xref: /linux/tools/power/x86/turbostat/turbostat.c (revision bcc8eda6d34934d80b96adb8dc4ff5dfc632a53a)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * turbostat -- show CPU frequency and C-state residency
4  * on modern Intel and AMD processors.
5  *
6  * Copyright (c) 2024 Intel Corporation.
7  * Len Brown <len.brown@intel.com>
8  */
9 
10 #define _GNU_SOURCE
11 #include MSRHEADER
12 
13 // copied from arch/x86/include/asm/cpu_device_id.h
14 #define VFM_MODEL_BIT	0
15 #define VFM_FAMILY_BIT	8
16 #define VFM_VENDOR_BIT	16
17 #define VFM_RSVD_BIT	24
18 
19 #define	VFM_MODEL_MASK	GENMASK(VFM_FAMILY_BIT - 1, VFM_MODEL_BIT)
20 #define	VFM_FAMILY_MASK	GENMASK(VFM_VENDOR_BIT - 1, VFM_FAMILY_BIT)
21 #define	VFM_VENDOR_MASK	GENMASK(VFM_RSVD_BIT - 1, VFM_VENDOR_BIT)
22 
23 #define VFM_MODEL(vfm)	(((vfm) & VFM_MODEL_MASK) >> VFM_MODEL_BIT)
24 #define VFM_FAMILY(vfm)	(((vfm) & VFM_FAMILY_MASK) >> VFM_FAMILY_BIT)
25 #define VFM_VENDOR(vfm)	(((vfm) & VFM_VENDOR_MASK) >> VFM_VENDOR_BIT)
26 
27 #define	VFM_MAKE(_vendor, _family, _model) (	\
28 	((_model) << VFM_MODEL_BIT) |		\
29 	((_family) << VFM_FAMILY_BIT) |		\
30 	((_vendor) << VFM_VENDOR_BIT)		\
31 )
32 // end copied section
33 
34 #define 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 enum counter_scope { SCOPE_CPU, SCOPE_CORE, SCOPE_PACKAGE };
99 enum counter_type { COUNTER_ITEMS, COUNTER_CYCLES, COUNTER_SECONDS, COUNTER_USEC, COUNTER_K2M };
100 enum counter_format { FORMAT_RAW, FORMAT_DELTA, FORMAT_PERCENT, FORMAT_AVERAGE };
101 enum counter_source { COUNTER_SOURCE_NONE, COUNTER_SOURCE_PERF, COUNTER_SOURCE_MSR };
102 
103 struct perf_counter_info {
104 	struct perf_counter_info *next;
105 
106 	/* How to open the counter / What counter it is. */
107 	char device[PERF_DEV_NAME_BYTES];
108 	char event[PERF_EVT_NAME_BYTES];
109 
110 	/* How to show/format the counter. */
111 	char name[PERF_NAME_BYTES];
112 	unsigned int width;
113 	enum counter_scope scope;
114 	enum counter_type type;
115 	enum counter_format format;
116 	double scale;
117 
118 	/* For reading the counter. */
119 	int *fd_perf_per_domain;
120 	size_t num_domains;
121 };
122 
123 struct sysfs_path {
124 	char path[PATH_BYTES];
125 	int id;
126 	struct sysfs_path *next;
127 };
128 
129 struct msr_counter {
130 	unsigned int msr_num;
131 	char name[NAME_BYTES];
132 	struct sysfs_path *sp;
133 	unsigned int width;
134 	enum counter_type type;
135 	enum counter_format format;
136 	struct msr_counter *next;
137 	unsigned int flags;
138 #define	FLAGS_HIDE	(1 << 0)
139 #define	FLAGS_SHOW	(1 << 1)
140 #define	SYSFS_PERCPU	(1 << 1)
141 };
142 
143 struct msr_counter bic[] = {
144 	{ 0x0, "usec", NULL, 0, 0, 0, NULL, 0 },
145 	{ 0x0, "Time_Of_Day_Seconds", NULL, 0, 0, 0, NULL, 0 },
146 	{ 0x0, "Package", NULL, 0, 0, 0, NULL, 0 },
147 	{ 0x0, "Node", NULL, 0, 0, 0, NULL, 0 },
148 	{ 0x0, "Avg_MHz", NULL, 0, 0, 0, NULL, 0 },
149 	{ 0x0, "Busy%", NULL, 0, 0, 0, NULL, 0 },
150 	{ 0x0, "Bzy_MHz", NULL, 0, 0, 0, NULL, 0 },
151 	{ 0x0, "TSC_MHz", NULL, 0, 0, 0, NULL, 0 },
152 	{ 0x0, "IRQ", NULL, 0, 0, 0, NULL, 0 },
153 	{ 0x0, "SMI", NULL, 32, 0, FORMAT_DELTA, NULL, 0 },
154 	{ 0x0, "sysfs", NULL, 0, 0, 0, NULL, 0 },
155 	{ 0x0, "CPU%c1", NULL, 0, 0, 0, NULL, 0 },
156 	{ 0x0, "CPU%c3", NULL, 0, 0, 0, NULL, 0 },
157 	{ 0x0, "CPU%c6", NULL, 0, 0, 0, NULL, 0 },
158 	{ 0x0, "CPU%c7", NULL, 0, 0, 0, NULL, 0 },
159 	{ 0x0, "ThreadC", NULL, 0, 0, 0, NULL, 0 },
160 	{ 0x0, "CoreTmp", NULL, 0, 0, 0, NULL, 0 },
161 	{ 0x0, "CoreCnt", NULL, 0, 0, 0, NULL, 0 },
162 	{ 0x0, "PkgTmp", NULL, 0, 0, 0, NULL, 0 },
163 	{ 0x0, "GFX%rc6", NULL, 0, 0, 0, NULL, 0 },
164 	{ 0x0, "GFXMHz", NULL, 0, 0, 0, NULL, 0 },
165 	{ 0x0, "Pkg%pc2", NULL, 0, 0, 0, NULL, 0 },
166 	{ 0x0, "Pkg%pc3", NULL, 0, 0, 0, NULL, 0 },
167 	{ 0x0, "Pkg%pc6", NULL, 0, 0, 0, NULL, 0 },
168 	{ 0x0, "Pkg%pc7", NULL, 0, 0, 0, NULL, 0 },
169 	{ 0x0, "Pkg%pc8", NULL, 0, 0, 0, NULL, 0 },
170 	{ 0x0, "Pkg%pc9", NULL, 0, 0, 0, NULL, 0 },
171 	{ 0x0, "Pk%pc10", NULL, 0, 0, 0, NULL, 0 },
172 	{ 0x0, "CPU%LPI", NULL, 0, 0, 0, NULL, 0 },
173 	{ 0x0, "SYS%LPI", NULL, 0, 0, 0, NULL, 0 },
174 	{ 0x0, "PkgWatt", NULL, 0, 0, 0, NULL, 0 },
175 	{ 0x0, "CorWatt", NULL, 0, 0, 0, NULL, 0 },
176 	{ 0x0, "GFXWatt", NULL, 0, 0, 0, NULL, 0 },
177 	{ 0x0, "PkgCnt", NULL, 0, 0, 0, NULL, 0 },
178 	{ 0x0, "RAMWatt", NULL, 0, 0, 0, NULL, 0 },
179 	{ 0x0, "PKG_%", NULL, 0, 0, 0, NULL, 0 },
180 	{ 0x0, "RAM_%", NULL, 0, 0, 0, NULL, 0 },
181 	{ 0x0, "Pkg_J", NULL, 0, 0, 0, NULL, 0 },
182 	{ 0x0, "Cor_J", NULL, 0, 0, 0, NULL, 0 },
183 	{ 0x0, "GFX_J", NULL, 0, 0, 0, NULL, 0 },
184 	{ 0x0, "RAM_J", NULL, 0, 0, 0, NULL, 0 },
185 	{ 0x0, "Mod%c6", NULL, 0, 0, 0, NULL, 0 },
186 	{ 0x0, "Totl%C0", NULL, 0, 0, 0, NULL, 0 },
187 	{ 0x0, "Any%C0", NULL, 0, 0, 0, NULL, 0 },
188 	{ 0x0, "GFX%C0", NULL, 0, 0, 0, NULL, 0 },
189 	{ 0x0, "CPUGFX%", NULL, 0, 0, 0, NULL, 0 },
190 	{ 0x0, "Core", NULL, 0, 0, 0, NULL, 0 },
191 	{ 0x0, "CPU", NULL, 0, 0, 0, NULL, 0 },
192 	{ 0x0, "APIC", NULL, 0, 0, 0, NULL, 0 },
193 	{ 0x0, "X2APIC", NULL, 0, 0, 0, NULL, 0 },
194 	{ 0x0, "Die", NULL, 0, 0, 0, NULL, 0 },
195 	{ 0x0, "GFXAMHz", NULL, 0, 0, 0, NULL, 0 },
196 	{ 0x0, "IPC", NULL, 0, 0, 0, NULL, 0 },
197 	{ 0x0, "CoreThr", NULL, 0, 0, 0, NULL, 0 },
198 	{ 0x0, "UncMHz", NULL, 0, 0, 0, NULL, 0 },
199 	{ 0x0, "SAM%mc6", NULL, 0, 0, 0, NULL, 0 },
200 	{ 0x0, "SAMMHz", NULL, 0, 0, 0, NULL, 0 },
201 	{ 0x0, "SAMAMHz", NULL, 0, 0, 0, NULL, 0 },
202 	{ 0x0, "Die%c6", NULL, 0, 0, 0, NULL, 0 },
203 	{ 0x0, "SysWatt", NULL, 0, 0, 0, NULL, 0 },
204 	{ 0x0, "Sys_J", NULL, 0, 0, 0, NULL, 0 },
205 };
206 
207 #define MAX_BIC (sizeof(bic) / sizeof(struct msr_counter))
208 #define	BIC_USEC	(1ULL << 0)
209 #define	BIC_TOD		(1ULL << 1)
210 #define	BIC_Package	(1ULL << 2)
211 #define	BIC_Node	(1ULL << 3)
212 #define	BIC_Avg_MHz	(1ULL << 4)
213 #define	BIC_Busy	(1ULL << 5)
214 #define	BIC_Bzy_MHz	(1ULL << 6)
215 #define	BIC_TSC_MHz	(1ULL << 7)
216 #define	BIC_IRQ		(1ULL << 8)
217 #define	BIC_SMI		(1ULL << 9)
218 #define	BIC_sysfs	(1ULL << 10)
219 #define	BIC_CPU_c1	(1ULL << 11)
220 #define	BIC_CPU_c3	(1ULL << 12)
221 #define	BIC_CPU_c6	(1ULL << 13)
222 #define	BIC_CPU_c7	(1ULL << 14)
223 #define	BIC_ThreadC	(1ULL << 15)
224 #define	BIC_CoreTmp	(1ULL << 16)
225 #define	BIC_CoreCnt	(1ULL << 17)
226 #define	BIC_PkgTmp	(1ULL << 18)
227 #define	BIC_GFX_rc6	(1ULL << 19)
228 #define	BIC_GFXMHz	(1ULL << 20)
229 #define	BIC_Pkgpc2	(1ULL << 21)
230 #define	BIC_Pkgpc3	(1ULL << 22)
231 #define	BIC_Pkgpc6	(1ULL << 23)
232 #define	BIC_Pkgpc7	(1ULL << 24)
233 #define	BIC_Pkgpc8	(1ULL << 25)
234 #define	BIC_Pkgpc9	(1ULL << 26)
235 #define	BIC_Pkgpc10	(1ULL << 27)
236 #define BIC_CPU_LPI	(1ULL << 28)
237 #define BIC_SYS_LPI	(1ULL << 29)
238 #define	BIC_PkgWatt	(1ULL << 30)
239 #define	BIC_CorWatt	(1ULL << 31)
240 #define	BIC_GFXWatt	(1ULL << 32)
241 #define	BIC_PkgCnt	(1ULL << 33)
242 #define	BIC_RAMWatt	(1ULL << 34)
243 #define	BIC_PKG__	(1ULL << 35)
244 #define	BIC_RAM__	(1ULL << 36)
245 #define	BIC_Pkg_J	(1ULL << 37)
246 #define	BIC_Cor_J	(1ULL << 38)
247 #define	BIC_GFX_J	(1ULL << 39)
248 #define	BIC_RAM_J	(1ULL << 40)
249 #define	BIC_Mod_c6	(1ULL << 41)
250 #define	BIC_Totl_c0	(1ULL << 42)
251 #define	BIC_Any_c0	(1ULL << 43)
252 #define	BIC_GFX_c0	(1ULL << 44)
253 #define	BIC_CPUGFX	(1ULL << 45)
254 #define	BIC_Core	(1ULL << 46)
255 #define	BIC_CPU		(1ULL << 47)
256 #define	BIC_APIC	(1ULL << 48)
257 #define	BIC_X2APIC	(1ULL << 49)
258 #define	BIC_Die		(1ULL << 50)
259 #define	BIC_GFXACTMHz	(1ULL << 51)
260 #define	BIC_IPC		(1ULL << 52)
261 #define	BIC_CORE_THROT_CNT	(1ULL << 53)
262 #define	BIC_UNCORE_MHZ		(1ULL << 54)
263 #define	BIC_SAM_mc6		(1ULL << 55)
264 #define	BIC_SAMMHz		(1ULL << 56)
265 #define	BIC_SAMACTMHz		(1ULL << 57)
266 #define	BIC_Diec6		(1ULL << 58)
267 #define	BIC_SysWatt		(1ULL << 59)
268 #define	BIC_Sys_J		(1ULL << 60)
269 
270 #define BIC_TOPOLOGY (BIC_Package | BIC_Node | BIC_CoreCnt | BIC_PkgCnt | BIC_Core | BIC_CPU | BIC_Die )
271 #define BIC_THERMAL_PWR ( BIC_CoreTmp | BIC_PkgTmp | BIC_PkgWatt | BIC_CorWatt | BIC_GFXWatt | BIC_RAMWatt | BIC_PKG__ | BIC_RAM__)
272 #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)
273 #define BIC_IDLE (BIC_sysfs | BIC_CPU_c1 | BIC_CPU_c3 | BIC_CPU_c6 | BIC_CPU_c7 | BIC_GFX_rc6 | BIC_Pkgpc2 | BIC_Pkgpc3 | BIC_Pkgpc6 | BIC_Pkgpc7 | BIC_Pkgpc8 | BIC_Pkgpc9 | BIC_Pkgpc10 | BIC_CPU_LPI | BIC_SYS_LPI | BIC_Mod_c6 | BIC_Totl_c0 | BIC_Any_c0 | BIC_GFX_c0 | BIC_CPUGFX | BIC_SAM_mc6 | BIC_Diec6)
274 #define BIC_OTHER ( BIC_IRQ | BIC_SMI | BIC_ThreadC | BIC_CoreTmp | BIC_IPC)
275 
276 #define BIC_DISABLED_BY_DEFAULT	(BIC_USEC | BIC_TOD | BIC_APIC | BIC_X2APIC | BIC_SysWatt | BIC_Sys_J)
277 
278 unsigned long long bic_enabled = (0xFFFFFFFFFFFFFFFFULL & ~BIC_DISABLED_BY_DEFAULT);
279 unsigned long long bic_present = BIC_USEC | BIC_TOD | BIC_sysfs | BIC_APIC | BIC_X2APIC;
280 
281 #define DO_BIC(COUNTER_NAME) (bic_enabled & bic_present & COUNTER_NAME)
282 #define DO_BIC_READ(COUNTER_NAME) (bic_present & COUNTER_NAME)
283 #define ENABLE_BIC(COUNTER_NAME) (bic_enabled |= COUNTER_NAME)
284 #define BIC_PRESENT(COUNTER_BIT) (bic_present |= COUNTER_BIT)
285 #define BIC_NOT_PRESENT(COUNTER_BIT) (bic_present &= ~COUNTER_BIT)
286 #define BIC_IS_ENABLED(COUNTER_BIT) (bic_enabled & COUNTER_BIT)
287 
288 /*
289  * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
290  * If you change the values, note they are used both in comparisons
291  * (>= PCL__7) and to index pkg_cstate_limit_strings[].
292  */
293 #define PCLUKN 0		/* Unknown */
294 #define PCLRSV 1		/* Reserved */
295 #define PCL__0 2		/* PC0 */
296 #define PCL__1 3		/* PC1 */
297 #define PCL__2 4		/* PC2 */
298 #define PCL__3 5		/* PC3 */
299 #define PCL__4 6		/* PC4 */
300 #define PCL__6 7		/* PC6 */
301 #define PCL_6N 8		/* PC6 No Retention */
302 #define PCL_6R 9		/* PC6 Retention */
303 #define PCL__7 10		/* PC7 */
304 #define PCL_7S 11		/* PC7 Shrink */
305 #define PCL__8 12		/* PC8 */
306 #define PCL__9 13		/* PC9 */
307 #define PCL_10 14		/* PC10 */
308 #define PCLUNL 15		/* Unlimited */
309 
310 struct amperf_group_fd;
311 
312 char *proc_stat = "/proc/stat";
313 FILE *outf;
314 int *fd_percpu;
315 int *fd_instr_count_percpu;
316 struct timeval interval_tv = { 5, 0 };
317 struct timespec interval_ts = { 5, 0 };
318 
319 unsigned int num_iterations;
320 unsigned int header_iterations;
321 unsigned int debug;
322 unsigned int quiet;
323 unsigned int shown;
324 unsigned int sums_need_wide_columns;
325 unsigned int rapl_joules;
326 unsigned int summary_only;
327 unsigned int list_header_only;
328 unsigned int dump_only;
329 unsigned int has_aperf;
330 unsigned int has_aperf_access;
331 unsigned int has_epb;
332 unsigned int has_turbo;
333 unsigned int is_hybrid;
334 unsigned int units = 1000000;	/* MHz etc */
335 unsigned int genuine_intel;
336 unsigned int authentic_amd;
337 unsigned int hygon_genuine;
338 unsigned int max_level, max_extended_level;
339 unsigned int has_invariant_tsc;
340 unsigned int aperf_mperf_multiplier = 1;
341 double bclk;
342 double base_hz;
343 unsigned int has_base_hz;
344 double tsc_tweak = 1.0;
345 unsigned int show_pkg_only;
346 unsigned int show_core_only;
347 char *output_buffer, *outp;
348 unsigned int do_dts;
349 unsigned int do_ptm;
350 unsigned int do_ipc;
351 unsigned long long cpuidle_cur_cpu_lpi_us;
352 unsigned long long cpuidle_cur_sys_lpi_us;
353 unsigned int tj_max;
354 unsigned int tj_max_override;
355 double rapl_power_units, rapl_time_units;
356 double rapl_dram_energy_units, rapl_energy_units;
357 double rapl_joule_counter_range;
358 unsigned int crystal_hz;
359 unsigned long long tsc_hz;
360 int base_cpu;
361 unsigned int has_hwp;		/* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
362 			/* IA32_HWP_REQUEST, IA32_HWP_STATUS */
363 unsigned int has_hwp_notify;	/* IA32_HWP_INTERRUPT */
364 unsigned int has_hwp_activity_window;	/* IA32_HWP_REQUEST[bits 41:32] */
365 unsigned int has_hwp_epp;	/* IA32_HWP_REQUEST[bits 31:24] */
366 unsigned int has_hwp_pkg;	/* IA32_HWP_REQUEST_PKG */
367 unsigned int first_counter_read = 1;
368 int ignore_stdin;
369 bool no_msr;
370 bool no_perf;
371 
372 enum gfx_sysfs_idx {
373 	GFX_rc6,
374 	GFX_MHz,
375 	GFX_ACTMHz,
376 	SAM_mc6,
377 	SAM_MHz,
378 	SAM_ACTMHz,
379 	GFX_MAX
380 };
381 
382 struct gfx_sysfs_info {
383 	FILE *fp;
384 	unsigned int val;
385 	unsigned long long val_ull;
386 };
387 
388 static struct gfx_sysfs_info gfx_info[GFX_MAX];
389 
390 int get_msr(int cpu, off_t offset, unsigned long long *msr);
391 int add_counter(unsigned int msr_num, char *path, char *name,
392 		unsigned int width, enum counter_scope scope,
393 		enum counter_type type, enum counter_format format, int flags, int package_num);
394 
395 /* Model specific support Start */
396 
397 /* List of features that may diverge among different platforms */
398 struct platform_features {
399 	bool has_msr_misc_feature_control;	/* MSR_MISC_FEATURE_CONTROL */
400 	bool has_msr_misc_pwr_mgmt;	/* MSR_MISC_PWR_MGMT */
401 	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 */
402 	bool has_config_tdp;	/* MSR_CONFIG_TDP_NOMINAL/LEVEL_1/LEVEL_2/CONTROL, MSR_TURBO_ACTIVATION_RATIO */
403 	int bclk_freq;		/* CPU base clock */
404 	int crystal_freq;	/* Crystal clock to use when not available from CPUID.15 */
405 	int supported_cstates;	/* Core cstates and Package cstates supported */
406 	int cst_limit;		/* MSR_PKG_CST_CONFIG_CONTROL */
407 	bool has_cst_auto_convension;	/* AUTOMATIC_CSTATE_CONVERSION bit in MSR_PKG_CST_CONFIG_CONTROL */
408 	bool has_irtl_msrs;	/* MSR_PKGC3/PKGC6/PKGC7/PKGC8/PKGC9/PKGC10_IRTL */
409 	bool has_msr_core_c1_res;	/* MSR_CORE_C1_RES */
410 	bool has_msr_module_c6_res_ms;	/* MSR_MODULE_C6_RES_MS */
411 	bool has_msr_c6_demotion_policy_config;	/* MSR_CC6_DEMOTION_POLICY_CONFIG/MSR_MC6_DEMOTION_POLICY_CONFIG */
412 	bool has_msr_atom_pkg_c6_residency;	/* MSR_ATOM_PKG_C6_RESIDENCY */
413 	bool has_msr_knl_core_c6_residency;	/* MSR_KNL_CORE_C6_RESIDENCY */
414 	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 */
415 	bool has_cst_prewake_bit;	/* Cstate prewake bit in MSR_IA32_POWER_CTL */
416 	int trl_msrs;		/* MSR_TURBO_RATIO_LIMIT/LIMIT1/LIMIT2/SECONDARY, Atom TRL MSRs */
417 	int plr_msrs;		/* MSR_CORE/GFX/RING_PERF_LIMIT_REASONS */
418 	int rapl_msrs;		/* RAPL PKG/DRAM/CORE/GFX MSRs, AMD RAPL MSRs */
419 	bool has_per_core_rapl;	/* Indicates cores energy collection is per-core, not per-package. AMD specific for now */
420 	bool has_rapl_divisor;	/* Divisor for Energy unit raw value from MSR_RAPL_POWER_UNIT */
421 	bool has_fixed_rapl_unit;	/* Fixed Energy Unit used for DRAM RAPL Domain */
422 	int rapl_quirk_tdp;	/* Hardcoded TDP value when cannot be retrieved from hardware */
423 	int tcc_offset_bits;	/* TCC Offset bits in MSR_IA32_TEMPERATURE_TARGET */
424 	bool enable_tsc_tweak;	/* Use CPU Base freq instead of TSC freq for aperf/mperf counter */
425 	bool need_perf_multiplier;	/* mperf/aperf multiplier */
426 };
427 
428 struct platform_data {
429 	unsigned int vfm;
430 	const struct platform_features *features;
431 };
432 
433 /* For BCLK */
434 enum bclk_freq {
435 	BCLK_100MHZ = 1,
436 	BCLK_133MHZ,
437 	BCLK_SLV,
438 };
439 
440 #define SLM_BCLK_FREQS 5
441 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0 };
442 
slm_bclk(void)443 double slm_bclk(void)
444 {
445 	unsigned long long msr = 3;
446 	unsigned int i;
447 	double freq;
448 
449 	if (get_msr(base_cpu, MSR_FSB_FREQ, &msr))
450 		fprintf(outf, "SLM BCLK: unknown\n");
451 
452 	i = msr & 0xf;
453 	if (i >= SLM_BCLK_FREQS) {
454 		fprintf(outf, "SLM BCLK[%d] invalid\n", i);
455 		i = 3;
456 	}
457 	freq = slm_freq_table[i];
458 
459 	if (!quiet)
460 		fprintf(outf, "SLM BCLK: %.1f Mhz\n", freq);
461 
462 	return freq;
463 }
464 
465 /* For Package cstate limit */
466 enum package_cstate_limit {
467 	CST_LIMIT_NHM = 1,
468 	CST_LIMIT_SNB,
469 	CST_LIMIT_HSW,
470 	CST_LIMIT_SKX,
471 	CST_LIMIT_ICX,
472 	CST_LIMIT_SLV,
473 	CST_LIMIT_AMT,
474 	CST_LIMIT_KNL,
475 	CST_LIMIT_GMT,
476 };
477 
478 /* For Turbo Ratio Limit MSRs */
479 enum turbo_ratio_limit_msrs {
480 	TRL_BASE = BIT(0),
481 	TRL_LIMIT1 = BIT(1),
482 	TRL_LIMIT2 = BIT(2),
483 	TRL_ATOM = BIT(3),
484 	TRL_KNL = BIT(4),
485 	TRL_CORECOUNT = BIT(5),
486 };
487 
488 /* For Perf Limit Reason MSRs */
489 enum perf_limit_reason_msrs {
490 	PLR_CORE = BIT(0),
491 	PLR_GFX = BIT(1),
492 	PLR_RING = BIT(2),
493 };
494 
495 /* For RAPL MSRs */
496 enum rapl_msrs {
497 	RAPL_PKG_POWER_LIMIT = BIT(0),	/* 0x610 MSR_PKG_POWER_LIMIT */
498 	RAPL_PKG_ENERGY_STATUS = BIT(1),	/* 0x611 MSR_PKG_ENERGY_STATUS */
499 	RAPL_PKG_PERF_STATUS = BIT(2),	/* 0x613 MSR_PKG_PERF_STATUS */
500 	RAPL_PKG_POWER_INFO = BIT(3),	/* 0x614 MSR_PKG_POWER_INFO */
501 	RAPL_DRAM_POWER_LIMIT = BIT(4),	/* 0x618 MSR_DRAM_POWER_LIMIT */
502 	RAPL_DRAM_ENERGY_STATUS = BIT(5),	/* 0x619 MSR_DRAM_ENERGY_STATUS */
503 	RAPL_DRAM_PERF_STATUS = BIT(6),	/* 0x61b MSR_DRAM_PERF_STATUS */
504 	RAPL_DRAM_POWER_INFO = BIT(7),	/* 0x61c MSR_DRAM_POWER_INFO */
505 	RAPL_CORE_POWER_LIMIT = BIT(8),	/* 0x638 MSR_PP0_POWER_LIMIT */
506 	RAPL_CORE_ENERGY_STATUS = BIT(9),	/* 0x639 MSR_PP0_ENERGY_STATUS */
507 	RAPL_CORE_POLICY = BIT(10),	/* 0x63a MSR_PP0_POLICY */
508 	RAPL_GFX_POWER_LIMIT = BIT(11),	/* 0x640 MSR_PP1_POWER_LIMIT */
509 	RAPL_GFX_ENERGY_STATUS = BIT(12),	/* 0x641 MSR_PP1_ENERGY_STATUS */
510 	RAPL_GFX_POLICY = BIT(13),	/* 0x642 MSR_PP1_POLICY */
511 	RAPL_AMD_PWR_UNIT = BIT(14),	/* 0xc0010299 MSR_AMD_RAPL_POWER_UNIT */
512 	RAPL_AMD_CORE_ENERGY_STAT = BIT(15),	/* 0xc001029a MSR_AMD_CORE_ENERGY_STATUS */
513 	RAPL_AMD_PKG_ENERGY_STAT = BIT(16),	/* 0xc001029b MSR_AMD_PKG_ENERGY_STATUS */
514 	RAPL_PLATFORM_ENERGY_LIMIT = BIT(17),	/* 0x64c MSR_PLATFORM_ENERGY_LIMIT */
515 	RAPL_PLATFORM_ENERGY_STATUS = BIT(18),	/* 0x64d MSR_PLATFORM_ENERGY_STATUS */
516 };
517 
518 #define RAPL_PKG	(RAPL_PKG_ENERGY_STATUS | RAPL_PKG_POWER_LIMIT)
519 #define RAPL_DRAM	(RAPL_DRAM_ENERGY_STATUS | RAPL_DRAM_POWER_LIMIT)
520 #define RAPL_CORE	(RAPL_CORE_ENERGY_STATUS | RAPL_CORE_POWER_LIMIT)
521 #define RAPL_GFX	(RAPL_GFX_POWER_LIMIT | RAPL_GFX_ENERGY_STATUS)
522 #define RAPL_PSYS	(RAPL_PLATFORM_ENERGY_STATUS | RAPL_PLATFORM_ENERGY_LIMIT)
523 
524 #define RAPL_PKG_ALL	(RAPL_PKG | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO)
525 #define RAPL_DRAM_ALL	(RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_DRAM_POWER_INFO)
526 #define RAPL_CORE_ALL	(RAPL_CORE | RAPL_CORE_POLICY)
527 #define RAPL_GFX_ALL	(RAPL_GFX | RAPL_GFX_POLIGY)
528 
529 #define RAPL_AMD_F17H	(RAPL_AMD_PWR_UNIT | RAPL_AMD_CORE_ENERGY_STAT | RAPL_AMD_PKG_ENERGY_STAT)
530 
531 /* For Cstates */
532 enum cstates {
533 	CC1 = BIT(0),
534 	CC3 = BIT(1),
535 	CC6 = BIT(2),
536 	CC7 = BIT(3),
537 	PC2 = BIT(4),
538 	PC3 = BIT(5),
539 	PC6 = BIT(6),
540 	PC7 = BIT(7),
541 	PC8 = BIT(8),
542 	PC9 = BIT(9),
543 	PC10 = BIT(10),
544 };
545 
546 static const struct platform_features nhm_features = {
547 	.has_msr_misc_pwr_mgmt = 1,
548 	.has_nhm_msrs = 1,
549 	.bclk_freq = BCLK_133MHZ,
550 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
551 	.cst_limit = CST_LIMIT_NHM,
552 	.trl_msrs = TRL_BASE,
553 };
554 
555 static const struct platform_features nhx_features = {
556 	.has_msr_misc_pwr_mgmt = 1,
557 	.has_nhm_msrs = 1,
558 	.bclk_freq = BCLK_133MHZ,
559 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
560 	.cst_limit = CST_LIMIT_NHM,
561 };
562 
563 static const struct platform_features snb_features = {
564 	.has_msr_misc_feature_control = 1,
565 	.has_msr_misc_pwr_mgmt = 1,
566 	.has_nhm_msrs = 1,
567 	.bclk_freq = BCLK_100MHZ,
568 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
569 	.cst_limit = CST_LIMIT_SNB,
570 	.has_irtl_msrs = 1,
571 	.trl_msrs = TRL_BASE,
572 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
573 };
574 
575 static const struct platform_features snx_features = {
576 	.has_msr_misc_feature_control = 1,
577 	.has_msr_misc_pwr_mgmt = 1,
578 	.has_nhm_msrs = 1,
579 	.bclk_freq = BCLK_100MHZ,
580 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
581 	.cst_limit = CST_LIMIT_SNB,
582 	.has_irtl_msrs = 1,
583 	.trl_msrs = TRL_BASE,
584 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM_ALL,
585 };
586 
587 static const struct platform_features ivb_features = {
588 	.has_msr_misc_feature_control = 1,
589 	.has_msr_misc_pwr_mgmt = 1,
590 	.has_nhm_msrs = 1,
591 	.has_config_tdp = 1,
592 	.bclk_freq = BCLK_100MHZ,
593 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
594 	.cst_limit = CST_LIMIT_SNB,
595 	.has_irtl_msrs = 1,
596 	.trl_msrs = TRL_BASE,
597 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
598 };
599 
600 static const struct platform_features ivx_features = {
601 	.has_msr_misc_feature_control = 1,
602 	.has_msr_misc_pwr_mgmt = 1,
603 	.has_nhm_msrs = 1,
604 	.bclk_freq = BCLK_100MHZ,
605 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
606 	.cst_limit = CST_LIMIT_SNB,
607 	.has_irtl_msrs = 1,
608 	.trl_msrs = TRL_BASE | TRL_LIMIT1,
609 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM_ALL,
610 };
611 
612 static const struct platform_features hsw_features = {
613 	.has_msr_misc_feature_control = 1,
614 	.has_msr_misc_pwr_mgmt = 1,
615 	.has_nhm_msrs = 1,
616 	.has_config_tdp = 1,
617 	.bclk_freq = BCLK_100MHZ,
618 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
619 	.cst_limit = CST_LIMIT_HSW,
620 	.has_irtl_msrs = 1,
621 	.trl_msrs = TRL_BASE,
622 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
623 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
624 };
625 
626 static const struct platform_features hsx_features = {
627 	.has_msr_misc_feature_control = 1,
628 	.has_msr_misc_pwr_mgmt = 1,
629 	.has_nhm_msrs = 1,
630 	.has_config_tdp = 1,
631 	.bclk_freq = BCLK_100MHZ,
632 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
633 	.cst_limit = CST_LIMIT_HSW,
634 	.has_irtl_msrs = 1,
635 	.trl_msrs = TRL_BASE | TRL_LIMIT1 | TRL_LIMIT2,
636 	.plr_msrs = PLR_CORE | PLR_RING,
637 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
638 	.has_fixed_rapl_unit = 1,
639 };
640 
641 static const struct platform_features hswl_features = {
642 	.has_msr_misc_feature_control = 1,
643 	.has_msr_misc_pwr_mgmt = 1,
644 	.has_nhm_msrs = 1,
645 	.has_config_tdp = 1,
646 	.bclk_freq = BCLK_100MHZ,
647 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
648 	.cst_limit = CST_LIMIT_HSW,
649 	.has_irtl_msrs = 1,
650 	.trl_msrs = TRL_BASE,
651 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
652 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
653 };
654 
655 static const struct platform_features hswg_features = {
656 	.has_msr_misc_feature_control = 1,
657 	.has_msr_misc_pwr_mgmt = 1,
658 	.has_nhm_msrs = 1,
659 	.has_config_tdp = 1,
660 	.bclk_freq = BCLK_100MHZ,
661 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
662 	.cst_limit = CST_LIMIT_HSW,
663 	.has_irtl_msrs = 1,
664 	.trl_msrs = TRL_BASE,
665 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
666 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
667 };
668 
669 static const struct platform_features bdw_features = {
670 	.has_msr_misc_feature_control = 1,
671 	.has_msr_misc_pwr_mgmt = 1,
672 	.has_nhm_msrs = 1,
673 	.has_config_tdp = 1,
674 	.bclk_freq = BCLK_100MHZ,
675 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
676 	.cst_limit = CST_LIMIT_HSW,
677 	.has_irtl_msrs = 1,
678 	.trl_msrs = TRL_BASE,
679 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
680 };
681 
682 static const struct platform_features bdwg_features = {
683 	.has_msr_misc_feature_control = 1,
684 	.has_msr_misc_pwr_mgmt = 1,
685 	.has_nhm_msrs = 1,
686 	.has_config_tdp = 1,
687 	.bclk_freq = BCLK_100MHZ,
688 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
689 	.cst_limit = CST_LIMIT_HSW,
690 	.has_irtl_msrs = 1,
691 	.trl_msrs = TRL_BASE,
692 	.rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
693 };
694 
695 static const struct platform_features bdx_features = {
696 	.has_msr_misc_feature_control = 1,
697 	.has_msr_misc_pwr_mgmt = 1,
698 	.has_nhm_msrs = 1,
699 	.has_config_tdp = 1,
700 	.bclk_freq = BCLK_100MHZ,
701 	.supported_cstates = CC1 | CC3 | CC6 | PC2 | PC3 | PC6,
702 	.cst_limit = CST_LIMIT_HSW,
703 	.has_irtl_msrs = 1,
704 	.has_cst_auto_convension = 1,
705 	.trl_msrs = TRL_BASE,
706 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
707 	.has_fixed_rapl_unit = 1,
708 };
709 
710 static const struct platform_features skl_features = {
711 	.has_msr_misc_feature_control = 1,
712 	.has_msr_misc_pwr_mgmt = 1,
713 	.has_nhm_msrs = 1,
714 	.has_config_tdp = 1,
715 	.bclk_freq = BCLK_100MHZ,
716 	.crystal_freq = 24000000,
717 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
718 	.cst_limit = CST_LIMIT_HSW,
719 	.has_irtl_msrs = 1,
720 	.has_ext_cst_msrs = 1,
721 	.trl_msrs = TRL_BASE,
722 	.tcc_offset_bits = 6,
723 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX | RAPL_PSYS,
724 	.enable_tsc_tweak = 1,
725 };
726 
727 static const struct platform_features cnl_features = {
728 	.has_msr_misc_feature_control = 1,
729 	.has_msr_misc_pwr_mgmt = 1,
730 	.has_nhm_msrs = 1,
731 	.has_config_tdp = 1,
732 	.bclk_freq = BCLK_100MHZ,
733 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
734 	.cst_limit = CST_LIMIT_HSW,
735 	.has_irtl_msrs = 1,
736 	.has_msr_core_c1_res = 1,
737 	.has_ext_cst_msrs = 1,
738 	.trl_msrs = TRL_BASE,
739 	.tcc_offset_bits = 6,
740 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX | RAPL_PSYS,
741 	.enable_tsc_tweak = 1,
742 };
743 
744 /* Copied from cnl_features, with PC7/PC9 removed */
745 static const struct platform_features adl_features = {
746 	.has_msr_misc_feature_control	= cnl_features.has_msr_misc_feature_control,
747 	.has_msr_misc_pwr_mgmt		= cnl_features.has_msr_misc_pwr_mgmt,
748 	.has_nhm_msrs			= cnl_features.has_nhm_msrs,
749 	.has_config_tdp			= cnl_features.has_config_tdp,
750 	.bclk_freq			= cnl_features.bclk_freq,
751 	.supported_cstates		= CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC8 | PC10,
752 	.cst_limit			= cnl_features.cst_limit,
753 	.has_irtl_msrs			= cnl_features.has_irtl_msrs,
754 	.has_msr_core_c1_res		= cnl_features.has_msr_core_c1_res,
755 	.has_ext_cst_msrs		= cnl_features.has_ext_cst_msrs,
756 	.trl_msrs			= cnl_features.trl_msrs,
757 	.tcc_offset_bits		= cnl_features.tcc_offset_bits,
758 	.rapl_msrs			= cnl_features.rapl_msrs,
759 	.enable_tsc_tweak		= cnl_features.enable_tsc_tweak,
760 };
761 
762 /* Copied from adl_features, with PC3/PC8 removed */
763 static const struct platform_features lnl_features = {
764 	.has_msr_misc_feature_control	= adl_features.has_msr_misc_feature_control,
765 	.has_msr_misc_pwr_mgmt		= adl_features.has_msr_misc_pwr_mgmt,
766 	.has_nhm_msrs			= adl_features.has_nhm_msrs,
767 	.has_config_tdp			= adl_features.has_config_tdp,
768 	.bclk_freq			= adl_features.bclk_freq,
769 	.supported_cstates		= CC1 | CC6 | CC7 | PC2 | PC6 | PC10,
770 	.cst_limit			= adl_features.cst_limit,
771 	.has_irtl_msrs			= adl_features.has_irtl_msrs,
772 	.has_msr_core_c1_res		= adl_features.has_msr_core_c1_res,
773 	.has_ext_cst_msrs		= adl_features.has_ext_cst_msrs,
774 	.trl_msrs			= adl_features.trl_msrs,
775 	.tcc_offset_bits		= adl_features.tcc_offset_bits,
776 	.rapl_msrs			= adl_features.rapl_msrs,
777 	.enable_tsc_tweak		= adl_features.enable_tsc_tweak,
778 };
779 
780 static const struct platform_features skx_features = {
781 	.has_msr_misc_feature_control = 1,
782 	.has_msr_misc_pwr_mgmt = 1,
783 	.has_nhm_msrs = 1,
784 	.has_config_tdp = 1,
785 	.bclk_freq = BCLK_100MHZ,
786 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
787 	.cst_limit = CST_LIMIT_SKX,
788 	.has_irtl_msrs = 1,
789 	.has_cst_auto_convension = 1,
790 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
791 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
792 	.has_fixed_rapl_unit = 1,
793 };
794 
795 static const struct platform_features icx_features = {
796 	.has_msr_misc_feature_control = 1,
797 	.has_msr_misc_pwr_mgmt = 1,
798 	.has_nhm_msrs = 1,
799 	.has_config_tdp = 1,
800 	.bclk_freq = BCLK_100MHZ,
801 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
802 	.cst_limit = CST_LIMIT_ICX,
803 	.has_msr_core_c1_res = 1,
804 	.has_irtl_msrs = 1,
805 	.has_cst_prewake_bit = 1,
806 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
807 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
808 	.has_fixed_rapl_unit = 1,
809 };
810 
811 static const struct platform_features spr_features = {
812 	.has_msr_misc_feature_control = 1,
813 	.has_msr_misc_pwr_mgmt = 1,
814 	.has_nhm_msrs = 1,
815 	.has_config_tdp = 1,
816 	.bclk_freq = BCLK_100MHZ,
817 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
818 	.cst_limit = CST_LIMIT_SKX,
819 	.has_msr_core_c1_res = 1,
820 	.has_irtl_msrs = 1,
821 	.has_cst_prewake_bit = 1,
822 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
823 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
824 };
825 
826 static const struct platform_features srf_features = {
827 	.has_msr_misc_feature_control = 1,
828 	.has_msr_misc_pwr_mgmt = 1,
829 	.has_nhm_msrs = 1,
830 	.has_config_tdp = 1,
831 	.bclk_freq = BCLK_100MHZ,
832 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
833 	.cst_limit = CST_LIMIT_SKX,
834 	.has_msr_core_c1_res = 1,
835 	.has_msr_module_c6_res_ms = 1,
836 	.has_irtl_msrs = 1,
837 	.has_cst_prewake_bit = 1,
838 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
839 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
840 };
841 
842 static const struct platform_features grr_features = {
843 	.has_msr_misc_feature_control = 1,
844 	.has_msr_misc_pwr_mgmt = 1,
845 	.has_nhm_msrs = 1,
846 	.has_config_tdp = 1,
847 	.bclk_freq = BCLK_100MHZ,
848 	.supported_cstates = CC1 | CC6,
849 	.cst_limit = CST_LIMIT_SKX,
850 	.has_msr_core_c1_res = 1,
851 	.has_msr_module_c6_res_ms = 1,
852 	.has_irtl_msrs = 1,
853 	.has_cst_prewake_bit = 1,
854 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
855 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
856 };
857 
858 static const struct platform_features slv_features = {
859 	.has_nhm_msrs = 1,
860 	.bclk_freq = BCLK_SLV,
861 	.supported_cstates = CC1 | CC6 | PC6,
862 	.cst_limit = CST_LIMIT_SLV,
863 	.has_msr_core_c1_res = 1,
864 	.has_msr_module_c6_res_ms = 1,
865 	.has_msr_c6_demotion_policy_config = 1,
866 	.has_msr_atom_pkg_c6_residency = 1,
867 	.trl_msrs = TRL_ATOM,
868 	.rapl_msrs = RAPL_PKG | RAPL_CORE,
869 	.has_rapl_divisor = 1,
870 	.rapl_quirk_tdp = 30,
871 };
872 
873 static const struct platform_features slvd_features = {
874 	.has_msr_misc_pwr_mgmt = 1,
875 	.has_nhm_msrs = 1,
876 	.bclk_freq = BCLK_SLV,
877 	.supported_cstates = CC1 | CC6 | PC3 | PC6,
878 	.cst_limit = CST_LIMIT_SLV,
879 	.has_msr_atom_pkg_c6_residency = 1,
880 	.trl_msrs = TRL_BASE,
881 	.rapl_msrs = RAPL_PKG | RAPL_CORE,
882 	.rapl_quirk_tdp = 30,
883 };
884 
885 static const struct platform_features amt_features = {
886 	.has_nhm_msrs = 1,
887 	.bclk_freq = BCLK_133MHZ,
888 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
889 	.cst_limit = CST_LIMIT_AMT,
890 	.trl_msrs = TRL_BASE,
891 };
892 
893 static const struct platform_features gmt_features = {
894 	.has_msr_misc_pwr_mgmt = 1,
895 	.has_nhm_msrs = 1,
896 	.bclk_freq = BCLK_100MHZ,
897 	.crystal_freq = 19200000,
898 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
899 	.cst_limit = CST_LIMIT_GMT,
900 	.has_irtl_msrs = 1,
901 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
902 	.rapl_msrs = RAPL_PKG | RAPL_PKG_POWER_INFO,
903 };
904 
905 static const struct platform_features gmtd_features = {
906 	.has_msr_misc_pwr_mgmt = 1,
907 	.has_nhm_msrs = 1,
908 	.bclk_freq = BCLK_100MHZ,
909 	.crystal_freq = 25000000,
910 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
911 	.cst_limit = CST_LIMIT_GMT,
912 	.has_irtl_msrs = 1,
913 	.has_msr_core_c1_res = 1,
914 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
915 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_CORE_ENERGY_STATUS,
916 };
917 
918 static const struct platform_features gmtp_features = {
919 	.has_msr_misc_pwr_mgmt = 1,
920 	.has_nhm_msrs = 1,
921 	.bclk_freq = BCLK_100MHZ,
922 	.crystal_freq = 19200000,
923 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
924 	.cst_limit = CST_LIMIT_GMT,
925 	.has_irtl_msrs = 1,
926 	.trl_msrs = TRL_BASE,
927 	.rapl_msrs = RAPL_PKG | RAPL_PKG_POWER_INFO,
928 };
929 
930 static const struct platform_features tmt_features = {
931 	.has_msr_misc_pwr_mgmt = 1,
932 	.has_nhm_msrs = 1,
933 	.bclk_freq = BCLK_100MHZ,
934 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
935 	.cst_limit = CST_LIMIT_GMT,
936 	.has_irtl_msrs = 1,
937 	.trl_msrs = TRL_BASE,
938 	.rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX,
939 	.enable_tsc_tweak = 1,
940 };
941 
942 static const struct platform_features tmtd_features = {
943 	.has_msr_misc_pwr_mgmt = 1,
944 	.has_nhm_msrs = 1,
945 	.bclk_freq = BCLK_100MHZ,
946 	.supported_cstates = CC1 | CC6,
947 	.cst_limit = CST_LIMIT_GMT,
948 	.has_irtl_msrs = 1,
949 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
950 	.rapl_msrs = RAPL_PKG_ALL,
951 };
952 
953 static const struct platform_features knl_features = {
954 	.has_msr_misc_pwr_mgmt = 1,
955 	.has_nhm_msrs = 1,
956 	.has_config_tdp = 1,
957 	.bclk_freq = BCLK_100MHZ,
958 	.supported_cstates = CC1 | CC6 | PC3 | PC6,
959 	.cst_limit = CST_LIMIT_KNL,
960 	.has_msr_knl_core_c6_residency = 1,
961 	.trl_msrs = TRL_KNL,
962 	.rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
963 	.has_fixed_rapl_unit = 1,
964 	.need_perf_multiplier = 1,
965 };
966 
967 static const struct platform_features default_features = {
968 };
969 
970 static const struct platform_features amd_features_with_rapl = {
971 	.rapl_msrs = RAPL_AMD_F17H,
972 	.has_per_core_rapl = 1,
973 	.rapl_quirk_tdp = 280,	/* This is the max stock TDP of HEDT/Server Fam17h+ chips */
974 };
975 
976 static const struct platform_data turbostat_pdata[] = {
977 	{ INTEL_NEHALEM, &nhm_features },
978 	{ INTEL_NEHALEM_G, &nhm_features },
979 	{ INTEL_NEHALEM_EP, &nhm_features },
980 	{ INTEL_NEHALEM_EX, &nhx_features },
981 	{ INTEL_WESTMERE, &nhm_features },
982 	{ INTEL_WESTMERE_EP, &nhm_features },
983 	{ INTEL_WESTMERE_EX, &nhx_features },
984 	{ INTEL_SANDYBRIDGE, &snb_features },
985 	{ INTEL_SANDYBRIDGE_X, &snx_features },
986 	{ INTEL_IVYBRIDGE, &ivb_features },
987 	{ INTEL_IVYBRIDGE_X, &ivx_features },
988 	{ INTEL_HASWELL, &hsw_features },
989 	{ INTEL_HASWELL_X, &hsx_features },
990 	{ INTEL_HASWELL_L, &hswl_features },
991 	{ INTEL_HASWELL_G, &hswg_features },
992 	{ INTEL_BROADWELL, &bdw_features },
993 	{ INTEL_BROADWELL_G, &bdwg_features },
994 	{ INTEL_BROADWELL_X, &bdx_features },
995 	{ INTEL_BROADWELL_D, &bdx_features },
996 	{ INTEL_SKYLAKE_L, &skl_features },
997 	{ INTEL_SKYLAKE, &skl_features },
998 	{ INTEL_SKYLAKE_X, &skx_features },
999 	{ INTEL_KABYLAKE_L, &skl_features },
1000 	{ INTEL_KABYLAKE, &skl_features },
1001 	{ INTEL_COMETLAKE, &skl_features },
1002 	{ INTEL_COMETLAKE_L, &skl_features },
1003 	{ INTEL_CANNONLAKE_L, &cnl_features },
1004 	{ INTEL_ICELAKE_X, &icx_features },
1005 	{ INTEL_ICELAKE_D, &icx_features },
1006 	{ INTEL_ICELAKE_L, &cnl_features },
1007 	{ INTEL_ICELAKE_NNPI, &cnl_features },
1008 	{ INTEL_ROCKETLAKE, &cnl_features },
1009 	{ INTEL_TIGERLAKE_L, &cnl_features },
1010 	{ INTEL_TIGERLAKE, &cnl_features },
1011 	{ INTEL_SAPPHIRERAPIDS_X, &spr_features },
1012 	{ INTEL_EMERALDRAPIDS_X, &spr_features },
1013 	{ INTEL_GRANITERAPIDS_X, &spr_features },
1014 	{ INTEL_GRANITERAPIDS_D, &spr_features },
1015 	{ INTEL_LAKEFIELD, &cnl_features },
1016 	{ INTEL_ALDERLAKE, &adl_features },
1017 	{ INTEL_ALDERLAKE_L, &adl_features },
1018 	{ INTEL_RAPTORLAKE, &adl_features },
1019 	{ INTEL_RAPTORLAKE_P, &adl_features },
1020 	{ INTEL_RAPTORLAKE_S, &adl_features },
1021 	{ INTEL_METEORLAKE, &adl_features },
1022 	{ INTEL_METEORLAKE_L, &adl_features },
1023 	{ INTEL_ARROWLAKE_H, &adl_features },
1024 	{ INTEL_ARROWLAKE_U, &adl_features },
1025 	{ INTEL_ARROWLAKE, &adl_features },
1026 	{ INTEL_LUNARLAKE_M, &lnl_features },
1027 	{ INTEL_ATOM_SILVERMONT, &slv_features },
1028 	{ INTEL_ATOM_SILVERMONT_D, &slvd_features },
1029 	{ INTEL_ATOM_AIRMONT, &amt_features },
1030 	{ INTEL_ATOM_GOLDMONT, &gmt_features },
1031 	{ INTEL_ATOM_GOLDMONT_D, &gmtd_features },
1032 	{ INTEL_ATOM_GOLDMONT_PLUS, &gmtp_features },
1033 	{ INTEL_ATOM_TREMONT_D, &tmtd_features },
1034 	{ INTEL_ATOM_TREMONT, &tmt_features },
1035 	{ INTEL_ATOM_TREMONT_L, &tmt_features },
1036 	{ INTEL_ATOM_GRACEMONT, &adl_features },
1037 	{ INTEL_ATOM_CRESTMONT_X, &srf_features },
1038 	{ INTEL_ATOM_CRESTMONT, &grr_features },
1039 	{ INTEL_XEON_PHI_KNL, &knl_features },
1040 	{ INTEL_XEON_PHI_KNM, &knl_features },
1041 	/*
1042 	 * Missing support for
1043 	 * INTEL_ICELAKE
1044 	 * INTEL_ATOM_SILVERMONT_MID
1045 	 * INTEL_ATOM_AIRMONT_MID
1046 	 * INTEL_ATOM_AIRMONT_NP
1047 	 */
1048 	{ 0, NULL },
1049 };
1050 
1051 static const struct platform_features *platform;
1052 
probe_platform_features(unsigned int family,unsigned int model)1053 void probe_platform_features(unsigned int family, unsigned int model)
1054 {
1055 	int i;
1056 
1057 	platform = &default_features;
1058 
1059 	if (authentic_amd || hygon_genuine) {
1060 		if (max_extended_level >= 0x80000007) {
1061 			unsigned int eax, ebx, ecx, edx;
1062 
1063 			__cpuid(0x80000007, eax, ebx, ecx, edx);
1064 			/* RAPL (Fam 17h+) */
1065 			if ((edx & (1 << 14)) && family >= 0x17)
1066 				platform = &amd_features_with_rapl;
1067 		}
1068 		return;
1069 	}
1070 
1071 	if (!genuine_intel)
1072 		return;
1073 
1074 	for (i = 0; turbostat_pdata[i].features; i++) {
1075 		if (VFM_FAMILY(turbostat_pdata[i].vfm) == family && VFM_MODEL(turbostat_pdata[i].vfm) == model) {
1076 			platform = turbostat_pdata[i].features;
1077 			return;
1078 		}
1079 	}
1080 }
1081 
1082 /* Model specific support End */
1083 
1084 #define	TJMAX_DEFAULT	100
1085 
1086 /* MSRs that are not yet in the kernel-provided header. */
1087 #define MSR_RAPL_PWR_UNIT	0xc0010299
1088 #define MSR_CORE_ENERGY_STAT	0xc001029a
1089 #define MSR_PKG_ENERGY_STAT	0xc001029b
1090 
1091 #define MAX(a, b) ((a) > (b) ? (a) : (b))
1092 
1093 int backwards_count;
1094 char *progname;
1095 
1096 #define CPU_SUBSET_MAXCPUS	1024	/* need to use before probe... */
1097 cpu_set_t *cpu_present_set, *cpu_effective_set, *cpu_allowed_set, *cpu_affinity_set, *cpu_subset;
1098 size_t cpu_present_setsize, cpu_effective_setsize, cpu_allowed_setsize, cpu_affinity_setsize, cpu_subset_size;
1099 #define MAX_ADDED_THREAD_COUNTERS 24
1100 #define MAX_ADDED_CORE_COUNTERS 8
1101 #define MAX_ADDED_PACKAGE_COUNTERS 16
1102 #define PMT_MAX_ADDED_THREAD_COUNTERS 24
1103 #define PMT_MAX_ADDED_CORE_COUNTERS 8
1104 #define PMT_MAX_ADDED_PACKAGE_COUNTERS 16
1105 #define BITMASK_SIZE 32
1106 
1107 #define ZERO_ARRAY(arr) (memset(arr, 0, sizeof(arr)) + __must_be_array(arr))
1108 
1109 /* Indexes used to map data read from perf and MSRs into global variables */
1110 enum rapl_rci_index {
1111 	RAPL_RCI_INDEX_ENERGY_PKG = 0,
1112 	RAPL_RCI_INDEX_ENERGY_CORES = 1,
1113 	RAPL_RCI_INDEX_DRAM = 2,
1114 	RAPL_RCI_INDEX_GFX = 3,
1115 	RAPL_RCI_INDEX_PKG_PERF_STATUS = 4,
1116 	RAPL_RCI_INDEX_DRAM_PERF_STATUS = 5,
1117 	RAPL_RCI_INDEX_CORE_ENERGY = 6,
1118 	RAPL_RCI_INDEX_ENERGY_PLATFORM = 7,
1119 	NUM_RAPL_COUNTERS,
1120 };
1121 
1122 enum rapl_unit {
1123 	RAPL_UNIT_INVALID,
1124 	RAPL_UNIT_JOULES,
1125 	RAPL_UNIT_WATTS,
1126 };
1127 
1128 struct rapl_counter_info_t {
1129 	unsigned long long data[NUM_RAPL_COUNTERS];
1130 	enum counter_source source[NUM_RAPL_COUNTERS];
1131 	unsigned long long flags[NUM_RAPL_COUNTERS];
1132 	double scale[NUM_RAPL_COUNTERS];
1133 	enum rapl_unit unit[NUM_RAPL_COUNTERS];
1134 	unsigned long long msr[NUM_RAPL_COUNTERS];
1135 	unsigned long long msr_mask[NUM_RAPL_COUNTERS];
1136 	int msr_shift[NUM_RAPL_COUNTERS];
1137 
1138 	int fd_perf;
1139 };
1140 
1141 /* struct rapl_counter_info_t for each RAPL domain */
1142 struct rapl_counter_info_t *rapl_counter_info_perdomain;
1143 unsigned int rapl_counter_info_perdomain_size;
1144 
1145 #define RAPL_COUNTER_FLAG_PLATFORM_COUNTER (1u << 0)
1146 #define RAPL_COUNTER_FLAG_USE_MSR_SUM (1u << 1)
1147 
1148 struct rapl_counter_arch_info {
1149 	int feature_mask;	/* Mask for testing if the counter is supported on host */
1150 	const char *perf_subsys;
1151 	const char *perf_name;
1152 	unsigned long long msr;
1153 	unsigned long long msr_mask;
1154 	int msr_shift;		/* Positive mean shift right, negative mean shift left */
1155 	double *platform_rapl_msr_scale;	/* Scale applied to values read by MSR (platform dependent, filled at runtime) */
1156 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1157 	unsigned long long bic;
1158 	double compat_scale;	/* Some counters require constant scaling to be in the same range as other, similar ones */
1159 	unsigned long long flags;
1160 };
1161 
1162 static const struct rapl_counter_arch_info rapl_counter_arch_infos[] = {
1163 	{
1164 	 .feature_mask = RAPL_PKG,
1165 	 .perf_subsys = "power",
1166 	 .perf_name = "energy-pkg",
1167 	 .msr = MSR_PKG_ENERGY_STATUS,
1168 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1169 	 .msr_shift = 0,
1170 	 .platform_rapl_msr_scale = &rapl_energy_units,
1171 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1172 	 .bic = BIC_PkgWatt | BIC_Pkg_J,
1173 	 .compat_scale = 1.0,
1174 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1175 	  },
1176 	{
1177 	 .feature_mask = RAPL_AMD_F17H,
1178 	 .perf_subsys = "power",
1179 	 .perf_name = "energy-pkg",
1180 	 .msr = MSR_PKG_ENERGY_STAT,
1181 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1182 	 .msr_shift = 0,
1183 	 .platform_rapl_msr_scale = &rapl_energy_units,
1184 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1185 	 .bic = BIC_PkgWatt | BIC_Pkg_J,
1186 	 .compat_scale = 1.0,
1187 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1188 	  },
1189 	{
1190 	 .feature_mask = RAPL_CORE_ENERGY_STATUS,
1191 	 .perf_subsys = "power",
1192 	 .perf_name = "energy-cores",
1193 	 .msr = MSR_PP0_ENERGY_STATUS,
1194 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1195 	 .msr_shift = 0,
1196 	 .platform_rapl_msr_scale = &rapl_energy_units,
1197 	 .rci_index = RAPL_RCI_INDEX_ENERGY_CORES,
1198 	 .bic = BIC_CorWatt | BIC_Cor_J,
1199 	 .compat_scale = 1.0,
1200 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1201 	  },
1202 	{
1203 	 .feature_mask = RAPL_DRAM,
1204 	 .perf_subsys = "power",
1205 	 .perf_name = "energy-ram",
1206 	 .msr = MSR_DRAM_ENERGY_STATUS,
1207 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1208 	 .msr_shift = 0,
1209 	 .platform_rapl_msr_scale = &rapl_dram_energy_units,
1210 	 .rci_index = RAPL_RCI_INDEX_DRAM,
1211 	 .bic = BIC_RAMWatt | BIC_RAM_J,
1212 	 .compat_scale = 1.0,
1213 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1214 	  },
1215 	{
1216 	 .feature_mask = RAPL_GFX,
1217 	 .perf_subsys = "power",
1218 	 .perf_name = "energy-gpu",
1219 	 .msr = MSR_PP1_ENERGY_STATUS,
1220 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1221 	 .msr_shift = 0,
1222 	 .platform_rapl_msr_scale = &rapl_energy_units,
1223 	 .rci_index = RAPL_RCI_INDEX_GFX,
1224 	 .bic = BIC_GFXWatt | BIC_GFX_J,
1225 	 .compat_scale = 1.0,
1226 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1227 	  },
1228 	{
1229 	 .feature_mask = RAPL_PKG_PERF_STATUS,
1230 	 .perf_subsys = NULL,
1231 	 .perf_name = NULL,
1232 	 .msr = MSR_PKG_PERF_STATUS,
1233 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1234 	 .msr_shift = 0,
1235 	 .platform_rapl_msr_scale = &rapl_time_units,
1236 	 .rci_index = RAPL_RCI_INDEX_PKG_PERF_STATUS,
1237 	 .bic = BIC_PKG__,
1238 	 .compat_scale = 100.0,
1239 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1240 	  },
1241 	{
1242 	 .feature_mask = RAPL_DRAM_PERF_STATUS,
1243 	 .perf_subsys = NULL,
1244 	 .perf_name = NULL,
1245 	 .msr = MSR_DRAM_PERF_STATUS,
1246 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1247 	 .msr_shift = 0,
1248 	 .platform_rapl_msr_scale = &rapl_time_units,
1249 	 .rci_index = RAPL_RCI_INDEX_DRAM_PERF_STATUS,
1250 	 .bic = BIC_RAM__,
1251 	 .compat_scale = 100.0,
1252 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1253 	  },
1254 	{
1255 	 .feature_mask = RAPL_AMD_F17H,
1256 	 .perf_subsys = NULL,
1257 	 .perf_name = NULL,
1258 	 .msr = MSR_CORE_ENERGY_STAT,
1259 	 .msr_mask = 0xFFFFFFFF,
1260 	 .msr_shift = 0,
1261 	 .platform_rapl_msr_scale = &rapl_energy_units,
1262 	 .rci_index = RAPL_RCI_INDEX_CORE_ENERGY,
1263 	 .bic = BIC_CorWatt | BIC_Cor_J,
1264 	 .compat_scale = 1.0,
1265 	 .flags = 0,
1266 	  },
1267 	{
1268 	 .feature_mask = RAPL_PSYS,
1269 	 .perf_subsys = "power",
1270 	 .perf_name = "energy-psys",
1271 	 .msr = MSR_PLATFORM_ENERGY_STATUS,
1272 	 .msr_mask = 0x00000000FFFFFFFF,
1273 	 .msr_shift = 0,
1274 	 .platform_rapl_msr_scale = &rapl_energy_units,
1275 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PLATFORM,
1276 	 .bic = BIC_SysWatt | BIC_Sys_J,
1277 	 .compat_scale = 1.0,
1278 	 .flags = RAPL_COUNTER_FLAG_PLATFORM_COUNTER | RAPL_COUNTER_FLAG_USE_MSR_SUM,
1279 	  },
1280 };
1281 
1282 struct rapl_counter {
1283 	unsigned long long raw_value;
1284 	enum rapl_unit unit;
1285 	double scale;
1286 };
1287 
1288 /* Indexes used to map data read from perf and MSRs into global variables */
1289 enum ccstate_rci_index {
1290 	CCSTATE_RCI_INDEX_C1_RESIDENCY = 0,
1291 	CCSTATE_RCI_INDEX_C3_RESIDENCY = 1,
1292 	CCSTATE_RCI_INDEX_C6_RESIDENCY = 2,
1293 	CCSTATE_RCI_INDEX_C7_RESIDENCY = 3,
1294 	PCSTATE_RCI_INDEX_C2_RESIDENCY = 4,
1295 	PCSTATE_RCI_INDEX_C3_RESIDENCY = 5,
1296 	PCSTATE_RCI_INDEX_C6_RESIDENCY = 6,
1297 	PCSTATE_RCI_INDEX_C7_RESIDENCY = 7,
1298 	PCSTATE_RCI_INDEX_C8_RESIDENCY = 8,
1299 	PCSTATE_RCI_INDEX_C9_RESIDENCY = 9,
1300 	PCSTATE_RCI_INDEX_C10_RESIDENCY = 10,
1301 	NUM_CSTATE_COUNTERS,
1302 };
1303 
1304 struct cstate_counter_info_t {
1305 	unsigned long long data[NUM_CSTATE_COUNTERS];
1306 	enum counter_source source[NUM_CSTATE_COUNTERS];
1307 	unsigned long long msr[NUM_CSTATE_COUNTERS];
1308 	int fd_perf_core;
1309 	int fd_perf_pkg;
1310 };
1311 
1312 struct cstate_counter_info_t *ccstate_counter_info;
1313 unsigned int ccstate_counter_info_size;
1314 
1315 #define CSTATE_COUNTER_FLAG_COLLECT_PER_CORE   (1u << 0)
1316 #define CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD ((1u << 1) | CSTATE_COUNTER_FLAG_COLLECT_PER_CORE)
1317 #define CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY (1u << 2)
1318 
1319 struct cstate_counter_arch_info {
1320 	int feature_mask;	/* Mask for testing if the counter is supported on host */
1321 	const char *perf_subsys;
1322 	const char *perf_name;
1323 	unsigned long long msr;
1324 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1325 	unsigned long long bic;
1326 	unsigned long long flags;
1327 	int pkg_cstate_limit;
1328 };
1329 
1330 static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
1331 	{
1332 	 .feature_mask = CC1,
1333 	 .perf_subsys = "cstate_core",
1334 	 .perf_name = "c1-residency",
1335 	 .msr = MSR_CORE_C1_RES,
1336 	 .rci_index = CCSTATE_RCI_INDEX_C1_RESIDENCY,
1337 	 .bic = BIC_CPU_c1,
1338 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD,
1339 	 .pkg_cstate_limit = 0,
1340 	  },
1341 	{
1342 	 .feature_mask = CC3,
1343 	 .perf_subsys = "cstate_core",
1344 	 .perf_name = "c3-residency",
1345 	 .msr = MSR_CORE_C3_RESIDENCY,
1346 	 .rci_index = CCSTATE_RCI_INDEX_C3_RESIDENCY,
1347 	 .bic = BIC_CPU_c3,
1348 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1349 	 .pkg_cstate_limit = 0,
1350 	  },
1351 	{
1352 	 .feature_mask = CC6,
1353 	 .perf_subsys = "cstate_core",
1354 	 .perf_name = "c6-residency",
1355 	 .msr = MSR_CORE_C6_RESIDENCY,
1356 	 .rci_index = CCSTATE_RCI_INDEX_C6_RESIDENCY,
1357 	 .bic = BIC_CPU_c6,
1358 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1359 	 .pkg_cstate_limit = 0,
1360 	  },
1361 	{
1362 	 .feature_mask = CC7,
1363 	 .perf_subsys = "cstate_core",
1364 	 .perf_name = "c7-residency",
1365 	 .msr = MSR_CORE_C7_RESIDENCY,
1366 	 .rci_index = CCSTATE_RCI_INDEX_C7_RESIDENCY,
1367 	 .bic = BIC_CPU_c7,
1368 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1369 	 .pkg_cstate_limit = 0,
1370 	  },
1371 	{
1372 	 .feature_mask = PC2,
1373 	 .perf_subsys = "cstate_pkg",
1374 	 .perf_name = "c2-residency",
1375 	 .msr = MSR_PKG_C2_RESIDENCY,
1376 	 .rci_index = PCSTATE_RCI_INDEX_C2_RESIDENCY,
1377 	 .bic = BIC_Pkgpc2,
1378 	 .flags = 0,
1379 	 .pkg_cstate_limit = PCL__2,
1380 	  },
1381 	{
1382 	 .feature_mask = PC3,
1383 	 .perf_subsys = "cstate_pkg",
1384 	 .perf_name = "c3-residency",
1385 	 .msr = MSR_PKG_C3_RESIDENCY,
1386 	 .rci_index = PCSTATE_RCI_INDEX_C3_RESIDENCY,
1387 	 .bic = BIC_Pkgpc3,
1388 	 .flags = 0,
1389 	 .pkg_cstate_limit = PCL__3,
1390 	  },
1391 	{
1392 	 .feature_mask = PC6,
1393 	 .perf_subsys = "cstate_pkg",
1394 	 .perf_name = "c6-residency",
1395 	 .msr = MSR_PKG_C6_RESIDENCY,
1396 	 .rci_index = PCSTATE_RCI_INDEX_C6_RESIDENCY,
1397 	 .bic = BIC_Pkgpc6,
1398 	 .flags = 0,
1399 	 .pkg_cstate_limit = PCL__6,
1400 	  },
1401 	{
1402 	 .feature_mask = PC7,
1403 	 .perf_subsys = "cstate_pkg",
1404 	 .perf_name = "c7-residency",
1405 	 .msr = MSR_PKG_C7_RESIDENCY,
1406 	 .rci_index = PCSTATE_RCI_INDEX_C7_RESIDENCY,
1407 	 .bic = BIC_Pkgpc7,
1408 	 .flags = 0,
1409 	 .pkg_cstate_limit = PCL__7,
1410 	  },
1411 	{
1412 	 .feature_mask = PC8,
1413 	 .perf_subsys = "cstate_pkg",
1414 	 .perf_name = "c8-residency",
1415 	 .msr = MSR_PKG_C8_RESIDENCY,
1416 	 .rci_index = PCSTATE_RCI_INDEX_C8_RESIDENCY,
1417 	 .bic = BIC_Pkgpc8,
1418 	 .flags = 0,
1419 	 .pkg_cstate_limit = PCL__8,
1420 	  },
1421 	{
1422 	 .feature_mask = PC9,
1423 	 .perf_subsys = "cstate_pkg",
1424 	 .perf_name = "c9-residency",
1425 	 .msr = MSR_PKG_C9_RESIDENCY,
1426 	 .rci_index = PCSTATE_RCI_INDEX_C9_RESIDENCY,
1427 	 .bic = BIC_Pkgpc9,
1428 	 .flags = 0,
1429 	 .pkg_cstate_limit = PCL__9,
1430 	  },
1431 	{
1432 	 .feature_mask = PC10,
1433 	 .perf_subsys = "cstate_pkg",
1434 	 .perf_name = "c10-residency",
1435 	 .msr = MSR_PKG_C10_RESIDENCY,
1436 	 .rci_index = PCSTATE_RCI_INDEX_C10_RESIDENCY,
1437 	 .bic = BIC_Pkgpc10,
1438 	 .flags = 0,
1439 	 .pkg_cstate_limit = PCL_10,
1440 	  },
1441 };
1442 
1443 /* Indexes used to map data read from perf and MSRs into global variables */
1444 enum msr_rci_index {
1445 	MSR_RCI_INDEX_APERF = 0,
1446 	MSR_RCI_INDEX_MPERF = 1,
1447 	MSR_RCI_INDEX_SMI = 2,
1448 	NUM_MSR_COUNTERS,
1449 };
1450 
1451 struct msr_counter_info_t {
1452 	unsigned long long data[NUM_MSR_COUNTERS];
1453 	enum counter_source source[NUM_MSR_COUNTERS];
1454 	unsigned long long msr[NUM_MSR_COUNTERS];
1455 	unsigned long long msr_mask[NUM_MSR_COUNTERS];
1456 	int fd_perf;
1457 };
1458 
1459 struct msr_counter_info_t *msr_counter_info;
1460 unsigned int msr_counter_info_size;
1461 
1462 struct msr_counter_arch_info {
1463 	const char *perf_subsys;
1464 	const char *perf_name;
1465 	unsigned long long msr;
1466 	unsigned long long msr_mask;
1467 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1468 	bool needed;
1469 	bool present;
1470 };
1471 
1472 enum msr_arch_info_index {
1473 	MSR_ARCH_INFO_APERF_INDEX = 0,
1474 	MSR_ARCH_INFO_MPERF_INDEX = 1,
1475 	MSR_ARCH_INFO_SMI_INDEX = 2,
1476 };
1477 
1478 static struct msr_counter_arch_info msr_counter_arch_infos[] = {
1479 	[MSR_ARCH_INFO_APERF_INDEX] = {
1480 				       .perf_subsys = "msr",
1481 				       .perf_name = "aperf",
1482 				       .msr = MSR_IA32_APERF,
1483 				       .msr_mask = 0xFFFFFFFFFFFFFFFF,
1484 				       .rci_index = MSR_RCI_INDEX_APERF,
1485 				        },
1486 
1487 	[MSR_ARCH_INFO_MPERF_INDEX] = {
1488 				       .perf_subsys = "msr",
1489 				       .perf_name = "mperf",
1490 				       .msr = MSR_IA32_MPERF,
1491 				       .msr_mask = 0xFFFFFFFFFFFFFFFF,
1492 				       .rci_index = MSR_RCI_INDEX_MPERF,
1493 				        },
1494 
1495 	[MSR_ARCH_INFO_SMI_INDEX] = {
1496 				     .perf_subsys = "msr",
1497 				     .perf_name = "smi",
1498 				     .msr = MSR_SMI_COUNT,
1499 				     .msr_mask = 0xFFFFFFFF,
1500 				     .rci_index = MSR_RCI_INDEX_SMI,
1501 				      },
1502 };
1503 
1504 /* Can be redefined when compiling, useful for testing. */
1505 #ifndef SYSFS_TELEM_PATH
1506 #define SYSFS_TELEM_PATH "/sys/class/intel_pmt"
1507 #endif
1508 
1509 #define PMT_COUNTER_MTL_DC6_OFFSET 120
1510 #define PMT_COUNTER_MTL_DC6_LSB    0
1511 #define PMT_COUNTER_MTL_DC6_MSB    63
1512 #define PMT_MTL_DC6_GUID           0x1a067102
1513 
1514 #define PMT_COUNTER_NAME_SIZE_BYTES      16
1515 #define PMT_COUNTER_TYPE_NAME_SIZE_BYTES 32
1516 
1517 struct pmt_mmio {
1518 	struct pmt_mmio *next;
1519 
1520 	unsigned int guid;
1521 	unsigned int size;
1522 
1523 	/* Base pointer to the mmaped memory. */
1524 	void *mmio_base;
1525 
1526 	/*
1527 	 * Offset to be applied to the mmio_base
1528 	 * to get the beginning of the PMT counters for given GUID.
1529 	 */
1530 	unsigned long pmt_offset;
1531 } *pmt_mmios;
1532 
1533 enum pmt_datatype {
1534 	PMT_TYPE_RAW,
1535 	PMT_TYPE_XTAL_TIME,
1536 };
1537 
1538 struct pmt_domain_info {
1539 	/*
1540 	 * Pointer to the MMIO obtained by applying a counter offset
1541 	 * to the mmio_base of the mmaped region for the given GUID.
1542 	 *
1543 	 * This is where to read the raw value of the counter from.
1544 	 */
1545 	unsigned long *pcounter;
1546 };
1547 
1548 struct pmt_counter {
1549 	struct pmt_counter *next;
1550 
1551 	/* PMT metadata */
1552 	char name[PMT_COUNTER_NAME_SIZE_BYTES];
1553 	enum pmt_datatype type;
1554 	enum counter_scope scope;
1555 	unsigned int lsb;
1556 	unsigned int msb;
1557 
1558 	/* BIC-like metadata */
1559 	enum counter_format format;
1560 
1561 	unsigned int num_domains;
1562 	struct pmt_domain_info *domains;
1563 };
1564 
pmt_counter_get_width(const struct pmt_counter * p)1565 unsigned int pmt_counter_get_width(const struct pmt_counter *p)
1566 {
1567 	return (p->msb - p->lsb) + 1;
1568 }
1569 
pmt_counter_resize_(struct pmt_counter * pcounter,unsigned int new_size)1570 void pmt_counter_resize_(struct pmt_counter *pcounter, unsigned int new_size)
1571 {
1572 	struct pmt_domain_info *new_mem;
1573 
1574 	new_mem = (struct pmt_domain_info *)reallocarray(pcounter->domains, new_size, sizeof(*pcounter->domains));
1575 	if (!new_mem) {
1576 		fprintf(stderr, "%s: failed to allocate memory for PMT counters\n", __func__);
1577 		exit(1);
1578 	}
1579 
1580 	/* Zero initialize just allocated memory. */
1581 	const size_t num_new_domains = new_size - pcounter->num_domains;
1582 
1583 	memset(&new_mem[pcounter->num_domains], 0, num_new_domains * sizeof(*pcounter->domains));
1584 
1585 	pcounter->num_domains = new_size;
1586 	pcounter->domains = new_mem;
1587 }
1588 
pmt_counter_resize(struct pmt_counter * pcounter,unsigned int new_size)1589 void pmt_counter_resize(struct pmt_counter *pcounter, unsigned int new_size)
1590 {
1591 	/*
1592 	 * Allocate more memory ahead of time.
1593 	 *
1594 	 * Always allocate space for at least 8 elements
1595 	 * and double the size when growing.
1596 	 */
1597 	if (new_size < 8)
1598 		new_size = 8;
1599 	new_size = MAX(new_size, pcounter->num_domains * 2);
1600 
1601 	pmt_counter_resize_(pcounter, new_size);
1602 }
1603 
1604 struct thread_data {
1605 	struct timeval tv_begin;
1606 	struct timeval tv_end;
1607 	struct timeval tv_delta;
1608 	unsigned long long tsc;
1609 	unsigned long long aperf;
1610 	unsigned long long mperf;
1611 	unsigned long long c1;
1612 	unsigned long long instr_count;
1613 	unsigned long long irq_count;
1614 	unsigned int smi_count;
1615 	unsigned int cpu_id;
1616 	unsigned int apic_id;
1617 	unsigned int x2apic_id;
1618 	unsigned int flags;
1619 	bool is_atom;
1620 	unsigned long long counter[MAX_ADDED_THREAD_COUNTERS];
1621 	unsigned long long perf_counter[MAX_ADDED_THREAD_COUNTERS];
1622 	unsigned long long pmt_counter[PMT_MAX_ADDED_THREAD_COUNTERS];
1623 } *thread_even, *thread_odd;
1624 
1625 struct core_data {
1626 	int base_cpu;
1627 	unsigned long long c3;
1628 	unsigned long long c6;
1629 	unsigned long long c7;
1630 	unsigned long long mc6_us;	/* duplicate as per-core for now, even though per module */
1631 	unsigned int core_temp_c;
1632 	struct rapl_counter core_energy;	/* MSR_CORE_ENERGY_STAT */
1633 	unsigned int core_id;
1634 	unsigned long long core_throt_cnt;
1635 	unsigned long long counter[MAX_ADDED_CORE_COUNTERS];
1636 	unsigned long long perf_counter[MAX_ADDED_CORE_COUNTERS];
1637 	unsigned long long pmt_counter[PMT_MAX_ADDED_CORE_COUNTERS];
1638 } *core_even, *core_odd;
1639 
1640 struct pkg_data {
1641 	int base_cpu;
1642 	unsigned long long pc2;
1643 	unsigned long long pc3;
1644 	unsigned long long pc6;
1645 	unsigned long long pc7;
1646 	unsigned long long pc8;
1647 	unsigned long long pc9;
1648 	unsigned long long pc10;
1649 	long long cpu_lpi;
1650 	long long sys_lpi;
1651 	unsigned long long pkg_wtd_core_c0;
1652 	unsigned long long pkg_any_core_c0;
1653 	unsigned long long pkg_any_gfxe_c0;
1654 	unsigned long long pkg_both_core_gfxe_c0;
1655 	long long gfx_rc6_ms;
1656 	unsigned int gfx_mhz;
1657 	unsigned int gfx_act_mhz;
1658 	long long sam_mc6_ms;
1659 	unsigned int sam_mhz;
1660 	unsigned int sam_act_mhz;
1661 	unsigned int package_id;
1662 	struct rapl_counter energy_pkg;	/* MSR_PKG_ENERGY_STATUS */
1663 	struct rapl_counter energy_dram;	/* MSR_DRAM_ENERGY_STATUS */
1664 	struct rapl_counter energy_cores;	/* MSR_PP0_ENERGY_STATUS */
1665 	struct rapl_counter energy_gfx;	/* MSR_PP1_ENERGY_STATUS */
1666 	struct rapl_counter rapl_pkg_perf_status;	/* MSR_PKG_PERF_STATUS */
1667 	struct rapl_counter rapl_dram_perf_status;	/* MSR_DRAM_PERF_STATUS */
1668 	unsigned int pkg_temp_c;
1669 	unsigned int uncore_mhz;
1670 	unsigned long long die_c6;
1671 	unsigned long long counter[MAX_ADDED_PACKAGE_COUNTERS];
1672 	unsigned long long perf_counter[MAX_ADDED_PACKAGE_COUNTERS];
1673 	unsigned long long pmt_counter[PMT_MAX_ADDED_PACKAGE_COUNTERS];
1674 } *package_even, *package_odd;
1675 
1676 #define ODD_COUNTERS thread_odd, core_odd, package_odd
1677 #define EVEN_COUNTERS thread_even, core_even, package_even
1678 
1679 #define GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no)	      \
1680 	((thread_base) +						      \
1681 	 ((pkg_no) *							      \
1682 	  topo.nodes_per_pkg * topo.cores_per_node * topo.threads_per_core) + \
1683 	 ((node_no) * topo.cores_per_node * topo.threads_per_core) +	      \
1684 	 ((core_no) * topo.threads_per_core) +				      \
1685 	 (thread_no))
1686 
1687 #define GET_CORE(core_base, core_no, node_no, pkg_no)			\
1688 	((core_base) +							\
1689 	 ((pkg_no) *  topo.nodes_per_pkg * topo.cores_per_node) +	\
1690 	 ((node_no) * topo.cores_per_node) +				\
1691 	 (core_no))
1692 
1693 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
1694 
1695 /*
1696  * The accumulated sum of MSR is defined as a monotonic
1697  * increasing MSR, it will be accumulated periodically,
1698  * despite its register's bit width.
1699  */
1700 enum {
1701 	IDX_PKG_ENERGY,
1702 	IDX_DRAM_ENERGY,
1703 	IDX_PP0_ENERGY,
1704 	IDX_PP1_ENERGY,
1705 	IDX_PKG_PERF,
1706 	IDX_DRAM_PERF,
1707 	IDX_PSYS_ENERGY,
1708 	IDX_COUNT,
1709 };
1710 
1711 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr);
1712 
1713 struct msr_sum_array {
1714 	/* get_msr_sum() = sum + (get_msr() - last) */
1715 	struct {
1716 		/*The accumulated MSR value is updated by the timer */
1717 		unsigned long long sum;
1718 		/*The MSR footprint recorded in last timer */
1719 		unsigned long long last;
1720 	} entries[IDX_COUNT];
1721 };
1722 
1723 /* The percpu MSR sum array.*/
1724 struct msr_sum_array *per_cpu_msr_sum;
1725 
idx_to_offset(int idx)1726 off_t idx_to_offset(int idx)
1727 {
1728 	off_t offset;
1729 
1730 	switch (idx) {
1731 	case IDX_PKG_ENERGY:
1732 		if (platform->rapl_msrs & RAPL_AMD_F17H)
1733 			offset = MSR_PKG_ENERGY_STAT;
1734 		else
1735 			offset = MSR_PKG_ENERGY_STATUS;
1736 		break;
1737 	case IDX_DRAM_ENERGY:
1738 		offset = MSR_DRAM_ENERGY_STATUS;
1739 		break;
1740 	case IDX_PP0_ENERGY:
1741 		offset = MSR_PP0_ENERGY_STATUS;
1742 		break;
1743 	case IDX_PP1_ENERGY:
1744 		offset = MSR_PP1_ENERGY_STATUS;
1745 		break;
1746 	case IDX_PKG_PERF:
1747 		offset = MSR_PKG_PERF_STATUS;
1748 		break;
1749 	case IDX_DRAM_PERF:
1750 		offset = MSR_DRAM_PERF_STATUS;
1751 		break;
1752 	case IDX_PSYS_ENERGY:
1753 		offset = MSR_PLATFORM_ENERGY_STATUS;
1754 		break;
1755 	default:
1756 		offset = -1;
1757 	}
1758 	return offset;
1759 }
1760 
offset_to_idx(off_t offset)1761 int offset_to_idx(off_t offset)
1762 {
1763 	int idx;
1764 
1765 	switch (offset) {
1766 	case MSR_PKG_ENERGY_STATUS:
1767 	case MSR_PKG_ENERGY_STAT:
1768 		idx = IDX_PKG_ENERGY;
1769 		break;
1770 	case MSR_DRAM_ENERGY_STATUS:
1771 		idx = IDX_DRAM_ENERGY;
1772 		break;
1773 	case MSR_PP0_ENERGY_STATUS:
1774 		idx = IDX_PP0_ENERGY;
1775 		break;
1776 	case MSR_PP1_ENERGY_STATUS:
1777 		idx = IDX_PP1_ENERGY;
1778 		break;
1779 	case MSR_PKG_PERF_STATUS:
1780 		idx = IDX_PKG_PERF;
1781 		break;
1782 	case MSR_DRAM_PERF_STATUS:
1783 		idx = IDX_DRAM_PERF;
1784 		break;
1785 	case MSR_PLATFORM_ENERGY_STATUS:
1786 		idx = IDX_PSYS_ENERGY;
1787 		break;
1788 	default:
1789 		idx = -1;
1790 	}
1791 	return idx;
1792 }
1793 
idx_valid(int idx)1794 int idx_valid(int idx)
1795 {
1796 	switch (idx) {
1797 	case IDX_PKG_ENERGY:
1798 		return platform->rapl_msrs & (RAPL_PKG | RAPL_AMD_F17H);
1799 	case IDX_DRAM_ENERGY:
1800 		return platform->rapl_msrs & RAPL_DRAM;
1801 	case IDX_PP0_ENERGY:
1802 		return platform->rapl_msrs & RAPL_CORE_ENERGY_STATUS;
1803 	case IDX_PP1_ENERGY:
1804 		return platform->rapl_msrs & RAPL_GFX;
1805 	case IDX_PKG_PERF:
1806 		return platform->rapl_msrs & RAPL_PKG_PERF_STATUS;
1807 	case IDX_DRAM_PERF:
1808 		return platform->rapl_msrs & RAPL_DRAM_PERF_STATUS;
1809 	case IDX_PSYS_ENERGY:
1810 		return platform->rapl_msrs & RAPL_PSYS;
1811 	default:
1812 		return 0;
1813 	}
1814 }
1815 
1816 struct sys_counters {
1817 	/* MSR added counters */
1818 	unsigned int added_thread_counters;
1819 	unsigned int added_core_counters;
1820 	unsigned int added_package_counters;
1821 	struct msr_counter *tp;
1822 	struct msr_counter *cp;
1823 	struct msr_counter *pp;
1824 
1825 	/* perf added counters */
1826 	unsigned int added_thread_perf_counters;
1827 	unsigned int added_core_perf_counters;
1828 	unsigned int added_package_perf_counters;
1829 	struct perf_counter_info *perf_tp;
1830 	struct perf_counter_info *perf_cp;
1831 	struct perf_counter_info *perf_pp;
1832 
1833 	struct pmt_counter *pmt_tp;
1834 	struct pmt_counter *pmt_cp;
1835 	struct pmt_counter *pmt_pp;
1836 } sys;
1837 
free_msr_counters_(struct msr_counter ** pp)1838 static size_t free_msr_counters_(struct msr_counter **pp)
1839 {
1840 	struct msr_counter *p = NULL;
1841 	size_t num_freed = 0;
1842 
1843 	while (*pp) {
1844 		p = *pp;
1845 
1846 		if (p->msr_num != 0) {
1847 			*pp = p->next;
1848 
1849 			free(p);
1850 			++num_freed;
1851 
1852 			continue;
1853 		}
1854 
1855 		pp = &p->next;
1856 	}
1857 
1858 	return num_freed;
1859 }
1860 
1861 /*
1862  * Free all added counters accessed via msr.
1863  */
free_sys_msr_counters(void)1864 static void free_sys_msr_counters(void)
1865 {
1866 	/* Thread counters */
1867 	sys.added_thread_counters -= free_msr_counters_(&sys.tp);
1868 
1869 	/* Core counters */
1870 	sys.added_core_counters -= free_msr_counters_(&sys.cp);
1871 
1872 	/* Package counters */
1873 	sys.added_package_counters -= free_msr_counters_(&sys.pp);
1874 }
1875 
1876 struct system_summary {
1877 	struct thread_data threads;
1878 	struct core_data cores;
1879 	struct pkg_data packages;
1880 } average;
1881 
1882 struct platform_counters {
1883 	struct rapl_counter energy_psys;	/* MSR_PLATFORM_ENERGY_STATUS */
1884 } platform_counters_odd, platform_counters_even;
1885 
1886 struct cpu_topology {
1887 	int physical_package_id;
1888 	int die_id;
1889 	int logical_cpu_id;
1890 	int physical_node_id;
1891 	int logical_node_id;	/* 0-based count within the package */
1892 	int physical_core_id;
1893 	int thread_id;
1894 	int type;
1895 	cpu_set_t *put_ids;	/* Processing Unit/Thread IDs */
1896 } *cpus;
1897 
1898 struct topo_params {
1899 	int num_packages;
1900 	int num_die;
1901 	int num_cpus;
1902 	int num_cores;
1903 	int allowed_packages;
1904 	int allowed_cpus;
1905 	int allowed_cores;
1906 	int max_cpu_num;
1907 	int max_core_id;
1908 	int max_package_id;
1909 	int max_die_id;
1910 	int max_node_num;
1911 	int nodes_per_pkg;
1912 	int cores_per_node;
1913 	int threads_per_core;
1914 } topo;
1915 
1916 struct timeval tv_even, tv_odd, tv_delta;
1917 
1918 int *irq_column_2_cpu;		/* /proc/interrupts column numbers */
1919 int *irqs_per_cpu;		/* indexed by cpu_num */
1920 
1921 void setup_all_buffers(bool startup);
1922 
1923 char *sys_lpi_file;
1924 char *sys_lpi_file_sysfs = "/sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us";
1925 char *sys_lpi_file_debugfs = "/sys/kernel/debug/pmc_core/slp_s0_residency_usec";
1926 
cpu_is_not_present(int cpu)1927 int cpu_is_not_present(int cpu)
1928 {
1929 	return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
1930 }
1931 
cpu_is_not_allowed(int cpu)1932 int cpu_is_not_allowed(int cpu)
1933 {
1934 	return !CPU_ISSET_S(cpu, cpu_allowed_setsize, cpu_allowed_set);
1935 }
1936 
1937 /*
1938  * run func(thread, core, package) in topology order
1939  * skip non-present cpus
1940  */
1941 
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)1942 int for_all_cpus(int (func) (struct thread_data *, struct core_data *, struct pkg_data *),
1943 		 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
1944 {
1945 	int retval, pkg_no, core_no, thread_no, node_no;
1946 
1947 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1948 		for (node_no = 0; node_no < topo.nodes_per_pkg; node_no++) {
1949 			for (core_no = 0; core_no < topo.cores_per_node; ++core_no) {
1950 				for (thread_no = 0; thread_no < topo.threads_per_core; ++thread_no) {
1951 					struct thread_data *t;
1952 					struct core_data *c;
1953 					struct pkg_data *p;
1954 					t = GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no);
1955 
1956 					if (cpu_is_not_allowed(t->cpu_id))
1957 						continue;
1958 
1959 					c = GET_CORE(core_base, core_no, node_no, pkg_no);
1960 					p = GET_PKG(pkg_base, pkg_no);
1961 
1962 					retval = func(t, c, p);
1963 					if (retval)
1964 						return retval;
1965 				}
1966 			}
1967 		}
1968 	}
1969 	return 0;
1970 }
1971 
is_cpu_first_thread_in_core(struct thread_data * t,struct core_data * c,struct pkg_data * p)1972 int is_cpu_first_thread_in_core(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1973 {
1974 	UNUSED(p);
1975 
1976 	return ((int)t->cpu_id == c->base_cpu || c->base_cpu < 0);
1977 }
1978 
is_cpu_first_core_in_package(struct thread_data * t,struct core_data * c,struct pkg_data * p)1979 int is_cpu_first_core_in_package(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1980 {
1981 	UNUSED(c);
1982 
1983 	return ((int)t->cpu_id == p->base_cpu || p->base_cpu < 0);
1984 }
1985 
is_cpu_first_thread_in_package(struct thread_data * t,struct core_data * c,struct pkg_data * p)1986 int is_cpu_first_thread_in_package(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1987 {
1988 	return is_cpu_first_thread_in_core(t, c, p) && is_cpu_first_core_in_package(t, c, p);
1989 }
1990 
cpu_migrate(int cpu)1991 int cpu_migrate(int cpu)
1992 {
1993 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1994 	CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
1995 	if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
1996 		return -1;
1997 	else
1998 		return 0;
1999 }
2000 
get_msr_fd(int cpu)2001 int get_msr_fd(int cpu)
2002 {
2003 	char pathname[32];
2004 	int fd;
2005 
2006 	fd = fd_percpu[cpu];
2007 
2008 	if (fd)
2009 		return fd;
2010 
2011 	sprintf(pathname, "/dev/cpu/%d/msr", cpu);
2012 	fd = open(pathname, O_RDONLY);
2013 	if (fd < 0)
2014 		err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, "
2015 		    "or run with --no-msr, or run as root", pathname);
2016 
2017 	fd_percpu[cpu] = fd;
2018 
2019 	return fd;
2020 }
2021 
bic_disable_msr_access(void)2022 static void bic_disable_msr_access(void)
2023 {
2024 	const unsigned long bic_msrs = BIC_Mod_c6 | BIC_CoreTmp |
2025 	    BIC_Totl_c0 | BIC_Any_c0 | BIC_GFX_c0 | BIC_CPUGFX | BIC_PkgTmp;
2026 
2027 	bic_enabled &= ~bic_msrs;
2028 
2029 	free_sys_msr_counters();
2030 }
2031 
perf_event_open(struct perf_event_attr * hw_event,pid_t pid,int cpu,int group_fd,unsigned long flags)2032 static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid, int cpu, int group_fd, unsigned long flags)
2033 {
2034 	assert(!no_perf);
2035 
2036 	return syscall(__NR_perf_event_open, hw_event, pid, cpu, group_fd, flags);
2037 }
2038 
open_perf_counter(int cpu,unsigned int type,unsigned int config,int group_fd,__u64 read_format)2039 static long open_perf_counter(int cpu, unsigned int type, unsigned int config, int group_fd, __u64 read_format)
2040 {
2041 	struct perf_event_attr attr;
2042 	const pid_t pid = -1;
2043 	const unsigned long flags = 0;
2044 
2045 	assert(!no_perf);
2046 
2047 	memset(&attr, 0, sizeof(struct perf_event_attr));
2048 
2049 	attr.type = type;
2050 	attr.size = sizeof(struct perf_event_attr);
2051 	attr.config = config;
2052 	attr.disabled = 0;
2053 	attr.sample_type = PERF_SAMPLE_IDENTIFIER;
2054 	attr.read_format = read_format;
2055 
2056 	const int fd = perf_event_open(&attr, pid, cpu, group_fd, flags);
2057 
2058 	return fd;
2059 }
2060 
get_instr_count_fd(int cpu)2061 int get_instr_count_fd(int cpu)
2062 {
2063 	if (fd_instr_count_percpu[cpu])
2064 		return fd_instr_count_percpu[cpu];
2065 
2066 	fd_instr_count_percpu[cpu] = open_perf_counter(cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, -1, 0);
2067 
2068 	return fd_instr_count_percpu[cpu];
2069 }
2070 
get_msr(int cpu,off_t offset,unsigned long long * msr)2071 int get_msr(int cpu, off_t offset, unsigned long long *msr)
2072 {
2073 	ssize_t retval;
2074 
2075 	assert(!no_msr);
2076 
2077 	retval = pread(get_msr_fd(cpu), msr, sizeof(*msr), offset);
2078 
2079 	if (retval != sizeof *msr)
2080 		err(-1, "cpu%d: msr offset 0x%llx read failed", cpu, (unsigned long long)offset);
2081 
2082 	return 0;
2083 }
2084 
probe_msr(int cpu,off_t offset)2085 int probe_msr(int cpu, off_t offset)
2086 {
2087 	ssize_t retval;
2088 	unsigned long long dummy;
2089 
2090 	assert(!no_msr);
2091 
2092 	retval = pread(get_msr_fd(cpu), &dummy, sizeof(dummy), offset);
2093 
2094 	if (retval != sizeof(dummy))
2095 		return 1;
2096 
2097 	return 0;
2098 }
2099 
2100 /* Convert CPU ID to domain ID for given added perf counter. */
cpu_to_domain(const struct perf_counter_info * pc,int cpu)2101 unsigned int cpu_to_domain(const struct perf_counter_info *pc, int cpu)
2102 {
2103 	switch (pc->scope) {
2104 	case SCOPE_CPU:
2105 		return cpu;
2106 
2107 	case SCOPE_CORE:
2108 		return cpus[cpu].physical_core_id;
2109 
2110 	case SCOPE_PACKAGE:
2111 		return cpus[cpu].physical_package_id;
2112 	}
2113 
2114 	__builtin_unreachable();
2115 }
2116 
2117 #define MAX_DEFERRED 16
2118 char *deferred_add_names[MAX_DEFERRED];
2119 char *deferred_skip_names[MAX_DEFERRED];
2120 int deferred_add_index;
2121 int deferred_skip_index;
2122 
2123 /*
2124  * HIDE_LIST - hide this list of counters, show the rest [default]
2125  * SHOW_LIST - show this list of counters, hide the rest
2126  */
2127 enum show_hide_mode { SHOW_LIST, HIDE_LIST } global_show_hide_mode = HIDE_LIST;
2128 
help(void)2129 void help(void)
2130 {
2131 	fprintf(outf,
2132 		"Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
2133 		"\n"
2134 		"Turbostat forks the specified COMMAND and prints statistics\n"
2135 		"when COMMAND completes.\n"
2136 		"If no COMMAND is specified, turbostat wakes every 5-seconds\n"
2137 		"to print statistics, until interrupted.\n"
2138 		"  -a, --add	add a counter\n"
2139 		"		  eg. --add msr0x10,u64,cpu,delta,MY_TSC\n"
2140 		"		  eg. --add perf/cstate_pkg/c2-residency,package,delta,percent,perfPC2\n"
2141 		"		  eg. --add pmt,name=XTAL,type=raw,domain=package0,offset=0,lsb=0,msb=63,guid=0x1a067102\n"
2142 		"  -c, --cpu	cpu-set	limit output to summary plus cpu-set:\n"
2143 		"		  {core | package | j,k,l..m,n-p }\n"
2144 		"  -d, --debug	displays usec, Time_Of_Day_Seconds and more debugging\n"
2145 		"		debug messages are printed to stderr\n"
2146 		"  -D, --Dump	displays the raw counter values\n"
2147 		"  -e, --enable	[all | column]\n"
2148 		"		shows all or the specified disabled column\n"
2149 		"  -H, --hide [column|column,column,...]\n"
2150 		"		hide the specified column(s)\n"
2151 		"  -i, --interval sec.subsec\n"
2152 		"		Override default 5-second measurement interval\n"
2153 		"  -J, --Joules	displays energy in Joules instead of Watts\n"
2154 		"  -l, --list	list column headers only\n"
2155 		"  -M, --no-msr Disable all uses of the MSR driver\n"
2156 		"  -P, --no-perf Disable all uses of the perf API\n"
2157 		"  -n, --num_iterations num\n"
2158 		"		number of the measurement iterations\n"
2159 		"  -N, --header_iterations num\n"
2160 		"		print header every num iterations\n"
2161 		"  -o, --out file\n"
2162 		"		create or truncate \"file\" for all output\n"
2163 		"  -q, --quiet	skip decoding system configuration header\n"
2164 		"  -s, --show [column|column,column,...]\n"
2165 		"		show only the specified column(s)\n"
2166 		"  -S, --Summary\n"
2167 		"		limits output to 1-line system summary per interval\n"
2168 		"  -T, --TCC temperature\n"
2169 		"		sets the Thermal Control Circuit temperature in\n"
2170 		"		  degrees Celsius\n"
2171 		"  -h, --help	print this help message\n"
2172 		"  -v, --version	print version information\n" "\n" "For more help, run \"man turbostat\"\n");
2173 }
2174 
2175 /*
2176  * bic_lookup
2177  * for all the strings in comma separate name_list,
2178  * set the approprate bit in return value.
2179  */
bic_lookup(char * name_list,enum show_hide_mode mode)2180 unsigned long long bic_lookup(char *name_list, enum show_hide_mode mode)
2181 {
2182 	unsigned int i;
2183 	unsigned long long retval = 0;
2184 
2185 	while (name_list) {
2186 		char *comma;
2187 
2188 		comma = strchr(name_list, ',');
2189 
2190 		if (comma)
2191 			*comma = '\0';
2192 
2193 		for (i = 0; i < MAX_BIC; ++i) {
2194 			if (!strcmp(name_list, bic[i].name)) {
2195 				retval |= (1ULL << i);
2196 				break;
2197 			}
2198 			if (!strcmp(name_list, "all")) {
2199 				retval |= ~0;
2200 				break;
2201 			} else if (!strcmp(name_list, "topology")) {
2202 				retval |= BIC_TOPOLOGY;
2203 				break;
2204 			} else if (!strcmp(name_list, "power")) {
2205 				retval |= BIC_THERMAL_PWR;
2206 				break;
2207 			} else if (!strcmp(name_list, "idle")) {
2208 				retval |= BIC_IDLE;
2209 				break;
2210 			} else if (!strcmp(name_list, "frequency")) {
2211 				retval |= BIC_FREQUENCY;
2212 				break;
2213 			} else if (!strcmp(name_list, "other")) {
2214 				retval |= BIC_OTHER;
2215 				break;
2216 			}
2217 
2218 		}
2219 		if (i == MAX_BIC) {
2220 			if (mode == SHOW_LIST) {
2221 				deferred_add_names[deferred_add_index++] = name_list;
2222 				if (deferred_add_index >= MAX_DEFERRED) {
2223 					fprintf(stderr, "More than max %d un-recognized --add options '%s'\n",
2224 						MAX_DEFERRED, name_list);
2225 					help();
2226 					exit(1);
2227 				}
2228 			} else {
2229 				deferred_skip_names[deferred_skip_index++] = name_list;
2230 				if (debug)
2231 					fprintf(stderr, "deferred \"%s\"\n", name_list);
2232 				if (deferred_skip_index >= MAX_DEFERRED) {
2233 					fprintf(stderr, "More than max %d un-recognized --skip options '%s'\n",
2234 						MAX_DEFERRED, name_list);
2235 					help();
2236 					exit(1);
2237 				}
2238 			}
2239 		}
2240 
2241 		name_list = comma;
2242 		if (name_list)
2243 			name_list++;
2244 
2245 	}
2246 	return retval;
2247 }
2248 
print_header(char * delim)2249 void print_header(char *delim)
2250 {
2251 	struct msr_counter *mp;
2252 	struct perf_counter_info *pp;
2253 	struct pmt_counter *ppmt;
2254 	int printed = 0;
2255 
2256 	if (DO_BIC(BIC_USEC))
2257 		outp += sprintf(outp, "%susec", (printed++ ? delim : ""));
2258 	if (DO_BIC(BIC_TOD))
2259 		outp += sprintf(outp, "%sTime_Of_Day_Seconds", (printed++ ? delim : ""));
2260 	if (DO_BIC(BIC_Package))
2261 		outp += sprintf(outp, "%sPackage", (printed++ ? delim : ""));
2262 	if (DO_BIC(BIC_Die))
2263 		outp += sprintf(outp, "%sDie", (printed++ ? delim : ""));
2264 	if (DO_BIC(BIC_Node))
2265 		outp += sprintf(outp, "%sNode", (printed++ ? delim : ""));
2266 	if (DO_BIC(BIC_Core))
2267 		outp += sprintf(outp, "%sCore", (printed++ ? delim : ""));
2268 	if (DO_BIC(BIC_CPU))
2269 		outp += sprintf(outp, "%sCPU", (printed++ ? delim : ""));
2270 	if (DO_BIC(BIC_APIC))
2271 		outp += sprintf(outp, "%sAPIC", (printed++ ? delim : ""));
2272 	if (DO_BIC(BIC_X2APIC))
2273 		outp += sprintf(outp, "%sX2APIC", (printed++ ? delim : ""));
2274 	if (DO_BIC(BIC_Avg_MHz))
2275 		outp += sprintf(outp, "%sAvg_MHz", (printed++ ? delim : ""));
2276 	if (DO_BIC(BIC_Busy))
2277 		outp += sprintf(outp, "%sBusy%%", (printed++ ? delim : ""));
2278 	if (DO_BIC(BIC_Bzy_MHz))
2279 		outp += sprintf(outp, "%sBzy_MHz", (printed++ ? delim : ""));
2280 	if (DO_BIC(BIC_TSC_MHz))
2281 		outp += sprintf(outp, "%sTSC_MHz", (printed++ ? delim : ""));
2282 
2283 	if (DO_BIC(BIC_IPC))
2284 		outp += sprintf(outp, "%sIPC", (printed++ ? delim : ""));
2285 
2286 	if (DO_BIC(BIC_IRQ)) {
2287 		if (sums_need_wide_columns)
2288 			outp += sprintf(outp, "%s     IRQ", (printed++ ? delim : ""));
2289 		else
2290 			outp += sprintf(outp, "%sIRQ", (printed++ ? delim : ""));
2291 	}
2292 
2293 	if (DO_BIC(BIC_SMI))
2294 		outp += sprintf(outp, "%sSMI", (printed++ ? delim : ""));
2295 
2296 	for (mp = sys.tp; mp; mp = mp->next) {
2297 
2298 		if (mp->format == FORMAT_RAW) {
2299 			if (mp->width == 64)
2300 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), mp->name);
2301 			else
2302 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), mp->name);
2303 		} else {
2304 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2305 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), mp->name);
2306 			else
2307 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), mp->name);
2308 		}
2309 	}
2310 
2311 	for (pp = sys.perf_tp; pp; pp = pp->next) {
2312 
2313 		if (pp->format == FORMAT_RAW) {
2314 			if (pp->width == 64)
2315 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), pp->name);
2316 			else
2317 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), pp->name);
2318 		} else {
2319 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2320 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), pp->name);
2321 			else
2322 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), pp->name);
2323 		}
2324 	}
2325 
2326 	ppmt = sys.pmt_tp;
2327 	while (ppmt) {
2328 		switch (ppmt->type) {
2329 		case PMT_TYPE_RAW:
2330 			if (pmt_counter_get_width(ppmt) <= 32)
2331 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), ppmt->name);
2332 			else
2333 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), ppmt->name);
2334 
2335 			break;
2336 
2337 		case PMT_TYPE_XTAL_TIME:
2338 			outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), ppmt->name);
2339 			break;
2340 		}
2341 
2342 		ppmt = ppmt->next;
2343 	}
2344 
2345 	if (DO_BIC(BIC_CPU_c1))
2346 		outp += sprintf(outp, "%sCPU%%c1", (printed++ ? delim : ""));
2347 	if (DO_BIC(BIC_CPU_c3))
2348 		outp += sprintf(outp, "%sCPU%%c3", (printed++ ? delim : ""));
2349 	if (DO_BIC(BIC_CPU_c6))
2350 		outp += sprintf(outp, "%sCPU%%c6", (printed++ ? delim : ""));
2351 	if (DO_BIC(BIC_CPU_c7))
2352 		outp += sprintf(outp, "%sCPU%%c7", (printed++ ? delim : ""));
2353 
2354 	if (DO_BIC(BIC_Mod_c6))
2355 		outp += sprintf(outp, "%sMod%%c6", (printed++ ? delim : ""));
2356 
2357 	if (DO_BIC(BIC_CoreTmp))
2358 		outp += sprintf(outp, "%sCoreTmp", (printed++ ? delim : ""));
2359 
2360 	if (DO_BIC(BIC_CORE_THROT_CNT))
2361 		outp += sprintf(outp, "%sCoreThr", (printed++ ? delim : ""));
2362 
2363 	if (platform->rapl_msrs && !rapl_joules) {
2364 		if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
2365 			outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
2366 	} else if (platform->rapl_msrs && rapl_joules) {
2367 		if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
2368 			outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
2369 	}
2370 
2371 	for (mp = sys.cp; mp; mp = mp->next) {
2372 		if (mp->format == FORMAT_RAW) {
2373 			if (mp->width == 64)
2374 				outp += sprintf(outp, "%s%18.18s", delim, mp->name);
2375 			else
2376 				outp += sprintf(outp, "%s%10.10s", delim, mp->name);
2377 		} else {
2378 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2379 				outp += sprintf(outp, "%s%8s", delim, mp->name);
2380 			else
2381 				outp += sprintf(outp, "%s%s", delim, mp->name);
2382 		}
2383 	}
2384 
2385 	for (pp = sys.perf_cp; pp; pp = pp->next) {
2386 
2387 		if (pp->format == FORMAT_RAW) {
2388 			if (pp->width == 64)
2389 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), pp->name);
2390 			else
2391 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), pp->name);
2392 		} else {
2393 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2394 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), pp->name);
2395 			else
2396 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), pp->name);
2397 		}
2398 	}
2399 
2400 	ppmt = sys.pmt_cp;
2401 	while (ppmt) {
2402 		switch (ppmt->type) {
2403 		case PMT_TYPE_RAW:
2404 			if (pmt_counter_get_width(ppmt) <= 32)
2405 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), ppmt->name);
2406 			else
2407 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), ppmt->name);
2408 
2409 			break;
2410 
2411 		case PMT_TYPE_XTAL_TIME:
2412 			outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), ppmt->name);
2413 			break;
2414 		}
2415 
2416 		ppmt = ppmt->next;
2417 	}
2418 
2419 	if (DO_BIC(BIC_PkgTmp))
2420 		outp += sprintf(outp, "%sPkgTmp", (printed++ ? delim : ""));
2421 
2422 	if (DO_BIC(BIC_GFX_rc6))
2423 		outp += sprintf(outp, "%sGFX%%rc6", (printed++ ? delim : ""));
2424 
2425 	if (DO_BIC(BIC_GFXMHz))
2426 		outp += sprintf(outp, "%sGFXMHz", (printed++ ? delim : ""));
2427 
2428 	if (DO_BIC(BIC_GFXACTMHz))
2429 		outp += sprintf(outp, "%sGFXAMHz", (printed++ ? delim : ""));
2430 
2431 	if (DO_BIC(BIC_SAM_mc6))
2432 		outp += sprintf(outp, "%sSAM%%mc6", (printed++ ? delim : ""));
2433 
2434 	if (DO_BIC(BIC_SAMMHz))
2435 		outp += sprintf(outp, "%sSAMMHz", (printed++ ? delim : ""));
2436 
2437 	if (DO_BIC(BIC_SAMACTMHz))
2438 		outp += sprintf(outp, "%sSAMAMHz", (printed++ ? delim : ""));
2439 
2440 	if (DO_BIC(BIC_Totl_c0))
2441 		outp += sprintf(outp, "%sTotl%%C0", (printed++ ? delim : ""));
2442 	if (DO_BIC(BIC_Any_c0))
2443 		outp += sprintf(outp, "%sAny%%C0", (printed++ ? delim : ""));
2444 	if (DO_BIC(BIC_GFX_c0))
2445 		outp += sprintf(outp, "%sGFX%%C0", (printed++ ? delim : ""));
2446 	if (DO_BIC(BIC_CPUGFX))
2447 		outp += sprintf(outp, "%sCPUGFX%%", (printed++ ? delim : ""));
2448 
2449 	if (DO_BIC(BIC_Pkgpc2))
2450 		outp += sprintf(outp, "%sPkg%%pc2", (printed++ ? delim : ""));
2451 	if (DO_BIC(BIC_Pkgpc3))
2452 		outp += sprintf(outp, "%sPkg%%pc3", (printed++ ? delim : ""));
2453 	if (DO_BIC(BIC_Pkgpc6))
2454 		outp += sprintf(outp, "%sPkg%%pc6", (printed++ ? delim : ""));
2455 	if (DO_BIC(BIC_Pkgpc7))
2456 		outp += sprintf(outp, "%sPkg%%pc7", (printed++ ? delim : ""));
2457 	if (DO_BIC(BIC_Pkgpc8))
2458 		outp += sprintf(outp, "%sPkg%%pc8", (printed++ ? delim : ""));
2459 	if (DO_BIC(BIC_Pkgpc9))
2460 		outp += sprintf(outp, "%sPkg%%pc9", (printed++ ? delim : ""));
2461 	if (DO_BIC(BIC_Pkgpc10))
2462 		outp += sprintf(outp, "%sPk%%pc10", (printed++ ? delim : ""));
2463 	if (DO_BIC(BIC_Diec6))
2464 		outp += sprintf(outp, "%sDie%%c6", (printed++ ? delim : ""));
2465 	if (DO_BIC(BIC_CPU_LPI))
2466 		outp += sprintf(outp, "%sCPU%%LPI", (printed++ ? delim : ""));
2467 	if (DO_BIC(BIC_SYS_LPI))
2468 		outp += sprintf(outp, "%sSYS%%LPI", (printed++ ? delim : ""));
2469 
2470 	if (platform->rapl_msrs && !rapl_joules) {
2471 		if (DO_BIC(BIC_PkgWatt))
2472 			outp += sprintf(outp, "%sPkgWatt", (printed++ ? delim : ""));
2473 		if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
2474 			outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
2475 		if (DO_BIC(BIC_GFXWatt))
2476 			outp += sprintf(outp, "%sGFXWatt", (printed++ ? delim : ""));
2477 		if (DO_BIC(BIC_RAMWatt))
2478 			outp += sprintf(outp, "%sRAMWatt", (printed++ ? delim : ""));
2479 		if (DO_BIC(BIC_PKG__))
2480 			outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
2481 		if (DO_BIC(BIC_RAM__))
2482 			outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
2483 	} else if (platform->rapl_msrs && rapl_joules) {
2484 		if (DO_BIC(BIC_Pkg_J))
2485 			outp += sprintf(outp, "%sPkg_J", (printed++ ? delim : ""));
2486 		if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
2487 			outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
2488 		if (DO_BIC(BIC_GFX_J))
2489 			outp += sprintf(outp, "%sGFX_J", (printed++ ? delim : ""));
2490 		if (DO_BIC(BIC_RAM_J))
2491 			outp += sprintf(outp, "%sRAM_J", (printed++ ? delim : ""));
2492 		if (DO_BIC(BIC_PKG__))
2493 			outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
2494 		if (DO_BIC(BIC_RAM__))
2495 			outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
2496 	}
2497 	if (DO_BIC(BIC_UNCORE_MHZ))
2498 		outp += sprintf(outp, "%sUncMHz", (printed++ ? delim : ""));
2499 
2500 	for (mp = sys.pp; mp; mp = mp->next) {
2501 		if (mp->format == FORMAT_RAW) {
2502 			if (mp->width == 64)
2503 				outp += sprintf(outp, "%s%18.18s", delim, mp->name);
2504 			else if (mp->width == 32)
2505 				outp += sprintf(outp, "%s%10.10s", delim, mp->name);
2506 			else
2507 				outp += sprintf(outp, "%s%7.7s", delim, mp->name);
2508 		} else {
2509 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2510 				outp += sprintf(outp, "%s%8s", delim, mp->name);
2511 			else
2512 				outp += sprintf(outp, "%s%7.7s", delim, mp->name);
2513 		}
2514 	}
2515 
2516 	for (pp = sys.perf_pp; pp; pp = pp->next) {
2517 
2518 		if (pp->format == FORMAT_RAW) {
2519 			if (pp->width == 64)
2520 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), pp->name);
2521 			else
2522 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), pp->name);
2523 		} else {
2524 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2525 				outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), pp->name);
2526 			else
2527 				outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), pp->name);
2528 		}
2529 	}
2530 
2531 	ppmt = sys.pmt_pp;
2532 	while (ppmt) {
2533 		switch (ppmt->type) {
2534 		case PMT_TYPE_RAW:
2535 			if (pmt_counter_get_width(ppmt) <= 32)
2536 				outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), ppmt->name);
2537 			else
2538 				outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), ppmt->name);
2539 
2540 			break;
2541 
2542 		case PMT_TYPE_XTAL_TIME:
2543 			outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), ppmt->name);
2544 			break;
2545 		}
2546 
2547 		ppmt = ppmt->next;
2548 	}
2549 
2550 	if (DO_BIC(BIC_SysWatt))
2551 		outp += sprintf(outp, "%sSysWatt", (printed++ ? delim : ""));
2552 	if (DO_BIC(BIC_Sys_J))
2553 		outp += sprintf(outp, "%sSys_J", (printed++ ? delim : ""));
2554 
2555 	outp += sprintf(outp, "\n");
2556 }
2557 
dump_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)2558 int dump_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2559 {
2560 	int i;
2561 	struct msr_counter *mp;
2562 	struct platform_counters *pplat_cnt = p == package_odd ? &platform_counters_odd : &platform_counters_even;
2563 
2564 	outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
2565 
2566 	if (t) {
2567 		outp += sprintf(outp, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags);
2568 		outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
2569 		outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
2570 		outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
2571 		outp += sprintf(outp, "c1: %016llX\n", t->c1);
2572 
2573 		if (DO_BIC(BIC_IPC))
2574 			outp += sprintf(outp, "IPC: %lld\n", t->instr_count);
2575 
2576 		if (DO_BIC(BIC_IRQ))
2577 			outp += sprintf(outp, "IRQ: %lld\n", t->irq_count);
2578 		if (DO_BIC(BIC_SMI))
2579 			outp += sprintf(outp, "SMI: %d\n", t->smi_count);
2580 
2581 		for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
2582 			outp +=
2583 			    sprintf(outp, "tADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num,
2584 				    t->counter[i], mp->sp->path);
2585 		}
2586 	}
2587 
2588 	if (c && is_cpu_first_thread_in_core(t, c, p)) {
2589 		outp += sprintf(outp, "core: %d\n", c->core_id);
2590 		outp += sprintf(outp, "c3: %016llX\n", c->c3);
2591 		outp += sprintf(outp, "c6: %016llX\n", c->c6);
2592 		outp += sprintf(outp, "c7: %016llX\n", c->c7);
2593 		outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
2594 		outp += sprintf(outp, "cpu_throt_count: %016llX\n", c->core_throt_cnt);
2595 
2596 		const unsigned long long energy_value = c->core_energy.raw_value * c->core_energy.scale;
2597 		const double energy_scale = c->core_energy.scale;
2598 
2599 		if (c->core_energy.unit == RAPL_UNIT_JOULES)
2600 			outp += sprintf(outp, "Joules: %0llX (scale: %lf)\n", energy_value, energy_scale);
2601 
2602 		for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
2603 			outp +=
2604 			    sprintf(outp, "cADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num,
2605 				    c->counter[i], mp->sp->path);
2606 		}
2607 		outp += sprintf(outp, "mc6_us: %016llX\n", c->mc6_us);
2608 	}
2609 
2610 	if (p && is_cpu_first_core_in_package(t, c, p)) {
2611 		outp += sprintf(outp, "package: %d\n", p->package_id);
2612 
2613 		outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
2614 		outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
2615 		outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
2616 		outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
2617 
2618 		outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
2619 		if (DO_BIC(BIC_Pkgpc3))
2620 			outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
2621 		if (DO_BIC(BIC_Pkgpc6))
2622 			outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
2623 		if (DO_BIC(BIC_Pkgpc7))
2624 			outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
2625 		outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
2626 		outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
2627 		outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
2628 		outp += sprintf(outp, "cpu_lpi: %016llX\n", p->cpu_lpi);
2629 		outp += sprintf(outp, "sys_lpi: %016llX\n", p->sys_lpi);
2630 		outp += sprintf(outp, "Joules PKG: %0llX\n", p->energy_pkg.raw_value);
2631 		outp += sprintf(outp, "Joules COR: %0llX\n", p->energy_cores.raw_value);
2632 		outp += sprintf(outp, "Joules GFX: %0llX\n", p->energy_gfx.raw_value);
2633 		outp += sprintf(outp, "Joules RAM: %0llX\n", p->energy_dram.raw_value);
2634 		outp += sprintf(outp, "Joules PSYS: %0llX\n", pplat_cnt->energy_psys.raw_value);
2635 		outp += sprintf(outp, "Throttle PKG: %0llX\n", p->rapl_pkg_perf_status.raw_value);
2636 		outp += sprintf(outp, "Throttle RAM: %0llX\n", p->rapl_dram_perf_status.raw_value);
2637 		outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
2638 
2639 		for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
2640 			outp +=
2641 			    sprintf(outp, "pADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num,
2642 				    p->counter[i], mp->sp->path);
2643 		}
2644 	}
2645 
2646 	outp += sprintf(outp, "\n");
2647 
2648 	return 0;
2649 }
2650 
rapl_counter_get_value(const struct rapl_counter * c,enum rapl_unit desired_unit,double interval)2651 double rapl_counter_get_value(const struct rapl_counter *c, enum rapl_unit desired_unit, double interval)
2652 {
2653 	assert(desired_unit != RAPL_UNIT_INVALID);
2654 
2655 	/*
2656 	 * For now we don't expect anything other than joules,
2657 	 * so just simplify the logic.
2658 	 */
2659 	assert(c->unit == RAPL_UNIT_JOULES);
2660 
2661 	const double scaled = c->raw_value * c->scale;
2662 
2663 	if (desired_unit == RAPL_UNIT_WATTS)
2664 		return scaled / interval;
2665 	return scaled;
2666 }
2667 
2668 /*
2669  * column formatting convention & formats
2670  */
format_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)2671 int format_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2672 {
2673 	static int count;
2674 
2675 	struct platform_counters *pplat_cnt = NULL;
2676 	double interval_float, tsc;
2677 	char *fmt8;
2678 	int i;
2679 	struct msr_counter *mp;
2680 	struct perf_counter_info *pp;
2681 	struct pmt_counter *ppmt;
2682 	char *delim = "\t";
2683 	int printed = 0;
2684 
2685 	if (t == &average.threads) {
2686 		pplat_cnt = count & 1 ? &platform_counters_odd : &platform_counters_even;
2687 		++count;
2688 	}
2689 
2690 	/* if showing only 1st thread in core and this isn't one, bail out */
2691 	if (show_core_only && !is_cpu_first_thread_in_core(t, c, p))
2692 		return 0;
2693 
2694 	/* if showing only 1st thread in pkg and this isn't one, bail out */
2695 	if (show_pkg_only && !is_cpu_first_core_in_package(t, c, p))
2696 		return 0;
2697 
2698 	/*if not summary line and --cpu is used */
2699 	if ((t != &average.threads) && (cpu_subset && !CPU_ISSET_S(t->cpu_id, cpu_subset_size, cpu_subset)))
2700 		return 0;
2701 
2702 	if (DO_BIC(BIC_USEC)) {
2703 		/* on each row, print how many usec each timestamp took to gather */
2704 		struct timeval tv;
2705 
2706 		timersub(&t->tv_end, &t->tv_begin, &tv);
2707 		outp += sprintf(outp, "%5ld\t", tv.tv_sec * 1000000 + tv.tv_usec);
2708 	}
2709 
2710 	/* Time_Of_Day_Seconds: on each row, print sec.usec last timestamp taken */
2711 	if (DO_BIC(BIC_TOD))
2712 		outp += sprintf(outp, "%10ld.%06ld\t", t->tv_end.tv_sec, t->tv_end.tv_usec);
2713 
2714 	interval_float = t->tv_delta.tv_sec + t->tv_delta.tv_usec / 1000000.0;
2715 
2716 	tsc = t->tsc * tsc_tweak;
2717 
2718 	/* topo columns, print blanks on 1st (average) line */
2719 	if (t == &average.threads) {
2720 		if (DO_BIC(BIC_Package))
2721 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2722 		if (DO_BIC(BIC_Die))
2723 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2724 		if (DO_BIC(BIC_Node))
2725 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2726 		if (DO_BIC(BIC_Core))
2727 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2728 		if (DO_BIC(BIC_CPU))
2729 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2730 		if (DO_BIC(BIC_APIC))
2731 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2732 		if (DO_BIC(BIC_X2APIC))
2733 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2734 	} else {
2735 		if (DO_BIC(BIC_Package)) {
2736 			if (p)
2737 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->package_id);
2738 			else
2739 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2740 		}
2741 		if (DO_BIC(BIC_Die)) {
2742 			if (c)
2743 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].die_id);
2744 			else
2745 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2746 		}
2747 		if (DO_BIC(BIC_Node)) {
2748 			if (t)
2749 				outp += sprintf(outp, "%s%d",
2750 						(printed++ ? delim : ""), cpus[t->cpu_id].physical_node_id);
2751 			else
2752 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2753 		}
2754 		if (DO_BIC(BIC_Core)) {
2755 			if (c)
2756 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_id);
2757 			else
2758 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
2759 		}
2760 		if (DO_BIC(BIC_CPU))
2761 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->cpu_id);
2762 		if (DO_BIC(BIC_APIC))
2763 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->apic_id);
2764 		if (DO_BIC(BIC_X2APIC))
2765 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->x2apic_id);
2766 	}
2767 
2768 	if (DO_BIC(BIC_Avg_MHz))
2769 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 / units * t->aperf / interval_float);
2770 
2771 	if (DO_BIC(BIC_Busy))
2772 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->mperf / tsc);
2773 
2774 	if (DO_BIC(BIC_Bzy_MHz)) {
2775 		if (has_base_hz)
2776 			outp +=
2777 			    sprintf(outp, "%s%.0f", (printed++ ? delim : ""), base_hz / units * t->aperf / t->mperf);
2778 		else
2779 			outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""),
2780 					tsc / units * t->aperf / t->mperf / interval_float);
2781 	}
2782 
2783 	if (DO_BIC(BIC_TSC_MHz))
2784 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 * t->tsc / units / interval_float);
2785 
2786 	if (DO_BIC(BIC_IPC))
2787 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 1.0 * t->instr_count / t->aperf);
2788 
2789 	/* IRQ */
2790 	if (DO_BIC(BIC_IRQ)) {
2791 		if (sums_need_wide_columns)
2792 			outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->irq_count);
2793 		else
2794 			outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->irq_count);
2795 	}
2796 
2797 	/* SMI */
2798 	if (DO_BIC(BIC_SMI))
2799 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->smi_count);
2800 
2801 	/* Added counters */
2802 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
2803 		if (mp->format == FORMAT_RAW) {
2804 			if (mp->width == 32)
2805 				outp +=
2806 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int)t->counter[i]);
2807 			else
2808 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->counter[i]);
2809 		} else if (mp->format == FORMAT_DELTA) {
2810 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2811 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->counter[i]);
2812 			else
2813 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->counter[i]);
2814 		} else if (mp->format == FORMAT_PERCENT) {
2815 			if (mp->type == COUNTER_USEC)
2816 				outp +=
2817 				    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2818 					    t->counter[i] / interval_float / 10000);
2819 			else
2820 				outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->counter[i] / tsc);
2821 		}
2822 	}
2823 
2824 	/* Added perf counters */
2825 	for (i = 0, pp = sys.perf_tp; pp; ++i, pp = pp->next) {
2826 		if (pp->format == FORMAT_RAW) {
2827 			if (pp->width == 32)
2828 				outp +=
2829 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
2830 					    (unsigned int)t->perf_counter[i]);
2831 			else
2832 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->perf_counter[i]);
2833 		} else if (pp->format == FORMAT_DELTA) {
2834 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2835 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->perf_counter[i]);
2836 			else
2837 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->perf_counter[i]);
2838 		} else if (pp->format == FORMAT_PERCENT) {
2839 			if (pp->type == COUNTER_USEC)
2840 				outp +=
2841 				    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2842 					    t->perf_counter[i] / interval_float / 10000);
2843 			else
2844 				outp +=
2845 				    sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->perf_counter[i] / tsc);
2846 		}
2847 	}
2848 
2849 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
2850 		const unsigned long value_raw = t->pmt_counter[i];
2851 		const double value_converted = 100.0 * value_raw / crystal_hz / interval_float;
2852 		switch (ppmt->type) {
2853 		case PMT_TYPE_RAW:
2854 			if (pmt_counter_get_width(ppmt) <= 32)
2855 				outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
2856 						(unsigned int)t->pmt_counter[i]);
2857 			else
2858 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->pmt_counter[i]);
2859 
2860 			break;
2861 
2862 		case PMT_TYPE_XTAL_TIME:
2863 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
2864 			break;
2865 		}
2866 	}
2867 
2868 	/* C1 */
2869 	if (DO_BIC(BIC_CPU_c1))
2870 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->c1 / tsc);
2871 
2872 	/* print per-core data only for 1st thread in core */
2873 	if (!is_cpu_first_thread_in_core(t, c, p))
2874 		goto done;
2875 
2876 	if (DO_BIC(BIC_CPU_c3))
2877 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c3 / tsc);
2878 	if (DO_BIC(BIC_CPU_c6))
2879 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c6 / tsc);
2880 	if (DO_BIC(BIC_CPU_c7))
2881 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c7 / tsc);
2882 
2883 	/* Mod%c6 */
2884 	if (DO_BIC(BIC_Mod_c6))
2885 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->mc6_us / tsc);
2886 
2887 	if (DO_BIC(BIC_CoreTmp))
2888 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_temp_c);
2889 
2890 	/* Core throttle count */
2891 	if (DO_BIC(BIC_CORE_THROT_CNT))
2892 		outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->core_throt_cnt);
2893 
2894 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
2895 		if (mp->format == FORMAT_RAW) {
2896 			if (mp->width == 32)
2897 				outp +=
2898 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int)c->counter[i]);
2899 			else
2900 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->counter[i]);
2901 		} else if (mp->format == FORMAT_DELTA) {
2902 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2903 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), c->counter[i]);
2904 			else
2905 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->counter[i]);
2906 		} else if (mp->format == FORMAT_PERCENT) {
2907 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->counter[i] / tsc);
2908 		}
2909 	}
2910 
2911 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
2912 		if (pp->format == FORMAT_RAW) {
2913 			if (pp->width == 32)
2914 				outp +=
2915 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
2916 					    (unsigned int)c->perf_counter[i]);
2917 			else
2918 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->perf_counter[i]);
2919 		} else if (pp->format == FORMAT_DELTA) {
2920 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
2921 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), c->perf_counter[i]);
2922 			else
2923 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->perf_counter[i]);
2924 		} else if (pp->format == FORMAT_PERCENT) {
2925 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->perf_counter[i] / tsc);
2926 		}
2927 	}
2928 
2929 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
2930 		const unsigned long value_raw = c->pmt_counter[i];
2931 		const double value_converted = 100.0 * value_raw / crystal_hz / interval_float;
2932 		switch (ppmt->type) {
2933 		case PMT_TYPE_RAW:
2934 			if (pmt_counter_get_width(ppmt) <= 32)
2935 				outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
2936 						(unsigned int)c->pmt_counter[i]);
2937 			else
2938 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->pmt_counter[i]);
2939 
2940 			break;
2941 
2942 		case PMT_TYPE_XTAL_TIME:
2943 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
2944 			break;
2945 		}
2946 	}
2947 
2948 	fmt8 = "%s%.2f";
2949 
2950 	if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
2951 		outp +=
2952 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
2953 			    rapl_counter_get_value(&c->core_energy, RAPL_UNIT_WATTS, interval_float));
2954 	if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
2955 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
2956 				rapl_counter_get_value(&c->core_energy, RAPL_UNIT_JOULES, interval_float));
2957 
2958 	/* print per-package data only for 1st core in package */
2959 	if (!is_cpu_first_core_in_package(t, c, p))
2960 		goto done;
2961 
2962 	/* PkgTmp */
2963 	if (DO_BIC(BIC_PkgTmp))
2964 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->pkg_temp_c);
2965 
2966 	/* GFXrc6 */
2967 	if (DO_BIC(BIC_GFX_rc6)) {
2968 		if (p->gfx_rc6_ms == -1) {	/* detect GFX counter reset */
2969 			outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
2970 		} else {
2971 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2972 					p->gfx_rc6_ms / 10.0 / interval_float);
2973 		}
2974 	}
2975 
2976 	/* GFXMHz */
2977 	if (DO_BIC(BIC_GFXMHz))
2978 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_mhz);
2979 
2980 	/* GFXACTMHz */
2981 	if (DO_BIC(BIC_GFXACTMHz))
2982 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_act_mhz);
2983 
2984 	/* SAMmc6 */
2985 	if (DO_BIC(BIC_SAM_mc6)) {
2986 		if (p->sam_mc6_ms == -1) {	/* detect GFX counter reset */
2987 			outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
2988 		} else {
2989 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
2990 					p->sam_mc6_ms / 10.0 / interval_float);
2991 		}
2992 	}
2993 
2994 	/* SAMMHz */
2995 	if (DO_BIC(BIC_SAMMHz))
2996 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->sam_mhz);
2997 
2998 	/* SAMACTMHz */
2999 	if (DO_BIC(BIC_SAMACTMHz))
3000 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->sam_act_mhz);
3001 
3002 	/* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
3003 	if (DO_BIC(BIC_Totl_c0))
3004 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_wtd_core_c0 / tsc);
3005 	if (DO_BIC(BIC_Any_c0))
3006 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_core_c0 / tsc);
3007 	if (DO_BIC(BIC_GFX_c0))
3008 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_gfxe_c0 / tsc);
3009 	if (DO_BIC(BIC_CPUGFX))
3010 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_both_core_gfxe_c0 / tsc);
3011 
3012 	if (DO_BIC(BIC_Pkgpc2))
3013 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc2 / tsc);
3014 	if (DO_BIC(BIC_Pkgpc3))
3015 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc3 / tsc);
3016 	if (DO_BIC(BIC_Pkgpc6))
3017 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc6 / tsc);
3018 	if (DO_BIC(BIC_Pkgpc7))
3019 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc7 / tsc);
3020 	if (DO_BIC(BIC_Pkgpc8))
3021 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc8 / tsc);
3022 	if (DO_BIC(BIC_Pkgpc9))
3023 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc9 / tsc);
3024 	if (DO_BIC(BIC_Pkgpc10))
3025 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc10 / tsc);
3026 
3027 	if (DO_BIC(BIC_Diec6))
3028 		outp +=
3029 		    sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->die_c6 / crystal_hz / interval_float);
3030 
3031 	if (DO_BIC(BIC_CPU_LPI)) {
3032 		if (p->cpu_lpi >= 0)
3033 			outp +=
3034 			    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
3035 				    100.0 * p->cpu_lpi / 1000000.0 / interval_float);
3036 		else
3037 			outp += sprintf(outp, "%s(neg)", (printed++ ? delim : ""));
3038 	}
3039 	if (DO_BIC(BIC_SYS_LPI)) {
3040 		if (p->sys_lpi >= 0)
3041 			outp +=
3042 			    sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
3043 				    100.0 * p->sys_lpi / 1000000.0 / interval_float);
3044 		else
3045 			outp += sprintf(outp, "%s(neg)", (printed++ ? delim : ""));
3046 	}
3047 
3048 	if (DO_BIC(BIC_PkgWatt))
3049 		outp +=
3050 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3051 			    rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_WATTS, interval_float));
3052 	if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
3053 		outp +=
3054 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3055 			    rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_WATTS, interval_float));
3056 	if (DO_BIC(BIC_GFXWatt))
3057 		outp +=
3058 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3059 			    rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_WATTS, interval_float));
3060 	if (DO_BIC(BIC_RAMWatt))
3061 		outp +=
3062 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3063 			    rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_WATTS, interval_float));
3064 	if (DO_BIC(BIC_Pkg_J))
3065 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3066 				rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_JOULES, interval_float));
3067 	if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
3068 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3069 				rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_JOULES, interval_float));
3070 	if (DO_BIC(BIC_GFX_J))
3071 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3072 				rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_JOULES, interval_float));
3073 	if (DO_BIC(BIC_RAM_J))
3074 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3075 				rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_JOULES, interval_float));
3076 	if (DO_BIC(BIC_PKG__))
3077 		outp +=
3078 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3079 			    rapl_counter_get_value(&p->rapl_pkg_perf_status, RAPL_UNIT_WATTS, interval_float));
3080 	if (DO_BIC(BIC_RAM__))
3081 		outp +=
3082 		    sprintf(outp, fmt8, (printed++ ? delim : ""),
3083 			    rapl_counter_get_value(&p->rapl_dram_perf_status, RAPL_UNIT_WATTS, interval_float));
3084 	/* UncMHz */
3085 	if (DO_BIC(BIC_UNCORE_MHZ))
3086 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->uncore_mhz);
3087 
3088 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3089 		if (mp->format == FORMAT_RAW) {
3090 			if (mp->width == 32)
3091 				outp +=
3092 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int)p->counter[i]);
3093 			else
3094 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->counter[i]);
3095 		} else if (mp->format == FORMAT_DELTA) {
3096 			if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
3097 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), p->counter[i]);
3098 			else
3099 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), p->counter[i]);
3100 		} else if (mp->format == FORMAT_PERCENT) {
3101 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->counter[i] / tsc);
3102 		} else if (mp->type == COUNTER_K2M)
3103 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), (unsigned int)p->counter[i] / 1000);
3104 	}
3105 
3106 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3107 		if (pp->format == FORMAT_RAW) {
3108 			if (pp->width == 32)
3109 				outp +=
3110 				    sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
3111 					    (unsigned int)p->perf_counter[i]);
3112 			else
3113 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->perf_counter[i]);
3114 		} else if (pp->format == FORMAT_DELTA) {
3115 			if ((pp->type == COUNTER_ITEMS) && sums_need_wide_columns)
3116 				outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), p->perf_counter[i]);
3117 			else
3118 				outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), p->perf_counter[i]);
3119 		} else if (pp->format == FORMAT_PERCENT) {
3120 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->perf_counter[i] / tsc);
3121 		} else if (pp->type == COUNTER_K2M) {
3122 			outp +=
3123 			    sprintf(outp, "%s%d", (printed++ ? delim : ""), (unsigned int)p->perf_counter[i] / 1000);
3124 		}
3125 	}
3126 
3127 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3128 		const unsigned long value_raw = p->pmt_counter[i];
3129 		const double value_converted = 100.0 * value_raw / crystal_hz / interval_float;
3130 		switch (ppmt->type) {
3131 		case PMT_TYPE_RAW:
3132 			if (pmt_counter_get_width(ppmt) <= 32)
3133 				outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""),
3134 						(unsigned int)p->pmt_counter[i]);
3135 			else
3136 				outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->pmt_counter[i]);
3137 
3138 			break;
3139 
3140 		case PMT_TYPE_XTAL_TIME:
3141 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
3142 			break;
3143 		}
3144 	}
3145 
3146 	if (DO_BIC(BIC_SysWatt) && (t == &average.threads))
3147 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3148 				rapl_counter_get_value(&pplat_cnt->energy_psys, RAPL_UNIT_WATTS, interval_float));
3149 	if (DO_BIC(BIC_Sys_J) && (t == &average.threads))
3150 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
3151 				rapl_counter_get_value(&pplat_cnt->energy_psys, RAPL_UNIT_JOULES, interval_float));
3152 
3153 done:
3154 	if (*(outp - 1) != '\n')
3155 		outp += sprintf(outp, "\n");
3156 
3157 	return 0;
3158 }
3159 
flush_output_stdout(void)3160 void flush_output_stdout(void)
3161 {
3162 	FILE *filep;
3163 
3164 	if (outf == stderr)
3165 		filep = stdout;
3166 	else
3167 		filep = outf;
3168 
3169 	fputs(output_buffer, filep);
3170 	fflush(filep);
3171 
3172 	outp = output_buffer;
3173 }
3174 
flush_output_stderr(void)3175 void flush_output_stderr(void)
3176 {
3177 	fputs(output_buffer, outf);
3178 	fflush(outf);
3179 	outp = output_buffer;
3180 }
3181 
format_all_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)3182 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3183 {
3184 	static int count;
3185 
3186 	if ((!count || (header_iterations && !(count % header_iterations))) || !summary_only)
3187 		print_header("\t");
3188 
3189 	format_counters(&average.threads, &average.cores, &average.packages);
3190 
3191 	count++;
3192 
3193 	if (summary_only)
3194 		return;
3195 
3196 	for_all_cpus(format_counters, t, c, p);
3197 }
3198 
3199 #define DELTA_WRAP32(new, old)			\
3200 	old = ((((unsigned long long)new << 32) - ((unsigned long long)old << 32)) >> 32);
3201 
delta_package(struct pkg_data * new,struct pkg_data * old)3202 int delta_package(struct pkg_data *new, struct pkg_data *old)
3203 {
3204 	int i;
3205 	struct msr_counter *mp;
3206 	struct perf_counter_info *pp;
3207 	struct pmt_counter *ppmt;
3208 
3209 	if (DO_BIC(BIC_Totl_c0))
3210 		old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
3211 	if (DO_BIC(BIC_Any_c0))
3212 		old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
3213 	if (DO_BIC(BIC_GFX_c0))
3214 		old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
3215 	if (DO_BIC(BIC_CPUGFX))
3216 		old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
3217 
3218 	old->pc2 = new->pc2 - old->pc2;
3219 	if (DO_BIC(BIC_Pkgpc3))
3220 		old->pc3 = new->pc3 - old->pc3;
3221 	if (DO_BIC(BIC_Pkgpc6))
3222 		old->pc6 = new->pc6 - old->pc6;
3223 	if (DO_BIC(BIC_Pkgpc7))
3224 		old->pc7 = new->pc7 - old->pc7;
3225 	old->pc8 = new->pc8 - old->pc8;
3226 	old->pc9 = new->pc9 - old->pc9;
3227 	old->pc10 = new->pc10 - old->pc10;
3228 	old->die_c6 = new->die_c6 - old->die_c6;
3229 	old->cpu_lpi = new->cpu_lpi - old->cpu_lpi;
3230 	old->sys_lpi = new->sys_lpi - old->sys_lpi;
3231 	old->pkg_temp_c = new->pkg_temp_c;
3232 
3233 	/* flag an error when rc6 counter resets/wraps */
3234 	if (old->gfx_rc6_ms > new->gfx_rc6_ms)
3235 		old->gfx_rc6_ms = -1;
3236 	else
3237 		old->gfx_rc6_ms = new->gfx_rc6_ms - old->gfx_rc6_ms;
3238 
3239 	old->uncore_mhz = new->uncore_mhz;
3240 	old->gfx_mhz = new->gfx_mhz;
3241 	old->gfx_act_mhz = new->gfx_act_mhz;
3242 
3243 	/* flag an error when mc6 counter resets/wraps */
3244 	if (old->sam_mc6_ms > new->sam_mc6_ms)
3245 		old->sam_mc6_ms = -1;
3246 	else
3247 		old->sam_mc6_ms = new->sam_mc6_ms - old->sam_mc6_ms;
3248 
3249 	old->sam_mhz = new->sam_mhz;
3250 	old->sam_act_mhz = new->sam_act_mhz;
3251 
3252 	old->energy_pkg.raw_value = new->energy_pkg.raw_value - old->energy_pkg.raw_value;
3253 	old->energy_cores.raw_value = new->energy_cores.raw_value - old->energy_cores.raw_value;
3254 	old->energy_gfx.raw_value = new->energy_gfx.raw_value - old->energy_gfx.raw_value;
3255 	old->energy_dram.raw_value = new->energy_dram.raw_value - old->energy_dram.raw_value;
3256 	old->rapl_pkg_perf_status.raw_value = new->rapl_pkg_perf_status.raw_value - old->rapl_pkg_perf_status.raw_value;
3257 	old->rapl_dram_perf_status.raw_value =
3258 	    new->rapl_dram_perf_status.raw_value - old->rapl_dram_perf_status.raw_value;
3259 
3260 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3261 		if (mp->format == FORMAT_RAW)
3262 			old->counter[i] = new->counter[i];
3263 		else if (mp->format == FORMAT_AVERAGE)
3264 			old->counter[i] = new->counter[i];
3265 		else
3266 			old->counter[i] = new->counter[i] - old->counter[i];
3267 	}
3268 
3269 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3270 		if (pp->format == FORMAT_RAW)
3271 			old->perf_counter[i] = new->perf_counter[i];
3272 		else if (pp->format == FORMAT_AVERAGE)
3273 			old->perf_counter[i] = new->perf_counter[i];
3274 		else
3275 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3276 	}
3277 
3278 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3279 		if (ppmt->format == FORMAT_RAW)
3280 			old->pmt_counter[i] = new->pmt_counter[i];
3281 		else
3282 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3283 	}
3284 
3285 	return 0;
3286 }
3287 
delta_core(struct core_data * new,struct core_data * old)3288 void delta_core(struct core_data *new, struct core_data *old)
3289 {
3290 	int i;
3291 	struct msr_counter *mp;
3292 	struct perf_counter_info *pp;
3293 	struct pmt_counter *ppmt;
3294 
3295 	old->c3 = new->c3 - old->c3;
3296 	old->c6 = new->c6 - old->c6;
3297 	old->c7 = new->c7 - old->c7;
3298 	old->core_temp_c = new->core_temp_c;
3299 	old->core_throt_cnt = new->core_throt_cnt;
3300 	old->mc6_us = new->mc6_us - old->mc6_us;
3301 
3302 	DELTA_WRAP32(new->core_energy.raw_value, old->core_energy.raw_value);
3303 
3304 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3305 		if (mp->format == FORMAT_RAW)
3306 			old->counter[i] = new->counter[i];
3307 		else
3308 			old->counter[i] = new->counter[i] - old->counter[i];
3309 	}
3310 
3311 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3312 		if (pp->format == FORMAT_RAW)
3313 			old->perf_counter[i] = new->perf_counter[i];
3314 		else
3315 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3316 	}
3317 
3318 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3319 		if (ppmt->format == FORMAT_RAW)
3320 			old->pmt_counter[i] = new->pmt_counter[i];
3321 		else
3322 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3323 	}
3324 }
3325 
soft_c1_residency_display(int bic)3326 int soft_c1_residency_display(int bic)
3327 {
3328 	if (!DO_BIC(BIC_CPU_c1) || platform->has_msr_core_c1_res)
3329 		return 0;
3330 
3331 	return DO_BIC_READ(bic);
3332 }
3333 
3334 /*
3335  * old = new - old
3336  */
delta_thread(struct thread_data * new,struct thread_data * old,struct core_data * core_delta)3337 int delta_thread(struct thread_data *new, struct thread_data *old, struct core_data *core_delta)
3338 {
3339 	int i;
3340 	struct msr_counter *mp;
3341 	struct perf_counter_info *pp;
3342 	struct pmt_counter *ppmt;
3343 
3344 	/* we run cpuid just the 1st time, copy the results */
3345 	if (DO_BIC(BIC_APIC))
3346 		new->apic_id = old->apic_id;
3347 	if (DO_BIC(BIC_X2APIC))
3348 		new->x2apic_id = old->x2apic_id;
3349 
3350 	/*
3351 	 * the timestamps from start of measurement interval are in "old"
3352 	 * the timestamp from end of measurement interval are in "new"
3353 	 * over-write old w/ new so we can print end of interval values
3354 	 */
3355 
3356 	timersub(&new->tv_begin, &old->tv_begin, &old->tv_delta);
3357 	old->tv_begin = new->tv_begin;
3358 	old->tv_end = new->tv_end;
3359 
3360 	old->tsc = new->tsc - old->tsc;
3361 
3362 	/* check for TSC < 1 Mcycles over interval */
3363 	if (old->tsc < (1000 * 1000))
3364 		errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
3365 		     "You can disable all c-states by booting with \"idle=poll\"\n"
3366 		     "or just the deep ones with \"processor.max_cstate=1\"");
3367 
3368 	old->c1 = new->c1 - old->c1;
3369 
3370 	if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz) || DO_BIC(BIC_IPC)
3371 	    || soft_c1_residency_display(BIC_Avg_MHz)) {
3372 		if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
3373 			old->aperf = new->aperf - old->aperf;
3374 			old->mperf = new->mperf - old->mperf;
3375 		} else {
3376 			return -1;
3377 		}
3378 	}
3379 
3380 	if (platform->has_msr_core_c1_res) {
3381 		/*
3382 		 * Some models have a dedicated C1 residency MSR,
3383 		 * which should be more accurate than the derivation below.
3384 		 */
3385 	} else {
3386 		/*
3387 		 * As counter collection is not atomic,
3388 		 * it is possible for mperf's non-halted cycles + idle states
3389 		 * to exceed TSC's all cycles: show c1 = 0% in that case.
3390 		 */
3391 		if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > (old->tsc * tsc_tweak))
3392 			old->c1 = 0;
3393 		else {
3394 			/* normal case, derive c1 */
3395 			old->c1 = (old->tsc * tsc_tweak) - old->mperf - core_delta->c3
3396 			    - core_delta->c6 - core_delta->c7;
3397 		}
3398 	}
3399 
3400 	if (old->mperf == 0) {
3401 		if (debug > 1)
3402 			fprintf(outf, "cpu%d MPERF 0!\n", old->cpu_id);
3403 		old->mperf = 1;	/* divide by 0 protection */
3404 	}
3405 
3406 	if (DO_BIC(BIC_IPC))
3407 		old->instr_count = new->instr_count - old->instr_count;
3408 
3409 	if (DO_BIC(BIC_IRQ))
3410 		old->irq_count = new->irq_count - old->irq_count;
3411 
3412 	if (DO_BIC(BIC_SMI))
3413 		old->smi_count = new->smi_count - old->smi_count;
3414 
3415 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3416 		if (mp->format == FORMAT_RAW)
3417 			old->counter[i] = new->counter[i];
3418 		else
3419 			old->counter[i] = new->counter[i] - old->counter[i];
3420 	}
3421 
3422 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
3423 		if (pp->format == FORMAT_RAW)
3424 			old->perf_counter[i] = new->perf_counter[i];
3425 		else
3426 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3427 	}
3428 
3429 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3430 		if (ppmt->format == FORMAT_RAW)
3431 			old->pmt_counter[i] = new->pmt_counter[i];
3432 		else
3433 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3434 	}
3435 
3436 	return 0;
3437 }
3438 
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)3439 int delta_cpu(struct thread_data *t, struct core_data *c,
3440 	      struct pkg_data *p, struct thread_data *t2, struct core_data *c2, struct pkg_data *p2)
3441 {
3442 	int retval = 0;
3443 
3444 	/* calculate core delta only for 1st thread in core */
3445 	if (is_cpu_first_thread_in_core(t, c, p))
3446 		delta_core(c, c2);
3447 
3448 	/* always calculate thread delta */
3449 	retval = delta_thread(t, t2, c2);	/* c2 is core delta */
3450 	if (retval)
3451 		return retval;
3452 
3453 	/* calculate package delta only for 1st core in package */
3454 	if (is_cpu_first_core_in_package(t, c, p))
3455 		retval = delta_package(p, p2);
3456 
3457 	return retval;
3458 }
3459 
delta_platform(struct platform_counters * new,struct platform_counters * old)3460 void delta_platform(struct platform_counters *new, struct platform_counters *old)
3461 {
3462 	old->energy_psys.raw_value = new->energy_psys.raw_value - old->energy_psys.raw_value;
3463 }
3464 
rapl_counter_clear(struct rapl_counter * c)3465 void rapl_counter_clear(struct rapl_counter *c)
3466 {
3467 	c->raw_value = 0;
3468 	c->scale = 0.0;
3469 	c->unit = RAPL_UNIT_INVALID;
3470 }
3471 
clear_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)3472 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3473 {
3474 	int i;
3475 	struct msr_counter *mp;
3476 
3477 	t->tv_begin.tv_sec = 0;
3478 	t->tv_begin.tv_usec = 0;
3479 	t->tv_end.tv_sec = 0;
3480 	t->tv_end.tv_usec = 0;
3481 	t->tv_delta.tv_sec = 0;
3482 	t->tv_delta.tv_usec = 0;
3483 
3484 	t->tsc = 0;
3485 	t->aperf = 0;
3486 	t->mperf = 0;
3487 	t->c1 = 0;
3488 
3489 	t->instr_count = 0;
3490 
3491 	t->irq_count = 0;
3492 	t->smi_count = 0;
3493 
3494 	c->c3 = 0;
3495 	c->c6 = 0;
3496 	c->c7 = 0;
3497 	c->mc6_us = 0;
3498 	c->core_temp_c = 0;
3499 	rapl_counter_clear(&c->core_energy);
3500 	c->core_throt_cnt = 0;
3501 
3502 	p->pkg_wtd_core_c0 = 0;
3503 	p->pkg_any_core_c0 = 0;
3504 	p->pkg_any_gfxe_c0 = 0;
3505 	p->pkg_both_core_gfxe_c0 = 0;
3506 
3507 	p->pc2 = 0;
3508 	if (DO_BIC(BIC_Pkgpc3))
3509 		p->pc3 = 0;
3510 	if (DO_BIC(BIC_Pkgpc6))
3511 		p->pc6 = 0;
3512 	if (DO_BIC(BIC_Pkgpc7))
3513 		p->pc7 = 0;
3514 	p->pc8 = 0;
3515 	p->pc9 = 0;
3516 	p->pc10 = 0;
3517 	p->die_c6 = 0;
3518 	p->cpu_lpi = 0;
3519 	p->sys_lpi = 0;
3520 
3521 	rapl_counter_clear(&p->energy_pkg);
3522 	rapl_counter_clear(&p->energy_dram);
3523 	rapl_counter_clear(&p->energy_cores);
3524 	rapl_counter_clear(&p->energy_gfx);
3525 	rapl_counter_clear(&p->rapl_pkg_perf_status);
3526 	rapl_counter_clear(&p->rapl_dram_perf_status);
3527 	p->pkg_temp_c = 0;
3528 
3529 	p->gfx_rc6_ms = 0;
3530 	p->uncore_mhz = 0;
3531 	p->gfx_mhz = 0;
3532 	p->gfx_act_mhz = 0;
3533 	p->sam_mc6_ms = 0;
3534 	p->sam_mhz = 0;
3535 	p->sam_act_mhz = 0;
3536 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next)
3537 		t->counter[i] = 0;
3538 
3539 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next)
3540 		c->counter[i] = 0;
3541 
3542 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next)
3543 		p->counter[i] = 0;
3544 
3545 	memset(&t->perf_counter[0], 0, sizeof(t->perf_counter));
3546 	memset(&c->perf_counter[0], 0, sizeof(c->perf_counter));
3547 	memset(&p->perf_counter[0], 0, sizeof(p->perf_counter));
3548 
3549 	memset(&t->pmt_counter[0], 0, ARRAY_SIZE(t->pmt_counter));
3550 	memset(&c->pmt_counter[0], 0, ARRAY_SIZE(c->pmt_counter));
3551 	memset(&p->pmt_counter[0], 0, ARRAY_SIZE(p->pmt_counter));
3552 }
3553 
rapl_counter_accumulate(struct rapl_counter * dst,const struct rapl_counter * src)3554 void rapl_counter_accumulate(struct rapl_counter *dst, const struct rapl_counter *src)
3555 {
3556 	/* Copy unit and scale from src if dst is not initialized */
3557 	if (dst->unit == RAPL_UNIT_INVALID) {
3558 		dst->unit = src->unit;
3559 		dst->scale = src->scale;
3560 	}
3561 
3562 	assert(dst->unit == src->unit);
3563 	assert(dst->scale == src->scale);
3564 
3565 	dst->raw_value += src->raw_value;
3566 }
3567 
sum_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)3568 int sum_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3569 {
3570 	int i;
3571 	struct msr_counter *mp;
3572 	struct perf_counter_info *pp;
3573 	struct pmt_counter *ppmt;
3574 
3575 	/* copy un-changing apic_id's */
3576 	if (DO_BIC(BIC_APIC))
3577 		average.threads.apic_id = t->apic_id;
3578 	if (DO_BIC(BIC_X2APIC))
3579 		average.threads.x2apic_id = t->x2apic_id;
3580 
3581 	/* remember first tv_begin */
3582 	if (average.threads.tv_begin.tv_sec == 0)
3583 		average.threads.tv_begin = t->tv_begin;
3584 
3585 	/* remember last tv_end */
3586 	average.threads.tv_end = t->tv_end;
3587 
3588 	average.threads.tsc += t->tsc;
3589 	average.threads.aperf += t->aperf;
3590 	average.threads.mperf += t->mperf;
3591 	average.threads.c1 += t->c1;
3592 
3593 	average.threads.instr_count += t->instr_count;
3594 
3595 	average.threads.irq_count += t->irq_count;
3596 	average.threads.smi_count += t->smi_count;
3597 
3598 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3599 		if (mp->format == FORMAT_RAW)
3600 			continue;
3601 		average.threads.counter[i] += t->counter[i];
3602 	}
3603 
3604 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
3605 		if (pp->format == FORMAT_RAW)
3606 			continue;
3607 		average.threads.perf_counter[i] += t->perf_counter[i];
3608 	}
3609 
3610 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3611 		average.threads.pmt_counter[i] += t->pmt_counter[i];
3612 	}
3613 
3614 	/* sum per-core values only for 1st thread in core */
3615 	if (!is_cpu_first_thread_in_core(t, c, p))
3616 		return 0;
3617 
3618 	average.cores.c3 += c->c3;
3619 	average.cores.c6 += c->c6;
3620 	average.cores.c7 += c->c7;
3621 	average.cores.mc6_us += c->mc6_us;
3622 
3623 	average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
3624 	average.cores.core_throt_cnt = MAX(average.cores.core_throt_cnt, c->core_throt_cnt);
3625 
3626 	rapl_counter_accumulate(&average.cores.core_energy, &c->core_energy);
3627 
3628 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3629 		if (mp->format == FORMAT_RAW)
3630 			continue;
3631 		average.cores.counter[i] += c->counter[i];
3632 	}
3633 
3634 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3635 		if (pp->format == FORMAT_RAW)
3636 			continue;
3637 		average.cores.perf_counter[i] += c->perf_counter[i];
3638 	}
3639 
3640 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3641 		average.cores.pmt_counter[i] += c->pmt_counter[i];
3642 	}
3643 
3644 	/* sum per-pkg values only for 1st core in pkg */
3645 	if (!is_cpu_first_core_in_package(t, c, p))
3646 		return 0;
3647 
3648 	if (DO_BIC(BIC_Totl_c0))
3649 		average.packages.pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
3650 	if (DO_BIC(BIC_Any_c0))
3651 		average.packages.pkg_any_core_c0 += p->pkg_any_core_c0;
3652 	if (DO_BIC(BIC_GFX_c0))
3653 		average.packages.pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
3654 	if (DO_BIC(BIC_CPUGFX))
3655 		average.packages.pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
3656 
3657 	average.packages.pc2 += p->pc2;
3658 	if (DO_BIC(BIC_Pkgpc3))
3659 		average.packages.pc3 += p->pc3;
3660 	if (DO_BIC(BIC_Pkgpc6))
3661 		average.packages.pc6 += p->pc6;
3662 	if (DO_BIC(BIC_Pkgpc7))
3663 		average.packages.pc7 += p->pc7;
3664 	average.packages.pc8 += p->pc8;
3665 	average.packages.pc9 += p->pc9;
3666 	average.packages.pc10 += p->pc10;
3667 	average.packages.die_c6 += p->die_c6;
3668 
3669 	average.packages.cpu_lpi = p->cpu_lpi;
3670 	average.packages.sys_lpi = p->sys_lpi;
3671 
3672 	rapl_counter_accumulate(&average.packages.energy_pkg, &p->energy_pkg);
3673 	rapl_counter_accumulate(&average.packages.energy_dram, &p->energy_dram);
3674 	rapl_counter_accumulate(&average.packages.energy_cores, &p->energy_cores);
3675 	rapl_counter_accumulate(&average.packages.energy_gfx, &p->energy_gfx);
3676 
3677 	average.packages.gfx_rc6_ms = p->gfx_rc6_ms;
3678 	average.packages.uncore_mhz = p->uncore_mhz;
3679 	average.packages.gfx_mhz = p->gfx_mhz;
3680 	average.packages.gfx_act_mhz = p->gfx_act_mhz;
3681 	average.packages.sam_mc6_ms = p->sam_mc6_ms;
3682 	average.packages.sam_mhz = p->sam_mhz;
3683 	average.packages.sam_act_mhz = p->sam_act_mhz;
3684 
3685 	average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
3686 
3687 	rapl_counter_accumulate(&average.packages.rapl_pkg_perf_status, &p->rapl_pkg_perf_status);
3688 	rapl_counter_accumulate(&average.packages.rapl_dram_perf_status, &p->rapl_dram_perf_status);
3689 
3690 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3691 		if ((mp->format == FORMAT_RAW) && (topo.num_packages == 0))
3692 			average.packages.counter[i] = p->counter[i];
3693 		else
3694 			average.packages.counter[i] += p->counter[i];
3695 	}
3696 
3697 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3698 		if ((pp->format == FORMAT_RAW) && (topo.num_packages == 0))
3699 			average.packages.perf_counter[i] = p->perf_counter[i];
3700 		else
3701 			average.packages.perf_counter[i] += p->perf_counter[i];
3702 	}
3703 
3704 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3705 		average.packages.pmt_counter[i] += p->pmt_counter[i];
3706 	}
3707 
3708 	return 0;
3709 }
3710 
3711 /*
3712  * sum the counters for all cpus in the system
3713  * compute the weighted average
3714  */
compute_average(struct thread_data * t,struct core_data * c,struct pkg_data * p)3715 void compute_average(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3716 {
3717 	int i;
3718 	struct msr_counter *mp;
3719 	struct perf_counter_info *pp;
3720 	struct pmt_counter *ppmt;
3721 
3722 	clear_counters(&average.threads, &average.cores, &average.packages);
3723 
3724 	for_all_cpus(sum_counters, t, c, p);
3725 
3726 	/* Use the global time delta for the average. */
3727 	average.threads.tv_delta = tv_delta;
3728 
3729 	average.threads.tsc /= topo.allowed_cpus;
3730 	average.threads.aperf /= topo.allowed_cpus;
3731 	average.threads.mperf /= topo.allowed_cpus;
3732 	average.threads.instr_count /= topo.allowed_cpus;
3733 	average.threads.c1 /= topo.allowed_cpus;
3734 
3735 	if (average.threads.irq_count > 9999999)
3736 		sums_need_wide_columns = 1;
3737 
3738 	average.cores.c3 /= topo.allowed_cores;
3739 	average.cores.c6 /= topo.allowed_cores;
3740 	average.cores.c7 /= topo.allowed_cores;
3741 	average.cores.mc6_us /= topo.allowed_cores;
3742 
3743 	if (DO_BIC(BIC_Totl_c0))
3744 		average.packages.pkg_wtd_core_c0 /= topo.allowed_packages;
3745 	if (DO_BIC(BIC_Any_c0))
3746 		average.packages.pkg_any_core_c0 /= topo.allowed_packages;
3747 	if (DO_BIC(BIC_GFX_c0))
3748 		average.packages.pkg_any_gfxe_c0 /= topo.allowed_packages;
3749 	if (DO_BIC(BIC_CPUGFX))
3750 		average.packages.pkg_both_core_gfxe_c0 /= topo.allowed_packages;
3751 
3752 	average.packages.pc2 /= topo.allowed_packages;
3753 	if (DO_BIC(BIC_Pkgpc3))
3754 		average.packages.pc3 /= topo.allowed_packages;
3755 	if (DO_BIC(BIC_Pkgpc6))
3756 		average.packages.pc6 /= topo.allowed_packages;
3757 	if (DO_BIC(BIC_Pkgpc7))
3758 		average.packages.pc7 /= topo.allowed_packages;
3759 
3760 	average.packages.pc8 /= topo.allowed_packages;
3761 	average.packages.pc9 /= topo.allowed_packages;
3762 	average.packages.pc10 /= topo.allowed_packages;
3763 	average.packages.die_c6 /= topo.allowed_packages;
3764 
3765 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3766 		if (mp->format == FORMAT_RAW)
3767 			continue;
3768 		if (mp->type == COUNTER_ITEMS) {
3769 			if (average.threads.counter[i] > 9999999)
3770 				sums_need_wide_columns = 1;
3771 			continue;
3772 		}
3773 		average.threads.counter[i] /= topo.allowed_cpus;
3774 	}
3775 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3776 		if (mp->format == FORMAT_RAW)
3777 			continue;
3778 		if (mp->type == COUNTER_ITEMS) {
3779 			if (average.cores.counter[i] > 9999999)
3780 				sums_need_wide_columns = 1;
3781 		}
3782 		average.cores.counter[i] /= topo.allowed_cores;
3783 	}
3784 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3785 		if (mp->format == FORMAT_RAW)
3786 			continue;
3787 		if (mp->type == COUNTER_ITEMS) {
3788 			if (average.packages.counter[i] > 9999999)
3789 				sums_need_wide_columns = 1;
3790 		}
3791 		average.packages.counter[i] /= topo.allowed_packages;
3792 	}
3793 
3794 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
3795 		if (pp->format == FORMAT_RAW)
3796 			continue;
3797 		if (pp->type == COUNTER_ITEMS) {
3798 			if (average.threads.perf_counter[i] > 9999999)
3799 				sums_need_wide_columns = 1;
3800 			continue;
3801 		}
3802 		average.threads.perf_counter[i] /= topo.allowed_cpus;
3803 	}
3804 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3805 		if (pp->format == FORMAT_RAW)
3806 			continue;
3807 		if (pp->type == COUNTER_ITEMS) {
3808 			if (average.cores.perf_counter[i] > 9999999)
3809 				sums_need_wide_columns = 1;
3810 		}
3811 		average.cores.perf_counter[i] /= topo.allowed_cores;
3812 	}
3813 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3814 		if (pp->format == FORMAT_RAW)
3815 			continue;
3816 		if (pp->type == COUNTER_ITEMS) {
3817 			if (average.packages.perf_counter[i] > 9999999)
3818 				sums_need_wide_columns = 1;
3819 		}
3820 		average.packages.perf_counter[i] /= topo.allowed_packages;
3821 	}
3822 
3823 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3824 		average.threads.pmt_counter[i] /= topo.allowed_cpus;
3825 	}
3826 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3827 		average.cores.pmt_counter[i] /= topo.allowed_cores;
3828 	}
3829 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3830 		average.packages.pmt_counter[i] /= topo.allowed_packages;
3831 	}
3832 }
3833 
rdtsc(void)3834 static unsigned long long rdtsc(void)
3835 {
3836 	unsigned int low, high;
3837 
3838 	asm volatile ("rdtsc":"=a" (low), "=d"(high));
3839 
3840 	return low | ((unsigned long long)high) << 32;
3841 }
3842 
3843 /*
3844  * Open a file, and exit on failure
3845  */
fopen_or_die(const char * path,const char * mode)3846 FILE *fopen_or_die(const char *path, const char *mode)
3847 {
3848 	FILE *filep = fopen(path, mode);
3849 
3850 	if (!filep)
3851 		err(1, "%s: open failed", path);
3852 	return filep;
3853 }
3854 
3855 /*
3856  * snapshot_sysfs_counter()
3857  *
3858  * return snapshot of given counter
3859  */
snapshot_sysfs_counter(char * path)3860 unsigned long long snapshot_sysfs_counter(char *path)
3861 {
3862 	FILE *fp;
3863 	int retval;
3864 	unsigned long long counter;
3865 
3866 	fp = fopen_or_die(path, "r");
3867 
3868 	retval = fscanf(fp, "%lld", &counter);
3869 	if (retval != 1)
3870 		err(1, "snapshot_sysfs_counter(%s)", path);
3871 
3872 	fclose(fp);
3873 
3874 	return counter;
3875 }
3876 
get_mp(int cpu,struct msr_counter * mp,unsigned long long * counterp,char * counter_path)3877 int get_mp(int cpu, struct msr_counter *mp, unsigned long long *counterp, char *counter_path)
3878 {
3879 	if (mp->msr_num != 0) {
3880 		assert(!no_msr);
3881 		if (get_msr(cpu, mp->msr_num, counterp))
3882 			return -1;
3883 	} else {
3884 		char path[128 + PATH_BYTES];
3885 
3886 		if (mp->flags & SYSFS_PERCPU) {
3887 			sprintf(path, "/sys/devices/system/cpu/cpu%d/%s", cpu, mp->sp->path);
3888 
3889 			*counterp = snapshot_sysfs_counter(path);
3890 		} else {
3891 			*counterp = snapshot_sysfs_counter(counter_path);
3892 		}
3893 	}
3894 
3895 	return 0;
3896 }
3897 
get_legacy_uncore_mhz(int package)3898 unsigned long long get_legacy_uncore_mhz(int package)
3899 {
3900 	char path[128];
3901 	int die;
3902 	static int warn_once;
3903 
3904 	/*
3905 	 * for this package, use the first die_id that exists
3906 	 */
3907 	for (die = 0; die <= topo.max_die_id; ++die) {
3908 
3909 		sprintf(path, "/sys/devices/system/cpu/intel_uncore_frequency/package_%02d_die_%02d/current_freq_khz",
3910 			package, die);
3911 
3912 		if (access(path, R_OK) == 0)
3913 			return (snapshot_sysfs_counter(path) / 1000);
3914 	}
3915 	if (!warn_once) {
3916 		warnx("BUG: %s: No %s", __func__, path);
3917 		warn_once = 1;
3918 	}
3919 
3920 	return 0;
3921 }
3922 
get_epb(int cpu)3923 int get_epb(int cpu)
3924 {
3925 	char path[128 + PATH_BYTES];
3926 	unsigned long long msr;
3927 	int ret, epb = -1;
3928 	FILE *fp;
3929 
3930 	sprintf(path, "/sys/devices/system/cpu/cpu%d/power/energy_perf_bias", cpu);
3931 
3932 	fp = fopen(path, "r");
3933 	if (!fp)
3934 		goto msr_fallback;
3935 
3936 	ret = fscanf(fp, "%d", &epb);
3937 	if (ret != 1)
3938 		err(1, "%s(%s)", __func__, path);
3939 
3940 	fclose(fp);
3941 
3942 	return epb;
3943 
3944 msr_fallback:
3945 	if (no_msr)
3946 		return -1;
3947 
3948 	get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr);
3949 
3950 	return msr & 0xf;
3951 }
3952 
get_apic_id(struct thread_data * t)3953 void get_apic_id(struct thread_data *t)
3954 {
3955 	unsigned int eax, ebx, ecx, edx;
3956 
3957 	if (DO_BIC(BIC_APIC)) {
3958 		eax = ebx = ecx = edx = 0;
3959 		__cpuid(1, eax, ebx, ecx, edx);
3960 
3961 		t->apic_id = (ebx >> 24) & 0xff;
3962 	}
3963 
3964 	if (!DO_BIC(BIC_X2APIC))
3965 		return;
3966 
3967 	if (authentic_amd || hygon_genuine) {
3968 		unsigned int topology_extensions;
3969 
3970 		if (max_extended_level < 0x8000001e)
3971 			return;
3972 
3973 		eax = ebx = ecx = edx = 0;
3974 		__cpuid(0x80000001, eax, ebx, ecx, edx);
3975 		topology_extensions = ecx & (1 << 22);
3976 
3977 		if (topology_extensions == 0)
3978 			return;
3979 
3980 		eax = ebx = ecx = edx = 0;
3981 		__cpuid(0x8000001e, eax, ebx, ecx, edx);
3982 
3983 		t->x2apic_id = eax;
3984 		return;
3985 	}
3986 
3987 	if (!genuine_intel)
3988 		return;
3989 
3990 	if (max_level < 0xb)
3991 		return;
3992 
3993 	ecx = 0;
3994 	__cpuid(0xb, eax, ebx, ecx, edx);
3995 	t->x2apic_id = edx;
3996 
3997 	if (debug && (t->apic_id != (t->x2apic_id & 0xff)))
3998 		fprintf(outf, "cpu%d: BIOS BUG: apic 0x%x x2apic 0x%x\n", t->cpu_id, t->apic_id, t->x2apic_id);
3999 }
4000 
get_core_throt_cnt(int cpu,unsigned long long * cnt)4001 int get_core_throt_cnt(int cpu, unsigned long long *cnt)
4002 {
4003 	char path[128 + PATH_BYTES];
4004 	unsigned long long tmp;
4005 	FILE *fp;
4006 	int ret;
4007 
4008 	sprintf(path, "/sys/devices/system/cpu/cpu%d/thermal_throttle/core_throttle_count", cpu);
4009 	fp = fopen(path, "r");
4010 	if (!fp)
4011 		return -1;
4012 	ret = fscanf(fp, "%lld", &tmp);
4013 	fclose(fp);
4014 	if (ret != 1)
4015 		return -1;
4016 	*cnt = tmp;
4017 
4018 	return 0;
4019 }
4020 
4021 struct amperf_group_fd {
4022 	int aperf;		/* Also the group descriptor */
4023 	int mperf;
4024 };
4025 
read_perf_counter_info(const char * const path,const char * const parse_format,void * value_ptr)4026 static int read_perf_counter_info(const char *const path, const char *const parse_format, void *value_ptr)
4027 {
4028 	int fdmt;
4029 	int bytes_read;
4030 	char buf[64];
4031 	int ret = -1;
4032 
4033 	fdmt = open(path, O_RDONLY, 0);
4034 	if (fdmt == -1) {
4035 		if (debug)
4036 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4037 		ret = -1;
4038 		goto cleanup_and_exit;
4039 	}
4040 
4041 	bytes_read = read(fdmt, buf, sizeof(buf) - 1);
4042 	if (bytes_read <= 0 || bytes_read >= (int)sizeof(buf)) {
4043 		if (debug)
4044 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4045 		ret = -1;
4046 		goto cleanup_and_exit;
4047 	}
4048 
4049 	buf[bytes_read] = '\0';
4050 
4051 	if (sscanf(buf, parse_format, value_ptr) != 1) {
4052 		if (debug)
4053 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4054 		ret = -1;
4055 		goto cleanup_and_exit;
4056 	}
4057 
4058 	ret = 0;
4059 
4060 cleanup_and_exit:
4061 	close(fdmt);
4062 	return ret;
4063 }
4064 
read_perf_counter_info_n(const char * const path,const char * const parse_format)4065 static unsigned int read_perf_counter_info_n(const char *const path, const char *const parse_format)
4066 {
4067 	unsigned int v;
4068 	int status;
4069 
4070 	status = read_perf_counter_info(path, parse_format, &v);
4071 	if (status)
4072 		v = -1;
4073 
4074 	return v;
4075 }
4076 
read_perf_type(const char * subsys)4077 static unsigned int read_perf_type(const char *subsys)
4078 {
4079 	const char *const path_format = "/sys/bus/event_source/devices/%s/type";
4080 	const char *const format = "%u";
4081 	char path[128];
4082 
4083 	snprintf(path, sizeof(path), path_format, subsys);
4084 
4085 	return read_perf_counter_info_n(path, format);
4086 }
4087 
read_perf_config(const char * subsys,const char * event_name)4088 static unsigned int read_perf_config(const char *subsys, const char *event_name)
4089 {
4090 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s";
4091 	FILE *fconfig = NULL;
4092 	char path[128];
4093 	char config_str[64];
4094 	unsigned int config;
4095 	unsigned int umask;
4096 	bool has_config = false;
4097 	bool has_umask = false;
4098 	unsigned int ret = -1;
4099 
4100 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4101 
4102 	fconfig = fopen(path, "r");
4103 	if (!fconfig)
4104 		return -1;
4105 
4106 	if (fgets(config_str, ARRAY_SIZE(config_str), fconfig) != config_str)
4107 		goto cleanup_and_exit;
4108 
4109 	for (char *pconfig_str = &config_str[0]; pconfig_str;) {
4110 		if (sscanf(pconfig_str, "event=%x", &config) == 1) {
4111 			has_config = true;
4112 			goto next;
4113 		}
4114 
4115 		if (sscanf(pconfig_str, "umask=%x", &umask) == 1) {
4116 			has_umask = true;
4117 			goto next;
4118 		}
4119 
4120 next:
4121 		pconfig_str = strchr(pconfig_str, ',');
4122 		if (pconfig_str) {
4123 			*pconfig_str = '\0';
4124 			++pconfig_str;
4125 		}
4126 	}
4127 
4128 	if (!has_umask)
4129 		umask = 0;
4130 
4131 	if (has_config)
4132 		ret = (umask << 8) | config;
4133 
4134 cleanup_and_exit:
4135 	fclose(fconfig);
4136 	return ret;
4137 }
4138 
read_perf_rapl_unit(const char * subsys,const char * event_name)4139 static unsigned int read_perf_rapl_unit(const char *subsys, const char *event_name)
4140 {
4141 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.unit";
4142 	const char *const format = "%s";
4143 	char path[128];
4144 	char unit_buffer[16];
4145 
4146 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4147 
4148 	read_perf_counter_info(path, format, &unit_buffer);
4149 	if (strcmp("Joules", unit_buffer) == 0)
4150 		return RAPL_UNIT_JOULES;
4151 
4152 	return RAPL_UNIT_INVALID;
4153 }
4154 
read_perf_scale(const char * subsys,const char * event_name)4155 static double read_perf_scale(const char *subsys, const char *event_name)
4156 {
4157 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.scale";
4158 	const char *const format = "%lf";
4159 	char path[128];
4160 	double scale;
4161 
4162 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4163 
4164 	if (read_perf_counter_info(path, format, &scale))
4165 		return 0.0;
4166 
4167 	return scale;
4168 }
4169 
rapl_counter_info_count_perf(const struct rapl_counter_info_t * rci)4170 size_t rapl_counter_info_count_perf(const struct rapl_counter_info_t *rci)
4171 {
4172 	size_t ret = 0;
4173 
4174 	for (int i = 0; i < NUM_RAPL_COUNTERS; ++i)
4175 		if (rci->source[i] == COUNTER_SOURCE_PERF)
4176 			++ret;
4177 
4178 	return ret;
4179 }
4180 
cstate_counter_info_count_perf(const struct cstate_counter_info_t * cci)4181 static size_t cstate_counter_info_count_perf(const struct cstate_counter_info_t *cci)
4182 {
4183 	size_t ret = 0;
4184 
4185 	for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i)
4186 		if (cci->source[i] == COUNTER_SOURCE_PERF)
4187 			++ret;
4188 
4189 	return ret;
4190 }
4191 
write_rapl_counter(struct rapl_counter * rc,struct rapl_counter_info_t * rci,unsigned int idx)4192 void write_rapl_counter(struct rapl_counter *rc, struct rapl_counter_info_t *rci, unsigned int idx)
4193 {
4194 	if (rci->source[idx] == COUNTER_SOURCE_NONE)
4195 		return;
4196 
4197 	rc->raw_value = rci->data[idx];
4198 	rc->unit = rci->unit[idx];
4199 	rc->scale = rci->scale[idx];
4200 }
4201 
get_rapl_counters(int cpu,unsigned int domain,struct core_data * c,struct pkg_data * p)4202 int get_rapl_counters(int cpu, unsigned int domain, struct core_data *c, struct pkg_data *p)
4203 {
4204 	struct platform_counters *pplat_cnt = p == package_odd ? &platform_counters_odd : &platform_counters_even;
4205 	unsigned long long perf_data[NUM_RAPL_COUNTERS + 1];
4206 	struct rapl_counter_info_t *rci;
4207 
4208 	if (debug >= 2)
4209 		fprintf(stderr, "%s: cpu%d domain%d\n", __func__, cpu, domain);
4210 
4211 	assert(rapl_counter_info_perdomain);
4212 	assert(domain < rapl_counter_info_perdomain_size);
4213 
4214 	rci = &rapl_counter_info_perdomain[domain];
4215 
4216 	/*
4217 	 * If we have any perf counters to read, read them all now, in bulk
4218 	 */
4219 	if (rci->fd_perf != -1) {
4220 		size_t num_perf_counters = rapl_counter_info_count_perf(rci);
4221 		const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
4222 		const ssize_t actual_read_size = read(rci->fd_perf, &perf_data[0], sizeof(perf_data));
4223 
4224 		if (actual_read_size != expected_read_size)
4225 			err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size,
4226 			    actual_read_size);
4227 	}
4228 
4229 	for (unsigned int i = 0, pi = 1; i < NUM_RAPL_COUNTERS; ++i) {
4230 		switch (rci->source[i]) {
4231 		case COUNTER_SOURCE_NONE:
4232 			rci->data[i] = 0;
4233 			break;
4234 
4235 		case COUNTER_SOURCE_PERF:
4236 			assert(pi < ARRAY_SIZE(perf_data));
4237 			assert(rci->fd_perf != -1);
4238 
4239 			if (debug >= 2)
4240 				fprintf(stderr, "Reading rapl counter via perf at %u (%llu %e %lf)\n",
4241 					i, perf_data[pi], rci->scale[i], perf_data[pi] * rci->scale[i]);
4242 
4243 			rci->data[i] = perf_data[pi];
4244 
4245 			++pi;
4246 			break;
4247 
4248 		case COUNTER_SOURCE_MSR:
4249 			if (debug >= 2)
4250 				fprintf(stderr, "Reading rapl counter via msr at %u\n", i);
4251 
4252 			assert(!no_msr);
4253 			if (rci->flags[i] & RAPL_COUNTER_FLAG_USE_MSR_SUM) {
4254 				if (get_msr_sum(cpu, rci->msr[i], &rci->data[i]))
4255 					return -13 - i;
4256 			} else {
4257 				if (get_msr(cpu, rci->msr[i], &rci->data[i]))
4258 					return -13 - i;
4259 			}
4260 
4261 			rci->data[i] &= rci->msr_mask[i];
4262 			if (rci->msr_shift[i] >= 0)
4263 				rci->data[i] >>= abs(rci->msr_shift[i]);
4264 			else
4265 				rci->data[i] <<= abs(rci->msr_shift[i]);
4266 
4267 			break;
4268 		}
4269 	}
4270 
4271 	BUILD_BUG_ON(NUM_RAPL_COUNTERS != 8);
4272 	write_rapl_counter(&p->energy_pkg, rci, RAPL_RCI_INDEX_ENERGY_PKG);
4273 	write_rapl_counter(&p->energy_cores, rci, RAPL_RCI_INDEX_ENERGY_CORES);
4274 	write_rapl_counter(&p->energy_dram, rci, RAPL_RCI_INDEX_DRAM);
4275 	write_rapl_counter(&p->energy_gfx, rci, RAPL_RCI_INDEX_GFX);
4276 	write_rapl_counter(&p->rapl_pkg_perf_status, rci, RAPL_RCI_INDEX_PKG_PERF_STATUS);
4277 	write_rapl_counter(&p->rapl_dram_perf_status, rci, RAPL_RCI_INDEX_DRAM_PERF_STATUS);
4278 	write_rapl_counter(&c->core_energy, rci, RAPL_RCI_INDEX_CORE_ENERGY);
4279 	write_rapl_counter(&pplat_cnt->energy_psys, rci, RAPL_RCI_INDEX_ENERGY_PLATFORM);
4280 
4281 	return 0;
4282 }
4283 
find_sysfs_path_by_id(struct sysfs_path * sp,int id)4284 char *find_sysfs_path_by_id(struct sysfs_path *sp, int id)
4285 {
4286 	while (sp) {
4287 		if (sp->id == id)
4288 			return (sp->path);
4289 		sp = sp->next;
4290 	}
4291 	if (debug)
4292 		warnx("%s: id%d not found", __func__, id);
4293 	return NULL;
4294 }
4295 
get_cstate_counters(unsigned int cpu,struct thread_data * t,struct core_data * c,struct pkg_data * p)4296 int get_cstate_counters(unsigned int cpu, struct thread_data *t, struct core_data *c, struct pkg_data *p)
4297 {
4298 	/*
4299 	 * Overcommit memory a little bit here,
4300 	 * but skip calculating exact sizes for the buffers.
4301 	 */
4302 	unsigned long long perf_data[NUM_CSTATE_COUNTERS];
4303 	unsigned long long perf_data_core[NUM_CSTATE_COUNTERS + 1];
4304 	unsigned long long perf_data_pkg[NUM_CSTATE_COUNTERS + 1];
4305 
4306 	struct cstate_counter_info_t *cci;
4307 
4308 	if (debug >= 2)
4309 		fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
4310 
4311 	assert(ccstate_counter_info);
4312 	assert(cpu <= ccstate_counter_info_size);
4313 
4314 	ZERO_ARRAY(perf_data);
4315 	ZERO_ARRAY(perf_data_core);
4316 	ZERO_ARRAY(perf_data_pkg);
4317 
4318 	cci = &ccstate_counter_info[cpu];
4319 
4320 	/*
4321 	 * If we have any perf counters to read, read them all now, in bulk
4322 	 */
4323 	const size_t num_perf_counters = cstate_counter_info_count_perf(cci);
4324 	ssize_t expected_read_size = num_perf_counters * sizeof(unsigned long long);
4325 	ssize_t actual_read_size_core = 0, actual_read_size_pkg = 0;
4326 
4327 	if (cci->fd_perf_core != -1) {
4328 		/* Each descriptor read begins with number of counters read. */
4329 		expected_read_size += sizeof(unsigned long long);
4330 
4331 		actual_read_size_core = read(cci->fd_perf_core, &perf_data_core[0], sizeof(perf_data_core));
4332 
4333 		if (actual_read_size_core <= 0)
4334 			err(-1, "%s: read perf %s: %ld", __func__, "core", actual_read_size_core);
4335 	}
4336 
4337 	if (cci->fd_perf_pkg != -1) {
4338 		/* Each descriptor read begins with number of counters read. */
4339 		expected_read_size += sizeof(unsigned long long);
4340 
4341 		actual_read_size_pkg = read(cci->fd_perf_pkg, &perf_data_pkg[0], sizeof(perf_data_pkg));
4342 
4343 		if (actual_read_size_pkg <= 0)
4344 			err(-1, "%s: read perf %s: %ld", __func__, "pkg", actual_read_size_pkg);
4345 	}
4346 
4347 	const ssize_t actual_read_size_total = actual_read_size_core + actual_read_size_pkg;
4348 
4349 	if (actual_read_size_total != expected_read_size)
4350 		err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size, actual_read_size_total);
4351 
4352 	/*
4353 	 * Copy ccstate and pcstate data into unified buffer.
4354 	 *
4355 	 * Skip first element from core and pkg buffers.
4356 	 * Kernel puts there how many counters were read.
4357 	 */
4358 	const size_t num_core_counters = perf_data_core[0];
4359 	const size_t num_pkg_counters = perf_data_pkg[0];
4360 
4361 	assert(num_perf_counters == num_core_counters + num_pkg_counters);
4362 
4363 	/* Copy ccstate perf data */
4364 	memcpy(&perf_data[0], &perf_data_core[1], num_core_counters * sizeof(unsigned long long));
4365 
4366 	/* Copy pcstate perf data */
4367 	memcpy(&perf_data[num_core_counters], &perf_data_pkg[1], num_pkg_counters * sizeof(unsigned long long));
4368 
4369 	for (unsigned int i = 0, pi = 0; i < NUM_CSTATE_COUNTERS; ++i) {
4370 		switch (cci->source[i]) {
4371 		case COUNTER_SOURCE_NONE:
4372 			break;
4373 
4374 		case COUNTER_SOURCE_PERF:
4375 			assert(pi < ARRAY_SIZE(perf_data));
4376 			assert(cci->fd_perf_core != -1 || cci->fd_perf_pkg != -1);
4377 
4378 			if (debug >= 2)
4379 				fprintf(stderr, "cstate via %s %u: %llu\n", "perf", i, perf_data[pi]);
4380 
4381 			cci->data[i] = perf_data[pi];
4382 
4383 			++pi;
4384 			break;
4385 
4386 		case COUNTER_SOURCE_MSR:
4387 			assert(!no_msr);
4388 			if (get_msr(cpu, cci->msr[i], &cci->data[i]))
4389 				return -13 - i;
4390 
4391 			if (debug >= 2)
4392 				fprintf(stderr, "cstate via %s0x%llx %u: %llu\n", "msr", cci->msr[i], i, cci->data[i]);
4393 
4394 			break;
4395 		}
4396 	}
4397 
4398 	/*
4399 	 * Helper to write the data only if the source of
4400 	 * the counter for the current cpu is not none.
4401 	 *
4402 	 * Otherwise we would overwrite core data with 0 (default value),
4403 	 * when invoked for the thread sibling.
4404 	 */
4405 #define PERF_COUNTER_WRITE_DATA(out_counter, index) do {	\
4406 	if (cci->source[index] != COUNTER_SOURCE_NONE)		\
4407 		out_counter = cci->data[index];			\
4408 } while (0)
4409 
4410 	BUILD_BUG_ON(NUM_CSTATE_COUNTERS != 11);
4411 
4412 	PERF_COUNTER_WRITE_DATA(t->c1, CCSTATE_RCI_INDEX_C1_RESIDENCY);
4413 	PERF_COUNTER_WRITE_DATA(c->c3, CCSTATE_RCI_INDEX_C3_RESIDENCY);
4414 	PERF_COUNTER_WRITE_DATA(c->c6, CCSTATE_RCI_INDEX_C6_RESIDENCY);
4415 	PERF_COUNTER_WRITE_DATA(c->c7, CCSTATE_RCI_INDEX_C7_RESIDENCY);
4416 
4417 	PERF_COUNTER_WRITE_DATA(p->pc2, PCSTATE_RCI_INDEX_C2_RESIDENCY);
4418 	PERF_COUNTER_WRITE_DATA(p->pc3, PCSTATE_RCI_INDEX_C3_RESIDENCY);
4419 	PERF_COUNTER_WRITE_DATA(p->pc6, PCSTATE_RCI_INDEX_C6_RESIDENCY);
4420 	PERF_COUNTER_WRITE_DATA(p->pc7, PCSTATE_RCI_INDEX_C7_RESIDENCY);
4421 	PERF_COUNTER_WRITE_DATA(p->pc8, PCSTATE_RCI_INDEX_C8_RESIDENCY);
4422 	PERF_COUNTER_WRITE_DATA(p->pc9, PCSTATE_RCI_INDEX_C9_RESIDENCY);
4423 	PERF_COUNTER_WRITE_DATA(p->pc10, PCSTATE_RCI_INDEX_C10_RESIDENCY);
4424 
4425 #undef PERF_COUNTER_WRITE_DATA
4426 
4427 	return 0;
4428 }
4429 
msr_counter_info_count_perf(const struct msr_counter_info_t * mci)4430 size_t msr_counter_info_count_perf(const struct msr_counter_info_t *mci)
4431 {
4432 	size_t ret = 0;
4433 
4434 	for (int i = 0; i < NUM_MSR_COUNTERS; ++i)
4435 		if (mci->source[i] == COUNTER_SOURCE_PERF)
4436 			++ret;
4437 
4438 	return ret;
4439 }
4440 
get_smi_aperf_mperf(unsigned int cpu,struct thread_data * t)4441 int get_smi_aperf_mperf(unsigned int cpu, struct thread_data *t)
4442 {
4443 	unsigned long long perf_data[NUM_MSR_COUNTERS + 1];
4444 
4445 	struct msr_counter_info_t *mci;
4446 
4447 	if (debug >= 2)
4448 		fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
4449 
4450 	assert(msr_counter_info);
4451 	assert(cpu <= msr_counter_info_size);
4452 
4453 	mci = &msr_counter_info[cpu];
4454 
4455 	ZERO_ARRAY(perf_data);
4456 	ZERO_ARRAY(mci->data);
4457 
4458 	if (mci->fd_perf != -1) {
4459 		const size_t num_perf_counters = msr_counter_info_count_perf(mci);
4460 		const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
4461 		const ssize_t actual_read_size = read(mci->fd_perf, &perf_data[0], sizeof(perf_data));
4462 
4463 		if (actual_read_size != expected_read_size)
4464 			err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size,
4465 			    actual_read_size);
4466 	}
4467 
4468 	for (unsigned int i = 0, pi = 1; i < NUM_MSR_COUNTERS; ++i) {
4469 		switch (mci->source[i]) {
4470 		case COUNTER_SOURCE_NONE:
4471 			break;
4472 
4473 		case COUNTER_SOURCE_PERF:
4474 			assert(pi < ARRAY_SIZE(perf_data));
4475 			assert(mci->fd_perf != -1);
4476 
4477 			if (debug >= 2)
4478 				fprintf(stderr, "Reading msr counter via perf at %u: %llu\n", i, perf_data[pi]);
4479 
4480 			mci->data[i] = perf_data[pi];
4481 
4482 			++pi;
4483 			break;
4484 
4485 		case COUNTER_SOURCE_MSR:
4486 			assert(!no_msr);
4487 
4488 			if (get_msr(cpu, mci->msr[i], &mci->data[i]))
4489 				return -2 - i;
4490 
4491 			mci->data[i] &= mci->msr_mask[i];
4492 
4493 			if (debug >= 2)
4494 				fprintf(stderr, "Reading msr counter via msr at %u: %llu\n", i, mci->data[i]);
4495 
4496 			break;
4497 		}
4498 	}
4499 
4500 	BUILD_BUG_ON(NUM_MSR_COUNTERS != 3);
4501 	t->aperf = mci->data[MSR_RCI_INDEX_APERF];
4502 	t->mperf = mci->data[MSR_RCI_INDEX_MPERF];
4503 	t->smi_count = mci->data[MSR_RCI_INDEX_SMI];
4504 
4505 	return 0;
4506 }
4507 
perf_counter_info_read_values(struct perf_counter_info * pp,int cpu,unsigned long long * out,size_t out_size)4508 int perf_counter_info_read_values(struct perf_counter_info *pp, int cpu, unsigned long long *out, size_t out_size)
4509 {
4510 	unsigned int domain;
4511 	unsigned long long value;
4512 	int fd_counter;
4513 
4514 	for (size_t i = 0; pp; ++i, pp = pp->next) {
4515 		domain = cpu_to_domain(pp, cpu);
4516 		assert(domain < pp->num_domains);
4517 
4518 		fd_counter = pp->fd_perf_per_domain[domain];
4519 
4520 		if (fd_counter == -1)
4521 			continue;
4522 
4523 		if (read(fd_counter, &value, sizeof(value)) != sizeof(value))
4524 			return 1;
4525 
4526 		assert(i < out_size);
4527 		out[i] = value * pp->scale;
4528 	}
4529 
4530 	return 0;
4531 }
4532 
pmt_gen_value_mask(unsigned int lsb,unsigned int msb)4533 unsigned long pmt_gen_value_mask(unsigned int lsb, unsigned int msb)
4534 {
4535 	unsigned long mask;
4536 
4537 	if (msb == 63)
4538 		mask = 0xffffffffffffffff;
4539 	else
4540 		mask = ((1 << (msb + 1)) - 1);
4541 
4542 	mask -= (1 << lsb) - 1;
4543 
4544 	return mask;
4545 }
4546 
pmt_read_counter(struct pmt_counter * ppmt,unsigned int domain_id)4547 unsigned long pmt_read_counter(struct pmt_counter *ppmt, unsigned int domain_id)
4548 {
4549 	assert(domain_id < ppmt->num_domains);
4550 
4551 	const unsigned long *pmmio = ppmt->domains[domain_id].pcounter;
4552 	const unsigned long value = pmmio ? *pmmio : 0;
4553 	const unsigned long value_mask = pmt_gen_value_mask(ppmt->lsb, ppmt->msb);
4554 	const unsigned long value_shift = ppmt->lsb;
4555 
4556 	return (value & value_mask) >> value_shift;
4557 }
4558 
4559 /*
4560  * get_counters(...)
4561  * migrate to cpu
4562  * acquire and record local counters for that cpu
4563  */
get_counters(struct thread_data * t,struct core_data * c,struct pkg_data * p)4564 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4565 {
4566 	int cpu = t->cpu_id;
4567 	unsigned long long msr;
4568 	struct msr_counter *mp;
4569 	struct pmt_counter *pp;
4570 	int i;
4571 	int status;
4572 
4573 	if (cpu_migrate(cpu)) {
4574 		fprintf(outf, "%s: Could not migrate to CPU %d\n", __func__, cpu);
4575 		return -1;
4576 	}
4577 
4578 	gettimeofday(&t->tv_begin, (struct timezone *)NULL);
4579 
4580 	if (first_counter_read)
4581 		get_apic_id(t);
4582 
4583 	t->tsc = rdtsc();	/* we are running on local CPU of interest */
4584 
4585 	get_smi_aperf_mperf(cpu, t);
4586 
4587 	if (DO_BIC(BIC_IPC))
4588 		if (read(get_instr_count_fd(cpu), &t->instr_count, sizeof(long long)) != sizeof(long long))
4589 			return -4;
4590 
4591 	if (DO_BIC(BIC_IRQ))
4592 		t->irq_count = irqs_per_cpu[cpu];
4593 
4594 	get_cstate_counters(cpu, t, c, p);
4595 
4596 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
4597 		if (get_mp(cpu, mp, &t->counter[i], mp->sp->path))
4598 			return -10;
4599 	}
4600 
4601 	if (perf_counter_info_read_values(sys.perf_tp, cpu, t->perf_counter, MAX_ADDED_THREAD_COUNTERS))
4602 		return -10;
4603 
4604 	for (i = 0, pp = sys.pmt_tp; pp; i++, pp = pp->next)
4605 		t->pmt_counter[i] = pmt_read_counter(pp, t->cpu_id);
4606 
4607 	/* collect core counters only for 1st thread in core */
4608 	if (!is_cpu_first_thread_in_core(t, c, p))
4609 		goto done;
4610 
4611 	if (platform->has_per_core_rapl) {
4612 		status = get_rapl_counters(cpu, c->core_id, c, p);
4613 		if (status != 0)
4614 			return status;
4615 	}
4616 
4617 	if (DO_BIC(BIC_CPU_c7) && t->is_atom) {
4618 		/*
4619 		 * For Atom CPUs that has core cstate deeper than c6,
4620 		 * MSR_CORE_C6_RESIDENCY returns residency of cc6 and deeper.
4621 		 * Minus CC7 (and deeper cstates) residency to get
4622 		 * accturate cc6 residency.
4623 		 */
4624 		c->c6 -= c->c7;
4625 	}
4626 
4627 	if (DO_BIC(BIC_Mod_c6))
4628 		if (get_msr(cpu, MSR_MODULE_C6_RES_MS, &c->mc6_us))
4629 			return -8;
4630 
4631 	if (DO_BIC(BIC_CoreTmp)) {
4632 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
4633 			return -9;
4634 		c->core_temp_c = tj_max - ((msr >> 16) & 0x7F);
4635 	}
4636 
4637 	if (DO_BIC(BIC_CORE_THROT_CNT))
4638 		get_core_throt_cnt(cpu, &c->core_throt_cnt);
4639 
4640 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
4641 		if (get_mp(cpu, mp, &c->counter[i], mp->sp->path))
4642 			return -10;
4643 	}
4644 
4645 	if (perf_counter_info_read_values(sys.perf_cp, cpu, c->perf_counter, MAX_ADDED_CORE_COUNTERS))
4646 		return -10;
4647 
4648 	for (i = 0, pp = sys.pmt_cp; pp; i++, pp = pp->next)
4649 		c->pmt_counter[i] = pmt_read_counter(pp, c->core_id);
4650 
4651 	/* collect package counters only for 1st core in package */
4652 	if (!is_cpu_first_core_in_package(t, c, p))
4653 		goto done;
4654 
4655 	if (DO_BIC(BIC_Totl_c0)) {
4656 		if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
4657 			return -10;
4658 	}
4659 	if (DO_BIC(BIC_Any_c0)) {
4660 		if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
4661 			return -11;
4662 	}
4663 	if (DO_BIC(BIC_GFX_c0)) {
4664 		if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
4665 			return -12;
4666 	}
4667 	if (DO_BIC(BIC_CPUGFX)) {
4668 		if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
4669 			return -13;
4670 	}
4671 
4672 	if (DO_BIC(BIC_CPU_LPI))
4673 		p->cpu_lpi = cpuidle_cur_cpu_lpi_us;
4674 	if (DO_BIC(BIC_SYS_LPI))
4675 		p->sys_lpi = cpuidle_cur_sys_lpi_us;
4676 
4677 	if (!platform->has_per_core_rapl) {
4678 		status = get_rapl_counters(cpu, p->package_id, c, p);
4679 		if (status != 0)
4680 			return status;
4681 	}
4682 
4683 	if (DO_BIC(BIC_PkgTmp)) {
4684 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
4685 			return -17;
4686 		p->pkg_temp_c = tj_max - ((msr >> 16) & 0x7F);
4687 	}
4688 
4689 	if (DO_BIC(BIC_UNCORE_MHZ))
4690 		p->uncore_mhz = get_legacy_uncore_mhz(p->package_id);
4691 
4692 	if (DO_BIC(BIC_GFX_rc6))
4693 		p->gfx_rc6_ms = gfx_info[GFX_rc6].val_ull;
4694 
4695 	if (DO_BIC(BIC_GFXMHz))
4696 		p->gfx_mhz = gfx_info[GFX_MHz].val;
4697 
4698 	if (DO_BIC(BIC_GFXACTMHz))
4699 		p->gfx_act_mhz = gfx_info[GFX_ACTMHz].val;
4700 
4701 	if (DO_BIC(BIC_SAM_mc6))
4702 		p->sam_mc6_ms = gfx_info[SAM_mc6].val_ull;
4703 
4704 	if (DO_BIC(BIC_SAMMHz))
4705 		p->sam_mhz = gfx_info[SAM_MHz].val;
4706 
4707 	if (DO_BIC(BIC_SAMACTMHz))
4708 		p->sam_act_mhz = gfx_info[SAM_ACTMHz].val;
4709 
4710 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
4711 		char *path = NULL;
4712 
4713 		if (mp->msr_num == 0) {
4714 			path = find_sysfs_path_by_id(mp->sp, p->package_id);
4715 			if (path == NULL) {
4716 				warnx("%s: package_id %d not found", __func__, p->package_id);
4717 				return -10;
4718 			}
4719 		}
4720 		if (get_mp(cpu, mp, &p->counter[i], path))
4721 			return -10;
4722 	}
4723 
4724 	if (perf_counter_info_read_values(sys.perf_pp, cpu, p->perf_counter, MAX_ADDED_PACKAGE_COUNTERS))
4725 		return -10;
4726 
4727 	for (i = 0, pp = sys.pmt_pp; pp; i++, pp = pp->next)
4728 		p->pmt_counter[i] = pmt_read_counter(pp, p->package_id);
4729 
4730 done:
4731 	gettimeofday(&t->tv_end, (struct timezone *)NULL);
4732 
4733 	return 0;
4734 }
4735 
4736 int pkg_cstate_limit = PCLUKN;
4737 char *pkg_cstate_limit_strings[] = { "unknown", "reserved", "pc0", "pc1", "pc2",
4738 	"pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "pc10", "unlimited"
4739 };
4740 
4741 int nhm_pkg_cstate_limits[16] =
4742     { PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4743 	PCLRSV, PCLRSV
4744 };
4745 
4746 int snb_pkg_cstate_limits[16] =
4747     { PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4748 	PCLRSV, PCLRSV
4749 };
4750 
4751 int hsw_pkg_cstate_limits[16] =
4752     { PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4753 	PCLRSV, PCLRSV
4754 };
4755 
4756 int slv_pkg_cstate_limits[16] =
4757     { PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4758 	PCL__6, PCL__7
4759 };
4760 
4761 int amt_pkg_cstate_limits[16] =
4762     { PCLUNL, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4763 	PCLRSV, PCLRSV
4764 };
4765 
4766 int phi_pkg_cstate_limits[16] =
4767     { PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4768 	PCLRSV, PCLRSV
4769 };
4770 
4771 int glm_pkg_cstate_limits[16] =
4772     { PCLUNL, PCL__1, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCL_10, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4773 	PCLRSV, PCLRSV
4774 };
4775 
4776 int skx_pkg_cstate_limits[16] =
4777     { PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4778 	PCLRSV, PCLRSV
4779 };
4780 
4781 int icx_pkg_cstate_limits[16] =
4782     { PCL__0, PCL__2, PCL__6, PCL__6, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
4783 	PCLRSV, PCLRSV
4784 };
4785 
probe_cst_limit(void)4786 void probe_cst_limit(void)
4787 {
4788 	unsigned long long msr;
4789 	int *pkg_cstate_limits;
4790 
4791 	if (!platform->has_nhm_msrs || no_msr)
4792 		return;
4793 
4794 	switch (platform->cst_limit) {
4795 	case CST_LIMIT_NHM:
4796 		pkg_cstate_limits = nhm_pkg_cstate_limits;
4797 		break;
4798 	case CST_LIMIT_SNB:
4799 		pkg_cstate_limits = snb_pkg_cstate_limits;
4800 		break;
4801 	case CST_LIMIT_HSW:
4802 		pkg_cstate_limits = hsw_pkg_cstate_limits;
4803 		break;
4804 	case CST_LIMIT_SKX:
4805 		pkg_cstate_limits = skx_pkg_cstate_limits;
4806 		break;
4807 	case CST_LIMIT_ICX:
4808 		pkg_cstate_limits = icx_pkg_cstate_limits;
4809 		break;
4810 	case CST_LIMIT_SLV:
4811 		pkg_cstate_limits = slv_pkg_cstate_limits;
4812 		break;
4813 	case CST_LIMIT_AMT:
4814 		pkg_cstate_limits = amt_pkg_cstate_limits;
4815 		break;
4816 	case CST_LIMIT_KNL:
4817 		pkg_cstate_limits = phi_pkg_cstate_limits;
4818 		break;
4819 	case CST_LIMIT_GMT:
4820 		pkg_cstate_limits = glm_pkg_cstate_limits;
4821 		break;
4822 	default:
4823 		return;
4824 	}
4825 
4826 	get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
4827 	pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
4828 }
4829 
dump_platform_info(void)4830 static void dump_platform_info(void)
4831 {
4832 	unsigned long long msr;
4833 	unsigned int ratio;
4834 
4835 	if (!platform->has_nhm_msrs || no_msr)
4836 		return;
4837 
4838 	get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
4839 
4840 	fprintf(outf, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", base_cpu, msr);
4841 
4842 	ratio = (msr >> 40) & 0xFF;
4843 	fprintf(outf, "%d * %.1f = %.1f MHz max efficiency frequency\n", ratio, bclk, ratio * bclk);
4844 
4845 	ratio = (msr >> 8) & 0xFF;
4846 	fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n", ratio, bclk, ratio * bclk);
4847 }
4848 
dump_power_ctl(void)4849 static void dump_power_ctl(void)
4850 {
4851 	unsigned long long msr;
4852 
4853 	if (!platform->has_nhm_msrs || no_msr)
4854 		return;
4855 
4856 	get_msr(base_cpu, MSR_IA32_POWER_CTL, &msr);
4857 	fprintf(outf, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
4858 		base_cpu, msr, msr & 0x2 ? "EN" : "DIS");
4859 
4860 	/* C-state Pre-wake Disable (CSTATE_PREWAKE_DISABLE) */
4861 	if (platform->has_cst_prewake_bit)
4862 		fprintf(outf, "C-state Pre-wake: %sabled\n", msr & 0x40000000 ? "DIS" : "EN");
4863 
4864 	return;
4865 }
4866 
dump_turbo_ratio_limit2(void)4867 static void dump_turbo_ratio_limit2(void)
4868 {
4869 	unsigned long long msr;
4870 	unsigned int ratio;
4871 
4872 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
4873 
4874 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu, msr);
4875 
4876 	ratio = (msr >> 8) & 0xFF;
4877 	if (ratio)
4878 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 18 active cores\n", ratio, bclk, ratio * bclk);
4879 
4880 	ratio = (msr >> 0) & 0xFF;
4881 	if (ratio)
4882 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 17 active cores\n", ratio, bclk, ratio * bclk);
4883 	return;
4884 }
4885 
dump_turbo_ratio_limit1(void)4886 static void dump_turbo_ratio_limit1(void)
4887 {
4888 	unsigned long long msr;
4889 	unsigned int ratio;
4890 
4891 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
4892 
4893 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, msr);
4894 
4895 	ratio = (msr >> 56) & 0xFF;
4896 	if (ratio)
4897 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 16 active cores\n", ratio, bclk, ratio * bclk);
4898 
4899 	ratio = (msr >> 48) & 0xFF;
4900 	if (ratio)
4901 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 15 active cores\n", ratio, bclk, ratio * bclk);
4902 
4903 	ratio = (msr >> 40) & 0xFF;
4904 	if (ratio)
4905 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 14 active cores\n", ratio, bclk, ratio * bclk);
4906 
4907 	ratio = (msr >> 32) & 0xFF;
4908 	if (ratio)
4909 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 13 active cores\n", ratio, bclk, ratio * bclk);
4910 
4911 	ratio = (msr >> 24) & 0xFF;
4912 	if (ratio)
4913 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 12 active cores\n", ratio, bclk, ratio * bclk);
4914 
4915 	ratio = (msr >> 16) & 0xFF;
4916 	if (ratio)
4917 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 11 active cores\n", ratio, bclk, ratio * bclk);
4918 
4919 	ratio = (msr >> 8) & 0xFF;
4920 	if (ratio)
4921 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 10 active cores\n", ratio, bclk, ratio * bclk);
4922 
4923 	ratio = (msr >> 0) & 0xFF;
4924 	if (ratio)
4925 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 9 active cores\n", ratio, bclk, ratio * bclk);
4926 	return;
4927 }
4928 
dump_turbo_ratio_limits(int trl_msr_offset)4929 static void dump_turbo_ratio_limits(int trl_msr_offset)
4930 {
4931 	unsigned long long msr, core_counts;
4932 	int shift;
4933 
4934 	get_msr(base_cpu, trl_msr_offset, &msr);
4935 	fprintf(outf, "cpu%d: MSR_%sTURBO_RATIO_LIMIT: 0x%08llx\n",
4936 		base_cpu, trl_msr_offset == MSR_SECONDARY_TURBO_RATIO_LIMIT ? "SECONDARY_" : "", msr);
4937 
4938 	if (platform->trl_msrs & TRL_CORECOUNT) {
4939 		get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &core_counts);
4940 		fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, core_counts);
4941 	} else {
4942 		core_counts = 0x0807060504030201;
4943 	}
4944 
4945 	for (shift = 56; shift >= 0; shift -= 8) {
4946 		unsigned int ratio, group_size;
4947 
4948 		ratio = (msr >> shift) & 0xFF;
4949 		group_size = (core_counts >> shift) & 0xFF;
4950 		if (ratio)
4951 			fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
4952 				ratio, bclk, ratio * bclk, group_size);
4953 	}
4954 
4955 	return;
4956 }
4957 
dump_atom_turbo_ratio_limits(void)4958 static void dump_atom_turbo_ratio_limits(void)
4959 {
4960 	unsigned long long msr;
4961 	unsigned int ratio;
4962 
4963 	get_msr(base_cpu, MSR_ATOM_CORE_RATIOS, &msr);
4964 	fprintf(outf, "cpu%d: MSR_ATOM_CORE_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
4965 
4966 	ratio = (msr >> 0) & 0x3F;
4967 	if (ratio)
4968 		fprintf(outf, "%d * %.1f = %.1f MHz minimum operating frequency\n", ratio, bclk, ratio * bclk);
4969 
4970 	ratio = (msr >> 8) & 0x3F;
4971 	if (ratio)
4972 		fprintf(outf, "%d * %.1f = %.1f MHz low frequency mode (LFM)\n", ratio, bclk, ratio * bclk);
4973 
4974 	ratio = (msr >> 16) & 0x3F;
4975 	if (ratio)
4976 		fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n", ratio, bclk, ratio * bclk);
4977 
4978 	get_msr(base_cpu, MSR_ATOM_CORE_TURBO_RATIOS, &msr);
4979 	fprintf(outf, "cpu%d: MSR_ATOM_CORE_TURBO_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
4980 
4981 	ratio = (msr >> 24) & 0x3F;
4982 	if (ratio)
4983 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 4 active cores\n", ratio, bclk, ratio * bclk);
4984 
4985 	ratio = (msr >> 16) & 0x3F;
4986 	if (ratio)
4987 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 3 active cores\n", ratio, bclk, ratio * bclk);
4988 
4989 	ratio = (msr >> 8) & 0x3F;
4990 	if (ratio)
4991 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 2 active cores\n", ratio, bclk, ratio * bclk);
4992 
4993 	ratio = (msr >> 0) & 0x3F;
4994 	if (ratio)
4995 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 1 active core\n", ratio, bclk, ratio * bclk);
4996 }
4997 
dump_knl_turbo_ratio_limits(void)4998 static void dump_knl_turbo_ratio_limits(void)
4999 {
5000 	const unsigned int buckets_no = 7;
5001 
5002 	unsigned long long msr;
5003 	int delta_cores, delta_ratio;
5004 	int i, b_nr;
5005 	unsigned int cores[buckets_no];
5006 	unsigned int ratio[buckets_no];
5007 
5008 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
5009 
5010 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu, msr);
5011 
5012 	/*
5013 	 * Turbo encoding in KNL is as follows:
5014 	 * [0] -- Reserved
5015 	 * [7:1] -- Base value of number of active cores of bucket 1.
5016 	 * [15:8] -- Base value of freq ratio of bucket 1.
5017 	 * [20:16] -- +ve delta of number of active cores of bucket 2.
5018 	 * i.e. active cores of bucket 2 =
5019 	 * active cores of bucket 1 + delta
5020 	 * [23:21] -- Negative delta of freq ratio of bucket 2.
5021 	 * i.e. freq ratio of bucket 2 =
5022 	 * freq ratio of bucket 1 - delta
5023 	 * [28:24]-- +ve delta of number of active cores of bucket 3.
5024 	 * [31:29]-- -ve delta of freq ratio of bucket 3.
5025 	 * [36:32]-- +ve delta of number of active cores of bucket 4.
5026 	 * [39:37]-- -ve delta of freq ratio of bucket 4.
5027 	 * [44:40]-- +ve delta of number of active cores of bucket 5.
5028 	 * [47:45]-- -ve delta of freq ratio of bucket 5.
5029 	 * [52:48]-- +ve delta of number of active cores of bucket 6.
5030 	 * [55:53]-- -ve delta of freq ratio of bucket 6.
5031 	 * [60:56]-- +ve delta of number of active cores of bucket 7.
5032 	 * [63:61]-- -ve delta of freq ratio of bucket 7.
5033 	 */
5034 
5035 	b_nr = 0;
5036 	cores[b_nr] = (msr & 0xFF) >> 1;
5037 	ratio[b_nr] = (msr >> 8) & 0xFF;
5038 
5039 	for (i = 16; i < 64; i += 8) {
5040 		delta_cores = (msr >> i) & 0x1F;
5041 		delta_ratio = (msr >> (i + 5)) & 0x7;
5042 
5043 		cores[b_nr + 1] = cores[b_nr] + delta_cores;
5044 		ratio[b_nr + 1] = ratio[b_nr] - delta_ratio;
5045 		b_nr++;
5046 	}
5047 
5048 	for (i = buckets_no - 1; i >= 0; i--)
5049 		if (i > 0 ? ratio[i] != ratio[i - 1] : 1)
5050 			fprintf(outf,
5051 				"%d * %.1f = %.1f MHz max turbo %d active cores\n",
5052 				ratio[i], bclk, ratio[i] * bclk, cores[i]);
5053 }
5054 
dump_cst_cfg(void)5055 static void dump_cst_cfg(void)
5056 {
5057 	unsigned long long msr;
5058 
5059 	if (!platform->has_nhm_msrs || no_msr)
5060 		return;
5061 
5062 	get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
5063 
5064 	fprintf(outf, "cpu%d: MSR_PKG_CST_CONFIG_CONTROL: 0x%08llx", base_cpu, msr);
5065 
5066 	fprintf(outf, " (%s%s%s%s%slocked, pkg-cstate-limit=%d (%s)",
5067 		(msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
5068 		(msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
5069 		(msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
5070 		(msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
5071 		(msr & (1 << 15)) ? "" : "UN", (unsigned int)msr & 0xF, pkg_cstate_limit_strings[pkg_cstate_limit]);
5072 
5073 #define AUTOMATIC_CSTATE_CONVERSION		(1UL << 16)
5074 	if (platform->has_cst_auto_convension) {
5075 		fprintf(outf, ", automatic c-state conversion=%s", (msr & AUTOMATIC_CSTATE_CONVERSION) ? "on" : "off");
5076 	}
5077 
5078 	fprintf(outf, ")\n");
5079 
5080 	return;
5081 }
5082 
dump_config_tdp(void)5083 static void dump_config_tdp(void)
5084 {
5085 	unsigned long long msr;
5086 
5087 	get_msr(base_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
5088 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu, msr);
5089 	fprintf(outf, " (base_ratio=%d)\n", (unsigned int)msr & 0xFF);
5090 
5091 	get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
5092 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu, msr);
5093 	if (msr) {
5094 		fprintf(outf, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
5095 		fprintf(outf, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
5096 		fprintf(outf, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
5097 		fprintf(outf, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0x7FFF);
5098 	}
5099 	fprintf(outf, ")\n");
5100 
5101 	get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
5102 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu, msr);
5103 	if (msr) {
5104 		fprintf(outf, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
5105 		fprintf(outf, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
5106 		fprintf(outf, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
5107 		fprintf(outf, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0x7FFF);
5108 	}
5109 	fprintf(outf, ")\n");
5110 
5111 	get_msr(base_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
5112 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu, msr);
5113 	if ((msr) & 0x3)
5114 		fprintf(outf, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
5115 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
5116 	fprintf(outf, ")\n");
5117 
5118 	get_msr(base_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
5119 	fprintf(outf, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu, msr);
5120 	fprintf(outf, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xFF);
5121 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
5122 	fprintf(outf, ")\n");
5123 }
5124 
5125 unsigned int irtl_time_units[] = { 1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
5126 
print_irtl(void)5127 void print_irtl(void)
5128 {
5129 	unsigned long long msr;
5130 
5131 	if (!platform->has_irtl_msrs || no_msr)
5132 		return;
5133 
5134 	if (platform->supported_cstates & PC3) {
5135 		get_msr(base_cpu, MSR_PKGC3_IRTL, &msr);
5136 		fprintf(outf, "cpu%d: MSR_PKGC3_IRTL: 0x%08llx (", base_cpu, msr);
5137 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5138 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5139 	}
5140 
5141 	if (platform->supported_cstates & PC6) {
5142 		get_msr(base_cpu, MSR_PKGC6_IRTL, &msr);
5143 		fprintf(outf, "cpu%d: MSR_PKGC6_IRTL: 0x%08llx (", base_cpu, msr);
5144 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5145 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5146 	}
5147 
5148 	if (platform->supported_cstates & PC7) {
5149 		get_msr(base_cpu, MSR_PKGC7_IRTL, &msr);
5150 		fprintf(outf, "cpu%d: MSR_PKGC7_IRTL: 0x%08llx (", base_cpu, msr);
5151 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5152 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5153 	}
5154 
5155 	if (platform->supported_cstates & PC8) {
5156 		get_msr(base_cpu, MSR_PKGC8_IRTL, &msr);
5157 		fprintf(outf, "cpu%d: MSR_PKGC8_IRTL: 0x%08llx (", base_cpu, msr);
5158 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5159 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5160 	}
5161 
5162 	if (platform->supported_cstates & PC9) {
5163 		get_msr(base_cpu, MSR_PKGC9_IRTL, &msr);
5164 		fprintf(outf, "cpu%d: MSR_PKGC9_IRTL: 0x%08llx (", base_cpu, msr);
5165 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5166 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5167 	}
5168 
5169 	if (platform->supported_cstates & PC10) {
5170 		get_msr(base_cpu, MSR_PKGC10_IRTL, &msr);
5171 		fprintf(outf, "cpu%d: MSR_PKGC10_IRTL: 0x%08llx (", base_cpu, msr);
5172 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
5173 			(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5174 	}
5175 }
5176 
free_fd_percpu(void)5177 void free_fd_percpu(void)
5178 {
5179 	int i;
5180 
5181 	if (!fd_percpu)
5182 		return;
5183 
5184 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
5185 		if (fd_percpu[i] != 0)
5186 			close(fd_percpu[i]);
5187 	}
5188 
5189 	free(fd_percpu);
5190 	fd_percpu = NULL;
5191 }
5192 
free_fd_instr_count_percpu(void)5193 void free_fd_instr_count_percpu(void)
5194 {
5195 	if (!fd_instr_count_percpu)
5196 		return;
5197 
5198 	for (int i = 0; i < topo.max_cpu_num + 1; ++i) {
5199 		if (fd_instr_count_percpu[i] != 0)
5200 			close(fd_instr_count_percpu[i]);
5201 	}
5202 
5203 	free(fd_instr_count_percpu);
5204 	fd_instr_count_percpu = NULL;
5205 }
5206 
free_fd_cstate(void)5207 void free_fd_cstate(void)
5208 {
5209 	if (!ccstate_counter_info)
5210 		return;
5211 
5212 	const int counter_info_num = ccstate_counter_info_size;
5213 
5214 	for (int counter_id = 0; counter_id < counter_info_num; ++counter_id) {
5215 		if (ccstate_counter_info[counter_id].fd_perf_core != -1)
5216 			close(ccstate_counter_info[counter_id].fd_perf_core);
5217 
5218 		if (ccstate_counter_info[counter_id].fd_perf_pkg != -1)
5219 			close(ccstate_counter_info[counter_id].fd_perf_pkg);
5220 	}
5221 
5222 	free(ccstate_counter_info);
5223 	ccstate_counter_info = NULL;
5224 	ccstate_counter_info_size = 0;
5225 }
5226 
free_fd_msr(void)5227 void free_fd_msr(void)
5228 {
5229 	if (!msr_counter_info)
5230 		return;
5231 
5232 	for (int cpu = 0; cpu < topo.max_cpu_num; ++cpu) {
5233 		if (msr_counter_info[cpu].fd_perf != -1)
5234 			close(msr_counter_info[cpu].fd_perf);
5235 	}
5236 
5237 	free(msr_counter_info);
5238 	msr_counter_info = NULL;
5239 	msr_counter_info_size = 0;
5240 }
5241 
free_fd_rapl_percpu(void)5242 void free_fd_rapl_percpu(void)
5243 {
5244 	if (!rapl_counter_info_perdomain)
5245 		return;
5246 
5247 	const int num_domains = rapl_counter_info_perdomain_size;
5248 
5249 	for (int domain_id = 0; domain_id < num_domains; ++domain_id) {
5250 		if (rapl_counter_info_perdomain[domain_id].fd_perf != -1)
5251 			close(rapl_counter_info_perdomain[domain_id].fd_perf);
5252 	}
5253 
5254 	free(rapl_counter_info_perdomain);
5255 	rapl_counter_info_perdomain = NULL;
5256 	rapl_counter_info_perdomain_size = 0;
5257 }
5258 
free_fd_added_perf_counters_(struct perf_counter_info * pp)5259 void free_fd_added_perf_counters_(struct perf_counter_info *pp)
5260 {
5261 	if (!pp)
5262 		return;
5263 
5264 	if (!pp->fd_perf_per_domain)
5265 		return;
5266 
5267 	while (pp) {
5268 		for (size_t domain = 0; domain < pp->num_domains; ++domain) {
5269 			if (pp->fd_perf_per_domain[domain] != -1) {
5270 				close(pp->fd_perf_per_domain[domain]);
5271 				pp->fd_perf_per_domain[domain] = -1;
5272 			}
5273 		}
5274 
5275 		free(pp->fd_perf_per_domain);
5276 		pp->fd_perf_per_domain = NULL;
5277 
5278 		pp = pp->next;
5279 	}
5280 }
5281 
free_fd_added_perf_counters(void)5282 void free_fd_added_perf_counters(void)
5283 {
5284 	free_fd_added_perf_counters_(sys.perf_tp);
5285 	free_fd_added_perf_counters_(sys.perf_cp);
5286 	free_fd_added_perf_counters_(sys.perf_pp);
5287 }
5288 
free_all_buffers(void)5289 void free_all_buffers(void)
5290 {
5291 	int i;
5292 
5293 	CPU_FREE(cpu_present_set);
5294 	cpu_present_set = NULL;
5295 	cpu_present_setsize = 0;
5296 
5297 	CPU_FREE(cpu_effective_set);
5298 	cpu_effective_set = NULL;
5299 	cpu_effective_setsize = 0;
5300 
5301 	CPU_FREE(cpu_allowed_set);
5302 	cpu_allowed_set = NULL;
5303 	cpu_allowed_setsize = 0;
5304 
5305 	CPU_FREE(cpu_affinity_set);
5306 	cpu_affinity_set = NULL;
5307 	cpu_affinity_setsize = 0;
5308 
5309 	free(thread_even);
5310 	free(core_even);
5311 	free(package_even);
5312 
5313 	thread_even = NULL;
5314 	core_even = NULL;
5315 	package_even = NULL;
5316 
5317 	free(thread_odd);
5318 	free(core_odd);
5319 	free(package_odd);
5320 
5321 	thread_odd = NULL;
5322 	core_odd = NULL;
5323 	package_odd = NULL;
5324 
5325 	free(output_buffer);
5326 	output_buffer = NULL;
5327 	outp = NULL;
5328 
5329 	free_fd_percpu();
5330 	free_fd_instr_count_percpu();
5331 	free_fd_msr();
5332 	free_fd_rapl_percpu();
5333 	free_fd_cstate();
5334 	free_fd_added_perf_counters();
5335 
5336 	free(irq_column_2_cpu);
5337 	free(irqs_per_cpu);
5338 
5339 	for (i = 0; i <= topo.max_cpu_num; ++i) {
5340 		if (cpus[i].put_ids)
5341 			CPU_FREE(cpus[i].put_ids);
5342 	}
5343 	free(cpus);
5344 }
5345 
5346 /*
5347  * Parse a file containing a single int.
5348  * Return 0 if file can not be opened
5349  * Exit if file can be opened, but can not be parsed
5350  */
parse_int_file(const char * fmt,...)5351 int parse_int_file(const char *fmt, ...)
5352 {
5353 	va_list args;
5354 	char path[PATH_MAX];
5355 	FILE *filep;
5356 	int value;
5357 
5358 	va_start(args, fmt);
5359 	vsnprintf(path, sizeof(path), fmt, args);
5360 	va_end(args);
5361 	filep = fopen(path, "r");
5362 	if (!filep)
5363 		return 0;
5364 	if (fscanf(filep, "%d", &value) != 1)
5365 		err(1, "%s: failed to parse number from file", path);
5366 	fclose(filep);
5367 	return value;
5368 }
5369 
5370 /*
5371  * cpu_is_first_core_in_package(cpu)
5372  * return 1 if given CPU is 1st core in package
5373  */
cpu_is_first_core_in_package(int cpu)5374 int cpu_is_first_core_in_package(int cpu)
5375 {
5376 	return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
5377 }
5378 
get_physical_package_id(int cpu)5379 int get_physical_package_id(int cpu)
5380 {
5381 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
5382 }
5383 
get_die_id(int cpu)5384 int get_die_id(int cpu)
5385 {
5386 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/die_id", cpu);
5387 }
5388 
get_core_id(int cpu)5389 int get_core_id(int cpu)
5390 {
5391 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
5392 }
5393 
set_node_data(void)5394 void set_node_data(void)
5395 {
5396 	int pkg, node, lnode, cpu, cpux;
5397 	int cpu_count;
5398 
5399 	/* initialize logical_node_id */
5400 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu)
5401 		cpus[cpu].logical_node_id = -1;
5402 
5403 	cpu_count = 0;
5404 	for (pkg = 0; pkg < topo.num_packages; pkg++) {
5405 		lnode = 0;
5406 		for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
5407 			if (cpus[cpu].physical_package_id != pkg)
5408 				continue;
5409 			/* find a cpu with an unset logical_node_id */
5410 			if (cpus[cpu].logical_node_id != -1)
5411 				continue;
5412 			cpus[cpu].logical_node_id = lnode;
5413 			node = cpus[cpu].physical_node_id;
5414 			cpu_count++;
5415 			/*
5416 			 * find all matching cpus on this pkg and set
5417 			 * the logical_node_id
5418 			 */
5419 			for (cpux = cpu; cpux <= topo.max_cpu_num; cpux++) {
5420 				if ((cpus[cpux].physical_package_id == pkg) && (cpus[cpux].physical_node_id == node)) {
5421 					cpus[cpux].logical_node_id = lnode;
5422 					cpu_count++;
5423 				}
5424 			}
5425 			lnode++;
5426 			if (lnode > topo.nodes_per_pkg)
5427 				topo.nodes_per_pkg = lnode;
5428 		}
5429 		if (cpu_count >= topo.max_cpu_num)
5430 			break;
5431 	}
5432 }
5433 
get_physical_node_id(struct cpu_topology * thiscpu)5434 int get_physical_node_id(struct cpu_topology *thiscpu)
5435 {
5436 	char path[80];
5437 	FILE *filep;
5438 	int i;
5439 	int cpu = thiscpu->logical_cpu_id;
5440 
5441 	for (i = 0; i <= topo.max_cpu_num; i++) {
5442 		sprintf(path, "/sys/devices/system/cpu/cpu%d/node%i/cpulist", cpu, i);
5443 		filep = fopen(path, "r");
5444 		if (!filep)
5445 			continue;
5446 		fclose(filep);
5447 		return i;
5448 	}
5449 	return -1;
5450 }
5451 
parse_cpu_str(char * cpu_str,cpu_set_t * cpu_set,int cpu_set_size)5452 static int parse_cpu_str(char *cpu_str, cpu_set_t *cpu_set, int cpu_set_size)
5453 {
5454 	unsigned int start, end;
5455 	char *next = cpu_str;
5456 
5457 	while (next && *next) {
5458 
5459 		if (*next == '-')	/* no negative cpu numbers */
5460 			return 1;
5461 
5462 		if (*next == '\0' || *next == '\n')
5463 			break;
5464 
5465 		start = strtoul(next, &next, 10);
5466 
5467 		if (start >= CPU_SUBSET_MAXCPUS)
5468 			return 1;
5469 		CPU_SET_S(start, cpu_set_size, cpu_set);
5470 
5471 		if (*next == '\0' || *next == '\n')
5472 			break;
5473 
5474 		if (*next == ',') {
5475 			next += 1;
5476 			continue;
5477 		}
5478 
5479 		if (*next == '-') {
5480 			next += 1;	/* start range */
5481 		} else if (*next == '.') {
5482 			next += 1;
5483 			if (*next == '.')
5484 				next += 1;	/* start range */
5485 			else
5486 				return 1;
5487 		}
5488 
5489 		end = strtoul(next, &next, 10);
5490 		if (end <= start)
5491 			return 1;
5492 
5493 		while (++start <= end) {
5494 			if (start >= CPU_SUBSET_MAXCPUS)
5495 				return 1;
5496 			CPU_SET_S(start, cpu_set_size, cpu_set);
5497 		}
5498 
5499 		if (*next == ',')
5500 			next += 1;
5501 		else if (*next != '\0' && *next != '\n')
5502 			return 1;
5503 	}
5504 
5505 	return 0;
5506 }
5507 
get_thread_siblings(struct cpu_topology * thiscpu)5508 int get_thread_siblings(struct cpu_topology *thiscpu)
5509 {
5510 	char path[80], character;
5511 	FILE *filep;
5512 	unsigned long map;
5513 	int so, shift, sib_core;
5514 	int cpu = thiscpu->logical_cpu_id;
5515 	int offset = topo.max_cpu_num + 1;
5516 	size_t size;
5517 	int thread_id = 0;
5518 
5519 	thiscpu->put_ids = CPU_ALLOC((topo.max_cpu_num + 1));
5520 	if (thiscpu->thread_id < 0)
5521 		thiscpu->thread_id = thread_id++;
5522 	if (!thiscpu->put_ids)
5523 		return -1;
5524 
5525 	size = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
5526 	CPU_ZERO_S(size, thiscpu->put_ids);
5527 
5528 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", cpu);
5529 	filep = fopen(path, "r");
5530 
5531 	if (!filep) {
5532 		warnx("%s: open failed", path);
5533 		return -1;
5534 	}
5535 	do {
5536 		offset -= BITMASK_SIZE;
5537 		if (fscanf(filep, "%lx%c", &map, &character) != 2)
5538 			err(1, "%s: failed to parse file", path);
5539 		for (shift = 0; shift < BITMASK_SIZE; shift++) {
5540 			if ((map >> shift) & 0x1) {
5541 				so = shift + offset;
5542 				sib_core = get_core_id(so);
5543 				if (sib_core == thiscpu->physical_core_id) {
5544 					CPU_SET_S(so, size, thiscpu->put_ids);
5545 					if ((so != cpu) && (cpus[so].thread_id < 0))
5546 						cpus[so].thread_id = thread_id++;
5547 				}
5548 			}
5549 		}
5550 	} while (character == ',');
5551 	fclose(filep);
5552 
5553 	return CPU_COUNT_S(size, thiscpu->put_ids);
5554 }
5555 
5556 /*
5557  * run func(thread, core, package) in topology order
5558  * skip non-present cpus
5559  */
5560 
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)5561 int for_all_cpus_2(int (func) (struct thread_data *, struct core_data *,
5562 			       struct pkg_data *, struct thread_data *, struct core_data *,
5563 			       struct pkg_data *), struct thread_data *thread_base,
5564 		   struct core_data *core_base, struct pkg_data *pkg_base,
5565 		   struct thread_data *thread_base2, struct core_data *core_base2, struct pkg_data *pkg_base2)
5566 {
5567 	int retval, pkg_no, node_no, core_no, thread_no;
5568 
5569 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
5570 		for (node_no = 0; node_no < topo.nodes_per_pkg; ++node_no) {
5571 			for (core_no = 0; core_no < topo.cores_per_node; ++core_no) {
5572 				for (thread_no = 0; thread_no < topo.threads_per_core; ++thread_no) {
5573 					struct thread_data *t, *t2;
5574 					struct core_data *c, *c2;
5575 					struct pkg_data *p, *p2;
5576 
5577 					t = GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no);
5578 
5579 					if (cpu_is_not_allowed(t->cpu_id))
5580 						continue;
5581 
5582 					t2 = GET_THREAD(thread_base2, thread_no, core_no, node_no, pkg_no);
5583 
5584 					c = GET_CORE(core_base, core_no, node_no, pkg_no);
5585 					c2 = GET_CORE(core_base2, core_no, node_no, pkg_no);
5586 
5587 					p = GET_PKG(pkg_base, pkg_no);
5588 					p2 = GET_PKG(pkg_base2, pkg_no);
5589 
5590 					retval = func(t, c, p, t2, c2, p2);
5591 					if (retval)
5592 						return retval;
5593 				}
5594 			}
5595 		}
5596 	}
5597 	return 0;
5598 }
5599 
5600 /*
5601  * run func(cpu) on every cpu in /proc/stat
5602  * return max_cpu number
5603  */
for_all_proc_cpus(int (func)(int))5604 int for_all_proc_cpus(int (func) (int))
5605 {
5606 	FILE *fp;
5607 	int cpu_num;
5608 	int retval;
5609 
5610 	fp = fopen_or_die(proc_stat, "r");
5611 
5612 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
5613 	if (retval != 0)
5614 		err(1, "%s: failed to parse format", proc_stat);
5615 
5616 	while (1) {
5617 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
5618 		if (retval != 1)
5619 			break;
5620 
5621 		retval = func(cpu_num);
5622 		if (retval) {
5623 			fclose(fp);
5624 			return (retval);
5625 		}
5626 	}
5627 	fclose(fp);
5628 	return 0;
5629 }
5630 
5631 #define PATH_EFFECTIVE_CPUS	"/sys/fs/cgroup/cpuset.cpus.effective"
5632 
5633 static char cpu_effective_str[1024];
5634 
update_effective_str(bool startup)5635 static int update_effective_str(bool startup)
5636 {
5637 	FILE *fp;
5638 	char *pos;
5639 	char buf[1024];
5640 	int ret;
5641 
5642 	if (cpu_effective_str[0] == '\0' && !startup)
5643 		return 0;
5644 
5645 	fp = fopen(PATH_EFFECTIVE_CPUS, "r");
5646 	if (!fp)
5647 		return 0;
5648 
5649 	pos = fgets(buf, 1024, fp);
5650 	if (!pos)
5651 		err(1, "%s: file read failed\n", PATH_EFFECTIVE_CPUS);
5652 
5653 	fclose(fp);
5654 
5655 	ret = strncmp(cpu_effective_str, buf, 1024);
5656 	if (!ret)
5657 		return 0;
5658 
5659 	strncpy(cpu_effective_str, buf, 1024);
5660 	return 1;
5661 }
5662 
update_effective_set(bool startup)5663 static void update_effective_set(bool startup)
5664 {
5665 	update_effective_str(startup);
5666 
5667 	if (parse_cpu_str(cpu_effective_str, cpu_effective_set, cpu_effective_setsize))
5668 		err(1, "%s: cpu str malformat %s\n", PATH_EFFECTIVE_CPUS, cpu_effective_str);
5669 }
5670 
5671 void linux_perf_init(void);
5672 void msr_perf_init(void);
5673 void rapl_perf_init(void);
5674 void cstate_perf_init(void);
5675 void added_perf_counters_init(void);
5676 void pmt_init(void);
5677 
re_initialize(void)5678 void re_initialize(void)
5679 {
5680 	free_all_buffers();
5681 	setup_all_buffers(false);
5682 	linux_perf_init();
5683 	msr_perf_init();
5684 	rapl_perf_init();
5685 	cstate_perf_init();
5686 	added_perf_counters_init();
5687 	pmt_init();
5688 	fprintf(outf, "turbostat: re-initialized with num_cpus %d, allowed_cpus %d\n", topo.num_cpus,
5689 		topo.allowed_cpus);
5690 }
5691 
set_max_cpu_num(void)5692 void set_max_cpu_num(void)
5693 {
5694 	FILE *filep;
5695 	int base_cpu;
5696 	unsigned long dummy;
5697 	char pathname[64];
5698 
5699 	base_cpu = sched_getcpu();
5700 	if (base_cpu < 0)
5701 		err(1, "cannot find calling cpu ID");
5702 	sprintf(pathname, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", base_cpu);
5703 
5704 	filep = fopen_or_die(pathname, "r");
5705 	topo.max_cpu_num = 0;
5706 	while (fscanf(filep, "%lx,", &dummy) == 1)
5707 		topo.max_cpu_num += BITMASK_SIZE;
5708 	fclose(filep);
5709 	topo.max_cpu_num--;	/* 0 based */
5710 }
5711 
5712 /*
5713  * count_cpus()
5714  * remember the last one seen, it will be the max
5715  */
count_cpus(int cpu)5716 int count_cpus(int cpu)
5717 {
5718 	UNUSED(cpu);
5719 
5720 	topo.num_cpus++;
5721 	return 0;
5722 }
5723 
mark_cpu_present(int cpu)5724 int mark_cpu_present(int cpu)
5725 {
5726 	CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
5727 	return 0;
5728 }
5729 
init_thread_id(int cpu)5730 int init_thread_id(int cpu)
5731 {
5732 	cpus[cpu].thread_id = -1;
5733 	return 0;
5734 }
5735 
set_my_cpu_type(void)5736 int set_my_cpu_type(void)
5737 {
5738 	unsigned int eax, ebx, ecx, edx;
5739 	unsigned int max_level;
5740 
5741 	__cpuid(0, max_level, ebx, ecx, edx);
5742 
5743 	if (max_level < CPUID_LEAF_MODEL_ID)
5744 		return 0;
5745 
5746 	__cpuid(CPUID_LEAF_MODEL_ID, eax, ebx, ecx, edx);
5747 
5748 	return (eax >> CPUID_LEAF_MODEL_ID_CORE_TYPE_SHIFT);
5749 }
5750 
set_cpu_hybrid_type(int cpu)5751 int set_cpu_hybrid_type(int cpu)
5752 {
5753 	if (cpu_migrate(cpu))
5754 		return -1;
5755 
5756 	int type = set_my_cpu_type();
5757 
5758 	cpus[cpu].type = type;
5759 	return 0;
5760 }
5761 
5762 /*
5763  * snapshot_proc_interrupts()
5764  *
5765  * read and record summary of /proc/interrupts
5766  *
5767  * return 1 if config change requires a restart, else return 0
5768  */
snapshot_proc_interrupts(void)5769 int snapshot_proc_interrupts(void)
5770 {
5771 	static FILE *fp;
5772 	int column, retval;
5773 
5774 	if (fp == NULL)
5775 		fp = fopen_or_die("/proc/interrupts", "r");
5776 	else
5777 		rewind(fp);
5778 
5779 	/* read 1st line of /proc/interrupts to get cpu* name for each column */
5780 	for (column = 0; column < topo.num_cpus; ++column) {
5781 		int cpu_number;
5782 
5783 		retval = fscanf(fp, " CPU%d", &cpu_number);
5784 		if (retval != 1)
5785 			break;
5786 
5787 		if (cpu_number > topo.max_cpu_num) {
5788 			warn("/proc/interrupts: cpu%d: > %d", cpu_number, topo.max_cpu_num);
5789 			return 1;
5790 		}
5791 
5792 		irq_column_2_cpu[column] = cpu_number;
5793 		irqs_per_cpu[cpu_number] = 0;
5794 	}
5795 
5796 	/* read /proc/interrupt count lines and sum up irqs per cpu */
5797 	while (1) {
5798 		int column;
5799 		char buf[64];
5800 
5801 		retval = fscanf(fp, " %s:", buf);	/* flush irq# "N:" */
5802 		if (retval != 1)
5803 			break;
5804 
5805 		/* read the count per cpu */
5806 		for (column = 0; column < topo.num_cpus; ++column) {
5807 
5808 			int cpu_number, irq_count;
5809 
5810 			retval = fscanf(fp, " %d", &irq_count);
5811 			if (retval != 1)
5812 				break;
5813 
5814 			cpu_number = irq_column_2_cpu[column];
5815 			irqs_per_cpu[cpu_number] += irq_count;
5816 
5817 		}
5818 
5819 		while (getc(fp) != '\n') ;	/* flush interrupt description */
5820 
5821 	}
5822 	return 0;
5823 }
5824 
5825 /*
5826  * snapshot_graphics()
5827  *
5828  * record snapshot of specified graphics sysfs knob
5829  *
5830  * return 1 if config change requires a restart, else return 0
5831  */
snapshot_graphics(int idx)5832 int snapshot_graphics(int idx)
5833 {
5834 	int retval;
5835 
5836 	rewind(gfx_info[idx].fp);
5837 
5838 	switch (idx) {
5839 	case GFX_rc6:
5840 	case SAM_mc6:
5841 		retval = fscanf(gfx_info[idx].fp, "%lld", &gfx_info[idx].val_ull);
5842 		if (retval != 1)
5843 			err(1, "rc6");
5844 		return 0;
5845 	case GFX_MHz:
5846 	case GFX_ACTMHz:
5847 	case SAM_MHz:
5848 	case SAM_ACTMHz:
5849 		retval = fscanf(gfx_info[idx].fp, "%d", &gfx_info[idx].val);
5850 		if (retval != 1)
5851 			err(1, "MHz");
5852 		return 0;
5853 	default:
5854 		return -EINVAL;
5855 	}
5856 }
5857 
5858 /*
5859  * snapshot_cpu_lpi()
5860  *
5861  * record snapshot of
5862  * /sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us
5863  */
snapshot_cpu_lpi_us(void)5864 int snapshot_cpu_lpi_us(void)
5865 {
5866 	FILE *fp;
5867 	int retval;
5868 
5869 	fp = fopen_or_die("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", "r");
5870 
5871 	retval = fscanf(fp, "%lld", &cpuidle_cur_cpu_lpi_us);
5872 	if (retval != 1) {
5873 		fprintf(stderr, "Disabling Low Power Idle CPU output\n");
5874 		BIC_NOT_PRESENT(BIC_CPU_LPI);
5875 		fclose(fp);
5876 		return -1;
5877 	}
5878 
5879 	fclose(fp);
5880 
5881 	return 0;
5882 }
5883 
5884 /*
5885  * snapshot_sys_lpi()
5886  *
5887  * record snapshot of sys_lpi_file
5888  */
snapshot_sys_lpi_us(void)5889 int snapshot_sys_lpi_us(void)
5890 {
5891 	FILE *fp;
5892 	int retval;
5893 
5894 	fp = fopen_or_die(sys_lpi_file, "r");
5895 
5896 	retval = fscanf(fp, "%lld", &cpuidle_cur_sys_lpi_us);
5897 	if (retval != 1) {
5898 		fprintf(stderr, "Disabling Low Power Idle System output\n");
5899 		BIC_NOT_PRESENT(BIC_SYS_LPI);
5900 		fclose(fp);
5901 		return -1;
5902 	}
5903 	fclose(fp);
5904 
5905 	return 0;
5906 }
5907 
5908 /*
5909  * snapshot /proc and /sys files
5910  *
5911  * return 1 if configuration restart needed, else return 0
5912  */
snapshot_proc_sysfs_files(void)5913 int snapshot_proc_sysfs_files(void)
5914 {
5915 	if (DO_BIC(BIC_IRQ))
5916 		if (snapshot_proc_interrupts())
5917 			return 1;
5918 
5919 	if (DO_BIC(BIC_GFX_rc6))
5920 		snapshot_graphics(GFX_rc6);
5921 
5922 	if (DO_BIC(BIC_GFXMHz))
5923 		snapshot_graphics(GFX_MHz);
5924 
5925 	if (DO_BIC(BIC_GFXACTMHz))
5926 		snapshot_graphics(GFX_ACTMHz);
5927 
5928 	if (DO_BIC(BIC_SAM_mc6))
5929 		snapshot_graphics(SAM_mc6);
5930 
5931 	if (DO_BIC(BIC_SAMMHz))
5932 		snapshot_graphics(SAM_MHz);
5933 
5934 	if (DO_BIC(BIC_SAMACTMHz))
5935 		snapshot_graphics(SAM_ACTMHz);
5936 
5937 	if (DO_BIC(BIC_CPU_LPI))
5938 		snapshot_cpu_lpi_us();
5939 
5940 	if (DO_BIC(BIC_SYS_LPI))
5941 		snapshot_sys_lpi_us();
5942 
5943 	return 0;
5944 }
5945 
5946 int exit_requested;
5947 
signal_handler(int signal)5948 static void signal_handler(int signal)
5949 {
5950 	switch (signal) {
5951 	case SIGINT:
5952 		exit_requested = 1;
5953 		if (debug)
5954 			fprintf(stderr, " SIGINT\n");
5955 		break;
5956 	case SIGUSR1:
5957 		if (debug > 1)
5958 			fprintf(stderr, "SIGUSR1\n");
5959 		break;
5960 	}
5961 }
5962 
setup_signal_handler(void)5963 void setup_signal_handler(void)
5964 {
5965 	struct sigaction sa;
5966 
5967 	memset(&sa, 0, sizeof(sa));
5968 
5969 	sa.sa_handler = &signal_handler;
5970 
5971 	if (sigaction(SIGINT, &sa, NULL) < 0)
5972 		err(1, "sigaction SIGINT");
5973 	if (sigaction(SIGUSR1, &sa, NULL) < 0)
5974 		err(1, "sigaction SIGUSR1");
5975 }
5976 
do_sleep(void)5977 void do_sleep(void)
5978 {
5979 	struct timeval tout;
5980 	struct timespec rest;
5981 	fd_set readfds;
5982 	int retval;
5983 
5984 	FD_ZERO(&readfds);
5985 	FD_SET(0, &readfds);
5986 
5987 	if (ignore_stdin) {
5988 		nanosleep(&interval_ts, NULL);
5989 		return;
5990 	}
5991 
5992 	tout = interval_tv;
5993 	retval = select(1, &readfds, NULL, NULL, &tout);
5994 
5995 	if (retval == 1) {
5996 		switch (getc(stdin)) {
5997 		case 'q':
5998 			exit_requested = 1;
5999 			break;
6000 		case EOF:
6001 			/*
6002 			 * 'stdin' is a pipe closed on the other end. There
6003 			 * won't be any further input.
6004 			 */
6005 			ignore_stdin = 1;
6006 			/* Sleep the rest of the time */
6007 			rest.tv_sec = (tout.tv_sec + tout.tv_usec / 1000000);
6008 			rest.tv_nsec = (tout.tv_usec % 1000000) * 1000;
6009 			nanosleep(&rest, NULL);
6010 		}
6011 	}
6012 }
6013 
get_msr_sum(int cpu,off_t offset,unsigned long long * msr)6014 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr)
6015 {
6016 	int ret, idx;
6017 	unsigned long long msr_cur, msr_last;
6018 
6019 	assert(!no_msr);
6020 
6021 	if (!per_cpu_msr_sum)
6022 		return 1;
6023 
6024 	idx = offset_to_idx(offset);
6025 	if (idx < 0)
6026 		return idx;
6027 	/* get_msr_sum() = sum + (get_msr() - last) */
6028 	ret = get_msr(cpu, offset, &msr_cur);
6029 	if (ret)
6030 		return ret;
6031 	msr_last = per_cpu_msr_sum[cpu].entries[idx].last;
6032 	DELTA_WRAP32(msr_cur, msr_last);
6033 	*msr = msr_last + per_cpu_msr_sum[cpu].entries[idx].sum;
6034 
6035 	return 0;
6036 }
6037 
6038 timer_t timerid;
6039 
6040 /* Timer callback, update the sum of MSRs periodically. */
update_msr_sum(struct thread_data * t,struct core_data * c,struct pkg_data * p)6041 static int update_msr_sum(struct thread_data *t, struct core_data *c, struct pkg_data *p)
6042 {
6043 	int i, ret;
6044 	int cpu = t->cpu_id;
6045 
6046 	UNUSED(c);
6047 	UNUSED(p);
6048 
6049 	assert(!no_msr);
6050 
6051 	for (i = IDX_PKG_ENERGY; i < IDX_COUNT; i++) {
6052 		unsigned long long msr_cur, msr_last;
6053 		off_t offset;
6054 
6055 		if (!idx_valid(i))
6056 			continue;
6057 		offset = idx_to_offset(i);
6058 		if (offset < 0)
6059 			continue;
6060 		ret = get_msr(cpu, offset, &msr_cur);
6061 		if (ret) {
6062 			fprintf(outf, "Can not update msr(0x%llx)\n", (unsigned long long)offset);
6063 			continue;
6064 		}
6065 
6066 		msr_last = per_cpu_msr_sum[cpu].entries[i].last;
6067 		per_cpu_msr_sum[cpu].entries[i].last = msr_cur & 0xffffffff;
6068 
6069 		DELTA_WRAP32(msr_cur, msr_last);
6070 		per_cpu_msr_sum[cpu].entries[i].sum += msr_last;
6071 	}
6072 	return 0;
6073 }
6074 
msr_record_handler(union sigval v)6075 static void msr_record_handler(union sigval v)
6076 {
6077 	UNUSED(v);
6078 
6079 	for_all_cpus(update_msr_sum, EVEN_COUNTERS);
6080 }
6081 
msr_sum_record(void)6082 void msr_sum_record(void)
6083 {
6084 	struct itimerspec its;
6085 	struct sigevent sev;
6086 
6087 	per_cpu_msr_sum = calloc(topo.max_cpu_num + 1, sizeof(struct msr_sum_array));
6088 	if (!per_cpu_msr_sum) {
6089 		fprintf(outf, "Can not allocate memory for long time MSR.\n");
6090 		return;
6091 	}
6092 	/*
6093 	 * Signal handler might be restricted, so use thread notifier instead.
6094 	 */
6095 	memset(&sev, 0, sizeof(struct sigevent));
6096 	sev.sigev_notify = SIGEV_THREAD;
6097 	sev.sigev_notify_function = msr_record_handler;
6098 
6099 	sev.sigev_value.sival_ptr = &timerid;
6100 	if (timer_create(CLOCK_REALTIME, &sev, &timerid) == -1) {
6101 		fprintf(outf, "Can not create timer.\n");
6102 		goto release_msr;
6103 	}
6104 
6105 	its.it_value.tv_sec = 0;
6106 	its.it_value.tv_nsec = 1;
6107 	/*
6108 	 * A wraparound time has been calculated early.
6109 	 * Some sources state that the peak power for a
6110 	 * microprocessor is usually 1.5 times the TDP rating,
6111 	 * use 2 * TDP for safety.
6112 	 */
6113 	its.it_interval.tv_sec = rapl_joule_counter_range / 2;
6114 	its.it_interval.tv_nsec = 0;
6115 
6116 	if (timer_settime(timerid, 0, &its, NULL) == -1) {
6117 		fprintf(outf, "Can not set timer.\n");
6118 		goto release_timer;
6119 	}
6120 	return;
6121 
6122 release_timer:
6123 	timer_delete(timerid);
6124 release_msr:
6125 	free(per_cpu_msr_sum);
6126 }
6127 
6128 /*
6129  * set_my_sched_priority(pri)
6130  * return previous priority on success
6131  * return value < -20 on failure
6132  */
set_my_sched_priority(int priority)6133 int set_my_sched_priority(int priority)
6134 {
6135 	int retval;
6136 	int original_priority;
6137 
6138 	errno = 0;
6139 	original_priority = getpriority(PRIO_PROCESS, 0);
6140 	if (errno && (original_priority == -1))
6141 		return -21;
6142 
6143 	retval = setpriority(PRIO_PROCESS, 0, priority);
6144 	if (retval)
6145 		return -21;
6146 
6147 	errno = 0;
6148 	retval = getpriority(PRIO_PROCESS, 0);
6149 	if (retval != priority)
6150 		return -21;
6151 
6152 	return original_priority;
6153 }
6154 
turbostat_loop()6155 void turbostat_loop()
6156 {
6157 	int retval;
6158 	int restarted = 0;
6159 	unsigned int done_iters = 0;
6160 
6161 	setup_signal_handler();
6162 
6163 	/*
6164 	 * elevate own priority for interval mode
6165 	 *
6166 	 * ignore on error - we probably don't have permission to set it, but
6167 	 * it's not a big deal
6168 	 */
6169 	set_my_sched_priority(-20);
6170 
6171 restart:
6172 	restarted++;
6173 
6174 	snapshot_proc_sysfs_files();
6175 	retval = for_all_cpus(get_counters, EVEN_COUNTERS);
6176 	first_counter_read = 0;
6177 	if (retval < -1) {
6178 		exit(retval);
6179 	} else if (retval == -1) {
6180 		if (restarted > 10) {
6181 			exit(retval);
6182 		}
6183 		re_initialize();
6184 		goto restart;
6185 	}
6186 	restarted = 0;
6187 	done_iters = 0;
6188 	gettimeofday(&tv_even, (struct timezone *)NULL);
6189 
6190 	while (1) {
6191 		if (for_all_proc_cpus(cpu_is_not_present)) {
6192 			re_initialize();
6193 			goto restart;
6194 		}
6195 		if (update_effective_str(false)) {
6196 			re_initialize();
6197 			goto restart;
6198 		}
6199 		do_sleep();
6200 		if (snapshot_proc_sysfs_files())
6201 			goto restart;
6202 		retval = for_all_cpus(get_counters, ODD_COUNTERS);
6203 		if (retval < -1) {
6204 			exit(retval);
6205 		} else if (retval == -1) {
6206 			re_initialize();
6207 			goto restart;
6208 		}
6209 		gettimeofday(&tv_odd, (struct timezone *)NULL);
6210 		timersub(&tv_odd, &tv_even, &tv_delta);
6211 		if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) {
6212 			re_initialize();
6213 			goto restart;
6214 		}
6215 		delta_platform(&platform_counters_odd, &platform_counters_even);
6216 		compute_average(EVEN_COUNTERS);
6217 		format_all_counters(EVEN_COUNTERS);
6218 		flush_output_stdout();
6219 		if (exit_requested)
6220 			break;
6221 		if (num_iterations && ++done_iters >= num_iterations)
6222 			break;
6223 		do_sleep();
6224 		if (snapshot_proc_sysfs_files())
6225 			goto restart;
6226 		retval = for_all_cpus(get_counters, EVEN_COUNTERS);
6227 		if (retval < -1) {
6228 			exit(retval);
6229 		} else if (retval == -1) {
6230 			re_initialize();
6231 			goto restart;
6232 		}
6233 		gettimeofday(&tv_even, (struct timezone *)NULL);
6234 		timersub(&tv_even, &tv_odd, &tv_delta);
6235 		if (for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) {
6236 			re_initialize();
6237 			goto restart;
6238 		}
6239 		delta_platform(&platform_counters_even, &platform_counters_odd);
6240 		compute_average(ODD_COUNTERS);
6241 		format_all_counters(ODD_COUNTERS);
6242 		flush_output_stdout();
6243 		if (exit_requested)
6244 			break;
6245 		if (num_iterations && ++done_iters >= num_iterations)
6246 			break;
6247 	}
6248 }
6249 
check_dev_msr()6250 void check_dev_msr()
6251 {
6252 	struct stat sb;
6253 	char pathname[32];
6254 
6255 	if (no_msr)
6256 		return;
6257 
6258 	sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
6259 	if (stat(pathname, &sb))
6260 		if (system("/sbin/modprobe msr > /dev/null 2>&1"))
6261 			no_msr = 1;
6262 }
6263 
6264 /*
6265  * check for CAP_SYS_RAWIO
6266  * return 0 on success
6267  * return 1 on fail
6268  */
check_for_cap_sys_rawio(void)6269 int check_for_cap_sys_rawio(void)
6270 {
6271 	cap_t caps;
6272 	cap_flag_value_t cap_flag_value;
6273 	int ret = 0;
6274 
6275 	caps = cap_get_proc();
6276 	if (caps == NULL)
6277 		return 1;
6278 
6279 	if (cap_get_flag(caps, CAP_SYS_RAWIO, CAP_EFFECTIVE, &cap_flag_value)) {
6280 		ret = 1;
6281 		goto free_and_exit;
6282 	}
6283 
6284 	if (cap_flag_value != CAP_SET) {
6285 		ret = 1;
6286 		goto free_and_exit;
6287 	}
6288 
6289 free_and_exit:
6290 	if (cap_free(caps) == -1)
6291 		err(-6, "cap_free\n");
6292 
6293 	return ret;
6294 }
6295 
check_msr_permission(void)6296 void check_msr_permission(void)
6297 {
6298 	int failed = 0;
6299 	char pathname[32];
6300 
6301 	if (no_msr)
6302 		return;
6303 
6304 	/* check for CAP_SYS_RAWIO */
6305 	failed += check_for_cap_sys_rawio();
6306 
6307 	/* test file permissions */
6308 	sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
6309 	if (euidaccess(pathname, R_OK)) {
6310 		failed++;
6311 	}
6312 
6313 	if (failed) {
6314 		warnx("Failed to access %s. Some of the counters may not be available\n"
6315 		      "\tRun as root to enable them or use %s to disable the access explicitly", pathname, "--no-msr");
6316 		no_msr = 1;
6317 	}
6318 }
6319 
probe_bclk(void)6320 void probe_bclk(void)
6321 {
6322 	unsigned long long msr;
6323 	unsigned int base_ratio;
6324 
6325 	if (!platform->has_nhm_msrs || no_msr)
6326 		return;
6327 
6328 	if (platform->bclk_freq == BCLK_100MHZ)
6329 		bclk = 100.00;
6330 	else if (platform->bclk_freq == BCLK_133MHZ)
6331 		bclk = 133.33;
6332 	else if (platform->bclk_freq == BCLK_SLV)
6333 		bclk = slm_bclk();
6334 	else
6335 		return;
6336 
6337 	get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
6338 	base_ratio = (msr >> 8) & 0xFF;
6339 
6340 	base_hz = base_ratio * bclk * 1000000;
6341 	has_base_hz = 1;
6342 
6343 	if (platform->enable_tsc_tweak)
6344 		tsc_tweak = base_hz / tsc_hz;
6345 }
6346 
remove_underbar(char * s)6347 static void remove_underbar(char *s)
6348 {
6349 	char *to = s;
6350 
6351 	while (*s) {
6352 		if (*s != '_')
6353 			*to++ = *s;
6354 		s++;
6355 	}
6356 
6357 	*to = 0;
6358 }
6359 
dump_turbo_ratio_info(void)6360 static void dump_turbo_ratio_info(void)
6361 {
6362 	if (!has_turbo)
6363 		return;
6364 
6365 	if (!platform->has_nhm_msrs || no_msr)
6366 		return;
6367 
6368 	if (platform->trl_msrs & TRL_LIMIT2)
6369 		dump_turbo_ratio_limit2();
6370 
6371 	if (platform->trl_msrs & TRL_LIMIT1)
6372 		dump_turbo_ratio_limit1();
6373 
6374 	if (platform->trl_msrs & TRL_BASE) {
6375 		dump_turbo_ratio_limits(MSR_TURBO_RATIO_LIMIT);
6376 
6377 		if (is_hybrid)
6378 			dump_turbo_ratio_limits(MSR_SECONDARY_TURBO_RATIO_LIMIT);
6379 	}
6380 
6381 	if (platform->trl_msrs & TRL_ATOM)
6382 		dump_atom_turbo_ratio_limits();
6383 
6384 	if (platform->trl_msrs & TRL_KNL)
6385 		dump_knl_turbo_ratio_limits();
6386 
6387 	if (platform->has_config_tdp)
6388 		dump_config_tdp();
6389 }
6390 
read_sysfs_int(char * path)6391 static int read_sysfs_int(char *path)
6392 {
6393 	FILE *input;
6394 	int retval = -1;
6395 
6396 	input = fopen(path, "r");
6397 	if (input == NULL) {
6398 		if (debug)
6399 			fprintf(outf, "NSFOD %s\n", path);
6400 		return (-1);
6401 	}
6402 	if (fscanf(input, "%d", &retval) != 1)
6403 		err(1, "%s: failed to read int from file", path);
6404 	fclose(input);
6405 
6406 	return (retval);
6407 }
6408 
dump_sysfs_file(char * path)6409 static void dump_sysfs_file(char *path)
6410 {
6411 	FILE *input;
6412 	char cpuidle_buf[64];
6413 
6414 	input = fopen(path, "r");
6415 	if (input == NULL) {
6416 		if (debug)
6417 			fprintf(outf, "NSFOD %s\n", path);
6418 		return;
6419 	}
6420 	if (!fgets(cpuidle_buf, sizeof(cpuidle_buf), input))
6421 		err(1, "%s: failed to read file", path);
6422 	fclose(input);
6423 
6424 	fprintf(outf, "%s: %s", strrchr(path, '/') + 1, cpuidle_buf);
6425 }
6426 
probe_intel_uncore_frequency_legacy(void)6427 static void probe_intel_uncore_frequency_legacy(void)
6428 {
6429 	int i, j;
6430 	char path[256];
6431 
6432 	for (i = 0; i < topo.num_packages; ++i) {
6433 		for (j = 0; j <= topo.max_die_id; ++j) {
6434 			int k, l;
6435 			char path_base[128];
6436 
6437 			sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/package_%02d_die_%02d", i,
6438 				j);
6439 
6440 			if (access(path_base, R_OK))
6441 				continue;
6442 
6443 			BIC_PRESENT(BIC_UNCORE_MHZ);
6444 
6445 			if (quiet)
6446 				return;
6447 
6448 			sprintf(path, "%s/min_freq_khz", path_base);
6449 			k = read_sysfs_int(path);
6450 			sprintf(path, "%s/max_freq_khz", path_base);
6451 			l = read_sysfs_int(path);
6452 			fprintf(outf, "Uncore Frequency package%d die%d: %d - %d MHz ", i, j, k / 1000, l / 1000);
6453 
6454 			sprintf(path, "%s/initial_min_freq_khz", path_base);
6455 			k = read_sysfs_int(path);
6456 			sprintf(path, "%s/initial_max_freq_khz", path_base);
6457 			l = read_sysfs_int(path);
6458 			fprintf(outf, "(%d - %d MHz)", k / 1000, l / 1000);
6459 
6460 			sprintf(path, "%s/current_freq_khz", path_base);
6461 			k = read_sysfs_int(path);
6462 			fprintf(outf, " %d MHz\n", k / 1000);
6463 		}
6464 	}
6465 }
6466 
probe_intel_uncore_frequency_cluster(void)6467 static void probe_intel_uncore_frequency_cluster(void)
6468 {
6469 	int i, uncore_max_id;
6470 	char path[256];
6471 	char path_base[128];
6472 
6473 	if (access("/sys/devices/system/cpu/intel_uncore_frequency/uncore00/current_freq_khz", R_OK))
6474 		return;
6475 
6476 	for (uncore_max_id = 0;; ++uncore_max_id) {
6477 
6478 		sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/uncore%02d", uncore_max_id);
6479 
6480 		/* uncore## start at 00 and skips no numbers, so stop upon first missing */
6481 		if (access(path_base, R_OK)) {
6482 			uncore_max_id -= 1;
6483 			break;
6484 		}
6485 	}
6486 	for (i = uncore_max_id; i >= 0; --i) {
6487 		int k, l;
6488 		int package_id, domain_id, cluster_id;
6489 		char name_buf[16];
6490 
6491 		sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/uncore%02d", i);
6492 
6493 		if (access(path_base, R_OK))
6494 			err(1, "%s: %s\n", __func__, path_base);
6495 
6496 		sprintf(path, "%s/package_id", path_base);
6497 		package_id = read_sysfs_int(path);
6498 
6499 		sprintf(path, "%s/domain_id", path_base);
6500 		domain_id = read_sysfs_int(path);
6501 
6502 		sprintf(path, "%s/fabric_cluster_id", path_base);
6503 		cluster_id = read_sysfs_int(path);
6504 
6505 		sprintf(path, "%s/current_freq_khz", path_base);
6506 		sprintf(name_buf, "UMHz%d.%d", domain_id, cluster_id);
6507 
6508 		add_counter(0, path, name_buf, 0, SCOPE_PACKAGE, COUNTER_K2M, FORMAT_AVERAGE, 0, package_id);
6509 
6510 		if (quiet)
6511 			continue;
6512 
6513 		sprintf(path, "%s/min_freq_khz", path_base);
6514 		k = read_sysfs_int(path);
6515 		sprintf(path, "%s/max_freq_khz", path_base);
6516 		l = read_sysfs_int(path);
6517 		fprintf(outf, "Uncore Frequency package%d domain%d cluster%d: %d - %d MHz ", package_id, domain_id,
6518 			cluster_id, k / 1000, l / 1000);
6519 
6520 		sprintf(path, "%s/initial_min_freq_khz", path_base);
6521 		k = read_sysfs_int(path);
6522 		sprintf(path, "%s/initial_max_freq_khz", path_base);
6523 		l = read_sysfs_int(path);
6524 		fprintf(outf, "(%d - %d MHz)", k / 1000, l / 1000);
6525 
6526 		sprintf(path, "%s/current_freq_khz", path_base);
6527 		k = read_sysfs_int(path);
6528 		fprintf(outf, " %d MHz\n", k / 1000);
6529 	}
6530 }
6531 
probe_intel_uncore_frequency(void)6532 static void probe_intel_uncore_frequency(void)
6533 {
6534 	if (!genuine_intel)
6535 		return;
6536 
6537 	if (access("/sys/devices/system/cpu/intel_uncore_frequency/uncore00", R_OK) == 0)
6538 		probe_intel_uncore_frequency_cluster();
6539 	else
6540 		probe_intel_uncore_frequency_legacy();
6541 }
6542 
set_graphics_fp(char * path,int idx)6543 static void set_graphics_fp(char *path, int idx)
6544 {
6545 	if (!access(path, R_OK))
6546 		gfx_info[idx].fp = fopen_or_die(path, "r");
6547 }
6548 
6549 /* Enlarge this if there are /sys/class/drm/card2 ... */
6550 #define GFX_MAX_CARDS	2
6551 
probe_graphics(void)6552 static void probe_graphics(void)
6553 {
6554 	char path[PATH_MAX];
6555 	int i;
6556 
6557 	/* Xe graphics sysfs knobs */
6558 	if (!access("/sys/class/drm/card0/device/tile0/gt0/gtidle/idle_residency_ms", R_OK)) {
6559 		FILE *fp;
6560 		char buf[8];
6561 		bool gt0_is_gt;
6562 
6563 		fp = fopen("/sys/class/drm/card0/device/tile0/gt0/gtidle/name", "r");
6564 		if (!fp)
6565 			goto next;
6566 
6567 		if (!fread(buf, sizeof(char), 7, fp)) {
6568 			fclose(fp);
6569 			goto next;
6570 		}
6571 		fclose(fp);
6572 
6573 		if (!strncmp(buf, "gt0-rc", strlen("gt0-rc")))
6574 			gt0_is_gt = true;
6575 		else if (!strncmp(buf, "gt0-mc", strlen("gt0-mc")))
6576 			gt0_is_gt = false;
6577 		else
6578 			goto next;
6579 
6580 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/gtidle/idle_residency_ms", gt0_is_gt ? GFX_rc6 : SAM_mc6);
6581 
6582 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/freq0/cur_freq", gt0_is_gt ? GFX_MHz : SAM_MHz);
6583 
6584 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/freq0/act_freq", gt0_is_gt ? GFX_ACTMHz : SAM_ACTMHz);
6585 
6586 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/gtidle/idle_residency_ms", gt0_is_gt ? SAM_mc6 : GFX_rc6);
6587 
6588 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/freq0/cur_freq", gt0_is_gt ? SAM_MHz : GFX_MHz);
6589 
6590 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/freq0/act_freq", gt0_is_gt ? SAM_ACTMHz : GFX_ACTMHz);
6591 
6592 		goto end;
6593 	}
6594 
6595 next:
6596 	/* New i915 graphics sysfs knobs */
6597 	for (i = 0; i < GFX_MAX_CARDS; i++) {
6598 		snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rc6_residency_ms", i);
6599 		if (!access(path, R_OK))
6600 			break;
6601 	}
6602 
6603 	if (i == GFX_MAX_CARDS)
6604 		goto legacy_i915;
6605 
6606 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rc6_residency_ms", i);
6607 	set_graphics_fp(path, GFX_rc6);
6608 
6609 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rps_cur_freq_mhz", i);
6610 	set_graphics_fp(path, GFX_MHz);
6611 
6612 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rps_act_freq_mhz", i);
6613 	set_graphics_fp(path, GFX_ACTMHz);
6614 
6615 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rc6_residency_ms", i);
6616 	set_graphics_fp(path, SAM_mc6);
6617 
6618 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rps_cur_freq_mhz", i);
6619 	set_graphics_fp(path, SAM_MHz);
6620 
6621 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rps_act_freq_mhz", i);
6622 	set_graphics_fp(path, SAM_ACTMHz);
6623 
6624 	goto end;
6625 
6626 legacy_i915:
6627 	/* Fall back to traditional i915 graphics sysfs knobs */
6628 	set_graphics_fp("/sys/class/drm/card0/power/rc6_residency_ms", GFX_rc6);
6629 
6630 	set_graphics_fp("/sys/class/drm/card0/gt_cur_freq_mhz", GFX_MHz);
6631 	if (!gfx_info[GFX_MHz].fp)
6632 		set_graphics_fp("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", GFX_MHz);
6633 
6634 	set_graphics_fp("/sys/class/drm/card0/gt_act_freq_mhz", GFX_ACTMHz);
6635 	if (!gfx_info[GFX_ACTMHz].fp)
6636 		set_graphics_fp("/sys/class/graphics/fb0/device/drm/card0/gt_act_freq_mhz", GFX_ACTMHz);
6637 
6638 end:
6639 	if (gfx_info[GFX_rc6].fp)
6640 		BIC_PRESENT(BIC_GFX_rc6);
6641 	if (gfx_info[GFX_MHz].fp)
6642 		BIC_PRESENT(BIC_GFXMHz);
6643 	if (gfx_info[GFX_ACTMHz].fp)
6644 		BIC_PRESENT(BIC_GFXACTMHz);
6645 	if (gfx_info[SAM_mc6].fp)
6646 		BIC_PRESENT(BIC_SAM_mc6);
6647 	if (gfx_info[SAM_MHz].fp)
6648 		BIC_PRESENT(BIC_SAMMHz);
6649 	if (gfx_info[SAM_ACTMHz].fp)
6650 		BIC_PRESENT(BIC_SAMACTMHz);
6651 }
6652 
dump_sysfs_cstate_config(void)6653 static void dump_sysfs_cstate_config(void)
6654 {
6655 	char path[64];
6656 	char name_buf[16];
6657 	char desc[64];
6658 	FILE *input;
6659 	int state;
6660 	char *sp;
6661 
6662 	if (access("/sys/devices/system/cpu/cpuidle", R_OK)) {
6663 		fprintf(outf, "cpuidle not loaded\n");
6664 		return;
6665 	}
6666 
6667 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_driver");
6668 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor");
6669 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor_ro");
6670 
6671 	for (state = 0; state < 10; ++state) {
6672 
6673 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", base_cpu, state);
6674 		input = fopen(path, "r");
6675 		if (input == NULL)
6676 			continue;
6677 		if (!fgets(name_buf, sizeof(name_buf), input))
6678 			err(1, "%s: failed to read file", path);
6679 
6680 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
6681 		sp = strchr(name_buf, '-');
6682 		if (!sp)
6683 			sp = strchrnul(name_buf, '\n');
6684 		*sp = '\0';
6685 		fclose(input);
6686 
6687 		remove_underbar(name_buf);
6688 
6689 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/desc", base_cpu, state);
6690 		input = fopen(path, "r");
6691 		if (input == NULL)
6692 			continue;
6693 		if (!fgets(desc, sizeof(desc), input))
6694 			err(1, "%s: failed to read file", path);
6695 
6696 		fprintf(outf, "cpu%d: %s: %s", base_cpu, name_buf, desc);
6697 		fclose(input);
6698 	}
6699 }
6700 
dump_sysfs_pstate_config(void)6701 static void dump_sysfs_pstate_config(void)
6702 {
6703 	char path[64];
6704 	char driver_buf[64];
6705 	char governor_buf[64];
6706 	FILE *input;
6707 	int turbo;
6708 
6709 	sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_driver", base_cpu);
6710 	input = fopen(path, "r");
6711 	if (input == NULL) {
6712 		fprintf(outf, "NSFOD %s\n", path);
6713 		return;
6714 	}
6715 	if (!fgets(driver_buf, sizeof(driver_buf), input))
6716 		err(1, "%s: failed to read file", path);
6717 	fclose(input);
6718 
6719 	sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_governor", base_cpu);
6720 	input = fopen(path, "r");
6721 	if (input == NULL) {
6722 		fprintf(outf, "NSFOD %s\n", path);
6723 		return;
6724 	}
6725 	if (!fgets(governor_buf, sizeof(governor_buf), input))
6726 		err(1, "%s: failed to read file", path);
6727 	fclose(input);
6728 
6729 	fprintf(outf, "cpu%d: cpufreq driver: %s", base_cpu, driver_buf);
6730 	fprintf(outf, "cpu%d: cpufreq governor: %s", base_cpu, governor_buf);
6731 
6732 	sprintf(path, "/sys/devices/system/cpu/cpufreq/boost");
6733 	input = fopen(path, "r");
6734 	if (input != NULL) {
6735 		if (fscanf(input, "%d", &turbo) != 1)
6736 			err(1, "%s: failed to parse number from file", path);
6737 		fprintf(outf, "cpufreq boost: %d\n", turbo);
6738 		fclose(input);
6739 	}
6740 
6741 	sprintf(path, "/sys/devices/system/cpu/intel_pstate/no_turbo");
6742 	input = fopen(path, "r");
6743 	if (input != NULL) {
6744 		if (fscanf(input, "%d", &turbo) != 1)
6745 			err(1, "%s: failed to parse number from file", path);
6746 		fprintf(outf, "cpufreq intel_pstate no_turbo: %d\n", turbo);
6747 		fclose(input);
6748 	}
6749 }
6750 
6751 /*
6752  * print_epb()
6753  * Decode the ENERGY_PERF_BIAS MSR
6754  */
print_epb(struct thread_data * t,struct core_data * c,struct pkg_data * p)6755 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
6756 {
6757 	char *epb_string;
6758 	int cpu, epb;
6759 
6760 	UNUSED(c);
6761 	UNUSED(p);
6762 
6763 	if (!has_epb)
6764 		return 0;
6765 
6766 	cpu = t->cpu_id;
6767 
6768 	/* EPB is per-package */
6769 	if (!is_cpu_first_thread_in_package(t, c, p))
6770 		return 0;
6771 
6772 	if (cpu_migrate(cpu)) {
6773 		fprintf(outf, "print_epb: Could not migrate to CPU %d\n", cpu);
6774 		return -1;
6775 	}
6776 
6777 	epb = get_epb(cpu);
6778 	if (epb < 0)
6779 		return 0;
6780 
6781 	switch (epb) {
6782 	case ENERGY_PERF_BIAS_PERFORMANCE:
6783 		epb_string = "performance";
6784 		break;
6785 	case ENERGY_PERF_BIAS_NORMAL:
6786 		epb_string = "balanced";
6787 		break;
6788 	case ENERGY_PERF_BIAS_POWERSAVE:
6789 		epb_string = "powersave";
6790 		break;
6791 	default:
6792 		epb_string = "custom";
6793 		break;
6794 	}
6795 	fprintf(outf, "cpu%d: EPB: %d (%s)\n", cpu, epb, epb_string);
6796 
6797 	return 0;
6798 }
6799 
6800 /*
6801  * print_hwp()
6802  * Decode the MSR_HWP_CAPABILITIES
6803  */
print_hwp(struct thread_data * t,struct core_data * c,struct pkg_data * p)6804 int print_hwp(struct thread_data *t, struct core_data *c, struct pkg_data *p)
6805 {
6806 	unsigned long long msr;
6807 	int cpu;
6808 
6809 	UNUSED(c);
6810 	UNUSED(p);
6811 
6812 	if (no_msr)
6813 		return 0;
6814 
6815 	if (!has_hwp)
6816 		return 0;
6817 
6818 	cpu = t->cpu_id;
6819 
6820 	/* MSR_HWP_CAPABILITIES is per-package */
6821 	if (!is_cpu_first_thread_in_package(t, c, p))
6822 		return 0;
6823 
6824 	if (cpu_migrate(cpu)) {
6825 		fprintf(outf, "print_hwp: Could not migrate to CPU %d\n", cpu);
6826 		return -1;
6827 	}
6828 
6829 	if (get_msr(cpu, MSR_PM_ENABLE, &msr))
6830 		return 0;
6831 
6832 	fprintf(outf, "cpu%d: MSR_PM_ENABLE: 0x%08llx (%sHWP)\n", cpu, msr, (msr & (1 << 0)) ? "" : "No-");
6833 
6834 	/* MSR_PM_ENABLE[1] == 1 if HWP is enabled and MSRs visible */
6835 	if ((msr & (1 << 0)) == 0)
6836 		return 0;
6837 
6838 	if (get_msr(cpu, MSR_HWP_CAPABILITIES, &msr))
6839 		return 0;
6840 
6841 	fprintf(outf, "cpu%d: MSR_HWP_CAPABILITIES: 0x%08llx "
6842 		"(high %d guar %d eff %d low %d)\n",
6843 		cpu, msr,
6844 		(unsigned int)HWP_HIGHEST_PERF(msr),
6845 		(unsigned int)HWP_GUARANTEED_PERF(msr),
6846 		(unsigned int)HWP_MOSTEFFICIENT_PERF(msr), (unsigned int)HWP_LOWEST_PERF(msr));
6847 
6848 	if (get_msr(cpu, MSR_HWP_REQUEST, &msr))
6849 		return 0;
6850 
6851 	fprintf(outf, "cpu%d: MSR_HWP_REQUEST: 0x%08llx "
6852 		"(min %d max %d des %d epp 0x%x window 0x%x pkg 0x%x)\n",
6853 		cpu, msr,
6854 		(unsigned int)(((msr) >> 0) & 0xff),
6855 		(unsigned int)(((msr) >> 8) & 0xff),
6856 		(unsigned int)(((msr) >> 16) & 0xff),
6857 		(unsigned int)(((msr) >> 24) & 0xff),
6858 		(unsigned int)(((msr) >> 32) & 0xff3), (unsigned int)(((msr) >> 42) & 0x1));
6859 
6860 	if (has_hwp_pkg) {
6861 		if (get_msr(cpu, MSR_HWP_REQUEST_PKG, &msr))
6862 			return 0;
6863 
6864 		fprintf(outf, "cpu%d: MSR_HWP_REQUEST_PKG: 0x%08llx "
6865 			"(min %d max %d des %d epp 0x%x window 0x%x)\n",
6866 			cpu, msr,
6867 			(unsigned int)(((msr) >> 0) & 0xff),
6868 			(unsigned int)(((msr) >> 8) & 0xff),
6869 			(unsigned int)(((msr) >> 16) & 0xff),
6870 			(unsigned int)(((msr) >> 24) & 0xff), (unsigned int)(((msr) >> 32) & 0xff3));
6871 	}
6872 	if (has_hwp_notify) {
6873 		if (get_msr(cpu, MSR_HWP_INTERRUPT, &msr))
6874 			return 0;
6875 
6876 		fprintf(outf, "cpu%d: MSR_HWP_INTERRUPT: 0x%08llx "
6877 			"(%s_Guaranteed_Perf_Change, %s_Excursion_Min)\n",
6878 			cpu, msr, ((msr) & 0x1) ? "EN" : "Dis", ((msr) & 0x2) ? "EN" : "Dis");
6879 	}
6880 	if (get_msr(cpu, MSR_HWP_STATUS, &msr))
6881 		return 0;
6882 
6883 	fprintf(outf, "cpu%d: MSR_HWP_STATUS: 0x%08llx "
6884 		"(%sGuaranteed_Perf_Change, %sExcursion_Min)\n",
6885 		cpu, msr, ((msr) & 0x1) ? "" : "No-", ((msr) & 0x4) ? "" : "No-");
6886 
6887 	return 0;
6888 }
6889 
6890 /*
6891  * print_perf_limit()
6892  */
print_perf_limit(struct thread_data * t,struct core_data * c,struct pkg_data * p)6893 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
6894 {
6895 	unsigned long long msr;
6896 	int cpu;
6897 
6898 	UNUSED(c);
6899 	UNUSED(p);
6900 
6901 	if (no_msr)
6902 		return 0;
6903 
6904 	cpu = t->cpu_id;
6905 
6906 	/* per-package */
6907 	if (!is_cpu_first_thread_in_package(t, c, p))
6908 		return 0;
6909 
6910 	if (cpu_migrate(cpu)) {
6911 		fprintf(outf, "print_perf_limit: Could not migrate to CPU %d\n", cpu);
6912 		return -1;
6913 	}
6914 
6915 	if (platform->plr_msrs & PLR_CORE) {
6916 		get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
6917 		fprintf(outf, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
6918 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
6919 			(msr & 1 << 15) ? "bit15, " : "",
6920 			(msr & 1 << 14) ? "bit14, " : "",
6921 			(msr & 1 << 13) ? "Transitions, " : "",
6922 			(msr & 1 << 12) ? "MultiCoreTurbo, " : "",
6923 			(msr & 1 << 11) ? "PkgPwrL2, " : "",
6924 			(msr & 1 << 10) ? "PkgPwrL1, " : "",
6925 			(msr & 1 << 9) ? "CorePwr, " : "",
6926 			(msr & 1 << 8) ? "Amps, " : "",
6927 			(msr & 1 << 6) ? "VR-Therm, " : "",
6928 			(msr & 1 << 5) ? "Auto-HWP, " : "",
6929 			(msr & 1 << 4) ? "Graphics, " : "",
6930 			(msr & 1 << 2) ? "bit2, " : "",
6931 			(msr & 1 << 1) ? "ThermStatus, " : "", (msr & 1 << 0) ? "PROCHOT, " : "");
6932 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
6933 			(msr & 1 << 31) ? "bit31, " : "",
6934 			(msr & 1 << 30) ? "bit30, " : "",
6935 			(msr & 1 << 29) ? "Transitions, " : "",
6936 			(msr & 1 << 28) ? "MultiCoreTurbo, " : "",
6937 			(msr & 1 << 27) ? "PkgPwrL2, " : "",
6938 			(msr & 1 << 26) ? "PkgPwrL1, " : "",
6939 			(msr & 1 << 25) ? "CorePwr, " : "",
6940 			(msr & 1 << 24) ? "Amps, " : "",
6941 			(msr & 1 << 22) ? "VR-Therm, " : "",
6942 			(msr & 1 << 21) ? "Auto-HWP, " : "",
6943 			(msr & 1 << 20) ? "Graphics, " : "",
6944 			(msr & 1 << 18) ? "bit18, " : "",
6945 			(msr & 1 << 17) ? "ThermStatus, " : "", (msr & 1 << 16) ? "PROCHOT, " : "");
6946 
6947 	}
6948 	if (platform->plr_msrs & PLR_GFX) {
6949 		get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
6950 		fprintf(outf, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
6951 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s)",
6952 			(msr & 1 << 0) ? "PROCHOT, " : "",
6953 			(msr & 1 << 1) ? "ThermStatus, " : "",
6954 			(msr & 1 << 4) ? "Graphics, " : "",
6955 			(msr & 1 << 6) ? "VR-Therm, " : "",
6956 			(msr & 1 << 8) ? "Amps, " : "",
6957 			(msr & 1 << 9) ? "GFXPwr, " : "",
6958 			(msr & 1 << 10) ? "PkgPwrL1, " : "", (msr & 1 << 11) ? "PkgPwrL2, " : "");
6959 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s)\n",
6960 			(msr & 1 << 16) ? "PROCHOT, " : "",
6961 			(msr & 1 << 17) ? "ThermStatus, " : "",
6962 			(msr & 1 << 20) ? "Graphics, " : "",
6963 			(msr & 1 << 22) ? "VR-Therm, " : "",
6964 			(msr & 1 << 24) ? "Amps, " : "",
6965 			(msr & 1 << 25) ? "GFXPwr, " : "",
6966 			(msr & 1 << 26) ? "PkgPwrL1, " : "", (msr & 1 << 27) ? "PkgPwrL2, " : "");
6967 	}
6968 	if (platform->plr_msrs & PLR_RING) {
6969 		get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
6970 		fprintf(outf, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
6971 		fprintf(outf, " (Active: %s%s%s%s%s%s)",
6972 			(msr & 1 << 0) ? "PROCHOT, " : "",
6973 			(msr & 1 << 1) ? "ThermStatus, " : "",
6974 			(msr & 1 << 6) ? "VR-Therm, " : "",
6975 			(msr & 1 << 8) ? "Amps, " : "",
6976 			(msr & 1 << 10) ? "PkgPwrL1, " : "", (msr & 1 << 11) ? "PkgPwrL2, " : "");
6977 		fprintf(outf, " (Logged: %s%s%s%s%s%s)\n",
6978 			(msr & 1 << 16) ? "PROCHOT, " : "",
6979 			(msr & 1 << 17) ? "ThermStatus, " : "",
6980 			(msr & 1 << 22) ? "VR-Therm, " : "",
6981 			(msr & 1 << 24) ? "Amps, " : "",
6982 			(msr & 1 << 26) ? "PkgPwrL1, " : "", (msr & 1 << 27) ? "PkgPwrL2, " : "");
6983 	}
6984 	return 0;
6985 }
6986 
6987 #define	RAPL_POWER_GRANULARITY	0x7FFF	/* 15 bit power granularity */
6988 #define	RAPL_TIME_GRANULARITY	0x3F	/* 6 bit time granularity */
6989 
get_quirk_tdp(void)6990 double get_quirk_tdp(void)
6991 {
6992 	if (platform->rapl_quirk_tdp)
6993 		return platform->rapl_quirk_tdp;
6994 
6995 	return 135.0;
6996 }
6997 
get_tdp_intel(void)6998 double get_tdp_intel(void)
6999 {
7000 	unsigned long long msr;
7001 
7002 	if (platform->rapl_msrs & RAPL_PKG_POWER_INFO)
7003 		if (!get_msr(base_cpu, MSR_PKG_POWER_INFO, &msr))
7004 			return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
7005 	return get_quirk_tdp();
7006 }
7007 
get_tdp_amd(void)7008 double get_tdp_amd(void)
7009 {
7010 	return get_quirk_tdp();
7011 }
7012 
rapl_probe_intel(void)7013 void rapl_probe_intel(void)
7014 {
7015 	unsigned long long msr;
7016 	unsigned int time_unit;
7017 	double tdp;
7018 	const unsigned long long bic_watt_bits = BIC_SysWatt | BIC_PkgWatt | BIC_CorWatt | BIC_RAMWatt | BIC_GFXWatt;
7019 	const unsigned long long bic_joules_bits = BIC_Sys_J | BIC_Pkg_J | BIC_Cor_J | BIC_RAM_J | BIC_GFX_J;
7020 
7021 	if (rapl_joules)
7022 		bic_enabled &= ~bic_watt_bits;
7023 	else
7024 		bic_enabled &= ~bic_joules_bits;
7025 
7026 	if (!(platform->rapl_msrs & RAPL_PKG_PERF_STATUS))
7027 		bic_enabled &= ~BIC_PKG__;
7028 	if (!(platform->rapl_msrs & RAPL_DRAM_PERF_STATUS))
7029 		bic_enabled &= ~BIC_RAM__;
7030 
7031 	/* units on package 0, verify later other packages match */
7032 	if (get_msr(base_cpu, MSR_RAPL_POWER_UNIT, &msr))
7033 		return;
7034 
7035 	rapl_power_units = 1.0 / (1 << (msr & 0xF));
7036 	if (platform->has_rapl_divisor)
7037 		rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
7038 	else
7039 		rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
7040 
7041 	if (platform->has_fixed_rapl_unit)
7042 		rapl_dram_energy_units = (15.3 / 1000000);
7043 	else
7044 		rapl_dram_energy_units = rapl_energy_units;
7045 
7046 	time_unit = msr >> 16 & 0xF;
7047 	if (time_unit == 0)
7048 		time_unit = 0xA;
7049 
7050 	rapl_time_units = 1.0 / (1 << (time_unit));
7051 
7052 	tdp = get_tdp_intel();
7053 
7054 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
7055 	if (!quiet)
7056 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
7057 }
7058 
rapl_probe_amd(void)7059 void rapl_probe_amd(void)
7060 {
7061 	unsigned long long msr;
7062 	double tdp;
7063 	const unsigned long long bic_watt_bits = BIC_PkgWatt | BIC_CorWatt;
7064 	const unsigned long long bic_joules_bits = BIC_Pkg_J | BIC_Cor_J;
7065 
7066 	if (rapl_joules)
7067 		bic_enabled &= ~bic_watt_bits;
7068 	else
7069 		bic_enabled &= ~bic_joules_bits;
7070 
7071 	if (get_msr(base_cpu, MSR_RAPL_PWR_UNIT, &msr))
7072 		return;
7073 
7074 	rapl_time_units = ldexp(1.0, -(msr >> 16 & 0xf));
7075 	rapl_energy_units = ldexp(1.0, -(msr >> 8 & 0x1f));
7076 	rapl_power_units = ldexp(1.0, -(msr & 0xf));
7077 
7078 	tdp = get_tdp_amd();
7079 
7080 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
7081 	if (!quiet)
7082 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
7083 }
7084 
print_power_limit_msr(int cpu,unsigned long long msr,char * label)7085 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
7086 {
7087 	fprintf(outf, "cpu%d: %s: %sabled (%0.3f Watts, %f sec, clamp %sabled)\n",
7088 		cpu, label,
7089 		((msr >> 15) & 1) ? "EN" : "DIS",
7090 		((msr >> 0) & 0x7FFF) * rapl_power_units,
7091 		(1.0 + (((msr >> 22) & 0x3) / 4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
7092 		(((msr >> 16) & 1) ? "EN" : "DIS"));
7093 
7094 	return;
7095 }
7096 
print_rapl(struct thread_data * t,struct core_data * c,struct pkg_data * p)7097 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7098 {
7099 	unsigned long long msr;
7100 	const char *msr_name;
7101 	int cpu;
7102 
7103 	UNUSED(c);
7104 	UNUSED(p);
7105 
7106 	if (!platform->rapl_msrs)
7107 		return 0;
7108 
7109 	/* RAPL counters are per package, so print only for 1st thread/package */
7110 	if (!is_cpu_first_thread_in_package(t, c, p))
7111 		return 0;
7112 
7113 	cpu = t->cpu_id;
7114 	if (cpu_migrate(cpu)) {
7115 		fprintf(outf, "print_rapl: Could not migrate to CPU %d\n", cpu);
7116 		return -1;
7117 	}
7118 
7119 	if (platform->rapl_msrs & RAPL_AMD_F17H) {
7120 		msr_name = "MSR_RAPL_PWR_UNIT";
7121 		if (get_msr(cpu, MSR_RAPL_PWR_UNIT, &msr))
7122 			return -1;
7123 	} else {
7124 		msr_name = "MSR_RAPL_POWER_UNIT";
7125 		if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
7126 			return -1;
7127 	}
7128 
7129 	fprintf(outf, "cpu%d: %s: 0x%08llx (%f Watts, %f Joules, %f sec.)\n", cpu, msr_name, msr,
7130 		rapl_power_units, rapl_energy_units, rapl_time_units);
7131 
7132 	if (platform->rapl_msrs & RAPL_PKG_POWER_INFO) {
7133 
7134 		if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
7135 			return -5;
7136 
7137 		fprintf(outf, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
7138 			cpu, msr,
7139 			((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7140 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7141 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7142 			((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
7143 
7144 	}
7145 	if (platform->rapl_msrs & RAPL_PKG) {
7146 
7147 		if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
7148 			return -9;
7149 
7150 		fprintf(outf, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
7151 			cpu, msr, (msr >> 63) & 1 ? "" : "UN");
7152 
7153 		print_power_limit_msr(cpu, msr, "PKG Limit #1");
7154 		fprintf(outf, "cpu%d: PKG Limit #2: %sabled (%0.3f Watts, %f* sec, clamp %sabled)\n",
7155 			cpu,
7156 			((msr >> 47) & 1) ? "EN" : "DIS",
7157 			((msr >> 32) & 0x7FFF) * rapl_power_units,
7158 			(1.0 + (((msr >> 54) & 0x3) / 4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
7159 			((msr >> 48) & 1) ? "EN" : "DIS");
7160 
7161 		if (get_msr(cpu, MSR_VR_CURRENT_CONFIG, &msr))
7162 			return -9;
7163 
7164 		fprintf(outf, "cpu%d: MSR_VR_CURRENT_CONFIG: 0x%08llx\n", cpu, msr);
7165 		fprintf(outf, "cpu%d: PKG Limit #4: %f Watts (%slocked)\n",
7166 			cpu, ((msr >> 0) & 0x1FFF) * rapl_power_units, (msr >> 31) & 1 ? "" : "UN");
7167 	}
7168 
7169 	if (platform->rapl_msrs & RAPL_DRAM_POWER_INFO) {
7170 		if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
7171 			return -6;
7172 
7173 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
7174 			cpu, msr,
7175 			((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7176 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7177 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
7178 			((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
7179 	}
7180 	if (platform->rapl_msrs & RAPL_DRAM) {
7181 		if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
7182 			return -9;
7183 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
7184 			cpu, msr, (msr >> 31) & 1 ? "" : "UN");
7185 
7186 		print_power_limit_msr(cpu, msr, "DRAM Limit");
7187 	}
7188 	if (platform->rapl_msrs & RAPL_CORE_POLICY) {
7189 		if (get_msr(cpu, MSR_PP0_POLICY, &msr))
7190 			return -7;
7191 
7192 		fprintf(outf, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
7193 	}
7194 	if (platform->rapl_msrs & RAPL_CORE_POWER_LIMIT) {
7195 		if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
7196 			return -9;
7197 		fprintf(outf, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
7198 			cpu, msr, (msr >> 31) & 1 ? "" : "UN");
7199 		print_power_limit_msr(cpu, msr, "Cores Limit");
7200 	}
7201 	if (platform->rapl_msrs & RAPL_GFX) {
7202 		if (get_msr(cpu, MSR_PP1_POLICY, &msr))
7203 			return -8;
7204 
7205 		fprintf(outf, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
7206 
7207 		if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
7208 			return -9;
7209 		fprintf(outf, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
7210 			cpu, msr, (msr >> 31) & 1 ? "" : "UN");
7211 		print_power_limit_msr(cpu, msr, "GFX Limit");
7212 	}
7213 	return 0;
7214 }
7215 
7216 /*
7217  * probe_rapl()
7218  *
7219  * sets rapl_power_units, rapl_energy_units, rapl_time_units
7220  */
probe_rapl(void)7221 void probe_rapl(void)
7222 {
7223 	if (!platform->rapl_msrs || no_msr)
7224 		return;
7225 
7226 	if (genuine_intel)
7227 		rapl_probe_intel();
7228 	if (authentic_amd || hygon_genuine)
7229 		rapl_probe_amd();
7230 
7231 	if (quiet)
7232 		return;
7233 
7234 	for_all_cpus(print_rapl, ODD_COUNTERS);
7235 }
7236 
7237 /*
7238  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
7239  * the Thermal Control Circuit (TCC) activates.
7240  * This is usually equal to tjMax.
7241  *
7242  * Older processors do not have this MSR, so there we guess,
7243  * but also allow cmdline over-ride with -T.
7244  *
7245  * Several MSR temperature values are in units of degrees-C
7246  * below this value, including the Digital Thermal Sensor (DTS),
7247  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
7248  */
set_temperature_target(struct thread_data * t,struct core_data * c,struct pkg_data * p)7249 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7250 {
7251 	unsigned long long msr;
7252 	unsigned int tcc_default, tcc_offset;
7253 	int cpu;
7254 
7255 	UNUSED(c);
7256 	UNUSED(p);
7257 
7258 	/* tj_max is used only for dts or ptm */
7259 	if (!(do_dts || do_ptm))
7260 		return 0;
7261 
7262 	/* this is a per-package concept */
7263 	if (!is_cpu_first_thread_in_package(t, c, p))
7264 		return 0;
7265 
7266 	cpu = t->cpu_id;
7267 	if (cpu_migrate(cpu)) {
7268 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
7269 		return -1;
7270 	}
7271 
7272 	if (tj_max_override != 0) {
7273 		tj_max = tj_max_override;
7274 		fprintf(outf, "cpu%d: Using cmdline TCC Target (%d C)\n", cpu, tj_max);
7275 		return 0;
7276 	}
7277 
7278 	/* Temperature Target MSR is Nehalem and newer only */
7279 	if (!platform->has_nhm_msrs || no_msr)
7280 		goto guess;
7281 
7282 	if (get_msr(base_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
7283 		goto guess;
7284 
7285 	tcc_default = (msr >> 16) & 0xFF;
7286 
7287 	if (!quiet) {
7288 		int bits = platform->tcc_offset_bits;
7289 		unsigned long long enabled = 0;
7290 
7291 		if (bits && !get_msr(base_cpu, MSR_PLATFORM_INFO, &enabled))
7292 			enabled = (enabled >> 30) & 1;
7293 
7294 		if (bits && enabled) {
7295 			tcc_offset = (msr >> 24) & GENMASK(bits - 1, 0);
7296 			fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C) (%d default - %d offset)\n",
7297 				cpu, msr, tcc_default - tcc_offset, tcc_default, tcc_offset);
7298 		} else {
7299 			fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n", cpu, msr, tcc_default);
7300 		}
7301 	}
7302 
7303 	if (!tcc_default)
7304 		goto guess;
7305 
7306 	tj_max = tcc_default;
7307 
7308 	return 0;
7309 
7310 guess:
7311 	tj_max = TJMAX_DEFAULT;
7312 	fprintf(outf, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n", cpu, tj_max);
7313 
7314 	return 0;
7315 }
7316 
print_thermal(struct thread_data * t,struct core_data * c,struct pkg_data * p)7317 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7318 {
7319 	unsigned long long msr;
7320 	unsigned int dts, dts2;
7321 	int cpu;
7322 
7323 	UNUSED(c);
7324 	UNUSED(p);
7325 
7326 	if (no_msr)
7327 		return 0;
7328 
7329 	if (!(do_dts || do_ptm))
7330 		return 0;
7331 
7332 	cpu = t->cpu_id;
7333 
7334 	/* DTS is per-core, no need to print for each thread */
7335 	if (!is_cpu_first_thread_in_core(t, c, p))
7336 		return 0;
7337 
7338 	if (cpu_migrate(cpu)) {
7339 		fprintf(outf, "print_thermal: Could not migrate to CPU %d\n", cpu);
7340 		return -1;
7341 	}
7342 
7343 	if (do_ptm && is_cpu_first_core_in_package(t, c, p)) {
7344 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
7345 			return 0;
7346 
7347 		dts = (msr >> 16) & 0x7F;
7348 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n", cpu, msr, tj_max - dts);
7349 
7350 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
7351 			return 0;
7352 
7353 		dts = (msr >> 16) & 0x7F;
7354 		dts2 = (msr >> 8) & 0x7F;
7355 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
7356 			cpu, msr, tj_max - dts, tj_max - dts2);
7357 	}
7358 
7359 	if (do_dts && debug) {
7360 		unsigned int resolution;
7361 
7362 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
7363 			return 0;
7364 
7365 		dts = (msr >> 16) & 0x7F;
7366 		resolution = (msr >> 27) & 0xF;
7367 		fprintf(outf, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
7368 			cpu, msr, tj_max - dts, resolution);
7369 
7370 		if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
7371 			return 0;
7372 
7373 		dts = (msr >> 16) & 0x7F;
7374 		dts2 = (msr >> 8) & 0x7F;
7375 		fprintf(outf, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
7376 			cpu, msr, tj_max - dts, tj_max - dts2);
7377 	}
7378 
7379 	return 0;
7380 }
7381 
probe_thermal(void)7382 void probe_thermal(void)
7383 {
7384 	if (!access("/sys/devices/system/cpu/cpu0/thermal_throttle/core_throttle_count", R_OK))
7385 		BIC_PRESENT(BIC_CORE_THROT_CNT);
7386 	else
7387 		BIC_NOT_PRESENT(BIC_CORE_THROT_CNT);
7388 
7389 	for_all_cpus(set_temperature_target, ODD_COUNTERS);
7390 
7391 	if (quiet)
7392 		return;
7393 
7394 	for_all_cpus(print_thermal, ODD_COUNTERS);
7395 }
7396 
get_cpu_type(struct thread_data * t,struct core_data * c,struct pkg_data * p)7397 int get_cpu_type(struct thread_data *t, struct core_data *c, struct pkg_data *p)
7398 {
7399 	unsigned int eax, ebx, ecx, edx;
7400 
7401 	UNUSED(c);
7402 	UNUSED(p);
7403 
7404 	if (!genuine_intel)
7405 		return 0;
7406 
7407 	if (cpu_migrate(t->cpu_id)) {
7408 		fprintf(outf, "Could not migrate to CPU %d\n", t->cpu_id);
7409 		return -1;
7410 	}
7411 
7412 	if (max_level < 0x1a)
7413 		return 0;
7414 
7415 	__cpuid(0x1a, eax, ebx, ecx, edx);
7416 	eax = (eax >> 24) & 0xFF;
7417 	if (eax == 0x20)
7418 		t->is_atom = true;
7419 	return 0;
7420 }
7421 
decode_feature_control_msr(void)7422 void decode_feature_control_msr(void)
7423 {
7424 	unsigned long long msr;
7425 
7426 	if (no_msr)
7427 		return;
7428 
7429 	if (!get_msr(base_cpu, MSR_IA32_FEAT_CTL, &msr))
7430 		fprintf(outf, "cpu%d: MSR_IA32_FEATURE_CONTROL: 0x%08llx (%sLocked %s)\n",
7431 			base_cpu, msr, msr & FEAT_CTL_LOCKED ? "" : "UN-", msr & (1 << 18) ? "SGX" : "");
7432 }
7433 
decode_misc_enable_msr(void)7434 void decode_misc_enable_msr(void)
7435 {
7436 	unsigned long long msr;
7437 
7438 	if (no_msr)
7439 		return;
7440 
7441 	if (!genuine_intel)
7442 		return;
7443 
7444 	if (!get_msr(base_cpu, MSR_IA32_MISC_ENABLE, &msr))
7445 		fprintf(outf, "cpu%d: MSR_IA32_MISC_ENABLE: 0x%08llx (%sTCC %sEIST %sMWAIT %sPREFETCH %sTURBO)\n",
7446 			base_cpu, msr,
7447 			msr & MSR_IA32_MISC_ENABLE_TM1 ? "" : "No-",
7448 			msr & MSR_IA32_MISC_ENABLE_ENHANCED_SPEEDSTEP ? "" : "No-",
7449 			msr & MSR_IA32_MISC_ENABLE_MWAIT ? "" : "No-",
7450 			msr & MSR_IA32_MISC_ENABLE_PREFETCH_DISABLE ? "No-" : "",
7451 			msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ? "No-" : "");
7452 }
7453 
decode_misc_feature_control(void)7454 void decode_misc_feature_control(void)
7455 {
7456 	unsigned long long msr;
7457 
7458 	if (no_msr)
7459 		return;
7460 
7461 	if (!platform->has_msr_misc_feature_control)
7462 		return;
7463 
7464 	if (!get_msr(base_cpu, MSR_MISC_FEATURE_CONTROL, &msr))
7465 		fprintf(outf,
7466 			"cpu%d: MSR_MISC_FEATURE_CONTROL: 0x%08llx (%sL2-Prefetch %sL2-Prefetch-pair %sL1-Prefetch %sL1-IP-Prefetch)\n",
7467 			base_cpu, msr, msr & (0 << 0) ? "No-" : "", msr & (1 << 0) ? "No-" : "",
7468 			msr & (2 << 0) ? "No-" : "", msr & (3 << 0) ? "No-" : "");
7469 }
7470 
7471 /*
7472  * Decode MSR_MISC_PWR_MGMT
7473  *
7474  * Decode the bits according to the Nehalem documentation
7475  * bit[0] seems to continue to have same meaning going forward
7476  * bit[1] less so...
7477  */
decode_misc_pwr_mgmt_msr(void)7478 void decode_misc_pwr_mgmt_msr(void)
7479 {
7480 	unsigned long long msr;
7481 
7482 	if (no_msr)
7483 		return;
7484 
7485 	if (!platform->has_msr_misc_pwr_mgmt)
7486 		return;
7487 
7488 	if (!get_msr(base_cpu, MSR_MISC_PWR_MGMT, &msr))
7489 		fprintf(outf, "cpu%d: MSR_MISC_PWR_MGMT: 0x%08llx (%sable-EIST_Coordination %sable-EPB %sable-OOB)\n",
7490 			base_cpu, msr,
7491 			msr & (1 << 0) ? "DIS" : "EN", msr & (1 << 1) ? "EN" : "DIS", msr & (1 << 8) ? "EN" : "DIS");
7492 }
7493 
7494 /*
7495  * Decode MSR_CC6_DEMOTION_POLICY_CONFIG, MSR_MC6_DEMOTION_POLICY_CONFIG
7496  *
7497  * This MSRs are present on Silvermont processors,
7498  * Intel Atom processor E3000 series (Baytrail), and friends.
7499  */
decode_c6_demotion_policy_msr(void)7500 void decode_c6_demotion_policy_msr(void)
7501 {
7502 	unsigned long long msr;
7503 
7504 	if (no_msr)
7505 		return;
7506 
7507 	if (!platform->has_msr_c6_demotion_policy_config)
7508 		return;
7509 
7510 	if (!get_msr(base_cpu, MSR_CC6_DEMOTION_POLICY_CONFIG, &msr))
7511 		fprintf(outf, "cpu%d: MSR_CC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-CC6-Demotion)\n",
7512 			base_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
7513 
7514 	if (!get_msr(base_cpu, MSR_MC6_DEMOTION_POLICY_CONFIG, &msr))
7515 		fprintf(outf, "cpu%d: MSR_MC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-MC6-Demotion)\n",
7516 			base_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
7517 }
7518 
print_dev_latency(void)7519 void print_dev_latency(void)
7520 {
7521 	char *path = "/dev/cpu_dma_latency";
7522 	int fd;
7523 	int value;
7524 	int retval;
7525 
7526 	fd = open(path, O_RDONLY);
7527 	if (fd < 0) {
7528 		if (debug)
7529 			warnx("Read %s failed", path);
7530 		return;
7531 	}
7532 
7533 	retval = read(fd, (void *)&value, sizeof(int));
7534 	if (retval != sizeof(int)) {
7535 		warn("read failed %s", path);
7536 		close(fd);
7537 		return;
7538 	}
7539 	fprintf(outf, "/dev/cpu_dma_latency: %d usec (%s)\n", value, value == 2000000000 ? "default" : "constrained");
7540 
7541 	close(fd);
7542 }
7543 
has_instr_count_access(void)7544 static int has_instr_count_access(void)
7545 {
7546 	int fd;
7547 	int has_access;
7548 
7549 	if (no_perf)
7550 		return 0;
7551 
7552 	fd = open_perf_counter(base_cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, -1, 0);
7553 	has_access = fd != -1;
7554 
7555 	if (fd != -1)
7556 		close(fd);
7557 
7558 	if (!has_access)
7559 		warnx("Failed to access %s. Some of the counters may not be available\n"
7560 		      "\tRun as root to enable them or use %s to disable the access explicitly",
7561 		      "instructions retired perf counter", "--no-perf");
7562 
7563 	return has_access;
7564 }
7565 
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_)7566 int add_rapl_perf_counter_(int cpu, struct rapl_counter_info_t *rci, const struct rapl_counter_arch_info *cai,
7567 			   double *scale_, enum rapl_unit *unit_)
7568 {
7569 	if (no_perf)
7570 		return -1;
7571 
7572 	const double scale = read_perf_scale(cai->perf_subsys, cai->perf_name);
7573 
7574 	if (scale == 0.0)
7575 		return -1;
7576 
7577 	const enum rapl_unit unit = read_perf_rapl_unit(cai->perf_subsys, cai->perf_name);
7578 
7579 	if (unit == RAPL_UNIT_INVALID)
7580 		return -1;
7581 
7582 	const unsigned int rapl_type = read_perf_type(cai->perf_subsys);
7583 	const unsigned int rapl_energy_pkg_config = read_perf_config(cai->perf_subsys, cai->perf_name);
7584 
7585 	const int fd_counter =
7586 	    open_perf_counter(cpu, rapl_type, rapl_energy_pkg_config, rci->fd_perf, PERF_FORMAT_GROUP);
7587 	if (fd_counter == -1)
7588 		return -1;
7589 
7590 	/* If it's the first counter opened, make it a group descriptor */
7591 	if (rci->fd_perf == -1)
7592 		rci->fd_perf = fd_counter;
7593 
7594 	*scale_ = scale;
7595 	*unit_ = unit;
7596 	return fd_counter;
7597 }
7598 
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)7599 int add_rapl_perf_counter(int cpu, struct rapl_counter_info_t *rci, const struct rapl_counter_arch_info *cai,
7600 			  double *scale, enum rapl_unit *unit)
7601 {
7602 	int ret = add_rapl_perf_counter_(cpu, rci, cai, scale, unit);
7603 
7604 	if (debug >= 2)
7605 		fprintf(stderr, "%s: %d (cpu: %d)\n", __func__, ret, cpu);
7606 
7607 	return ret;
7608 }
7609 
7610 /*
7611  * Linux-perf manages the HW instructions-retired counter
7612  * by enabling when requested, and hiding rollover
7613  */
linux_perf_init(void)7614 void linux_perf_init(void)
7615 {
7616 	if (access("/proc/sys/kernel/perf_event_paranoid", F_OK))
7617 		return;
7618 
7619 	if (BIC_IS_ENABLED(BIC_IPC) && has_aperf) {
7620 		fd_instr_count_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
7621 		if (fd_instr_count_percpu == NULL)
7622 			err(-1, "calloc fd_instr_count_percpu");
7623 	}
7624 }
7625 
rapl_perf_init(void)7626 void rapl_perf_init(void)
7627 {
7628 	const unsigned int num_domains = (platform->has_per_core_rapl ? topo.max_core_id : topo.max_package_id) + 1;
7629 	bool *domain_visited = calloc(num_domains, sizeof(bool));
7630 
7631 	rapl_counter_info_perdomain = calloc(num_domains, sizeof(*rapl_counter_info_perdomain));
7632 	if (rapl_counter_info_perdomain == NULL)
7633 		err(-1, "calloc rapl_counter_info_percpu");
7634 	rapl_counter_info_perdomain_size = num_domains;
7635 
7636 	/*
7637 	 * Initialize rapl_counter_info_percpu
7638 	 */
7639 	for (unsigned int domain_id = 0; domain_id < num_domains; ++domain_id) {
7640 		struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[domain_id];
7641 
7642 		rci->fd_perf = -1;
7643 		for (size_t i = 0; i < NUM_RAPL_COUNTERS; ++i) {
7644 			rci->data[i] = 0;
7645 			rci->source[i] = COUNTER_SOURCE_NONE;
7646 		}
7647 	}
7648 
7649 	/*
7650 	 * Open/probe the counters
7651 	 * If can't get it via perf, fallback to MSR
7652 	 */
7653 	for (size_t i = 0; i < ARRAY_SIZE(rapl_counter_arch_infos); ++i) {
7654 
7655 		const struct rapl_counter_arch_info *const cai = &rapl_counter_arch_infos[i];
7656 		bool has_counter = 0;
7657 		double scale;
7658 		enum rapl_unit unit;
7659 		unsigned int next_domain;
7660 
7661 		memset(domain_visited, 0, num_domains * sizeof(*domain_visited));
7662 
7663 		for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
7664 
7665 			if (cpu_is_not_allowed(cpu))
7666 				continue;
7667 
7668 			/* Skip already seen and handled RAPL domains */
7669 			next_domain =
7670 			    platform->has_per_core_rapl ? cpus[cpu].physical_core_id : cpus[cpu].physical_package_id;
7671 
7672 			assert(next_domain < num_domains);
7673 
7674 			if (domain_visited[next_domain])
7675 				continue;
7676 
7677 			domain_visited[next_domain] = 1;
7678 
7679 			if ((cai->flags & RAPL_COUNTER_FLAG_PLATFORM_COUNTER) && (cpu != base_cpu))
7680 				continue;
7681 
7682 			struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[next_domain];
7683 
7684 			/* Check if the counter is enabled and accessible */
7685 			if (BIC_IS_ENABLED(cai->bic) && (platform->rapl_msrs & cai->feature_mask)) {
7686 
7687 				/* Use perf API for this counter */
7688 				if (!no_perf && cai->perf_name
7689 				    && add_rapl_perf_counter(cpu, rci, cai, &scale, &unit) != -1) {
7690 					rci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
7691 					rci->scale[cai->rci_index] = scale * cai->compat_scale;
7692 					rci->unit[cai->rci_index] = unit;
7693 					rci->flags[cai->rci_index] = cai->flags;
7694 
7695 					/* Use MSR for this counter */
7696 				} else if (!no_msr && cai->msr && probe_msr(cpu, cai->msr) == 0) {
7697 					rci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
7698 					rci->msr[cai->rci_index] = cai->msr;
7699 					rci->msr_mask[cai->rci_index] = cai->msr_mask;
7700 					rci->msr_shift[cai->rci_index] = cai->msr_shift;
7701 					rci->unit[cai->rci_index] = RAPL_UNIT_JOULES;
7702 					rci->scale[cai->rci_index] = *cai->platform_rapl_msr_scale * cai->compat_scale;
7703 					rci->flags[cai->rci_index] = cai->flags;
7704 				}
7705 			}
7706 
7707 			if (rci->source[cai->rci_index] != COUNTER_SOURCE_NONE)
7708 				has_counter = 1;
7709 		}
7710 
7711 		/* If any CPU has access to the counter, make it present */
7712 		if (has_counter)
7713 			BIC_PRESENT(cai->bic);
7714 	}
7715 
7716 	free(domain_visited);
7717 }
7718 
7719 /* Assumes msr_counter_info is populated */
has_amperf_access(void)7720 static int has_amperf_access(void)
7721 {
7722 	return msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].present &&
7723 	    msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].present;
7724 }
7725 
get_cstate_perf_group_fd(struct cstate_counter_info_t * cci,const char * group_name)7726 int *get_cstate_perf_group_fd(struct cstate_counter_info_t *cci, const char *group_name)
7727 {
7728 	if (strcmp(group_name, "cstate_core") == 0)
7729 		return &cci->fd_perf_core;
7730 
7731 	if (strcmp(group_name, "cstate_pkg") == 0)
7732 		return &cci->fd_perf_pkg;
7733 
7734 	return NULL;
7735 }
7736 
add_cstate_perf_counter_(int cpu,struct cstate_counter_info_t * cci,const struct cstate_counter_arch_info * cai)7737 int add_cstate_perf_counter_(int cpu, struct cstate_counter_info_t *cci, const struct cstate_counter_arch_info *cai)
7738 {
7739 	if (no_perf)
7740 		return -1;
7741 
7742 	int *pfd_group = get_cstate_perf_group_fd(cci, cai->perf_subsys);
7743 
7744 	if (pfd_group == NULL)
7745 		return -1;
7746 
7747 	const unsigned int type = read_perf_type(cai->perf_subsys);
7748 	const unsigned int config = read_perf_config(cai->perf_subsys, cai->perf_name);
7749 
7750 	const int fd_counter = open_perf_counter(cpu, type, config, *pfd_group, PERF_FORMAT_GROUP);
7751 
7752 	if (fd_counter == -1)
7753 		return -1;
7754 
7755 	/* If it's the first counter opened, make it a group descriptor */
7756 	if (*pfd_group == -1)
7757 		*pfd_group = fd_counter;
7758 
7759 	return fd_counter;
7760 }
7761 
add_cstate_perf_counter(int cpu,struct cstate_counter_info_t * cci,const struct cstate_counter_arch_info * cai)7762 int add_cstate_perf_counter(int cpu, struct cstate_counter_info_t *cci, const struct cstate_counter_arch_info *cai)
7763 {
7764 	int ret = add_cstate_perf_counter_(cpu, cci, cai);
7765 
7766 	if (debug >= 2)
7767 		fprintf(stderr, "%s: %d (cpu: %d)\n", __func__, ret, cpu);
7768 
7769 	return ret;
7770 }
7771 
add_msr_perf_counter_(int cpu,struct msr_counter_info_t * cci,const struct msr_counter_arch_info * cai)7772 int add_msr_perf_counter_(int cpu, struct msr_counter_info_t *cci, const struct msr_counter_arch_info *cai)
7773 {
7774 	if (no_perf)
7775 		return -1;
7776 
7777 	const unsigned int type = read_perf_type(cai->perf_subsys);
7778 	const unsigned int config = read_perf_config(cai->perf_subsys, cai->perf_name);
7779 
7780 	const int fd_counter = open_perf_counter(cpu, type, config, cci->fd_perf, PERF_FORMAT_GROUP);
7781 
7782 	if (fd_counter == -1)
7783 		return -1;
7784 
7785 	/* If it's the first counter opened, make it a group descriptor */
7786 	if (cci->fd_perf == -1)
7787 		cci->fd_perf = fd_counter;
7788 
7789 	return fd_counter;
7790 }
7791 
add_msr_perf_counter(int cpu,struct msr_counter_info_t * cci,const struct msr_counter_arch_info * cai)7792 int add_msr_perf_counter(int cpu, struct msr_counter_info_t *cci, const struct msr_counter_arch_info *cai)
7793 {
7794 	int ret = add_msr_perf_counter_(cpu, cci, cai);
7795 
7796 	if (debug)
7797 		fprintf(stderr, "%s: %s/%s: %d (cpu: %d)\n", __func__, cai->perf_subsys, cai->perf_name, ret, cpu);
7798 
7799 	return ret;
7800 }
7801 
msr_perf_init_(void)7802 void msr_perf_init_(void)
7803 {
7804 	const int mci_num = topo.max_cpu_num + 1;
7805 
7806 	msr_counter_info = calloc(mci_num, sizeof(*msr_counter_info));
7807 	if (!msr_counter_info)
7808 		err(1, "calloc msr_counter_info");
7809 	msr_counter_info_size = mci_num;
7810 
7811 	for (int cpu = 0; cpu < mci_num; ++cpu)
7812 		msr_counter_info[cpu].fd_perf = -1;
7813 
7814 	for (int cidx = 0; cidx < NUM_MSR_COUNTERS; ++cidx) {
7815 
7816 		struct msr_counter_arch_info *cai = &msr_counter_arch_infos[cidx];
7817 
7818 		cai->present = false;
7819 
7820 		for (int cpu = 0; cpu < mci_num; ++cpu) {
7821 
7822 			struct msr_counter_info_t *const cci = &msr_counter_info[cpu];
7823 
7824 			if (cpu_is_not_allowed(cpu))
7825 				continue;
7826 
7827 			if (cai->needed) {
7828 				/* Use perf API for this counter */
7829 				if (!no_perf && cai->perf_name && add_msr_perf_counter(cpu, cci, cai) != -1) {
7830 					cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
7831 					cai->present = true;
7832 
7833 					/* User MSR for this counter */
7834 				} else if (!no_msr && cai->msr && probe_msr(cpu, cai->msr) == 0) {
7835 					cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
7836 					cci->msr[cai->rci_index] = cai->msr;
7837 					cci->msr_mask[cai->rci_index] = cai->msr_mask;
7838 					cai->present = true;
7839 				}
7840 			}
7841 		}
7842 	}
7843 }
7844 
7845 /* Initialize data for reading perf counters from the MSR group. */
msr_perf_init(void)7846 void msr_perf_init(void)
7847 {
7848 	bool need_amperf = false, need_smi = false;
7849 	const bool need_soft_c1 = (!platform->has_msr_core_c1_res) && (platform->supported_cstates & CC1);
7850 
7851 	need_amperf = BIC_IS_ENABLED(BIC_Avg_MHz) || BIC_IS_ENABLED(BIC_Busy) || BIC_IS_ENABLED(BIC_Bzy_MHz)
7852 	    || BIC_IS_ENABLED(BIC_IPC) || need_soft_c1;
7853 
7854 	if (BIC_IS_ENABLED(BIC_SMI))
7855 		need_smi = true;
7856 
7857 	/* Enable needed counters */
7858 	msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].needed = need_amperf;
7859 	msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].needed = need_amperf;
7860 	msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].needed = need_smi;
7861 
7862 	msr_perf_init_();
7863 
7864 	const bool has_amperf = has_amperf_access();
7865 	const bool has_smi = msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].present;
7866 
7867 	has_aperf_access = has_amperf;
7868 
7869 	if (has_amperf) {
7870 		BIC_PRESENT(BIC_Avg_MHz);
7871 		BIC_PRESENT(BIC_Busy);
7872 		BIC_PRESENT(BIC_Bzy_MHz);
7873 		BIC_PRESENT(BIC_SMI);
7874 	}
7875 
7876 	if (has_smi)
7877 		BIC_PRESENT(BIC_SMI);
7878 }
7879 
cstate_perf_init_(bool soft_c1)7880 void cstate_perf_init_(bool soft_c1)
7881 {
7882 	bool has_counter;
7883 	bool *cores_visited = NULL, *pkg_visited = NULL;
7884 	const int cores_visited_elems = topo.max_core_id + 1;
7885 	const int pkg_visited_elems = topo.max_package_id + 1;
7886 	const int cci_num = topo.max_cpu_num + 1;
7887 
7888 	ccstate_counter_info = calloc(cci_num, sizeof(*ccstate_counter_info));
7889 	if (!ccstate_counter_info)
7890 		err(1, "calloc ccstate_counter_arch_info");
7891 	ccstate_counter_info_size = cci_num;
7892 
7893 	cores_visited = calloc(cores_visited_elems, sizeof(*cores_visited));
7894 	if (!cores_visited)
7895 		err(1, "calloc cores_visited");
7896 
7897 	pkg_visited = calloc(pkg_visited_elems, sizeof(*pkg_visited));
7898 	if (!pkg_visited)
7899 		err(1, "calloc pkg_visited");
7900 
7901 	/* Initialize cstate_counter_info_percpu */
7902 	for (int cpu = 0; cpu < cci_num; ++cpu) {
7903 		ccstate_counter_info[cpu].fd_perf_core = -1;
7904 		ccstate_counter_info[cpu].fd_perf_pkg = -1;
7905 	}
7906 
7907 	for (int cidx = 0; cidx < NUM_CSTATE_COUNTERS; ++cidx) {
7908 		has_counter = false;
7909 		memset(cores_visited, 0, cores_visited_elems * sizeof(*cores_visited));
7910 		memset(pkg_visited, 0, pkg_visited_elems * sizeof(*pkg_visited));
7911 
7912 		const struct cstate_counter_arch_info *cai = &ccstate_counter_arch_infos[cidx];
7913 
7914 		for (int cpu = 0; cpu < cci_num; ++cpu) {
7915 
7916 			struct cstate_counter_info_t *const cci = &ccstate_counter_info[cpu];
7917 
7918 			if (cpu_is_not_allowed(cpu))
7919 				continue;
7920 
7921 			const int core_id = cpus[cpu].physical_core_id;
7922 			const int pkg_id = cpus[cpu].physical_package_id;
7923 
7924 			assert(core_id < cores_visited_elems);
7925 			assert(pkg_id < pkg_visited_elems);
7926 
7927 			const bool per_thread = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD;
7928 			const bool per_core = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_CORE;
7929 
7930 			if (!per_thread && cores_visited[core_id])
7931 				continue;
7932 
7933 			if (!per_core && pkg_visited[pkg_id])
7934 				continue;
7935 
7936 			const bool counter_needed = BIC_IS_ENABLED(cai->bic) ||
7937 			    (soft_c1 && (cai->flags & CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY));
7938 			const bool counter_supported = (platform->supported_cstates & cai->feature_mask);
7939 
7940 			if (counter_needed && counter_supported) {
7941 				/* Use perf API for this counter */
7942 				if (!no_perf && cai->perf_name && add_cstate_perf_counter(cpu, cci, cai) != -1) {
7943 
7944 					cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
7945 
7946 					/* User MSR for this counter */
7947 				} else if (!no_msr && cai->msr && pkg_cstate_limit >= cai->pkg_cstate_limit
7948 					   && probe_msr(cpu, cai->msr) == 0) {
7949 					cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
7950 					cci->msr[cai->rci_index] = cai->msr;
7951 				}
7952 			}
7953 
7954 			if (cci->source[cai->rci_index] != COUNTER_SOURCE_NONE) {
7955 				has_counter = true;
7956 				cores_visited[core_id] = true;
7957 				pkg_visited[pkg_id] = true;
7958 			}
7959 		}
7960 
7961 		/* If any CPU has access to the counter, make it present */
7962 		if (has_counter)
7963 			BIC_PRESENT(cai->bic);
7964 	}
7965 
7966 	free(cores_visited);
7967 	free(pkg_visited);
7968 }
7969 
cstate_perf_init(void)7970 void cstate_perf_init(void)
7971 {
7972 	/*
7973 	 * If we don't have a C1 residency MSR, we calculate it "in software",
7974 	 * but we need APERF, MPERF too.
7975 	 */
7976 	const bool soft_c1 = !platform->has_msr_core_c1_res && has_amperf_access()
7977 	    && platform->supported_cstates & CC1;
7978 
7979 	if (soft_c1)
7980 		BIC_PRESENT(BIC_CPU_c1);
7981 
7982 	cstate_perf_init_(soft_c1);
7983 }
7984 
probe_cstates(void)7985 void probe_cstates(void)
7986 {
7987 	probe_cst_limit();
7988 
7989 	if (platform->has_msr_module_c6_res_ms)
7990 		BIC_PRESENT(BIC_Mod_c6);
7991 
7992 	if (platform->has_ext_cst_msrs && !no_msr) {
7993 		BIC_PRESENT(BIC_Totl_c0);
7994 		BIC_PRESENT(BIC_Any_c0);
7995 		BIC_PRESENT(BIC_GFX_c0);
7996 		BIC_PRESENT(BIC_CPUGFX);
7997 	}
7998 
7999 	if (quiet)
8000 		return;
8001 
8002 	dump_power_ctl();
8003 	dump_cst_cfg();
8004 	decode_c6_demotion_policy_msr();
8005 	print_dev_latency();
8006 	dump_sysfs_cstate_config();
8007 	print_irtl();
8008 }
8009 
probe_lpi(void)8010 void probe_lpi(void)
8011 {
8012 	if (!access("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", R_OK))
8013 		BIC_PRESENT(BIC_CPU_LPI);
8014 	else
8015 		BIC_NOT_PRESENT(BIC_CPU_LPI);
8016 
8017 	if (!access(sys_lpi_file_sysfs, R_OK)) {
8018 		sys_lpi_file = sys_lpi_file_sysfs;
8019 		BIC_PRESENT(BIC_SYS_LPI);
8020 	} else if (!access(sys_lpi_file_debugfs, R_OK)) {
8021 		sys_lpi_file = sys_lpi_file_debugfs;
8022 		BIC_PRESENT(BIC_SYS_LPI);
8023 	} else {
8024 		sys_lpi_file_sysfs = NULL;
8025 		BIC_NOT_PRESENT(BIC_SYS_LPI);
8026 	}
8027 
8028 }
8029 
probe_pstates(void)8030 void probe_pstates(void)
8031 {
8032 	probe_bclk();
8033 
8034 	if (quiet)
8035 		return;
8036 
8037 	dump_platform_info();
8038 	dump_turbo_ratio_info();
8039 	dump_sysfs_pstate_config();
8040 	decode_misc_pwr_mgmt_msr();
8041 
8042 	for_all_cpus(print_hwp, ODD_COUNTERS);
8043 	for_all_cpus(print_epb, ODD_COUNTERS);
8044 	for_all_cpus(print_perf_limit, ODD_COUNTERS);
8045 }
8046 
process_cpuid()8047 void process_cpuid()
8048 {
8049 	unsigned int eax, ebx, ecx, edx;
8050 	unsigned int fms, family, model, stepping, ecx_flags, edx_flags;
8051 	unsigned long long ucode_patch = 0;
8052 	bool ucode_patch_valid = false;
8053 
8054 	eax = ebx = ecx = edx = 0;
8055 
8056 	__cpuid(0, max_level, ebx, ecx, edx);
8057 
8058 	if (ebx == 0x756e6547 && ecx == 0x6c65746e && edx == 0x49656e69)
8059 		genuine_intel = 1;
8060 	else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65)
8061 		authentic_amd = 1;
8062 	else if (ebx == 0x6f677948 && ecx == 0x656e6975 && edx == 0x6e65476e)
8063 		hygon_genuine = 1;
8064 
8065 	if (!quiet)
8066 		fprintf(outf, "CPUID(0): %.4s%.4s%.4s 0x%x CPUID levels\n",
8067 			(char *)&ebx, (char *)&edx, (char *)&ecx, max_level);
8068 
8069 	__cpuid(1, fms, ebx, ecx, edx);
8070 	family = (fms >> 8) & 0xf;
8071 	model = (fms >> 4) & 0xf;
8072 	stepping = fms & 0xf;
8073 	if (family == 0xf)
8074 		family += (fms >> 20) & 0xff;
8075 	if (family >= 6)
8076 		model += ((fms >> 16) & 0xf) << 4;
8077 	ecx_flags = ecx;
8078 	edx_flags = edx;
8079 
8080 	if (!no_msr) {
8081 		if (get_msr(sched_getcpu(), MSR_IA32_UCODE_REV, &ucode_patch))
8082 			warnx("get_msr(UCODE)");
8083 		else
8084 			ucode_patch_valid = true;
8085 	}
8086 
8087 	/*
8088 	 * check max extended function levels of CPUID.
8089 	 * This is needed to check for invariant TSC.
8090 	 * This check is valid for both Intel and AMD.
8091 	 */
8092 	ebx = ecx = edx = 0;
8093 	__cpuid(0x80000000, max_extended_level, ebx, ecx, edx);
8094 
8095 	if (!quiet) {
8096 		fprintf(outf, "CPUID(1): family:model:stepping 0x%x:%x:%x (%d:%d:%d)",
8097 			family, model, stepping, family, model, stepping);
8098 		if (ucode_patch_valid)
8099 			fprintf(outf, " microcode 0x%x", (unsigned int)((ucode_patch >> 32) & 0xFFFFFFFF));
8100 		fputc('\n', outf);
8101 
8102 		fprintf(outf, "CPUID(0x80000000): max_extended_levels: 0x%x\n", max_extended_level);
8103 		fprintf(outf, "CPUID(1): %s %s %s %s %s %s %s %s %s %s\n",
8104 			ecx_flags & (1 << 0) ? "SSE3" : "-",
8105 			ecx_flags & (1 << 3) ? "MONITOR" : "-",
8106 			ecx_flags & (1 << 6) ? "SMX" : "-",
8107 			ecx_flags & (1 << 7) ? "EIST" : "-",
8108 			ecx_flags & (1 << 8) ? "TM2" : "-",
8109 			edx_flags & (1 << 4) ? "TSC" : "-",
8110 			edx_flags & (1 << 5) ? "MSR" : "-",
8111 			edx_flags & (1 << 22) ? "ACPI-TM" : "-",
8112 			edx_flags & (1 << 28) ? "HT" : "-", edx_flags & (1 << 29) ? "TM" : "-");
8113 	}
8114 
8115 	probe_platform_features(family, model);
8116 
8117 	if (!(edx_flags & (1 << 5)))
8118 		errx(1, "CPUID: no MSR");
8119 
8120 	if (max_extended_level >= 0x80000007) {
8121 
8122 		/*
8123 		 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
8124 		 * this check is valid for both Intel and AMD
8125 		 */
8126 		__cpuid(0x80000007, eax, ebx, ecx, edx);
8127 		has_invariant_tsc = edx & (1 << 8);
8128 	}
8129 
8130 	/*
8131 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
8132 	 * this check is valid for both Intel and AMD
8133 	 */
8134 
8135 	__cpuid(0x6, eax, ebx, ecx, edx);
8136 	has_aperf = ecx & (1 << 0);
8137 	do_dts = eax & (1 << 0);
8138 	if (do_dts)
8139 		BIC_PRESENT(BIC_CoreTmp);
8140 	has_turbo = eax & (1 << 1);
8141 	do_ptm = eax & (1 << 6);
8142 	if (do_ptm)
8143 		BIC_PRESENT(BIC_PkgTmp);
8144 	has_hwp = eax & (1 << 7);
8145 	has_hwp_notify = eax & (1 << 8);
8146 	has_hwp_activity_window = eax & (1 << 9);
8147 	has_hwp_epp = eax & (1 << 10);
8148 	has_hwp_pkg = eax & (1 << 11);
8149 	has_epb = ecx & (1 << 3);
8150 
8151 	if (!quiet)
8152 		fprintf(outf, "CPUID(6): %sAPERF, %sTURBO, %sDTS, %sPTM, %sHWP, "
8153 			"%sHWPnotify, %sHWPwindow, %sHWPepp, %sHWPpkg, %sEPB\n",
8154 			has_aperf ? "" : "No-",
8155 			has_turbo ? "" : "No-",
8156 			do_dts ? "" : "No-",
8157 			do_ptm ? "" : "No-",
8158 			has_hwp ? "" : "No-",
8159 			has_hwp_notify ? "" : "No-",
8160 			has_hwp_activity_window ? "" : "No-",
8161 			has_hwp_epp ? "" : "No-", has_hwp_pkg ? "" : "No-", has_epb ? "" : "No-");
8162 
8163 	if (!quiet)
8164 		decode_misc_enable_msr();
8165 
8166 	if (max_level >= 0x7 && !quiet) {
8167 		int has_sgx;
8168 
8169 		ecx = 0;
8170 
8171 		__cpuid_count(0x7, 0, eax, ebx, ecx, edx);
8172 
8173 		has_sgx = ebx & (1 << 2);
8174 
8175 		is_hybrid = edx & (1 << 15);
8176 
8177 		fprintf(outf, "CPUID(7): %sSGX %sHybrid\n", has_sgx ? "" : "No-", is_hybrid ? "" : "No-");
8178 
8179 		if (has_sgx)
8180 			decode_feature_control_msr();
8181 	}
8182 
8183 	if (max_level >= 0x15) {
8184 		unsigned int eax_crystal;
8185 		unsigned int ebx_tsc;
8186 
8187 		/*
8188 		 * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
8189 		 */
8190 		eax_crystal = ebx_tsc = crystal_hz = edx = 0;
8191 		__cpuid(0x15, eax_crystal, ebx_tsc, crystal_hz, edx);
8192 
8193 		if (ebx_tsc != 0) {
8194 			if (!quiet && (ebx != 0))
8195 				fprintf(outf, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n",
8196 					eax_crystal, ebx_tsc, crystal_hz);
8197 
8198 			if (crystal_hz == 0)
8199 				crystal_hz = platform->crystal_freq;
8200 
8201 			if (crystal_hz) {
8202 				tsc_hz = (unsigned long long)crystal_hz *ebx_tsc / eax_crystal;
8203 				if (!quiet)
8204 					fprintf(outf, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n",
8205 						tsc_hz / 1000000, crystal_hz, ebx_tsc, eax_crystal);
8206 			}
8207 		}
8208 	}
8209 	if (max_level >= 0x16) {
8210 		unsigned int base_mhz, max_mhz, bus_mhz, edx;
8211 
8212 		/*
8213 		 * CPUID 16H Base MHz, Max MHz, Bus MHz
8214 		 */
8215 		base_mhz = max_mhz = bus_mhz = edx = 0;
8216 
8217 		__cpuid(0x16, base_mhz, max_mhz, bus_mhz, edx);
8218 
8219 		bclk = bus_mhz;
8220 
8221 		base_hz = base_mhz * 1000000;
8222 		has_base_hz = 1;
8223 
8224 		if (platform->enable_tsc_tweak)
8225 			tsc_tweak = base_hz / tsc_hz;
8226 
8227 		if (!quiet)
8228 			fprintf(outf, "CPUID(0x16): base_mhz: %d max_mhz: %d bus_mhz: %d\n",
8229 				base_mhz, max_mhz, bus_mhz);
8230 	}
8231 
8232 	if (has_aperf)
8233 		aperf_mperf_multiplier = platform->need_perf_multiplier ? 1024 : 1;
8234 
8235 	BIC_PRESENT(BIC_IRQ);
8236 	BIC_PRESENT(BIC_TSC_MHz);
8237 }
8238 
counter_info_init(void)8239 static void counter_info_init(void)
8240 {
8241 	for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i) {
8242 		struct cstate_counter_arch_info *const cai = &ccstate_counter_arch_infos[i];
8243 
8244 		if (platform->has_msr_knl_core_c6_residency && cai->msr == MSR_CORE_C6_RESIDENCY)
8245 			cai->msr = MSR_KNL_CORE_C6_RESIDENCY;
8246 
8247 		if (!platform->has_msr_core_c1_res && cai->msr == MSR_CORE_C1_RES)
8248 			cai->msr = 0;
8249 
8250 		if (platform->has_msr_atom_pkg_c6_residency && cai->msr == MSR_PKG_C6_RESIDENCY)
8251 			cai->msr = MSR_ATOM_PKG_C6_RESIDENCY;
8252 	}
8253 
8254 	for (int i = 0; i < NUM_MSR_COUNTERS; ++i) {
8255 		msr_counter_arch_infos[i].present = false;
8256 		msr_counter_arch_infos[i].needed = false;
8257 	}
8258 }
8259 
probe_pm_features(void)8260 void probe_pm_features(void)
8261 {
8262 	probe_pstates();
8263 
8264 	probe_cstates();
8265 
8266 	probe_lpi();
8267 
8268 	probe_intel_uncore_frequency();
8269 
8270 	probe_graphics();
8271 
8272 	probe_rapl();
8273 
8274 	probe_thermal();
8275 
8276 	if (platform->has_nhm_msrs && !no_msr)
8277 		BIC_PRESENT(BIC_SMI);
8278 
8279 	if (!quiet)
8280 		decode_misc_feature_control();
8281 }
8282 
8283 /*
8284  * in /dev/cpu/ return success for names that are numbers
8285  * ie. filter out ".", "..", "microcode".
8286  */
dir_filter(const struct dirent * dirp)8287 int dir_filter(const struct dirent *dirp)
8288 {
8289 	if (isdigit(dirp->d_name[0]))
8290 		return 1;
8291 	else
8292 		return 0;
8293 }
8294 
topology_probe(bool startup)8295 void topology_probe(bool startup)
8296 {
8297 	int i;
8298 	int max_core_id = 0;
8299 	int max_package_id = 0;
8300 	int max_siblings = 0;
8301 
8302 	/* Initialize num_cpus, max_cpu_num */
8303 	set_max_cpu_num();
8304 	topo.num_cpus = 0;
8305 	for_all_proc_cpus(count_cpus);
8306 	if (!summary_only)
8307 		BIC_PRESENT(BIC_CPU);
8308 
8309 	if (debug > 1)
8310 		fprintf(outf, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
8311 
8312 	cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
8313 	if (cpus == NULL)
8314 		err(1, "calloc cpus");
8315 
8316 	/*
8317 	 * Allocate and initialize cpu_present_set
8318 	 */
8319 	cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
8320 	if (cpu_present_set == NULL)
8321 		err(3, "CPU_ALLOC");
8322 	cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8323 	CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
8324 	for_all_proc_cpus(mark_cpu_present);
8325 
8326 	/*
8327 	 * Allocate and initialize cpu_effective_set
8328 	 */
8329 	cpu_effective_set = CPU_ALLOC((topo.max_cpu_num + 1));
8330 	if (cpu_effective_set == NULL)
8331 		err(3, "CPU_ALLOC");
8332 	cpu_effective_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8333 	CPU_ZERO_S(cpu_effective_setsize, cpu_effective_set);
8334 	update_effective_set(startup);
8335 
8336 	/*
8337 	 * Allocate and initialize cpu_allowed_set
8338 	 */
8339 	cpu_allowed_set = CPU_ALLOC((topo.max_cpu_num + 1));
8340 	if (cpu_allowed_set == NULL)
8341 		err(3, "CPU_ALLOC");
8342 	cpu_allowed_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8343 	CPU_ZERO_S(cpu_allowed_setsize, cpu_allowed_set);
8344 
8345 	/*
8346 	 * Validate and update cpu_allowed_set.
8347 	 *
8348 	 * Make sure all cpus in cpu_subset are also in cpu_present_set during startup.
8349 	 * Give a warning when cpus in cpu_subset become unavailable at runtime.
8350 	 * Give a warning when cpus are not effective because of cgroup setting.
8351 	 *
8352 	 * cpu_allowed_set is the intersection of cpu_present_set/cpu_effective_set/cpu_subset.
8353 	 */
8354 	for (i = 0; i < CPU_SUBSET_MAXCPUS; ++i) {
8355 		if (cpu_subset && !CPU_ISSET_S(i, cpu_subset_size, cpu_subset))
8356 			continue;
8357 
8358 		if (!CPU_ISSET_S(i, cpu_present_setsize, cpu_present_set)) {
8359 			if (cpu_subset) {
8360 				/* cpus in cpu_subset must be in cpu_present_set during startup */
8361 				if (startup)
8362 					err(1, "cpu%d not present", i);
8363 				else
8364 					fprintf(stderr, "cpu%d not present\n", i);
8365 			}
8366 			continue;
8367 		}
8368 
8369 		if (CPU_COUNT_S(cpu_effective_setsize, cpu_effective_set)) {
8370 			if (!CPU_ISSET_S(i, cpu_effective_setsize, cpu_effective_set)) {
8371 				fprintf(stderr, "cpu%d not effective\n", i);
8372 				continue;
8373 			}
8374 		}
8375 
8376 		CPU_SET_S(i, cpu_allowed_setsize, cpu_allowed_set);
8377 	}
8378 
8379 	if (!CPU_COUNT_S(cpu_allowed_setsize, cpu_allowed_set))
8380 		err(-ENODEV, "No valid cpus found");
8381 	sched_setaffinity(0, cpu_allowed_setsize, cpu_allowed_set);
8382 
8383 	/*
8384 	 * Allocate and initialize cpu_affinity_set
8385 	 */
8386 	cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
8387 	if (cpu_affinity_set == NULL)
8388 		err(3, "CPU_ALLOC");
8389 	cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
8390 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
8391 
8392 	for_all_proc_cpus(init_thread_id);
8393 
8394 	for_all_proc_cpus(set_cpu_hybrid_type);
8395 
8396 	/*
8397 	 * For online cpus
8398 	 * find max_core_id, max_package_id
8399 	 */
8400 	for (i = 0; i <= topo.max_cpu_num; ++i) {
8401 		int siblings;
8402 
8403 		if (cpu_is_not_present(i)) {
8404 			if (debug > 1)
8405 				fprintf(outf, "cpu%d NOT PRESENT\n", i);
8406 			continue;
8407 		}
8408 
8409 		cpus[i].logical_cpu_id = i;
8410 
8411 		/* get package information */
8412 		cpus[i].physical_package_id = get_physical_package_id(i);
8413 		if (cpus[i].physical_package_id > max_package_id)
8414 			max_package_id = cpus[i].physical_package_id;
8415 
8416 		/* get die information */
8417 		cpus[i].die_id = get_die_id(i);
8418 		if (cpus[i].die_id > topo.max_die_id)
8419 			topo.max_die_id = cpus[i].die_id;
8420 
8421 		/* get numa node information */
8422 		cpus[i].physical_node_id = get_physical_node_id(&cpus[i]);
8423 		if (cpus[i].physical_node_id > topo.max_node_num)
8424 			topo.max_node_num = cpus[i].physical_node_id;
8425 
8426 		/* get core information */
8427 		cpus[i].physical_core_id = get_core_id(i);
8428 		if (cpus[i].physical_core_id > max_core_id)
8429 			max_core_id = cpus[i].physical_core_id;
8430 
8431 		/* get thread information */
8432 		siblings = get_thread_siblings(&cpus[i]);
8433 		if (siblings > max_siblings)
8434 			max_siblings = siblings;
8435 		if (cpus[i].thread_id == 0)
8436 			topo.num_cores++;
8437 	}
8438 	topo.max_core_id = max_core_id;
8439 	topo.max_package_id = max_package_id;
8440 
8441 	topo.cores_per_node = max_core_id + 1;
8442 	if (debug > 1)
8443 		fprintf(outf, "max_core_id %d, sizing for %d cores per package\n", max_core_id, topo.cores_per_node);
8444 	if (!summary_only)
8445 		BIC_PRESENT(BIC_Core);
8446 
8447 	topo.num_die = topo.max_die_id + 1;
8448 	if (debug > 1)
8449 		fprintf(outf, "max_die_id %d, sizing for %d die\n", topo.max_die_id, topo.num_die);
8450 	if (!summary_only && topo.num_die > 1)
8451 		BIC_PRESENT(BIC_Die);
8452 
8453 	topo.num_packages = max_package_id + 1;
8454 	if (debug > 1)
8455 		fprintf(outf, "max_package_id %d, sizing for %d packages\n", max_package_id, topo.num_packages);
8456 	if (!summary_only && topo.num_packages > 1)
8457 		BIC_PRESENT(BIC_Package);
8458 
8459 	set_node_data();
8460 	if (debug > 1)
8461 		fprintf(outf, "nodes_per_pkg %d\n", topo.nodes_per_pkg);
8462 	if (!summary_only && topo.nodes_per_pkg > 1)
8463 		BIC_PRESENT(BIC_Node);
8464 
8465 	topo.threads_per_core = max_siblings;
8466 	if (debug > 1)
8467 		fprintf(outf, "max_siblings %d\n", max_siblings);
8468 
8469 	if (debug < 1)
8470 		return;
8471 
8472 	for (i = 0; i <= topo.max_cpu_num; ++i) {
8473 		if (cpu_is_not_present(i))
8474 			continue;
8475 		fprintf(outf,
8476 			"cpu %d pkg %d die %d node %d lnode %d core %d thread %d\n",
8477 			i, cpus[i].physical_package_id, cpus[i].die_id,
8478 			cpus[i].physical_node_id, cpus[i].logical_node_id, cpus[i].physical_core_id, cpus[i].thread_id);
8479 	}
8480 
8481 }
8482 
allocate_counters(struct thread_data ** t,struct core_data ** c,struct pkg_data ** p)8483 void allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
8484 {
8485 	int i;
8486 	int num_cores = topo.cores_per_node * topo.nodes_per_pkg * topo.num_packages;
8487 	int num_threads = topo.threads_per_core * num_cores;
8488 
8489 	*t = calloc(num_threads, sizeof(struct thread_data));
8490 	if (*t == NULL)
8491 		goto error;
8492 
8493 	for (i = 0; i < num_threads; i++)
8494 		(*t)[i].cpu_id = -1;
8495 
8496 	*c = calloc(num_cores, sizeof(struct core_data));
8497 	if (*c == NULL)
8498 		goto error;
8499 
8500 	for (i = 0; i < num_cores; i++) {
8501 		(*c)[i].core_id = -1;
8502 		(*c)[i].base_cpu = -1;
8503 	}
8504 
8505 	*p = calloc(topo.num_packages, sizeof(struct pkg_data));
8506 	if (*p == NULL)
8507 		goto error;
8508 
8509 	for (i = 0; i < topo.num_packages; i++) {
8510 		(*p)[i].package_id = i;
8511 		(*p)[i].base_cpu = -1;
8512 	}
8513 
8514 	return;
8515 error:
8516 	err(1, "calloc counters");
8517 }
8518 
8519 /*
8520  * init_counter()
8521  *
8522  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
8523  */
init_counter(struct thread_data * thread_base,struct core_data * core_base,struct pkg_data * pkg_base,int cpu_id)8524 void init_counter(struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base, int cpu_id)
8525 {
8526 	int pkg_id = cpus[cpu_id].physical_package_id;
8527 	int node_id = cpus[cpu_id].logical_node_id;
8528 	int core_id = cpus[cpu_id].physical_core_id;
8529 	int thread_id = cpus[cpu_id].thread_id;
8530 	struct thread_data *t;
8531 	struct core_data *c;
8532 	struct pkg_data *p;
8533 
8534 	/* Workaround for systems where physical_node_id==-1
8535 	 * and logical_node_id==(-1 - topo.num_cpus)
8536 	 */
8537 	if (node_id < 0)
8538 		node_id = 0;
8539 
8540 	t = GET_THREAD(thread_base, thread_id, core_id, node_id, pkg_id);
8541 	c = GET_CORE(core_base, core_id, node_id, pkg_id);
8542 	p = GET_PKG(pkg_base, pkg_id);
8543 
8544 	t->cpu_id = cpu_id;
8545 	if (!cpu_is_not_allowed(cpu_id)) {
8546 		if (c->base_cpu < 0)
8547 			c->base_cpu = t->cpu_id;
8548 		if (p->base_cpu < 0)
8549 			p->base_cpu = t->cpu_id;
8550 	}
8551 
8552 	c->core_id = core_id;
8553 	p->package_id = pkg_id;
8554 }
8555 
initialize_counters(int cpu_id)8556 int initialize_counters(int cpu_id)
8557 {
8558 	init_counter(EVEN_COUNTERS, cpu_id);
8559 	init_counter(ODD_COUNTERS, cpu_id);
8560 	return 0;
8561 }
8562 
allocate_output_buffer()8563 void allocate_output_buffer()
8564 {
8565 	output_buffer = calloc(1, (1 + topo.num_cpus) * 2048);
8566 	outp = output_buffer;
8567 	if (outp == NULL)
8568 		err(-1, "calloc output buffer");
8569 }
8570 
allocate_fd_percpu(void)8571 void allocate_fd_percpu(void)
8572 {
8573 	fd_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8574 	if (fd_percpu == NULL)
8575 		err(-1, "calloc fd_percpu");
8576 }
8577 
allocate_irq_buffers(void)8578 void allocate_irq_buffers(void)
8579 {
8580 	irq_column_2_cpu = calloc(topo.num_cpus, sizeof(int));
8581 	if (irq_column_2_cpu == NULL)
8582 		err(-1, "calloc %d", topo.num_cpus);
8583 
8584 	irqs_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8585 	if (irqs_per_cpu == NULL)
8586 		err(-1, "calloc %d", topo.max_cpu_num + 1);
8587 }
8588 
update_topo(struct thread_data * t,struct core_data * c,struct pkg_data * p)8589 int update_topo(struct thread_data *t, struct core_data *c, struct pkg_data *p)
8590 {
8591 	topo.allowed_cpus++;
8592 	if ((int)t->cpu_id == c->base_cpu)
8593 		topo.allowed_cores++;
8594 	if ((int)t->cpu_id == p->base_cpu)
8595 		topo.allowed_packages++;
8596 
8597 	return 0;
8598 }
8599 
topology_update(void)8600 void topology_update(void)
8601 {
8602 	topo.allowed_cpus = 0;
8603 	topo.allowed_cores = 0;
8604 	topo.allowed_packages = 0;
8605 	for_all_cpus(update_topo, ODD_COUNTERS);
8606 }
8607 
setup_all_buffers(bool startup)8608 void setup_all_buffers(bool startup)
8609 {
8610 	topology_probe(startup);
8611 	allocate_irq_buffers();
8612 	allocate_fd_percpu();
8613 	allocate_counters(&thread_even, &core_even, &package_even);
8614 	allocate_counters(&thread_odd, &core_odd, &package_odd);
8615 	allocate_output_buffer();
8616 	for_all_proc_cpus(initialize_counters);
8617 	topology_update();
8618 }
8619 
set_base_cpu(void)8620 void set_base_cpu(void)
8621 {
8622 	int i;
8623 
8624 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
8625 		if (cpu_is_not_allowed(i))
8626 			continue;
8627 		base_cpu = i;
8628 		if (debug > 1)
8629 			fprintf(outf, "base_cpu = %d\n", base_cpu);
8630 		return;
8631 	}
8632 	err(-ENODEV, "No valid cpus found");
8633 }
8634 
has_added_counters(void)8635 bool has_added_counters(void)
8636 {
8637 	/*
8638 	 * It only makes sense to call this after the command line is parsed,
8639 	 * otherwise sys structure is not populated.
8640 	 */
8641 
8642 	return sys.added_core_counters | sys.added_thread_counters | sys.added_package_counters;
8643 }
8644 
check_msr_access(void)8645 void check_msr_access(void)
8646 {
8647 	check_dev_msr();
8648 	check_msr_permission();
8649 
8650 	if (no_msr)
8651 		bic_disable_msr_access();
8652 }
8653 
check_perf_access(void)8654 void check_perf_access(void)
8655 {
8656 	if (no_perf || !BIC_IS_ENABLED(BIC_IPC) || !has_instr_count_access())
8657 		bic_enabled &= ~BIC_IPC;
8658 }
8659 
perf_has_hybrid_devices(void)8660 bool perf_has_hybrid_devices(void)
8661 {
8662 	/*
8663 	 *  0: unknown
8664 	 *  1: has separate perf device for p and e core
8665 	 * -1: doesn't have separate perf device for p and e core
8666 	 */
8667 	static int cached;
8668 
8669 	if (cached > 0)
8670 		return true;
8671 
8672 	if (cached < 0)
8673 		return false;
8674 
8675 	if (access("/sys/bus/event_source/devices/cpu_core", F_OK)) {
8676 		cached = -1;
8677 		return false;
8678 	}
8679 
8680 	if (access("/sys/bus/event_source/devices/cpu_atom", F_OK)) {
8681 		cached = -1;
8682 		return false;
8683 	}
8684 
8685 	cached = 1;
8686 	return true;
8687 }
8688 
added_perf_counters_init_(struct perf_counter_info * pinfo)8689 int added_perf_counters_init_(struct perf_counter_info *pinfo)
8690 {
8691 	size_t num_domains = 0;
8692 	unsigned int next_domain;
8693 	bool *domain_visited;
8694 	unsigned int perf_type, perf_config;
8695 	double perf_scale;
8696 	int fd_perf;
8697 
8698 	if (!pinfo)
8699 		return 0;
8700 
8701 	const size_t max_num_domains = MAX(topo.max_cpu_num + 1, MAX(topo.max_core_id + 1, topo.max_package_id + 1));
8702 
8703 	domain_visited = calloc(max_num_domains, sizeof(*domain_visited));
8704 
8705 	while (pinfo) {
8706 		switch (pinfo->scope) {
8707 		case SCOPE_CPU:
8708 			num_domains = topo.max_cpu_num + 1;
8709 			break;
8710 
8711 		case SCOPE_CORE:
8712 			num_domains = topo.max_core_id + 1;
8713 			break;
8714 
8715 		case SCOPE_PACKAGE:
8716 			num_domains = topo.max_package_id + 1;
8717 			break;
8718 		}
8719 
8720 		/* Allocate buffer for file descriptor for each domain. */
8721 		pinfo->fd_perf_per_domain = calloc(num_domains, sizeof(*pinfo->fd_perf_per_domain));
8722 		if (!pinfo->fd_perf_per_domain)
8723 			errx(1, "%s: alloc %s", __func__, "fd_perf_per_domain");
8724 
8725 		for (size_t i = 0; i < num_domains; ++i)
8726 			pinfo->fd_perf_per_domain[i] = -1;
8727 
8728 		pinfo->num_domains = num_domains;
8729 		pinfo->scale = 1.0;
8730 
8731 		memset(domain_visited, 0, max_num_domains * sizeof(*domain_visited));
8732 
8733 		for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
8734 
8735 			next_domain = cpu_to_domain(pinfo, cpu);
8736 
8737 			assert(next_domain < num_domains);
8738 
8739 			if (cpu_is_not_allowed(cpu))
8740 				continue;
8741 
8742 			if (domain_visited[next_domain])
8743 				continue;
8744 
8745 			/*
8746 			 * Intel hybrid platforms expose different perf devices for P and E cores.
8747 			 * Instead of one, "/sys/bus/event_source/devices/cpu" device, there are
8748 			 * "/sys/bus/event_source/devices/{cpu_core,cpu_atom}".
8749 			 *
8750 			 * This makes it more complicated to the user, because most of the counters
8751 			 * are available on both and have to be handled manually, otherwise.
8752 			 *
8753 			 * Code below, allow user to use the old "cpu" name, which is translated accordingly.
8754 			 */
8755 			const char *perf_device = pinfo->device;
8756 
8757 			if (strcmp(perf_device, "cpu") == 0 && perf_has_hybrid_devices()) {
8758 				switch (cpus[cpu].type) {
8759 				case INTEL_PCORE_TYPE:
8760 					perf_device = "cpu_core";
8761 					break;
8762 
8763 				case INTEL_ECORE_TYPE:
8764 					perf_device = "cpu_atom";
8765 					break;
8766 
8767 				default: /* Don't change, we will probably fail and report a problem soon. */
8768 					break;
8769 				}
8770 			}
8771 
8772 			perf_type = read_perf_type(perf_device);
8773 			if (perf_type == (unsigned int)-1) {
8774 				warnx("%s: perf/%s/%s: failed to read %s",
8775 				      __func__, perf_device, pinfo->event, "type");
8776 				continue;
8777 			}
8778 
8779 			perf_config = read_perf_config(perf_device, pinfo->event);
8780 			if (perf_config == (unsigned int)-1) {
8781 				warnx("%s: perf/%s/%s: failed to read %s",
8782 				      __func__, perf_device, pinfo->event, "config");
8783 				continue;
8784 			}
8785 
8786 			/* Scale is not required, some counters just don't have it. */
8787 			perf_scale = read_perf_scale(perf_device, pinfo->event);
8788 			if (perf_scale == 0.0)
8789 				perf_scale = 1.0;
8790 
8791 			fd_perf = open_perf_counter(cpu, perf_type, perf_config, -1, 0);
8792 			if (fd_perf == -1) {
8793 				warnx("%s: perf/%s/%s: failed to open counter on cpu%d",
8794 				      __func__, perf_device, pinfo->event, cpu);
8795 				continue;
8796 			}
8797 
8798 			domain_visited[next_domain] = 1;
8799 			pinfo->fd_perf_per_domain[next_domain] = fd_perf;
8800 			pinfo->scale = perf_scale;
8801 
8802 			if (debug)
8803 				fprintf(stderr, "Add perf/%s/%s cpu%d: %d\n",
8804 					perf_device, pinfo->event, cpu, pinfo->fd_perf_per_domain[next_domain]);
8805 		}
8806 
8807 		pinfo = pinfo->next;
8808 	}
8809 
8810 	free(domain_visited);
8811 
8812 	return 0;
8813 }
8814 
added_perf_counters_init(void)8815 void added_perf_counters_init(void)
8816 {
8817 	if (added_perf_counters_init_(sys.perf_tp))
8818 		errx(1, "%s: %s", __func__, "thread");
8819 
8820 	if (added_perf_counters_init_(sys.perf_cp))
8821 		errx(1, "%s: %s", __func__, "core");
8822 
8823 	if (added_perf_counters_init_(sys.perf_pp))
8824 		errx(1, "%s: %s", __func__, "package");
8825 }
8826 
parse_telem_info_file(int fd_dir,const char * info_filename,const char * format,unsigned long * output)8827 int parse_telem_info_file(int fd_dir, const char *info_filename, const char *format, unsigned long *output)
8828 {
8829 	int fd_telem_info;
8830 	FILE *file_telem_info;
8831 	unsigned long value;
8832 
8833 	fd_telem_info = openat(fd_dir, info_filename, O_RDONLY);
8834 	if (fd_telem_info == -1)
8835 		return -1;
8836 
8837 	file_telem_info = fdopen(fd_telem_info, "r");
8838 	if (file_telem_info == NULL) {
8839 		close(fd_telem_info);
8840 		return -1;
8841 	}
8842 
8843 	if (fscanf(file_telem_info, format, &value) != 1) {
8844 		fclose(file_telem_info);
8845 		return -1;
8846 	}
8847 
8848 	fclose(file_telem_info);
8849 
8850 	*output = value;
8851 
8852 	return 0;
8853 }
8854 
pmt_mmio_open(unsigned int target_guid)8855 struct pmt_mmio *pmt_mmio_open(unsigned int target_guid)
8856 {
8857 	DIR *dirp;
8858 	struct dirent *entry;
8859 	struct stat st;
8860 	unsigned int telem_idx;
8861 	int fd_telem_dir, fd_pmt;
8862 	unsigned long guid, size, offset;
8863 	size_t mmap_size;
8864 	void *mmio;
8865 	struct pmt_mmio *ret = NULL;
8866 
8867 	if (stat(SYSFS_TELEM_PATH, &st) == -1)
8868 		return NULL;
8869 
8870 	dirp = opendir(SYSFS_TELEM_PATH);
8871 	if (dirp == NULL)
8872 		return NULL;
8873 
8874 	for (;;) {
8875 		entry = readdir(dirp);
8876 
8877 		if (entry == NULL)
8878 			break;
8879 
8880 		if (strcmp(entry->d_name, ".") == 0)
8881 			continue;
8882 
8883 		if (strcmp(entry->d_name, "..") == 0)
8884 			continue;
8885 
8886 		if (sscanf(entry->d_name, "telem%u", &telem_idx) != 1)
8887 			continue;
8888 
8889 		if (fstatat(dirfd(dirp), entry->d_name, &st, 0) == -1) {
8890 			break;
8891 		}
8892 
8893 		if (!S_ISDIR(st.st_mode))
8894 			continue;
8895 
8896 		fd_telem_dir = openat(dirfd(dirp), entry->d_name, O_RDONLY);
8897 		if (fd_telem_dir == -1) {
8898 			break;
8899 		}
8900 
8901 		if (parse_telem_info_file(fd_telem_dir, "guid", "%lx", &guid)) {
8902 			close(fd_telem_dir);
8903 			break;
8904 		}
8905 
8906 		if (parse_telem_info_file(fd_telem_dir, "size", "%lu", &size)) {
8907 			close(fd_telem_dir);
8908 			break;
8909 		}
8910 
8911 		if (guid != target_guid) {
8912 			close(fd_telem_dir);
8913 			continue;
8914 		}
8915 
8916 		if (parse_telem_info_file(fd_telem_dir, "offset", "%lu", &offset)) {
8917 			close(fd_telem_dir);
8918 			break;
8919 		}
8920 
8921 		assert(offset == 0);
8922 
8923 		fd_pmt = openat(fd_telem_dir, "telem", O_RDONLY);
8924 		if (fd_pmt == -1)
8925 			goto loop_cleanup_and_break;
8926 
8927 		mmap_size = (size + 0x1000UL) & (~0x1000UL);
8928 		mmio = mmap(0, mmap_size, PROT_READ, MAP_SHARED, fd_pmt, 0);
8929 		if (mmio != MAP_FAILED) {
8930 
8931 			if (debug)
8932 				fprintf(stderr, "%s: 0x%lx mmaped at: %p\n", __func__, guid, mmio);
8933 
8934 			ret = calloc(1, sizeof(*ret));
8935 
8936 			if (!ret) {
8937 				fprintf(stderr, "%s: Failed to allocate pmt_mmio\n", __func__);
8938 				exit(1);
8939 			}
8940 
8941 			ret->guid = guid;
8942 			ret->mmio_base = mmio;
8943 			ret->pmt_offset = offset;
8944 			ret->size = size;
8945 
8946 			ret->next = pmt_mmios;
8947 			pmt_mmios = ret;
8948 		}
8949 
8950 loop_cleanup_and_break:
8951 		close(fd_pmt);
8952 		close(fd_telem_dir);
8953 		break;
8954 	}
8955 
8956 	closedir(dirp);
8957 
8958 	return ret;
8959 }
8960 
pmt_mmio_find(unsigned int guid)8961 struct pmt_mmio *pmt_mmio_find(unsigned int guid)
8962 {
8963 	struct pmt_mmio *pmmio = pmt_mmios;
8964 
8965 	while (pmmio) {
8966 		if (pmmio->guid == guid)
8967 			return pmmio;
8968 
8969 		pmmio = pmmio->next;
8970 	}
8971 
8972 	return NULL;
8973 }
8974 
pmt_get_counter_pointer(struct pmt_mmio * pmmio,unsigned long counter_offset)8975 void *pmt_get_counter_pointer(struct pmt_mmio *pmmio, unsigned long counter_offset)
8976 {
8977 	char *ret;
8978 
8979 	/* Get base of mmaped PMT file. */
8980 	ret = (char *)pmmio->mmio_base;
8981 
8982 	/*
8983 	 * Apply PMT MMIO offset to obtain beginning of the mmaped telemetry data.
8984 	 * It's not guaranteed that the mmaped memory begins with the telemetry data
8985 	 *      - we might have to apply the offset first.
8986 	 */
8987 	ret += pmmio->pmt_offset;
8988 
8989 	/* Apply the counter offset to get the address to the mmaped counter. */
8990 	ret += counter_offset;
8991 
8992 	return ret;
8993 }
8994 
pmt_add_guid(unsigned int guid)8995 struct pmt_mmio *pmt_add_guid(unsigned int guid)
8996 {
8997 	struct pmt_mmio *ret;
8998 
8999 	ret = pmt_mmio_find(guid);
9000 	if (!ret)
9001 		ret = pmt_mmio_open(guid);
9002 
9003 	return ret;
9004 }
9005 
9006 enum pmt_open_mode {
9007 	PMT_OPEN_TRY,		/* Open failure is not an error. */
9008 	PMT_OPEN_REQUIRED,	/* Open failure is a fatal error. */
9009 };
9010 
pmt_find_counter(struct pmt_counter * pcounter,const char * name)9011 struct pmt_counter *pmt_find_counter(struct pmt_counter *pcounter, const char *name)
9012 {
9013 	while (pcounter) {
9014 		if (strcmp(pcounter->name, name) == 0)
9015 			break;
9016 
9017 		pcounter = pcounter->next;
9018 	}
9019 
9020 	return pcounter;
9021 }
9022 
pmt_get_scope_root(enum counter_scope scope)9023 struct pmt_counter **pmt_get_scope_root(enum counter_scope scope)
9024 {
9025 	switch (scope) {
9026 	case SCOPE_CPU:
9027 		return &sys.pmt_tp;
9028 	case SCOPE_CORE:
9029 		return &sys.pmt_cp;
9030 	case SCOPE_PACKAGE:
9031 		return &sys.pmt_pp;
9032 	}
9033 
9034 	__builtin_unreachable();
9035 }
9036 
pmt_counter_add_domain(struct pmt_counter * pcounter,unsigned long * pmmio,unsigned int domain_id)9037 void pmt_counter_add_domain(struct pmt_counter *pcounter, unsigned long *pmmio, unsigned int domain_id)
9038 {
9039 	/* Make sure the new domain fits. */
9040 	if (domain_id >= pcounter->num_domains)
9041 		pmt_counter_resize(pcounter, domain_id + 1);
9042 
9043 	assert(pcounter->domains);
9044 	assert(domain_id < pcounter->num_domains);
9045 
9046 	pcounter->domains[domain_id].pcounter = pmmio;
9047 }
9048 
pmt_add_counter(unsigned int guid,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)9049 int pmt_add_counter(unsigned int guid, const char *name, enum pmt_datatype type,
9050 		    unsigned int lsb, unsigned int msb, unsigned int offset, enum counter_scope scope,
9051 		    enum counter_format format, unsigned int domain_id, enum pmt_open_mode mode)
9052 {
9053 	struct pmt_mmio *mmio;
9054 	struct pmt_counter *pcounter;
9055 	struct pmt_counter **const pmt_root = pmt_get_scope_root(scope);
9056 	bool new_counter = false;
9057 	int conflict = 0;
9058 
9059 	if (lsb > msb) {
9060 		fprintf(stderr, "%s: %s: `%s` must be satisfied\n", __func__, "lsb <= msb", name);
9061 		exit(1);
9062 	}
9063 
9064 	if (msb >= 64) {
9065 		fprintf(stderr, "%s: %s: `%s` must be satisfied\n", __func__, "msb < 64", name);
9066 		exit(1);
9067 	}
9068 
9069 	mmio = pmt_add_guid(guid);
9070 	if (!mmio) {
9071 		if (mode != PMT_OPEN_TRY) {
9072 			fprintf(stderr, "%s: failed to map PMT MMIO for guid %x\n", __func__, guid);
9073 			exit(1);
9074 		}
9075 
9076 		return 1;
9077 	}
9078 
9079 	if (offset >= mmio->size) {
9080 		if (mode != PMT_OPEN_TRY) {
9081 			fprintf(stderr, "%s: offset %u outside of PMT MMIO size %u\n", __func__, offset, mmio->size);
9082 			exit(1);
9083 		}
9084 
9085 		return 1;
9086 	}
9087 
9088 	pcounter = pmt_find_counter(*pmt_root, name);
9089 	if (!pcounter) {
9090 		pcounter = calloc(1, sizeof(*pcounter));
9091 		new_counter = true;
9092 	}
9093 
9094 	if (new_counter) {
9095 		strncpy(pcounter->name, name, ARRAY_SIZE(pcounter->name) - 1);
9096 		pcounter->type = type;
9097 		pcounter->scope = scope;
9098 		pcounter->lsb = lsb;
9099 		pcounter->msb = msb;
9100 		pcounter->format = format;
9101 	} else {
9102 		conflict += pcounter->type != type;
9103 		conflict += pcounter->scope != scope;
9104 		conflict += pcounter->lsb != lsb;
9105 		conflict += pcounter->msb != msb;
9106 		conflict += pcounter->format != format;
9107 	}
9108 
9109 	if (conflict) {
9110 		fprintf(stderr, "%s: conflicting parameters for the PMT counter with the same name %s\n",
9111 			__func__, name);
9112 		exit(1);
9113 	}
9114 
9115 	pmt_counter_add_domain(pcounter, pmt_get_counter_pointer(mmio, offset), domain_id);
9116 
9117 	if (new_counter) {
9118 		pcounter->next = *pmt_root;
9119 		*pmt_root = pcounter;
9120 	}
9121 
9122 	return 0;
9123 }
9124 
pmt_init(void)9125 void pmt_init(void)
9126 {
9127 	if (BIC_IS_ENABLED(BIC_Diec6)) {
9128 		pmt_add_counter(PMT_MTL_DC6_GUID, "Die%c6", PMT_TYPE_XTAL_TIME, PMT_COUNTER_MTL_DC6_LSB,
9129 				PMT_COUNTER_MTL_DC6_MSB, PMT_COUNTER_MTL_DC6_OFFSET, SCOPE_PACKAGE, FORMAT_DELTA,
9130 				0, PMT_OPEN_TRY);
9131 	}
9132 }
9133 
turbostat_init()9134 void turbostat_init()
9135 {
9136 	setup_all_buffers(true);
9137 	set_base_cpu();
9138 	check_msr_access();
9139 	check_perf_access();
9140 	process_cpuid();
9141 	counter_info_init();
9142 	probe_pm_features();
9143 	msr_perf_init();
9144 	linux_perf_init();
9145 	rapl_perf_init();
9146 	cstate_perf_init();
9147 	added_perf_counters_init();
9148 	pmt_init();
9149 
9150 	for_all_cpus(get_cpu_type, ODD_COUNTERS);
9151 	for_all_cpus(get_cpu_type, EVEN_COUNTERS);
9152 
9153 	if (BIC_IS_ENABLED(BIC_IPC) && has_aperf_access && get_instr_count_fd(base_cpu) != -1)
9154 		BIC_PRESENT(BIC_IPC);
9155 
9156 	/*
9157 	 * If TSC tweak is needed, but couldn't get it,
9158 	 * disable more BICs, since it can't be reported accurately.
9159 	 */
9160 	if (platform->enable_tsc_tweak && !has_base_hz) {
9161 		bic_enabled &= ~BIC_Busy;
9162 		bic_enabled &= ~BIC_Bzy_MHz;
9163 	}
9164 }
9165 
fork_it(char ** argv)9166 int fork_it(char **argv)
9167 {
9168 	pid_t child_pid;
9169 	int status;
9170 
9171 	snapshot_proc_sysfs_files();
9172 	status = for_all_cpus(get_counters, EVEN_COUNTERS);
9173 	first_counter_read = 0;
9174 	if (status)
9175 		exit(status);
9176 	gettimeofday(&tv_even, (struct timezone *)NULL);
9177 
9178 	child_pid = fork();
9179 	if (!child_pid) {
9180 		/* child */
9181 		execvp(argv[0], argv);
9182 		err(errno, "exec %s", argv[0]);
9183 	} else {
9184 
9185 		/* parent */
9186 		if (child_pid == -1)
9187 			err(1, "fork");
9188 
9189 		signal(SIGINT, SIG_IGN);
9190 		signal(SIGQUIT, SIG_IGN);
9191 		if (waitpid(child_pid, &status, 0) == -1)
9192 			err(status, "waitpid");
9193 
9194 		if (WIFEXITED(status))
9195 			status = WEXITSTATUS(status);
9196 	}
9197 	/*
9198 	 * n.b. fork_it() does not check for errors from for_all_cpus()
9199 	 * because re-starting is problematic when forking
9200 	 */
9201 	snapshot_proc_sysfs_files();
9202 	for_all_cpus(get_counters, ODD_COUNTERS);
9203 	gettimeofday(&tv_odd, (struct timezone *)NULL);
9204 	timersub(&tv_odd, &tv_even, &tv_delta);
9205 	if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS))
9206 		fprintf(outf, "%s: Counter reset detected\n", progname);
9207 	else {
9208 		compute_average(EVEN_COUNTERS);
9209 		format_all_counters(EVEN_COUNTERS);
9210 	}
9211 
9212 	fprintf(outf, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec / 1000000.0);
9213 
9214 	flush_output_stderr();
9215 
9216 	return status;
9217 }
9218 
get_and_dump_counters(void)9219 int get_and_dump_counters(void)
9220 {
9221 	int status;
9222 
9223 	snapshot_proc_sysfs_files();
9224 	status = for_all_cpus(get_counters, ODD_COUNTERS);
9225 	if (status)
9226 		return status;
9227 
9228 	status = for_all_cpus(dump_counters, ODD_COUNTERS);
9229 	if (status)
9230 		return status;
9231 
9232 	flush_output_stdout();
9233 
9234 	return status;
9235 }
9236 
print_version()9237 void print_version()
9238 {
9239 	fprintf(outf, "turbostat version 2024.11.30 - Len Brown <lenb@kernel.org>\n");
9240 }
9241 
9242 #define COMMAND_LINE_SIZE 2048
9243 
print_bootcmd(void)9244 void print_bootcmd(void)
9245 {
9246 	char bootcmd[COMMAND_LINE_SIZE];
9247 	FILE *fp;
9248 	int ret;
9249 
9250 	memset(bootcmd, 0, COMMAND_LINE_SIZE);
9251 	fp = fopen("/proc/cmdline", "r");
9252 	if (!fp)
9253 		return;
9254 
9255 	ret = fread(bootcmd, sizeof(char), COMMAND_LINE_SIZE - 1, fp);
9256 	if (ret) {
9257 		bootcmd[ret] = '\0';
9258 		/* the last character is already '\n' */
9259 		fprintf(outf, "Kernel command line: %s", bootcmd);
9260 	}
9261 
9262 	fclose(fp);
9263 }
9264 
find_msrp_by_name(struct msr_counter * head,char * name)9265 struct msr_counter *find_msrp_by_name(struct msr_counter *head, char *name)
9266 {
9267 	struct msr_counter *mp;
9268 
9269 	for (mp = head; mp; mp = mp->next) {
9270 		if (debug)
9271 			fprintf(stderr, "%s: %s %s\n", __func__, name, mp->name);
9272 		if (!strncmp(name, mp->name, strlen(mp->name)))
9273 			return mp;
9274 	}
9275 	return NULL;
9276 }
9277 
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)9278 int add_counter(unsigned int msr_num, char *path, char *name,
9279 		unsigned int width, enum counter_scope scope,
9280 		enum counter_type type, enum counter_format format, int flags, int id)
9281 {
9282 	struct msr_counter *msrp;
9283 
9284 	if (no_msr && msr_num)
9285 		errx(1, "Requested MSR counter 0x%x, but in --no-msr mode", msr_num);
9286 
9287 	if (debug)
9288 		fprintf(stderr, "%s(msr%d, %s, %s, width%d, scope%d, type%d, format%d, flags%x, id%d)\n",
9289 			__func__, msr_num, path, name, width, scope, type, format, flags, id);
9290 
9291 	switch (scope) {
9292 
9293 	case SCOPE_CPU:
9294 		msrp = find_msrp_by_name(sys.tp, name);
9295 		if (msrp) {
9296 			if (debug)
9297 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
9298 			break;
9299 		}
9300 		if (sys.added_thread_counters++ >= MAX_ADDED_THREAD_COUNTERS) {
9301 			warnx("ignoring thread counter %s", name);
9302 			return -1;
9303 		}
9304 		break;
9305 	case SCOPE_CORE:
9306 		msrp = find_msrp_by_name(sys.cp, name);
9307 		if (msrp) {
9308 			if (debug)
9309 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
9310 			break;
9311 		}
9312 		if (sys.added_core_counters++ >= MAX_ADDED_CORE_COUNTERS) {
9313 			warnx("ignoring core counter %s", name);
9314 			return -1;
9315 		}
9316 		break;
9317 	case SCOPE_PACKAGE:
9318 		msrp = find_msrp_by_name(sys.pp, name);
9319 		if (msrp) {
9320 			if (debug)
9321 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
9322 			break;
9323 		}
9324 		if (sys.added_package_counters++ >= MAX_ADDED_PACKAGE_COUNTERS) {
9325 			warnx("ignoring package counter %s", name);
9326 			return -1;
9327 		}
9328 		break;
9329 	default:
9330 		warnx("ignoring counter %s with unknown scope", name);
9331 		return -1;
9332 	}
9333 
9334 	if (msrp == NULL) {
9335 		msrp = calloc(1, sizeof(struct msr_counter));
9336 		if (msrp == NULL)
9337 			err(-1, "calloc msr_counter");
9338 
9339 		msrp->msr_num = msr_num;
9340 		strncpy(msrp->name, name, NAME_BYTES - 1);
9341 		msrp->width = width;
9342 		msrp->type = type;
9343 		msrp->format = format;
9344 		msrp->flags = flags;
9345 
9346 		switch (scope) {
9347 		case SCOPE_CPU:
9348 			msrp->next = sys.tp;
9349 			sys.tp = msrp;
9350 			break;
9351 		case SCOPE_CORE:
9352 			msrp->next = sys.cp;
9353 			sys.cp = msrp;
9354 			break;
9355 		case SCOPE_PACKAGE:
9356 			msrp->next = sys.pp;
9357 			sys.pp = msrp;
9358 			break;
9359 		}
9360 	}
9361 
9362 	if (path) {
9363 		struct sysfs_path *sp;
9364 
9365 		sp = calloc(1, sizeof(struct sysfs_path));
9366 		if (sp == NULL) {
9367 			perror("calloc");
9368 			exit(1);
9369 		}
9370 		strncpy(sp->path, path, PATH_BYTES - 1);
9371 		sp->id = id;
9372 		sp->next = msrp->sp;
9373 		msrp->sp = sp;
9374 	}
9375 
9376 	return 0;
9377 }
9378 
9379 /*
9380  * Initialize the fields used for identifying and opening the counter.
9381  *
9382  * Defer the initialization of any runtime buffers for actually reading
9383  * the counters for when we initialize all perf counters, so we can later
9384  * easily call re_initialize().
9385  */
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)9386 struct perf_counter_info *make_perf_counter_info(const char *perf_device,
9387 						 const char *perf_event,
9388 						 const char *name,
9389 						 unsigned int width,
9390 						 enum counter_scope scope,
9391 						 enum counter_type type, enum counter_format format)
9392 {
9393 	struct perf_counter_info *pinfo;
9394 
9395 	pinfo = calloc(1, sizeof(*pinfo));
9396 	if (!pinfo)
9397 		errx(1, "%s: Failed to allocate %s/%s\n", __func__, perf_device, perf_event);
9398 
9399 	strncpy(pinfo->device, perf_device, ARRAY_SIZE(pinfo->device) - 1);
9400 	strncpy(pinfo->event, perf_event, ARRAY_SIZE(pinfo->event) - 1);
9401 
9402 	strncpy(pinfo->name, name, ARRAY_SIZE(pinfo->name) - 1);
9403 	pinfo->width = width;
9404 	pinfo->scope = scope;
9405 	pinfo->type = type;
9406 	pinfo->format = format;
9407 
9408 	return pinfo;
9409 }
9410 
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)9411 int add_perf_counter(const char *perf_device, const char *perf_event, const char *name_buffer, unsigned int width,
9412 		     enum counter_scope scope, enum counter_type type, enum counter_format format)
9413 {
9414 	struct perf_counter_info *pinfo;
9415 
9416 	switch (scope) {
9417 	case SCOPE_CPU:
9418 		if (sys.added_thread_perf_counters >= MAX_ADDED_THREAD_COUNTERS) {
9419 			warnx("ignoring thread counter perf/%s/%s", perf_device, perf_event);
9420 			return -1;
9421 		}
9422 		break;
9423 
9424 	case SCOPE_CORE:
9425 		if (sys.added_core_perf_counters >= MAX_ADDED_CORE_COUNTERS) {
9426 			warnx("ignoring core counter perf/%s/%s", perf_device, perf_event);
9427 			return -1;
9428 		}
9429 		break;
9430 
9431 	case SCOPE_PACKAGE:
9432 		if (sys.added_package_perf_counters >= MAX_ADDED_PACKAGE_COUNTERS) {
9433 			warnx("ignoring package counter perf/%s/%s", perf_device, perf_event);
9434 			return -1;
9435 		}
9436 		break;
9437 	}
9438 
9439 	pinfo = make_perf_counter_info(perf_device, perf_event, name_buffer, width, scope, type, format);
9440 
9441 	if (!pinfo)
9442 		return -1;
9443 
9444 	switch (scope) {
9445 	case SCOPE_CPU:
9446 		pinfo->next = sys.perf_tp;
9447 		sys.perf_tp = pinfo;
9448 		++sys.added_thread_perf_counters;
9449 		break;
9450 
9451 	case SCOPE_CORE:
9452 		pinfo->next = sys.perf_cp;
9453 		sys.perf_cp = pinfo;
9454 		++sys.added_core_perf_counters;
9455 		break;
9456 
9457 	case SCOPE_PACKAGE:
9458 		pinfo->next = sys.perf_pp;
9459 		sys.perf_pp = pinfo;
9460 		++sys.added_package_perf_counters;
9461 		break;
9462 	}
9463 
9464 	// FIXME: we might not have debug here yet
9465 	if (debug)
9466 		fprintf(stderr, "%s: %s/%s, name: %s, scope%d\n",
9467 			__func__, pinfo->device, pinfo->event, pinfo->name, pinfo->scope);
9468 
9469 	return 0;
9470 }
9471 
parse_add_command_msr(char * add_command)9472 void parse_add_command_msr(char *add_command)
9473 {
9474 	int msr_num = 0;
9475 	char *path = NULL;
9476 	char perf_device[PERF_DEV_NAME_BYTES] = "";
9477 	char perf_event[PERF_EVT_NAME_BYTES] = "";
9478 	char name_buffer[PERF_NAME_BYTES] = "";
9479 	int width = 64;
9480 	int fail = 0;
9481 	enum counter_scope scope = SCOPE_CPU;
9482 	enum counter_type type = COUNTER_CYCLES;
9483 	enum counter_format format = FORMAT_DELTA;
9484 
9485 	while (add_command) {
9486 
9487 		if (sscanf(add_command, "msr0x%x", &msr_num) == 1)
9488 			goto next;
9489 
9490 		if (sscanf(add_command, "msr%d", &msr_num) == 1)
9491 			goto next;
9492 
9493 		BUILD_BUG_ON(ARRAY_SIZE(perf_device) <= 31);
9494 		BUILD_BUG_ON(ARRAY_SIZE(perf_event) <= 31);
9495 		if (sscanf(add_command, "perf/%31[^/]/%31[^,]", &perf_device[0], &perf_event[0]) == 2)
9496 			goto next;
9497 
9498 		if (*add_command == '/') {
9499 			path = add_command;
9500 			goto next;
9501 		}
9502 
9503 		if (sscanf(add_command, "u%d", &width) == 1) {
9504 			if ((width == 32) || (width == 64))
9505 				goto next;
9506 			width = 64;
9507 		}
9508 		if (!strncmp(add_command, "cpu", strlen("cpu"))) {
9509 			scope = SCOPE_CPU;
9510 			goto next;
9511 		}
9512 		if (!strncmp(add_command, "core", strlen("core"))) {
9513 			scope = SCOPE_CORE;
9514 			goto next;
9515 		}
9516 		if (!strncmp(add_command, "package", strlen("package"))) {
9517 			scope = SCOPE_PACKAGE;
9518 			goto next;
9519 		}
9520 		if (!strncmp(add_command, "cycles", strlen("cycles"))) {
9521 			type = COUNTER_CYCLES;
9522 			goto next;
9523 		}
9524 		if (!strncmp(add_command, "seconds", strlen("seconds"))) {
9525 			type = COUNTER_SECONDS;
9526 			goto next;
9527 		}
9528 		if (!strncmp(add_command, "usec", strlen("usec"))) {
9529 			type = COUNTER_USEC;
9530 			goto next;
9531 		}
9532 		if (!strncmp(add_command, "raw", strlen("raw"))) {
9533 			format = FORMAT_RAW;
9534 			goto next;
9535 		}
9536 		if (!strncmp(add_command, "delta", strlen("delta"))) {
9537 			format = FORMAT_DELTA;
9538 			goto next;
9539 		}
9540 		if (!strncmp(add_command, "percent", strlen("percent"))) {
9541 			format = FORMAT_PERCENT;
9542 			goto next;
9543 		}
9544 
9545 		BUILD_BUG_ON(ARRAY_SIZE(name_buffer) <= 18);
9546 		if (sscanf(add_command, "%18s,%*s", name_buffer) == 1) {
9547 			char *eos;
9548 
9549 			eos = strchr(name_buffer, ',');
9550 			if (eos)
9551 				*eos = '\0';
9552 			goto next;
9553 		}
9554 
9555 next:
9556 		add_command = strchr(add_command, ',');
9557 		if (add_command) {
9558 			*add_command = '\0';
9559 			add_command++;
9560 		}
9561 
9562 	}
9563 	if ((msr_num == 0) && (path == NULL) && (perf_device[0] == '\0' || perf_event[0] == '\0')) {
9564 		fprintf(stderr, "--add: (msrDDD | msr0xXXX | /path_to_counter | perf/device/event ) required\n");
9565 		fail++;
9566 	}
9567 
9568 	/* Test for non-empty perf_device and perf_event */
9569 	const bool is_perf_counter = perf_device[0] && perf_event[0];
9570 
9571 	/* generate default column header */
9572 	if (*name_buffer == '\0') {
9573 		if (is_perf_counter) {
9574 			snprintf(name_buffer, ARRAY_SIZE(name_buffer), "perf/%s", perf_event);
9575 		} else {
9576 			if (width == 32)
9577 				sprintf(name_buffer, "M0x%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
9578 			else
9579 				sprintf(name_buffer, "M0X%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
9580 		}
9581 	}
9582 
9583 	if (is_perf_counter) {
9584 		if (add_perf_counter(perf_device, perf_event, name_buffer, width, scope, type, format))
9585 			fail++;
9586 	} else {
9587 		if (add_counter(msr_num, path, name_buffer, width, scope, type, format, 0, 0))
9588 			fail++;
9589 	}
9590 
9591 	if (fail) {
9592 		help();
9593 		exit(1);
9594 	}
9595 }
9596 
starts_with(const char * str,const char * prefix)9597 bool starts_with(const char *str, const char *prefix)
9598 {
9599 	return strncmp(prefix, str, strlen(prefix)) == 0;
9600 }
9601 
parse_add_command_pmt(char * add_command)9602 void parse_add_command_pmt(char *add_command)
9603 {
9604 	char *name = NULL;
9605 	char *type_name = NULL;
9606 	char *format_name = NULL;
9607 	unsigned int offset;
9608 	unsigned int lsb;
9609 	unsigned int msb;
9610 	unsigned int guid;
9611 	unsigned int domain_id;
9612 	enum counter_scope scope = 0;
9613 	enum pmt_datatype type = PMT_TYPE_RAW;
9614 	enum counter_format format = FORMAT_RAW;
9615 	bool has_offset = false;
9616 	bool has_lsb = false;
9617 	bool has_msb = false;
9618 	bool has_format = true;	/* Format has a default value. */
9619 	bool has_guid = false;
9620 	bool has_scope = false;
9621 	bool has_type = true;	/* Type has a default value. */
9622 
9623 	/* Consume the "pmt," prefix. */
9624 	add_command = strchr(add_command, ',');
9625 	if (!add_command) {
9626 		help();
9627 		exit(1);
9628 	}
9629 	++add_command;
9630 
9631 	while (add_command) {
9632 		if (starts_with(add_command, "name=")) {
9633 			name = add_command + strlen("name=");
9634 			goto next;
9635 		}
9636 
9637 		if (starts_with(add_command, "type=")) {
9638 			type_name = add_command + strlen("type=");
9639 			goto next;
9640 		}
9641 
9642 		if (starts_with(add_command, "domain=")) {
9643 			const size_t prefix_len = strlen("domain=");
9644 
9645 			if (sscanf(add_command + prefix_len, "cpu%u", &domain_id) == 1) {
9646 				scope = SCOPE_CPU;
9647 				has_scope = true;
9648 			} else if (sscanf(add_command + prefix_len, "core%u", &domain_id) == 1) {
9649 				scope = SCOPE_CORE;
9650 				has_scope = true;
9651 			} else if (sscanf(add_command + prefix_len, "package%u", &domain_id) == 1) {
9652 				scope = SCOPE_PACKAGE;
9653 				has_scope = true;
9654 			}
9655 
9656 			if (!has_scope) {
9657 				printf("%s: invalid value for scope. Expected cpu%%u, core%%u or package%%u.\n",
9658 				       __func__);
9659 				exit(1);
9660 			}
9661 
9662 			goto next;
9663 		}
9664 
9665 		if (starts_with(add_command, "format=")) {
9666 			format_name = add_command + strlen("format=");
9667 			goto next;
9668 		}
9669 
9670 		if (sscanf(add_command, "offset=%u", &offset) == 1) {
9671 			has_offset = true;
9672 			goto next;
9673 		}
9674 
9675 		if (sscanf(add_command, "lsb=%u", &lsb) == 1) {
9676 			has_lsb = true;
9677 			goto next;
9678 		}
9679 
9680 		if (sscanf(add_command, "msb=%u", &msb) == 1) {
9681 			has_msb = true;
9682 			goto next;
9683 		}
9684 
9685 		if (sscanf(add_command, "guid=%x", &guid) == 1) {
9686 			has_guid = true;
9687 			goto next;
9688 		}
9689 
9690 next:
9691 		add_command = strchr(add_command, ',');
9692 		if (add_command) {
9693 			*add_command = '\0';
9694 			add_command++;
9695 		}
9696 	}
9697 
9698 	if (!name) {
9699 		printf("%s: missing %s\n", __func__, "name");
9700 		exit(1);
9701 	}
9702 
9703 	if (strlen(name) >= PMT_COUNTER_NAME_SIZE_BYTES) {
9704 		printf("%s: name has to be at most %d characters long\n", __func__, PMT_COUNTER_NAME_SIZE_BYTES);
9705 		exit(1);
9706 	}
9707 
9708 	if (format_name) {
9709 		has_format = false;
9710 
9711 		if (strcmp("raw", format_name) == 0) {
9712 			format = FORMAT_RAW;
9713 			has_format = true;
9714 		}
9715 
9716 		if (strcmp("delta", format_name) == 0) {
9717 			format = FORMAT_DELTA;
9718 			has_format = true;
9719 		}
9720 
9721 		if (!has_format) {
9722 			fprintf(stderr, "%s: Invalid format %s. Expected raw or delta\n", __func__, format_name);
9723 			exit(1);
9724 		}
9725 	}
9726 
9727 	if (type_name) {
9728 		has_type = false;
9729 
9730 		if (strcmp("raw", type_name) == 0) {
9731 			type = PMT_TYPE_RAW;
9732 			has_type = true;
9733 		}
9734 
9735 		if (strcmp("txtal_time", type_name) == 0) {
9736 			type = PMT_TYPE_XTAL_TIME;
9737 			has_type = true;
9738 		}
9739 
9740 		if (!has_type) {
9741 			printf("%s: invalid %s: %s\n", __func__, "type", type_name);
9742 			exit(1);
9743 		}
9744 	}
9745 
9746 	if (!has_offset) {
9747 		printf("%s : missing %s\n", __func__, "offset");
9748 		exit(1);
9749 	}
9750 
9751 	if (!has_lsb) {
9752 		printf("%s: missing %s\n", __func__, "lsb");
9753 		exit(1);
9754 	}
9755 
9756 	if (!has_msb) {
9757 		printf("%s: missing %s\n", __func__, "msb");
9758 		exit(1);
9759 	}
9760 
9761 	if (!has_guid) {
9762 		printf("%s: missing %s\n", __func__, "guid");
9763 		exit(1);
9764 	}
9765 
9766 	if (!has_scope) {
9767 		printf("%s: missing %s\n", __func__, "scope");
9768 		exit(1);
9769 	}
9770 
9771 	if (lsb > msb) {
9772 		printf("%s: lsb > msb doesn't make sense\n", __func__);
9773 		exit(1);
9774 	}
9775 
9776 	pmt_add_counter(guid, name, type, lsb, msb, offset, scope, format, domain_id, PMT_OPEN_REQUIRED);
9777 }
9778 
parse_add_command(char * add_command)9779 void parse_add_command(char *add_command)
9780 {
9781 	if (strncmp(add_command, "pmt", strlen("pmt")) == 0)
9782 		return parse_add_command_pmt(add_command);
9783 	return parse_add_command_msr(add_command);
9784 }
9785 
is_deferred_add(char * name)9786 int is_deferred_add(char *name)
9787 {
9788 	int i;
9789 
9790 	for (i = 0; i < deferred_add_index; ++i)
9791 		if (!strcmp(name, deferred_add_names[i]))
9792 			return 1;
9793 	return 0;
9794 }
9795 
is_deferred_skip(char * name)9796 int is_deferred_skip(char *name)
9797 {
9798 	int i;
9799 
9800 	for (i = 0; i < deferred_skip_index; ++i)
9801 		if (!strcmp(name, deferred_skip_names[i]))
9802 			return 1;
9803 	return 0;
9804 }
9805 
probe_sysfs(void)9806 void probe_sysfs(void)
9807 {
9808 	char path[64];
9809 	char name_buf[16];
9810 	FILE *input;
9811 	int state;
9812 	char *sp;
9813 
9814 	for (state = 10; state >= 0; --state) {
9815 
9816 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", base_cpu, state);
9817 		input = fopen(path, "r");
9818 		if (input == NULL)
9819 			continue;
9820 		if (!fgets(name_buf, sizeof(name_buf), input))
9821 			err(1, "%s: failed to read file", path);
9822 
9823 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
9824 		sp = strchr(name_buf, '-');
9825 		if (!sp)
9826 			sp = strchrnul(name_buf, '\n');
9827 		*sp = '%';
9828 		*(sp + 1) = '\0';
9829 
9830 		remove_underbar(name_buf);
9831 
9832 		fclose(input);
9833 
9834 		sprintf(path, "cpuidle/state%d/time", state);
9835 
9836 		if (!DO_BIC(BIC_sysfs) && !is_deferred_add(name_buf))
9837 			continue;
9838 
9839 		if (is_deferred_skip(name_buf))
9840 			continue;
9841 
9842 		add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_USEC, FORMAT_PERCENT, SYSFS_PERCPU, 0);
9843 	}
9844 
9845 	for (state = 10; state >= 0; --state) {
9846 
9847 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", base_cpu, state);
9848 		input = fopen(path, "r");
9849 		if (input == NULL)
9850 			continue;
9851 		if (!fgets(name_buf, sizeof(name_buf), input))
9852 			err(1, "%s: failed to read file", path);
9853 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
9854 		sp = strchr(name_buf, '-');
9855 		if (!sp)
9856 			sp = strchrnul(name_buf, '\n');
9857 		*sp = '\0';
9858 		fclose(input);
9859 
9860 		remove_underbar(name_buf);
9861 
9862 		sprintf(path, "cpuidle/state%d/usage", state);
9863 
9864 		if (!DO_BIC(BIC_sysfs) && !is_deferred_add(name_buf))
9865 			continue;
9866 
9867 		if (is_deferred_skip(name_buf))
9868 			continue;
9869 
9870 		add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS, FORMAT_DELTA, SYSFS_PERCPU, 0);
9871 	}
9872 
9873 }
9874 
9875 /*
9876  * parse cpuset with following syntax
9877  * 1,2,4..6,8-10 and set bits in cpu_subset
9878  */
parse_cpu_command(char * optarg)9879 void parse_cpu_command(char *optarg)
9880 {
9881 	if (!strcmp(optarg, "core")) {
9882 		if (cpu_subset)
9883 			goto error;
9884 		show_core_only++;
9885 		return;
9886 	}
9887 	if (!strcmp(optarg, "package")) {
9888 		if (cpu_subset)
9889 			goto error;
9890 		show_pkg_only++;
9891 		return;
9892 	}
9893 	if (show_core_only || show_pkg_only)
9894 		goto error;
9895 
9896 	cpu_subset = CPU_ALLOC(CPU_SUBSET_MAXCPUS);
9897 	if (cpu_subset == NULL)
9898 		err(3, "CPU_ALLOC");
9899 	cpu_subset_size = CPU_ALLOC_SIZE(CPU_SUBSET_MAXCPUS);
9900 
9901 	CPU_ZERO_S(cpu_subset_size, cpu_subset);
9902 
9903 	if (parse_cpu_str(optarg, cpu_subset, cpu_subset_size))
9904 		goto error;
9905 
9906 	return;
9907 
9908 error:
9909 	fprintf(stderr, "\"--cpu %s\" malformed\n", optarg);
9910 	help();
9911 	exit(-1);
9912 }
9913 
cmdline(int argc,char ** argv)9914 void cmdline(int argc, char **argv)
9915 {
9916 	int opt;
9917 	int option_index = 0;
9918 	static struct option long_options[] = {
9919 		{ "add", required_argument, 0, 'a' },
9920 		{ "cpu", required_argument, 0, 'c' },
9921 		{ "Dump", no_argument, 0, 'D' },
9922 		{ "debug", no_argument, 0, 'd' },	/* internal, not documented */
9923 		{ "enable", required_argument, 0, 'e' },
9924 		{ "interval", required_argument, 0, 'i' },
9925 		{ "IPC", no_argument, 0, 'I' },
9926 		{ "num_iterations", required_argument, 0, 'n' },
9927 		{ "header_iterations", required_argument, 0, 'N' },
9928 		{ "help", no_argument, 0, 'h' },
9929 		{ "hide", required_argument, 0, 'H' },	// meh, -h taken by --help
9930 		{ "Joules", no_argument, 0, 'J' },
9931 		{ "list", no_argument, 0, 'l' },
9932 		{ "out", required_argument, 0, 'o' },
9933 		{ "quiet", no_argument, 0, 'q' },
9934 		{ "no-msr", no_argument, 0, 'M' },
9935 		{ "no-perf", no_argument, 0, 'P' },
9936 		{ "show", required_argument, 0, 's' },
9937 		{ "Summary", no_argument, 0, 'S' },
9938 		{ "TCC", required_argument, 0, 'T' },
9939 		{ "version", no_argument, 0, 'v' },
9940 		{ 0, 0, 0, 0 }
9941 	};
9942 
9943 	progname = argv[0];
9944 
9945 	/*
9946 	 * Parse some options early, because they may make other options invalid,
9947 	 * like adding the MSR counter with --add and at the same time using --no-msr.
9948 	 */
9949 	while ((opt = getopt_long_only(argc, argv, "+MPn:", long_options, &option_index)) != -1) {
9950 		switch (opt) {
9951 		case 'M':
9952 			no_msr = 1;
9953 			break;
9954 		case 'P':
9955 			no_perf = 1;
9956 			break;
9957 		default:
9958 			break;
9959 		}
9960 	}
9961 	optind = 0;
9962 
9963 	while ((opt = getopt_long_only(argc, argv, "+C:c:Dde:hi:Jn:o:qMST:v", long_options, &option_index)) != -1) {
9964 		switch (opt) {
9965 		case 'a':
9966 			parse_add_command(optarg);
9967 			break;
9968 		case 'c':
9969 			parse_cpu_command(optarg);
9970 			break;
9971 		case 'D':
9972 			dump_only++;
9973 			/*
9974 			 * Force the no_perf early to prevent using it as a source.
9975 			 * User asks for raw values, but perf returns them relative
9976 			 * to the opening of the file descriptor.
9977 			 */
9978 			no_perf = 1;
9979 			break;
9980 		case 'e':
9981 			/* --enable specified counter */
9982 			bic_enabled = bic_enabled | bic_lookup(optarg, SHOW_LIST);
9983 			break;
9984 		case 'd':
9985 			debug++;
9986 			ENABLE_BIC(BIC_DISABLED_BY_DEFAULT);
9987 			break;
9988 		case 'H':
9989 			/*
9990 			 * --hide: do not show those specified
9991 			 *  multiple invocations simply clear more bits in enabled mask
9992 			 */
9993 			bic_enabled &= ~bic_lookup(optarg, HIDE_LIST);
9994 			break;
9995 		case 'h':
9996 		default:
9997 			help();
9998 			exit(1);
9999 		case 'i':
10000 			{
10001 				double interval = strtod(optarg, NULL);
10002 
10003 				if (interval < 0.001) {
10004 					fprintf(outf, "interval %f seconds is too small\n", interval);
10005 					exit(2);
10006 				}
10007 
10008 				interval_tv.tv_sec = interval_ts.tv_sec = interval;
10009 				interval_tv.tv_usec = (interval - interval_tv.tv_sec) * 1000000;
10010 				interval_ts.tv_nsec = (interval - interval_ts.tv_sec) * 1000000000;
10011 			}
10012 			break;
10013 		case 'J':
10014 			rapl_joules++;
10015 			break;
10016 		case 'l':
10017 			ENABLE_BIC(BIC_DISABLED_BY_DEFAULT);
10018 			list_header_only++;
10019 			quiet++;
10020 			break;
10021 		case 'o':
10022 			outf = fopen_or_die(optarg, "w");
10023 			break;
10024 		case 'q':
10025 			quiet = 1;
10026 			break;
10027 		case 'M':
10028 		case 'P':
10029 			/* Parsed earlier */
10030 			break;
10031 		case 'n':
10032 			num_iterations = strtod(optarg, NULL);
10033 
10034 			if (num_iterations <= 0) {
10035 				fprintf(outf, "iterations %d should be positive number\n", num_iterations);
10036 				exit(2);
10037 			}
10038 			break;
10039 		case 'N':
10040 			header_iterations = strtod(optarg, NULL);
10041 
10042 			if (header_iterations <= 0) {
10043 				fprintf(outf, "iterations %d should be positive number\n", header_iterations);
10044 				exit(2);
10045 			}
10046 			break;
10047 		case 's':
10048 			/*
10049 			 * --show: show only those specified
10050 			 *  The 1st invocation will clear and replace the enabled mask
10051 			 *  subsequent invocations can add to it.
10052 			 */
10053 			if (shown == 0)
10054 				bic_enabled = bic_lookup(optarg, SHOW_LIST);
10055 			else
10056 				bic_enabled |= bic_lookup(optarg, SHOW_LIST);
10057 			shown = 1;
10058 			break;
10059 		case 'S':
10060 			summary_only++;
10061 			break;
10062 		case 'T':
10063 			tj_max_override = atoi(optarg);
10064 			break;
10065 		case 'v':
10066 			print_version();
10067 			exit(0);
10068 			break;
10069 		}
10070 	}
10071 }
10072 
set_rlimit(void)10073 void set_rlimit(void)
10074 {
10075 	struct rlimit limit;
10076 
10077 	if (getrlimit(RLIMIT_NOFILE, &limit) < 0)
10078 		err(1, "Failed to get rlimit");
10079 
10080 	if (limit.rlim_max < MAX_NOFILE)
10081 		limit.rlim_max = MAX_NOFILE;
10082 	if (limit.rlim_cur < MAX_NOFILE)
10083 		limit.rlim_cur = MAX_NOFILE;
10084 
10085 	if (setrlimit(RLIMIT_NOFILE, &limit) < 0)
10086 		err(1, "Failed to set rlimit");
10087 }
10088 
main(int argc,char ** argv)10089 int main(int argc, char **argv)
10090 {
10091 	int fd, ret;
10092 
10093 	fd = open("/sys/fs/cgroup/cgroup.procs", O_WRONLY);
10094 	if (fd < 0)
10095 		goto skip_cgroup_setting;
10096 
10097 	ret = write(fd, "0\n", 2);
10098 	if (ret == -1)
10099 		perror("Can't update cgroup\n");
10100 
10101 	close(fd);
10102 
10103 skip_cgroup_setting:
10104 	outf = stderr;
10105 	cmdline(argc, argv);
10106 
10107 	if (!quiet) {
10108 		print_version();
10109 		print_bootcmd();
10110 	}
10111 
10112 	probe_sysfs();
10113 
10114 	if (!getuid())
10115 		set_rlimit();
10116 
10117 	turbostat_init();
10118 
10119 	if (!no_msr)
10120 		msr_sum_record();
10121 
10122 	/* dump counters and exit */
10123 	if (dump_only)
10124 		return get_and_dump_counters();
10125 
10126 	/* list header and exit */
10127 	if (list_header_only) {
10128 		print_header(",");
10129 		flush_output_stdout();
10130 		return 0;
10131 	}
10132 
10133 	/*
10134 	 * if any params left, it must be a command to fork
10135 	 */
10136 	if (argc - optind)
10137 		return fork_it(argv + optind);
10138 	else
10139 		turbostat_loop();
10140 
10141 	return 0;
10142 }
10143