158 
9159 		mmap_size = ROUND_UP_TO_PAGE_SIZE(size);
9160 		mmio = mmap(0, mmap_size, PROT_READ, MAP_SHARED, fd_pmt, 0);
9161 		if (mmio != MAP_FAILED) {
9162 			if (debug)
9163 				fprintf(stderr, "%s: 0x%lx mmaped at: %p\n", __func__, guid, mmio);
9164 
9165 			new_pmt = calloc(1, sizeof(*new_pmt));
9166 
9167 			if (!new_pmt) {
9168 				fprintf(stderr, "%s: Failed to allocate pmt_mmio\n", __func__);
9169 				exit(1);
9170 			}
9171 
9172 			/*
9173 			 * Create linked list of mmaped regions,
9174 			 * but preserve the ordering from sysfs.
9175 			 * Ordering is important for the user to
9176 			 * use the seq=%u parameter when adding a counter.
9177 			 */
9178 			new_pmt->guid = guid;
9179 			new_pmt->mmio_base = mmio;
9180 			new_pmt->pmt_offset = offset;
9181 			new_pmt->size = size;
9182 			new_pmt->next = pmt_mmios;
9183 
9184 			if (last)
9185 				last->next = new_pmt;
9186 			else
9187 				head = new_pmt;
9188 
9189 			last = new_pmt;
9190 		}
9191 
9192 loop_cleanup_and_break:
9193 		close(fd_pmt);
9194 		close(fd_telem_dir);
9195 	}
9196 
9197 	pmt_diriter_remove(&pmt_iter);
9198 
9199 	/*
9200 	 * If we found something, stick just
9201 	 * created linked list to the front.
9202 	 */
9203 	if (head)
9204 		pmt_mmios = head;
9205 
9206 	return head;
9207 }
9208 
pmt_mmio_find(unsigned int guid)9209 struct pmt_mmio *pmt_mmio_find(unsigned int guid)
9210 {
9211 	struct pmt_mmio *pmmio = pmt_mmios;
9212 
9213 	while (pmmio) {
9214 		if (pmmio->guid == guid)
9215 			return pmmio;
9216 
9217 		pmmio = pmmio->next;
9218 	}
9219 
9220 	return NULL;
9221 }
9222 
pmt_get_counter_pointer(struct pmt_mmio * pmmio,unsigned long counter_offset)9223 void *pmt_get_counter_pointer(struct pmt_mmio *pmmio, unsigned long counter_offset)
9224 {
9225 	char *ret;
9226 
9227 	/* Get base of mmaped PMT file. */
9228 	ret = (char *)pmmio->mmio_base;
9229 
9230 	/*
9231 	 * Apply PMT MMIO offset to obtain beginning of the mmaped telemetry data.
9232 	 * It's not guaranteed that the mmaped memory begins with the telemetry data
9233 	 *      - we might have to apply the offset first.
9234 	 */
9235 	ret += pmmio->pmt_offset;
9236 
9237 	/* Apply the counter offset to get the address to the mmaped counter. */
9238 	ret += counter_offset;
9239 
9240 	return ret;
9241 }
9242 
pmt_add_guid(unsigned int guid,unsigned int seq)9243 struct pmt_mmio *pmt_add_guid(unsigned int guid, unsigned int seq)
9244 {
9245 	struct pmt_mmio *ret;
9246 
9247 	ret = pmt_mmio_find(guid);
9248 	if (!ret)
9249 		ret = pmt_mmio_open(guid);
9250 
9251 	while (ret && seq) {
9252 		ret = ret->next;
9253 		--seq;
9254 	}
9255 
9256 	return ret;
9257 }
9258 
9259 enum pmt_open_mode {
9260 	PMT_OPEN_TRY,		/* Open failure is not an error. */
9261 	PMT_OPEN_REQUIRED,	/* Open failure is a fatal error. */
9262 };
9263 
pmt_find_counter(struct pmt_counter * pcounter,const char * name)9264 struct pmt_counter *pmt_find_counter(struct pmt_counter *pcounter, const char *name)
9265 {
9266 	while (pcounter) {
9267 		if (strcmp(pcounter->name, name) == 0)
9268 			break;
9269 
9270 		pcounter = pcounter->next;
9271 	}
9272 
9273 	return pcounter;
9274 }
9275 
pmt_get_scope_root(enum counter_scope scope)9276 struct pmt_counter **pmt_get_scope_root(enum counter_scope scope)
9277 {
9278 	switch (scope) {
9279 	case SCOPE_CPU:
9280 		return &sys.pmt_tp;
9281 	case SCOPE_CORE:
9282 		return &sys.pmt_cp;
9283 	case SCOPE_PACKAGE:
9284 		return &sys.pmt_pp;
9285 	}
9286 
9287 	__builtin_unreachable();
9288 }
9289 
pmt_counter_add_domain(struct pmt_counter * pcounter,unsigned long * pmmio,unsigned int domain_id)9290 void pmt_counter_add_domain(struct pmt_counter *pcounter, unsigned long *pmmio, unsigned int domain_id)
9291 {
9292 	/* Make sure the new domain fits. */
9293 	if (domain_id >= pcounter->num_domains)
9294 		pmt_counter_resize(pcounter, domain_id + 1);
9295 
9296 	assert(pcounter->domains);
9297 	assert(domain_id < pcounter->num_domains);
9298 
9299 	pcounter->domains[domain_id].pcounter = pmmio;
9300 }
9301 
pmt_add_counter(unsigned int guid,unsigned int seq,const char * name,enum pmt_datatype type,unsigned int lsb,unsigned int msb,unsigned int offset,enum counter_scope scope,enum counter_format format,unsigned int domain_id,enum pmt_open_mode mode)9302 int pmt_add_counter(unsigned int guid, unsigned int seq, const char *name, enum pmt_datatype type,
9303 		    unsigned int lsb, unsigned int msb, unsigned int offset, enum counter_scope scope,
9304 		    enum counter_format format, unsigned int domain_id, enum pmt_open_mode mode)
9305 {
9306 	struct pmt_mmio *mmio;
9307 	struct pmt_counter *pcounter;
9308 	struct pmt_counter **const pmt_root = pmt_get_scope_root(scope);
9309 	bool new_counter = false;
9310 	int conflict = 0;
9311 
9312 	if (lsb > msb) {
9313 		fprintf(stderr, "%s: %s: `%s` must be satisfied\n", __func__, "lsb <= msb", name);
9314 		exit(1);
9315 	}
9316 
9317 	if (msb >= 64) {
9318 		fprintf(stderr, "%s: %s: `%s` must be satisfied\n", __func__, "msb < 64", name);
9319 		exit(1);
9320 	}
9321 
9322 	mmio = pmt_add_guid(guid, seq);
9323 	if (!mmio) {
9324 		if (mode != PMT_OPEN_TRY) {
9325 			fprintf(stderr, "%s: failed to map PMT MMIO for guid %x, seq %u\n", __func__, guid, seq);
9326 			exit(1);
9327 		}
9328 
9329 		return 1;
9330 	}
9331 
9332 	if (offset >= mmio->size) {
9333 		if (mode != PMT_OPEN_TRY) {
9334 			fprintf(stderr, "%s: offset %u outside of PMT MMIO size %u\n", __func__, offset, mmio->size);
9335 			exit(1);
9336 		}
9337 
9338 		return 1;
9339 	}
9340 
9341 	pcounter = pmt_find_counter(*pmt_root, name);
9342 	if (!pcounter) {
9343 		pcounter = calloc(1, sizeof(*pcounter));
9344 		new_counter = true;
9345 	}
9346 
9347 	if (new_counter) {
9348 		strncpy(pcounter->name, name, ARRAY_SIZE(pcounter->name) - 1);
9349 		pcounter->type = type;
9350 		pcounter->scope = scope;
9351 		pcounter->lsb = lsb;
9352 		pcounter->msb = msb;
9353 		pcounter->format = format;
9354 	} else {
9355 		conflict += pcounter->type != type;
9356 		conflict += pcounter->scope != scope;
9357 		conflict += pcounter->lsb != lsb;
9358 		conflict += pcounter->msb != msb;
9359 		conflict += pcounter->format != format;
9360 	}
9361 
9362 	if (conflict) {
9363 		fprintf(stderr, "%s: conflicting parameters for the PMT counter with the same name %s\n",
9364 			__func__, name);
9365 		exit(1);
9366 	}
9367 
9368 	pmt_counter_add_domain(pcounter, pmt_get_counter_pointer(mmio, offset), domain_id);
9369 
9370 	if (new_counter) {
9371 		pcounter->next = *pmt_root;
9372 		*pmt_root = pcounter;
9373 	}
9374 
9375 	return 0;
9376 }
9377 
pmt_init(void)9378 void pmt_init(void)
9379 {
9380 	int cpu_num;
9381 	unsigned long seq, offset, mod_num;
9382 
9383 	if (BIC_IS_ENABLED(BIC_Diec6)) {
9384 		pmt_add_counter(PMT_MTL_DC6_GUID, PMT_MTL_DC6_SEQ, "Die%c6", PMT_TYPE_XTAL_TIME,
9385 				PMT_COUNTER_MTL_DC6_LSB, PMT_COUNTER_MTL_DC6_MSB, PMT_COUNTER_MTL_DC6_OFFSET,
9386 				SCOPE_PACKAGE, FORMAT_DELTA, 0, PMT_OPEN_TRY);
9387 	}
9388 
9389 	if (BIC_IS_ENABLED(BIC_CPU_c1e)) {
9390 		seq = 0;
9391 		offset = PMT_COUNTER_CWF_MC1E_OFFSET_BASE;
9392 		mod_num = 0;	/* Relative module number for current PMT file. */
9393 
9394 		/* Open the counter for each CPU. */
9395 		for (cpu_num = 0; cpu_num < topo.max_cpu_num;) {
9396 
9397 			if (cpu_is_not_allowed(cpu_num))
9398 				goto next_loop_iter;
9399 
9400 			/*
9401 			 * Set the scope to CPU, even though CWF report the counter per module.
9402 			 * CPUs inside the same module will read from the same location, instead of reporting zeros.
9403 			 *
9404 			 * CWF with newer firmware might require a PMT_TYPE_XTAL_TIME intead of PMT_TYPE_TCORE_CLOCK.
9405 			 */
9406 			pmt_add_counter(PMT_CWF_MC1E_GUID, seq, "CPU%c1e", PMT_TYPE_TCORE_CLOCK,
9407 					PMT_COUNTER_CWF_MC1E_LSB, PMT_COUNTER_CWF_MC1E_MSB, offset, SCOPE_CPU,
9408 					FORMAT_DELTA, cpu_num, PMT_OPEN_TRY);
9409 
9410 			/*
9411 			 * Rather complex logic for each time we go to the next loop iteration,
9412 			 * so keep it as a label.
9413 			 */
9414 next_loop_iter:
9415 			/*
9416 			 * Advance the cpu number and check if we should also advance offset to
9417 			 * the next counter inside the PMT file.
9418 			 *
9419 			 * On Clearwater Forest platform, the counter is reported per module,
9420 			 * so open the same counter for all of the CPUs inside the module.
9421 			 * That way, reported table show the correct value for all of the CPUs inside the module,
9422 			 * instead of zeros.
9423 			 */
9424 			++cpu_num;
9425 			if (cpu_num % PMT_COUNTER_CWF_CPUS_PER_MODULE == 0) {
9426 				offset += PMT_COUNTER_CWF_MC1E_OFFSET_INCREMENT;
9427 				++mod_num;
9428 			}
9429 
9430 			/*
9431 			 * There are PMT_COUNTER_CWF_MC1E_NUM_MODULES_PER_FILE in each PMT file.
9432 			 *
9433 			 * If that number is reached, seq must be incremented to advance to the next file in a sequence.
9434 			 * Offset inside that file and a module counter has to be reset.
9435 			 */
9436 			if (mod_num == PMT_COUNTER_CWF_MC1E_NUM_MODULES_PER_FILE) {
9437 				++seq;
9438 				offset = PMT_COUNTER_CWF_MC1E_OFFSET_BASE;
9439 				mod_num = 0;
9440 			}
9441 		}
9442 	}
9443 }
9444 
turbostat_init()9445 void turbostat_init()
9446 {
9447 	setup_all_buffers(true);
9448 	set_base_cpu();
9449 	check_msr_access();
9450 	check_perf_access();
9451 	process_cpuid();
9452 	counter_info_init();
9453 	probe_pm_features();
9454 	msr_perf_init();
9455 	linux_perf_init();
9456 	rapl_perf_init();
9457 	cstate_perf_init();
9458 	added_perf_counters_init();
9459 	pmt_init();
9460 
9461 	for_all_cpus(get_cpu_type, ODD_COUNTERS);
9462 	for_all_cpus(get_cpu_type, EVEN_COUNTERS);
9463 
9464 	if (BIC_IS_ENABLED(BIC_IPC) && has_aperf_access && get_instr_count_fd(base_cpu) != -1)
9465 		BIC_PRESENT(BIC_IPC);
9466 
9467 	/*
9468 	 * If TSC tweak is needed, but couldn't get it,
9469 	 * disable more BICs, since it can't be reported accurately.
9470 	 */
9471 	if (platform->enable_tsc_tweak && !has_base_hz) {
9472 		bic_enabled &= ~BIC_Busy;
9473 		bic_enabled &= ~BIC_Bzy_MHz;
9474 	}
9475 }
9476 
affinitize_child(void)9477 void affinitize_child(void)
9478 {
9479 	/* Prefer cpu_possible_set, if available */
9480 	if (sched_setaffinity(0, cpu_possible_setsize, cpu_possible_set)) {
9481 		warn("sched_setaffinity cpu_possible_set");
9482 
9483 		/* Otherwise, allow child to run on same cpu set as turbostat */
9484 		if (sched_setaffinity(0, cpu_allowed_setsize, cpu_allowed_set))
9485 			warn("sched_setaffinity cpu_allowed_set");
9486 	}
9487 }
9488 
fork_it(char ** argv)9489 int fork_it(char **argv)
9490 {
9491 	pid_t child_pid;
9492 	int status;
9493 
9494 	snapshot_proc_sysfs_files();
9495 	status = for_all_cpus(get_counters, EVEN_COUNTERS);
9496 	first_counter_read = 0;
9497 	if (status)
9498 		exit(status);
9499 	gettimeofday(&tv_even, (struct timezone *)NULL);
9500 
9501 	child_pid = fork();
9502 	if (!child_pid) {
9503 		/* child */
9504 		affinitize_child();
9505 		execvp(argv[0], argv);
9506 		err(errno, "exec %s", argv[0]);
9507 	} else {
9508 
9509 		/* parent */
9510 		if (child_pid == -1)
9511 			err(1, "fork");
9512 
9513 		signal(SIGINT, SIG_IGN);
9514 		signal(SIGQUIT, SIG_IGN);
9515 		if (waitpid(child_pid, &status, 0) == -1)
9516 			err(status, "waitpid");
9517 
9518 		if (WIFEXITED(status))
9519 			status = WEXITSTATUS(status);
9520 	}
9521 	/*
9522 	 * n.b. fork_it() does not check for errors from for_all_cpus()
9523 	 * because re-starting is problematic when forking
9524 	 */
9525 	snapshot_proc_sysfs_files();
9526 	for_all_cpus(get_counters, ODD_COUNTERS);
9527 	gettimeofday(&tv_odd, (struct timezone *)NULL);
9528 	timersub(&tv_odd, &tv_even, &tv_delta);
9529 	if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS))
9530 		fprintf(outf, "%s: Counter reset detected\n", progname);
9531 
9532 	compute_average(EVEN_COUNTERS);
9533 	format_all_counters(EVEN_COUNTERS);
9534 
9535 	fprintf(outf, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec / 1000000.0);
9536 
9537 	flush_output_stderr();
9538 
9539 	return status;
9540 }
9541 
get_and_dump_counters(void)9542 int get_and_dump_counters(void)
9543 {
9544 	int status;
9545 
9546 	snapshot_proc_sysfs_files();
9547 	status = for_all_cpus(get_counters, ODD_COUNTERS);
9548 	if (status)
9549 		return status;
9550 
9551 	status = for_all_cpus(dump_counters, ODD_COUNTERS);
9552 	if (status)
9553 		return status;
9554 
9555 	flush_output_stdout();
9556 
9557 	return status;
9558 }
9559 
print_version()9560 void print_version()
9561 {
9562 	fprintf(outf, "turbostat version 2025.02.02 - Len Brown <lenb@kernel.org>\n");
9563 }
9564 
9565 #define COMMAND_LINE_SIZE 2048
9566 
print_bootcmd(void)9567 void print_bootcmd(void)
9568 {
9569 	char bootcmd[COMMAND_LINE_SIZE];
9570 	FILE *fp;
9571 	int ret;
9572 
9573 	memset(bootcmd, 0, COMMAND_LINE_SIZE);
9574 	fp = fopen("/proc/cmdline", "r");
9575 	if (!fp)
9576 		return;
9577 
9578 	ret = fread(bootcmd, sizeof(char), COMMAND_LINE_SIZE - 1, fp);
9579 	if (ret) {
9580 		bootcmd[ret] = '\0';
9581 		/* the last character is already '\n' */
9582 		fprintf(outf, "Kernel command line: %s", bootcmd);
9583 	}
9584 
9585 	fclose(fp);
9586 }
9587 
find_msrp_by_name(struct msr_counter * head,char * name)9588 struct msr_counter *find_msrp_by_name(struct msr_counter *head, char *name)
9589 {
9590 	struct msr_counter *mp;
9591 
9592 	for (mp = head; mp; mp = mp->next) {
9593 		if (debug)
9594 			fprintf(stderr, "%s: %s %s\n", __func__, name, mp->name);
9595 		if (!strncmp(name, mp->name, strlen(mp->name)))
9596 			return mp;
9597 	}
9598 	return NULL;
9599 }
9600 
add_counter(unsigned int msr_num,char * path,char * name,unsigned int width,enum counter_scope scope,enum counter_type type,enum counter_format format,int flags,int id)9601 int add_counter(unsigned int msr_num, char *path, char *name,
9602 		unsigned int width, enum counter_scope scope,
9603 		enum counter_type type, enum counter_format format, int flags, int id)
9604 {
9605 	struct msr_counter *msrp;
9606 
9607 	if (no_msr && msr_num)
9608 		errx(1, "Requested MSR counter 0x%x, but in --no-msr mode", msr_num);
9609 
9610 	if (debug)
9611 		fprintf(stderr, "%s(msr%d, %s, %s, width%d, scope%d, type%d, format%d, flags%x, id%d)\n",
9612 			__func__, msr_num, path, name, width, scope, type, format, flags, id);
9613 
9614 	switch (scope) {
9615 
9616 	case SCOPE_CPU:
9617 		msrp = find_msrp_by_name(sys.tp, name);
9618 		if (msrp) {
9619 			if (debug)
9620 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
9621 			break;
9622 		}
9623 		if (sys.added_thread_counters++ >= MAX_ADDED_THREAD_COUNTERS) {
9624 			warnx("ignoring thread counter %s", name);
9625 			return -1;
9626 		}
9627 		break;
9628 	case SCOPE_CORE:
9629 		msrp = find_msrp_by_name(sys.cp, name);
9630 		if (msrp) {
9631 			if (debug)
9632 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
9633 			break;
9634 		}
9635 		if (sys.added_core_counters++ >= MAX_ADDED_CORE_COUNTERS) {
9636 			warnx("ignoring core counter %s", name);
9637 			return -1;
9638 		}
9639 		break;
9640 	case SCOPE_PACKAGE:
9641 		msrp = find_msrp_by_name(sys.pp, name);
9642 		if (msrp) {
9643 			if (debug)
9644 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
9645 			break;
9646 		}
9647 		if (sys.added_package_counters++ >= MAX_ADDED_PACKAGE_COUNTERS) {
9648 			warnx("ignoring package counter %s", name);
9649 			return -1;
9650 		}
9651 		break;
9652 	default:
9653 		warnx("ignoring counter %s with unknown scope", name);
9654 		return -1;
9655 	}
9656 
9657 	if (msrp == NULL) {
9658 		msrp = calloc(1, sizeof(struct msr_counter));
9659 		if (msrp == NULL)
9660 			err(-1, "calloc msr_counter");
9661 
9662 		msrp->msr_num = msr_num;
9663 		strncpy(msrp->name, name, NAME_BYTES - 1);
9664 		msrp->width = width;
9665 		msrp->type = type;
9666 		msrp->format = format;
9667 		msrp->flags = flags;
9668 
9669 		switch (scope) {
9670 		case SCOPE_CPU:
9671 			msrp->next = sys.tp;
9672 			sys.tp = msrp;
9673 			break;
9674 		case SCOPE_CORE:
9675 			msrp->next = sys.cp;
9676 			sys.cp = msrp;
9677 			break;
9678 		case SCOPE_PACKAGE:
9679 			msrp->next = sys.pp;
9680 			sys.pp = msrp;
9681 			break;
9682 		}
9683 	}
9684 
9685 	if (path) {
9686 		struct sysfs_path *sp;
9687 
9688 		sp = calloc(1, sizeof(struct sysfs_path));
9689 		if (sp == NULL) {
9690 			perror("calloc");
9691 			exit(1);
9692 		}
9693 		strncpy(sp->path, path, PATH_BYTES - 1);
9694 		sp->id = id;
9695 		sp->next = msrp->sp;
9696 		msrp->sp = sp;
9697 	}
9698 
9699 	return 0;
9700 }
9701 
9702 /*
9703  * Initialize the fields used for identifying and opening the counter.
9704  *
9705  * Defer the initialization of any runtime buffers for actually reading
9706  * the counters for when we initialize all perf counters, so we can later
9707  * easily call re_initialize().
9708  */
make_perf_counter_info(const char * perf_device,const char * perf_event,const char * name,unsigned int width,enum counter_scope scope,enum counter_type type,enum counter_format format)9709 struct perf_counter_info *make_perf_counter_info(const char *perf_device,
9710 						 const char *perf_event,
9711 						 const char *name,
9712 						 unsigned int width,
9713 						 enum counter_scope scope,
9714 						 enum counter_type type, enum counter_format format)
9715 {
9716 	struct perf_counter_info *pinfo;
9717 
9718 	pinfo = calloc(1, sizeof(*pinfo));
9719 	if (!pinfo)
9720 		errx(1, "%s: Failed to allocate %s/%s\n", __func__, perf_device, perf_event);
9721 
9722 	strncpy(pinfo->device, perf_device, ARRAY_SIZE(pinfo->device) - 1);
9723 	strncpy(pinfo->event, perf_event, ARRAY_SIZE(pinfo->event) - 1);
9724 
9725 	strncpy(pinfo->name, name, ARRAY_SIZE(pinfo->name) - 1);
9726 	pinfo->width = width;
9727 	pinfo->scope = scope;
9728 	pinfo->type = type;
9729 	pinfo->format = format;
9730 
9731 	return pinfo;
9732 }
9733 
add_perf_counter(const char * perf_device,const char * perf_event,const char * name_buffer,unsigned int width,enum counter_scope scope,enum counter_type type,enum counter_format format)9734 int add_perf_counter(const char *perf_device, const char *perf_event, const char *name_buffer, unsigned int width,
9735 		     enum counter_scope scope, enum counter_type type, enum counter_format format)
9736 {
9737 	struct perf_counter_info *pinfo;
9738 
9739 	switch (scope) {
9740 	case SCOPE_CPU:
9741 		if (sys.added_thread_perf_counters >= MAX_ADDED_THREAD_COUNTERS) {
9742 			warnx("ignoring thread counter perf/%s/%s", perf_device, perf_event);
9743 			return -1;
9744 		}
9745 		break;
9746 
9747 	case SCOPE_CORE:
9748 		if (sys.added_core_perf_counters >= MAX_ADDED_CORE_COUNTERS) {
9749 			warnx("ignoring core counter perf/%s/%s", perf_device, perf_event);
9750 			return -1;
9751 		}
9752 		break;
9753 
9754 	case SCOPE_PACKAGE:
9755 		if (sys.added_package_perf_counters >= MAX_ADDED_PACKAGE_COUNTERS) {
9756 			warnx("ignoring package counter perf/%s/%s", perf_device, perf_event);
9757 			return -1;
9758 		}
9759 		break;
9760 	}
9761 
9762 	pinfo = make_perf_counter_info(perf_device, perf_event, name_buffer, width, scope, type, format);
9763 
9764 	if (!pinfo)
9765 		return -1;
9766 
9767 	switch (scope) {
9768 	case SCOPE_CPU:
9769 		pinfo->next = sys.perf_tp;
9770 		sys.perf_tp = pinfo;
9771 		++sys.added_thread_perf_counters;
9772 		break;
9773 
9774 	case SCOPE_CORE:
9775 		pinfo->next = sys.perf_cp;
9776 		sys.perf_cp = pinfo;
9777 		++sys.added_core_perf_counters;
9778 		break;
9779 
9780 	case SCOPE_PACKAGE:
9781 		pinfo->next = sys.perf_pp;
9782 		sys.perf_pp = pinfo;
9783 		++sys.added_package_perf_counters;
9784 		break;
9785 	}
9786 
9787 	// FIXME: we might not have debug here yet
9788 	if (debug)
9789 		fprintf(stderr, "%s: %s/%s, name: %s, scope%d\n",
9790 			__func__, pinfo->device, pinfo->event, pinfo->name, pinfo->scope);
9791 
9792 	return 0;
9793 }
9794 
parse_add_command_msr(char * add_command)9795 void parse_add_command_msr(char *add_command)
9796 {
9797 	int msr_num = 0;
9798 	char *path = NULL;
9799 	char perf_device[PERF_DEV_NAME_BYTES] = "";
9800 	char perf_event[PERF_EVT_NAME_BYTES] = "";
9801 	char name_buffer[PERF_NAME_BYTES] = "";
9802 	int width = 64;
9803 	int fail = 0;
9804 	enum counter_scope scope = SCOPE_CPU;
9805 	enum counter_type type = COUNTER_CYCLES;
9806 	enum counter_format format = FORMAT_DELTA;
9807 
9808 	while (add_command) {
9809 
9810 		if (sscanf(add_command, "msr0x%x", &msr_num) == 1)
9811 			goto next;
9812 
9813 		if (sscanf(add_command, "msr%d", &msr_num) == 1)
9814 			goto next;
9815 
9816 		BUILD_BUG_ON(ARRAY_SIZE(perf_device) <= 31);
9817 		BUILD_BUG_ON(ARRAY_SIZE(perf_event) <= 31);
9818 		if (sscanf(add_command, "perf/%31[^/]/%31[^,]", &perf_device[0], &perf_event[0]) == 2)
9819 			goto next;
9820 
9821 		if (*add_command == '/') {
9822 			path = add_command;
9823 			goto next;
9824 		}
9825 
9826 		if (sscanf(add_command, "u%d", &width) == 1) {
9827 			if ((width == 32) || (width == 64))
9828 				goto next;
9829 			width = 64;
9830 		}
9831 		if (!strncmp(add_command, "cpu", strlen("cpu"))) {
9832 			scope = SCOPE_CPU;
9833 			goto next;
9834 		}
9835 		if (!strncmp(add_command, "core", strlen("core"))) {
9836 			scope = SCOPE_CORE;
9837 			goto next;
9838 		}
9839 		if (!strncmp(add_command, "package", strlen("package"))) {
9840 			scope = SCOPE_PACKAGE;
9841 			goto next;
9842 		}
9843 		if (!strncmp(add_command, "cycles", strlen("cycles"))) {
9844 			type = COUNTER_CYCLES;
9845 			goto next;
9846 		}
9847 		if (!strncmp(add_command, "seconds", strlen("seconds"))) {
9848 			type = COUNTER_SECONDS;
9849 			goto next;
9850 		}
9851 		if (!strncmp(add_command, "usec", strlen("usec"))) {
9852 			type = COUNTER_USEC;
9853 			goto next;
9854 		}
9855 		if (!strncmp(add_command, "raw", strlen("raw"))) {
9856 			format = FORMAT_RAW;
9857 			goto next;
9858 		}
9859 		if (!strncmp(add_command, "delta", strlen("delta"))) {
9860 			format = FORMAT_DELTA;
9861 			goto next;
9862 		}
9863 		if (!strncmp(add_command, "percent", strlen("percent"))) {
9864 			format = FORMAT_PERCENT;
9865 			goto next;
9866 		}
9867 
9868 		BUILD_BUG_ON(ARRAY_SIZE(name_buffer) <= 18);
9869 		if (sscanf(add_command, "%18s,%*s", name_buffer) == 1) {
9870 			char *eos;
9871 
9872 			eos = strchr(name_buffer, ',');
9873 			if (eos)
9874 				*eos = '\0';
9875 			goto next;
9876 		}
9877 
9878 next:
9879 		add_command = strchr(add_command, ',');
9880 		if (add_command) {
9881 			*add_command = '\0';
9882 			add_command++;
9883 		}
9884 
9885 	}
9886 	if ((msr_num == 0) && (path == NULL) && (perf_device[0] == '\0' || perf_event[0] == '\0')) {
9887 		fprintf(stderr, "--add: (msrDDD | msr0xXXX | /path_to_counter | perf/device/event) required\n");
9888 		fail++;
9889 	}
9890 
9891 	/* Test for non-empty perf_device and perf_event */
9892 	const bool is_perf_counter = perf_device[0] && perf_event[0];
9893 
9894 	/* generate default column header */
9895 	if (*name_buffer == '\0') {
9896 		if (is_perf_counter) {
9897 			snprintf(name_buffer, ARRAY_SIZE(name_buffer), "perf/%s", perf_event);
9898 		} else {
9899 			if (width == 32)
9900 				sprintf(name_buffer, "M0x%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
9901 			else
9902 				sprintf(name_buffer, "M0X%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
9903 		}
9904 	}
9905 
9906 	if (is_perf_counter) {
9907 		if (add_perf_counter(perf_device, perf_event, name_buffer, width, scope, type, format))
9908 			fail++;
9909 	} else {
9910 		if (add_counter(msr_num, path, name_buffer, width, scope, type, format, 0, 0))
9911 			fail++;
9912 	}
9913 
9914 	if (fail) {
9915 		help();
9916 		exit(1);
9917 	}
9918 }
9919 
starts_with(const char * str,const char * prefix)9920 bool starts_with(const char *str, const char *prefix)
9921 {
9922 	return strncmp(prefix, str, strlen(prefix)) == 0;
9923 }
9924 
pmt_parse_from_path(const char * target_path,unsigned int * out_guid,unsigned int * out_seq)9925 int pmt_parse_from_path(const char *target_path, unsigned int *out_guid, unsigned int *out_seq)
9926 {
9927 	struct pmt_diriter_t pmt_iter;
9928 	const struct dirent *dirname;
9929 	struct stat stat, target_stat;
9930 	int fd_telem_dir = -1;
9931 	int fd_target_dir;
9932 	unsigned int seq = 0;
9933 	unsigned long guid, target_guid;
9934 	int ret = -1;
9935 
9936 	fd_target_dir = open(target_path, O_RDONLY | O_DIRECTORY);
9937 	if (fd_target_dir == -1) {
9938 		return -1;
9939 	}
9940 
9941 	if (fstat(fd_target_dir, &target_stat) == -1) {
9942 		fprintf(stderr, "%s: Failed to stat the target: %s", __func__, strerror(errno));
9943 		exit(1);
9944 	}
9945 
9946 	if (parse_telem_info_file(fd_target_dir, "guid", "%lx", &target_guid)) {
9947 		fprintf(stderr, "%s: Failed to parse the target guid file: %s", __func__, strerror(errno));
9948 		exit(1);
9949 	}
9950 
9951 	close(fd_target_dir);
9952 
9953 	pmt_diriter_init(&pmt_iter);
9954 
9955 	for (dirname = pmt_diriter_begin(&pmt_iter, SYSFS_TELEM_PATH); dirname != NULL;
9956 	     dirname = pmt_diriter_next(&pmt_iter)) {
9957 
9958 		fd_telem_dir = openat(dirfd(pmt_iter.dir), dirname->d_name, O_RDONLY | O_DIRECTORY);
9959 		if (fd_telem_dir == -1)
9960 			continue;
9961 
9962 		if (parse_telem_info_file(fd_telem_dir, "guid", "%lx", &guid)) {
9963 			fprintf(stderr, "%s: Failed to parse the guid file: %s", __func__, strerror(errno));
9964 			continue;
9965 		}
9966 
9967 		if (fstat(fd_telem_dir, &stat) == -1) {
9968 			fprintf(stderr, "%s: Failed to stat %s directory: %s", __func__,
9969 				dirname->d_name, strerror(errno));
9970 			continue;
9971 		}
9972 
9973 		/*
9974 		 * If reached the same directory as target, exit the loop.
9975 		 * Seq has the correct value now.
9976 		 */
9977 		if (stat.st_dev == target_stat.st_dev && stat.st_ino == target_stat.st_ino) {
9978 			ret = 0;
9979 			break;
9980 		}
9981 
9982 		/*
9983 		 * If reached directory with the same guid,
9984 		 * but it's not the target directory yet,
9985 		 * increment seq and continue the search.
9986 		 */
9987 		if (guid == target_guid)
9988 			++seq;
9989 
9990 		close(fd_telem_dir);
9991 		fd_telem_dir = -1;
9992 	}
9993 
9994 	pmt_diriter_remove(&pmt_iter);
9995 
9996 	if (fd_telem_dir != -1)
9997 		close(fd_telem_dir);
9998 
9999 	if (!ret) {
10000 		*out_guid = target_guid;
10001 		*out_seq = seq;
10002 	}
10003 
10004 	return ret;
10005 }
10006 
parse_add_command_pmt(char * add_command)10007 void parse_add_command_pmt(char *add_command)
10008 {
10009 	char *name = NULL;
10010 	char *type_name = NULL;
10011 	char *format_name = NULL;
10012 	char *direct_path = NULL;
10013 	static const char direct_path_prefix[] = "path=";
10014 	unsigned int offset;
10015 	unsigned int lsb;
10016 	unsigned int msb;
10017 	unsigned int guid;
10018 	unsigned int seq = 0; /* By default, pick first file in a sequence with a given GUID. */
10019 	unsigned int domain_id;
10020 	enum counter_scope scope = 0;
10021 	enum pmt_datatype type = PMT_TYPE_RAW;
10022 	enum counter_format format = FORMAT_RAW;
10023 	bool has_offset = false;
10024 	bool has_lsb = false;
10025 	bool has_msb = false;
10026 	bool has_format = true;	/* Format has a default value. */
10027 	bool has_guid = false;
10028 	bool has_scope = false;
10029 	bool has_type = true;	/* Type has a default value. */
10030 
10031 	/* Consume the "pmt," prefix. */
10032 	add_command = strchr(add_command, ',');
10033 	if (!add_command) {
10034 		help();
10035 		exit(1);
10036 	}
10037 	++add_command;
10038 
10039 	while (add_command) {
10040 		if (starts_with(add_command, "name=")) {
10041 			name = add_command + strlen("name=");
10042 			goto next;
10043 		}
10044 
10045 		if (starts_with(add_command, "type=")) {
10046 			type_name = add_command + strlen("type=");
10047 			goto next;
10048 		}
10049 
10050 		if (starts_with(add_command, "domain=")) {
10051 			const size_t prefix_len = strlen("domain=");
10052 
10053 			if (sscanf(add_command + prefix_len, "cpu%u", &domain_id) == 1) {
10054 				scope = SCOPE_CPU;
10055 				has_scope = true;
10056 			} else if (sscanf(add_command + prefix_len, "core%u", &domain_id) == 1) {
10057 				scope = SCOPE_CORE;
10058 				has_scope = true;
10059 			} else if (sscanf(add_command + prefix_len, "package%u", &domain_id) == 1) {
10060 				scope = SCOPE_PACKAGE;
10061 				has_scope = true;
10062 			}
10063 
10064 			if (!has_scope) {
10065 				printf("%s: invalid value for scope. Expected cpu%%u, core%%u or package%%u.\n",
10066 				       __func__);
10067 				exit(1);
10068 			}
10069 
10070 			goto next;
10071 		}
10072 
10073 		if (starts_with(add_command, "format=")) {
10074 			format_name = add_command + strlen("format=");
10075 			goto next;
10076 		}
10077 
10078 		if (sscanf(add_command, "offset=%u", &offset) == 1) {
10079 			has_offset = true;
10080 			goto next;
10081 		}
10082 
10083 		if (sscanf(add_command, "lsb=%u", &lsb) == 1) {
10084 			has_lsb = true;
10085 			goto next;
10086 		}
10087 
10088 		if (sscanf(add_command, "msb=%u", &msb) == 1) {
10089 			has_msb = true;
10090 			goto next;
10091 		}
10092 
10093 		if (sscanf(add_command, "guid=%x", &guid) == 1) {
10094 			has_guid = true;
10095 			goto next;
10096 		}
10097 
10098 		if (sscanf(add_command, "seq=%x", &seq) == 1)
10099 			goto next;
10100 
10101 		if (strncmp(add_command, direct_path_prefix, strlen(direct_path_prefix)) == 0) {
10102 			direct_path = add_command + strlen(direct_path_prefix);
10103 			goto next;
10104 		}
10105 next:
10106 		add_command = strchr(add_command, ',');
10107 		if (add_command) {
10108 			*add_command = '\0';
10109 			add_command++;
10110 		}
10111 	}
10112 
10113 	if (!name) {
10114 		printf("%s: missing %s\n", __func__, "name");
10115 		exit(1);
10116 	}
10117 
10118 	if (strlen(name) >= PMT_COUNTER_NAME_SIZE_BYTES) {
10119 		printf("%s: name has to be at most %d characters long\n", __func__, PMT_COUNTER_NAME_SIZE_BYTES);
10120 		exit(1);
10121 	}
10122 
10123 	if (format_name) {
10124 		has_format = false;
10125 
10126 		if (strcmp("raw", format_name) == 0) {
10127 			format = FORMAT_RAW;
10128 			has_format = true;
10129 		}
10130 
10131 		if (strcmp("delta", format_name) == 0) {
10132 			format = FORMAT_DELTA;
10133 			has_format = true;
10134 		}
10135 
10136 		if (!has_format) {
10137 			fprintf(stderr, "%s: Invalid format %s. Expected raw or delta\n", __func__, format_name);
10138 			exit(1);
10139 		}
10140 	}
10141 
10142 	if (type_name) {
10143 		has_type = false;
10144 
10145 		if (strcmp("raw", type_name) == 0) {
10146 			type = PMT_TYPE_RAW;
10147 			has_type = true;
10148 		}
10149 
10150 		if (strcmp("txtal_time", type_name) == 0) {
10151 			type = PMT_TYPE_XTAL_TIME;
10152 			has_type = true;
10153 		}
10154 
10155 		if (strcmp("tcore_clock", type_name) == 0) {
10156 			type = PMT_TYPE_TCORE_CLOCK;
10157 			has_type = true;
10158 		}
10159 
10160 		if (!has_type) {
10161 			printf("%s: invalid %s: %s\n", __func__, "type", type_name);
10162 			exit(1);
10163 		}
10164 	}
10165 
10166 	if (!has_offset) {
10167 		printf("%s : missing %s\n", __func__, "offset");
10168 		exit(1);
10169 	}
10170 
10171 	if (!has_lsb) {
10172 		printf("%s: missing %s\n", __func__, "lsb");
10173 		exit(1);
10174 	}
10175 
10176 	if (!has_msb) {
10177 		printf("%s: missing %s\n", __func__, "msb");
10178 		exit(1);
10179 	}
10180 
10181 	if (direct_path && has_guid) {
10182 		printf("%s: path and guid+seq parameters are mutually exclusive\n"
10183 		       "notice: passed guid=0x%x and path=%s\n", __func__, guid, direct_path);
10184 		exit(1);
10185 	}
10186 
10187 	if (direct_path) {
10188 		if (pmt_parse_from_path(direct_path, &guid, &seq)) {
10189 			printf("%s: failed to parse PMT file from %s\n", __func__, direct_path);
10190 			exit(1);
10191 		}
10192 
10193 		/* GUID was just infered from the direct path. */
10194 		has_guid = true;
10195 	}
10196 
10197 	if (!has_guid) {
10198 		printf("%s: missing %s\n", __func__, "guid or path");
10199 		exit(1);
10200 	}
10201 
10202 	if (!has_scope) {
10203 		printf("%s: missing %s\n", __func__, "scope");
10204 		exit(1);
10205 	}
10206 
10207 	if (lsb > msb) {
10208 		printf("%s: lsb > msb doesn't make sense\n", __func__);
10209 		exit(1);
10210 	}
10211 
10212 	pmt_add_counter(guid, seq, name, type, lsb, msb, offset, scope, format, domain_id, PMT_OPEN_REQUIRED);
10213 }
10214 
parse_add_command(char * add_command)10215 void parse_add_command(char *add_command)
10216 {
10217 	if (strncmp(add_command, "pmt", strlen("pmt")) == 0)
10218 		return parse_add_command_pmt(add_command);
10219 	return parse_add_command_msr(add_command);
10220 }
10221 
is_deferred_add(char * name)10222 int is_deferred_add(char *name)
10223 {
10224 	int i;
10225 
10226 	for (i = 0; i < deferred_add_index; ++i)
10227 		if (!strcmp(name, deferred_add_names[i]))
10228 			return 1;
10229 	return 0;
10230 }
10231 
is_deferred_skip(char * name)10232 int is_deferred_skip(char *name)
10233 {
10234 	int i;
10235 
10236 	for (i = 0; i < deferred_skip_index; ++i)
10237 		if (!strcmp(name, deferred_skip_names[i]))
10238 			return 1;
10239 	return 0;
10240 }
10241 
probe_sysfs(void)10242 void probe_sysfs(void)
10243 {
10244 	char path[64];
10245 	char name_buf[16];
10246 	FILE *input;
10247 	int state;
10248 	char *sp;
10249 
10250 	for (state = 10; state >= 0; --state) {
10251 
10252 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", base_cpu, state);
10253 		input = fopen(path, "r");
10254 		if (input == NULL)
10255 			continue;
10256 		if (!fgets(name_buf, sizeof(name_buf), input))
10257 			err(1, "%s: failed to read file", path);
10258 
10259 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
10260 		sp = strchr(name_buf, '-');
10261 		if (!sp)
10262 			sp = strchrnul(name_buf, '\n');
10263 		*sp = '%';
10264 		*(sp + 1) = '\0';
10265 
10266 		remove_underbar(name_buf);
10267 
10268 		fclose(input);
10269 
10270 		sprintf(path, "cpuidle/state%d/time", state);
10271 
10272 		if (!DO_BIC(BIC_sysfs) && !is_deferred_add(name_buf))
10273 			continue;
10274 
10275 		if (is_deferred_skip(name_buf))
10276 			continue;
10277 
10278 		add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_USEC, FORMAT_PERCENT, SYSFS_PERCPU, 0);
10279 	}
10280 
10281 	for (state = 10; state >= 0; --state) {
10282 
10283 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", base_cpu, state);
10284 		input = fopen(path, "r");
10285 		if (input == NULL)
10286 			continue;
10287 		if (!fgets(name_buf, sizeof(name_buf), input))
10288 			err(1, "%s: failed to read file", path);
10289 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
10290 		sp = strchr(name_buf, '-');
10291 		if (!sp)
10292 			sp = strchrnul(name_buf, '\n');
10293 		*sp = '\0';
10294 		fclose(input);
10295 
10296 		remove_underbar(name_buf);
10297 
10298 		sprintf(path, "cpuidle/state%d/usage", state);
10299 
10300 		if (!DO_BIC(BIC_sysfs) && !is_deferred_add(name_buf))
10301 			continue;
10302 
10303 		if (is_deferred_skip(name_buf))
10304 			continue;
10305 
10306 		add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS, FORMAT_DELTA, SYSFS_PERCPU, 0);
10307 	}
10308 
10309 }
10310 
10311 /*
10312  * parse cpuset with following syntax
10313  * 1,2,4..6,8-10 and set bits in cpu_subset
10314  */
parse_cpu_command(char * optarg)10315 void parse_cpu_command(char *optarg)
10316 {
10317 	if (!strcmp(optarg, "core")) {
10318 		if (cpu_subset)
10319 			goto error;
10320 		show_core_only++;
10321 		return;
10322 	}
10323 	if (!strcmp(optarg, "package")) {
10324 		if (cpu_subset)
10325 			goto error;
10326 		show_pkg_only++;
10327 		return;
10328 	}
10329 	if (show_core_only || show_pkg_only)
10330 		goto error;
10331 
10332 	cpu_subset = CPU_ALLOC(CPU_SUBSET_MAXCPUS);
10333 	if (cpu_subset == NULL)
10334 		err(3, "CPU_ALLOC");
10335 	cpu_subset_size = CPU_ALLOC_SIZE(CPU_SUBSET_MAXCPUS);
10336 
10337 	CPU_ZERO_S(cpu_subset_size, cpu_subset);
10338 
10339 	if (parse_cpu_str(optarg, cpu_subset, cpu_subset_size))
10340 		goto error;
10341 
10342 	return;
10343 
10344 error:
10345 	fprintf(stderr, "\"--cpu %s\" malformed\n", optarg);
10346 	help();
10347 	exit(-1);
10348 }
10349 
cmdline(int argc,char ** argv)10350 void cmdline(int argc, char **argv)
10351 {
10352 	int opt;
10353 	int option_index = 0;
10354 	static struct option long_options[] = {
10355 		{ "add", required_argument, 0, 'a' },
10356 		{ "cpu", required_argument, 0, 'c' },
10357 		{ "Dump", no_argument, 0, 'D' },
10358 		{ "debug", no_argument, 0, 'd' },	/* internal, not documented */
10359 		{ "enable", required_argument, 0, 'e' },
10360 		{ "force", no_argument, 0, 'f' },
10361 		{ "interval", required_argument, 0, 'i' },
10362 		{ "IPC", no_argument, 0, 'I' },
10363 		{ "num_iterations", required_argument, 0, 'n' },
10364 		{ "header_iterations", required_argument, 0, 'N' },
10365 		{ "help", no_argument, 0, 'h' },
10366 		{ "hide", required_argument, 0, 'H' },	// meh, -h taken by --help
10367 		{ "Joules", no_argument, 0, 'J' },
10368 		{ "list", no_argument, 0, 'l' },
10369 		{ "out", required_argument, 0, 'o' },
10370 		{ "quiet", no_argument, 0, 'q' },
10371 		{ "no-msr", no_argument, 0, 'M' },
10372 		{ "no-perf", no_argument, 0, 'P' },
10373 		{ "show", required_argument, 0, 's' },
10374 		{ "Summary", no_argument, 0, 'S' },
10375 		{ "TCC", required_argument, 0, 'T' },
10376 		{ "version", no_argument, 0, 'v' },
10377 		{ 0, 0, 0, 0 }
10378 	};
10379 
10380 	progname = argv[0];
10381 
10382 	/*
10383 	 * Parse some options early, because they may make other options invalid,
10384 	 * like adding the MSR counter with --add and at the same time using --no-msr.
10385 	 */
10386 	while ((opt = getopt_long_only(argc, argv, "+MPn:", long_options, &option_index)) != -1) {
10387 		switch (opt) {
10388 		case 'M':
10389 			no_msr = 1;
10390 			break;
10391 		case 'P':
10392 			no_perf = 1;
10393 			break;
10394 		default:
10395 			break;
10396 		}
10397 	}
10398 	optind = 0;
10399 
10400 	while ((opt = getopt_long_only(argc, argv, "+C:c:Dde:hi:Jn:o:qMST:v", long_options, &option_index)) != -1) {
10401 		switch (opt) {
10402 		case 'a':
10403 			parse_add_command(optarg);
10404 			break;
10405 		case 'c':
10406 			parse_cpu_command(optarg);
10407 			break;
10408 		case 'D':
10409 			dump_only++;
10410 			/*
10411 			 * Force the no_perf early to prevent using it as a source.
10412 			 * User asks for raw values, but perf returns them relative
10413 			 * to the opening of the file descriptor.
10414 			 */
10415 			no_perf = 1;
10416 			break;
10417 		case 'e':
10418 			/* --enable specified counter */
10419 			bic_enabled = bic_enabled | bic_lookup(optarg, SHOW_LIST);
10420 			break;
10421 		case 'f':
10422 			force_load++;
10423 			break;
10424 		case 'd':
10425 			debug++;
10426 			ENABLE_BIC(BIC_DISABLED_BY_DEFAULT);
10427 			break;
10428 		case 'H':
10429 			/*
10430 			 * --hide: do not show those specified
10431 			 *  multiple invocations simply clear more bits in enabled mask
10432 			 */
10433 			bic_enabled &= ~bic_lookup(optarg, HIDE_LIST);
10434 			break;
10435 		case 'h':
10436 		default:
10437 			help();
10438 			exit(1);
10439 		case 'i':
10440 			{
10441 				double interval = strtod(optarg, NULL);
10442 
10443 				if (interval < 0.001) {
10444 					fprintf(outf, "interval %f seconds is too small\n", interval);
10445 					exit(2);
10446 				}
10447 
10448 				interval_tv.tv_sec = interval_ts.tv_sec = interval;
10449 				interval_tv.tv_usec = (interval - interval_tv.tv_sec) * 1000000;
10450 				interval_ts.tv_nsec = (interval - interval_ts.tv_sec) * 1000000000;
10451 			}
10452 			break;
10453 		case 'J':
10454 			rapl_joules++;
10455 			break;
10456 		case 'l':
10457 			ENABLE_BIC(BIC_DISABLED_BY_DEFAULT);
10458 			list_header_only++;
10459 			quiet++;
10460 			break;
10461 		case 'o':
10462 			outf = fopen_or_die(optarg, "w");
10463 			break;
10464 		case 'q':
10465 			quiet = 1;
10466 			break;
10467 		case 'M':
10468 		case 'P':
10469 			/* Parsed earlier */
10470 			break;
10471 		case 'n':
10472 			num_iterations = strtod(optarg, NULL);
10473 
10474 			if (num_iterations <= 0) {
10475 				fprintf(outf, "iterations %d should be positive number\n", num_iterations);
10476 				exit(2);
10477 			}
10478 			break;
10479 		case 'N':
10480 			header_iterations = strtod(optarg, NULL);
10481 
10482 			if (header_iterations <= 0) {
10483 				fprintf(outf, "iterations %d should be positive number\n", header_iterations);
10484 				exit(2);
10485 			}
10486 			break;
10487 		case 's':
10488 			/*
10489 			 * --show: show only those specified
10490 			 *  The 1st invocation will clear and replace the enabled mask
10491 			 *  subsequent invocations can add to it.
10492 			 */
10493 			if (shown == 0)
10494 				bic_enabled = bic_lookup(optarg, SHOW_LIST);
10495 			else
10496 				bic_enabled |= bic_lookup(optarg, SHOW_LIST);
10497 			shown = 1;
10498 			break;
10499 		case 'S':
10500 			summary_only++;
10501 			break;
10502 		case 'T':
10503 			tj_max_override = atoi(optarg);
10504 			break;
10505 		case 'v':
10506 			print_version();
10507 			exit(0);
10508 			break;
10509 		}
10510 	}
10511 }
10512 
set_rlimit(void)10513 void set_rlimit(void)
10514 {
10515 	struct rlimit limit;
10516 
10517 	if (getrlimit(RLIMIT_NOFILE, &limit) < 0)
10518 		err(1, "Failed to get rlimit");
10519 
10520 	if (limit.rlim_max < MAX_NOFILE)
10521 		limit.rlim_max = MAX_NOFILE;
10522 	if (limit.rlim_cur < MAX_NOFILE)
10523 		limit.rlim_cur = MAX_NOFILE;
10524 
10525 	if (setrlimit(RLIMIT_NOFILE, &limit) < 0)
10526 		err(1, "Failed to set rlimit");
10527 }
10528 
main(int argc,char ** argv)10529 int main(int argc, char **argv)
10530 {
10531 	int fd, ret;
10532 
10533 	fd = open("/sys/fs/cgroup/cgroup.procs", O_WRONLY);
10534 	if (fd < 0)
10535 		goto skip_cgroup_setting;
10536 
10537 	ret = write(fd, "0\n", 2);
10538 	if (ret == -1)
10539 		perror("Can't update cgroup\n");
10540 
10541 	close(fd);
10542 
10543 skip_cgroup_setting:
10544 	outf = stderr;
10545 	cmdline(argc, argv);
10546 
10547 	if (!quiet) {
10548 		print_version();
10549 		print_bootcmd();
10550 	}
10551 
10552 	probe_sysfs();
10553 
10554 	if (!getuid())
10555 		set_rlimit();
10556 
10557 	turbostat_init();
10558 
10559 	if (!no_msr)
10560 		msr_sum_record();
10561 
10562 	/* dump counters and exit */
10563 	if (dump_only)
10564 		return get_and_dump_counters();
10565 
10566 	/* list header and exit */
10567 	if (list_header_only) {
10568 		print_header(",");
10569 		flush_output_stdout();
10570 		return 0;
10571 	}
10572 
10573 	/*
10574 	 * if any params left, it must be a command to fork
10575 	 */
10576 	if (argc - optind)
10577 		return fork_it(argv + optind);
10578 	else
10579 		turbostat_loop();
10580 
10581 	return 0;
10582 }
10583