xref: /linux/tools/power/x86/turbostat/turbostat.c (revision 2ff1bc41ef9133a52c116b36e5e88430353a8ba5)
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) 2010 - 2026 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 #include <limits.h>
71 
72 #define UNUSED(x) (void)(x)
73 
74 /*
75  * This list matches the column headers, except
76  * 1. built-in only, the sysfs counters are not here -- we learn of those at run-time
77  * 2. Core and CPU are moved to the end, we can't have strings that contain them
78  *    matching on them for --show and --hide.
79  */
80 
81 /*
82  * buffer size used by sscanf() for added column names
83  * Usually truncated to 7 characters, but also handles 18 columns for raw 64-bit counters
84  */
85 #define	NAME_BYTES 20
86 #define PATH_BYTES 128
87 #define PERF_NAME_BYTES 128
88 
89 #define MAX_NOFILE 0x8000
90 
91 #define COUNTER_KIND_PERF_PREFIX "perf/"
92 #define COUNTER_KIND_PERF_PREFIX_LEN strlen(COUNTER_KIND_PERF_PREFIX)
93 #define PERF_DEV_NAME_BYTES 32
94 #define PERF_EVT_NAME_BYTES 32
95 
96 #define INTEL_ECORE_TYPE	0x20
97 #define INTEL_PCORE_TYPE	0x40
98 
99 #define ROUND_UP_TO_PAGE_SIZE(n) (((n) + 0x1000UL-1UL) & ~(0x1000UL-1UL))
100 
101 enum counter_scope { SCOPE_CPU, SCOPE_CORE, SCOPE_PACKAGE };
102 enum counter_type { COUNTER_ITEMS, COUNTER_CYCLES, COUNTER_SECONDS, COUNTER_USEC, COUNTER_K2M };
103 enum counter_format { FORMAT_RAW, FORMAT_DELTA, FORMAT_PERCENT, FORMAT_AVERAGE };
104 enum counter_source { COUNTER_SOURCE_NONE, COUNTER_SOURCE_PERF, COUNTER_SOURCE_MSR };
105 
106 struct perf_counter_info {
107 	struct perf_counter_info *next;
108 
109 	/* How to open the counter / What counter it is. */
110 	char device[PERF_DEV_NAME_BYTES];
111 	char event[PERF_EVT_NAME_BYTES];
112 
113 	/* How to show/format the counter. */
114 	char name[PERF_NAME_BYTES];
115 	unsigned int width;
116 	enum counter_scope scope;
117 	enum counter_type type;
118 	enum counter_format format;
119 	double scale;
120 
121 	/* For reading the counter. */
122 	int *fd_perf_per_domain;
123 	size_t num_domains;
124 };
125 
126 struct sysfs_path {
127 	char path[PATH_BYTES];
128 	int id;
129 	struct sysfs_path *next;
130 };
131 
132 struct msr_counter {
133 	unsigned int msr_num;
134 	char name[NAME_BYTES];
135 	struct sysfs_path *sp;
136 	unsigned int width;
137 	enum counter_type type;
138 	enum counter_format format;
139 	struct msr_counter *next;
140 	unsigned int flags;
141 #define	FLAGS_HIDE	(1 << 0)
142 #define	FLAGS_SHOW	(1 << 1)
143 #define	SYSFS_PERCPU	(1 << 1)
144 };
145 static int use_android_msr_path;
146 
147 struct msr_counter bic[] = {
148 	{ 0x0, "usec", NULL, 0, 0, 0, NULL, 0 },
149 	{ 0x0, "Time_Of_Day_Seconds", NULL, 0, 0, 0, NULL, 0 },
150 	{ 0x0, "Package", NULL, 0, 0, 0, NULL, 0 },
151 	{ 0x0, "Node", NULL, 0, 0, 0, NULL, 0 },
152 	{ 0x0, "Avg_MHz", NULL, 0, 0, 0, NULL, 0 },
153 	{ 0x0, "Busy%", NULL, 0, 0, 0, NULL, 0 },
154 	{ 0x0, "Bzy_MHz", NULL, 0, 0, 0, NULL, 0 },
155 	{ 0x0, "TSC_MHz", NULL, 0, 0, 0, NULL, 0 },
156 	{ 0x0, "IRQ", NULL, 0, 0, 0, NULL, 0 },
157 	{ 0x0, "SMI", NULL, 32, 0, FORMAT_DELTA, NULL, 0 },
158 	{ 0x0, "cpuidle", NULL, 0, 0, 0, NULL, 0 },
159 	{ 0x0, "CPU%c1", NULL, 0, 0, 0, NULL, 0 },
160 	{ 0x0, "CPU%c3", NULL, 0, 0, 0, NULL, 0 },
161 	{ 0x0, "CPU%c6", NULL, 0, 0, 0, NULL, 0 },
162 	{ 0x0, "CPU%c7", NULL, 0, 0, 0, NULL, 0 },
163 	{ 0x0, "ThreadC", NULL, 0, 0, 0, NULL, 0 },
164 	{ 0x0, "CoreTmp", NULL, 0, 0, 0, NULL, 0 },
165 	{ 0x0, "CoreCnt", NULL, 0, 0, 0, NULL, 0 },
166 	{ 0x0, "PkgTmp", NULL, 0, 0, 0, NULL, 0 },
167 	{ 0x0, "GFX%rc6", NULL, 0, 0, 0, NULL, 0 },
168 	{ 0x0, "GFXMHz", NULL, 0, 0, 0, NULL, 0 },
169 	{ 0x0, "Pkg%pc2", NULL, 0, 0, 0, NULL, 0 },
170 	{ 0x0, "Pkg%pc3", NULL, 0, 0, 0, NULL, 0 },
171 	{ 0x0, "Pkg%pc6", NULL, 0, 0, 0, NULL, 0 },
172 	{ 0x0, "Pkg%pc7", NULL, 0, 0, 0, NULL, 0 },
173 	{ 0x0, "Pkg%pc8", NULL, 0, 0, 0, NULL, 0 },
174 	{ 0x0, "Pkg%pc9", NULL, 0, 0, 0, NULL, 0 },
175 	{ 0x0, "Pk%pc10", NULL, 0, 0, 0, NULL, 0 },
176 	{ 0x0, "CPU%LPI", NULL, 0, 0, 0, NULL, 0 },
177 	{ 0x0, "SYS%LPI", NULL, 0, 0, 0, NULL, 0 },
178 	{ 0x0, "PkgWatt", NULL, 0, 0, 0, NULL, 0 },
179 	{ 0x0, "CorWatt", NULL, 0, 0, 0, NULL, 0 },
180 	{ 0x0, "GFXWatt", NULL, 0, 0, 0, NULL, 0 },
181 	{ 0x0, "PkgCnt", NULL, 0, 0, 0, NULL, 0 },
182 	{ 0x0, "RAMWatt", NULL, 0, 0, 0, NULL, 0 },
183 	{ 0x0, "PKG_%", NULL, 0, 0, 0, NULL, 0 },
184 	{ 0x0, "RAM_%", NULL, 0, 0, 0, NULL, 0 },
185 	{ 0x0, "Pkg_J", NULL, 0, 0, 0, NULL, 0 },
186 	{ 0x0, "Cor_J", NULL, 0, 0, 0, NULL, 0 },
187 	{ 0x0, "GFX_J", NULL, 0, 0, 0, NULL, 0 },
188 	{ 0x0, "RAM_J", NULL, 0, 0, 0, NULL, 0 },
189 	{ 0x0, "Mod%c6", NULL, 0, 0, 0, NULL, 0 },
190 	{ 0x0, "Totl%C0", NULL, 0, 0, 0, NULL, 0 },
191 	{ 0x0, "Any%C0", NULL, 0, 0, 0, NULL, 0 },
192 	{ 0x0, "GFX%C0", NULL, 0, 0, 0, NULL, 0 },
193 	{ 0x0, "CPUGFX%", NULL, 0, 0, 0, NULL, 0 },
194 	{ 0x0, "Module", NULL, 0, 0, 0, NULL, 0 },
195 	{ 0x0, "Core", NULL, 0, 0, 0, NULL, 0 },
196 	{ 0x0, "CPU", NULL, 0, 0, 0, NULL, 0 },
197 	{ 0x0, "APIC", NULL, 0, 0, 0, NULL, 0 },
198 	{ 0x0, "X2APIC", NULL, 0, 0, 0, NULL, 0 },
199 	{ 0x0, "Die", NULL, 0, 0, 0, NULL, 0 },
200 	{ 0x0, "L3", NULL, 0, 0, 0, NULL, 0 },
201 	{ 0x0, "GFXAMHz", NULL, 0, 0, 0, NULL, 0 },
202 	{ 0x0, "IPC", NULL, 0, 0, 0, NULL, 0 },
203 	{ 0x0, "CoreThr", NULL, 0, 0, 0, NULL, 0 },
204 	{ 0x0, "UncMHz", NULL, 0, 0, 0, NULL, 0 },
205 	{ 0x0, "SAM%mc6", NULL, 0, 0, 0, NULL, 0 },
206 	{ 0x0, "SAMMHz", NULL, 0, 0, 0, NULL, 0 },
207 	{ 0x0, "SAMAMHz", NULL, 0, 0, 0, NULL, 0 },
208 	{ 0x0, "Die%c6", NULL, 0, 0, 0, NULL, 0 },
209 	{ 0x0, "SysWatt", NULL, 0, 0, 0, NULL, 0 },
210 	{ 0x0, "Sys_J", NULL, 0, 0, 0, NULL, 0 },
211 	{ 0x0, "NMI", NULL, 0, 0, 0, NULL, 0 },
212 	{ 0x0, "CPU%c1e", NULL, 0, 0, 0, NULL, 0 },
213 	{ 0x0, "pct_idle", NULL, 0, 0, 0, NULL, 0 },
214 	{ 0x0, "LLCMRPS", NULL, 0, 0, 0, NULL, 0 },
215 	{ 0x0, "LLC%hit", NULL, 0, 0, 0, NULL, 0 },
216 	{ 0x0, "L2MRPS", NULL, 0, 0, 0, NULL, 0 },
217 	{ 0x0, "L2%hit", NULL, 0, 0, 0, NULL, 0 },
218 };
219 
220 /* n.b. bic_names must match the order in bic[], above */
221 enum bic_names {
222 	BIC_USEC,
223 	BIC_TOD,
224 	BIC_Package,
225 	BIC_Node,
226 	BIC_Avg_MHz,
227 	BIC_Busy,
228 	BIC_Bzy_MHz,
229 	BIC_TSC_MHz,
230 	BIC_IRQ,
231 	BIC_SMI,
232 	BIC_cpuidle,
233 	BIC_CPU_c1,
234 	BIC_CPU_c3,
235 	BIC_CPU_c6,
236 	BIC_CPU_c7,
237 	BIC_ThreadC,
238 	BIC_CoreTmp,
239 	BIC_CoreCnt,
240 	BIC_PkgTmp,
241 	BIC_GFX_rc6,
242 	BIC_GFXMHz,
243 	BIC_Pkgpc2,
244 	BIC_Pkgpc3,
245 	BIC_Pkgpc6,
246 	BIC_Pkgpc7,
247 	BIC_Pkgpc8,
248 	BIC_Pkgpc9,
249 	BIC_Pkgpc10,
250 	BIC_CPU_LPI,
251 	BIC_SYS_LPI,
252 	BIC_PkgWatt,
253 	BIC_CorWatt,
254 	BIC_GFXWatt,
255 	BIC_PkgCnt,
256 	BIC_RAMWatt,
257 	BIC_PKG__,
258 	BIC_RAM__,
259 	BIC_Pkg_J,
260 	BIC_Cor_J,
261 	BIC_GFX_J,
262 	BIC_RAM_J,
263 	BIC_Mod_c6,
264 	BIC_Totl_c0,
265 	BIC_Any_c0,
266 	BIC_GFX_c0,
267 	BIC_CPUGFX,
268 	BIC_Module,
269 	BIC_Core,
270 	BIC_CPU,
271 	BIC_APIC,
272 	BIC_X2APIC,
273 	BIC_Die,
274 	BIC_L3,
275 	BIC_GFXACTMHz,
276 	BIC_IPC,
277 	BIC_CORE_THROT_CNT,
278 	BIC_UNCORE_MHZ,
279 	BIC_SAM_mc6,
280 	BIC_SAMMHz,
281 	BIC_SAMACTMHz,
282 	BIC_Diec6,
283 	BIC_SysWatt,
284 	BIC_Sys_J,
285 	BIC_NMI,
286 	BIC_CPU_c1e,
287 	BIC_pct_idle,
288 	BIC_LLC_MRPS,
289 	BIC_LLC_HIT,
290 	BIC_L2_MRPS,
291 	BIC_L2_HIT,
292 	MAX_BIC
293 };
294 
print_bic_set(char * s,cpu_set_t * set)295 void print_bic_set(char *s, cpu_set_t *set)
296 {
297 	int i;
298 
299 	assert(MAX_BIC < CPU_SETSIZE);
300 
301 	printf("%s:", s);
302 
303 	for (i = 0; i < MAX_BIC; ++i) {
304 
305 		if (CPU_ISSET(i, set))
306 			printf(" %s", bic[i].name);
307 	}
308 	putchar('\n');
309 }
310 
311 static cpu_set_t bic_group_topology;
312 static cpu_set_t bic_group_thermal_pwr;
313 static cpu_set_t bic_group_frequency;
314 static cpu_set_t bic_group_hw_idle;
315 static cpu_set_t bic_group_sw_idle;
316 static cpu_set_t bic_group_idle;
317 static cpu_set_t bic_group_cache;
318 static cpu_set_t bic_group_other;
319 static cpu_set_t bic_group_disabled_by_default;
320 static cpu_set_t bic_enabled;
321 static cpu_set_t bic_present;
322 
323 /* modify */
324 #define BIC_INIT(set) CPU_ZERO(set)
325 
326 #define SET_BIC(COUNTER_NUMBER, set) CPU_SET(COUNTER_NUMBER, set)
327 #define CLR_BIC(COUNTER_NUMBER, set) CPU_CLR(COUNTER_NUMBER, set)
328 
329 #define BIC_PRESENT(COUNTER_NUMBER) SET_BIC(COUNTER_NUMBER, &bic_present)
330 #define BIC_NOT_PRESENT(COUNTER_NUMBER) CPU_CLR(COUNTER_NUMBER, &bic_present)
331 
332 /* test */
333 #define BIC_IS_ENABLED(COUNTER_NUMBER) CPU_ISSET(COUNTER_NUMBER, &bic_enabled)
334 #define DO_BIC_READ(COUNTER_NUMBER) CPU_ISSET(COUNTER_NUMBER, &bic_present)
335 #define DO_BIC(COUNTER_NUMBER) (CPU_ISSET(COUNTER_NUMBER, &bic_enabled) && CPU_ISSET(COUNTER_NUMBER, &bic_present))
336 
bic_set_all(cpu_set_t * set)337 static void bic_set_all(cpu_set_t *set)
338 {
339 	int i;
340 
341 	assert(MAX_BIC < CPU_SETSIZE);
342 
343 	for (i = 0; i < MAX_BIC; ++i)
344 		SET_BIC(i, set);
345 }
346 
347 /*
348  * bic_clear_bits()
349  * clear all the bits from "clr" in "dst"
350  */
bic_clear_bits(cpu_set_t * dst,cpu_set_t * clr)351 static void bic_clear_bits(cpu_set_t *dst, cpu_set_t *clr)
352 {
353 	int i;
354 
355 	assert(MAX_BIC < CPU_SETSIZE);
356 
357 	for (i = 0; i < MAX_BIC; ++i)
358 		if (CPU_ISSET(i, clr))
359 			CLR_BIC(i, dst);
360 }
361 
bic_groups_init(void)362 static void bic_groups_init(void)
363 {
364 	BIC_INIT(&bic_group_topology);
365 	SET_BIC(BIC_Package, &bic_group_topology);
366 	SET_BIC(BIC_Node, &bic_group_topology);
367 	SET_BIC(BIC_CoreCnt, &bic_group_topology);
368 	SET_BIC(BIC_PkgCnt, &bic_group_topology);
369 	SET_BIC(BIC_Module, &bic_group_topology);
370 	SET_BIC(BIC_Core, &bic_group_topology);
371 	SET_BIC(BIC_CPU, &bic_group_topology);
372 	SET_BIC(BIC_Die, &bic_group_topology);
373 	SET_BIC(BIC_L3, &bic_group_topology);
374 
375 	BIC_INIT(&bic_group_thermal_pwr);
376 	SET_BIC(BIC_CoreTmp, &bic_group_thermal_pwr);
377 	SET_BIC(BIC_PkgTmp, &bic_group_thermal_pwr);
378 	SET_BIC(BIC_PkgWatt, &bic_group_thermal_pwr);
379 	SET_BIC(BIC_CorWatt, &bic_group_thermal_pwr);
380 	SET_BIC(BIC_GFXWatt, &bic_group_thermal_pwr);
381 	SET_BIC(BIC_RAMWatt, &bic_group_thermal_pwr);
382 	SET_BIC(BIC_PKG__, &bic_group_thermal_pwr);
383 	SET_BIC(BIC_RAM__, &bic_group_thermal_pwr);
384 	SET_BIC(BIC_SysWatt, &bic_group_thermal_pwr);
385 
386 	BIC_INIT(&bic_group_frequency);
387 	SET_BIC(BIC_Avg_MHz, &bic_group_frequency);
388 	SET_BIC(BIC_Busy, &bic_group_frequency);
389 	SET_BIC(BIC_Bzy_MHz, &bic_group_frequency);
390 	SET_BIC(BIC_TSC_MHz, &bic_group_frequency);
391 	SET_BIC(BIC_GFXMHz, &bic_group_frequency);
392 	SET_BIC(BIC_GFXACTMHz, &bic_group_frequency);
393 	SET_BIC(BIC_SAMMHz, &bic_group_frequency);
394 	SET_BIC(BIC_SAMACTMHz, &bic_group_frequency);
395 	SET_BIC(BIC_UNCORE_MHZ, &bic_group_frequency);
396 
397 	BIC_INIT(&bic_group_hw_idle);
398 	SET_BIC(BIC_Busy, &bic_group_hw_idle);
399 	SET_BIC(BIC_CPU_c1, &bic_group_hw_idle);
400 	SET_BIC(BIC_CPU_c3, &bic_group_hw_idle);
401 	SET_BIC(BIC_CPU_c6, &bic_group_hw_idle);
402 	SET_BIC(BIC_CPU_c7, &bic_group_hw_idle);
403 	SET_BIC(BIC_GFX_rc6, &bic_group_hw_idle);
404 	SET_BIC(BIC_Pkgpc2, &bic_group_hw_idle);
405 	SET_BIC(BIC_Pkgpc3, &bic_group_hw_idle);
406 	SET_BIC(BIC_Pkgpc6, &bic_group_hw_idle);
407 	SET_BIC(BIC_Pkgpc7, &bic_group_hw_idle);
408 	SET_BIC(BIC_Pkgpc8, &bic_group_hw_idle);
409 	SET_BIC(BIC_Pkgpc9, &bic_group_hw_idle);
410 	SET_BIC(BIC_Pkgpc10, &bic_group_hw_idle);
411 	SET_BIC(BIC_CPU_LPI, &bic_group_hw_idle);
412 	SET_BIC(BIC_SYS_LPI, &bic_group_hw_idle);
413 	SET_BIC(BIC_Mod_c6, &bic_group_hw_idle);
414 	SET_BIC(BIC_Totl_c0, &bic_group_hw_idle);
415 	SET_BIC(BIC_Any_c0, &bic_group_hw_idle);
416 	SET_BIC(BIC_GFX_c0, &bic_group_hw_idle);
417 	SET_BIC(BIC_CPUGFX, &bic_group_hw_idle);
418 	SET_BIC(BIC_SAM_mc6, &bic_group_hw_idle);
419 	SET_BIC(BIC_Diec6, &bic_group_hw_idle);
420 
421 	BIC_INIT(&bic_group_sw_idle);
422 	SET_BIC(BIC_Busy, &bic_group_sw_idle);
423 	SET_BIC(BIC_cpuidle, &bic_group_sw_idle);
424 	SET_BIC(BIC_pct_idle, &bic_group_sw_idle);
425 
426 	BIC_INIT(&bic_group_idle);
427 
428 	CPU_OR(&bic_group_idle, &bic_group_idle, &bic_group_hw_idle);
429 	SET_BIC(BIC_pct_idle, &bic_group_idle);
430 
431 	BIC_INIT(&bic_group_cache);
432 	SET_BIC(BIC_LLC_MRPS, &bic_group_cache);
433 	SET_BIC(BIC_LLC_HIT, &bic_group_cache);
434 	SET_BIC(BIC_L2_MRPS, &bic_group_cache);
435 	SET_BIC(BIC_L2_HIT, &bic_group_cache);
436 
437 	BIC_INIT(&bic_group_other);
438 	SET_BIC(BIC_IRQ, &bic_group_other);
439 	SET_BIC(BIC_NMI, &bic_group_other);
440 	SET_BIC(BIC_SMI, &bic_group_other);
441 	SET_BIC(BIC_ThreadC, &bic_group_other);
442 	SET_BIC(BIC_CoreTmp, &bic_group_other);
443 	SET_BIC(BIC_IPC, &bic_group_other);
444 
445 	BIC_INIT(&bic_group_disabled_by_default);
446 	SET_BIC(BIC_USEC, &bic_group_disabled_by_default);
447 	SET_BIC(BIC_TOD, &bic_group_disabled_by_default);
448 	SET_BIC(BIC_cpuidle, &bic_group_disabled_by_default);
449 	SET_BIC(BIC_APIC, &bic_group_disabled_by_default);
450 	SET_BIC(BIC_X2APIC, &bic_group_disabled_by_default);
451 
452 	BIC_INIT(&bic_enabled);
453 	bic_set_all(&bic_enabled);
454 	bic_clear_bits(&bic_enabled, &bic_group_disabled_by_default);
455 
456 	BIC_INIT(&bic_present);
457 	SET_BIC(BIC_USEC, &bic_present);
458 	SET_BIC(BIC_TOD, &bic_present);
459 	SET_BIC(BIC_cpuidle, &bic_present);
460 	SET_BIC(BIC_APIC, &bic_present);
461 	SET_BIC(BIC_X2APIC, &bic_present);
462 	SET_BIC(BIC_pct_idle, &bic_present);
463 }
464 
465 /*
466  * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
467  * If you change the values, note they are used both in comparisons
468  * (>= PCL__7) and to index pkg_cstate_limit_strings[].
469  */
470 #define PCLUKN 0		/* Unknown */
471 #define PCLRSV 1		/* Reserved */
472 #define PCL__0 2		/* PC0 */
473 #define PCL__1 3		/* PC1 */
474 #define PCL__2 4		/* PC2 */
475 #define PCL__3 5		/* PC3 */
476 #define PCL__4 6		/* PC4 */
477 #define PCL__6 7		/* PC6 */
478 #define PCL_6N 8		/* PC6 No Retention */
479 #define PCL_6R 9		/* PC6 Retention */
480 #define PCL__7 10		/* PC7 */
481 #define PCL_7S 11		/* PC7 Shrink */
482 #define PCL__8 12		/* PC8 */
483 #define PCL__9 13		/* PC9 */
484 #define PCL_10 14		/* PC10 */
485 #define PCLUNL 15		/* Unlimited */
486 
487 char *proc_stat = "/proc/stat";
488 FILE *outf;
489 int *fd_percpu;
490 int *fd_instr_count_percpu;
491 int *fd_llc_percpu;
492 int *fd_l2_percpu;
493 struct timeval interval_tv = { 5, 0 };
494 struct timespec interval_ts = { 5, 0 };
495 
496 unsigned int num_iterations;
497 unsigned int header_iterations;
498 unsigned int debug;
499 unsigned int quiet;
500 unsigned int shown;
501 unsigned int sums_need_wide_columns;
502 unsigned int rapl_joules;
503 unsigned int valid_rapl_msrs;
504 unsigned int summary_only;
505 unsigned int list_header_only;
506 unsigned int dump_only;
507 unsigned int force_load;
508 unsigned int cpuid_has_aperf_mperf;
509 unsigned int cpuid_has_hv;
510 unsigned int has_aperf_access;
511 unsigned int has_epb;
512 unsigned int has_turbo;
513 unsigned int is_hybrid;
514 unsigned int units = 1000000;	/* MHz etc */
515 unsigned int genuine_intel;
516 unsigned int authentic_amd;
517 unsigned int hygon_genuine;
518 unsigned int max_level, max_extended_level;
519 unsigned int has_invariant_tsc;
520 unsigned int aperf_mperf_multiplier = 1;
521 double bclk;
522 double base_hz;
523 unsigned int has_base_hz;
524 double tsc_tweak = 1.0;
525 unsigned int show_pkg_only;
526 unsigned int show_core_only;
527 char *output_buffer, *outp;
528 unsigned int do_dts;
529 unsigned int do_ptm;
530 unsigned int do_ipc;
531 unsigned long long cpuidle_cur_cpu_lpi_us;
532 unsigned long long cpuidle_cur_sys_lpi_us;
533 unsigned int tj_max;
534 unsigned int tj_max_override;
535 double rapl_power_units, rapl_time_units;
536 double rapl_dram_energy_units, rapl_energy_units, rapl_psys_energy_units;
537 double rapl_joule_counter_range;
538 unsigned int crystal_hz;
539 unsigned long long tsc_hz;
540 int master_cpu;
541 unsigned int has_hwp;		/* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
542 			/* IA32_HWP_REQUEST, IA32_HWP_STATUS */
543 unsigned int has_hwp_notify;	/* IA32_HWP_INTERRUPT */
544 unsigned int has_hwp_activity_window;	/* IA32_HWP_REQUEST[bits 41:32] */
545 unsigned int has_hwp_epp;	/* IA32_HWP_REQUEST[bits 31:24] */
546 unsigned int has_hwp_pkg;	/* IA32_HWP_REQUEST_PKG */
547 unsigned int first_counter_read = 1;
548 
549 static struct timeval procsysfs_tv_begin;
550 
551 int ignore_stdin;
552 bool no_msr;
553 bool no_perf;
554 
555 enum gfx_sysfs_idx {
556 	GFX_rc6,
557 	GFX_MHz,
558 	GFX_ACTMHz,
559 	SAM_mc6,
560 	SAM_MHz,
561 	SAM_ACTMHz,
562 	GFX_MAX
563 };
564 
565 struct gfx_sysfs_info {
566 	FILE *fp;
567 	unsigned int val;
568 	unsigned long long val_ull;
569 };
570 
571 static struct gfx_sysfs_info gfx_info[GFX_MAX];
572 
573 int get_msr(int cpu, off_t offset, unsigned long long *msr);
574 int add_counter(unsigned int msr_num, char *path, char *name,
575 		unsigned int width, enum counter_scope scope, enum counter_type type, enum counter_format format, int flags, int package_num);
576 
577 /* Model specific support Start */
578 
579 /* List of features that may diverge among different platforms */
580 struct platform_features {
581 	bool has_msr_misc_feature_control;	/* MSR_MISC_FEATURE_CONTROL */
582 	bool has_msr_misc_pwr_mgmt;	/* MSR_MISC_PWR_MGMT */
583 	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 */
584 	bool has_config_tdp;	/* MSR_CONFIG_TDP_NOMINAL/LEVEL_1/LEVEL_2/CONTROL, MSR_TURBO_ACTIVATION_RATIO */
585 	int bclk_freq;		/* CPU base clock */
586 	int crystal_freq;	/* Crystal clock to use when not available from CPUID.15 */
587 	int supported_cstates;	/* Core cstates and Package cstates supported */
588 	int cst_limit;		/* MSR_PKG_CST_CONFIG_CONTROL */
589 	bool has_cst_auto_convension;	/* AUTOMATIC_CSTATE_CONVERSION bit in MSR_PKG_CST_CONFIG_CONTROL */
590 	bool has_irtl_msrs;	/* MSR_PKGC3/PKGC6/PKGC7/PKGC8/PKGC9/PKGC10_IRTL */
591 	bool has_msr_core_c1_res;	/* MSR_CORE_C1_RES */
592 	bool has_msr_module_c6_res_ms;	/* MSR_MODULE_C6_RES_MS */
593 	bool has_msr_c6_demotion_policy_config;	/* MSR_CC6_DEMOTION_POLICY_CONFIG/MSR_MC6_DEMOTION_POLICY_CONFIG */
594 	bool has_msr_atom_pkg_c6_residency;	/* MSR_ATOM_PKG_C6_RESIDENCY */
595 	bool has_msr_knl_core_c6_residency;	/* MSR_KNL_CORE_C6_RESIDENCY */
596 	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 */
597 	bool has_cst_prewake_bit;	/* Cstate prewake bit in MSR_IA32_POWER_CTL */
598 	int trl_msrs;		/* MSR_TURBO_RATIO_LIMIT/LIMIT1/LIMIT2/SECONDARY, Atom TRL MSRs */
599 	int plr_msrs;		/* MSR_CORE/GFX/RING_PERF_LIMIT_REASONS */
600 	int plat_rapl_msrs;	/* RAPL PKG/DRAM/CORE/GFX MSRs, AMD RAPL MSRs */
601 	bool has_per_core_rapl;	/* Indicates cores energy collection is per-core, not per-package. AMD specific for now */
602 	bool has_rapl_divisor;	/* Divisor for Energy unit raw value from MSR_RAPL_POWER_UNIT */
603 	bool has_fixed_rapl_unit;	/* Fixed Energy Unit used for DRAM RAPL Domain */
604 	bool has_fixed_rapl_psys_unit;	/* Fixed Energy Unit used for PSYS RAPL Domain */
605 	int rapl_quirk_tdp;	/* Hardcoded TDP value when cannot be retrieved from hardware */
606 	int tcc_offset_bits;	/* TCC Offset bits in MSR_IA32_TEMPERATURE_TARGET */
607 	bool enable_tsc_tweak;	/* Use CPU Base freq instead of TSC freq for aperf/mperf counter */
608 	bool need_perf_multiplier;	/* mperf/aperf multiplier */
609 };
610 
611 struct platform_data {
612 	unsigned int vfm;
613 	const struct platform_features *features;
614 };
615 
616 /* For BCLK */
617 enum bclk_freq {
618 	BCLK_100MHZ = 1,
619 	BCLK_133MHZ,
620 	BCLK_SLV,
621 };
622 
623 #define SLM_BCLK_FREQS 5
624 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0 };
625 
slm_bclk(void)626 double slm_bclk(void)
627 {
628 	unsigned long long msr = 3;
629 	unsigned int i;
630 	double freq;
631 
632 	if (get_msr(master_cpu, MSR_FSB_FREQ, &msr))
633 		fprintf(outf, "SLM BCLK: unknown\n");
634 
635 	i = msr & 0xf;
636 	if (i >= SLM_BCLK_FREQS) {
637 		fprintf(outf, "SLM BCLK[%d] invalid\n", i);
638 		i = 3;
639 	}
640 	freq = slm_freq_table[i];
641 
642 	if (!quiet)
643 		fprintf(outf, "SLM BCLK: %.1f Mhz\n", freq);
644 
645 	return freq;
646 }
647 
648 /* For Package cstate limit */
649 enum package_cstate_limit {
650 	CST_LIMIT_NHM = 1,
651 	CST_LIMIT_SNB,
652 	CST_LIMIT_HSW,
653 	CST_LIMIT_SKX,
654 	CST_LIMIT_ICX,
655 	CST_LIMIT_SLV,
656 	CST_LIMIT_AMT,
657 	CST_LIMIT_KNL,
658 	CST_LIMIT_GMT,
659 };
660 
661 /* For Turbo Ratio Limit MSRs */
662 enum turbo_ratio_limit_msrs {
663 	TRL_BASE = BIT(0),
664 	TRL_LIMIT1 = BIT(1),
665 	TRL_LIMIT2 = BIT(2),
666 	TRL_ATOM = BIT(3),
667 	TRL_KNL = BIT(4),
668 	TRL_CORECOUNT = BIT(5),
669 };
670 
671 /* For Perf Limit Reason MSRs */
672 enum perf_limit_reason_msrs {
673 	PLR_CORE = BIT(0),
674 	PLR_GFX = BIT(1),
675 	PLR_RING = BIT(2),
676 };
677 
678 /* For RAPL MSRs */
679 enum rapl_msrs {
680 	RAPL_PKG_POWER_LIMIT = BIT(0),	/* 0x610 MSR_PKG_POWER_LIMIT */
681 	RAPL_PKG_ENERGY_STATUS = BIT(1),	/* 0x611 MSR_PKG_ENERGY_STATUS */
682 	RAPL_PKG_PERF_STATUS = BIT(2),	/* 0x613 MSR_PKG_PERF_STATUS */
683 	RAPL_PKG_POWER_INFO = BIT(3),	/* 0x614 MSR_PKG_POWER_INFO */
684 	RAPL_DRAM_POWER_LIMIT = BIT(4),	/* 0x618 MSR_DRAM_POWER_LIMIT */
685 	RAPL_DRAM_ENERGY_STATUS = BIT(5),	/* 0x619 MSR_DRAM_ENERGY_STATUS */
686 	RAPL_DRAM_PERF_STATUS = BIT(6),	/* 0x61b MSR_DRAM_PERF_STATUS */
687 	RAPL_DRAM_POWER_INFO = BIT(7),	/* 0x61c MSR_DRAM_POWER_INFO */
688 	RAPL_CORE_POWER_LIMIT = BIT(8),	/* 0x638 MSR_PP0_POWER_LIMIT */
689 	RAPL_CORE_ENERGY_STATUS = BIT(9),	/* 0x639 MSR_PP0_ENERGY_STATUS */
690 	RAPL_CORE_POLICY = BIT(10),	/* 0x63a MSR_PP0_POLICY */
691 	RAPL_GFX_POWER_LIMIT = BIT(11),	/* 0x640 MSR_PP1_POWER_LIMIT */
692 	RAPL_GFX_ENERGY_STATUS = BIT(12),	/* 0x641 MSR_PP1_ENERGY_STATUS */
693 	RAPL_GFX_POLICY = BIT(13),	/* 0x642 MSR_PP1_POLICY */
694 	RAPL_AMD_PWR_UNIT = BIT(14),	/* 0xc0010299 MSR_AMD_RAPL_POWER_UNIT */
695 	RAPL_AMD_CORE_ENERGY_STAT = BIT(15),	/* 0xc001029a MSR_AMD_CORE_ENERGY_STATUS */
696 	RAPL_AMD_PKG_ENERGY_STAT = BIT(16),	/* 0xc001029b MSR_AMD_PKG_ENERGY_STATUS */
697 	RAPL_PLATFORM_ENERGY_LIMIT = BIT(17),	/* 0x64c MSR_PLATFORM_ENERGY_LIMIT */
698 	RAPL_PLATFORM_ENERGY_STATUS = BIT(18),	/* 0x64d MSR_PLATFORM_ENERGY_STATUS */
699 };
700 
701 #define RAPL_PKG	(RAPL_PKG_ENERGY_STATUS | RAPL_PKG_POWER_LIMIT)
702 #define RAPL_DRAM	(RAPL_DRAM_ENERGY_STATUS | RAPL_DRAM_POWER_LIMIT)
703 #define RAPL_CORE	(RAPL_CORE_ENERGY_STATUS | RAPL_CORE_POWER_LIMIT)
704 #define RAPL_GFX	(RAPL_GFX_POWER_LIMIT | RAPL_GFX_ENERGY_STATUS)
705 #define RAPL_PSYS	(RAPL_PLATFORM_ENERGY_STATUS | RAPL_PLATFORM_ENERGY_LIMIT)
706 
707 #define RAPL_PKG_ALL	(RAPL_PKG | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO)
708 #define RAPL_DRAM_ALL	(RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_DRAM_POWER_INFO)
709 #define RAPL_CORE_ALL	(RAPL_CORE | RAPL_CORE_POLICY)
710 #define RAPL_GFX_ALL	(RAPL_GFX | RAPL_GFX_POLICY)
711 
712 #define RAPL_AMD_F17H	(RAPL_AMD_PWR_UNIT | RAPL_AMD_CORE_ENERGY_STAT | RAPL_AMD_PKG_ENERGY_STAT)
713 
714 /* For Cstates */
715 enum cstates {
716 	CC1 = BIT(0),
717 	CC3 = BIT(1),
718 	CC6 = BIT(2),
719 	CC7 = BIT(3),
720 	PC2 = BIT(4),
721 	PC3 = BIT(5),
722 	PC6 = BIT(6),
723 	PC7 = BIT(7),
724 	PC8 = BIT(8),
725 	PC9 = BIT(9),
726 	PC10 = BIT(10),
727 };
728 
729 static const struct platform_features nhm_features = {
730 	.has_msr_misc_pwr_mgmt = 1,
731 	.has_nhm_msrs = 1,
732 	.bclk_freq = BCLK_133MHZ,
733 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
734 	.cst_limit = CST_LIMIT_NHM,
735 	.trl_msrs = TRL_BASE,
736 };
737 
738 static const struct platform_features nhx_features = {
739 	.has_msr_misc_pwr_mgmt = 1,
740 	.has_nhm_msrs = 1,
741 	.bclk_freq = BCLK_133MHZ,
742 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
743 	.cst_limit = CST_LIMIT_NHM,
744 };
745 
746 static const struct platform_features snb_features = {
747 	.has_msr_misc_feature_control = 1,
748 	.has_msr_misc_pwr_mgmt = 1,
749 	.has_nhm_msrs = 1,
750 	.bclk_freq = BCLK_100MHZ,
751 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
752 	.cst_limit = CST_LIMIT_SNB,
753 	.has_irtl_msrs = 1,
754 	.trl_msrs = TRL_BASE,
755 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
756 };
757 
758 static const struct platform_features snx_features = {
759 	.has_msr_misc_feature_control = 1,
760 	.has_msr_misc_pwr_mgmt = 1,
761 	.has_nhm_msrs = 1,
762 	.bclk_freq = BCLK_100MHZ,
763 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
764 	.cst_limit = CST_LIMIT_SNB,
765 	.has_irtl_msrs = 1,
766 	.trl_msrs = TRL_BASE,
767 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM_ALL,
768 };
769 
770 static const struct platform_features ivb_features = {
771 	.has_msr_misc_feature_control = 1,
772 	.has_msr_misc_pwr_mgmt = 1,
773 	.has_nhm_msrs = 1,
774 	.has_config_tdp = 1,
775 	.bclk_freq = BCLK_100MHZ,
776 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
777 	.cst_limit = CST_LIMIT_SNB,
778 	.has_irtl_msrs = 1,
779 	.trl_msrs = TRL_BASE,
780 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
781 };
782 
783 static const struct platform_features ivx_features = {
784 	.has_msr_misc_feature_control = 1,
785 	.has_msr_misc_pwr_mgmt = 1,
786 	.has_nhm_msrs = 1,
787 	.bclk_freq = BCLK_100MHZ,
788 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
789 	.cst_limit = CST_LIMIT_SNB,
790 	.has_irtl_msrs = 1,
791 	.trl_msrs = TRL_BASE | TRL_LIMIT1,
792 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM_ALL,
793 };
794 
795 static const struct platform_features hsw_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 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
802 	.cst_limit = CST_LIMIT_HSW,
803 	.has_irtl_msrs = 1,
804 	.trl_msrs = TRL_BASE,
805 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
806 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
807 };
808 
809 static const struct platform_features hsx_features = {
810 	.has_msr_misc_feature_control = 1,
811 	.has_msr_misc_pwr_mgmt = 1,
812 	.has_nhm_msrs = 1,
813 	.has_config_tdp = 1,
814 	.bclk_freq = BCLK_100MHZ,
815 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
816 	.cst_limit = CST_LIMIT_HSW,
817 	.has_irtl_msrs = 1,
818 	.trl_msrs = TRL_BASE | TRL_LIMIT1 | TRL_LIMIT2,
819 	.plr_msrs = PLR_CORE | PLR_RING,
820 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
821 	.has_fixed_rapl_unit = 1,
822 };
823 
824 static const struct platform_features hswl_features = {
825 	.has_msr_misc_feature_control = 1,
826 	.has_msr_misc_pwr_mgmt = 1,
827 	.has_nhm_msrs = 1,
828 	.has_config_tdp = 1,
829 	.bclk_freq = BCLK_100MHZ,
830 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
831 	.cst_limit = CST_LIMIT_HSW,
832 	.has_irtl_msrs = 1,
833 	.trl_msrs = TRL_BASE,
834 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
835 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
836 };
837 
838 static const struct platform_features hswg_features = {
839 	.has_msr_misc_feature_control = 1,
840 	.has_msr_misc_pwr_mgmt = 1,
841 	.has_nhm_msrs = 1,
842 	.has_config_tdp = 1,
843 	.bclk_freq = BCLK_100MHZ,
844 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
845 	.cst_limit = CST_LIMIT_HSW,
846 	.has_irtl_msrs = 1,
847 	.trl_msrs = TRL_BASE,
848 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
849 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
850 };
851 
852 static const struct platform_features bdw_features = {
853 	.has_msr_misc_feature_control = 1,
854 	.has_msr_misc_pwr_mgmt = 1,
855 	.has_nhm_msrs = 1,
856 	.has_config_tdp = 1,
857 	.bclk_freq = BCLK_100MHZ,
858 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
859 	.cst_limit = CST_LIMIT_HSW,
860 	.has_irtl_msrs = 1,
861 	.trl_msrs = TRL_BASE,
862 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
863 };
864 
865 static const struct platform_features bdwg_features = {
866 	.has_msr_misc_feature_control = 1,
867 	.has_msr_misc_pwr_mgmt = 1,
868 	.has_nhm_msrs = 1,
869 	.has_config_tdp = 1,
870 	.bclk_freq = BCLK_100MHZ,
871 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
872 	.cst_limit = CST_LIMIT_HSW,
873 	.has_irtl_msrs = 1,
874 	.trl_msrs = TRL_BASE,
875 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
876 };
877 
878 static const struct platform_features bdx_features = {
879 	.has_msr_misc_feature_control = 1,
880 	.has_msr_misc_pwr_mgmt = 1,
881 	.has_nhm_msrs = 1,
882 	.has_config_tdp = 1,
883 	.bclk_freq = BCLK_100MHZ,
884 	.supported_cstates = CC1 | CC3 | CC6 | PC2 | PC3 | PC6,
885 	.cst_limit = CST_LIMIT_HSW,
886 	.has_irtl_msrs = 1,
887 	.has_cst_auto_convension = 1,
888 	.trl_msrs = TRL_BASE,
889 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
890 	.has_fixed_rapl_unit = 1,
891 };
892 
893 static const struct platform_features skl_features = {
894 	.has_msr_misc_feature_control = 1,
895 	.has_msr_misc_pwr_mgmt = 1,
896 	.has_nhm_msrs = 1,
897 	.has_config_tdp = 1,
898 	.bclk_freq = BCLK_100MHZ,
899 	.crystal_freq = 24000000,
900 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
901 	.cst_limit = CST_LIMIT_HSW,
902 	.has_irtl_msrs = 1,
903 	.has_ext_cst_msrs = 1,
904 	.trl_msrs = TRL_BASE,
905 	.tcc_offset_bits = 6,
906 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX | RAPL_PSYS,
907 	.enable_tsc_tweak = 1,
908 };
909 
910 static const struct platform_features cnl_features = {
911 	.has_msr_misc_feature_control = 1,
912 	.has_msr_misc_pwr_mgmt = 1,
913 	.has_nhm_msrs = 1,
914 	.has_config_tdp = 1,
915 	.bclk_freq = BCLK_100MHZ,
916 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
917 	.cst_limit = CST_LIMIT_HSW,
918 	.has_irtl_msrs = 1,
919 	.has_msr_core_c1_res = 1,
920 	.has_ext_cst_msrs = 1,
921 	.trl_msrs = TRL_BASE,
922 	.tcc_offset_bits = 6,
923 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX | RAPL_PSYS,
924 	.enable_tsc_tweak = 1,
925 };
926 
927 /* Copied from cnl_features, with PC7/PC9 removed */
928 static const struct platform_features adl_features = {
929 	.has_msr_misc_feature_control	= cnl_features.has_msr_misc_feature_control,
930 	.has_msr_misc_pwr_mgmt		= cnl_features.has_msr_misc_pwr_mgmt,
931 	.has_nhm_msrs			= cnl_features.has_nhm_msrs,
932 	.has_config_tdp			= cnl_features.has_config_tdp,
933 	.bclk_freq			= cnl_features.bclk_freq,
934 	.supported_cstates		= CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC8 | PC10,
935 	.cst_limit			= cnl_features.cst_limit,
936 	.has_irtl_msrs			= cnl_features.has_irtl_msrs,
937 	.has_msr_core_c1_res		= cnl_features.has_msr_core_c1_res,
938 	.has_ext_cst_msrs		= cnl_features.has_ext_cst_msrs,
939 	.trl_msrs			= cnl_features.trl_msrs,
940 	.tcc_offset_bits		= cnl_features.tcc_offset_bits,
941 	.plat_rapl_msrs			= cnl_features.plat_rapl_msrs,
942 	.enable_tsc_tweak		= cnl_features.enable_tsc_tweak,
943 };
944 
945 /* Copied from adl_features, with PC3/PC8 removed */
946 static const struct platform_features lnl_features = {
947 	.has_msr_misc_feature_control	= adl_features.has_msr_misc_feature_control,
948 	.has_msr_misc_pwr_mgmt		= adl_features.has_msr_misc_pwr_mgmt,
949 	.has_nhm_msrs			= adl_features.has_nhm_msrs,
950 	.has_config_tdp			= adl_features.has_config_tdp,
951 	.bclk_freq			= adl_features.bclk_freq,
952 	.supported_cstates		= CC1 | CC6 | CC7 | PC2 | PC6 | PC10,
953 	.cst_limit			= adl_features.cst_limit,
954 	.has_irtl_msrs			= adl_features.has_irtl_msrs,
955 	.has_msr_core_c1_res		= adl_features.has_msr_core_c1_res,
956 	.has_ext_cst_msrs		= adl_features.has_ext_cst_msrs,
957 	.trl_msrs			= adl_features.trl_msrs,
958 	.tcc_offset_bits		= adl_features.tcc_offset_bits,
959 	.plat_rapl_msrs			= adl_features.plat_rapl_msrs,
960 	.enable_tsc_tweak		= adl_features.enable_tsc_tweak,
961 };
962 
963 static const struct platform_features skx_features = {
964 	.has_msr_misc_feature_control = 1,
965 	.has_msr_misc_pwr_mgmt = 1,
966 	.has_nhm_msrs = 1,
967 	.has_config_tdp = 1,
968 	.bclk_freq = BCLK_100MHZ,
969 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
970 	.cst_limit = CST_LIMIT_SKX,
971 	.has_irtl_msrs = 1,
972 	.has_cst_auto_convension = 1,
973 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
974 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
975 	.has_fixed_rapl_unit = 1,
976 };
977 
978 static const struct platform_features icx_features = {
979 	.has_msr_misc_feature_control = 1,
980 	.has_msr_misc_pwr_mgmt = 1,
981 	.has_nhm_msrs = 1,
982 	.has_config_tdp = 1,
983 	.bclk_freq = BCLK_100MHZ,
984 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
985 	.cst_limit = CST_LIMIT_ICX,
986 	.has_msr_core_c1_res = 1,
987 	.has_irtl_msrs = 1,
988 	.has_cst_prewake_bit = 1,
989 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
990 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
991 	.has_fixed_rapl_unit = 1,
992 };
993 
994 static const struct platform_features spr_features = {
995 	.has_msr_misc_feature_control = 1,
996 	.has_msr_misc_pwr_mgmt = 1,
997 	.has_nhm_msrs = 1,
998 	.has_config_tdp = 1,
999 	.bclk_freq = BCLK_100MHZ,
1000 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
1001 	.cst_limit = CST_LIMIT_SKX,
1002 	.has_msr_core_c1_res = 1,
1003 	.has_irtl_msrs = 1,
1004 	.has_cst_prewake_bit = 1,
1005 	.has_fixed_rapl_psys_unit = 1,
1006 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1007 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
1008 };
1009 
1010 static const struct platform_features dmr_features = {
1011 	.has_msr_misc_feature_control	= spr_features.has_msr_misc_feature_control,
1012 	.has_msr_misc_pwr_mgmt		= spr_features.has_msr_misc_pwr_mgmt,
1013 	.has_nhm_msrs			= spr_features.has_nhm_msrs,
1014 	.bclk_freq			= spr_features.bclk_freq,
1015 	.supported_cstates		= spr_features.supported_cstates,
1016 	.cst_limit			= spr_features.cst_limit,
1017 	.has_msr_core_c1_res		= spr_features.has_msr_core_c1_res,
1018 	.has_cst_prewake_bit		= spr_features.has_cst_prewake_bit,
1019 	.has_fixed_rapl_psys_unit	= spr_features.has_fixed_rapl_psys_unit,
1020 	.trl_msrs			= spr_features.trl_msrs,
1021 	.has_msr_module_c6_res_ms	= 1,	/* DMR has Dual-Core-Module and MC6 MSR */
1022 	.plat_rapl_msrs			= 0,	/* DMR does not have RAPL MSRs */
1023 	.plr_msrs			= 0,	/* DMR does not have PLR  MSRs */
1024 	.has_irtl_msrs			= 0,	/* DMR does not have IRTL MSRs */
1025 	.has_config_tdp			= 0,	/* DMR does not have CTDP MSRs */
1026 };
1027 
1028 static const struct platform_features srf_features = {
1029 	.has_msr_misc_feature_control = 1,
1030 	.has_msr_misc_pwr_mgmt = 1,
1031 	.has_nhm_msrs = 1,
1032 	.has_config_tdp = 1,
1033 	.bclk_freq = BCLK_100MHZ,
1034 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
1035 	.cst_limit = CST_LIMIT_SKX,
1036 	.has_msr_core_c1_res = 1,
1037 	.has_msr_module_c6_res_ms = 1,
1038 	.has_irtl_msrs = 1,
1039 	.has_cst_prewake_bit = 1,
1040 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1041 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
1042 };
1043 
1044 static const struct platform_features grr_features = {
1045 	.has_msr_misc_feature_control = 1,
1046 	.has_msr_misc_pwr_mgmt = 1,
1047 	.has_nhm_msrs = 1,
1048 	.has_config_tdp = 1,
1049 	.bclk_freq = BCLK_100MHZ,
1050 	.supported_cstates = CC1 | CC6,
1051 	.cst_limit = CST_LIMIT_SKX,
1052 	.has_msr_core_c1_res = 1,
1053 	.has_msr_module_c6_res_ms = 1,
1054 	.has_irtl_msrs = 1,
1055 	.has_cst_prewake_bit = 1,
1056 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1057 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
1058 };
1059 
1060 static const struct platform_features slv_features = {
1061 	.has_nhm_msrs = 1,
1062 	.bclk_freq = BCLK_SLV,
1063 	.supported_cstates = CC1 | CC6 | PC6,
1064 	.cst_limit = CST_LIMIT_SLV,
1065 	.has_msr_core_c1_res = 1,
1066 	.has_msr_module_c6_res_ms = 1,
1067 	.has_msr_c6_demotion_policy_config = 1,
1068 	.has_msr_atom_pkg_c6_residency = 1,
1069 	.trl_msrs = TRL_ATOM,
1070 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE,
1071 	.has_rapl_divisor = 1,
1072 	.rapl_quirk_tdp = 30,
1073 };
1074 
1075 static const struct platform_features slvd_features = {
1076 	.has_msr_misc_pwr_mgmt = 1,
1077 	.has_nhm_msrs = 1,
1078 	.bclk_freq = BCLK_SLV,
1079 	.supported_cstates = CC1 | CC6 | PC3 | PC6,
1080 	.cst_limit = CST_LIMIT_SLV,
1081 	.has_msr_atom_pkg_c6_residency = 1,
1082 	.trl_msrs = TRL_BASE,
1083 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE,
1084 	.rapl_quirk_tdp = 30,
1085 };
1086 
1087 static const struct platform_features amt_features = {
1088 	.has_nhm_msrs = 1,
1089 	.bclk_freq = BCLK_133MHZ,
1090 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
1091 	.cst_limit = CST_LIMIT_AMT,
1092 	.trl_msrs = TRL_BASE,
1093 };
1094 
1095 static const struct platform_features gmt_features = {
1096 	.has_msr_misc_pwr_mgmt = 1,
1097 	.has_nhm_msrs = 1,
1098 	.bclk_freq = BCLK_100MHZ,
1099 	.crystal_freq = 19200000,
1100 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
1101 	.cst_limit = CST_LIMIT_GMT,
1102 	.has_irtl_msrs = 1,
1103 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1104 	.plat_rapl_msrs = RAPL_PKG | RAPL_PKG_POWER_INFO,
1105 };
1106 
1107 static const struct platform_features gmtd_features = {
1108 	.has_msr_misc_pwr_mgmt = 1,
1109 	.has_nhm_msrs = 1,
1110 	.bclk_freq = BCLK_100MHZ,
1111 	.crystal_freq = 25000000,
1112 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
1113 	.cst_limit = CST_LIMIT_GMT,
1114 	.has_irtl_msrs = 1,
1115 	.has_msr_core_c1_res = 1,
1116 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1117 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_CORE_ENERGY_STATUS,
1118 };
1119 
1120 static const struct platform_features gmtp_features = {
1121 	.has_msr_misc_pwr_mgmt = 1,
1122 	.has_nhm_msrs = 1,
1123 	.bclk_freq = BCLK_100MHZ,
1124 	.crystal_freq = 19200000,
1125 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
1126 	.cst_limit = CST_LIMIT_GMT,
1127 	.has_irtl_msrs = 1,
1128 	.trl_msrs = TRL_BASE,
1129 	.plat_rapl_msrs = RAPL_PKG | RAPL_PKG_POWER_INFO,
1130 };
1131 
1132 static const struct platform_features tmt_features = {
1133 	.has_msr_misc_pwr_mgmt = 1,
1134 	.has_nhm_msrs = 1,
1135 	.bclk_freq = BCLK_100MHZ,
1136 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
1137 	.cst_limit = CST_LIMIT_GMT,
1138 	.has_irtl_msrs = 1,
1139 	.trl_msrs = TRL_BASE,
1140 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX,
1141 	.enable_tsc_tweak = 1,
1142 };
1143 
1144 static const struct platform_features tmtd_features = {
1145 	.has_msr_misc_pwr_mgmt = 1,
1146 	.has_nhm_msrs = 1,
1147 	.bclk_freq = BCLK_100MHZ,
1148 	.supported_cstates = CC1 | CC6,
1149 	.cst_limit = CST_LIMIT_GMT,
1150 	.has_irtl_msrs = 1,
1151 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1152 	.plat_rapl_msrs = RAPL_PKG_ALL,
1153 };
1154 
1155 static const struct platform_features knl_features = {
1156 	.has_msr_misc_pwr_mgmt = 1,
1157 	.has_nhm_msrs = 1,
1158 	.has_config_tdp = 1,
1159 	.bclk_freq = BCLK_100MHZ,
1160 	.supported_cstates = CC1 | CC6 | PC3 | PC6,
1161 	.cst_limit = CST_LIMIT_KNL,
1162 	.has_msr_knl_core_c6_residency = 1,
1163 	.trl_msrs = TRL_KNL,
1164 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
1165 	.has_fixed_rapl_unit = 1,
1166 	.need_perf_multiplier = 1,
1167 };
1168 
1169 static const struct platform_features default_features = {
1170 };
1171 
1172 static const struct platform_features amd_features_with_rapl = {
1173 	.plat_rapl_msrs = RAPL_AMD_F17H,
1174 	.has_per_core_rapl = 1,
1175 	.rapl_quirk_tdp = 280,	/* This is the max stock TDP of HEDT/Server Fam17h+ chips */
1176 };
1177 
1178 static const struct platform_data turbostat_pdata[] = {
1179 	{ INTEL_NEHALEM, &nhm_features },
1180 	{ INTEL_NEHALEM_G, &nhm_features },
1181 	{ INTEL_NEHALEM_EP, &nhm_features },
1182 	{ INTEL_NEHALEM_EX, &nhx_features },
1183 	{ INTEL_WESTMERE, &nhm_features },
1184 	{ INTEL_WESTMERE_EP, &nhm_features },
1185 	{ INTEL_WESTMERE_EX, &nhx_features },
1186 	{ INTEL_SANDYBRIDGE, &snb_features },
1187 	{ INTEL_SANDYBRIDGE_X, &snx_features },
1188 	{ INTEL_IVYBRIDGE, &ivb_features },
1189 	{ INTEL_IVYBRIDGE_X, &ivx_features },
1190 	{ INTEL_HASWELL, &hsw_features },
1191 	{ INTEL_HASWELL_X, &hsx_features },
1192 	{ INTEL_HASWELL_L, &hswl_features },
1193 	{ INTEL_HASWELL_G, &hswg_features },
1194 	{ INTEL_BROADWELL, &bdw_features },
1195 	{ INTEL_BROADWELL_G, &bdwg_features },
1196 	{ INTEL_BROADWELL_X, &bdx_features },
1197 	{ INTEL_BROADWELL_D, &bdx_features },
1198 	{ INTEL_SKYLAKE_L, &skl_features },
1199 	{ INTEL_SKYLAKE, &skl_features },
1200 	{ INTEL_SKYLAKE_X, &skx_features },
1201 	{ INTEL_KABYLAKE_L, &skl_features },
1202 	{ INTEL_KABYLAKE, &skl_features },
1203 	{ INTEL_COMETLAKE, &skl_features },
1204 	{ INTEL_COMETLAKE_L, &skl_features },
1205 	{ INTEL_CANNONLAKE_L, &cnl_features },
1206 	{ INTEL_ICELAKE_X, &icx_features },
1207 	{ INTEL_ICELAKE_D, &icx_features },
1208 	{ INTEL_ICELAKE_L, &cnl_features },
1209 	{ INTEL_ICELAKE_NNPI, &cnl_features },
1210 	{ INTEL_ROCKETLAKE, &cnl_features },
1211 	{ INTEL_TIGERLAKE_L, &cnl_features },
1212 	{ INTEL_TIGERLAKE, &cnl_features },
1213 	{ INTEL_SAPPHIRERAPIDS_X, &spr_features },
1214 	{ INTEL_EMERALDRAPIDS_X, &spr_features },
1215 	{ INTEL_GRANITERAPIDS_X, &spr_features },
1216 	{ INTEL_GRANITERAPIDS_D, &spr_features },
1217 	{ INTEL_DIAMONDRAPIDS_X, &dmr_features },
1218 	{ INTEL_LAKEFIELD, &cnl_features },
1219 	{ INTEL_ALDERLAKE, &adl_features },
1220 	{ INTEL_ALDERLAKE_L, &adl_features },
1221 	{ INTEL_RAPTORLAKE, &adl_features },
1222 	{ INTEL_RAPTORLAKE_P, &adl_features },
1223 	{ INTEL_RAPTORLAKE_S, &adl_features },
1224 	{ INTEL_BARTLETTLAKE, &adl_features },
1225 	{ INTEL_METEORLAKE, &adl_features },
1226 	{ INTEL_METEORLAKE_L, &adl_features },
1227 	{ INTEL_ARROWLAKE_H, &adl_features },
1228 	{ INTEL_ARROWLAKE_U, &adl_features },
1229 	{ INTEL_ARROWLAKE, &adl_features },
1230 	{ INTEL_LUNARLAKE_M, &lnl_features },
1231 	{ INTEL_PANTHERLAKE_L, &lnl_features },
1232 	{ INTEL_NOVALAKE, &lnl_features },
1233 	{ INTEL_NOVALAKE_L, &lnl_features },
1234 	{ INTEL_WILDCATLAKE_L, &lnl_features },
1235 	{ INTEL_ATOM_SILVERMONT, &slv_features },
1236 	{ INTEL_ATOM_SILVERMONT_D, &slvd_features },
1237 	{ INTEL_ATOM_AIRMONT, &amt_features },
1238 	{ INTEL_ATOM_GOLDMONT, &gmt_features },
1239 	{ INTEL_ATOM_GOLDMONT_D, &gmtd_features },
1240 	{ INTEL_ATOM_GOLDMONT_PLUS, &gmtp_features },
1241 	{ INTEL_ATOM_TREMONT_D, &tmtd_features },
1242 	{ INTEL_ATOM_TREMONT, &tmt_features },
1243 	{ INTEL_ATOM_TREMONT_L, &tmt_features },
1244 	{ INTEL_ATOM_GRACEMONT, &adl_features },
1245 	{ INTEL_ATOM_CRESTMONT_X, &srf_features },
1246 	{ INTEL_ATOM_CRESTMONT, &grr_features },
1247 	{ INTEL_ATOM_DARKMONT_X, &srf_features },
1248 	{ INTEL_XEON_PHI_KNL, &knl_features },
1249 	{ INTEL_XEON_PHI_KNM, &knl_features },
1250 	/*
1251 	 * Missing support for
1252 	 * INTEL_ICELAKE
1253 	 * INTEL_ATOM_SILVERMONT_MID
1254 	 * INTEL_ATOM_SILVERMONT_MID2
1255 	 * INTEL_ATOM_AIRMONT_NP
1256 	 */
1257 	{ 0, NULL },
1258 };
1259 
1260 struct {
1261 	unsigned int uniform;
1262 	unsigned int pcore;
1263 	unsigned int ecore;
1264 	unsigned int lcore;
1265 } perf_pmu_types;
1266 
1267 /*
1268  * Events are enumerated in https://github.com/intel/perfmon
1269  * and tools/perf/pmu-events/arch/x86/.../cache.json
1270  */
1271 struct perf_l2_events {
1272 	unsigned long long refs;	/* L2_REQUEST.ALL */
1273 	unsigned long long hits;	/* L2_REQUEST.HIT */
1274 };
1275 
1276 struct perf_model_support {
1277 	unsigned int vfm;
1278 	struct perf_l2_events first;
1279 	struct perf_l2_events second;
1280 	struct perf_l2_events third;
1281 } *perf_model_support;
1282 
1283 /* Perf Cache Events */
1284 #define	PCE(ext_umask, umask)	(((unsigned long long) ext_umask) << 40 | umask << 8 | 0x24)
1285 
1286 /*
1287  * Enumerate up to three perf CPU PMU's in a system.
1288  * The first, second, and third columns are populated without skipping, describing
1289  * pcore, ecore, lcore PMUs, in order, if present.  (The associated PMU "type" field is
1290  * read from sysfs in all cases.)  Eg.
1291  *
1292  * non-hybrid:
1293  *	GNR: pcore, {}, {}
1294  *	ADL-N: ecore, {}, {}
1295  * hybrid:
1296  *	MTL: pcore, ecore, {}%
1297  *	ARL-H: pcore, ecore, lcore
1298  *	LNL: ecore, ecore%%, {}
1299  *
1300  * % MTL physical lcore share architecture and PMU with ecore, and are thus not enumerated separately.
1301  * %% LNL physical lcore is enumerated by perf as ecore
1302  */
1303 static struct perf_model_support turbostat_perf_model_support[] = {
1304 	{ INTEL_SAPPHIRERAPIDS_X, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, {}, {} },
1305 	{ INTEL_EMERALDRAPIDS_X, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, {}, {} },
1306 	{ INTEL_GRANITERAPIDS_X, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, {}, {} },
1307 	{ INTEL_GRANITERAPIDS_D, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, {}, {} },
1308 	{ INTEL_DIAMONDRAPIDS_X, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, {}, {} },
1309 
1310 	{ INTEL_ATOM_GRACEMONT, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {}, {} },	/* ADL-N */
1311 	{ INTEL_ATOM_CRESTMONT_X, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {}, {} },	/* SRF */
1312 	{ INTEL_ATOM_CRESTMONT, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {}, {} },	/* GRR */
1313 	{ INTEL_ATOM_DARKMONT_X, { PCE(0x01, 0xFF), PCE(0x01, 0xBF)}, {}, {} },	/* CWF */
1314 
1315 	{ INTEL_ALDERLAKE, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1316 	{ INTEL_ALDERLAKE, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1317 	{ INTEL_ALDERLAKE_L, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1318 	{ INTEL_RAPTORLAKE, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1319 	{ INTEL_RAPTORLAKE_P, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1320 	{ INTEL_RAPTORLAKE_S, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1321 	{ INTEL_METEORLAKE_L, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1322 	{ INTEL_METEORLAKE, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1323 	{ INTEL_ARROWLAKE_U, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1324 
1325 	{ INTEL_LUNARLAKE_M, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x00, 0x07), PCE(0x00, 0x02)}, {} },
1326 	{ INTEL_ARROWLAKE_H, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x00, 0x07), PCE(0x00, 0x02)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)} },
1327 	{ INTEL_ARROWLAKE, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x00, 0x07), PCE(0x00, 0x02)}, {} },
1328 
1329 	{ INTEL_PANTHERLAKE_L, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x01, 0xFF), PCE(0x01, 0xBF)}, {} },
1330 	{ INTEL_WILDCATLAKE_L, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x01, 0xFF), PCE(0x01, 0xBF)}, {} },
1331 
1332 	{ INTEL_NOVALAKE, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x01, 0xFF), PCE(0x01, 0xBF)}, {} },
1333 	{ INTEL_NOVALAKE_L, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x01, 0xFF), PCE(0x01, 0xBF)}, {} },
1334 
1335 	{ 0, {}, {}, {} }
1336 };
1337 
1338 static const struct platform_features *platform;
1339 
probe_platform_features(unsigned int family,unsigned int model)1340 void probe_platform_features(unsigned int family, unsigned int model)
1341 {
1342 	int i;
1343 
1344 	if (authentic_amd || hygon_genuine) {
1345 		/* fallback to default features on unsupported models */
1346 		force_load++;
1347 		if (max_extended_level >= 0x80000007) {
1348 			unsigned int eax, ebx, ecx, edx;
1349 
1350 			__cpuid(0x80000007, eax, ebx, ecx, edx);
1351 			/* RAPL (Fam 17h+) */
1352 			if ((edx & (1 << 14)) && family >= 0x17)
1353 				platform = &amd_features_with_rapl;
1354 		}
1355 		goto end;
1356 	}
1357 
1358 	if (!genuine_intel)
1359 		goto end;
1360 
1361 	for (i = 0; turbostat_pdata[i].features; i++) {
1362 		if (VFM_FAMILY(turbostat_pdata[i].vfm) == family && VFM_MODEL(turbostat_pdata[i].vfm) == model) {
1363 			platform = turbostat_pdata[i].features;
1364 			return;
1365 		}
1366 	}
1367 
1368 end:
1369 	if (force_load && !platform) {
1370 		fprintf(outf, "Forced to run on unsupported platform!\n");
1371 		platform = &default_features;
1372 	}
1373 
1374 	if (platform)
1375 		return;
1376 
1377 	fprintf(stderr, "Unsupported platform detected.\n\tSee RUN THE LATEST VERSION on turbostat(8)\n");
1378 	exit(1);
1379 }
1380 
init_perf_model_support(unsigned int family,unsigned int model)1381 void init_perf_model_support(unsigned int family, unsigned int model)
1382 {
1383 	int i;
1384 
1385 	if (!genuine_intel)
1386 		return;
1387 
1388 	for (i = 0; turbostat_perf_model_support[i].vfm; i++) {
1389 		if (VFM_FAMILY(turbostat_perf_model_support[i].vfm) == family && VFM_MODEL(turbostat_perf_model_support[i].vfm) == model) {
1390 			perf_model_support = &turbostat_perf_model_support[i];
1391 			return;
1392 		}
1393 	}
1394 }
1395 
1396 /* Model specific support End */
1397 
1398 #define	TJMAX_DEFAULT	100
1399 
1400 /* MSRs that are not yet in the kernel-provided header. */
1401 #define MSR_RAPL_PWR_UNIT	0xc0010299
1402 #define MSR_CORE_ENERGY_STAT	0xc001029a
1403 #define MSR_PKG_ENERGY_STAT	0xc001029b
1404 
1405 #define MAX(a, b) ((a) > (b) ? (a) : (b))
1406 
1407 int backwards_count;
1408 char *progname;
1409 
1410 #define CPU_SUBSET_MAXCPUS	8192	/* need to use before probe... */
1411 cpu_set_t *cpu_present_set, *cpu_possible_set, *cpu_effective_set, *cpu_allowed_set, *cpu_affinity_set, *cpu_subset;
1412 cpu_set_t *perf_pcore_set, *perf_ecore_set, *perf_lcore_set;
1413 size_t cpu_present_setsize, cpu_possible_setsize, cpu_effective_setsize, cpu_allowed_setsize, cpu_affinity_setsize, cpu_subset_size;
1414 #define MAX_ADDED_THREAD_COUNTERS 24
1415 #define MAX_ADDED_CORE_COUNTERS 8
1416 #define MAX_ADDED_PACKAGE_COUNTERS 16
1417 #define PMT_MAX_ADDED_THREAD_COUNTERS 24
1418 #define PMT_MAX_ADDED_CORE_COUNTERS 8
1419 #define PMT_MAX_ADDED_PACKAGE_COUNTERS 16
1420 #define BITMASK_SIZE 32
1421 
1422 #define ZERO_ARRAY(arr) (memset(arr, 0, sizeof(arr)) + __must_be_array(arr))
1423 
1424 /* Indexes used to map data read from perf and MSRs into global variables */
1425 enum rapl_rci_index {
1426 	RAPL_RCI_INDEX_ENERGY_PKG = 0,
1427 	RAPL_RCI_INDEX_ENERGY_CORES = 1,
1428 	RAPL_RCI_INDEX_DRAM = 2,
1429 	RAPL_RCI_INDEX_GFX = 3,
1430 	RAPL_RCI_INDEX_PKG_PERF_STATUS = 4,
1431 	RAPL_RCI_INDEX_DRAM_PERF_STATUS = 5,
1432 	RAPL_RCI_INDEX_CORE_ENERGY = 6,
1433 	RAPL_RCI_INDEX_ENERGY_PLATFORM = 7,
1434 	NUM_RAPL_COUNTERS,
1435 };
1436 
1437 enum rapl_unit {
1438 	RAPL_UNIT_INVALID,
1439 	RAPL_UNIT_JOULES,
1440 	RAPL_UNIT_WATTS,
1441 };
1442 
1443 struct rapl_counter_info_t {
1444 	unsigned long long data[NUM_RAPL_COUNTERS];
1445 	enum counter_source source[NUM_RAPL_COUNTERS];
1446 	unsigned long long flags[NUM_RAPL_COUNTERS];
1447 	double scale[NUM_RAPL_COUNTERS];
1448 	enum rapl_unit unit[NUM_RAPL_COUNTERS];
1449 	unsigned long long msr[NUM_RAPL_COUNTERS];
1450 	unsigned long long msr_mask[NUM_RAPL_COUNTERS];
1451 	int msr_shift[NUM_RAPL_COUNTERS];
1452 
1453 	int fd_perf;
1454 };
1455 
1456 /* struct rapl_counter_info_t for each RAPL domain */
1457 struct rapl_counter_info_t *rapl_counter_info_perdomain;
1458 unsigned int rapl_counter_info_perdomain_size;
1459 
1460 #define RAPL_COUNTER_FLAG_PLATFORM_COUNTER (1u << 0)
1461 #define RAPL_COUNTER_FLAG_USE_MSR_SUM (1u << 1)
1462 
1463 struct rapl_counter_arch_info {
1464 	int feature_mask;	/* Mask for testing if the counter is supported on host */
1465 	const char *perf_subsys;
1466 	const char *perf_name;
1467 	unsigned long long msr;
1468 	unsigned long long msr_mask;
1469 	int msr_shift;		/* Positive mean shift right, negative mean shift left */
1470 	double *platform_rapl_msr_scale;	/* Scale applied to values read by MSR (platform dependent, filled at runtime) */
1471 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1472 	unsigned int bic_number;
1473 	double compat_scale;	/* Some counters require constant scaling to be in the same range as other, similar ones */
1474 	unsigned long long flags;
1475 };
1476 
1477 static const struct rapl_counter_arch_info rapl_counter_arch_infos[] = {
1478 	{
1479 	 .feature_mask = RAPL_PKG,
1480 	 .perf_subsys = "power",
1481 	 .perf_name = "energy-pkg",
1482 	 .msr = MSR_PKG_ENERGY_STATUS,
1483 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1484 	 .msr_shift = 0,
1485 	 .platform_rapl_msr_scale = &rapl_energy_units,
1486 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1487 	 .bic_number = BIC_PkgWatt,
1488 	 .compat_scale = 1.0,
1489 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1490 	  },
1491 	{
1492 	 .feature_mask = RAPL_PKG,
1493 	 .perf_subsys = "power",
1494 	 .perf_name = "energy-pkg",
1495 	 .msr = MSR_PKG_ENERGY_STATUS,
1496 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1497 	 .msr_shift = 0,
1498 	 .platform_rapl_msr_scale = &rapl_energy_units,
1499 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1500 	 .bic_number = BIC_Pkg_J,
1501 	 .compat_scale = 1.0,
1502 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1503 	  },
1504 	{
1505 	 .feature_mask = RAPL_AMD_F17H,
1506 	 .perf_subsys = "power",
1507 	 .perf_name = "energy-pkg",
1508 	 .msr = MSR_PKG_ENERGY_STAT,
1509 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1510 	 .msr_shift = 0,
1511 	 .platform_rapl_msr_scale = &rapl_energy_units,
1512 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1513 	 .bic_number = BIC_PkgWatt,
1514 	 .compat_scale = 1.0,
1515 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1516 	  },
1517 	{
1518 	 .feature_mask = RAPL_AMD_F17H,
1519 	 .perf_subsys = "power",
1520 	 .perf_name = "energy-pkg",
1521 	 .msr = MSR_PKG_ENERGY_STAT,
1522 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1523 	 .msr_shift = 0,
1524 	 .platform_rapl_msr_scale = &rapl_energy_units,
1525 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1526 	 .bic_number = BIC_Pkg_J,
1527 	 .compat_scale = 1.0,
1528 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1529 	  },
1530 	{
1531 	 .feature_mask = RAPL_CORE_ENERGY_STATUS,
1532 	 .perf_subsys = "power",
1533 	 .perf_name = "energy-cores",
1534 	 .msr = MSR_PP0_ENERGY_STATUS,
1535 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1536 	 .msr_shift = 0,
1537 	 .platform_rapl_msr_scale = &rapl_energy_units,
1538 	 .rci_index = RAPL_RCI_INDEX_ENERGY_CORES,
1539 	 .bic_number = BIC_CorWatt,
1540 	 .compat_scale = 1.0,
1541 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1542 	  },
1543 	{
1544 	 .feature_mask = RAPL_CORE_ENERGY_STATUS,
1545 	 .perf_subsys = "power",
1546 	 .perf_name = "energy-cores",
1547 	 .msr = MSR_PP0_ENERGY_STATUS,
1548 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1549 	 .msr_shift = 0,
1550 	 .platform_rapl_msr_scale = &rapl_energy_units,
1551 	 .rci_index = RAPL_RCI_INDEX_ENERGY_CORES,
1552 	 .bic_number = BIC_Cor_J,
1553 	 .compat_scale = 1.0,
1554 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1555 	  },
1556 	{
1557 	 .feature_mask = RAPL_DRAM,
1558 	 .perf_subsys = "power",
1559 	 .perf_name = "energy-ram",
1560 	 .msr = MSR_DRAM_ENERGY_STATUS,
1561 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1562 	 .msr_shift = 0,
1563 	 .platform_rapl_msr_scale = &rapl_dram_energy_units,
1564 	 .rci_index = RAPL_RCI_INDEX_DRAM,
1565 	 .bic_number = BIC_RAMWatt,
1566 	 .compat_scale = 1.0,
1567 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1568 	  },
1569 	{
1570 	 .feature_mask = RAPL_DRAM,
1571 	 .perf_subsys = "power",
1572 	 .perf_name = "energy-ram",
1573 	 .msr = MSR_DRAM_ENERGY_STATUS,
1574 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1575 	 .msr_shift = 0,
1576 	 .platform_rapl_msr_scale = &rapl_dram_energy_units,
1577 	 .rci_index = RAPL_RCI_INDEX_DRAM,
1578 	 .bic_number = BIC_RAM_J,
1579 	 .compat_scale = 1.0,
1580 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1581 	  },
1582 	{
1583 	 .feature_mask = RAPL_GFX,
1584 	 .perf_subsys = "power",
1585 	 .perf_name = "energy-gpu",
1586 	 .msr = MSR_PP1_ENERGY_STATUS,
1587 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1588 	 .msr_shift = 0,
1589 	 .platform_rapl_msr_scale = &rapl_energy_units,
1590 	 .rci_index = RAPL_RCI_INDEX_GFX,
1591 	 .bic_number = BIC_GFXWatt,
1592 	 .compat_scale = 1.0,
1593 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1594 	  },
1595 	{
1596 	 .feature_mask = RAPL_GFX,
1597 	 .perf_subsys = "power",
1598 	 .perf_name = "energy-gpu",
1599 	 .msr = MSR_PP1_ENERGY_STATUS,
1600 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1601 	 .msr_shift = 0,
1602 	 .platform_rapl_msr_scale = &rapl_energy_units,
1603 	 .rci_index = RAPL_RCI_INDEX_GFX,
1604 	 .bic_number = BIC_GFX_J,
1605 	 .compat_scale = 1.0,
1606 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1607 	  },
1608 	{
1609 	 .feature_mask = RAPL_PKG_PERF_STATUS,
1610 	 .perf_subsys = NULL,
1611 	 .perf_name = NULL,
1612 	 .msr = MSR_PKG_PERF_STATUS,
1613 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1614 	 .msr_shift = 0,
1615 	 .platform_rapl_msr_scale = &rapl_time_units,
1616 	 .rci_index = RAPL_RCI_INDEX_PKG_PERF_STATUS,
1617 	 .bic_number = BIC_PKG__,
1618 	 .compat_scale = 100.0,
1619 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1620 	  },
1621 	{
1622 	 .feature_mask = RAPL_DRAM_PERF_STATUS,
1623 	 .perf_subsys = NULL,
1624 	 .perf_name = NULL,
1625 	 .msr = MSR_DRAM_PERF_STATUS,
1626 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1627 	 .msr_shift = 0,
1628 	 .platform_rapl_msr_scale = &rapl_time_units,
1629 	 .rci_index = RAPL_RCI_INDEX_DRAM_PERF_STATUS,
1630 	 .bic_number = BIC_RAM__,
1631 	 .compat_scale = 100.0,
1632 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1633 	  },
1634 	{
1635 	 .feature_mask = RAPL_AMD_F17H,
1636 	 .perf_subsys = NULL,
1637 	 .perf_name = NULL,
1638 	 .msr = MSR_CORE_ENERGY_STAT,
1639 	 .msr_mask = 0xFFFFFFFF,
1640 	 .msr_shift = 0,
1641 	 .platform_rapl_msr_scale = &rapl_energy_units,
1642 	 .rci_index = RAPL_RCI_INDEX_CORE_ENERGY,
1643 	 .bic_number = BIC_CorWatt,
1644 	 .compat_scale = 1.0,
1645 	 .flags = 0,
1646 	  },
1647 	{
1648 	 .feature_mask = RAPL_AMD_F17H,
1649 	 .perf_subsys = NULL,
1650 	 .perf_name = NULL,
1651 	 .msr = MSR_CORE_ENERGY_STAT,
1652 	 .msr_mask = 0xFFFFFFFF,
1653 	 .msr_shift = 0,
1654 	 .platform_rapl_msr_scale = &rapl_energy_units,
1655 	 .rci_index = RAPL_RCI_INDEX_CORE_ENERGY,
1656 	 .bic_number = BIC_Cor_J,
1657 	 .compat_scale = 1.0,
1658 	 .flags = 0,
1659 	  },
1660 	{
1661 	 .feature_mask = RAPL_PSYS,
1662 	 .perf_subsys = "power",
1663 	 .perf_name = "energy-psys",
1664 	 .msr = MSR_PLATFORM_ENERGY_STATUS,
1665 	 .msr_mask = 0x00000000FFFFFFFF,
1666 	 .msr_shift = 0,
1667 	 .platform_rapl_msr_scale = &rapl_psys_energy_units,
1668 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PLATFORM,
1669 	 .bic_number = BIC_SysWatt,
1670 	 .compat_scale = 1.0,
1671 	 .flags = RAPL_COUNTER_FLAG_PLATFORM_COUNTER | RAPL_COUNTER_FLAG_USE_MSR_SUM,
1672 	  },
1673 	{
1674 	 .feature_mask = RAPL_PSYS,
1675 	 .perf_subsys = "power",
1676 	 .perf_name = "energy-psys",
1677 	 .msr = MSR_PLATFORM_ENERGY_STATUS,
1678 	 .msr_mask = 0x00000000FFFFFFFF,
1679 	 .msr_shift = 0,
1680 	 .platform_rapl_msr_scale = &rapl_psys_energy_units,
1681 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PLATFORM,
1682 	 .bic_number = BIC_Sys_J,
1683 	 .compat_scale = 1.0,
1684 	 .flags = RAPL_COUNTER_FLAG_PLATFORM_COUNTER | RAPL_COUNTER_FLAG_USE_MSR_SUM,
1685 	  },
1686 };
1687 
1688 struct rapl_counter {
1689 	unsigned long long raw_value;
1690 	enum rapl_unit unit;
1691 	double scale;
1692 };
1693 
1694 /* Indexes used to map data read from perf and MSRs into global variables */
1695 enum ccstate_rci_index {
1696 	CCSTATE_RCI_INDEX_C1_RESIDENCY = 0,
1697 	CCSTATE_RCI_INDEX_C3_RESIDENCY = 1,
1698 	CCSTATE_RCI_INDEX_C6_RESIDENCY = 2,
1699 	CCSTATE_RCI_INDEX_C7_RESIDENCY = 3,
1700 	PCSTATE_RCI_INDEX_C2_RESIDENCY = 4,
1701 	PCSTATE_RCI_INDEX_C3_RESIDENCY = 5,
1702 	PCSTATE_RCI_INDEX_C6_RESIDENCY = 6,
1703 	PCSTATE_RCI_INDEX_C7_RESIDENCY = 7,
1704 	PCSTATE_RCI_INDEX_C8_RESIDENCY = 8,
1705 	PCSTATE_RCI_INDEX_C9_RESIDENCY = 9,
1706 	PCSTATE_RCI_INDEX_C10_RESIDENCY = 10,
1707 	NUM_CSTATE_COUNTERS,
1708 };
1709 
1710 struct cstate_counter_info_t {
1711 	unsigned long long data[NUM_CSTATE_COUNTERS];
1712 	enum counter_source source[NUM_CSTATE_COUNTERS];
1713 	unsigned long long msr[NUM_CSTATE_COUNTERS];
1714 	int fd_perf_core;
1715 	int fd_perf_pkg;
1716 };
1717 
1718 struct cstate_counter_info_t *ccstate_counter_info;
1719 unsigned int ccstate_counter_info_size;
1720 
1721 #define CSTATE_COUNTER_FLAG_COLLECT_PER_CORE   (1u << 0)
1722 #define CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD ((1u << 1) | CSTATE_COUNTER_FLAG_COLLECT_PER_CORE)
1723 #define CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY (1u << 2)
1724 
1725 struct cstate_counter_arch_info {
1726 	int feature_mask;	/* Mask for testing if the counter is supported on host */
1727 	const char *perf_subsys;
1728 	const char *perf_name;
1729 	unsigned long long msr;
1730 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1731 	unsigned int bic_number;
1732 	unsigned long long flags;
1733 	int pkg_cstate_limit;
1734 };
1735 
1736 static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
1737 	{
1738 	 .feature_mask = CC1,
1739 	 .perf_subsys = "cstate_core",
1740 	 .perf_name = "c1-residency",
1741 	 .msr = MSR_CORE_C1_RES,
1742 	 .rci_index = CCSTATE_RCI_INDEX_C1_RESIDENCY,
1743 	 .bic_number = BIC_CPU_c1,
1744 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD,
1745 	 .pkg_cstate_limit = 0,
1746 	  },
1747 	{
1748 	 .feature_mask = CC3,
1749 	 .perf_subsys = "cstate_core",
1750 	 .perf_name = "c3-residency",
1751 	 .msr = MSR_CORE_C3_RESIDENCY,
1752 	 .rci_index = CCSTATE_RCI_INDEX_C3_RESIDENCY,
1753 	 .bic_number = BIC_CPU_c3,
1754 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1755 	 .pkg_cstate_limit = 0,
1756 	  },
1757 	{
1758 	 .feature_mask = CC6,
1759 	 .perf_subsys = "cstate_core",
1760 	 .perf_name = "c6-residency",
1761 	 .msr = MSR_CORE_C6_RESIDENCY,
1762 	 .rci_index = CCSTATE_RCI_INDEX_C6_RESIDENCY,
1763 	 .bic_number = BIC_CPU_c6,
1764 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1765 	 .pkg_cstate_limit = 0,
1766 	  },
1767 	{
1768 	 .feature_mask = CC7,
1769 	 .perf_subsys = "cstate_core",
1770 	 .perf_name = "c7-residency",
1771 	 .msr = MSR_CORE_C7_RESIDENCY,
1772 	 .rci_index = CCSTATE_RCI_INDEX_C7_RESIDENCY,
1773 	 .bic_number = BIC_CPU_c7,
1774 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1775 	 .pkg_cstate_limit = 0,
1776 	  },
1777 	{
1778 	 .feature_mask = PC2,
1779 	 .perf_subsys = "cstate_pkg",
1780 	 .perf_name = "c2-residency",
1781 	 .msr = MSR_PKG_C2_RESIDENCY,
1782 	 .rci_index = PCSTATE_RCI_INDEX_C2_RESIDENCY,
1783 	 .bic_number = BIC_Pkgpc2,
1784 	 .flags = 0,
1785 	 .pkg_cstate_limit = PCL__2,
1786 	  },
1787 	{
1788 	 .feature_mask = PC3,
1789 	 .perf_subsys = "cstate_pkg",
1790 	 .perf_name = "c3-residency",
1791 	 .msr = MSR_PKG_C3_RESIDENCY,
1792 	 .rci_index = PCSTATE_RCI_INDEX_C3_RESIDENCY,
1793 	 .bic_number = BIC_Pkgpc3,
1794 	 .flags = 0,
1795 	 .pkg_cstate_limit = PCL__3,
1796 	  },
1797 	{
1798 	 .feature_mask = PC6,
1799 	 .perf_subsys = "cstate_pkg",
1800 	 .perf_name = "c6-residency",
1801 	 .msr = MSR_PKG_C6_RESIDENCY,
1802 	 .rci_index = PCSTATE_RCI_INDEX_C6_RESIDENCY,
1803 	 .bic_number = BIC_Pkgpc6,
1804 	 .flags = 0,
1805 	 .pkg_cstate_limit = PCL__6,
1806 	  },
1807 	{
1808 	 .feature_mask = PC7,
1809 	 .perf_subsys = "cstate_pkg",
1810 	 .perf_name = "c7-residency",
1811 	 .msr = MSR_PKG_C7_RESIDENCY,
1812 	 .rci_index = PCSTATE_RCI_INDEX_C7_RESIDENCY,
1813 	 .bic_number = BIC_Pkgpc7,
1814 	 .flags = 0,
1815 	 .pkg_cstate_limit = PCL__7,
1816 	  },
1817 	{
1818 	 .feature_mask = PC8,
1819 	 .perf_subsys = "cstate_pkg",
1820 	 .perf_name = "c8-residency",
1821 	 .msr = MSR_PKG_C8_RESIDENCY,
1822 	 .rci_index = PCSTATE_RCI_INDEX_C8_RESIDENCY,
1823 	 .bic_number = BIC_Pkgpc8,
1824 	 .flags = 0,
1825 	 .pkg_cstate_limit = PCL__8,
1826 	  },
1827 	{
1828 	 .feature_mask = PC9,
1829 	 .perf_subsys = "cstate_pkg",
1830 	 .perf_name = "c9-residency",
1831 	 .msr = MSR_PKG_C9_RESIDENCY,
1832 	 .rci_index = PCSTATE_RCI_INDEX_C9_RESIDENCY,
1833 	 .bic_number = BIC_Pkgpc9,
1834 	 .flags = 0,
1835 	 .pkg_cstate_limit = PCL__9,
1836 	  },
1837 	{
1838 	 .feature_mask = PC10,
1839 	 .perf_subsys = "cstate_pkg",
1840 	 .perf_name = "c10-residency",
1841 	 .msr = MSR_PKG_C10_RESIDENCY,
1842 	 .rci_index = PCSTATE_RCI_INDEX_C10_RESIDENCY,
1843 	 .bic_number = BIC_Pkgpc10,
1844 	 .flags = 0,
1845 	 .pkg_cstate_limit = PCL_10,
1846 	  },
1847 };
1848 
1849 /* Indexes used to map data read from perf and MSRs into global variables */
1850 enum msr_rci_index {
1851 	MSR_RCI_INDEX_APERF = 0,
1852 	MSR_RCI_INDEX_MPERF = 1,
1853 	MSR_RCI_INDEX_SMI = 2,
1854 	NUM_MSR_COUNTERS,
1855 };
1856 
1857 struct msr_counter_info_t {
1858 	unsigned long long data[NUM_MSR_COUNTERS];
1859 	enum counter_source source[NUM_MSR_COUNTERS];
1860 	unsigned long long msr[NUM_MSR_COUNTERS];
1861 	unsigned long long msr_mask[NUM_MSR_COUNTERS];
1862 	int fd_perf;
1863 };
1864 
1865 struct msr_counter_info_t *msr_counter_info;
1866 unsigned int msr_counter_info_size;
1867 
1868 struct msr_counter_arch_info {
1869 	const char *perf_subsys;
1870 	const char *perf_name;
1871 	unsigned long long msr;
1872 	unsigned long long msr_mask;
1873 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1874 	bool needed;
1875 	bool present;
1876 };
1877 
1878 enum msr_arch_info_index {
1879 	MSR_ARCH_INFO_APERF_INDEX = 0,
1880 	MSR_ARCH_INFO_MPERF_INDEX = 1,
1881 	MSR_ARCH_INFO_SMI_INDEX = 2,
1882 };
1883 
1884 static struct msr_counter_arch_info msr_counter_arch_infos[] = {
1885 	[MSR_ARCH_INFO_APERF_INDEX] = {
1886 				       .perf_subsys = "msr",
1887 				       .perf_name = "aperf",
1888 				       .msr = MSR_IA32_APERF,
1889 				       .msr_mask = 0xFFFFFFFFFFFFFFFF,
1890 				       .rci_index = MSR_RCI_INDEX_APERF,
1891 				        },
1892 
1893 	[MSR_ARCH_INFO_MPERF_INDEX] = {
1894 				       .perf_subsys = "msr",
1895 				       .perf_name = "mperf",
1896 				       .msr = MSR_IA32_MPERF,
1897 				       .msr_mask = 0xFFFFFFFFFFFFFFFF,
1898 				       .rci_index = MSR_RCI_INDEX_MPERF,
1899 				        },
1900 
1901 	[MSR_ARCH_INFO_SMI_INDEX] = {
1902 				     .perf_subsys = "msr",
1903 				     .perf_name = "smi",
1904 				     .msr = MSR_SMI_COUNT,
1905 				     .msr_mask = 0xFFFFFFFF,
1906 				     .rci_index = MSR_RCI_INDEX_SMI,
1907 				      },
1908 };
1909 
1910 /* Can be redefined when compiling, useful for testing. */
1911 #ifndef SYSFS_TELEM_PATH
1912 #define SYSFS_TELEM_PATH "/sys/class/intel_pmt"
1913 #endif
1914 
1915 #define PMT_COUNTER_MTL_DC6_OFFSET 120
1916 #define PMT_COUNTER_MTL_DC6_LSB    0
1917 #define PMT_COUNTER_MTL_DC6_MSB    63
1918 #define PMT_MTL_DC6_GUID           0x1a067102
1919 #define PMT_MTL_DC6_SEQ            0
1920 
1921 #define PMT_COUNTER_CWF_MC1E_OFFSET_BASE          20936
1922 #define PMT_COUNTER_CWF_MC1E_OFFSET_INCREMENT     24
1923 #define PMT_COUNTER_CWF_MC1E_NUM_MODULES_PER_FILE 12
1924 #define PMT_COUNTER_CWF_CPUS_PER_MODULE           4
1925 #define PMT_COUNTER_CWF_MC1E_LSB                  0
1926 #define PMT_COUNTER_CWF_MC1E_MSB                  63
1927 #define PMT_CWF_MC1E_GUID                         0x14421519
1928 
1929 unsigned long long tcore_clock_freq_hz = 800000000;
1930 
1931 #define PMT_COUNTER_NAME_SIZE_BYTES      16
1932 #define PMT_COUNTER_TYPE_NAME_SIZE_BYTES 32
1933 
1934 struct pmt_mmio {
1935 	struct pmt_mmio *next;
1936 
1937 	unsigned int guid;
1938 	unsigned int size;
1939 
1940 	/* Base pointer to the mmaped memory. */
1941 	void *mmio_base;
1942 
1943 	/*
1944 	 * Offset to be applied to the mmio_base
1945 	 * to get the beginning of the PMT counters for given GUID.
1946 	 */
1947 	unsigned long pmt_offset;
1948 } *pmt_mmios;
1949 
1950 enum pmt_datatype {
1951 	PMT_TYPE_RAW,
1952 	PMT_TYPE_XTAL_TIME,
1953 	PMT_TYPE_TCORE_CLOCK,
1954 };
1955 
1956 struct pmt_domain_info {
1957 	/*
1958 	 * Pointer to the MMIO obtained by applying a counter offset
1959 	 * to the mmio_base of the mmaped region for the given GUID.
1960 	 *
1961 	 * This is where to read the raw value of the counter from.
1962 	 */
1963 	unsigned long *pcounter;
1964 };
1965 
1966 struct pmt_counter {
1967 	struct pmt_counter *next;
1968 
1969 	/* PMT metadata */
1970 	char name[PMT_COUNTER_NAME_SIZE_BYTES];
1971 	enum pmt_datatype type;
1972 	enum counter_scope scope;
1973 	unsigned int lsb;
1974 	unsigned int msb;
1975 
1976 	/* BIC-like metadata */
1977 	enum counter_format format;
1978 
1979 	unsigned int num_domains;
1980 	struct pmt_domain_info *domains;
1981 };
1982 
1983 /*
1984  * PMT telemetry directory iterator.
1985  * Used to iterate telemetry files in sysfs in correct order.
1986  */
1987 struct pmt_diriter_t {
1988 	DIR *dir;
1989 	struct dirent **namelist;
1990 	unsigned int num_names;
1991 	unsigned int current_name_idx;
1992 };
1993 
pmt_telemdir_filter(const struct dirent * e)1994 int pmt_telemdir_filter(const struct dirent *e)
1995 {
1996 	unsigned int dummy;
1997 
1998 	return sscanf(e->d_name, "telem%u", &dummy);
1999 }
2000 
pmt_telemdir_sort(const struct dirent ** a,const struct dirent ** b)2001 int pmt_telemdir_sort(const struct dirent **a, const struct dirent **b)
2002 {
2003 	unsigned int aidx = 0, bidx = 0;
2004 
2005 	sscanf((*a)->d_name, "telem%u", &aidx);
2006 	sscanf((*b)->d_name, "telem%u", &bidx);
2007 
2008 	return (aidx > bidx) ? 1 : (aidx < bidx) ? -1 : 0;
2009 }
2010 
pmt_diriter_next(struct pmt_diriter_t * iter)2011 const struct dirent *pmt_diriter_next(struct pmt_diriter_t *iter)
2012 {
2013 	const struct dirent *ret = NULL;
2014 
2015 	if (!iter->dir)
2016 		return NULL;
2017 
2018 	if (iter->current_name_idx >= iter->num_names)
2019 		return NULL;
2020 
2021 	ret = iter->namelist[iter->current_name_idx];
2022 	++iter->current_name_idx;
2023 
2024 	return ret;
2025 }
2026 
pmt_diriter_begin(struct pmt_diriter_t * iter,const char * pmt_root_path)2027 const struct dirent *pmt_diriter_begin(struct pmt_diriter_t *iter, const char *pmt_root_path)
2028 {
2029 	int num_names = iter->num_names;
2030 
2031 	if (!iter->dir) {
2032 		iter->dir = opendir(pmt_root_path);
2033 		if (iter->dir == NULL)
2034 			return NULL;
2035 
2036 		num_names = scandir(pmt_root_path, &iter->namelist, pmt_telemdir_filter, pmt_telemdir_sort);
2037 		if (num_names == -1)
2038 			return NULL;
2039 	}
2040 
2041 	iter->current_name_idx = 0;
2042 	iter->num_names = num_names;
2043 
2044 	return pmt_diriter_next(iter);
2045 }
2046 
pmt_diriter_init(struct pmt_diriter_t * iter)2047 void pmt_diriter_init(struct pmt_diriter_t *iter)
2048 {
2049 	memset(iter, 0, sizeof(*iter));
2050 }
2051 
pmt_diriter_remove(struct pmt_diriter_t * iter)2052 void pmt_diriter_remove(struct pmt_diriter_t *iter)
2053 {
2054 	if (iter->namelist) {
2055 		for (unsigned int i = 0; i < iter->num_names; i++) {
2056 			free(iter->namelist[i]);
2057 			iter->namelist[i] = NULL;
2058 		}
2059 	}
2060 
2061 	free(iter->namelist);
2062 	iter->namelist = NULL;
2063 	iter->num_names = 0;
2064 	iter->current_name_idx = 0;
2065 
2066 	closedir(iter->dir);
2067 	iter->dir = NULL;
2068 }
2069 
pmt_counter_get_width(const struct pmt_counter * p)2070 unsigned int pmt_counter_get_width(const struct pmt_counter *p)
2071 {
2072 	return (p->msb - p->lsb) + 1;
2073 }
2074 
pmt_counter_resize_(struct pmt_counter * pcounter,unsigned int new_size)2075 void pmt_counter_resize_(struct pmt_counter *pcounter, unsigned int new_size)
2076 {
2077 	struct pmt_domain_info *new_mem;
2078 
2079 	new_mem = (struct pmt_domain_info *)reallocarray(pcounter->domains, new_size, sizeof(*pcounter->domains));
2080 	if (!new_mem) {
2081 		fprintf(stderr, "%s: failed to allocate memory for PMT counters\n", __func__);
2082 		exit(1);
2083 	}
2084 
2085 	/* Zero initialize just allocated memory. */
2086 	const size_t num_new_domains = new_size - pcounter->num_domains;
2087 
2088 	memset(&new_mem[pcounter->num_domains], 0, num_new_domains * sizeof(*pcounter->domains));
2089 
2090 	pcounter->num_domains = new_size;
2091 	pcounter->domains = new_mem;
2092 }
2093 
pmt_counter_resize(struct pmt_counter * pcounter,unsigned int new_size)2094 void pmt_counter_resize(struct pmt_counter *pcounter, unsigned int new_size)
2095 {
2096 	/*
2097 	 * Allocate more memory ahead of time.
2098 	 *
2099 	 * Always allocate space for at least 8 elements
2100 	 * and double the size when growing.
2101 	 */
2102 	if (new_size < 8)
2103 		new_size = 8;
2104 	new_size = MAX(new_size, pcounter->num_domains * 2);
2105 
2106 	pmt_counter_resize_(pcounter, new_size);
2107 }
2108 
2109 struct llc_stats {
2110 	unsigned long long references;
2111 	unsigned long long misses;
2112 };
2113 struct l2_stats {
2114 	unsigned long long references;
2115 	unsigned long long hits;
2116 };
2117 struct thread_data {
2118 	struct timeval tv_begin;
2119 	struct timeval tv_end;
2120 	struct timeval tv_delta;
2121 	unsigned long long tsc;
2122 	unsigned long long aperf;
2123 	unsigned long long mperf;
2124 	unsigned long long c1;
2125 	unsigned long long instr_count;
2126 	unsigned long long irq_count;
2127 	unsigned long long nmi_count;
2128 	unsigned int smi_count;
2129 	struct llc_stats llc;
2130 	struct l2_stats l2;
2131 	unsigned int cpu_id;
2132 	unsigned int apic_id;
2133 	unsigned int x2apic_id;
2134 	unsigned int flags;
2135 	bool is_atom;
2136 	unsigned long long counter[MAX_ADDED_THREAD_COUNTERS];
2137 	unsigned long long perf_counter[MAX_ADDED_THREAD_COUNTERS];
2138 	unsigned long long pmt_counter[PMT_MAX_ADDED_THREAD_COUNTERS];
2139 };
2140 
2141 struct core_data {
2142 	int first_cpu;
2143 	unsigned long long c3;
2144 	unsigned long long c6;
2145 	unsigned long long c7;
2146 	unsigned long long mc6_us;	/* duplicate as per-core for now, even though per module */
2147 	unsigned int core_temp_c;
2148 	struct rapl_counter core_energy;	/* MSR_CORE_ENERGY_STAT */
2149 	unsigned long long core_throt_cnt;
2150 	unsigned long long counter[MAX_ADDED_CORE_COUNTERS];
2151 	unsigned long long perf_counter[MAX_ADDED_CORE_COUNTERS];
2152 	unsigned long long pmt_counter[PMT_MAX_ADDED_CORE_COUNTERS];
2153 };
2154 
2155 struct pkg_data {
2156 	int first_cpu;
2157 	unsigned long long pc2;
2158 	unsigned long long pc3;
2159 	unsigned long long pc6;
2160 	unsigned long long pc7;
2161 	unsigned long long pc8;
2162 	unsigned long long pc9;
2163 	unsigned long long pc10;
2164 	long long cpu_lpi;
2165 	long long sys_lpi;
2166 	unsigned long long pkg_wtd_core_c0;
2167 	unsigned long long pkg_any_core_c0;
2168 	unsigned long long pkg_any_gfxe_c0;
2169 	unsigned long long pkg_both_core_gfxe_c0;
2170 	long long gfx_rc6_ms;
2171 	unsigned int gfx_mhz;
2172 	unsigned int gfx_act_mhz;
2173 	long long sam_mc6_ms;
2174 	unsigned int sam_mhz;
2175 	unsigned int sam_act_mhz;
2176 	struct rapl_counter energy_pkg;	/* MSR_PKG_ENERGY_STATUS */
2177 	struct rapl_counter energy_dram;	/* MSR_DRAM_ENERGY_STATUS */
2178 	struct rapl_counter energy_cores;	/* MSR_PP0_ENERGY_STATUS */
2179 	struct rapl_counter energy_gfx;	/* MSR_PP1_ENERGY_STATUS */
2180 	struct rapl_counter rapl_pkg_perf_status;	/* MSR_PKG_PERF_STATUS */
2181 	struct rapl_counter rapl_dram_perf_status;	/* MSR_DRAM_PERF_STATUS */
2182 	unsigned int pkg_temp_c;
2183 	unsigned int uncore_mhz;
2184 	unsigned long long die_c6;
2185 	unsigned long long counter[MAX_ADDED_PACKAGE_COUNTERS];
2186 	unsigned long long perf_counter[MAX_ADDED_PACKAGE_COUNTERS];
2187 	unsigned long long pmt_counter[PMT_MAX_ADDED_PACKAGE_COUNTERS];
2188 };
2189 
2190 #define ODD_COUNTERS odd.threads, odd.cores, odd.packages
2191 #define EVEN_COUNTERS even.threads, even.cores, even.packages
2192 
2193 /*
2194  * The accumulated sum of MSR is defined as a monotonic
2195  * increasing MSR, it will be accumulated periodically,
2196  * despite its register's bit width.
2197  */
2198 enum {
2199 	IDX_PKG_ENERGY,
2200 	IDX_DRAM_ENERGY,
2201 	IDX_PP0_ENERGY,
2202 	IDX_PP1_ENERGY,
2203 	IDX_PKG_PERF,
2204 	IDX_DRAM_PERF,
2205 	IDX_PSYS_ENERGY,
2206 	IDX_COUNT,
2207 };
2208 
2209 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr);
2210 
2211 struct msr_sum_array {
2212 	/* get_msr_sum() = sum + (get_msr() - last) */
2213 	struct {
2214 		/*The accumulated MSR value is updated by the timer */
2215 		unsigned long long sum;
2216 		/*The MSR footprint recorded in last timer */
2217 		unsigned long long last;
2218 	} entries[IDX_COUNT];
2219 };
2220 
2221 /* The percpu MSR sum array.*/
2222 struct msr_sum_array *per_cpu_msr_sum;
2223 
idx_to_offset(int idx)2224 off_t idx_to_offset(int idx)
2225 {
2226 	off_t offset;
2227 
2228 	switch (idx) {
2229 	case IDX_PKG_ENERGY:
2230 		if (platform->plat_rapl_msrs & RAPL_AMD_F17H)
2231 			offset = MSR_PKG_ENERGY_STAT;
2232 		else
2233 			offset = MSR_PKG_ENERGY_STATUS;
2234 		break;
2235 	case IDX_DRAM_ENERGY:
2236 		offset = MSR_DRAM_ENERGY_STATUS;
2237 		break;
2238 	case IDX_PP0_ENERGY:
2239 		offset = MSR_PP0_ENERGY_STATUS;
2240 		break;
2241 	case IDX_PP1_ENERGY:
2242 		offset = MSR_PP1_ENERGY_STATUS;
2243 		break;
2244 	case IDX_PKG_PERF:
2245 		offset = MSR_PKG_PERF_STATUS;
2246 		break;
2247 	case IDX_DRAM_PERF:
2248 		offset = MSR_DRAM_PERF_STATUS;
2249 		break;
2250 	case IDX_PSYS_ENERGY:
2251 		offset = MSR_PLATFORM_ENERGY_STATUS;
2252 		break;
2253 	default:
2254 		offset = -1;
2255 	}
2256 	return offset;
2257 }
2258 
offset_to_idx(off_t offset)2259 int offset_to_idx(off_t offset)
2260 {
2261 	int idx;
2262 
2263 	switch (offset) {
2264 	case MSR_PKG_ENERGY_STATUS:
2265 	case MSR_PKG_ENERGY_STAT:
2266 		idx = IDX_PKG_ENERGY;
2267 		break;
2268 	case MSR_DRAM_ENERGY_STATUS:
2269 		idx = IDX_DRAM_ENERGY;
2270 		break;
2271 	case MSR_PP0_ENERGY_STATUS:
2272 		idx = IDX_PP0_ENERGY;
2273 		break;
2274 	case MSR_PP1_ENERGY_STATUS:
2275 		idx = IDX_PP1_ENERGY;
2276 		break;
2277 	case MSR_PKG_PERF_STATUS:
2278 		idx = IDX_PKG_PERF;
2279 		break;
2280 	case MSR_DRAM_PERF_STATUS:
2281 		idx = IDX_DRAM_PERF;
2282 		break;
2283 	case MSR_PLATFORM_ENERGY_STATUS:
2284 		idx = IDX_PSYS_ENERGY;
2285 		break;
2286 	default:
2287 		idx = -1;
2288 	}
2289 	return idx;
2290 }
2291 
idx_valid(int idx)2292 int idx_valid(int idx)
2293 {
2294 	switch (idx) {
2295 	case IDX_PKG_ENERGY:
2296 		return valid_rapl_msrs & (RAPL_PKG | RAPL_AMD_F17H);
2297 	case IDX_DRAM_ENERGY:
2298 		return valid_rapl_msrs & RAPL_DRAM;
2299 	case IDX_PP0_ENERGY:
2300 		return valid_rapl_msrs & RAPL_CORE_ENERGY_STATUS;
2301 	case IDX_PP1_ENERGY:
2302 		return valid_rapl_msrs & RAPL_GFX;
2303 	case IDX_PKG_PERF:
2304 		return valid_rapl_msrs & RAPL_PKG_PERF_STATUS;
2305 	case IDX_DRAM_PERF:
2306 		return valid_rapl_msrs & RAPL_DRAM_PERF_STATUS;
2307 	case IDX_PSYS_ENERGY:
2308 		return valid_rapl_msrs & RAPL_PSYS;
2309 	default:
2310 		return 0;
2311 	}
2312 }
2313 
2314 struct sys_counters {
2315 	/* MSR added counters */
2316 	unsigned int added_thread_counters;
2317 	unsigned int added_core_counters;
2318 	unsigned int added_package_counters;
2319 	struct msr_counter *tp;
2320 	struct msr_counter *cp;
2321 	struct msr_counter *pp;
2322 
2323 	/* perf added counters */
2324 	unsigned int added_thread_perf_counters;
2325 	unsigned int added_core_perf_counters;
2326 	unsigned int added_package_perf_counters;
2327 	struct perf_counter_info *perf_tp;
2328 	struct perf_counter_info *perf_cp;
2329 	struct perf_counter_info *perf_pp;
2330 
2331 	struct pmt_counter *pmt_tp;
2332 	struct pmt_counter *pmt_cp;
2333 	struct pmt_counter *pmt_pp;
2334 } sys;
2335 
free_msr_counters_(struct msr_counter ** pp)2336 static size_t free_msr_counters_(struct msr_counter **pp)
2337 {
2338 	struct msr_counter *p = NULL;
2339 	size_t num_freed = 0;
2340 
2341 	while (*pp) {
2342 		p = *pp;
2343 
2344 		if (p->msr_num != 0) {
2345 			*pp = p->next;
2346 
2347 			free(p);
2348 			++num_freed;
2349 
2350 			continue;
2351 		}
2352 
2353 		pp = &p->next;
2354 	}
2355 
2356 	return num_freed;
2357 }
2358 
2359 /*
2360  * Free all added counters accessed via msr.
2361  */
free_sys_msr_counters(void)2362 static void free_sys_msr_counters(void)
2363 {
2364 	/* Thread counters */
2365 	sys.added_thread_counters -= free_msr_counters_(&sys.tp);
2366 
2367 	/* Core counters */
2368 	sys.added_core_counters -= free_msr_counters_(&sys.cp);
2369 
2370 	/* Package counters */
2371 	sys.added_package_counters -= free_msr_counters_(&sys.pp);
2372 }
2373 
2374 struct counters {
2375 	struct thread_data *threads;
2376 	struct core_data *cores;
2377 	struct pkg_data *packages;
2378 } average, even, odd;
2379 
2380 struct platform_counters {
2381 	struct rapl_counter energy_psys;	/* MSR_PLATFORM_ENERGY_STATUS */
2382 } platform_counters_odd, platform_counters_even;
2383 
2384 #define	MAX_HT_ID	3	/* support SMT-4 */
2385 
2386 struct cpu_topology {
2387 	int cpu_id;
2388 	int core_id;		/* unique within a package */
2389 	int module_id;
2390 	int package_id;
2391 	int die_id;
2392 	int l3_id;
2393 	int physical_node_id;
2394 	int logical_node_id;	/* 0-based count within the package */
2395 	int ht_id;		/* unique within a core */
2396 	int ht_sibling_cpu_id[MAX_HT_ID + 1];
2397 	int type;
2398 	cpu_set_t *put_ids;	/* Processing Unit/Thread IDs */
2399 } *cpus;
2400 
2401 struct topo_params {
2402 	int num_packages;
2403 	int num_die;
2404 	int num_cpus;
2405 	int num_cores;		/* system wide */
2406 	int allowed_packages;
2407 	int allowed_cpus;
2408 	int allowed_cores;
2409 	int max_cpu_num;
2410 	int max_core_id;	/* within a package */
2411 	int min_module_id;	/* system wide */
2412 	int max_module_id;	/* system wide */
2413 	int max_package_id;
2414 	int max_die_id;
2415 	int max_l3_id;
2416 	int max_node_num;
2417 	int nodes_per_pkg;
2418 	int cores_per_pkg;
2419 	int threads_per_core;
2420 } topo;
2421 
2422 struct timeval tv_even, tv_odd, tv_delta;
2423 
2424 int *irq_column_2_cpu;		/* /proc/interrupts column numbers */
2425 int *irqs_per_cpu;		/* indexed by cpu_num */
2426 int *nmi_per_cpu;		/* indexed by cpu_num */
2427 
2428 void setup_all_buffers(bool startup);
2429 
2430 char *sys_lpi_file;
2431 char *sys_lpi_file_sysfs = "/sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us";
2432 char *sys_lpi_file_debugfs = "/sys/kernel/debug/pmc_core/slp_s0_residency_usec";
2433 
cpu_is_not_present(int cpu)2434 int cpu_is_not_present(int cpu)
2435 {
2436 	if (cpu < 0)
2437 		return 1;
2438 
2439 	return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
2440 }
2441 
cpu_is_not_allowed(int cpu)2442 int cpu_is_not_allowed(int cpu)
2443 {
2444 	if (cpu < 0)
2445 		return 1;
2446 
2447 	return !CPU_ISSET_S(cpu, cpu_allowed_setsize, cpu_allowed_set);
2448 }
2449 
2450 #define GLOBAL_CORE_ID(core_id, pkg_id)	(core_id + pkg_id * (topo.max_core_id + 1))
2451 /*
2452  * run func(thread, core, package) in topology order
2453  * skip non-present cpus
2454  */
2455 
2456 #define PER_THREAD_PARAMS  struct thread_data *t, struct core_data *c, struct pkg_data *p
2457 
has_allowed_lower_ht_sibling(int cpu)2458 int has_allowed_lower_ht_sibling(int cpu)
2459 {
2460 	int i;
2461 
2462 	for (i = 0; i <= cpus[cpu].ht_id; ++i) {
2463 		int sibling_cpu_id = cpus[cpu].ht_sibling_cpu_id[i];
2464 
2465 		if (sibling_cpu_id == cpu)
2466 			return 0;
2467 
2468 		if (!cpu_is_not_allowed(sibling_cpu_id))
2469 			return 1;
2470 	}
2471 	return 0;
2472 }
2473 
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)2474 int for_all_cpus(int (func) (struct thread_data *, struct core_data *, struct pkg_data *),
2475 		 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
2476 {
2477 	int cpu, retval;
2478 
2479 	retval = 0;
2480 
2481 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
2482 		struct thread_data *t;
2483 		struct core_data *c;
2484 		struct pkg_data *p;
2485 
2486 		int pkg_id = cpus[cpu].package_id;
2487 
2488 		if (cpu_is_not_allowed(cpu))
2489 			continue;
2490 
2491 		if (has_allowed_lower_ht_sibling(cpu))	/* skip HT sibling */
2492 			continue;
2493 
2494 		t = &thread_base[cpu];
2495 		c = &core_base[GLOBAL_CORE_ID(cpus[cpu].core_id, pkg_id)];
2496 		p = &pkg_base[pkg_id];
2497 
2498 		retval |= func(t, c, p);
2499 
2500 		/* Handle other HT siblings now */
2501 		int i;
2502 
2503 		for (i = 0; i <= MAX_HT_ID; ++i) {
2504 			int sibling_cpu_id = cpus[cpu].ht_sibling_cpu_id[i];
2505 
2506 			if (sibling_cpu_id < 0)
2507 				break;
2508 
2509 			if (sibling_cpu_id == cpu)
2510 				continue;
2511 
2512 			if (cpu_is_not_allowed(sibling_cpu_id))
2513 				continue;
2514 
2515 			t = &thread_base[sibling_cpu_id];
2516 
2517 			retval |= func(t, c, p);
2518 		}
2519 	}
2520 	return retval;
2521 }
2522 
is_cpu_first_thread_in_core(struct thread_data * t,struct core_data * c)2523 int is_cpu_first_thread_in_core(struct thread_data *t, struct core_data *c)
2524 {
2525 	return ((int)t->cpu_id == c->first_cpu || c->first_cpu < 0);
2526 }
2527 
is_cpu_first_core_in_package(struct thread_data * t,struct pkg_data * p)2528 int is_cpu_first_core_in_package(struct thread_data *t, struct pkg_data *p)
2529 {
2530 	return ((int)t->cpu_id == p->first_cpu || p->first_cpu < 0);
2531 }
2532 
is_cpu_first_thread_in_package(struct thread_data * t,struct core_data * c,struct pkg_data * p)2533 int is_cpu_first_thread_in_package(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2534 {
2535 	return is_cpu_first_thread_in_core(t, c) && is_cpu_first_core_in_package(t, p);
2536 }
2537 
cpu_migrate(int cpu)2538 int cpu_migrate(int cpu)
2539 {
2540 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2541 	CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
2542 	if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
2543 		return -1;
2544 	else
2545 		return 0;
2546 }
2547 
get_msr_fd(int cpu)2548 int get_msr_fd(int cpu)
2549 {
2550 	char pathname[32];
2551 	int fd;
2552 
2553 	fd = fd_percpu[cpu];
2554 
2555 	if (fd)
2556 		return fd;
2557 	sprintf(pathname, use_android_msr_path ? "/dev/msr%d" : "/dev/cpu/%d/msr", cpu);
2558 	fd = open(pathname, O_RDONLY);
2559 	if (fd < 0)
2560 		err(-1, "%s open failed, try chown or chmod +r %s, "
2561 		    "or run with --no-msr, or run as root", pathname, use_android_msr_path ? "/dev/msr*" : "/dev/cpu/*/msr");
2562 	fd_percpu[cpu] = fd;
2563 
2564 	return fd;
2565 }
2566 
bic_disable_msr_access(void)2567 static void bic_disable_msr_access(void)
2568 {
2569 	CLR_BIC(BIC_Mod_c6, &bic_enabled);
2570 	CLR_BIC(BIC_CoreTmp, &bic_enabled);
2571 	CLR_BIC(BIC_Totl_c0, &bic_enabled);
2572 	CLR_BIC(BIC_Any_c0, &bic_enabled);
2573 	CLR_BIC(BIC_GFX_c0, &bic_enabled);
2574 	CLR_BIC(BIC_CPUGFX, &bic_enabled);
2575 	CLR_BIC(BIC_PkgTmp, &bic_enabled);
2576 
2577 	free_sys_msr_counters();
2578 }
2579 
bic_disable_perf_access(void)2580 static void bic_disable_perf_access(void)
2581 {
2582 	CLR_BIC(BIC_IPC, &bic_enabled);
2583 	CLR_BIC(BIC_LLC_MRPS, &bic_enabled);
2584 	CLR_BIC(BIC_LLC_HIT, &bic_enabled);
2585 	CLR_BIC(BIC_L2_MRPS, &bic_enabled);
2586 	CLR_BIC(BIC_L2_HIT, &bic_enabled);
2587 }
2588 
perf_event_open(struct perf_event_attr * hw_event,pid_t pid,int cpu,int group_fd,unsigned long flags)2589 static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid, int cpu, int group_fd, unsigned long flags)
2590 {
2591 	assert(!no_perf);
2592 
2593 	return syscall(__NR_perf_event_open, hw_event, pid, cpu, group_fd, flags);
2594 }
2595 
open_perf_counter(int cpu,unsigned int type,unsigned int config,int group_fd,__u64 read_format)2596 static long open_perf_counter(int cpu, unsigned int type, unsigned int config, int group_fd, __u64 read_format)
2597 {
2598 	struct perf_event_attr attr;
2599 	const pid_t pid = -1;
2600 	const unsigned long flags = 0;
2601 
2602 	assert(!no_perf);
2603 
2604 	memset(&attr, 0, sizeof(struct perf_event_attr));
2605 
2606 	attr.type = type;
2607 	attr.size = sizeof(struct perf_event_attr);
2608 	attr.config = config;
2609 	attr.disabled = 0;
2610 	attr.sample_type = PERF_SAMPLE_IDENTIFIER;
2611 	attr.read_format = read_format;
2612 
2613 	const int fd = perf_event_open(&attr, pid, cpu, group_fd, flags);
2614 
2615 	return fd;
2616 }
2617 
get_instr_count_fd(int cpu)2618 int get_instr_count_fd(int cpu)
2619 {
2620 	if (fd_instr_count_percpu[cpu])
2621 		return fd_instr_count_percpu[cpu];
2622 
2623 	fd_instr_count_percpu[cpu] = open_perf_counter(cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, -1, 0);
2624 
2625 	return fd_instr_count_percpu[cpu];
2626 }
2627 
get_msr(int cpu,off_t offset,unsigned long long * msr)2628 int get_msr(int cpu, off_t offset, unsigned long long *msr)
2629 {
2630 	ssize_t retval;
2631 
2632 	assert(!no_msr);
2633 
2634 	retval = pread(get_msr_fd(cpu), msr, sizeof(*msr), offset);
2635 
2636 	if (retval != sizeof *msr)
2637 		err(-1, "cpu%d: msr offset 0x%llx read failed", cpu, (unsigned long long)offset);
2638 
2639 	return 0;
2640 }
2641 
add_msr_counter(int cpu,off_t offset)2642 int add_msr_counter(int cpu, off_t offset)
2643 {
2644 	ssize_t retval;
2645 	unsigned long long value;
2646 
2647 	if (no_msr)
2648 		return -1;
2649 
2650 	if (!offset)
2651 		return -1;
2652 
2653 	retval = pread(get_msr_fd(cpu), &value, sizeof(value), offset);
2654 
2655 	/* if the read failed, the probe fails */
2656 	if (retval != sizeof(value))
2657 		return -1;
2658 
2659 	if (value == 0)
2660 		return 0;
2661 
2662 	return 1;
2663 }
2664 
add_rapl_msr_counter(int cpu,const struct rapl_counter_arch_info * cai)2665 int add_rapl_msr_counter(int cpu, const struct rapl_counter_arch_info *cai)
2666 {
2667 	int ret;
2668 
2669 	if (!(valid_rapl_msrs & cai->feature_mask))
2670 		return -1;
2671 
2672 	ret = add_msr_counter(cpu, cai->msr);
2673 	if (ret < 0)
2674 		return -1;
2675 
2676 	switch (cai->rci_index) {
2677 	case RAPL_RCI_INDEX_ENERGY_PKG:
2678 	case RAPL_RCI_INDEX_ENERGY_CORES:
2679 	case RAPL_RCI_INDEX_DRAM:
2680 	case RAPL_RCI_INDEX_GFX:
2681 	case RAPL_RCI_INDEX_ENERGY_PLATFORM:
2682 		if (ret == 0)
2683 			return 1;
2684 	}
2685 
2686 	/* PKG,DRAM_PERF_STATUS MSRs, can return any value */
2687 	return 1;
2688 }
2689 
2690 /* Convert CPU ID to domain ID for given added perf counter. */
cpu_to_domain(const struct perf_counter_info * pc,int cpu)2691 unsigned int cpu_to_domain(const struct perf_counter_info *pc, int cpu)
2692 {
2693 	switch (pc->scope) {
2694 	case SCOPE_CPU:
2695 		return cpu;
2696 
2697 	case SCOPE_CORE:
2698 		return cpus[cpu].core_id;
2699 
2700 	case SCOPE_PACKAGE:
2701 		return cpus[cpu].package_id;
2702 	}
2703 
2704 	__builtin_unreachable();
2705 }
2706 
2707 #define MAX_DEFERRED 16
2708 char *deferred_add_names[MAX_DEFERRED];
2709 char *deferred_skip_names[MAX_DEFERRED];
2710 int deferred_add_index;
2711 int deferred_skip_index;
2712 unsigned int deferred_add_consumed;
2713 unsigned int deferred_skip_consumed;
2714 
2715 /*
2716  * HIDE_LIST - hide this list of counters, show the rest [default]
2717  * SHOW_LIST - show this list of counters, hide the rest
2718  */
2719 enum show_hide_mode { SHOW_LIST, HIDE_LIST } global_show_hide_mode = HIDE_LIST;
2720 
help(void)2721 void help(void)
2722 {
2723 	fprintf(outf,
2724 		"Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
2725 		"\n"
2726 		"Turbostat forks the specified COMMAND and prints statistics\n"
2727 		"when COMMAND completes.\n"
2728 		"If no COMMAND is specified, turbostat wakes every 5-seconds\n"
2729 		"to print statistics, until interrupted.\n"
2730 		"  -a, --add counter\n"
2731 		"		add a counter\n"
2732 		"		  eg. --add msr0x10,u64,cpu,delta,MY_TSC\n"
2733 		"		  eg. --add perf/cstate_pkg/c2-residency,package,delta,percent,perfPC2\n"
2734 		"		  eg. --add pmt,name=XTAL,type=raw,domain=package0,offset=0,lsb=0,msb=63,guid=0x1a067102\n"
2735 		"  -c, --cpu cpu-set\n"
2736 		"		limit output to summary plus cpu-set:\n"
2737 		"		  {core | package | j,k,l..m,n-p }\n"
2738 		"  -d, --debug\n"
2739 		"		displays usec, Time_Of_Day_Seconds and more debugging\n"
2740 		"		debug messages are printed to stderr\n"
2741 		"  -D, --Dump\n"
2742 		"		displays the raw counter values\n"
2743 		"  -e, --enable [all | column]\n"
2744 		"		shows all or the specified disabled column\n"
2745 		"  -f, --force\n"
2746 		"		force load turbostat with minimum default features on unsupported platforms.\n"
2747 		"  -H, --hide [column | column,column,...]\n"
2748 		"		hide the specified column(s)\n"
2749 		"  -i, --interval sec.subsec\n"
2750 		"		override default 5-second measurement interval\n"
2751 		"  -J, --Joules\n"
2752 		"		displays energy in Joules instead of Watts\n"
2753 		"  -l, --list\n"
2754 		"		list column headers only\n"
2755 		"  -M, --no-msr\n"
2756 		"		disable all uses of the MSR driver\n"
2757 		"  -P, --no-perf\n"
2758 		"		disable all uses of the perf API\n"
2759 		"  -n, --num_iterations num\n"
2760 		"		number of the measurement iterations\n"
2761 		"  -N, --header_iterations num\n"
2762 		"		print header every num iterations\n"
2763 		"  -o, --out file\n"
2764 		"		create or truncate \"file\" for all output\n"
2765 		"  -q, --quiet\n"
2766 		"		skip decoding system configuration header\n"
2767 		"  -s, --show [column | column,column,...]\n"
2768 		"		show only the specified column(s)\n"
2769 		"  -S, --Summary\n"
2770 		"		limits output to 1-line system summary per interval\n"
2771 		"  -T, --TCC temperature\n"
2772 		"		sets the Thermal Control Circuit temperature in\n"
2773 		"		  degrees Celsius\n"
2774 		"  -h, --help\n"
2775 		"		print this help message\n  -v, --version\n\t\tprint version information\n\nFor more help, run \"man turbostat\"\n");
2776 }
2777 
2778 /*
2779  * bic_lookup
2780  * for all the strings in comma separate name_list,
2781  * set the approprate bit in return value.
2782  */
bic_lookup(cpu_set_t * ret_set,char * name_list,enum show_hide_mode mode)2783 void bic_lookup(cpu_set_t *ret_set, char *name_list, enum show_hide_mode mode)
2784 {
2785 	unsigned int i;
2786 
2787 	while (name_list) {
2788 		char *comma;
2789 
2790 		comma = strchr(name_list, ',');
2791 
2792 		if (comma)
2793 			*comma = '\0';
2794 
2795 		for (i = 0; i < MAX_BIC; ++i) {
2796 			if (!strcmp(name_list, bic[i].name)) {
2797 				SET_BIC(i, ret_set);
2798 				break;
2799 			}
2800 			if (!strcmp(name_list, "all")) {
2801 				bic_set_all(ret_set);
2802 				break;
2803 			} else if (!strcmp(name_list, "topology")) {
2804 				CPU_OR(ret_set, ret_set, &bic_group_topology);
2805 				break;
2806 			} else if (!strcmp(name_list, "power")) {
2807 				CPU_OR(ret_set, ret_set, &bic_group_thermal_pwr);
2808 				break;
2809 			} else if (!strcmp(name_list, "idle")) {
2810 				CPU_OR(ret_set, ret_set, &bic_group_idle);
2811 				break;
2812 			} else if (!strcmp(name_list, "cache")) {
2813 				CPU_OR(ret_set, ret_set, &bic_group_cache);
2814 				break;
2815 			} else if (!strcmp(name_list, "llc")) {
2816 				CPU_OR(ret_set, ret_set, &bic_group_cache);
2817 				break;
2818 			} else if (!strcmp(name_list, "swidle")) {
2819 				CPU_OR(ret_set, ret_set, &bic_group_sw_idle);
2820 				break;
2821 			} else if (!strcmp(name_list, "sysfs")) {	/* legacy compatibility */
2822 				CPU_OR(ret_set, ret_set, &bic_group_sw_idle);
2823 				break;
2824 			} else if (!strcmp(name_list, "hwidle")) {
2825 				CPU_OR(ret_set, ret_set, &bic_group_hw_idle);
2826 				break;
2827 			} else if (!strcmp(name_list, "frequency")) {
2828 				CPU_OR(ret_set, ret_set, &bic_group_frequency);
2829 				break;
2830 			} else if (!strcmp(name_list, "other")) {
2831 				CPU_OR(ret_set, ret_set, &bic_group_other);
2832 				break;
2833 			}
2834 		}
2835 		if (i == MAX_BIC) {
2836 			if (mode == SHOW_LIST) {
2837 				deferred_add_names[deferred_add_index++] = name_list;
2838 				if (deferred_add_index >= MAX_DEFERRED) {
2839 					fprintf(stderr, "More than max %d un-recognized --add options '%s'\n", MAX_DEFERRED, name_list);
2840 					help();
2841 					exit(1);
2842 				}
2843 			} else {
2844 				deferred_skip_names[deferred_skip_index++] = name_list;
2845 				if (debug)
2846 					fprintf(stderr, "deferred \"%s\"\n", name_list);
2847 				if (deferred_skip_index >= MAX_DEFERRED) {
2848 					fprintf(stderr, "More than max %d un-recognized --skip options '%s'\n", MAX_DEFERRED, name_list);
2849 					help();
2850 					exit(1);
2851 				}
2852 			}
2853 		}
2854 
2855 		name_list = comma;
2856 		if (name_list)
2857 			name_list++;
2858 
2859 	}
2860 }
2861 
2862 /*
2863  * print_name()
2864  * Print column header name for raw 64-bit counter in 16 columns (at least 8-char plus a tab)
2865  * Otherwise, allow the name + tab to fit within 8-coumn tab-stop.
2866  * In both cases, left justififed, just like other turbostat columns,
2867  * to allow the column values to consume the tab.
2868  *
2869  * Yes, 32-bit counters can overflow 8-columns, and
2870  * 64-bit counters can overflow 16-columns, but that is uncommon.
2871  */
print_name(int width,int * printed,char * delim,char * name,enum counter_type type,enum counter_format format)2872 static inline int print_name(int width, int *printed, char *delim, char *name, enum counter_type type, enum counter_format format)
2873 {
2874 	UNUSED(type);
2875 	char *sep = (*printed)++ ? delim : "";
2876 
2877 	if (format == FORMAT_RAW && width >= 64)
2878 		return sprintf(outp, "%s%-8s", sep, name);
2879 	else
2880 		return sprintf(outp, "%s%s", sep, name);
2881 }
2882 
print_hex_value(int width,int * printed,char * delim,unsigned long long value)2883 static inline int print_hex_value(int width, int *printed, char *delim, unsigned long long value)
2884 {
2885 	char *sep = (*printed)++ ? delim : "";
2886 
2887 	if (width <= 32)
2888 		return sprintf(outp, "%s%08llx", sep, value);
2889 	else
2890 		return sprintf(outp, "%s%016llx", sep, value);
2891 }
2892 
print_decimal_value(int width,int * printed,char * delim,unsigned long long value)2893 static inline int print_decimal_value(int width, int *printed, char *delim, unsigned long long value)
2894 {
2895 	char *sep = (*printed)++ ? delim : "";
2896 
2897 	UNUSED(width);
2898 
2899 	return sprintf(outp, "%s%lld", sep, value);
2900 }
2901 
print_float_value(int * printed,char * delim,double value)2902 static inline int print_float_value(int *printed, char *delim, double value)
2903 {
2904 	char *sep = (*printed)++ ? delim : "";
2905 
2906 	return sprintf(outp, "%s%0.2f", sep, value);
2907 }
2908 
print_header(char * delim)2909 void print_header(char *delim)
2910 {
2911 	struct msr_counter *mp;
2912 	struct perf_counter_info *pp;
2913 	struct pmt_counter *ppmt;
2914 	int printed = 0;
2915 
2916 	if (DO_BIC(BIC_USEC))
2917 		outp += sprintf(outp, "%susec", (printed++ ? delim : ""));
2918 	if (DO_BIC(BIC_TOD))
2919 		outp += sprintf(outp, "%sTime_Of_Day_Seconds", (printed++ ? delim : ""));
2920 	if (DO_BIC(BIC_Package))
2921 		outp += sprintf(outp, "%sPackage", (printed++ ? delim : ""));
2922 	if (DO_BIC(BIC_Die))
2923 		outp += sprintf(outp, "%sDie", (printed++ ? delim : ""));
2924 	if (DO_BIC(BIC_L3))
2925 		outp += sprintf(outp, "%sL3", (printed++ ? delim : ""));
2926 	if (DO_BIC(BIC_Node))
2927 		outp += sprintf(outp, "%sNode", (printed++ ? delim : ""));
2928 	if (DO_BIC(BIC_Module))
2929 		outp += sprintf(outp, "%sModule", (printed++ ? delim : ""));
2930 	if (DO_BIC(BIC_Core))
2931 		outp += sprintf(outp, "%sCore", (printed++ ? delim : ""));
2932 	if (DO_BIC(BIC_CPU))
2933 		outp += sprintf(outp, "%sCPU", (printed++ ? delim : ""));
2934 	if (DO_BIC(BIC_APIC))
2935 		outp += sprintf(outp, "%sAPIC", (printed++ ? delim : ""));
2936 	if (DO_BIC(BIC_X2APIC))
2937 		outp += sprintf(outp, "%sX2APIC", (printed++ ? delim : ""));
2938 	if (DO_BIC(BIC_Avg_MHz))
2939 		outp += sprintf(outp, "%sAvg_MHz", (printed++ ? delim : ""));
2940 	if (DO_BIC(BIC_Busy))
2941 		outp += sprintf(outp, "%sBusy%%", (printed++ ? delim : ""));
2942 	if (DO_BIC(BIC_Bzy_MHz))
2943 		outp += sprintf(outp, "%sBzy_MHz", (printed++ ? delim : ""));
2944 	if (DO_BIC(BIC_TSC_MHz))
2945 		outp += sprintf(outp, "%sTSC_MHz", (printed++ ? delim : ""));
2946 
2947 	if (DO_BIC(BIC_IPC))
2948 		outp += sprintf(outp, "%sIPC", (printed++ ? delim : ""));
2949 
2950 	if (DO_BIC(BIC_IRQ)) {
2951 		if (sums_need_wide_columns)
2952 			outp += sprintf(outp, "%s     IRQ", (printed++ ? delim : ""));
2953 		else
2954 			outp += sprintf(outp, "%sIRQ", (printed++ ? delim : ""));
2955 	}
2956 	if (DO_BIC(BIC_NMI)) {
2957 		if (sums_need_wide_columns)
2958 			outp += sprintf(outp, "%s     NMI", (printed++ ? delim : ""));
2959 		else
2960 			outp += sprintf(outp, "%sNMI", (printed++ ? delim : ""));
2961 	}
2962 
2963 	if (DO_BIC(BIC_SMI))
2964 		outp += sprintf(outp, "%sSMI", (printed++ ? delim : ""));
2965 
2966 	if (DO_BIC(BIC_LLC_MRPS))
2967 		outp += sprintf(outp, "%sLLCMRPS", (printed++ ? delim : ""));
2968 
2969 	if (DO_BIC(BIC_LLC_HIT))
2970 		outp += sprintf(outp, "%sLLC%%hit", (printed++ ? delim : ""));
2971 
2972 	if (DO_BIC(BIC_L2_MRPS))
2973 		outp += sprintf(outp, "%sL2MRPS", (printed++ ? delim : ""));
2974 
2975 	if (DO_BIC(BIC_L2_HIT))
2976 		outp += sprintf(outp, "%sL2%%hit", (printed++ ? delim : ""));
2977 
2978 	for (mp = sys.tp; mp; mp = mp->next)
2979 		outp += print_name(mp->width, &printed, delim, mp->name, mp->type, mp->format);
2980 
2981 	for (pp = sys.perf_tp; pp; pp = pp->next)
2982 		outp += print_name(pp->width, &printed, delim, pp->name, pp->type, pp->format);
2983 
2984 	ppmt = sys.pmt_tp;
2985 	while (ppmt) {
2986 		switch (ppmt->type) {
2987 		case PMT_TYPE_RAW:
2988 			outp += print_name(pmt_counter_get_width(ppmt), &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
2989 			break;
2990 
2991 		case PMT_TYPE_XTAL_TIME:
2992 		case PMT_TYPE_TCORE_CLOCK:
2993 			outp += print_name(32, &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
2994 			break;
2995 		}
2996 
2997 		ppmt = ppmt->next;
2998 	}
2999 
3000 	if (DO_BIC(BIC_CPU_c1))
3001 		outp += sprintf(outp, "%sCPU%%c1", (printed++ ? delim : ""));
3002 	if (DO_BIC(BIC_CPU_c3))
3003 		outp += sprintf(outp, "%sCPU%%c3", (printed++ ? delim : ""));
3004 	if (DO_BIC(BIC_CPU_c6))
3005 		outp += sprintf(outp, "%sCPU%%c6", (printed++ ? delim : ""));
3006 	if (DO_BIC(BIC_CPU_c7))
3007 		outp += sprintf(outp, "%sCPU%%c7", (printed++ ? delim : ""));
3008 
3009 	if (DO_BIC(BIC_Mod_c6))
3010 		outp += sprintf(outp, "%sMod%%c6", (printed++ ? delim : ""));
3011 
3012 	if (DO_BIC(BIC_CoreTmp))
3013 		outp += sprintf(outp, "%sCoreTmp", (printed++ ? delim : ""));
3014 
3015 	if (DO_BIC(BIC_CORE_THROT_CNT))
3016 		outp += sprintf(outp, "%sCoreThr", (printed++ ? delim : ""));
3017 
3018 	if (valid_rapl_msrs && !rapl_joules) {
3019 		if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
3020 			outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
3021 	} else if (valid_rapl_msrs && rapl_joules) {
3022 		if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
3023 			outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
3024 	}
3025 
3026 	for (mp = sys.cp; mp; mp = mp->next)
3027 		outp += print_name(mp->width, &printed, delim, mp->name, mp->type, mp->format);
3028 
3029 	for (pp = sys.perf_cp; pp; pp = pp->next)
3030 		outp += print_name(pp->width, &printed, delim, pp->name, pp->type, pp->format);
3031 
3032 	ppmt = sys.pmt_cp;
3033 	while (ppmt) {
3034 		switch (ppmt->type) {
3035 		case PMT_TYPE_RAW:
3036 			outp += print_name(pmt_counter_get_width(ppmt), &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
3037 
3038 			break;
3039 
3040 		case PMT_TYPE_XTAL_TIME:
3041 		case PMT_TYPE_TCORE_CLOCK:
3042 			outp += print_name(32, &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
3043 			break;
3044 		}
3045 
3046 		ppmt = ppmt->next;
3047 	}
3048 	if (DO_BIC(BIC_PkgTmp))
3049 		outp += sprintf(outp, "%sPkgTmp", (printed++ ? delim : ""));
3050 
3051 	if (DO_BIC(BIC_GFX_rc6))
3052 		outp += sprintf(outp, "%sGFX%%rc6", (printed++ ? delim : ""));
3053 
3054 	if (DO_BIC(BIC_GFXMHz))
3055 		outp += sprintf(outp, "%sGFXMHz", (printed++ ? delim : ""));
3056 
3057 	if (DO_BIC(BIC_GFXACTMHz))
3058 		outp += sprintf(outp, "%sGFXAMHz", (printed++ ? delim : ""));
3059 
3060 	if (DO_BIC(BIC_SAM_mc6))
3061 		outp += sprintf(outp, "%sSAM%%mc6", (printed++ ? delim : ""));
3062 
3063 	if (DO_BIC(BIC_SAMMHz))
3064 		outp += sprintf(outp, "%sSAMMHz", (printed++ ? delim : ""));
3065 
3066 	if (DO_BIC(BIC_SAMACTMHz))
3067 		outp += sprintf(outp, "%sSAMAMHz", (printed++ ? delim : ""));
3068 
3069 	if (DO_BIC(BIC_Totl_c0))
3070 		outp += sprintf(outp, "%sTotl%%C0", (printed++ ? delim : ""));
3071 	if (DO_BIC(BIC_Any_c0))
3072 		outp += sprintf(outp, "%sAny%%C0", (printed++ ? delim : ""));
3073 	if (DO_BIC(BIC_GFX_c0))
3074 		outp += sprintf(outp, "%sGFX%%C0", (printed++ ? delim : ""));
3075 	if (DO_BIC(BIC_CPUGFX))
3076 		outp += sprintf(outp, "%sCPUGFX%%", (printed++ ? delim : ""));
3077 
3078 	if (DO_BIC(BIC_Pkgpc2))
3079 		outp += sprintf(outp, "%sPkg%%pc2", (printed++ ? delim : ""));
3080 	if (DO_BIC(BIC_Pkgpc3))
3081 		outp += sprintf(outp, "%sPkg%%pc3", (printed++ ? delim : ""));
3082 	if (DO_BIC(BIC_Pkgpc6))
3083 		outp += sprintf(outp, "%sPkg%%pc6", (printed++ ? delim : ""));
3084 	if (DO_BIC(BIC_Pkgpc7))
3085 		outp += sprintf(outp, "%sPkg%%pc7", (printed++ ? delim : ""));
3086 	if (DO_BIC(BIC_Pkgpc8))
3087 		outp += sprintf(outp, "%sPkg%%pc8", (printed++ ? delim : ""));
3088 	if (DO_BIC(BIC_Pkgpc9))
3089 		outp += sprintf(outp, "%sPkg%%pc9", (printed++ ? delim : ""));
3090 	if (DO_BIC(BIC_Pkgpc10))
3091 		outp += sprintf(outp, "%sPk%%pc10", (printed++ ? delim : ""));
3092 	if (DO_BIC(BIC_Diec6))
3093 		outp += sprintf(outp, "%sDie%%c6", (printed++ ? delim : ""));
3094 	if (DO_BIC(BIC_CPU_LPI))
3095 		outp += sprintf(outp, "%sCPU%%LPI", (printed++ ? delim : ""));
3096 	if (DO_BIC(BIC_SYS_LPI))
3097 		outp += sprintf(outp, "%sSYS%%LPI", (printed++ ? delim : ""));
3098 
3099 	if (!rapl_joules) {
3100 		if (DO_BIC(BIC_PkgWatt))
3101 			outp += sprintf(outp, "%sPkgWatt", (printed++ ? delim : ""));
3102 		if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
3103 			outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
3104 		if (DO_BIC(BIC_GFXWatt))
3105 			outp += sprintf(outp, "%sGFXWatt", (printed++ ? delim : ""));
3106 		if (DO_BIC(BIC_RAMWatt))
3107 			outp += sprintf(outp, "%sRAMWatt", (printed++ ? delim : ""));
3108 		if (DO_BIC(BIC_PKG__))
3109 			outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
3110 		if (DO_BIC(BIC_RAM__))
3111 			outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
3112 	} else {
3113 		if (DO_BIC(BIC_Pkg_J))
3114 			outp += sprintf(outp, "%sPkg_J", (printed++ ? delim : ""));
3115 		if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
3116 			outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
3117 		if (DO_BIC(BIC_GFX_J))
3118 			outp += sprintf(outp, "%sGFX_J", (printed++ ? delim : ""));
3119 		if (DO_BIC(BIC_RAM_J))
3120 			outp += sprintf(outp, "%sRAM_J", (printed++ ? delim : ""));
3121 		if (DO_BIC(BIC_PKG__))
3122 			outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
3123 		if (DO_BIC(BIC_RAM__))
3124 			outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
3125 	}
3126 	if (DO_BIC(BIC_UNCORE_MHZ))
3127 		outp += sprintf(outp, "%sUncMHz", (printed++ ? delim : ""));
3128 
3129 	for (mp = sys.pp; mp; mp = mp->next)
3130 		outp += print_name(mp->width, &printed, delim, mp->name, mp->type, mp->format);
3131 
3132 	for (pp = sys.perf_pp; pp; pp = pp->next)
3133 		outp += print_name(pp->width, &printed, delim, pp->name, pp->type, pp->format);
3134 
3135 	ppmt = sys.pmt_pp;
3136 	while (ppmt) {
3137 		switch (ppmt->type) {
3138 		case PMT_TYPE_RAW:
3139 			outp += print_name(pmt_counter_get_width(ppmt), &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
3140 			break;
3141 
3142 		case PMT_TYPE_XTAL_TIME:
3143 		case PMT_TYPE_TCORE_CLOCK:
3144 			outp += print_name(32, &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
3145 			break;
3146 		}
3147 
3148 		ppmt = ppmt->next;
3149 	}
3150 
3151 	if (DO_BIC(BIC_SysWatt))
3152 		outp += sprintf(outp, "%sSysWatt", (printed++ ? delim : ""));
3153 	if (DO_BIC(BIC_Sys_J))
3154 		outp += sprintf(outp, "%sSys_J", (printed++ ? delim : ""));
3155 
3156 	outp += sprintf(outp, "\n");
3157 }
3158 
3159 /*
3160  * pct(numerator, denominator)
3161  *
3162  * Return sanity checked percentage (100.0 * numerator/denominotor)
3163  *
3164  * n < 0: nan
3165  * d <= 0: nan
3166  * n/d > 1.1: nan
3167  */
pct(double numerator,double denominator)3168 double pct(double numerator, double denominator)
3169 {
3170 	double retval;
3171 
3172 	if (numerator < 0)
3173 		return nan("");
3174 
3175 	if (denominator <= 0)
3176 		return nan("");
3177 
3178 	retval = 100.0 * numerator / denominator;
3179 
3180 	if (retval > 110.0)
3181 		return nan("");
3182 
3183 	return retval;
3184 }
3185 
dump_counters(PER_THREAD_PARAMS)3186 int dump_counters(PER_THREAD_PARAMS)
3187 {
3188 	int i;
3189 	struct msr_counter *mp;
3190 	struct platform_counters *pplat_cnt = p == odd.packages ? &platform_counters_odd : &platform_counters_even;
3191 
3192 	outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
3193 
3194 	if (t) {
3195 		outp += sprintf(outp, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags);
3196 		outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
3197 		outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
3198 		outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
3199 		outp += sprintf(outp, "c1: %016llX\n", t->c1);
3200 
3201 		if (DO_BIC(BIC_IPC))
3202 			outp += sprintf(outp, "IPC: %lld\n", t->instr_count);
3203 
3204 		if (DO_BIC(BIC_IRQ))
3205 			outp += sprintf(outp, "IRQ: %lld\n", t->irq_count);
3206 		if (DO_BIC(BIC_NMI))
3207 			outp += sprintf(outp, "IRQ: %lld\n", t->nmi_count);
3208 		if (DO_BIC(BIC_SMI))
3209 			outp += sprintf(outp, "SMI: %d\n", t->smi_count);
3210 
3211 		outp += sprintf(outp, "LLC refs: %lld", t->llc.references);
3212 		outp += sprintf(outp, "LLC miss: %lld", t->llc.misses);
3213 		outp += sprintf(outp, "LLC Hit%%: %.2f", pct((t->llc.references - t->llc.misses), t->llc.references));
3214 
3215 		outp += sprintf(outp, "L2 refs: %lld", t->l2.references);
3216 		outp += sprintf(outp, "L2 hits: %lld", t->l2.hits);
3217 		outp += sprintf(outp, "L2 Hit%%: %.2f", pct(t->l2.hits, t->l2.references));
3218 
3219 		for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3220 			outp += sprintf(outp, "tADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num, t->counter[i], mp->sp->path);
3221 		}
3222 	}
3223 
3224 	if (c && is_cpu_first_thread_in_core(t, c)) {
3225 		outp += sprintf(outp, "core: 0x%x\n", cpus[t->cpu_id].core_id);
3226 		outp += sprintf(outp, "c3: %016llX\n", c->c3);
3227 		outp += sprintf(outp, "c6: %016llX\n", c->c6);
3228 		outp += sprintf(outp, "c7: %016llX\n", c->c7);
3229 		outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
3230 		outp += sprintf(outp, "cpu_throt_count: %016llX\n", c->core_throt_cnt);
3231 
3232 		const unsigned long long energy_value = c->core_energy.raw_value * c->core_energy.scale;
3233 		const double energy_scale = c->core_energy.scale;
3234 
3235 		if (c->core_energy.unit == RAPL_UNIT_JOULES)
3236 			outp += sprintf(outp, "Joules: %0llX (scale: %lf)\n", energy_value, energy_scale);
3237 
3238 		for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3239 			outp += sprintf(outp, "cADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num, c->counter[i], mp->sp->path);
3240 		}
3241 		outp += sprintf(outp, "mc6_us: %016llX\n", c->mc6_us);
3242 	}
3243 
3244 	if (p && is_cpu_first_core_in_package(t, p)) {
3245 		outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
3246 		outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
3247 		outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
3248 		outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
3249 
3250 		outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
3251 		if (DO_BIC(BIC_Pkgpc3))
3252 			outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
3253 		if (DO_BIC(BIC_Pkgpc6))
3254 			outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
3255 		if (DO_BIC(BIC_Pkgpc7))
3256 			outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
3257 		outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
3258 		outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
3259 		outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
3260 		outp += sprintf(outp, "cpu_lpi: %016llX\n", p->cpu_lpi);
3261 		outp += sprintf(outp, "sys_lpi: %016llX\n", p->sys_lpi);
3262 		outp += sprintf(outp, "Joules PKG: %0llX\n", p->energy_pkg.raw_value);
3263 		outp += sprintf(outp, "Joules COR: %0llX\n", p->energy_cores.raw_value);
3264 		outp += sprintf(outp, "Joules GFX: %0llX\n", p->energy_gfx.raw_value);
3265 		outp += sprintf(outp, "Joules RAM: %0llX\n", p->energy_dram.raw_value);
3266 		outp += sprintf(outp, "Joules PSYS: %0llX\n", pplat_cnt->energy_psys.raw_value);
3267 		outp += sprintf(outp, "Throttle PKG: %0llX\n", p->rapl_pkg_perf_status.raw_value);
3268 		outp += sprintf(outp, "Throttle RAM: %0llX\n", p->rapl_dram_perf_status.raw_value);
3269 		outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
3270 
3271 		for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3272 			outp += sprintf(outp, "pADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num, p->counter[i], mp->sp->path);
3273 		}
3274 	}
3275 
3276 	outp += sprintf(outp, "\n");
3277 
3278 	return 0;
3279 }
3280 
rapl_counter_get_value(const struct rapl_counter * c,enum rapl_unit desired_unit,double interval)3281 double rapl_counter_get_value(const struct rapl_counter *c, enum rapl_unit desired_unit, double interval)
3282 {
3283 	assert(desired_unit != RAPL_UNIT_INVALID);
3284 
3285 	/*
3286 	 * For now we don't expect anything other than joules,
3287 	 * so just simplify the logic.
3288 	 */
3289 	assert(c->unit == RAPL_UNIT_JOULES);
3290 
3291 	const double scaled = c->raw_value * c->scale;
3292 
3293 	if (desired_unit == RAPL_UNIT_WATTS)
3294 		return scaled / interval;
3295 	return scaled;
3296 }
3297 
get_perf_llc_stats(int cpu,struct llc_stats * llc)3298 void get_perf_llc_stats(int cpu, struct llc_stats *llc)
3299 {
3300 	struct read_format {
3301 		unsigned long long num_read;
3302 		struct llc_stats llc;
3303 	} r;
3304 	const ssize_t expected_read_size = sizeof(r);
3305 	ssize_t actual_read_size;
3306 
3307 	actual_read_size = read(fd_llc_percpu[cpu], &r, expected_read_size);
3308 
3309 	if (actual_read_size == -1)
3310 		err(-1, "%s(cpu%d,) %d,,%ld", __func__, cpu, fd_llc_percpu[cpu], expected_read_size);
3311 
3312 	llc->references = r.llc.references;
3313 	llc->misses = r.llc.misses;
3314 	if (actual_read_size != expected_read_size)
3315 		warn("%s: failed to read perf_data (req %zu act %zu)", __func__, expected_read_size, actual_read_size);
3316 }
3317 
get_perf_l2_stats(int cpu,struct l2_stats * l2)3318 void get_perf_l2_stats(int cpu, struct l2_stats *l2)
3319 {
3320 	struct read_format {
3321 		unsigned long long num_read;
3322 		struct l2_stats l2;
3323 	} r;
3324 	const ssize_t expected_read_size = sizeof(r);
3325 	ssize_t actual_read_size;
3326 
3327 	actual_read_size = read(fd_l2_percpu[cpu], &r, expected_read_size);
3328 
3329 	if (actual_read_size == -1)
3330 		err(-1, "%s(cpu%d,) %d,,%ld", __func__, cpu, fd_l2_percpu[cpu], expected_read_size);
3331 
3332 	l2->references = r.l2.references;
3333 	l2->hits = r.l2.hits;
3334 	if (actual_read_size != expected_read_size)
3335 		warn("%s: cpu%d: failed to read(%d) perf_data (req %zu act %zu)", __func__, cpu, fd_l2_percpu[cpu], expected_read_size, actual_read_size);
3336 }
3337 
3338 /*
3339  * column formatting convention & formats
3340  */
format_counters(PER_THREAD_PARAMS)3341 int format_counters(PER_THREAD_PARAMS)
3342 {
3343 	static int count;
3344 
3345 	struct platform_counters *pplat_cnt = NULL;
3346 	double interval_float, tsc;
3347 	char *fmt8 = "%s%.2f";
3348 
3349 	int i;
3350 	struct msr_counter *mp;
3351 	struct perf_counter_info *pp;
3352 	struct pmt_counter *ppmt;
3353 	char *delim = "\t";
3354 	int printed = 0;
3355 
3356 	if (t == average.threads) {
3357 		pplat_cnt = count & 1 ? &platform_counters_odd : &platform_counters_even;
3358 		++count;
3359 	}
3360 
3361 	/* if showing only 1st thread in core and this isn't one, bail out */
3362 	if (show_core_only && !is_cpu_first_thread_in_core(t, c))
3363 		return 0;
3364 
3365 	/* if showing only 1st thread in pkg and this isn't one, bail out */
3366 	if (show_pkg_only && !is_cpu_first_core_in_package(t, p))
3367 		return 0;
3368 
3369 	/*if not summary line and --cpu is used */
3370 	if ((t != average.threads) && (cpu_subset && !CPU_ISSET_S(t->cpu_id, cpu_subset_size, cpu_subset)))
3371 		return 0;
3372 
3373 	if (DO_BIC(BIC_USEC)) {
3374 		/* on each row, print how many usec each timestamp took to gather */
3375 		struct timeval tv;
3376 
3377 		timersub(&t->tv_end, &t->tv_begin, &tv);
3378 		outp += sprintf(outp, "%5ld\t", tv.tv_sec * 1000000 + tv.tv_usec);
3379 	}
3380 
3381 	/* Time_Of_Day_Seconds: on each row, print sec.usec last timestamp taken */
3382 	if (DO_BIC(BIC_TOD))
3383 		outp += sprintf(outp, "%10ld.%06ld\t", t->tv_end.tv_sec, t->tv_end.tv_usec);
3384 
3385 	interval_float = t->tv_delta.tv_sec + t->tv_delta.tv_usec / 1000000.0;
3386 
3387 	tsc = t->tsc * tsc_tweak;
3388 
3389 	/* topo columns, print blanks on 1st (average) line */
3390 	if (t == average.threads) {
3391 		if (DO_BIC(BIC_Package))
3392 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3393 		if (DO_BIC(BIC_Die))
3394 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3395 		if (DO_BIC(BIC_L3))
3396 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3397 		if (DO_BIC(BIC_Node))
3398 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3399 		if (DO_BIC(BIC_Module))
3400 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3401 		if (DO_BIC(BIC_Core))
3402 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3403 		if (DO_BIC(BIC_CPU))
3404 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3405 		if (DO_BIC(BIC_APIC))
3406 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3407 		if (DO_BIC(BIC_X2APIC))
3408 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3409 	} else {
3410 		if (DO_BIC(BIC_Package)) {
3411 			if (p)
3412 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].package_id);
3413 			else
3414 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3415 		}
3416 		if (DO_BIC(BIC_Die)) {
3417 			if (c)
3418 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].die_id);
3419 			else
3420 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3421 		}
3422 		if (DO_BIC(BIC_L3)) {
3423 			if (c)
3424 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].l3_id);
3425 			else
3426 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3427 		}
3428 		if (DO_BIC(BIC_Node)) {
3429 			if (t)
3430 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].physical_node_id);
3431 			else
3432 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3433 		}
3434 		if (DO_BIC(BIC_Module)) {
3435 			if (c)
3436 				outp += sprintf(outp, "%s0x%x", (printed++ ? delim : ""), cpus[t->cpu_id].module_id);
3437 			else
3438 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3439 		}
3440 		if (DO_BIC(BIC_Core)) {
3441 			if (c)
3442 				outp += sprintf(outp, "%s0x%x", (printed++ ? delim : ""), cpus[t->cpu_id].core_id);
3443 			else
3444 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3445 		}
3446 		if (DO_BIC(BIC_CPU))
3447 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->cpu_id);
3448 		if (DO_BIC(BIC_APIC))
3449 			outp += sprintf(outp, "%s0x%x", (printed++ ? delim : ""), t->apic_id);
3450 		if (DO_BIC(BIC_X2APIC))
3451 			outp += sprintf(outp, "%s0x%x", (printed++ ? delim : ""), t->x2apic_id);
3452 	}
3453 
3454 	if (DO_BIC(BIC_Avg_MHz))
3455 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 / units * t->aperf / interval_float);
3456 
3457 	if (DO_BIC(BIC_Busy))
3458 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(t->mperf, tsc));
3459 
3460 	if (DO_BIC(BIC_Bzy_MHz)) {
3461 		if (has_base_hz)
3462 			outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), base_hz / units * t->aperf / t->mperf);
3463 		else
3464 			outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), tsc / units * t->aperf / t->mperf / interval_float);
3465 	}
3466 
3467 	if (DO_BIC(BIC_TSC_MHz))
3468 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 * t->tsc / units / interval_float);
3469 
3470 	if (DO_BIC(BIC_IPC))
3471 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 1.0 * t->instr_count / t->aperf);
3472 
3473 	/* IRQ */
3474 	if (DO_BIC(BIC_IRQ)) {
3475 		if (sums_need_wide_columns)
3476 			outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->irq_count);
3477 		else
3478 			outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->irq_count);
3479 	}
3480 
3481 	/* NMI */
3482 	if (DO_BIC(BIC_NMI)) {
3483 		if (sums_need_wide_columns)
3484 			outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->nmi_count);
3485 		else
3486 			outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->nmi_count);
3487 	}
3488 
3489 	/* SMI */
3490 	if (DO_BIC(BIC_SMI))
3491 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->smi_count);
3492 
3493 	/* LLC Stats */
3494 	if (DO_BIC(BIC_LLC_MRPS))
3495 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), t->llc.references / interval_float / 1000000);
3496 
3497 	if (DO_BIC(BIC_LLC_HIT))
3498 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), pct((t->llc.references - t->llc.misses), t->llc.references));
3499 
3500 	/* L2 Stats */
3501 	if (DO_BIC(BIC_L2_MRPS))
3502 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), t->l2.references / interval_float / 1000000);
3503 
3504 	if (DO_BIC(BIC_L2_HIT))
3505 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), pct(t->l2.hits, t->l2.references));
3506 
3507 	/* Added Thread Counters */
3508 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3509 		if (mp->format == FORMAT_RAW)
3510 			outp += print_hex_value(mp->width, &printed, delim, t->counter[i]);
3511 		else if (mp->format == FORMAT_DELTA || mp->format == FORMAT_AVERAGE)
3512 			outp += print_decimal_value(mp->width, &printed, delim, t->counter[i]);
3513 		else if (mp->format == FORMAT_PERCENT) {
3514 			if (mp->type == COUNTER_USEC)
3515 				outp += print_float_value(&printed, delim, t->counter[i] / interval_float / 10000);
3516 			else
3517 				outp += print_float_value(&printed, delim, pct(t->counter[i], tsc));
3518 		}
3519 	}
3520 
3521 	/* Added perf Thread Counters */
3522 	for (i = 0, pp = sys.perf_tp; pp; ++i, pp = pp->next) {
3523 		if (pp->format == FORMAT_RAW)
3524 			outp += print_hex_value(pp->width, &printed, delim, t->perf_counter[i]);
3525 		else if (pp->format == FORMAT_DELTA || pp->format == FORMAT_AVERAGE)
3526 			outp += print_decimal_value(pp->width, &printed, delim, t->perf_counter[i]);
3527 		else if (pp->format == FORMAT_PERCENT) {
3528 			if (pp->type == COUNTER_USEC)
3529 				outp += print_float_value(&printed, delim, t->perf_counter[i] / interval_float / 10000);
3530 			else
3531 				outp += print_float_value(&printed, delim, pct(t->perf_counter[i], tsc));
3532 		}
3533 	}
3534 
3535 	/* Added PMT Thread Counters */
3536 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3537 		const unsigned long value_raw = t->pmt_counter[i];
3538 		double value_converted;
3539 		switch (ppmt->type) {
3540 		case PMT_TYPE_RAW:
3541 			outp += print_hex_value(pmt_counter_get_width(ppmt), &printed, delim, t->pmt_counter[i]);
3542 			break;
3543 
3544 		case PMT_TYPE_XTAL_TIME:
3545 			value_converted = pct(value_raw / crystal_hz, interval_float);
3546 			outp += print_float_value(&printed, delim, value_converted);
3547 			break;
3548 
3549 		case PMT_TYPE_TCORE_CLOCK:
3550 			value_converted = pct(value_raw / tcore_clock_freq_hz, interval_float);
3551 			outp += print_float_value(&printed, delim, value_converted);
3552 		}
3553 	}
3554 
3555 	/* C1 */
3556 	if (DO_BIC(BIC_CPU_c1))
3557 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(t->c1, tsc));
3558 
3559 	/* print per-core data only for 1st thread in core */
3560 	if (!is_cpu_first_thread_in_core(t, c))
3561 		goto done;
3562 
3563 	if (DO_BIC(BIC_CPU_c3))
3564 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(c->c3, tsc));
3565 	if (DO_BIC(BIC_CPU_c6))
3566 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(c->c6, tsc));
3567 	if (DO_BIC(BIC_CPU_c7))
3568 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(c->c7, tsc));
3569 
3570 	/* Mod%c6 */
3571 	if (DO_BIC(BIC_Mod_c6))
3572 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(c->mc6_us, tsc));
3573 
3574 	if (DO_BIC(BIC_CoreTmp))
3575 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_temp_c);
3576 
3577 	/* Core throttle count */
3578 	if (DO_BIC(BIC_CORE_THROT_CNT))
3579 		outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->core_throt_cnt);
3580 
3581 	/* Added Core Counters */
3582 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3583 		if (mp->format == FORMAT_RAW)
3584 			outp += print_hex_value(mp->width, &printed, delim, c->counter[i]);
3585 		else if (mp->format == FORMAT_DELTA || mp->format == FORMAT_AVERAGE)
3586 			outp += print_decimal_value(mp->width, &printed, delim, c->counter[i]);
3587 		else if (mp->format == FORMAT_PERCENT)
3588 			outp += print_float_value(&printed, delim, pct(c->counter[i], tsc));
3589 	}
3590 
3591 	/* Added perf Core counters */
3592 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3593 		if (pp->format == FORMAT_RAW)
3594 			outp += print_hex_value(pp->width, &printed, delim, c->perf_counter[i]);
3595 		else if (pp->format == FORMAT_DELTA || pp->format == FORMAT_AVERAGE)
3596 			outp += print_decimal_value(pp->width, &printed, delim, c->perf_counter[i]);
3597 		else if (pp->format == FORMAT_PERCENT)
3598 			outp += print_float_value(&printed, delim, pct(c->perf_counter[i], tsc));
3599 	}
3600 
3601 	/* Added PMT Core counters */
3602 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3603 		const unsigned long value_raw = c->pmt_counter[i];
3604 		double value_converted;
3605 		switch (ppmt->type) {
3606 		case PMT_TYPE_RAW:
3607 			outp += print_hex_value(pmt_counter_get_width(ppmt), &printed, delim, c->pmt_counter[i]);
3608 			break;
3609 
3610 		case PMT_TYPE_XTAL_TIME:
3611 			value_converted = pct(value_raw / crystal_hz, interval_float);
3612 			outp += print_float_value(&printed, delim, value_converted);
3613 			break;
3614 
3615 		case PMT_TYPE_TCORE_CLOCK:
3616 			value_converted = pct(value_raw / tcore_clock_freq_hz, interval_float);
3617 			outp += print_float_value(&printed, delim, value_converted);
3618 		}
3619 	}
3620 
3621 	if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
3622 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&c->core_energy, RAPL_UNIT_WATTS, interval_float));
3623 	if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
3624 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&c->core_energy, RAPL_UNIT_JOULES, interval_float));
3625 
3626 	/* print per-package data only for 1st core in package */
3627 	if (!is_cpu_first_core_in_package(t, p))
3628 		goto done;
3629 
3630 	/* PkgTmp */
3631 	if (DO_BIC(BIC_PkgTmp))
3632 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->pkg_temp_c);
3633 
3634 	/* GFXrc6 */
3635 	if (DO_BIC(BIC_GFX_rc6)) {
3636 		if (p->gfx_rc6_ms == -1) {	/* detect GFX counter reset */
3637 			outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
3638 		} else {
3639 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), p->gfx_rc6_ms / 10.0 / interval_float);
3640 		}
3641 	}
3642 
3643 	/* GFXMHz */
3644 	if (DO_BIC(BIC_GFXMHz))
3645 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_mhz);
3646 
3647 	/* GFXACTMHz */
3648 	if (DO_BIC(BIC_GFXACTMHz))
3649 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_act_mhz);
3650 
3651 	/* SAMmc6 */
3652 	if (DO_BIC(BIC_SAM_mc6)) {
3653 		if (p->sam_mc6_ms == -1) {	/* detect GFX counter reset */
3654 			outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
3655 		} else {
3656 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), p->sam_mc6_ms / 10.0 / interval_float);
3657 		}
3658 	}
3659 
3660 	/* SAMMHz */
3661 	if (DO_BIC(BIC_SAMMHz))
3662 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->sam_mhz);
3663 
3664 	/* SAMACTMHz */
3665 	if (DO_BIC(BIC_SAMACTMHz))
3666 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->sam_act_mhz);
3667 
3668 	/* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
3669 	if (DO_BIC(BIC_Totl_c0))
3670 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100 * p->pkg_wtd_core_c0 / tsc);	/* can exceed 100% */
3671 	if (DO_BIC(BIC_Any_c0))
3672 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pkg_any_core_c0, tsc));
3673 	if (DO_BIC(BIC_GFX_c0))
3674 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pkg_any_gfxe_c0, tsc));
3675 	if (DO_BIC(BIC_CPUGFX))
3676 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pkg_both_core_gfxe_c0, tsc));
3677 
3678 	if (DO_BIC(BIC_Pkgpc2))
3679 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc2, tsc));
3680 	if (DO_BIC(BIC_Pkgpc3))
3681 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc3, tsc));
3682 	if (DO_BIC(BIC_Pkgpc6))
3683 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc6, tsc));
3684 	if (DO_BIC(BIC_Pkgpc7))
3685 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc7, tsc));
3686 	if (DO_BIC(BIC_Pkgpc8))
3687 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc8, tsc));
3688 	if (DO_BIC(BIC_Pkgpc9))
3689 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc9, tsc));
3690 	if (DO_BIC(BIC_Pkgpc10))
3691 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc10, tsc));
3692 
3693 	if (DO_BIC(BIC_Diec6))
3694 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->die_c6 / crystal_hz, interval_float));
3695 
3696 	if (DO_BIC(BIC_CPU_LPI)) {
3697 		if (p->cpu_lpi >= 0)
3698 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->cpu_lpi / 1000000.0, interval_float));
3699 		else
3700 			outp += sprintf(outp, "%s(neg)", (printed++ ? delim : ""));
3701 	}
3702 	if (DO_BIC(BIC_SYS_LPI)) {
3703 		if (p->sys_lpi >= 0)
3704 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->sys_lpi / 1000000.0, interval_float));
3705 		else
3706 			outp += sprintf(outp, "%s(neg)", (printed++ ? delim : ""));
3707 	}
3708 
3709 	if (DO_BIC(BIC_PkgWatt))
3710 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_WATTS, interval_float));
3711 	if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
3712 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_WATTS, interval_float));
3713 	if (DO_BIC(BIC_GFXWatt))
3714 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_WATTS, interval_float));
3715 	if (DO_BIC(BIC_RAMWatt))
3716 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_WATTS, interval_float));
3717 	if (DO_BIC(BIC_Pkg_J))
3718 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_JOULES, interval_float));
3719 	if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
3720 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_JOULES, interval_float));
3721 	if (DO_BIC(BIC_GFX_J))
3722 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_JOULES, interval_float));
3723 	if (DO_BIC(BIC_RAM_J))
3724 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_JOULES, interval_float));
3725 	if (DO_BIC(BIC_PKG__))
3726 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->rapl_pkg_perf_status, RAPL_UNIT_WATTS, interval_float));
3727 	if (DO_BIC(BIC_RAM__))
3728 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->rapl_dram_perf_status, RAPL_UNIT_WATTS, interval_float));
3729 	/* UncMHz */
3730 	if (DO_BIC(BIC_UNCORE_MHZ))
3731 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->uncore_mhz);
3732 
3733 	/* Added Package Counters */
3734 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3735 		if (mp->format == FORMAT_RAW)
3736 			outp += print_hex_value(mp->width, &printed, delim, p->counter[i]);
3737 		else if (mp->type == COUNTER_K2M)
3738 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), (unsigned int)p->counter[i] / 1000);
3739 		else if (mp->format == FORMAT_DELTA || mp->format == FORMAT_AVERAGE)
3740 			outp += print_decimal_value(mp->width, &printed, delim, p->counter[i]);
3741 		else if (mp->format == FORMAT_PERCENT)
3742 			outp += print_float_value(&printed, delim, pct(p->counter[i], tsc));
3743 	}
3744 
3745 	/* Added perf Package Counters */
3746 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3747 		if (pp->format == FORMAT_RAW)
3748 			outp += print_hex_value(pp->width, &printed, delim, p->perf_counter[i]);
3749 		else if (pp->type == COUNTER_K2M)
3750 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), (unsigned int)p->perf_counter[i] / 1000);
3751 		else if (pp->format == FORMAT_DELTA || pp->format == FORMAT_AVERAGE)
3752 			outp += print_decimal_value(pp->width, &printed, delim, p->perf_counter[i]);
3753 		else if (pp->format == FORMAT_PERCENT)
3754 			outp += print_float_value(&printed, delim, pct(p->perf_counter[i], tsc));
3755 	}
3756 
3757 	/* Added PMT Package Counters */
3758 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3759 		const unsigned long value_raw = p->pmt_counter[i];
3760 		double value_converted;
3761 		switch (ppmt->type) {
3762 		case PMT_TYPE_RAW:
3763 			outp += print_hex_value(pmt_counter_get_width(ppmt), &printed, delim, p->pmt_counter[i]);
3764 			break;
3765 
3766 		case PMT_TYPE_XTAL_TIME:
3767 			value_converted = pct(value_raw / crystal_hz, interval_float);
3768 			outp += print_float_value(&printed, delim, value_converted);
3769 			break;
3770 
3771 		case PMT_TYPE_TCORE_CLOCK:
3772 			value_converted = pct(value_raw / tcore_clock_freq_hz, interval_float);
3773 			outp += print_float_value(&printed, delim, value_converted);
3774 		}
3775 	}
3776 
3777 	if (DO_BIC(BIC_SysWatt) && (t == average.threads))
3778 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&pplat_cnt->energy_psys, RAPL_UNIT_WATTS, interval_float));
3779 	if (DO_BIC(BIC_Sys_J) && (t == average.threads))
3780 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&pplat_cnt->energy_psys, RAPL_UNIT_JOULES, interval_float));
3781 
3782 done:
3783 	if (*(outp - 1) != '\n')
3784 		outp += sprintf(outp, "\n");
3785 
3786 	return 0;
3787 }
3788 
flush_output_stdout(void)3789 void flush_output_stdout(void)
3790 {
3791 	FILE *filep;
3792 
3793 	if (outf == stderr)
3794 		filep = stdout;
3795 	else
3796 		filep = outf;
3797 
3798 	fputs(output_buffer, filep);
3799 	fflush(filep);
3800 
3801 	outp = output_buffer;
3802 }
3803 
flush_output_stderr(void)3804 void flush_output_stderr(void)
3805 {
3806 	fputs(output_buffer, outf);
3807 	fflush(outf);
3808 	outp = output_buffer;
3809 }
3810 
format_all_counters(PER_THREAD_PARAMS)3811 void format_all_counters(PER_THREAD_PARAMS)
3812 {
3813 	static int count;
3814 
3815 	if ((!count || (header_iterations && !(count % header_iterations))) || !summary_only)
3816 		print_header("\t");
3817 
3818 	format_counters(average.threads, average.cores, average.packages);
3819 
3820 	count++;
3821 
3822 	if (summary_only)
3823 		return;
3824 
3825 	for_all_cpus(format_counters, t, c, p);
3826 }
3827 
3828 #define DELTA_WRAP32(new, old)			\
3829 	old = ((((unsigned long long)new << 32) - ((unsigned long long)old << 32)) >> 32);
3830 
delta_package(struct pkg_data * new,struct pkg_data * old)3831 int delta_package(struct pkg_data *new, struct pkg_data *old)
3832 {
3833 	int i;
3834 	struct msr_counter *mp;
3835 	struct perf_counter_info *pp;
3836 	struct pmt_counter *ppmt;
3837 
3838 	if (DO_BIC(BIC_Totl_c0))
3839 		old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
3840 	if (DO_BIC(BIC_Any_c0))
3841 		old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
3842 	if (DO_BIC(BIC_GFX_c0))
3843 		old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
3844 	if (DO_BIC(BIC_CPUGFX))
3845 		old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
3846 
3847 	old->pc2 = new->pc2 - old->pc2;
3848 	if (DO_BIC(BIC_Pkgpc3))
3849 		old->pc3 = new->pc3 - old->pc3;
3850 	if (DO_BIC(BIC_Pkgpc6))
3851 		old->pc6 = new->pc6 - old->pc6;
3852 	if (DO_BIC(BIC_Pkgpc7))
3853 		old->pc7 = new->pc7 - old->pc7;
3854 	old->pc8 = new->pc8 - old->pc8;
3855 	old->pc9 = new->pc9 - old->pc9;
3856 	old->pc10 = new->pc10 - old->pc10;
3857 	old->die_c6 = new->die_c6 - old->die_c6;
3858 	old->cpu_lpi = new->cpu_lpi - old->cpu_lpi;
3859 	old->sys_lpi = new->sys_lpi - old->sys_lpi;
3860 	old->pkg_temp_c = new->pkg_temp_c;
3861 
3862 	/* flag an error when rc6 counter resets/wraps */
3863 	if (old->gfx_rc6_ms > new->gfx_rc6_ms)
3864 		old->gfx_rc6_ms = -1;
3865 	else
3866 		old->gfx_rc6_ms = new->gfx_rc6_ms - old->gfx_rc6_ms;
3867 
3868 	old->uncore_mhz = new->uncore_mhz;
3869 	old->gfx_mhz = new->gfx_mhz;
3870 	old->gfx_act_mhz = new->gfx_act_mhz;
3871 
3872 	/* flag an error when mc6 counter resets/wraps */
3873 	if (old->sam_mc6_ms > new->sam_mc6_ms)
3874 		old->sam_mc6_ms = -1;
3875 	else
3876 		old->sam_mc6_ms = new->sam_mc6_ms - old->sam_mc6_ms;
3877 
3878 	old->sam_mhz = new->sam_mhz;
3879 	old->sam_act_mhz = new->sam_act_mhz;
3880 
3881 	old->energy_pkg.raw_value = new->energy_pkg.raw_value - old->energy_pkg.raw_value;
3882 	old->energy_cores.raw_value = new->energy_cores.raw_value - old->energy_cores.raw_value;
3883 	old->energy_gfx.raw_value = new->energy_gfx.raw_value - old->energy_gfx.raw_value;
3884 	old->energy_dram.raw_value = new->energy_dram.raw_value - old->energy_dram.raw_value;
3885 	old->rapl_pkg_perf_status.raw_value = new->rapl_pkg_perf_status.raw_value - old->rapl_pkg_perf_status.raw_value;
3886 	old->rapl_dram_perf_status.raw_value = new->rapl_dram_perf_status.raw_value - old->rapl_dram_perf_status.raw_value;
3887 
3888 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3889 		if (mp->format == FORMAT_RAW)
3890 			old->counter[i] = new->counter[i];
3891 		else if (mp->format == FORMAT_AVERAGE)
3892 			old->counter[i] = new->counter[i];
3893 		else
3894 			old->counter[i] = new->counter[i] - old->counter[i];
3895 	}
3896 
3897 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3898 		if (pp->format == FORMAT_RAW)
3899 			old->perf_counter[i] = new->perf_counter[i];
3900 		else if (pp->format == FORMAT_AVERAGE)
3901 			old->perf_counter[i] = new->perf_counter[i];
3902 		else
3903 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3904 	}
3905 
3906 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3907 		if (ppmt->format == FORMAT_RAW)
3908 			old->pmt_counter[i] = new->pmt_counter[i];
3909 		else
3910 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3911 	}
3912 
3913 	return 0;
3914 }
3915 
delta_core(struct core_data * new,struct core_data * old)3916 void delta_core(struct core_data *new, struct core_data *old)
3917 {
3918 	int i;
3919 	struct msr_counter *mp;
3920 	struct perf_counter_info *pp;
3921 	struct pmt_counter *ppmt;
3922 
3923 	old->c3 = new->c3 - old->c3;
3924 	old->c6 = new->c6 - old->c6;
3925 	old->c7 = new->c7 - old->c7;
3926 	old->core_temp_c = new->core_temp_c;
3927 	old->core_throt_cnt = new->core_throt_cnt - old->core_throt_cnt;
3928 	old->mc6_us = new->mc6_us - old->mc6_us;
3929 
3930 	DELTA_WRAP32(new->core_energy.raw_value, old->core_energy.raw_value);
3931 
3932 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3933 		if (mp->format == FORMAT_RAW || mp->format == FORMAT_AVERAGE)
3934 			old->counter[i] = new->counter[i];
3935 		else
3936 			old->counter[i] = new->counter[i] - old->counter[i];
3937 	}
3938 
3939 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3940 		if (pp->format == FORMAT_RAW)
3941 			old->perf_counter[i] = new->perf_counter[i];
3942 		else
3943 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3944 	}
3945 
3946 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3947 		if (ppmt->format == FORMAT_RAW)
3948 			old->pmt_counter[i] = new->pmt_counter[i];
3949 		else
3950 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3951 	}
3952 }
3953 
soft_c1_residency_display(int bic)3954 int soft_c1_residency_display(int bic)
3955 {
3956 	if (!DO_BIC(BIC_CPU_c1) || platform->has_msr_core_c1_res)
3957 		return 0;
3958 
3959 	return DO_BIC_READ(bic);
3960 }
3961 
3962 /*
3963  * old = new - old
3964  */
delta_thread(struct thread_data * new,struct thread_data * old,struct core_data * core_delta)3965 int delta_thread(struct thread_data *new, struct thread_data *old, struct core_data *core_delta)
3966 {
3967 	int i;
3968 	struct msr_counter *mp;
3969 	struct perf_counter_info *pp;
3970 	struct pmt_counter *ppmt;
3971 
3972 	/* we run cpuid just the 1st time, copy the results */
3973 	if (DO_BIC(BIC_APIC))
3974 		new->apic_id = old->apic_id;
3975 	if (DO_BIC(BIC_X2APIC))
3976 		new->x2apic_id = old->x2apic_id;
3977 
3978 	/*
3979 	 * the timestamps from start of measurement interval are in "old"
3980 	 * the timestamp from end of measurement interval are in "new"
3981 	 * over-write old w/ new so we can print end of interval values
3982 	 */
3983 
3984 	timersub(&new->tv_begin, &old->tv_begin, &old->tv_delta);
3985 	old->tv_begin = new->tv_begin;
3986 	old->tv_end = new->tv_end;
3987 
3988 	old->tsc = new->tsc - old->tsc;
3989 
3990 	/* check for TSC < 1 Mcycles over interval */
3991 	if (old->tsc < (1000 * 1000))
3992 		errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
3993 		     "You can disable all c-states by booting with \"idle=poll\"\nor just the deep ones with \"processor.max_cstate=1\"");
3994 
3995 	old->c1 = new->c1 - old->c1;
3996 
3997 	if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz) || DO_BIC(BIC_IPC)
3998 	    || soft_c1_residency_display(BIC_Avg_MHz)) {
3999 		if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
4000 			old->aperf = new->aperf - old->aperf;
4001 			old->mperf = new->mperf - old->mperf;
4002 		} else {
4003 			return -1;
4004 		}
4005 	}
4006 
4007 	if (platform->has_msr_core_c1_res) {
4008 		/*
4009 		 * Some models have a dedicated C1 residency MSR,
4010 		 * which should be more accurate than the derivation below.
4011 		 */
4012 	} else {
4013 		/*
4014 		 * As counter collection is not atomic,
4015 		 * it is possible for mperf's non-halted cycles + idle states
4016 		 * to exceed TSC's all cycles: show c1 = 0% in that case.
4017 		 */
4018 		if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > (old->tsc * tsc_tweak))
4019 			old->c1 = 0;
4020 		else {
4021 			/* normal case, derive c1 */
4022 			old->c1 = (old->tsc * tsc_tweak) - old->mperf - core_delta->c3 - core_delta->c6 - core_delta->c7;
4023 		}
4024 	}
4025 
4026 	if (old->mperf == 0) {
4027 		if (debug > 1)
4028 			fprintf(outf, "cpu%d MPERF 0!\n", old->cpu_id);
4029 		old->mperf = 1;	/* divide by 0 protection */
4030 	}
4031 
4032 	if (DO_BIC(BIC_IPC))
4033 		old->instr_count = new->instr_count - old->instr_count;
4034 
4035 	if (DO_BIC(BIC_IRQ))
4036 		old->irq_count = new->irq_count - old->irq_count;
4037 
4038 	if (DO_BIC(BIC_NMI))
4039 		old->nmi_count = new->nmi_count - old->nmi_count;
4040 
4041 	if (DO_BIC(BIC_SMI))
4042 		old->smi_count = new->smi_count - old->smi_count;
4043 
4044 	if (DO_BIC(BIC_LLC_MRPS) || DO_BIC(BIC_LLC_HIT))
4045 		old->llc.references = new->llc.references - old->llc.references;
4046 
4047 	if (DO_BIC(BIC_LLC_HIT))
4048 		old->llc.misses = new->llc.misses - old->llc.misses;
4049 
4050 	if (DO_BIC(BIC_L2_MRPS) || DO_BIC(BIC_L2_HIT))
4051 		old->l2.references = new->l2.references - old->l2.references;
4052 
4053 	if (DO_BIC(BIC_L2_HIT))
4054 		old->l2.hits = new->l2.hits - old->l2.hits;
4055 
4056 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
4057 		if (mp->format == FORMAT_RAW || mp->format == FORMAT_AVERAGE)
4058 			old->counter[i] = new->counter[i];
4059 		else
4060 			old->counter[i] = new->counter[i] - old->counter[i];
4061 	}
4062 
4063 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
4064 		if (pp->format == FORMAT_RAW)
4065 			old->perf_counter[i] = new->perf_counter[i];
4066 		else
4067 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
4068 	}
4069 
4070 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
4071 		if (ppmt->format == FORMAT_RAW)
4072 			old->pmt_counter[i] = new->pmt_counter[i];
4073 		else
4074 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
4075 	}
4076 
4077 	return 0;
4078 }
4079 
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)4080 int 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)
4081 {
4082 	int retval = 0;
4083 
4084 	/* calculate core delta only for 1st thread in core */
4085 	if (is_cpu_first_thread_in_core(t, c))
4086 		delta_core(c, c2);
4087 
4088 	/* always calculate thread delta */
4089 	retval = delta_thread(t, t2, c2);	/* c2 is core delta */
4090 
4091 	/* calculate package delta only for 1st core in package */
4092 	if (is_cpu_first_core_in_package(t, p))
4093 		retval |= delta_package(p, p2);
4094 
4095 	return retval;
4096 }
4097 
delta_platform(struct platform_counters * new,struct platform_counters * old)4098 void delta_platform(struct platform_counters *new, struct platform_counters *old)
4099 {
4100 	old->energy_psys.raw_value = new->energy_psys.raw_value - old->energy_psys.raw_value;
4101 }
4102 
rapl_counter_clear(struct rapl_counter * c)4103 void rapl_counter_clear(struct rapl_counter *c)
4104 {
4105 	c->raw_value = 0;
4106 	c->scale = 0.0;
4107 	c->unit = RAPL_UNIT_INVALID;
4108 }
4109 
clear_counters(PER_THREAD_PARAMS)4110 void clear_counters(PER_THREAD_PARAMS)
4111 {
4112 	int i;
4113 	struct msr_counter *mp;
4114 
4115 	t->tv_begin.tv_sec = 0;
4116 	t->tv_begin.tv_usec = 0;
4117 	t->tv_end.tv_sec = 0;
4118 	t->tv_end.tv_usec = 0;
4119 	t->tv_delta.tv_sec = 0;
4120 	t->tv_delta.tv_usec = 0;
4121 
4122 	t->tsc = 0;
4123 	t->aperf = 0;
4124 	t->mperf = 0;
4125 	t->c1 = 0;
4126 
4127 	t->instr_count = 0;
4128 
4129 	t->irq_count = 0;
4130 	t->nmi_count = 0;
4131 	t->smi_count = 0;
4132 
4133 	t->llc.references = 0;
4134 	t->llc.misses = 0;
4135 
4136 	t->l2.references = 0;
4137 	t->l2.hits = 0;
4138 
4139 	c->c3 = 0;
4140 	c->c6 = 0;
4141 	c->c7 = 0;
4142 	c->mc6_us = 0;
4143 	c->core_temp_c = 0;
4144 	rapl_counter_clear(&c->core_energy);
4145 	c->core_throt_cnt = 0;
4146 
4147 	p->pkg_wtd_core_c0 = 0;
4148 	p->pkg_any_core_c0 = 0;
4149 	p->pkg_any_gfxe_c0 = 0;
4150 	p->pkg_both_core_gfxe_c0 = 0;
4151 
4152 	p->pc2 = 0;
4153 	if (DO_BIC(BIC_Pkgpc3))
4154 		p->pc3 = 0;
4155 	if (DO_BIC(BIC_Pkgpc6))
4156 		p->pc6 = 0;
4157 	if (DO_BIC(BIC_Pkgpc7))
4158 		p->pc7 = 0;
4159 	p->pc8 = 0;
4160 	p->pc9 = 0;
4161 	p->pc10 = 0;
4162 	p->die_c6 = 0;
4163 	p->cpu_lpi = 0;
4164 	p->sys_lpi = 0;
4165 
4166 	rapl_counter_clear(&p->energy_pkg);
4167 	rapl_counter_clear(&p->energy_dram);
4168 	rapl_counter_clear(&p->energy_cores);
4169 	rapl_counter_clear(&p->energy_gfx);
4170 	rapl_counter_clear(&p->rapl_pkg_perf_status);
4171 	rapl_counter_clear(&p->rapl_dram_perf_status);
4172 	p->pkg_temp_c = 0;
4173 
4174 	p->gfx_rc6_ms = 0;
4175 	p->uncore_mhz = 0;
4176 	p->gfx_mhz = 0;
4177 	p->gfx_act_mhz = 0;
4178 	p->sam_mc6_ms = 0;
4179 	p->sam_mhz = 0;
4180 	p->sam_act_mhz = 0;
4181 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next)
4182 		t->counter[i] = 0;
4183 
4184 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next)
4185 		c->counter[i] = 0;
4186 
4187 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next)
4188 		p->counter[i] = 0;
4189 
4190 	memset(&t->perf_counter[0], 0, sizeof(t->perf_counter));
4191 	memset(&c->perf_counter[0], 0, sizeof(c->perf_counter));
4192 	memset(&p->perf_counter[0], 0, sizeof(p->perf_counter));
4193 
4194 	memset(&t->pmt_counter[0], 0, ARRAY_SIZE(t->pmt_counter));
4195 	memset(&c->pmt_counter[0], 0, ARRAY_SIZE(c->pmt_counter));
4196 	memset(&p->pmt_counter[0], 0, ARRAY_SIZE(p->pmt_counter));
4197 }
4198 
rapl_counter_accumulate(struct rapl_counter * dst,const struct rapl_counter * src)4199 void rapl_counter_accumulate(struct rapl_counter *dst, const struct rapl_counter *src)
4200 {
4201 	/* Copy unit and scale from src if dst is not initialized */
4202 	if (dst->unit == RAPL_UNIT_INVALID) {
4203 		dst->unit = src->unit;
4204 		dst->scale = src->scale;
4205 	}
4206 
4207 	assert(dst->unit == src->unit);
4208 	assert(dst->scale == src->scale);
4209 
4210 	dst->raw_value += src->raw_value;
4211 }
4212 
sum_counters(PER_THREAD_PARAMS)4213 int sum_counters(PER_THREAD_PARAMS)
4214 {
4215 	int i;
4216 	struct msr_counter *mp;
4217 	struct perf_counter_info *pp;
4218 	struct pmt_counter *ppmt;
4219 
4220 	/* copy un-changing apic_id's */
4221 	if (DO_BIC(BIC_APIC))
4222 		average.threads->apic_id = t->apic_id;
4223 	if (DO_BIC(BIC_X2APIC))
4224 		average.threads->x2apic_id = t->x2apic_id;
4225 
4226 	/* remember first tv_begin */
4227 	if (average.threads->tv_begin.tv_sec == 0)
4228 		average.threads->tv_begin = procsysfs_tv_begin;
4229 
4230 	/* remember last tv_end */
4231 	average.threads->tv_end = t->tv_end;
4232 
4233 	average.threads->tsc += t->tsc;
4234 	average.threads->aperf += t->aperf;
4235 	average.threads->mperf += t->mperf;
4236 	average.threads->c1 += t->c1;
4237 
4238 	average.threads->instr_count += t->instr_count;
4239 
4240 	average.threads->irq_count += t->irq_count;
4241 	average.threads->nmi_count += t->nmi_count;
4242 	average.threads->smi_count += t->smi_count;
4243 
4244 	average.threads->llc.references += t->llc.references;
4245 	average.threads->llc.misses += t->llc.misses;
4246 
4247 	average.threads->l2.references += t->l2.references;
4248 	average.threads->l2.hits += t->l2.hits;
4249 
4250 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
4251 		if (mp->format == FORMAT_RAW)
4252 			continue;
4253 		average.threads->counter[i] += t->counter[i];
4254 	}
4255 
4256 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
4257 		if (pp->format == FORMAT_RAW)
4258 			continue;
4259 		average.threads->perf_counter[i] += t->perf_counter[i];
4260 	}
4261 
4262 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
4263 		average.threads->pmt_counter[i] += t->pmt_counter[i];
4264 	}
4265 
4266 	/* sum per-core values only for 1st thread in core */
4267 	if (!is_cpu_first_thread_in_core(t, c))
4268 		return 0;
4269 
4270 	average.cores->c3 += c->c3;
4271 	average.cores->c6 += c->c6;
4272 	average.cores->c7 += c->c7;
4273 	average.cores->mc6_us += c->mc6_us;
4274 
4275 	average.cores->core_temp_c = MAX(average.cores->core_temp_c, c->core_temp_c);
4276 	average.cores->core_throt_cnt = MAX(average.cores->core_throt_cnt, c->core_throt_cnt);
4277 
4278 	rapl_counter_accumulate(&average.cores->core_energy, &c->core_energy);
4279 
4280 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
4281 		if (mp->format == FORMAT_RAW)
4282 			continue;
4283 		average.cores->counter[i] += c->counter[i];
4284 	}
4285 
4286 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
4287 		if (pp->format == FORMAT_RAW)
4288 			continue;
4289 		average.cores->perf_counter[i] += c->perf_counter[i];
4290 	}
4291 
4292 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
4293 		average.cores->pmt_counter[i] += c->pmt_counter[i];
4294 	}
4295 
4296 	/* sum per-pkg values only for 1st core in pkg */
4297 	if (!is_cpu_first_core_in_package(t, p))
4298 		return 0;
4299 
4300 	if (DO_BIC(BIC_Totl_c0))
4301 		average.packages->pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
4302 	if (DO_BIC(BIC_Any_c0))
4303 		average.packages->pkg_any_core_c0 += p->pkg_any_core_c0;
4304 	if (DO_BIC(BIC_GFX_c0))
4305 		average.packages->pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
4306 	if (DO_BIC(BIC_CPUGFX))
4307 		average.packages->pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
4308 
4309 	average.packages->pc2 += p->pc2;
4310 	if (DO_BIC(BIC_Pkgpc3))
4311 		average.packages->pc3 += p->pc3;
4312 	if (DO_BIC(BIC_Pkgpc6))
4313 		average.packages->pc6 += p->pc6;
4314 	if (DO_BIC(BIC_Pkgpc7))
4315 		average.packages->pc7 += p->pc7;
4316 	average.packages->pc8 += p->pc8;
4317 	average.packages->pc9 += p->pc9;
4318 	average.packages->pc10 += p->pc10;
4319 	average.packages->die_c6 += p->die_c6;
4320 
4321 	average.packages->cpu_lpi = p->cpu_lpi;
4322 	average.packages->sys_lpi = p->sys_lpi;
4323 
4324 	rapl_counter_accumulate(&average.packages->energy_pkg, &p->energy_pkg);
4325 	rapl_counter_accumulate(&average.packages->energy_dram, &p->energy_dram);
4326 	rapl_counter_accumulate(&average.packages->energy_cores, &p->energy_cores);
4327 	rapl_counter_accumulate(&average.packages->energy_gfx, &p->energy_gfx);
4328 
4329 	average.packages->gfx_rc6_ms = p->gfx_rc6_ms;
4330 	average.packages->uncore_mhz = p->uncore_mhz;
4331 	average.packages->gfx_mhz = p->gfx_mhz;
4332 	average.packages->gfx_act_mhz = p->gfx_act_mhz;
4333 	average.packages->sam_mc6_ms = p->sam_mc6_ms;
4334 	average.packages->sam_mhz = p->sam_mhz;
4335 	average.packages->sam_act_mhz = p->sam_act_mhz;
4336 
4337 	average.packages->pkg_temp_c = MAX(average.packages->pkg_temp_c, p->pkg_temp_c);
4338 
4339 	rapl_counter_accumulate(&average.packages->rapl_pkg_perf_status, &p->rapl_pkg_perf_status);
4340 	rapl_counter_accumulate(&average.packages->rapl_dram_perf_status, &p->rapl_dram_perf_status);
4341 
4342 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
4343 		if ((mp->format == FORMAT_RAW) && (topo.num_packages == 0))
4344 			average.packages->counter[i] = p->counter[i];
4345 		else
4346 			average.packages->counter[i] += p->counter[i];
4347 	}
4348 
4349 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
4350 		if ((pp->format == FORMAT_RAW) && (topo.num_packages == 0))
4351 			average.packages->perf_counter[i] = p->perf_counter[i];
4352 		else
4353 			average.packages->perf_counter[i] += p->perf_counter[i];
4354 	}
4355 
4356 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
4357 		average.packages->pmt_counter[i] += p->pmt_counter[i];
4358 	}
4359 
4360 	return 0;
4361 }
4362 
4363 /*
4364  * sum the counters for all cpus in the system
4365  * compute the weighted average
4366  */
compute_average(PER_THREAD_PARAMS)4367 void compute_average(PER_THREAD_PARAMS)
4368 {
4369 	int i;
4370 	struct msr_counter *mp;
4371 	struct perf_counter_info *pp;
4372 	struct pmt_counter *ppmt;
4373 
4374 	clear_counters(average.threads, average.cores, average.packages);
4375 
4376 	for_all_cpus(sum_counters, t, c, p);
4377 
4378 	/* Use the global time delta for the average. */
4379 	average.threads->tv_delta = tv_delta;
4380 
4381 	average.threads->tsc /= topo.allowed_cpus;
4382 	average.threads->aperf /= topo.allowed_cpus;
4383 	average.threads->mperf /= topo.allowed_cpus;
4384 	average.threads->instr_count /= topo.allowed_cpus;
4385 	average.threads->c1 /= topo.allowed_cpus;
4386 
4387 	if (average.threads->irq_count > 9999999)
4388 		sums_need_wide_columns = 1;
4389 	if (average.threads->nmi_count > 9999999)
4390 		sums_need_wide_columns = 1;
4391 
4392 	average.cores->c3 /= topo.allowed_cores;
4393 	average.cores->c6 /= topo.allowed_cores;
4394 	average.cores->c7 /= topo.allowed_cores;
4395 	average.cores->mc6_us /= topo.allowed_cores;
4396 
4397 	if (DO_BIC(BIC_Totl_c0))
4398 		average.packages->pkg_wtd_core_c0 /= topo.allowed_packages;
4399 	if (DO_BIC(BIC_Any_c0))
4400 		average.packages->pkg_any_core_c0 /= topo.allowed_packages;
4401 	if (DO_BIC(BIC_GFX_c0))
4402 		average.packages->pkg_any_gfxe_c0 /= topo.allowed_packages;
4403 	if (DO_BIC(BIC_CPUGFX))
4404 		average.packages->pkg_both_core_gfxe_c0 /= topo.allowed_packages;
4405 
4406 	average.packages->pc2 /= topo.allowed_packages;
4407 	if (DO_BIC(BIC_Pkgpc3))
4408 		average.packages->pc3 /= topo.allowed_packages;
4409 	if (DO_BIC(BIC_Pkgpc6))
4410 		average.packages->pc6 /= topo.allowed_packages;
4411 	if (DO_BIC(BIC_Pkgpc7))
4412 		average.packages->pc7 /= topo.allowed_packages;
4413 
4414 	average.packages->pc8 /= topo.allowed_packages;
4415 	average.packages->pc9 /= topo.allowed_packages;
4416 	average.packages->pc10 /= topo.allowed_packages;
4417 	average.packages->die_c6 /= topo.allowed_packages;
4418 
4419 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
4420 		if (mp->format == FORMAT_RAW)
4421 			continue;
4422 		if (mp->type == COUNTER_ITEMS) {
4423 			if (average.threads->counter[i] > 9999999)
4424 				sums_need_wide_columns = 1;
4425 			continue;
4426 		}
4427 		average.threads->counter[i] /= topo.allowed_cpus;
4428 	}
4429 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
4430 		if (mp->format == FORMAT_RAW)
4431 			continue;
4432 		if (mp->type == COUNTER_ITEMS) {
4433 			if (average.cores->counter[i] > 9999999)
4434 				sums_need_wide_columns = 1;
4435 		}
4436 		average.cores->counter[i] /= topo.allowed_cores;
4437 	}
4438 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
4439 		if (mp->format == FORMAT_RAW)
4440 			continue;
4441 		if (mp->type == COUNTER_ITEMS) {
4442 			if (average.packages->counter[i] > 9999999)
4443 				sums_need_wide_columns = 1;
4444 		}
4445 		average.packages->counter[i] /= topo.allowed_packages;
4446 	}
4447 
4448 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
4449 		if (pp->format == FORMAT_RAW)
4450 			continue;
4451 		if (pp->type == COUNTER_ITEMS) {
4452 			if (average.threads->perf_counter[i] > 9999999)
4453 				sums_need_wide_columns = 1;
4454 			continue;
4455 		}
4456 		average.threads->perf_counter[i] /= topo.allowed_cpus;
4457 	}
4458 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
4459 		if (pp->format == FORMAT_RAW)
4460 			continue;
4461 		if (pp->type == COUNTER_ITEMS) {
4462 			if (average.cores->perf_counter[i] > 9999999)
4463 				sums_need_wide_columns = 1;
4464 		}
4465 		average.cores->perf_counter[i] /= topo.allowed_cores;
4466 	}
4467 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
4468 		if (pp->format == FORMAT_RAW)
4469 			continue;
4470 		if (pp->type == COUNTER_ITEMS) {
4471 			if (average.packages->perf_counter[i] > 9999999)
4472 				sums_need_wide_columns = 1;
4473 		}
4474 		average.packages->perf_counter[i] /= topo.allowed_packages;
4475 	}
4476 
4477 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
4478 		average.threads->pmt_counter[i] /= topo.allowed_cpus;
4479 	}
4480 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
4481 		average.cores->pmt_counter[i] /= topo.allowed_cores;
4482 	}
4483 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
4484 		average.packages->pmt_counter[i] /= topo.allowed_packages;
4485 	}
4486 }
4487 
rdtsc(void)4488 static unsigned long long rdtsc(void)
4489 {
4490 	unsigned int low, high;
4491 
4492 	asm volatile ("rdtsc":"=a" (low), "=d"(high));
4493 
4494 	return low | ((unsigned long long)high) << 32;
4495 }
4496 
4497 /*
4498  * Open a file, and exit on failure
4499  */
fopen_or_die(const char * path,const char * mode)4500 FILE *fopen_or_die(const char *path, const char *mode)
4501 {
4502 	FILE *filep = fopen(path, mode);
4503 
4504 	if (!filep)
4505 		err(1, "%s: open failed", path);
4506 	return filep;
4507 }
4508 
4509 /*
4510  * snapshot_sysfs_counter()
4511  *
4512  * return snapshot of given counter
4513  */
snapshot_sysfs_counter(char * path)4514 unsigned long long snapshot_sysfs_counter(char *path)
4515 {
4516 	FILE *fp;
4517 	int retval;
4518 	unsigned long long counter;
4519 
4520 	fp = fopen_or_die(path, "r");
4521 
4522 	retval = fscanf(fp, "%lld", &counter);
4523 	if (retval != 1)
4524 		err(1, "snapshot_sysfs_counter(%s)", path);
4525 
4526 	fclose(fp);
4527 
4528 	return counter;
4529 }
4530 
get_mp(int cpu,struct msr_counter * mp,unsigned long long * counterp,char * counter_path)4531 int get_mp(int cpu, struct msr_counter *mp, unsigned long long *counterp, char *counter_path)
4532 {
4533 	if (mp->msr_num != 0) {
4534 		assert(!no_msr);
4535 		if (get_msr(cpu, mp->msr_num, counterp))
4536 			return -1;
4537 	} else {
4538 		char path[128 + PATH_BYTES];
4539 
4540 		if (mp->flags & SYSFS_PERCPU) {
4541 			sprintf(path, "/sys/devices/system/cpu/cpu%d/%s", cpu, mp->sp->path);
4542 
4543 			*counterp = snapshot_sysfs_counter(path);
4544 		} else {
4545 			*counterp = snapshot_sysfs_counter(counter_path);
4546 		}
4547 	}
4548 
4549 	return 0;
4550 }
4551 
get_legacy_uncore_mhz(int package)4552 unsigned long long get_legacy_uncore_mhz(int package)
4553 {
4554 	char path[128];
4555 	int die;
4556 	static int warn_once;
4557 
4558 	/*
4559 	 * for this package, use the first die_id that exists
4560 	 */
4561 	for (die = 0; die <= topo.max_die_id; ++die) {
4562 
4563 		sprintf(path, "/sys/devices/system/cpu/intel_uncore_frequency/package_%02d_die_%02d/current_freq_khz", package, die);
4564 
4565 		if (access(path, R_OK) == 0)
4566 			return (snapshot_sysfs_counter(path) / 1000);
4567 	}
4568 	if (!warn_once) {
4569 		warnx("BUG: %s: No %s", __func__, path);
4570 		warn_once = 1;
4571 	}
4572 
4573 	return 0;
4574 }
4575 
get_epb(int cpu)4576 int get_epb(int cpu)
4577 {
4578 	char path[128 + PATH_BYTES];
4579 	unsigned long long msr;
4580 	int ret, epb = -1;
4581 	FILE *fp;
4582 
4583 	sprintf(path, "/sys/devices/system/cpu/cpu%d/power/energy_perf_bias", cpu);
4584 
4585 	fp = fopen(path, "r");
4586 	if (!fp)
4587 		goto msr_fallback;
4588 
4589 	ret = fscanf(fp, "%d", &epb);
4590 	if (ret != 1)
4591 		err(1, "%s(%s)", __func__, path);
4592 
4593 	fclose(fp);
4594 
4595 	return epb;
4596 
4597 msr_fallback:
4598 	if (no_msr)
4599 		return -1;
4600 
4601 	get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr);
4602 
4603 	return msr & 0xf;
4604 }
4605 
get_apic_id(struct thread_data * t)4606 void get_apic_id(struct thread_data *t)
4607 {
4608 	unsigned int eax, ebx, ecx, edx;
4609 
4610 	if (DO_BIC(BIC_APIC)) {
4611 		eax = ebx = ecx = edx = 0;
4612 		__cpuid(1, eax, ebx, ecx, edx);
4613 
4614 		t->apic_id = (ebx >> 24) & 0xff;
4615 	}
4616 
4617 	if (!DO_BIC(BIC_X2APIC))
4618 		return;
4619 
4620 	if (authentic_amd || hygon_genuine) {
4621 		unsigned int topology_extensions;
4622 
4623 		if (max_extended_level < 0x8000001e)
4624 			return;
4625 
4626 		eax = ebx = ecx = edx = 0;
4627 		__cpuid(0x80000001, eax, ebx, ecx, edx);
4628 		topology_extensions = ecx & (1 << 22);
4629 
4630 		if (topology_extensions == 0)
4631 			return;
4632 
4633 		eax = ebx = ecx = edx = 0;
4634 		__cpuid(0x8000001e, eax, ebx, ecx, edx);
4635 
4636 		t->x2apic_id = eax;
4637 		return;
4638 	}
4639 
4640 	if (!genuine_intel)
4641 		return;
4642 
4643 	if (max_level < 0xb)
4644 		return;
4645 
4646 	ecx = 0;
4647 	__cpuid(0xb, eax, ebx, ecx, edx);
4648 	t->x2apic_id = edx;
4649 
4650 	if (debug && (t->apic_id != (t->x2apic_id & 0xff)))
4651 		fprintf(outf, "cpu%d: BIOS BUG: apic 0x%x x2apic 0x%x\n", t->cpu_id, t->apic_id, t->x2apic_id);
4652 }
4653 
get_core_throt_cnt(int cpu,unsigned long long * cnt)4654 int get_core_throt_cnt(int cpu, unsigned long long *cnt)
4655 {
4656 	char path[128 + PATH_BYTES];
4657 	unsigned long long tmp;
4658 	FILE *fp;
4659 	int ret;
4660 
4661 	sprintf(path, "/sys/devices/system/cpu/cpu%d/thermal_throttle/core_throttle_count", cpu);
4662 	fp = fopen(path, "r");
4663 	if (!fp)
4664 		return -1;
4665 	ret = fscanf(fp, "%lld", &tmp);
4666 	fclose(fp);
4667 	if (ret != 1)
4668 		return -1;
4669 	*cnt = tmp;
4670 
4671 	return 0;
4672 }
4673 
read_perf_counter_info(const char * const path,const char * const parse_format,void * value_ptr)4674 static int read_perf_counter_info(const char *const path, const char *const parse_format, void *value_ptr)
4675 {
4676 	int fdmt;
4677 	int bytes_read;
4678 	char buf[64];
4679 	int ret = -1;
4680 
4681 	fdmt = open(path, O_RDONLY, 0);
4682 	if (fdmt == -1) {
4683 		if (debug)
4684 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4685 		ret = -1;
4686 		goto cleanup_and_exit;
4687 	}
4688 
4689 	bytes_read = read(fdmt, buf, sizeof(buf) - 1);
4690 	if (bytes_read <= 0 || bytes_read >= (int)sizeof(buf)) {
4691 		if (debug)
4692 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4693 		ret = -1;
4694 		goto cleanup_and_exit;
4695 	}
4696 
4697 	buf[bytes_read] = '\0';
4698 
4699 	if (sscanf(buf, parse_format, value_ptr) != 1) {
4700 		if (debug)
4701 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4702 		ret = -1;
4703 		goto cleanup_and_exit;
4704 	}
4705 
4706 	ret = 0;
4707 
4708 cleanup_and_exit:
4709 	close(fdmt);
4710 	return ret;
4711 }
4712 
read_perf_counter_info_n(const char * const path,const char * const parse_format)4713 static unsigned int read_perf_counter_info_n(const char *const path, const char *const parse_format)
4714 {
4715 	unsigned int v;
4716 	int status;
4717 
4718 	status = read_perf_counter_info(path, parse_format, &v);
4719 	if (status)
4720 		v = -1;
4721 
4722 	return v;
4723 }
4724 
read_perf_type(const char * subsys)4725 static unsigned int read_perf_type(const char *subsys)
4726 {
4727 	const char *const path_format = "/sys/bus/event_source/devices/%s/type";
4728 	const char *const format = "%u";
4729 	char path[128];
4730 
4731 	snprintf(path, sizeof(path), path_format, subsys);
4732 
4733 	return read_perf_counter_info_n(path, format);
4734 }
4735 
read_perf_config(const char * subsys,const char * event_name)4736 static unsigned int read_perf_config(const char *subsys, const char *event_name)
4737 {
4738 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s";
4739 	FILE *fconfig = NULL;
4740 	char path[128];
4741 	char config_str[64];
4742 	unsigned int config;
4743 	unsigned int umask;
4744 	bool has_config = false;
4745 	bool has_umask = false;
4746 	unsigned int ret = -1;
4747 
4748 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4749 
4750 	fconfig = fopen(path, "r");
4751 	if (!fconfig)
4752 		return -1;
4753 
4754 	if (fgets(config_str, ARRAY_SIZE(config_str), fconfig) != config_str)
4755 		goto cleanup_and_exit;
4756 
4757 	for (char *pconfig_str = &config_str[0]; pconfig_str;) {
4758 		if (sscanf(pconfig_str, "event=%x", &config) == 1) {
4759 			has_config = true;
4760 			goto next;
4761 		}
4762 
4763 		if (sscanf(pconfig_str, "umask=%x", &umask) == 1) {
4764 			has_umask = true;
4765 			goto next;
4766 		}
4767 
4768 next:
4769 		pconfig_str = strchr(pconfig_str, ',');
4770 		if (pconfig_str) {
4771 			*pconfig_str = '\0';
4772 			++pconfig_str;
4773 		}
4774 	}
4775 
4776 	if (!has_umask)
4777 		umask = 0;
4778 
4779 	if (has_config)
4780 		ret = (umask << 8) | config;
4781 
4782 cleanup_and_exit:
4783 	fclose(fconfig);
4784 	return ret;
4785 }
4786 
read_perf_rapl_unit(const char * subsys,const char * event_name)4787 static unsigned int read_perf_rapl_unit(const char *subsys, const char *event_name)
4788 {
4789 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.unit";
4790 	const char *const format = "%s";
4791 	char path[128];
4792 	char unit_buffer[16];
4793 
4794 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4795 
4796 	read_perf_counter_info(path, format, &unit_buffer);
4797 	if (strcmp("Joules", unit_buffer) == 0)
4798 		return RAPL_UNIT_JOULES;
4799 
4800 	return RAPL_UNIT_INVALID;
4801 }
4802 
read_perf_scale(const char * subsys,const char * event_name)4803 static double read_perf_scale(const char *subsys, const char *event_name)
4804 {
4805 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.scale";
4806 	const char *const format = "%lf";
4807 	char path[128];
4808 	double scale;
4809 
4810 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4811 
4812 	if (read_perf_counter_info(path, format, &scale))
4813 		return 0.0;
4814 
4815 	return scale;
4816 }
4817 
rapl_counter_info_count_perf(const struct rapl_counter_info_t * rci)4818 size_t rapl_counter_info_count_perf(const struct rapl_counter_info_t *rci)
4819 {
4820 	size_t ret = 0;
4821 
4822 	for (int i = 0; i < NUM_RAPL_COUNTERS; ++i)
4823 		if (rci->source[i] == COUNTER_SOURCE_PERF)
4824 			++ret;
4825 
4826 	return ret;
4827 }
4828 
cstate_counter_info_count_perf(const struct cstate_counter_info_t * cci)4829 static size_t cstate_counter_info_count_perf(const struct cstate_counter_info_t *cci)
4830 {
4831 	size_t ret = 0;
4832 
4833 	for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i)
4834 		if (cci->source[i] == COUNTER_SOURCE_PERF)
4835 			++ret;
4836 
4837 	return ret;
4838 }
4839 
write_rapl_counter(struct rapl_counter * rc,struct rapl_counter_info_t * rci,unsigned int idx)4840 void write_rapl_counter(struct rapl_counter *rc, struct rapl_counter_info_t *rci, unsigned int idx)
4841 {
4842 	if (rci->source[idx] == COUNTER_SOURCE_NONE)
4843 		return;
4844 
4845 	rc->raw_value = rci->data[idx];
4846 	rc->unit = rci->unit[idx];
4847 	rc->scale = rci->scale[idx];
4848 }
4849 
get_rapl_counters(int cpu,unsigned int domain,struct core_data * c,struct pkg_data * p)4850 int get_rapl_counters(int cpu, unsigned int domain, struct core_data *c, struct pkg_data *p)
4851 {
4852 	struct platform_counters *pplat_cnt = p == odd.packages ? &platform_counters_odd : &platform_counters_even;
4853 	unsigned long long perf_data[NUM_RAPL_COUNTERS + 1];
4854 	struct rapl_counter_info_t *rci;
4855 
4856 	if (debug >= 2)
4857 		fprintf(stderr, "%s: cpu%d domain%d\n", __func__, cpu, domain);
4858 
4859 	assert(rapl_counter_info_perdomain);
4860 	assert(domain < rapl_counter_info_perdomain_size);
4861 
4862 	rci = &rapl_counter_info_perdomain[domain];
4863 
4864 	/*
4865 	 * If we have any perf counters to read, read them all now, in bulk
4866 	 */
4867 	if (rci->fd_perf != -1) {
4868 		size_t num_perf_counters = rapl_counter_info_count_perf(rci);
4869 		const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
4870 		const ssize_t actual_read_size = read(rci->fd_perf, &perf_data[0], sizeof(perf_data));
4871 
4872 		if (actual_read_size != expected_read_size)
4873 			err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size, actual_read_size);
4874 	}
4875 
4876 	for (unsigned int i = 0, pi = 1; i < NUM_RAPL_COUNTERS; ++i) {
4877 		switch (rci->source[i]) {
4878 		case COUNTER_SOURCE_NONE:
4879 			rci->data[i] = 0;
4880 			break;
4881 
4882 		case COUNTER_SOURCE_PERF:
4883 			assert(pi < ARRAY_SIZE(perf_data));
4884 			assert(rci->fd_perf != -1);
4885 
4886 			if (debug >= 2)
4887 				fprintf(stderr, "Reading rapl counter via perf at %u (%llu %e %lf)\n",
4888 					i, perf_data[pi], rci->scale[i], perf_data[pi] * rci->scale[i]);
4889 
4890 			rci->data[i] = perf_data[pi];
4891 
4892 			++pi;
4893 			break;
4894 
4895 		case COUNTER_SOURCE_MSR:
4896 			if (debug >= 2)
4897 				fprintf(stderr, "Reading rapl counter via msr at %u\n", i);
4898 
4899 			assert(!no_msr);
4900 			if (rci->flags[i] & RAPL_COUNTER_FLAG_USE_MSR_SUM) {
4901 				if (get_msr_sum(cpu, rci->msr[i], &rci->data[i]))
4902 					return -13 - i;
4903 			} else {
4904 				if (get_msr(cpu, rci->msr[i], &rci->data[i]))
4905 					return -13 - i;
4906 			}
4907 
4908 			rci->data[i] &= rci->msr_mask[i];
4909 			if (rci->msr_shift[i] >= 0)
4910 				rci->data[i] >>= abs(rci->msr_shift[i]);
4911 			else
4912 				rci->data[i] <<= abs(rci->msr_shift[i]);
4913 
4914 			break;
4915 		}
4916 	}
4917 
4918 	BUILD_BUG_ON(NUM_RAPL_COUNTERS != 8);
4919 	write_rapl_counter(&p->energy_pkg, rci, RAPL_RCI_INDEX_ENERGY_PKG);
4920 	write_rapl_counter(&p->energy_cores, rci, RAPL_RCI_INDEX_ENERGY_CORES);
4921 	write_rapl_counter(&p->energy_dram, rci, RAPL_RCI_INDEX_DRAM);
4922 	write_rapl_counter(&p->energy_gfx, rci, RAPL_RCI_INDEX_GFX);
4923 	write_rapl_counter(&p->rapl_pkg_perf_status, rci, RAPL_RCI_INDEX_PKG_PERF_STATUS);
4924 	write_rapl_counter(&p->rapl_dram_perf_status, rci, RAPL_RCI_INDEX_DRAM_PERF_STATUS);
4925 	write_rapl_counter(&c->core_energy, rci, RAPL_RCI_INDEX_CORE_ENERGY);
4926 	write_rapl_counter(&pplat_cnt->energy_psys, rci, RAPL_RCI_INDEX_ENERGY_PLATFORM);
4927 
4928 	return 0;
4929 }
4930 
find_sysfs_path_by_id(struct sysfs_path * sp,int id)4931 char *find_sysfs_path_by_id(struct sysfs_path *sp, int id)
4932 {
4933 	while (sp) {
4934 		if (sp->id == id)
4935 			return (sp->path);
4936 		sp = sp->next;
4937 	}
4938 	if (debug)
4939 		warnx("%s: id%d not found", __func__, id);
4940 	return NULL;
4941 }
4942 
get_cstate_counters(unsigned int cpu,PER_THREAD_PARAMS)4943 int get_cstate_counters(unsigned int cpu, PER_THREAD_PARAMS)
4944 {
4945 	/*
4946 	 * Overcommit memory a little bit here,
4947 	 * but skip calculating exact sizes for the buffers.
4948 	 */
4949 	unsigned long long perf_data[NUM_CSTATE_COUNTERS];
4950 	unsigned long long perf_data_core[NUM_CSTATE_COUNTERS + 1];
4951 	unsigned long long perf_data_pkg[NUM_CSTATE_COUNTERS + 1];
4952 
4953 	struct cstate_counter_info_t *cci;
4954 
4955 	if (debug >= 2)
4956 		fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
4957 
4958 	assert(ccstate_counter_info);
4959 	assert(cpu <= ccstate_counter_info_size);
4960 
4961 	ZERO_ARRAY(perf_data);
4962 	ZERO_ARRAY(perf_data_core);
4963 	ZERO_ARRAY(perf_data_pkg);
4964 
4965 	cci = &ccstate_counter_info[cpu];
4966 
4967 	/*
4968 	 * If we have any perf counters to read, read them all now, in bulk
4969 	 */
4970 	const size_t num_perf_counters = cstate_counter_info_count_perf(cci);
4971 	ssize_t expected_read_size = num_perf_counters * sizeof(unsigned long long);
4972 	ssize_t actual_read_size_core = 0, actual_read_size_pkg = 0;
4973 
4974 	if (cci->fd_perf_core != -1) {
4975 		/* Each descriptor read begins with number of counters read. */
4976 		expected_read_size += sizeof(unsigned long long);
4977 
4978 		actual_read_size_core = read(cci->fd_perf_core, &perf_data_core[0], sizeof(perf_data_core));
4979 
4980 		if (actual_read_size_core <= 0)
4981 			err(-1, "%s: read perf %s: %ld", __func__, "core", actual_read_size_core);
4982 	}
4983 
4984 	if (cci->fd_perf_pkg != -1) {
4985 		/* Each descriptor read begins with number of counters read. */
4986 		expected_read_size += sizeof(unsigned long long);
4987 
4988 		actual_read_size_pkg = read(cci->fd_perf_pkg, &perf_data_pkg[0], sizeof(perf_data_pkg));
4989 
4990 		if (actual_read_size_pkg <= 0)
4991 			err(-1, "%s: read perf %s: %ld", __func__, "pkg", actual_read_size_pkg);
4992 	}
4993 
4994 	const ssize_t actual_read_size_total = actual_read_size_core + actual_read_size_pkg;
4995 
4996 	if (actual_read_size_total != expected_read_size)
4997 		err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size, actual_read_size_total);
4998 
4999 	/*
5000 	 * Copy ccstate and pcstate data into unified buffer.
5001 	 *
5002 	 * Skip first element from core and pkg buffers.
5003 	 * Kernel puts there how many counters were read.
5004 	 */
5005 	const size_t num_core_counters = perf_data_core[0];
5006 	const size_t num_pkg_counters = perf_data_pkg[0];
5007 
5008 	assert(num_perf_counters == num_core_counters + num_pkg_counters);
5009 
5010 	/* Copy ccstate perf data */
5011 	memcpy(&perf_data[0], &perf_data_core[1], num_core_counters * sizeof(unsigned long long));
5012 
5013 	/* Copy pcstate perf data */
5014 	memcpy(&perf_data[num_core_counters], &perf_data_pkg[1], num_pkg_counters * sizeof(unsigned long long));
5015 
5016 	for (unsigned int i = 0, pi = 0; i < NUM_CSTATE_COUNTERS; ++i) {
5017 		switch (cci->source[i]) {
5018 		case COUNTER_SOURCE_NONE:
5019 			break;
5020 
5021 		case COUNTER_SOURCE_PERF:
5022 			assert(pi < ARRAY_SIZE(perf_data));
5023 			assert(cci->fd_perf_core != -1 || cci->fd_perf_pkg != -1);
5024 
5025 			if (debug >= 2)
5026 				fprintf(stderr, "cstate via %s %u: %llu\n", "perf", i, perf_data[pi]);
5027 
5028 			cci->data[i] = perf_data[pi];
5029 
5030 			++pi;
5031 			break;
5032 
5033 		case COUNTER_SOURCE_MSR:
5034 			assert(!no_msr);
5035 			if (get_msr(cpu, cci->msr[i], &cci->data[i]))
5036 				return -13 - i;
5037 
5038 			if (debug >= 2)
5039 				fprintf(stderr, "cstate via %s0x%llx %u: %llu\n", "msr", cci->msr[i], i, cci->data[i]);
5040 
5041 			break;
5042 		}
5043 	}
5044 
5045 	/*
5046 	 * Helper to write the data only if the source of
5047 	 * the counter for the current cpu is not none.
5048 	 *
5049 	 * Otherwise we would overwrite core data with 0 (default value),
5050 	 * when invoked for the thread sibling.
5051 	 */
5052 #define PERF_COUNTER_WRITE_DATA(out_counter, index) do {	\
5053 	if (cci->source[index] != COUNTER_SOURCE_NONE)		\
5054 		out_counter = cci->data[index];			\
5055 } while (0)
5056 
5057 	BUILD_BUG_ON(NUM_CSTATE_COUNTERS != 11);
5058 
5059 	PERF_COUNTER_WRITE_DATA(t->c1, CCSTATE_RCI_INDEX_C1_RESIDENCY);
5060 	PERF_COUNTER_WRITE_DATA(c->c3, CCSTATE_RCI_INDEX_C3_RESIDENCY);
5061 	PERF_COUNTER_WRITE_DATA(c->c6, CCSTATE_RCI_INDEX_C6_RESIDENCY);
5062 	PERF_COUNTER_WRITE_DATA(c->c7, CCSTATE_RCI_INDEX_C7_RESIDENCY);
5063 
5064 	PERF_COUNTER_WRITE_DATA(p->pc2, PCSTATE_RCI_INDEX_C2_RESIDENCY);
5065 	PERF_COUNTER_WRITE_DATA(p->pc3, PCSTATE_RCI_INDEX_C3_RESIDENCY);
5066 	PERF_COUNTER_WRITE_DATA(p->pc6, PCSTATE_RCI_INDEX_C6_RESIDENCY);
5067 	PERF_COUNTER_WRITE_DATA(p->pc7, PCSTATE_RCI_INDEX_C7_RESIDENCY);
5068 	PERF_COUNTER_WRITE_DATA(p->pc8, PCSTATE_RCI_INDEX_C8_RESIDENCY);
5069 	PERF_COUNTER_WRITE_DATA(p->pc9, PCSTATE_RCI_INDEX_C9_RESIDENCY);
5070 	PERF_COUNTER_WRITE_DATA(p->pc10, PCSTATE_RCI_INDEX_C10_RESIDENCY);
5071 
5072 #undef PERF_COUNTER_WRITE_DATA
5073 
5074 	return 0;
5075 }
5076 
msr_counter_info_count_perf(const struct msr_counter_info_t * mci)5077 size_t msr_counter_info_count_perf(const struct msr_counter_info_t *mci)
5078 {
5079 	size_t ret = 0;
5080 
5081 	for (int i = 0; i < NUM_MSR_COUNTERS; ++i)
5082 		if (mci->source[i] == COUNTER_SOURCE_PERF)
5083 			++ret;
5084 
5085 	return ret;
5086 }
5087 
get_smi_aperf_mperf(unsigned int cpu,struct thread_data * t)5088 int get_smi_aperf_mperf(unsigned int cpu, struct thread_data *t)
5089 {
5090 	unsigned long long perf_data[NUM_MSR_COUNTERS + 1];
5091 
5092 	struct msr_counter_info_t *mci;
5093 
5094 	if (debug >= 2)
5095 		fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
5096 
5097 	assert(msr_counter_info);
5098 	assert(cpu <= msr_counter_info_size);
5099 
5100 	mci = &msr_counter_info[cpu];
5101 
5102 	ZERO_ARRAY(perf_data);
5103 	ZERO_ARRAY(mci->data);
5104 
5105 	if (mci->fd_perf != -1) {
5106 		const size_t num_perf_counters = msr_counter_info_count_perf(mci);
5107 		const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
5108 		const ssize_t actual_read_size = read(mci->fd_perf, &perf_data[0], sizeof(perf_data));
5109 
5110 		if (actual_read_size != expected_read_size)
5111 			err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size, actual_read_size);
5112 	}
5113 
5114 	for (unsigned int i = 0, pi = 1; i < NUM_MSR_COUNTERS; ++i) {
5115 		switch (mci->source[i]) {
5116 		case COUNTER_SOURCE_NONE:
5117 			break;
5118 
5119 		case COUNTER_SOURCE_PERF:
5120 			assert(pi < ARRAY_SIZE(perf_data));
5121 			assert(mci->fd_perf != -1);
5122 
5123 			if (debug >= 2)
5124 				fprintf(stderr, "Reading msr counter via perf at %u: %llu\n", i, perf_data[pi]);
5125 
5126 			mci->data[i] = perf_data[pi];
5127 
5128 			++pi;
5129 			break;
5130 
5131 		case COUNTER_SOURCE_MSR:
5132 			assert(!no_msr);
5133 
5134 			if (get_msr(cpu, mci->msr[i], &mci->data[i]))
5135 				return -2 - i;
5136 
5137 			mci->data[i] &= mci->msr_mask[i];
5138 
5139 			if (debug >= 2)
5140 				fprintf(stderr, "Reading msr counter via msr at %u: %llu\n", i, mci->data[i]);
5141 
5142 			break;
5143 		}
5144 	}
5145 
5146 	BUILD_BUG_ON(NUM_MSR_COUNTERS != 3);
5147 	t->aperf = mci->data[MSR_RCI_INDEX_APERF];
5148 	t->mperf = mci->data[MSR_RCI_INDEX_MPERF];
5149 	t->smi_count = mci->data[MSR_RCI_INDEX_SMI];
5150 
5151 	return 0;
5152 }
5153 
perf_counter_info_read_values(struct perf_counter_info * pp,int cpu,unsigned long long * out,size_t out_size)5154 int perf_counter_info_read_values(struct perf_counter_info *pp, int cpu, unsigned long long *out, size_t out_size)
5155 {
5156 	unsigned int domain;
5157 	unsigned long long value;
5158 	int fd_counter;
5159 
5160 	for (size_t i = 0; pp; ++i, pp = pp->next) {
5161 		domain = cpu_to_domain(pp, cpu);
5162 		assert(domain < pp->num_domains);
5163 
5164 		fd_counter = pp->fd_perf_per_domain[domain];
5165 
5166 		if (fd_counter == -1)
5167 			continue;
5168 
5169 		if (read(fd_counter, &value, sizeof(value)) != sizeof(value))
5170 			return 1;
5171 
5172 		assert(i < out_size);
5173 		out[i] = value * pp->scale;
5174 	}
5175 
5176 	return 0;
5177 }
5178 
pmt_gen_value_mask(unsigned int lsb,unsigned int msb)5179 unsigned long pmt_gen_value_mask(unsigned int lsb, unsigned int msb)
5180 {
5181 	unsigned long mask;
5182 
5183 	if (msb == 63)
5184 		mask = 0xffffffffffffffff;
5185 	else
5186 		mask = ((1 << (msb + 1)) - 1);
5187 
5188 	mask -= (1 << lsb) - 1;
5189 
5190 	return mask;
5191 }
5192 
pmt_read_counter(struct pmt_counter * ppmt,unsigned int domain_id)5193 unsigned long pmt_read_counter(struct pmt_counter *ppmt, unsigned int domain_id)
5194 {
5195 	if (domain_id >= ppmt->num_domains)
5196 		return 0;
5197 
5198 	const unsigned long *pmmio = ppmt->domains[domain_id].pcounter;
5199 	const unsigned long value = pmmio ? *pmmio : 0;
5200 	const unsigned long value_mask = pmt_gen_value_mask(ppmt->lsb, ppmt->msb);
5201 	const unsigned long value_shift = ppmt->lsb;
5202 
5203 	return (value & value_mask) >> value_shift;
5204 }
5205 
5206 /* Rapl domain enumeration helpers */
get_rapl_num_domains(void)5207 static inline int get_rapl_num_domains(void)
5208 {
5209 	if (!platform->has_per_core_rapl)
5210 		return topo.num_packages;
5211 
5212 	return GLOBAL_CORE_ID(topo.max_core_id, topo.num_packages) + 1;
5213 }
5214 
get_rapl_domain_id(int cpu)5215 static inline int get_rapl_domain_id(int cpu)
5216 {
5217 	if (!platform->has_per_core_rapl)
5218 		return cpus[cpu].package_id;
5219 
5220 	return GLOBAL_CORE_ID(cpus[cpu].core_id, cpus[cpu].package_id);
5221 }
5222 
5223 /*
5224  * get_counters(...)
5225  * migrate to cpu
5226  * acquire and record local counters for that cpu
5227  */
get_counters(PER_THREAD_PARAMS)5228 int get_counters(PER_THREAD_PARAMS)
5229 {
5230 	int cpu = t->cpu_id;
5231 	unsigned long long msr;
5232 	struct msr_counter *mp;
5233 	struct pmt_counter *pp;
5234 	int i;
5235 	int status;
5236 
5237 	if (cpu_migrate(cpu)) {
5238 		fprintf(outf, "%s: Could not migrate to CPU %d\n", __func__, cpu);
5239 		return -1;
5240 	}
5241 
5242 	gettimeofday(&t->tv_begin, (struct timezone *)NULL);
5243 
5244 	if (first_counter_read)
5245 		get_apic_id(t);
5246 
5247 	t->tsc = rdtsc();	/* we are running on local CPU of interest */
5248 
5249 	get_smi_aperf_mperf(cpu, t);
5250 
5251 	if (DO_BIC(BIC_LLC_MRPS) || DO_BIC(BIC_LLC_HIT))
5252 		get_perf_llc_stats(cpu, &t->llc);
5253 
5254 	if (DO_BIC(BIC_L2_MRPS) || DO_BIC(BIC_L2_HIT))
5255 		get_perf_l2_stats(cpu, &t->l2);
5256 
5257 	if (DO_BIC(BIC_IPC))
5258 		if (read(get_instr_count_fd(cpu), &t->instr_count, sizeof(long long)) != sizeof(long long))
5259 			return -4;
5260 
5261 	if (DO_BIC(BIC_IRQ))
5262 		t->irq_count = irqs_per_cpu[cpu];
5263 	if (DO_BIC(BIC_NMI))
5264 		t->nmi_count = nmi_per_cpu[cpu];
5265 
5266 	get_cstate_counters(cpu, t, c, p);
5267 
5268 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
5269 		if (get_mp(cpu, mp, &t->counter[i], mp->sp->path))
5270 			return -10;
5271 	}
5272 
5273 	if (perf_counter_info_read_values(sys.perf_tp, cpu, t->perf_counter, MAX_ADDED_THREAD_COUNTERS))
5274 		return -10;
5275 
5276 	for (i = 0, pp = sys.pmt_tp; pp; i++, pp = pp->next)
5277 		t->pmt_counter[i] = pmt_read_counter(pp, t->cpu_id);
5278 
5279 	/* collect core counters only for 1st thread in core */
5280 	if (!is_cpu_first_thread_in_core(t, c))
5281 		goto done;
5282 
5283 	if (platform->has_per_core_rapl) {
5284 		status = get_rapl_counters(cpu, get_rapl_domain_id(cpu), c, p);
5285 		if (status != 0)
5286 			return status;
5287 	}
5288 
5289 	if (DO_BIC(BIC_CPU_c7) && t->is_atom) {
5290 		/*
5291 		 * For Atom CPUs that has core cstate deeper than c6,
5292 		 * MSR_CORE_C6_RESIDENCY returns residency of cc6 and deeper.
5293 		 * Minus CC7 (and deeper cstates) residency to get
5294 		 * accturate cc6 residency.
5295 		 */
5296 		c->c6 -= c->c7;
5297 	}
5298 
5299 	if (DO_BIC(BIC_Mod_c6))
5300 		if (get_msr(cpu, MSR_MODULE_C6_RES_MS, &c->mc6_us))
5301 			return -8;
5302 
5303 	if (DO_BIC(BIC_CoreTmp)) {
5304 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
5305 			return -9;
5306 		c->core_temp_c = tj_max - ((msr >> 16) & 0x7F);
5307 	}
5308 
5309 	if (DO_BIC(BIC_CORE_THROT_CNT))
5310 		get_core_throt_cnt(cpu, &c->core_throt_cnt);
5311 
5312 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
5313 		if (get_mp(cpu, mp, &c->counter[i], mp->sp->path))
5314 			return -10;
5315 	}
5316 
5317 	if (perf_counter_info_read_values(sys.perf_cp, cpu, c->perf_counter, MAX_ADDED_CORE_COUNTERS))
5318 		return -10;
5319 
5320 	for (i = 0, pp = sys.pmt_cp; pp; i++, pp = pp->next)
5321 		c->pmt_counter[i] = pmt_read_counter(pp, cpus[t->cpu_id].core_id);
5322 
5323 	/* collect package counters only for 1st core in package */
5324 	if (!is_cpu_first_core_in_package(t, p))
5325 		goto done;
5326 
5327 	if (DO_BIC(BIC_Totl_c0)) {
5328 		if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
5329 			return -10;
5330 	}
5331 	if (DO_BIC(BIC_Any_c0)) {
5332 		if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
5333 			return -11;
5334 	}
5335 	if (DO_BIC(BIC_GFX_c0)) {
5336 		if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
5337 			return -12;
5338 	}
5339 	if (DO_BIC(BIC_CPUGFX)) {
5340 		if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
5341 			return -13;
5342 	}
5343 
5344 	if (DO_BIC(BIC_CPU_LPI))
5345 		p->cpu_lpi = cpuidle_cur_cpu_lpi_us;
5346 	if (DO_BIC(BIC_SYS_LPI))
5347 		p->sys_lpi = cpuidle_cur_sys_lpi_us;
5348 
5349 	if (!platform->has_per_core_rapl) {
5350 		status = get_rapl_counters(cpu, get_rapl_domain_id(cpu), c, p);
5351 		if (status != 0)
5352 			return status;
5353 	}
5354 
5355 	if (DO_BIC(BIC_PkgTmp)) {
5356 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
5357 			return -17;
5358 		p->pkg_temp_c = tj_max - ((msr >> 16) & 0x7F);
5359 	}
5360 
5361 	if (DO_BIC(BIC_UNCORE_MHZ))
5362 		p->uncore_mhz = get_legacy_uncore_mhz(cpus[t->cpu_id].package_id);
5363 
5364 	if (DO_BIC(BIC_GFX_rc6))
5365 		p->gfx_rc6_ms = gfx_info[GFX_rc6].val_ull;
5366 
5367 	if (DO_BIC(BIC_GFXMHz))
5368 		p->gfx_mhz = gfx_info[GFX_MHz].val;
5369 
5370 	if (DO_BIC(BIC_GFXACTMHz))
5371 		p->gfx_act_mhz = gfx_info[GFX_ACTMHz].val;
5372 
5373 	if (DO_BIC(BIC_SAM_mc6))
5374 		p->sam_mc6_ms = gfx_info[SAM_mc6].val_ull;
5375 
5376 	if (DO_BIC(BIC_SAMMHz))
5377 		p->sam_mhz = gfx_info[SAM_MHz].val;
5378 
5379 	if (DO_BIC(BIC_SAMACTMHz))
5380 		p->sam_act_mhz = gfx_info[SAM_ACTMHz].val;
5381 
5382 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
5383 		char *path = NULL;
5384 
5385 		if (mp->msr_num == 0) {
5386 			path = find_sysfs_path_by_id(mp->sp, cpus[t->cpu_id].package_id);
5387 			if (path == NULL) {
5388 				warnx("%s: package_id %d not found", __func__, cpus[t->cpu_id].package_id);
5389 				return -10;
5390 			}
5391 		}
5392 		if (get_mp(cpu, mp, &p->counter[i], path))
5393 			return -10;
5394 	}
5395 
5396 	if (perf_counter_info_read_values(sys.perf_pp, cpu, p->perf_counter, MAX_ADDED_PACKAGE_COUNTERS))
5397 		return -10;
5398 
5399 	for (i = 0, pp = sys.pmt_pp; pp; i++, pp = pp->next)
5400 		p->pmt_counter[i] = pmt_read_counter(pp, cpus[t->cpu_id].package_id);
5401 
5402 done:
5403 	gettimeofday(&t->tv_end, (struct timezone *)NULL);
5404 
5405 	return 0;
5406 }
5407 
5408 int pkg_cstate_limit = PCLUKN;
5409 char *pkg_cstate_limit_strings[] = { "unknown", "reserved", "pc0", "pc1", "pc2",
5410 	"pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "pc10", "unlimited"
5411 };
5412 
5413 int nhm_pkg_cstate_limits[16] = { PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5414 	PCLRSV, PCLRSV
5415 };
5416 
5417 int snb_pkg_cstate_limits[16] = { PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5418 	PCLRSV, PCLRSV
5419 };
5420 
5421 int hsw_pkg_cstate_limits[16] = { PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5422 	PCLRSV, PCLRSV
5423 };
5424 
5425 int slv_pkg_cstate_limits[16] = { PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5426 	PCL__6, PCL__7
5427 };
5428 
5429 int amt_pkg_cstate_limits[16] = { PCLUNL, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5430 	PCLRSV, PCLRSV
5431 };
5432 
5433 int phi_pkg_cstate_limits[16] = { PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5434 	PCLRSV, PCLRSV
5435 };
5436 
5437 int glm_pkg_cstate_limits[16] = { PCLUNL, PCL__1, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCL_10, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5438 	PCLRSV, PCLRSV
5439 };
5440 
5441 int skx_pkg_cstate_limits[16] = { PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5442 	PCLRSV, PCLRSV
5443 };
5444 
5445 int icx_pkg_cstate_limits[16] = { PCL__0, PCL__2, PCL__6, PCL__6, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5446 	PCLRSV, PCLRSV
5447 };
5448 
probe_cst_limit(void)5449 void probe_cst_limit(void)
5450 {
5451 	unsigned long long msr;
5452 	int *pkg_cstate_limits;
5453 
5454 	if (!platform->has_nhm_msrs || no_msr)
5455 		return;
5456 
5457 	switch (platform->cst_limit) {
5458 	case CST_LIMIT_NHM:
5459 		pkg_cstate_limits = nhm_pkg_cstate_limits;
5460 		break;
5461 	case CST_LIMIT_SNB:
5462 		pkg_cstate_limits = snb_pkg_cstate_limits;
5463 		break;
5464 	case CST_LIMIT_HSW:
5465 		pkg_cstate_limits = hsw_pkg_cstate_limits;
5466 		break;
5467 	case CST_LIMIT_SKX:
5468 		pkg_cstate_limits = skx_pkg_cstate_limits;
5469 		break;
5470 	case CST_LIMIT_ICX:
5471 		pkg_cstate_limits = icx_pkg_cstate_limits;
5472 		break;
5473 	case CST_LIMIT_SLV:
5474 		pkg_cstate_limits = slv_pkg_cstate_limits;
5475 		break;
5476 	case CST_LIMIT_AMT:
5477 		pkg_cstate_limits = amt_pkg_cstate_limits;
5478 		break;
5479 	case CST_LIMIT_KNL:
5480 		pkg_cstate_limits = phi_pkg_cstate_limits;
5481 		break;
5482 	case CST_LIMIT_GMT:
5483 		pkg_cstate_limits = glm_pkg_cstate_limits;
5484 		break;
5485 	default:
5486 		return;
5487 	}
5488 
5489 	get_msr(master_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
5490 	pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
5491 }
5492 
dump_platform_info(void)5493 static void dump_platform_info(void)
5494 {
5495 	unsigned long long msr;
5496 	unsigned int ratio;
5497 
5498 	if (!platform->has_nhm_msrs || no_msr)
5499 		return;
5500 
5501 	get_msr(master_cpu, MSR_PLATFORM_INFO, &msr);
5502 
5503 	fprintf(outf, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", master_cpu, msr);
5504 
5505 	ratio = (msr >> 40) & 0xFF;
5506 	fprintf(outf, "%d * %.1f = %.1f MHz max efficiency frequency\n", ratio, bclk, ratio * bclk);
5507 
5508 	ratio = (msr >> 8) & 0xFF;
5509 	fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n", ratio, bclk, ratio * bclk);
5510 }
5511 
dump_power_ctl(void)5512 static void dump_power_ctl(void)
5513 {
5514 	unsigned long long msr;
5515 
5516 	if (!platform->has_nhm_msrs || no_msr)
5517 		return;
5518 
5519 	get_msr(master_cpu, MSR_IA32_POWER_CTL, &msr);
5520 	fprintf(outf, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n", master_cpu, msr, msr & 0x2 ? "EN" : "DIS");
5521 
5522 	/* C-state Pre-wake Disable (CSTATE_PREWAKE_DISABLE) */
5523 	if (platform->has_cst_prewake_bit)
5524 		fprintf(outf, "C-state Pre-wake: %sabled\n", msr & 0x40000000 ? "DIS" : "EN");
5525 
5526 	return;
5527 }
5528 
dump_turbo_ratio_limit2(void)5529 static void dump_turbo_ratio_limit2(void)
5530 {
5531 	unsigned long long msr;
5532 	unsigned int ratio;
5533 
5534 	get_msr(master_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
5535 
5536 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", master_cpu, msr);
5537 
5538 	ratio = (msr >> 8) & 0xFF;
5539 	if (ratio)
5540 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 18 active cores\n", ratio, bclk, ratio * bclk);
5541 
5542 	ratio = (msr >> 0) & 0xFF;
5543 	if (ratio)
5544 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 17 active cores\n", ratio, bclk, ratio * bclk);
5545 	return;
5546 }
5547 
dump_turbo_ratio_limit1(void)5548 static void dump_turbo_ratio_limit1(void)
5549 {
5550 	unsigned long long msr;
5551 	unsigned int ratio;
5552 
5553 	get_msr(master_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
5554 
5555 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", master_cpu, msr);
5556 
5557 	ratio = (msr >> 56) & 0xFF;
5558 	if (ratio)
5559 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 16 active cores\n", ratio, bclk, ratio * bclk);
5560 
5561 	ratio = (msr >> 48) & 0xFF;
5562 	if (ratio)
5563 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 15 active cores\n", ratio, bclk, ratio * bclk);
5564 
5565 	ratio = (msr >> 40) & 0xFF;
5566 	if (ratio)
5567 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 14 active cores\n", ratio, bclk, ratio * bclk);
5568 
5569 	ratio = (msr >> 32) & 0xFF;
5570 	if (ratio)
5571 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 13 active cores\n", ratio, bclk, ratio * bclk);
5572 
5573 	ratio = (msr >> 24) & 0xFF;
5574 	if (ratio)
5575 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 12 active cores\n", ratio, bclk, ratio * bclk);
5576 
5577 	ratio = (msr >> 16) & 0xFF;
5578 	if (ratio)
5579 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 11 active cores\n", ratio, bclk, ratio * bclk);
5580 
5581 	ratio = (msr >> 8) & 0xFF;
5582 	if (ratio)
5583 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 10 active cores\n", ratio, bclk, ratio * bclk);
5584 
5585 	ratio = (msr >> 0) & 0xFF;
5586 	if (ratio)
5587 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 9 active cores\n", ratio, bclk, ratio * bclk);
5588 	return;
5589 }
5590 
dump_turbo_ratio_limits(int trl_msr_offset)5591 static void dump_turbo_ratio_limits(int trl_msr_offset)
5592 {
5593 	unsigned long long msr, core_counts;
5594 	int shift;
5595 
5596 	get_msr(master_cpu, trl_msr_offset, &msr);
5597 	fprintf(outf, "cpu%d: MSR_%sTURBO_RATIO_LIMIT: 0x%08llx\n", master_cpu, trl_msr_offset == MSR_SECONDARY_TURBO_RATIO_LIMIT ? "SECONDARY_" : "", msr);
5598 
5599 	if (platform->trl_msrs & TRL_CORECOUNT) {
5600 		get_msr(master_cpu, MSR_TURBO_RATIO_LIMIT1, &core_counts);
5601 		fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", master_cpu, core_counts);
5602 	} else {
5603 		core_counts = 0x0807060504030201;
5604 	}
5605 
5606 	for (shift = 56; shift >= 0; shift -= 8) {
5607 		unsigned int ratio, group_size;
5608 
5609 		ratio = (msr >> shift) & 0xFF;
5610 		group_size = (core_counts >> shift) & 0xFF;
5611 		if (ratio)
5612 			fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n", ratio, bclk, ratio * bclk, group_size);
5613 	}
5614 
5615 	return;
5616 }
5617 
dump_atom_turbo_ratio_limits(void)5618 static void dump_atom_turbo_ratio_limits(void)
5619 {
5620 	unsigned long long msr;
5621 	unsigned int ratio;
5622 
5623 	get_msr(master_cpu, MSR_ATOM_CORE_RATIOS, &msr);
5624 	fprintf(outf, "cpu%d: MSR_ATOM_CORE_RATIOS: 0x%08llx\n", master_cpu, msr & 0xFFFFFFFF);
5625 
5626 	ratio = (msr >> 0) & 0x3F;
5627 	if (ratio)
5628 		fprintf(outf, "%d * %.1f = %.1f MHz minimum operating frequency\n", ratio, bclk, ratio * bclk);
5629 
5630 	ratio = (msr >> 8) & 0x3F;
5631 	if (ratio)
5632 		fprintf(outf, "%d * %.1f = %.1f MHz low frequency mode (LFM)\n", ratio, bclk, ratio * bclk);
5633 
5634 	ratio = (msr >> 16) & 0x3F;
5635 	if (ratio)
5636 		fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n", ratio, bclk, ratio * bclk);
5637 
5638 	get_msr(master_cpu, MSR_ATOM_CORE_TURBO_RATIOS, &msr);
5639 	fprintf(outf, "cpu%d: MSR_ATOM_CORE_TURBO_RATIOS: 0x%08llx\n", master_cpu, msr & 0xFFFFFFFF);
5640 
5641 	ratio = (msr >> 24) & 0x3F;
5642 	if (ratio)
5643 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 4 active cores\n", ratio, bclk, ratio * bclk);
5644 
5645 	ratio = (msr >> 16) & 0x3F;
5646 	if (ratio)
5647 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 3 active cores\n", ratio, bclk, ratio * bclk);
5648 
5649 	ratio = (msr >> 8) & 0x3F;
5650 	if (ratio)
5651 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 2 active cores\n", ratio, bclk, ratio * bclk);
5652 
5653 	ratio = (msr >> 0) & 0x3F;
5654 	if (ratio)
5655 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 1 active core\n", ratio, bclk, ratio * bclk);
5656 }
5657 
dump_knl_turbo_ratio_limits(void)5658 static void dump_knl_turbo_ratio_limits(void)
5659 {
5660 	const unsigned int buckets_no = 7;
5661 
5662 	unsigned long long msr;
5663 	int delta_cores, delta_ratio;
5664 	int i, b_nr;
5665 	unsigned int cores[buckets_no];
5666 	unsigned int ratio[buckets_no];
5667 
5668 	get_msr(master_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
5669 
5670 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", master_cpu, msr);
5671 
5672 	/*
5673 	 * Turbo encoding in KNL is as follows:
5674 	 * [0] -- Reserved
5675 	 * [7:1] -- Base value of number of active cores of bucket 1.
5676 	 * [15:8] -- Base value of freq ratio of bucket 1.
5677 	 * [20:16] -- +ve delta of number of active cores of bucket 2.
5678 	 * i.e. active cores of bucket 2 =
5679 	 * active cores of bucket 1 + delta
5680 	 * [23:21] -- Negative delta of freq ratio of bucket 2.
5681 	 * i.e. freq ratio of bucket 2 =
5682 	 * freq ratio of bucket 1 - delta
5683 	 * [28:24]-- +ve delta of number of active cores of bucket 3.
5684 	 * [31:29]-- -ve delta of freq ratio of bucket 3.
5685 	 * [36:32]-- +ve delta of number of active cores of bucket 4.
5686 	 * [39:37]-- -ve delta of freq ratio of bucket 4.
5687 	 * [44:40]-- +ve delta of number of active cores of bucket 5.
5688 	 * [47:45]-- -ve delta of freq ratio of bucket 5.
5689 	 * [52:48]-- +ve delta of number of active cores of bucket 6.
5690 	 * [55:53]-- -ve delta of freq ratio of bucket 6.
5691 	 * [60:56]-- +ve delta of number of active cores of bucket 7.
5692 	 * [63:61]-- -ve delta of freq ratio of bucket 7.
5693 	 */
5694 
5695 	b_nr = 0;
5696 	cores[b_nr] = (msr & 0xFF) >> 1;
5697 	ratio[b_nr] = (msr >> 8) & 0xFF;
5698 
5699 	for (i = 16; i < 64; i += 8) {
5700 		delta_cores = (msr >> i) & 0x1F;
5701 		delta_ratio = (msr >> (i + 5)) & 0x7;
5702 
5703 		cores[b_nr + 1] = cores[b_nr] + delta_cores;
5704 		ratio[b_nr + 1] = ratio[b_nr] - delta_ratio;
5705 		b_nr++;
5706 	}
5707 
5708 	for (i = buckets_no - 1; i >= 0; i--)
5709 		if (i > 0 ? ratio[i] != ratio[i - 1] : 1)
5710 			fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n", ratio[i], bclk, ratio[i] * bclk, cores[i]);
5711 }
5712 
dump_cst_cfg(void)5713 static void dump_cst_cfg(void)
5714 {
5715 	unsigned long long msr;
5716 
5717 	if (!platform->has_nhm_msrs || no_msr)
5718 		return;
5719 
5720 	get_msr(master_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
5721 
5722 	fprintf(outf, "cpu%d: MSR_PKG_CST_CONFIG_CONTROL: 0x%08llx", master_cpu, msr);
5723 
5724 	fprintf(outf, " (%s%s%s%s%slocked, pkg-cstate-limit=%d (%s)",
5725 		(msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
5726 		(msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
5727 		(msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
5728 		(msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
5729 		(msr & (1 << 15)) ? "" : "UN", (unsigned int)msr & 0xF, pkg_cstate_limit_strings[pkg_cstate_limit]);
5730 
5731 #define AUTOMATIC_CSTATE_CONVERSION		(1UL << 16)
5732 	if (platform->has_cst_auto_convension) {
5733 		fprintf(outf, ", automatic c-state conversion=%s", (msr & AUTOMATIC_CSTATE_CONVERSION) ? "on" : "off");
5734 	}
5735 
5736 	fprintf(outf, ")\n");
5737 
5738 	return;
5739 }
5740 
dump_config_tdp(void)5741 static void dump_config_tdp(void)
5742 {
5743 	unsigned long long msr;
5744 
5745 	get_msr(master_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
5746 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", master_cpu, msr);
5747 	fprintf(outf, " (base_ratio=%d)\n", (unsigned int)msr & 0xFF);
5748 
5749 	get_msr(master_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
5750 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", master_cpu, msr);
5751 	if (msr) {
5752 		fprintf(outf, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
5753 		fprintf(outf, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
5754 		fprintf(outf, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
5755 		fprintf(outf, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0x7FFF);
5756 	}
5757 	fprintf(outf, ")\n");
5758 
5759 	get_msr(master_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
5760 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", master_cpu, msr);
5761 	if (msr) {
5762 		fprintf(outf, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
5763 		fprintf(outf, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
5764 		fprintf(outf, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
5765 		fprintf(outf, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0x7FFF);
5766 	}
5767 	fprintf(outf, ")\n");
5768 
5769 	get_msr(master_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
5770 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", master_cpu, msr);
5771 	if ((msr) & 0x3)
5772 		fprintf(outf, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
5773 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
5774 	fprintf(outf, ")\n");
5775 
5776 	get_msr(master_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
5777 	fprintf(outf, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", master_cpu, msr);
5778 	fprintf(outf, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xFF);
5779 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
5780 	fprintf(outf, ")\n");
5781 }
5782 
5783 unsigned int irtl_time_units[] = { 1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
5784 
print_irtl(void)5785 void print_irtl(void)
5786 {
5787 	unsigned long long msr;
5788 
5789 	if (!platform->has_irtl_msrs || no_msr)
5790 		return;
5791 
5792 	if (platform->supported_cstates & PC3) {
5793 		get_msr(master_cpu, MSR_PKGC3_IRTL, &msr);
5794 		fprintf(outf, "cpu%d: MSR_PKGC3_IRTL: 0x%08llx (", master_cpu, msr);
5795 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5796 	}
5797 
5798 	if (platform->supported_cstates & PC6) {
5799 		get_msr(master_cpu, MSR_PKGC6_IRTL, &msr);
5800 		fprintf(outf, "cpu%d: MSR_PKGC6_IRTL: 0x%08llx (", master_cpu, msr);
5801 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5802 	}
5803 
5804 	if (platform->supported_cstates & PC7) {
5805 		get_msr(master_cpu, MSR_PKGC7_IRTL, &msr);
5806 		fprintf(outf, "cpu%d: MSR_PKGC7_IRTL: 0x%08llx (", master_cpu, msr);
5807 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5808 	}
5809 
5810 	if (platform->supported_cstates & PC8) {
5811 		get_msr(master_cpu, MSR_PKGC8_IRTL, &msr);
5812 		fprintf(outf, "cpu%d: MSR_PKGC8_IRTL: 0x%08llx (", master_cpu, msr);
5813 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5814 	}
5815 
5816 	if (platform->supported_cstates & PC9) {
5817 		get_msr(master_cpu, MSR_PKGC9_IRTL, &msr);
5818 		fprintf(outf, "cpu%d: MSR_PKGC9_IRTL: 0x%08llx (", master_cpu, msr);
5819 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5820 	}
5821 
5822 	if (platform->supported_cstates & PC10) {
5823 		get_msr(master_cpu, MSR_PKGC10_IRTL, &msr);
5824 		fprintf(outf, "cpu%d: MSR_PKGC10_IRTL: 0x%08llx (", master_cpu, msr);
5825 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5826 	}
5827 }
5828 
free_fd_percpu(void)5829 void free_fd_percpu(void)
5830 {
5831 	int i;
5832 
5833 	if (!fd_percpu)
5834 		return;
5835 
5836 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
5837 		if (fd_percpu[i] != 0)
5838 			close(fd_percpu[i]);
5839 	}
5840 
5841 	free(fd_percpu);
5842 	fd_percpu = NULL;
5843 }
5844 
free_fd_instr_count_percpu(void)5845 void free_fd_instr_count_percpu(void)
5846 {
5847 	if (!fd_instr_count_percpu)
5848 		return;
5849 
5850 	for (int i = 0; i < topo.max_cpu_num + 1; ++i) {
5851 		if (fd_instr_count_percpu[i] != 0)
5852 			close(fd_instr_count_percpu[i]);
5853 	}
5854 
5855 	free(fd_instr_count_percpu);
5856 	fd_instr_count_percpu = NULL;
5857 }
5858 
free_fd_llc_percpu(void)5859 void free_fd_llc_percpu(void)
5860 {
5861 	if (!fd_llc_percpu)
5862 		return;
5863 
5864 	for (int i = 0; i < topo.max_cpu_num + 1; ++i) {
5865 		if (fd_llc_percpu[i] != 0)
5866 			close(fd_llc_percpu[i]);
5867 	}
5868 
5869 	free(fd_llc_percpu);
5870 	fd_llc_percpu = NULL;
5871 
5872 	BIC_NOT_PRESENT(BIC_LLC_MRPS);
5873 	BIC_NOT_PRESENT(BIC_LLC_HIT);
5874 }
5875 
free_fd_l2_percpu(void)5876 void free_fd_l2_percpu(void)
5877 {
5878 	if (!fd_l2_percpu)
5879 		return;
5880 
5881 	for (int i = 0; i < topo.max_cpu_num + 1; ++i) {
5882 		if (fd_l2_percpu[i] != 0)
5883 			close(fd_l2_percpu[i]);
5884 	}
5885 
5886 	free(fd_l2_percpu);
5887 	fd_l2_percpu = NULL;
5888 
5889 	BIC_NOT_PRESENT(BIC_L2_MRPS);
5890 	BIC_NOT_PRESENT(BIC_L2_HIT);
5891 }
5892 
free_fd_cstate(void)5893 void free_fd_cstate(void)
5894 {
5895 	if (!ccstate_counter_info)
5896 		return;
5897 
5898 	const int counter_info_num = ccstate_counter_info_size;
5899 
5900 	for (int counter_id = 0; counter_id < counter_info_num; ++counter_id) {
5901 		if (ccstate_counter_info[counter_id].fd_perf_core != -1)
5902 			close(ccstate_counter_info[counter_id].fd_perf_core);
5903 
5904 		if (ccstate_counter_info[counter_id].fd_perf_pkg != -1)
5905 			close(ccstate_counter_info[counter_id].fd_perf_pkg);
5906 	}
5907 
5908 	free(ccstate_counter_info);
5909 	ccstate_counter_info = NULL;
5910 	ccstate_counter_info_size = 0;
5911 }
5912 
free_fd_msr(void)5913 void free_fd_msr(void)
5914 {
5915 	if (!msr_counter_info)
5916 		return;
5917 
5918 	for (int cpu = 0; cpu < topo.max_cpu_num; ++cpu) {
5919 		if (msr_counter_info[cpu].fd_perf != -1)
5920 			close(msr_counter_info[cpu].fd_perf);
5921 	}
5922 
5923 	free(msr_counter_info);
5924 	msr_counter_info = NULL;
5925 	msr_counter_info_size = 0;
5926 }
5927 
free_fd_rapl_percpu(void)5928 void free_fd_rapl_percpu(void)
5929 {
5930 	if (!rapl_counter_info_perdomain)
5931 		return;
5932 
5933 	const int num_domains = rapl_counter_info_perdomain_size;
5934 
5935 	for (int domain_id = 0; domain_id < num_domains; ++domain_id) {
5936 		if (rapl_counter_info_perdomain[domain_id].fd_perf != -1)
5937 			close(rapl_counter_info_perdomain[domain_id].fd_perf);
5938 	}
5939 
5940 	free(rapl_counter_info_perdomain);
5941 	rapl_counter_info_perdomain = NULL;
5942 	rapl_counter_info_perdomain_size = 0;
5943 }
5944 
free_fd_added_perf_counters_(struct perf_counter_info * pp)5945 void free_fd_added_perf_counters_(struct perf_counter_info *pp)
5946 {
5947 	if (!pp)
5948 		return;
5949 
5950 	if (!pp->fd_perf_per_domain)
5951 		return;
5952 
5953 	while (pp) {
5954 		for (size_t domain = 0; domain < pp->num_domains; ++domain) {
5955 			if (pp->fd_perf_per_domain[domain] != -1) {
5956 				close(pp->fd_perf_per_domain[domain]);
5957 				pp->fd_perf_per_domain[domain] = -1;
5958 			}
5959 		}
5960 
5961 		free(pp->fd_perf_per_domain);
5962 		pp->fd_perf_per_domain = NULL;
5963 
5964 		pp = pp->next;
5965 	}
5966 }
5967 
free_fd_added_perf_counters(void)5968 void free_fd_added_perf_counters(void)
5969 {
5970 	free_fd_added_perf_counters_(sys.perf_tp);
5971 	free_fd_added_perf_counters_(sys.perf_cp);
5972 	free_fd_added_perf_counters_(sys.perf_pp);
5973 }
5974 
free_all_buffers(void)5975 void free_all_buffers(void)
5976 {
5977 	int i;
5978 
5979 	CPU_FREE(cpu_present_set);
5980 	cpu_present_set = NULL;
5981 	cpu_present_setsize = 0;
5982 
5983 	CPU_FREE(cpu_effective_set);
5984 	cpu_effective_set = NULL;
5985 	cpu_effective_setsize = 0;
5986 
5987 	CPU_FREE(cpu_allowed_set);
5988 	cpu_allowed_set = NULL;
5989 	cpu_allowed_setsize = 0;
5990 
5991 	CPU_FREE(cpu_affinity_set);
5992 	cpu_affinity_set = NULL;
5993 	cpu_affinity_setsize = 0;
5994 
5995 	if (perf_pcore_set) {
5996 		CPU_FREE(perf_pcore_set);
5997 		perf_pcore_set = NULL;
5998 	}
5999 
6000 	if (perf_ecore_set) {
6001 		CPU_FREE(perf_ecore_set);
6002 		perf_ecore_set = NULL;
6003 	}
6004 
6005 	if (perf_lcore_set) {
6006 		CPU_FREE(perf_lcore_set);
6007 		perf_lcore_set = NULL;
6008 	}
6009 
6010 	free(even.threads);
6011 	free(even.cores);
6012 	free(even.packages);
6013 
6014 	even.threads = NULL;
6015 	even.cores = NULL;
6016 	even.packages = NULL;
6017 
6018 	free(odd.threads);
6019 	free(odd.cores);
6020 	free(odd.packages);
6021 
6022 	odd.threads = NULL;
6023 	odd.cores = NULL;
6024 	odd.packages = NULL;
6025 
6026 	free(output_buffer);
6027 	output_buffer = NULL;
6028 	outp = NULL;
6029 
6030 	free_fd_percpu();
6031 	free_fd_instr_count_percpu();
6032 	free_fd_llc_percpu();
6033 	free_fd_l2_percpu();
6034 	free_fd_msr();
6035 	free_fd_rapl_percpu();
6036 	free_fd_cstate();
6037 	free_fd_added_perf_counters();
6038 
6039 	free(irq_column_2_cpu);
6040 	free(irqs_per_cpu);
6041 	free(nmi_per_cpu);
6042 
6043 	for (i = 0; i <= topo.max_cpu_num; ++i) {
6044 		if (cpus[i].put_ids)
6045 			CPU_FREE(cpus[i].put_ids);
6046 	}
6047 	free(cpus);
6048 }
6049 
6050 /*
6051  * Parse a file containing a single int.
6052  * Return 0 if file can not be opened
6053  * Exit if file can be opened, but can not be parsed
6054  */
parse_int_file(const char * fmt,...)6055 int parse_int_file(const char *fmt, ...)
6056 {
6057 	va_list args;
6058 	char path[PATH_MAX];
6059 	FILE *filep;
6060 	int value;
6061 
6062 	va_start(args, fmt);
6063 	vsnprintf(path, sizeof(path), fmt, args);
6064 	va_end(args);
6065 	filep = fopen(path, "r");
6066 	if (!filep)
6067 		return 0;
6068 	if (fscanf(filep, "%d", &value) != 1)
6069 		err(1, "%s: failed to parse number from file", path);
6070 	fclose(filep);
6071 	return value;
6072 }
6073 
6074 /*
6075  * cpu_is_first_core_in_package(cpu)
6076  * return 1 if given CPU is 1st core in package
6077  */
cpu_is_first_core_in_package(int cpu)6078 int cpu_is_first_core_in_package(int cpu)
6079 {
6080 	return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
6081 }
6082 
get_package_id(int cpu)6083 int get_package_id(int cpu)
6084 {
6085 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
6086 }
6087 
get_die_id(int cpu)6088 int get_die_id(int cpu)
6089 {
6090 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/die_id", cpu);
6091 }
6092 
get_l3_id(int cpu)6093 int get_l3_id(int cpu)
6094 {
6095 	return parse_int_file("/sys/devices/system/cpu/cpu%d/cache/index3/id", cpu);
6096 }
6097 
get_module_id(int cpu)6098 int get_module_id(int cpu)
6099 {
6100 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/cluster_id", cpu);
6101 }
6102 
get_core_id(int cpu)6103 int get_core_id(int cpu)
6104 {
6105 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
6106 }
6107 
set_node_data(void)6108 void set_node_data(void)
6109 {
6110 	int pkg, node, lnode, cpu, cpux;
6111 	int cpu_count;
6112 
6113 	/* initialize logical_node_id */
6114 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu)
6115 		cpus[cpu].logical_node_id = -1;
6116 
6117 	cpu_count = 0;
6118 	for (pkg = 0; pkg < topo.num_packages; pkg++) {
6119 		lnode = 0;
6120 		for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
6121 			if (cpus[cpu].package_id != pkg)
6122 				continue;
6123 			/* find a cpu with an unset logical_node_id */
6124 			if (cpus[cpu].logical_node_id != -1)
6125 				continue;
6126 			cpus[cpu].logical_node_id = lnode;
6127 			node = cpus[cpu].physical_node_id;
6128 			cpu_count++;
6129 			/*
6130 			 * find all matching cpus on this pkg and set
6131 			 * the logical_node_id
6132 			 */
6133 			for (cpux = cpu; cpux <= topo.max_cpu_num; cpux++) {
6134 				if ((cpus[cpux].package_id == pkg) && (cpus[cpux].physical_node_id == node)) {
6135 					cpus[cpux].logical_node_id = lnode;
6136 					cpu_count++;
6137 				}
6138 			}
6139 			lnode++;
6140 			if (lnode > topo.nodes_per_pkg)
6141 				topo.nodes_per_pkg = lnode;
6142 		}
6143 		if (cpu_count >= topo.max_cpu_num)
6144 			break;
6145 	}
6146 }
6147 
get_physical_node_id(struct cpu_topology * thiscpu)6148 int get_physical_node_id(struct cpu_topology *thiscpu)
6149 {
6150 	char path[80];
6151 	FILE *filep;
6152 	int i;
6153 	int cpu = thiscpu->cpu_id;
6154 
6155 	for (i = 0; i <= topo.max_cpu_num; i++) {
6156 		sprintf(path, "/sys/devices/system/cpu/cpu%d/node%i/cpulist", cpu, i);
6157 		filep = fopen(path, "r");
6158 		if (!filep)
6159 			continue;
6160 		fclose(filep);
6161 		return i;
6162 	}
6163 	return -1;
6164 }
6165 
parse_cpu_str(char * cpu_str,cpu_set_t * cpu_set,int cpu_set_size)6166 static int parse_cpu_str(char *cpu_str, cpu_set_t *cpu_set, int cpu_set_size)
6167 {
6168 	unsigned int start, end;
6169 	char *next = cpu_str;
6170 
6171 	while (next && *next) {
6172 
6173 		if (*next == '-')	/* no negative cpu numbers */
6174 			return 1;
6175 
6176 		if (*next == '\0' || *next == '\n')
6177 			break;
6178 
6179 		start = strtoul(next, &next, 10);
6180 
6181 		if (start >= CPU_SUBSET_MAXCPUS)
6182 			return 1;
6183 		CPU_SET_S(start, cpu_set_size, cpu_set);
6184 
6185 		if (*next == '\0' || *next == '\n')
6186 			break;
6187 
6188 		if (*next == ',') {
6189 			next += 1;
6190 			continue;
6191 		}
6192 
6193 		if (*next == '-') {
6194 			next += 1;	/* start range */
6195 		} else if (*next == '.') {
6196 			next += 1;
6197 			if (*next == '.')
6198 				next += 1;	/* start range */
6199 			else
6200 				return 1;
6201 		}
6202 
6203 		end = strtoul(next, &next, 10);
6204 		if (end <= start)
6205 			return 1;
6206 
6207 		while (++start <= end) {
6208 			if (start >= CPU_SUBSET_MAXCPUS)
6209 				return 1;
6210 			CPU_SET_S(start, cpu_set_size, cpu_set);
6211 		}
6212 
6213 		if (*next == ',')
6214 			next += 1;
6215 		else if (*next != '\0' && *next != '\n')
6216 			return 1;
6217 	}
6218 
6219 	return 0;
6220 }
6221 
6222 /*
6223  * run func(thread, core, package) in topology order
6224  * skip non-present cpus
6225  */
6226 
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)6227 int for_all_cpus_2(int (func) (struct thread_data *, struct core_data *,
6228 			       struct pkg_data *, struct thread_data *, struct core_data *,
6229 			       struct pkg_data *), struct thread_data *thread_base,
6230 		   struct core_data *core_base, struct pkg_data *pkg_base,
6231 		   struct thread_data *thread_base2, struct core_data *core_base2, struct pkg_data *pkg_base2)
6232 {
6233 	int cpu, retval;
6234 
6235 	retval = 0;
6236 
6237 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
6238 		struct thread_data *t, *t2;
6239 		struct core_data *c, *c2;
6240 		struct pkg_data *p, *p2;
6241 
6242 		if (cpu_is_not_allowed(cpu))
6243 			continue;
6244 
6245 		if (has_allowed_lower_ht_sibling(cpu))	/* skip HT sibling */
6246 			continue;
6247 
6248 		t = &thread_base[cpu];
6249 		t2 = &thread_base2[cpu];
6250 		c = &core_base[GLOBAL_CORE_ID(cpus[cpu].core_id, cpus[cpu].package_id)];
6251 		c2 = &core_base2[GLOBAL_CORE_ID(cpus[cpu].core_id, cpus[cpu].package_id)];
6252 		p = &pkg_base[cpus[cpu].package_id];
6253 		p2 = &pkg_base2[cpus[cpu].package_id];
6254 
6255 		retval |= func(t, c, p, t2, c2, p2);
6256 
6257 		/* Handle HT sibling now */
6258 		int i;
6259 
6260 		for (i = 0; i <= MAX_HT_ID; ++i) {
6261 			int sibling_cpu_id = cpus[cpu].ht_sibling_cpu_id[i];
6262 
6263 			if (sibling_cpu_id < 0)
6264 				break;
6265 
6266 			if (sibling_cpu_id == cpu)
6267 				continue;
6268 
6269 			if (cpu_is_not_allowed(sibling_cpu_id))
6270 				continue;
6271 
6272 			t = &thread_base[sibling_cpu_id];
6273 			t2 = &thread_base2[sibling_cpu_id];
6274 
6275 			retval |= func(t, c, p, t2, c2, p2);
6276 		}
6277 	}
6278 	return retval;
6279 }
6280 
6281 /*
6282  * run func(cpu) on every cpu in /proc/stat
6283  * return max_cpu number
6284  */
for_all_proc_cpus(int (func)(int))6285 int for_all_proc_cpus(int (func) (int))
6286 {
6287 	FILE *fp;
6288 	int cpu_num;
6289 	int retval;
6290 
6291 	fp = fopen_or_die(proc_stat, "r");
6292 
6293 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
6294 	if (retval != 0)
6295 		err(1, "%s: failed to parse format", proc_stat);
6296 
6297 	while (1) {
6298 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
6299 		if (retval != 1)
6300 			break;
6301 
6302 		retval = func(cpu_num);
6303 		if (retval) {
6304 			fclose(fp);
6305 			return (retval);
6306 		}
6307 	}
6308 	fclose(fp);
6309 	return 0;
6310 }
6311 
6312 #define PATH_EFFECTIVE_CPUS	"/sys/fs/cgroup/cpuset.cpus.effective"
6313 
6314 static char cpu_effective_str[1024];
6315 
update_effective_str(bool startup)6316 static int update_effective_str(bool startup)
6317 {
6318 	FILE *fp;
6319 	char *pos;
6320 	char buf[1024];
6321 	int ret;
6322 
6323 	if (cpu_effective_str[0] == '\0' && !startup)
6324 		return 0;
6325 
6326 	fp = fopen(PATH_EFFECTIVE_CPUS, "r");
6327 	if (!fp)
6328 		return 0;
6329 
6330 	pos = fgets(buf, 1024, fp);
6331 	if (!pos)
6332 		err(1, "%s: file read failed", PATH_EFFECTIVE_CPUS);
6333 
6334 	fclose(fp);
6335 
6336 	ret = strncmp(cpu_effective_str, buf, 1024);
6337 	if (!ret)
6338 		return 0;
6339 
6340 	strncpy(cpu_effective_str, buf, 1024);
6341 	return 1;
6342 }
6343 
update_effective_set(bool startup)6344 static void update_effective_set(bool startup)
6345 {
6346 	update_effective_str(startup);
6347 
6348 	if (parse_cpu_str(cpu_effective_str, cpu_effective_set, cpu_effective_setsize))
6349 		err(1, "%s: cpu str malformat %s", PATH_EFFECTIVE_CPUS, cpu_effective_str);
6350 }
6351 
6352 void linux_perf_init(void);
6353 void msr_perf_init(void);
6354 void rapl_perf_init(void);
6355 void cstate_perf_init(void);
6356 void perf_llc_init(void);
6357 void perf_l2_init(void);
6358 void added_perf_counters_init(void);
6359 void pmt_init(void);
6360 
re_initialize(void)6361 void re_initialize(void)
6362 {
6363 	free_all_buffers();
6364 	setup_all_buffers(false);
6365 	linux_perf_init();
6366 	msr_perf_init();
6367 	rapl_perf_init();
6368 	cstate_perf_init();
6369 	perf_llc_init();
6370 	perf_l2_init();
6371 	added_perf_counters_init();
6372 	pmt_init();
6373 	fprintf(outf, "turbostat: re-initialized with num_cpus %d, allowed_cpus %d\n", topo.num_cpus, topo.allowed_cpus);
6374 }
6375 
set_max_cpu_num(void)6376 void set_max_cpu_num(void)
6377 {
6378 	FILE *filep;
6379 	int current_cpu;
6380 	unsigned long dummy;
6381 	char pathname[64];
6382 
6383 	current_cpu = sched_getcpu();
6384 	if (current_cpu < 0)
6385 		err(1, "cannot find calling cpu ID");
6386 	sprintf(pathname, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", current_cpu);
6387 
6388 	filep = fopen_or_die(pathname, "r");
6389 	topo.max_cpu_num = 0;
6390 	while (fscanf(filep, "%lx,", &dummy) == 1)
6391 		topo.max_cpu_num += BITMASK_SIZE;
6392 	fclose(filep);
6393 	topo.max_cpu_num--;	/* 0 based */
6394 }
6395 
6396 /*
6397  * count_cpus()
6398  * remember the last one seen, it will be the max
6399  */
count_cpus(int cpu)6400 int count_cpus(int cpu)
6401 {
6402 	UNUSED(cpu);
6403 
6404 	topo.num_cpus++;
6405 	return 0;
6406 }
6407 
mark_cpu_present(int cpu)6408 int mark_cpu_present(int cpu)
6409 {
6410 	CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
6411 	return 0;
6412 }
6413 
clear_ht_id(int cpu)6414 int clear_ht_id(int cpu)
6415 {
6416 	int i;
6417 
6418 	cpus[cpu].ht_id = -1;
6419 	for (i = 0; i <= MAX_HT_ID; ++i)
6420 		cpus[cpu].ht_sibling_cpu_id[i] = -1;
6421 	return 0;
6422 }
6423 
set_my_cpu_type(void)6424 int set_my_cpu_type(void)
6425 {
6426 	unsigned int eax, ebx, ecx, edx;
6427 	unsigned int max_level;
6428 
6429 	__cpuid(0, max_level, ebx, ecx, edx);
6430 
6431 	if (max_level < CPUID_LEAF_MODEL_ID)
6432 		return 0;
6433 
6434 	__cpuid(CPUID_LEAF_MODEL_ID, eax, ebx, ecx, edx);
6435 
6436 	return (eax >> CPUID_LEAF_MODEL_ID_CORE_TYPE_SHIFT);
6437 }
6438 
set_cpu_hybrid_type(int cpu)6439 int set_cpu_hybrid_type(int cpu)
6440 {
6441 	if (cpu_migrate(cpu))
6442 		return -1;
6443 
6444 	int type = set_my_cpu_type();
6445 
6446 	cpus[cpu].type = type;
6447 	return 0;
6448 }
6449 
6450 /*
6451  * snapshot_proc_interrupts()
6452  *
6453  * read and record summary of /proc/interrupts
6454  *
6455  * return 1 if config change requires a restart, else return 0
6456  */
snapshot_proc_interrupts(void)6457 int snapshot_proc_interrupts(void)
6458 {
6459 	static FILE *fp;
6460 	int column, retval;
6461 
6462 	if (fp == NULL)
6463 		fp = fopen_or_die("/proc/interrupts", "r");
6464 	else
6465 		rewind(fp);
6466 
6467 	/* read 1st line of /proc/interrupts to get cpu* name for each column */
6468 	for (column = 0; column < topo.num_cpus; ++column) {
6469 		int cpu_number;
6470 
6471 		retval = fscanf(fp, " CPU%d", &cpu_number);
6472 		if (retval != 1)
6473 			break;
6474 
6475 		if (cpu_number > topo.max_cpu_num) {
6476 			warn("/proc/interrupts: cpu%d: > %d", cpu_number, topo.max_cpu_num);
6477 			return 1;
6478 		}
6479 
6480 		irq_column_2_cpu[column] = cpu_number;
6481 		irqs_per_cpu[cpu_number] = 0;
6482 		nmi_per_cpu[cpu_number] = 0;
6483 	}
6484 
6485 	/* read /proc/interrupt count lines and sum up irqs per cpu */
6486 	while (1) {
6487 		int column;
6488 		char buf[64];
6489 		int this_row_is_nmi = 0;
6490 
6491 		retval = fscanf(fp, " %s:", buf);	/* irq# "N:" */
6492 		if (retval != 1)
6493 			break;
6494 
6495 		if (strncmp(buf, "NMI", strlen("NMI")) == 0)
6496 			this_row_is_nmi = 1;
6497 
6498 		/* read the count per cpu */
6499 		for (column = 0; column < topo.num_cpus; ++column) {
6500 
6501 			int cpu_number, irq_count;
6502 
6503 			retval = fscanf(fp, " %d", &irq_count);
6504 
6505 			if (retval != 1)
6506 				break;
6507 
6508 			cpu_number = irq_column_2_cpu[column];
6509 			irqs_per_cpu[cpu_number] += irq_count;
6510 			if (this_row_is_nmi)
6511 				nmi_per_cpu[cpu_number] += irq_count;
6512 		}
6513 		while (getc(fp) != '\n') ;	/* flush interrupt description */
6514 
6515 	}
6516 	return 0;
6517 }
6518 
6519 /*
6520  * snapshot_graphics()
6521  *
6522  * record snapshot of specified graphics sysfs knob
6523  *
6524  * return 1 if config change requires a restart, else return 0
6525  */
snapshot_graphics(int idx)6526 int snapshot_graphics(int idx)
6527 {
6528 	int retval;
6529 
6530 	rewind(gfx_info[idx].fp);
6531 	fflush(gfx_info[idx].fp);
6532 
6533 	switch (idx) {
6534 	case GFX_rc6:
6535 	case SAM_mc6:
6536 		retval = fscanf(gfx_info[idx].fp, "%lld", &gfx_info[idx].val_ull);
6537 		if (retval != 1)
6538 			err(1, "rc6");
6539 		return 0;
6540 	case GFX_MHz:
6541 	case GFX_ACTMHz:
6542 	case SAM_MHz:
6543 	case SAM_ACTMHz:
6544 		retval = fscanf(gfx_info[idx].fp, "%d", &gfx_info[idx].val);
6545 		if (retval != 1)
6546 			err(1, "MHz");
6547 		return 0;
6548 	default:
6549 		return -EINVAL;
6550 	}
6551 }
6552 
6553 /*
6554  * snapshot_cpu_lpi()
6555  *
6556  * record snapshot of
6557  * /sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us
6558  */
snapshot_cpu_lpi_us(void)6559 int snapshot_cpu_lpi_us(void)
6560 {
6561 	FILE *fp;
6562 	int retval;
6563 
6564 	fp = fopen_or_die("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", "r");
6565 
6566 	retval = fscanf(fp, "%lld", &cpuidle_cur_cpu_lpi_us);
6567 	if (retval != 1) {
6568 		fprintf(stderr, "Disabling Low Power Idle CPU output\n");
6569 		BIC_NOT_PRESENT(BIC_CPU_LPI);
6570 		fclose(fp);
6571 		return -1;
6572 	}
6573 
6574 	fclose(fp);
6575 
6576 	return 0;
6577 }
6578 
6579 /*
6580  * snapshot_sys_lpi()
6581  *
6582  * record snapshot of sys_lpi_file
6583  */
snapshot_sys_lpi_us(void)6584 int snapshot_sys_lpi_us(void)
6585 {
6586 	FILE *fp;
6587 	int retval;
6588 
6589 	fp = fopen_or_die(sys_lpi_file, "r");
6590 
6591 	retval = fscanf(fp, "%lld", &cpuidle_cur_sys_lpi_us);
6592 	if (retval != 1) {
6593 		fprintf(stderr, "Disabling Low Power Idle System output\n");
6594 		BIC_NOT_PRESENT(BIC_SYS_LPI);
6595 		fclose(fp);
6596 		return -1;
6597 	}
6598 	fclose(fp);
6599 
6600 	return 0;
6601 }
6602 
6603 /*
6604  * snapshot /proc and /sys files
6605  *
6606  * return 1 if configuration restart needed, else return 0
6607  */
snapshot_proc_sysfs_files(void)6608 int snapshot_proc_sysfs_files(void)
6609 {
6610 	gettimeofday(&procsysfs_tv_begin, (struct timezone *)NULL);
6611 
6612 	if (DO_BIC(BIC_IRQ) || DO_BIC(BIC_NMI))
6613 		if (snapshot_proc_interrupts())
6614 			return 1;
6615 
6616 	if (DO_BIC(BIC_GFX_rc6))
6617 		snapshot_graphics(GFX_rc6);
6618 
6619 	if (DO_BIC(BIC_GFXMHz))
6620 		snapshot_graphics(GFX_MHz);
6621 
6622 	if (DO_BIC(BIC_GFXACTMHz))
6623 		snapshot_graphics(GFX_ACTMHz);
6624 
6625 	if (DO_BIC(BIC_SAM_mc6))
6626 		snapshot_graphics(SAM_mc6);
6627 
6628 	if (DO_BIC(BIC_SAMMHz))
6629 		snapshot_graphics(SAM_MHz);
6630 
6631 	if (DO_BIC(BIC_SAMACTMHz))
6632 		snapshot_graphics(SAM_ACTMHz);
6633 
6634 	if (DO_BIC(BIC_CPU_LPI))
6635 		snapshot_cpu_lpi_us();
6636 
6637 	if (DO_BIC(BIC_SYS_LPI))
6638 		snapshot_sys_lpi_us();
6639 
6640 	return 0;
6641 }
6642 
6643 int exit_requested;
6644 
signal_handler(int signal)6645 static void signal_handler(int signal)
6646 {
6647 	switch (signal) {
6648 	case SIGINT:
6649 		exit_requested = 1;
6650 		if (debug)
6651 			fprintf(stderr, " SIGINT\n");
6652 		break;
6653 	case SIGUSR1:
6654 		if (debug > 1)
6655 			fprintf(stderr, "SIGUSR1\n");
6656 		break;
6657 	}
6658 }
6659 
setup_signal_handler(void)6660 void setup_signal_handler(void)
6661 {
6662 	struct sigaction sa;
6663 
6664 	memset(&sa, 0, sizeof(sa));
6665 
6666 	sa.sa_handler = &signal_handler;
6667 
6668 	if (sigaction(SIGINT, &sa, NULL) < 0)
6669 		err(1, "sigaction SIGINT");
6670 	if (sigaction(SIGUSR1, &sa, NULL) < 0)
6671 		err(1, "sigaction SIGUSR1");
6672 }
6673 
do_sleep(void)6674 void do_sleep(void)
6675 {
6676 	struct timeval tout;
6677 	struct timespec rest;
6678 	fd_set readfds;
6679 	int retval;
6680 
6681 	FD_ZERO(&readfds);
6682 	FD_SET(0, &readfds);
6683 
6684 	if (ignore_stdin) {
6685 		nanosleep(&interval_ts, NULL);
6686 		return;
6687 	}
6688 
6689 	tout = interval_tv;
6690 	retval = select(1, &readfds, NULL, NULL, &tout);
6691 
6692 	if (retval == 1) {
6693 		switch (getc(stdin)) {
6694 		case 'q':
6695 			exit_requested = 1;
6696 			break;
6697 		case EOF:
6698 			/*
6699 			 * 'stdin' is a pipe closed on the other end. There
6700 			 * won't be any further input.
6701 			 */
6702 			ignore_stdin = 1;
6703 			/* Sleep the rest of the time */
6704 			rest.tv_sec = (tout.tv_sec + tout.tv_usec / 1000000);
6705 			rest.tv_nsec = (tout.tv_usec % 1000000) * 1000;
6706 			nanosleep(&rest, NULL);
6707 		}
6708 	}
6709 }
6710 
get_msr_sum(int cpu,off_t offset,unsigned long long * msr)6711 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr)
6712 {
6713 	int ret, idx;
6714 	unsigned long long msr_cur, msr_last;
6715 
6716 	assert(!no_msr);
6717 
6718 	if (!per_cpu_msr_sum)
6719 		return 1;
6720 
6721 	idx = offset_to_idx(offset);
6722 	if (idx < 0)
6723 		return idx;
6724 	/* get_msr_sum() = sum + (get_msr() - last) */
6725 	ret = get_msr(cpu, offset, &msr_cur);
6726 	if (ret)
6727 		return ret;
6728 	msr_last = per_cpu_msr_sum[cpu].entries[idx].last;
6729 	DELTA_WRAP32(msr_cur, msr_last);
6730 	*msr = msr_last + per_cpu_msr_sum[cpu].entries[idx].sum;
6731 
6732 	return 0;
6733 }
6734 
6735 timer_t timerid;
6736 
6737 /* Timer callback, update the sum of MSRs periodically. */
update_msr_sum(PER_THREAD_PARAMS)6738 static int update_msr_sum(PER_THREAD_PARAMS)
6739 {
6740 	int i, ret;
6741 	int cpu = t->cpu_id;
6742 
6743 	UNUSED(c);
6744 	UNUSED(p);
6745 
6746 	assert(!no_msr);
6747 
6748 	for (i = IDX_PKG_ENERGY; i < IDX_COUNT; i++) {
6749 		unsigned long long msr_cur, msr_last;
6750 		off_t offset;
6751 
6752 		if (!idx_valid(i))
6753 			continue;
6754 		offset = idx_to_offset(i);
6755 		if (offset < 0)
6756 			continue;
6757 		ret = get_msr(cpu, offset, &msr_cur);
6758 		if (ret) {
6759 			fprintf(outf, "Can not update msr(0x%llx)\n", (unsigned long long)offset);
6760 			continue;
6761 		}
6762 
6763 		msr_last = per_cpu_msr_sum[cpu].entries[i].last;
6764 		per_cpu_msr_sum[cpu].entries[i].last = msr_cur & 0xffffffff;
6765 
6766 		DELTA_WRAP32(msr_cur, msr_last);
6767 		per_cpu_msr_sum[cpu].entries[i].sum += msr_last;
6768 	}
6769 	return 0;
6770 }
6771 
msr_record_handler(union sigval v)6772 static void msr_record_handler(union sigval v)
6773 {
6774 	UNUSED(v);
6775 
6776 	for_all_cpus(update_msr_sum, EVEN_COUNTERS);
6777 }
6778 
msr_sum_record(void)6779 void msr_sum_record(void)
6780 {
6781 	struct itimerspec its;
6782 	struct sigevent sev;
6783 
6784 	per_cpu_msr_sum = calloc(topo.max_cpu_num + 1, sizeof(struct msr_sum_array));
6785 	if (!per_cpu_msr_sum) {
6786 		fprintf(outf, "Can not allocate memory for long time MSR.\n");
6787 		return;
6788 	}
6789 	/*
6790 	 * Signal handler might be restricted, so use thread notifier instead.
6791 	 */
6792 	memset(&sev, 0, sizeof(struct sigevent));
6793 	sev.sigev_notify = SIGEV_THREAD;
6794 	sev.sigev_notify_function = msr_record_handler;
6795 
6796 	sev.sigev_value.sival_ptr = &timerid;
6797 	if (timer_create(CLOCK_REALTIME, &sev, &timerid) == -1) {
6798 		fprintf(outf, "Can not create timer.\n");
6799 		goto release_msr;
6800 	}
6801 
6802 	its.it_value.tv_sec = 0;
6803 	its.it_value.tv_nsec = 1;
6804 	/*
6805 	 * A wraparound time has been calculated early.
6806 	 * Some sources state that the peak power for a
6807 	 * microprocessor is usually 1.5 times the TDP rating,
6808 	 * use 2 * TDP for safety.
6809 	 */
6810 	its.it_interval.tv_sec = rapl_joule_counter_range / 2;
6811 	its.it_interval.tv_nsec = 0;
6812 
6813 	if (timer_settime(timerid, 0, &its, NULL) == -1) {
6814 		fprintf(outf, "Can not set timer.\n");
6815 		goto release_timer;
6816 	}
6817 	return;
6818 
6819 release_timer:
6820 	timer_delete(timerid);
6821 release_msr:
6822 	free(per_cpu_msr_sum);
6823 	per_cpu_msr_sum = NULL;
6824 }
6825 
6826 /*
6827  * set_my_sched_priority(pri)
6828  * return previous priority on success
6829  * return value < -20 on failure
6830  */
set_my_sched_priority(int priority)6831 int set_my_sched_priority(int priority)
6832 {
6833 	int retval;
6834 	int original_priority;
6835 
6836 	errno = 0;
6837 	original_priority = getpriority(PRIO_PROCESS, 0);
6838 	if (errno && (original_priority == -1))
6839 		return -21;
6840 
6841 	retval = setpriority(PRIO_PROCESS, 0, priority);
6842 	if (retval)
6843 		return -21;
6844 
6845 	errno = 0;
6846 	retval = getpriority(PRIO_PROCESS, 0);
6847 	if (retval != priority)
6848 		return -21;
6849 
6850 	return original_priority;
6851 }
6852 
turbostat_loop()6853 void turbostat_loop()
6854 {
6855 	int retval;
6856 	int restarted = 0;
6857 	unsigned int done_iters = 0;
6858 
6859 	setup_signal_handler();
6860 
6861 	/*
6862 	 * elevate own priority for interval mode
6863 	 *
6864 	 * ignore on error - we probably don't have permission to set it, but
6865 	 * it's not a big deal
6866 	 */
6867 	set_my_sched_priority(-20);
6868 
6869 restart:
6870 	restarted++;
6871 
6872 	snapshot_proc_sysfs_files();
6873 	retval = for_all_cpus(get_counters, EVEN_COUNTERS);
6874 	first_counter_read = 0;
6875 	if (retval < -1) {
6876 		exit(retval);
6877 	} else if (retval == -1) {
6878 		if (restarted > 10) {
6879 			exit(retval);
6880 		}
6881 		re_initialize();
6882 		goto restart;
6883 	}
6884 	restarted = 0;
6885 	done_iters = 0;
6886 	gettimeofday(&tv_even, (struct timezone *)NULL);
6887 
6888 	while (1) {
6889 		if (for_all_proc_cpus(cpu_is_not_present)) {
6890 			re_initialize();
6891 			goto restart;
6892 		}
6893 		if (update_effective_str(false)) {
6894 			re_initialize();
6895 			goto restart;
6896 		}
6897 		do_sleep();
6898 		if (snapshot_proc_sysfs_files())
6899 			goto restart;
6900 		retval = for_all_cpus(get_counters, ODD_COUNTERS);
6901 		if (retval < -1) {
6902 			exit(retval);
6903 		} else if (retval == -1) {
6904 			re_initialize();
6905 			goto restart;
6906 		}
6907 		gettimeofday(&tv_odd, (struct timezone *)NULL);
6908 		timersub(&tv_odd, &tv_even, &tv_delta);
6909 		if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) {
6910 			re_initialize();
6911 			goto restart;
6912 		}
6913 		delta_platform(&platform_counters_odd, &platform_counters_even);
6914 		compute_average(EVEN_COUNTERS);
6915 		format_all_counters(EVEN_COUNTERS);
6916 		flush_output_stdout();
6917 		if (exit_requested)
6918 			break;
6919 		if (num_iterations && ++done_iters >= num_iterations)
6920 			break;
6921 		do_sleep();
6922 		if (snapshot_proc_sysfs_files())
6923 			goto restart;
6924 		retval = for_all_cpus(get_counters, EVEN_COUNTERS);
6925 		if (retval < -1) {
6926 			exit(retval);
6927 		} else if (retval == -1) {
6928 			re_initialize();
6929 			goto restart;
6930 		}
6931 		gettimeofday(&tv_even, (struct timezone *)NULL);
6932 		timersub(&tv_even, &tv_odd, &tv_delta);
6933 		if (for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) {
6934 			re_initialize();
6935 			goto restart;
6936 		}
6937 		delta_platform(&platform_counters_even, &platform_counters_odd);
6938 		compute_average(ODD_COUNTERS);
6939 		format_all_counters(ODD_COUNTERS);
6940 		flush_output_stdout();
6941 		if (exit_requested)
6942 			break;
6943 		if (num_iterations && ++done_iters >= num_iterations)
6944 			break;
6945 	}
6946 }
6947 
probe_dev_msr(void)6948 int probe_dev_msr(void)
6949 {
6950 	struct stat sb;
6951 	char pathname[32];
6952 
6953 	sprintf(pathname, "/dev/msr%d", master_cpu);
6954 	return !stat(pathname, &sb);
6955 }
6956 
probe_dev_cpu_msr(void)6957 int probe_dev_cpu_msr(void)
6958 {
6959 	struct stat sb;
6960 	char pathname[32];
6961 
6962 	sprintf(pathname, "/dev/cpu/%d/msr", master_cpu);
6963 	return !stat(pathname, &sb);
6964 }
6965 
probe_msr_driver(void)6966 int probe_msr_driver(void)
6967 {
6968 	if (probe_dev_msr()) {
6969 		use_android_msr_path = 1;
6970 		return 1;
6971 	}
6972 	return probe_dev_cpu_msr();
6973 }
6974 
check_msr_driver(void)6975 void check_msr_driver(void)
6976 {
6977 	if (probe_msr_driver())
6978 		return;
6979 
6980 	if (system("/sbin/modprobe msr > /dev/null 2>&1"))
6981 		no_msr = 1;
6982 
6983 	if (!probe_msr_driver())
6984 		no_msr = 1;
6985 }
6986 
6987 /*
6988  * check for CAP_SYS_RAWIO
6989  * return 0 on success
6990  * return 1 on fail
6991  */
check_for_cap_sys_rawio(void)6992 int check_for_cap_sys_rawio(void)
6993 {
6994 	cap_t caps;
6995 	cap_flag_value_t cap_flag_value;
6996 	int ret = 0;
6997 
6998 	caps = cap_get_proc();
6999 	if (caps == NULL) {
7000 		/*
7001 		 * CONFIG_MULTIUSER=n kernels have no cap_get_proc()
7002 		 * Allow them to continue and attempt to access MSRs
7003 		 */
7004 		if (errno == ENOSYS)
7005 			return 0;
7006 
7007 		return 1;
7008 	}
7009 
7010 	if (cap_get_flag(caps, CAP_SYS_RAWIO, CAP_EFFECTIVE, &cap_flag_value)) {
7011 		ret = 1;
7012 		goto free_and_exit;
7013 	}
7014 
7015 	if (cap_flag_value != CAP_SET) {
7016 		ret = 1;
7017 		goto free_and_exit;
7018 	}
7019 
7020 free_and_exit:
7021 	if (cap_free(caps) == -1)
7022 		err(-6, "cap_free");
7023 
7024 	return ret;
7025 }
7026 
check_msr_permission(void)7027 void check_msr_permission(void)
7028 {
7029 	int failed = 0;
7030 	char pathname[32];
7031 
7032 	if (no_msr)
7033 		return;
7034 
7035 	/* check for CAP_SYS_RAWIO */
7036 	failed += check_for_cap_sys_rawio();
7037 
7038 	/* test file permissions */
7039 	sprintf(pathname, use_android_msr_path ? "/dev/msr%d" : "/dev/cpu/%d/msr", master_cpu);
7040 	if (euidaccess(pathname, R_OK)) {
7041 		failed++;
7042 	}
7043 
7044 	if (failed) {
7045 		warnx("Failed to access %s. Some of the counters may not be available\n"
7046 		      "\tRun as root to enable them or use %s to disable the access explicitly", pathname, "--no-msr");
7047 		no_msr = 1;
7048 	}
7049 }
7050 
probe_bclk(void)7051 void probe_bclk(void)
7052 {
7053 	unsigned long long msr;
7054 	unsigned int base_ratio;
7055 
7056 	if (!platform->has_nhm_msrs || no_msr)
7057 		return;
7058 
7059 	if (platform->bclk_freq == BCLK_100MHZ)
7060 		bclk = 100.00;
7061 	else if (platform->bclk_freq == BCLK_133MHZ)
7062 		bclk = 133.33;
7063 	else if (platform->bclk_freq == BCLK_SLV)
7064 		bclk = slm_bclk();
7065 	else
7066 		return;
7067 
7068 	get_msr(master_cpu, MSR_PLATFORM_INFO, &msr);
7069 	base_ratio = (msr >> 8) & 0xFF;
7070 
7071 	base_hz = base_ratio * bclk * 1000000;
7072 	has_base_hz = 1;
7073 
7074 	if (platform->enable_tsc_tweak)
7075 		tsc_tweak = base_hz / tsc_hz;
7076 }
7077 
remove_underbar(char * s)7078 static void remove_underbar(char *s)
7079 {
7080 	char *to = s;
7081 
7082 	while (*s) {
7083 		if (*s != '_')
7084 			*to++ = *s;
7085 		s++;
7086 	}
7087 
7088 	*to = 0;
7089 }
7090 
dump_turbo_ratio_info(void)7091 static void dump_turbo_ratio_info(void)
7092 {
7093 	if (!has_turbo)
7094 		return;
7095 
7096 	if (!platform->has_nhm_msrs || no_msr)
7097 		return;
7098 
7099 	if (platform->trl_msrs & TRL_LIMIT2)
7100 		dump_turbo_ratio_limit2();
7101 
7102 	if (platform->trl_msrs & TRL_LIMIT1)
7103 		dump_turbo_ratio_limit1();
7104 
7105 	if (platform->trl_msrs & TRL_BASE) {
7106 		dump_turbo_ratio_limits(MSR_TURBO_RATIO_LIMIT);
7107 
7108 		if (is_hybrid)
7109 			dump_turbo_ratio_limits(MSR_SECONDARY_TURBO_RATIO_LIMIT);
7110 	}
7111 
7112 	if (platform->trl_msrs & TRL_ATOM)
7113 		dump_atom_turbo_ratio_limits();
7114 
7115 	if (platform->trl_msrs & TRL_KNL)
7116 		dump_knl_turbo_ratio_limits();
7117 
7118 	if (platform->has_config_tdp)
7119 		dump_config_tdp();
7120 }
7121 
read_sysfs_int(char * path)7122 static int read_sysfs_int(char *path)
7123 {
7124 	FILE *input;
7125 	int retval = -1;
7126 
7127 	input = fopen(path, "r");
7128 	if (input == NULL) {
7129 		if (debug)
7130 			fprintf(outf, "NSFOD %s\n", path);
7131 		return (-1);
7132 	}
7133 	if (fscanf(input, "%d", &retval) != 1)
7134 		err(1, "%s: failed to read int from file", path);
7135 	fclose(input);
7136 
7137 	return (retval);
7138 }
7139 
dump_sysfs_file(char * path)7140 static void dump_sysfs_file(char *path)
7141 {
7142 	FILE *input;
7143 	char cpuidle_buf[64];
7144 
7145 	input = fopen(path, "r");
7146 	if (input == NULL) {
7147 		if (debug)
7148 			fprintf(outf, "NSFOD %s\n", path);
7149 		return;
7150 	}
7151 	if (!fgets(cpuidle_buf, sizeof(cpuidle_buf), input))
7152 		err(1, "%s: failed to read file", path);
7153 	fclose(input);
7154 
7155 	fprintf(outf, "%s: %s", strrchr(path, '/') + 1, cpuidle_buf);
7156 }
7157 
probe_intel_uncore_frequency_legacy(void)7158 static void probe_intel_uncore_frequency_legacy(void)
7159 {
7160 	int i, j;
7161 	char path[256];
7162 
7163 	for (i = 0; i < topo.num_packages; ++i) {
7164 		for (j = 0; j <= topo.max_die_id; ++j) {
7165 			int k, l;
7166 			char path_base[128];
7167 
7168 			sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/package_%02d_die_%02d", i, j);
7169 
7170 			sprintf(path, "%s/current_freq_khz", path_base);
7171 			if (access(path, R_OK))
7172 				continue;
7173 
7174 			BIC_PRESENT(BIC_UNCORE_MHZ);
7175 
7176 			if (quiet)
7177 				return;
7178 
7179 			sprintf(path, "%s/min_freq_khz", path_base);
7180 			k = read_sysfs_int(path);
7181 			sprintf(path, "%s/max_freq_khz", path_base);
7182 			l = read_sysfs_int(path);
7183 			fprintf(outf, "Uncore Frequency package%d die%d: %d - %d MHz ", i, j, k / 1000, l / 1000);
7184 
7185 			sprintf(path, "%s/initial_min_freq_khz", path_base);
7186 			k = read_sysfs_int(path);
7187 			sprintf(path, "%s/initial_max_freq_khz", path_base);
7188 			l = read_sysfs_int(path);
7189 			fprintf(outf, "(%d - %d MHz)", k / 1000, l / 1000);
7190 
7191 			sprintf(path, "%s/current_freq_khz", path_base);
7192 			k = read_sysfs_int(path);
7193 			fprintf(outf, " %d MHz\n", k / 1000);
7194 		}
7195 	}
7196 }
7197 
probe_intel_uncore_frequency_cluster(void)7198 static void probe_intel_uncore_frequency_cluster(void)
7199 {
7200 	int i, uncore_max_id;
7201 	char path[256];
7202 	char path_base[128];
7203 
7204 	if (access("/sys/devices/system/cpu/intel_uncore_frequency/uncore00/current_freq_khz", R_OK))
7205 		return;
7206 
7207 	for (uncore_max_id = 0;; ++uncore_max_id) {
7208 
7209 		sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/uncore%02d", uncore_max_id);
7210 
7211 		/* uncore## start at 00 and skips no numbers, so stop upon first missing */
7212 		if (access(path_base, R_OK)) {
7213 			uncore_max_id -= 1;
7214 			break;
7215 		}
7216 	}
7217 	for (i = uncore_max_id; i >= 0; --i) {
7218 		int k, l;
7219 		int unc_pkg_id, domain_id, cluster_id;
7220 		char name_buf[16];
7221 
7222 		sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/uncore%02d", i);
7223 
7224 		if (access(path_base, R_OK))
7225 			err(1, "%s: %s", __func__, path_base);
7226 
7227 		sprintf(path, "%s/package_id", path_base);
7228 		unc_pkg_id = read_sysfs_int(path);
7229 
7230 		sprintf(path, "%s/domain_id", path_base);
7231 		domain_id = read_sysfs_int(path);
7232 
7233 		sprintf(path, "%s/fabric_cluster_id", path_base);
7234 		cluster_id = read_sysfs_int(path);
7235 
7236 		sprintf(path, "%s/current_freq_khz", path_base);
7237 		sprintf(name_buf, "UMHz%d.%d", domain_id, cluster_id);
7238 
7239 		/*
7240 		 * Once add_couter() is called, that counter is always read
7241 		 * and reported -- So it is effectively (enabled & present).
7242 		 * Only call add_counter() here if legacy BIC_UNCORE_MHZ (UncMHz)
7243 		 * is (enabled).  Since we are in this routine, we
7244 		 * know we will not probe and set (present) the legacy counter.
7245 		 *
7246 		 * This allows "--show/--hide UncMHz" to be effective for
7247 		 * the clustered MHz counters, as a group.
7248 		 */
7249 		if BIC_IS_ENABLED
7250 			(BIC_UNCORE_MHZ)
7251 			    add_counter(0, path, name_buf, 0, SCOPE_PACKAGE, COUNTER_K2M, FORMAT_AVERAGE, 0, unc_pkg_id);
7252 
7253 		if (quiet)
7254 			continue;
7255 
7256 		sprintf(path, "%s/min_freq_khz", path_base);
7257 		k = read_sysfs_int(path);
7258 		sprintf(path, "%s/max_freq_khz", path_base);
7259 		l = read_sysfs_int(path);
7260 		fprintf(outf, "Uncore Frequency package%d domain%d cluster%d: %d - %d MHz ", unc_pkg_id, domain_id, cluster_id, k / 1000, l / 1000);
7261 
7262 		sprintf(path, "%s/initial_min_freq_khz", path_base);
7263 		k = read_sysfs_int(path);
7264 		sprintf(path, "%s/initial_max_freq_khz", path_base);
7265 		l = read_sysfs_int(path);
7266 		fprintf(outf, "(%d - %d MHz)", k / 1000, l / 1000);
7267 
7268 		sprintf(path, "%s/current_freq_khz", path_base);
7269 		k = read_sysfs_int(path);
7270 		fprintf(outf, " %d MHz\n", k / 1000);
7271 	}
7272 }
7273 
probe_intel_uncore_frequency(void)7274 static void probe_intel_uncore_frequency(void)
7275 {
7276 	if (!genuine_intel)
7277 		return;
7278 
7279 	if (access("/sys/devices/system/cpu/intel_uncore_frequency/uncore00", R_OK) == 0)
7280 		probe_intel_uncore_frequency_cluster();
7281 	else
7282 		probe_intel_uncore_frequency_legacy();
7283 }
7284 
set_graphics_fp(char * path,int idx)7285 static void set_graphics_fp(char *path, int idx)
7286 {
7287 	if (!access(path, R_OK))
7288 		gfx_info[idx].fp = fopen_or_die(path, "r");
7289 }
7290 
7291 /* Enlarge this if there are /sys/class/drm/card2 ... */
7292 #define GFX_MAX_CARDS	2
7293 
probe_graphics(void)7294 static void probe_graphics(void)
7295 {
7296 	char path[PATH_MAX];
7297 	int i;
7298 
7299 	/* Xe graphics sysfs knobs */
7300 	if (!access("/sys/class/drm/card0/device/tile0/gt0/gtidle/idle_residency_ms", R_OK)) {
7301 		FILE *fp;
7302 		char buf[8];
7303 		bool gt0_is_gt;
7304 
7305 		fp = fopen("/sys/class/drm/card0/device/tile0/gt0/gtidle/name", "r");
7306 		if (!fp)
7307 			goto next;
7308 
7309 		if (!fread(buf, sizeof(char), 7, fp)) {
7310 			fclose(fp);
7311 			goto next;
7312 		}
7313 		fclose(fp);
7314 
7315 		if (!strncmp(buf, "gt0-rc", strlen("gt0-rc")))
7316 			gt0_is_gt = true;
7317 		else if (!strncmp(buf, "gt0-mc", strlen("gt0-mc")))
7318 			gt0_is_gt = false;
7319 		else
7320 			goto next;
7321 
7322 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/gtidle/idle_residency_ms", gt0_is_gt ? GFX_rc6 : SAM_mc6);
7323 
7324 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/freq0/cur_freq", gt0_is_gt ? GFX_MHz : SAM_MHz);
7325 
7326 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/freq0/act_freq", gt0_is_gt ? GFX_ACTMHz : SAM_ACTMHz);
7327 
7328 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/gtidle/idle_residency_ms", gt0_is_gt ? SAM_mc6 : GFX_rc6);
7329 
7330 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/freq0/cur_freq", gt0_is_gt ? SAM_MHz : GFX_MHz);
7331 
7332 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/freq0/act_freq", gt0_is_gt ? SAM_ACTMHz : GFX_ACTMHz);
7333 
7334 		goto end;
7335 	}
7336 
7337 next:
7338 	/* New i915 graphics sysfs knobs */
7339 	for (i = 0; i < GFX_MAX_CARDS; i++) {
7340 		snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rc6_residency_ms", i);
7341 		if (!access(path, R_OK))
7342 			break;
7343 	}
7344 
7345 	if (i == GFX_MAX_CARDS)
7346 		goto legacy_i915;
7347 
7348 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rc6_residency_ms", i);
7349 	set_graphics_fp(path, GFX_rc6);
7350 
7351 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rps_cur_freq_mhz", i);
7352 	set_graphics_fp(path, GFX_MHz);
7353 
7354 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rps_act_freq_mhz", i);
7355 	set_graphics_fp(path, GFX_ACTMHz);
7356 
7357 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rc6_residency_ms", i);
7358 	set_graphics_fp(path, SAM_mc6);
7359 
7360 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rps_cur_freq_mhz", i);
7361 	set_graphics_fp(path, SAM_MHz);
7362 
7363 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rps_act_freq_mhz", i);
7364 	set_graphics_fp(path, SAM_ACTMHz);
7365 
7366 	goto end;
7367 
7368 legacy_i915:
7369 	/* Fall back to traditional i915 graphics sysfs knobs */
7370 	set_graphics_fp("/sys/class/drm/card0/power/rc6_residency_ms", GFX_rc6);
7371 
7372 	set_graphics_fp("/sys/class/drm/card0/gt_cur_freq_mhz", GFX_MHz);
7373 	if (!gfx_info[GFX_MHz].fp)
7374 		set_graphics_fp("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", GFX_MHz);
7375 
7376 	set_graphics_fp("/sys/class/drm/card0/gt_act_freq_mhz", GFX_ACTMHz);
7377 	if (!gfx_info[GFX_ACTMHz].fp)
7378 		set_graphics_fp("/sys/class/graphics/fb0/device/drm/card0/gt_act_freq_mhz", GFX_ACTMHz);
7379 
7380 end:
7381 	if (gfx_info[GFX_rc6].fp)
7382 		BIC_PRESENT(BIC_GFX_rc6);
7383 	if (gfx_info[GFX_MHz].fp)
7384 		BIC_PRESENT(BIC_GFXMHz);
7385 	if (gfx_info[GFX_ACTMHz].fp)
7386 		BIC_PRESENT(BIC_GFXACTMHz);
7387 	if (gfx_info[SAM_mc6].fp)
7388 		BIC_PRESENT(BIC_SAM_mc6);
7389 	if (gfx_info[SAM_MHz].fp)
7390 		BIC_PRESENT(BIC_SAMMHz);
7391 	if (gfx_info[SAM_ACTMHz].fp)
7392 		BIC_PRESENT(BIC_SAMACTMHz);
7393 }
7394 
dump_sysfs_cstate_config(void)7395 static void dump_sysfs_cstate_config(void)
7396 {
7397 	char path[64];
7398 	char name_buf[16];
7399 	char desc[64];
7400 	FILE *input;
7401 	int state;
7402 	char *sp;
7403 
7404 	if (access("/sys/devices/system/cpu/cpuidle", R_OK)) {
7405 		fprintf(outf, "cpuidle not loaded\n");
7406 		return;
7407 	}
7408 
7409 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_driver");
7410 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor");
7411 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor_ro");
7412 
7413 	for (state = 0; state < 10; ++state) {
7414 
7415 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", master_cpu, state);
7416 		input = fopen(path, "r");
7417 		if (input == NULL)
7418 			continue;
7419 		if (!fgets(name_buf, sizeof(name_buf), input))
7420 			err(1, "%s: failed to read file", path);
7421 
7422 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
7423 		sp = strchr(name_buf, '-');
7424 		if (!sp)
7425 			sp = strchrnul(name_buf, '\n');
7426 		*sp = '\0';
7427 		fclose(input);
7428 
7429 		remove_underbar(name_buf);
7430 
7431 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/desc", master_cpu, state);
7432 		input = fopen(path, "r");
7433 		if (input == NULL)
7434 			continue;
7435 		if (!fgets(desc, sizeof(desc), input))
7436 			err(1, "%s: failed to read file", path);
7437 
7438 		fprintf(outf, "cpu%d: %s: %s", master_cpu, name_buf, desc);
7439 		fclose(input);
7440 	}
7441 }
7442 
dump_sysfs_pstate_config(void)7443 static void dump_sysfs_pstate_config(void)
7444 {
7445 	char path[64];
7446 	char driver_buf[64];
7447 	char governor_buf[64];
7448 	FILE *input;
7449 	int turbo;
7450 
7451 	sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_driver", master_cpu);
7452 	input = fopen(path, "r");
7453 	if (input == NULL) {
7454 		fprintf(outf, "NSFOD %s\n", path);
7455 		return;
7456 	}
7457 	if (!fgets(driver_buf, sizeof(driver_buf), input))
7458 		err(1, "%s: failed to read file", path);
7459 	fclose(input);
7460 
7461 	sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_governor", master_cpu);
7462 	input = fopen(path, "r");
7463 	if (input == NULL) {
7464 		fprintf(outf, "NSFOD %s\n", path);
7465 		return;
7466 	}
7467 	if (!fgets(governor_buf, sizeof(governor_buf), input))
7468 		err(1, "%s: failed to read file", path);
7469 	fclose(input);
7470 
7471 	fprintf(outf, "cpu%d: cpufreq driver: %s", master_cpu, driver_buf);
7472 	fprintf(outf, "cpu%d: cpufreq governor: %s", master_cpu, governor_buf);
7473 
7474 	sprintf(path, "/sys/devices/system/cpu/cpufreq/boost");
7475 	input = fopen(path, "r");
7476 	if (input != NULL) {
7477 		if (fscanf(input, "%d", &turbo) != 1)
7478 			err(1, "%s: failed to parse number from file", path);
7479 		fprintf(outf, "cpufreq boost: %d\n", turbo);
7480 		fclose(input);
7481 	}
7482 
7483 	sprintf(path, "/sys/devices/system/cpu/intel_pstate/no_turbo");
7484 	input = fopen(path, "r");
7485 	if (input != NULL) {
7486 		if (fscanf(input, "%d", &turbo) != 1)
7487 			err(1, "%s: failed to parse number from file", path);
7488 		fprintf(outf, "cpufreq intel_pstate no_turbo: %d\n", turbo);
7489 		fclose(input);
7490 	}
7491 }
7492 
7493 /*
7494  * print_epb()
7495  * Decode the ENERGY_PERF_BIAS MSR
7496  */
print_epb(PER_THREAD_PARAMS)7497 int print_epb(PER_THREAD_PARAMS)
7498 {
7499 	char *epb_string;
7500 	int cpu, epb;
7501 
7502 	UNUSED(c);
7503 	UNUSED(p);
7504 
7505 	if (!has_epb)
7506 		return 0;
7507 
7508 	cpu = t->cpu_id;
7509 
7510 	/* EPB is per-package */
7511 	if (!is_cpu_first_thread_in_package(t, c, p))
7512 		return 0;
7513 
7514 	if (cpu_migrate(cpu)) {
7515 		fprintf(outf, "print_epb: Could not migrate to CPU %d\n", cpu);
7516 		return -1;
7517 	}
7518 
7519 	epb = get_epb(cpu);
7520 	if (epb < 0)
7521 		return 0;
7522 
7523 	switch (epb) {
7524 	case ENERGY_PERF_BIAS_PERFORMANCE:
7525 		epb_string = "performance";
7526 		break;
7527 	case ENERGY_PERF_BIAS_NORMAL:
7528 		epb_string = "balanced";
7529 		break;
7530 	case ENERGY_PERF_BIAS_POWERSAVE:
7531 		epb_string = "powersave";
7532 		break;
7533 	default:
7534 		epb_string = "custom";
7535 		break;
7536 	}
7537 	fprintf(outf, "cpu%d: EPB: %d (%s)\n", cpu, epb, epb_string);
7538 
7539 	return 0;
7540 }
7541 
7542 /*
7543  * print_hwp()
7544  * Decode the MSR_HWP_CAPABILITIES
7545  */
print_hwp(PER_THREAD_PARAMS)7546 int print_hwp(PER_THREAD_PARAMS)
7547 {
7548 	unsigned long long msr;
7549 	int cpu;
7550 
7551 	UNUSED(c);
7552 	UNUSED(p);
7553 
7554 	if (no_msr)
7555 		return 0;
7556 
7557 	if (!has_hwp)
7558 		return 0;
7559 
7560 	cpu = t->cpu_id;
7561 
7562 	/* MSR_HWP_CAPABILITIES is per-package */
7563 	if (!is_cpu_first_thread_in_package(t, c, p))
7564 		return 0;
7565 
7566 	if (cpu_migrate(cpu)) {
7567 		fprintf(outf, "print_hwp: Could not migrate to CPU %d\n", cpu);
7568 		return -1;
7569 	}
7570 
7571 	if (get_msr(cpu, MSR_PM_ENABLE, &msr))
7572 		return 0;
7573 
7574 	fprintf(outf, "cpu%d: MSR_PM_ENABLE: 0x%08llx (%sHWP)\n", cpu, msr, (msr & (1 << 0)) ? "" : "No-");
7575 
7576 	/* MSR_PM_ENABLE[1] == 1 if HWP is enabled and MSRs visible */
7577 	if ((msr & (1 << 0)) == 0)
7578 		return 0;
7579 
7580 	if (get_msr(cpu, MSR_HWP_CAPABILITIES, &msr))
7581 		return 0;
7582 
7583 	fprintf(outf, "cpu%d: MSR_HWP_CAPABILITIES: 0x%08llx "
7584 		"(high %d guar %d eff %d low %d)\n",
7585 		cpu, msr,
7586 		(unsigned int)HWP_HIGHEST_PERF(msr),
7587 		(unsigned int)HWP_GUARANTEED_PERF(msr), (unsigned int)HWP_MOSTEFFICIENT_PERF(msr), (unsigned int)HWP_LOWEST_PERF(msr));
7588 
7589 	if (get_msr(cpu, MSR_HWP_REQUEST, &msr))
7590 		return 0;
7591 
7592 	fprintf(outf, "cpu%d: MSR_HWP_REQUEST: 0x%08llx "
7593 		"(min %d max %d des %d epp 0x%x window 0x%x pkg 0x%x)\n",
7594 		cpu, msr,
7595 		(unsigned int)(((msr) >> 0) & 0xff),
7596 		(unsigned int)(((msr) >> 8) & 0xff),
7597 		(unsigned int)(((msr) >> 16) & 0xff),
7598 		(unsigned int)(((msr) >> 24) & 0xff), (unsigned int)(((msr) >> 32) & 0xff3), (unsigned int)(((msr) >> 42) & 0x1));
7599 
7600 	if (has_hwp_pkg) {
7601 		if (get_msr(cpu, MSR_HWP_REQUEST_PKG, &msr))
7602 			return 0;
7603 
7604 		fprintf(outf, "cpu%d: MSR_HWP_REQUEST_PKG: 0x%08llx "
7605 			"(min %d max %d des %d epp 0x%x window 0x%x)\n",
7606 			cpu, msr,
7607 			(unsigned int)(((msr) >> 0) & 0xff),
7608 			(unsigned int)(((msr) >> 8) & 0xff),
7609 			(unsigned int)(((msr) >> 16) & 0xff), (unsigned int)(((msr) >> 24) & 0xff), (unsigned int)(((msr) >> 32) & 0xff3));
7610 	}
7611 	if (has_hwp_notify) {
7612 		if (get_msr(cpu, MSR_HWP_INTERRUPT, &msr))
7613 			return 0;
7614 
7615 		fprintf(outf, "cpu%d: MSR_HWP_INTERRUPT: 0x%08llx "
7616 			"(%s_Guaranteed_Perf_Change, %s_Excursion_Min)\n", cpu, msr, ((msr) & 0x1) ? "EN" : "Dis", ((msr) & 0x2) ? "EN" : "Dis");
7617 	}
7618 	if (get_msr(cpu, MSR_HWP_STATUS, &msr))
7619 		return 0;
7620 
7621 	fprintf(outf, "cpu%d: MSR_HWP_STATUS: 0x%08llx "
7622 		"(%sGuaranteed_Perf_Change, %sExcursion_Min)\n", cpu, msr, ((msr) & 0x1) ? "" : "No-", ((msr) & 0x4) ? "" : "No-");
7623 
7624 	return 0;
7625 }
7626 
7627 /*
7628  * print_perf_limit()
7629  */
print_perf_limit(PER_THREAD_PARAMS)7630 int print_perf_limit(PER_THREAD_PARAMS)
7631 {
7632 	unsigned long long msr;
7633 	int cpu;
7634 
7635 	UNUSED(c);
7636 	UNUSED(p);
7637 
7638 	if (no_msr)
7639 		return 0;
7640 
7641 	cpu = t->cpu_id;
7642 
7643 	/* per-package */
7644 	if (!is_cpu_first_thread_in_package(t, c, p))
7645 		return 0;
7646 
7647 	if (cpu_migrate(cpu)) {
7648 		fprintf(outf, "print_perf_limit: Could not migrate to CPU %d\n", cpu);
7649 		return -1;
7650 	}
7651 
7652 	if (platform->plr_msrs & PLR_CORE) {
7653 		get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
7654 		fprintf(outf, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
7655 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
7656 			(msr & 1 << 15) ? "bit15, " : "",
7657 			(msr & 1 << 14) ? "bit14, " : "",
7658 			(msr & 1 << 13) ? "Transitions, " : "",
7659 			(msr & 1 << 12) ? "MultiCoreTurbo, " : "",
7660 			(msr & 1 << 11) ? "PkgPwrL2, " : "",
7661 			(msr & 1 << 10) ? "PkgPwrL1, " : "",
7662 			(msr & 1 << 9) ? "CorePwr, " : "",
7663 			(msr & 1 << 8) ? "Amps, " : "",
7664 			(msr & 1 << 6) ? "VR-Therm, " : "",
7665 			(msr & 1 << 5) ? "Auto-HWP, " : "",
7666 			(msr & 1 << 4) ? "Graphics, " : "",
7667 			(msr & 1 << 2) ? "bit2, " : "", (msr & 1 << 1) ? "ThermStatus, " : "", (msr & 1 << 0) ? "PROCHOT, " : "");
7668 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
7669 			(msr & 1 << 31) ? "bit31, " : "",
7670 			(msr & 1 << 30) ? "bit30, " : "",
7671 			(msr & 1 << 29) ? "Transitions, " : "",
7672 			(msr & 1 << 28) ? "MultiCoreTurbo, " : "",
7673 			(msr & 1 << 27) ? "PkgPwrL2, " : "",
7674 			(msr & 1 << 26) ? "PkgPwrL1, " : "",
7675 			(msr & 1 << 25) ? "CorePwr, " : "",
7676 			(msr & 1 << 24) ? "Amps, " : "",
7677 			(msr & 1 << 22) ? "VR-Therm, " : "",
7678 			(msr & 1 << 21) ? "Auto-HWP, " : "",
7679 			(msr & 1 << 20) ? "Graphics, " : "",
7680 			(msr & 1 << 18) ? "bit18, " : "", (msr & 1 << 17) ? "ThermStatus, " : "", (msr & 1 << 16) ? "PROCHOT, " : "");
7681 
7682 	}
7683 	if (platform->plr_msrs & PLR_GFX) {
7684 		get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
7685 		fprintf(outf, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
7686 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s)",
7687 			(msr & 1 << 0) ? "PROCHOT, " : "",
7688 			(msr & 1 << 1) ? "ThermStatus, " : "",
7689 			(msr & 1 << 4) ? "Graphics, " : "",
7690 			(msr & 1 << 6) ? "VR-Therm, " : "",
7691 			(msr & 1 << 8) ? "Amps, " : "",
7692 			(msr & 1 << 9) ? "GFXPwr, " : "", (msr & 1 << 10) ? "PkgPwrL1, " : "", (msr & 1 << 11) ? "PkgPwrL2, " : "");
7693 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s)\n",
7694 			(msr & 1 << 16) ? "PROCHOT, " : "",
7695 			(msr & 1 << 17) ? "ThermStatus, " : "",
7696 			(msr & 1 << 20) ? "Graphics, " : "",
7697 			(msr & 1 << 22) ? "VR-Therm, " : "",
7698 			(msr & 1 << 24) ? "Amps, " : "",
7699 			(msr & 1 << 25) ? "GFXPwr, " : "", (msr & 1 << 26) ? "PkgPwrL1, " : "", (msr & 1 << 27) ? "PkgPwrL2, " : "");
7700 	}
7701 	if (platform->plr_msrs & PLR_RING) {
7702 		get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
7703 		fprintf(outf, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
7704 		fprintf(outf, " (Active: %s%s%s%s%s%s)",
7705 			(msr & 1 << 0) ? "PROCHOT, " : "",
7706 			(msr & 1 << 1) ? "ThermStatus, " : "",
7707 			(msr & 1 << 6) ? "VR-Therm, " : "",
7708 			(msr & 1 << 8) ? "Amps, " : "", (msr & 1 << 10) ? "PkgPwrL1, " : "", (msr & 1 << 11) ? "PkgPwrL2, " : "");
7709 		fprintf(outf, " (Logged: %s%s%s%s%s%s)\n",
7710 			(msr & 1 << 16) ? "PROCHOT, " : "",
7711 			(msr & 1 << 17) ? "ThermStatus, " : "",
7712 			(msr & 1 << 22) ? "VR-Therm, " : "",
7713 			(msr & 1 << 24) ? "Amps, " : "", (msr & 1 << 26) ? "PkgPwrL1, " : "", (msr & 1 << 27) ? "PkgPwrL2, " : "");
7714 	}
7715 	return 0;
7716 }
7717 
7718 #define	RAPL_POWER_GRANULARITY	0x7FFF	/* 15 bit power granularity */
7719 #define	RAPL_TIME_GRANULARITY	0x3F	/* 6 bit time granularity */
7720 
get_quirk_tdp(void)7721 double get_quirk_tdp(void)
7722 {
7723 	if (platform->rapl_quirk_tdp)
7724 		return platform->rapl_quirk_tdp;
7725 
7726 	return 135.0;
7727 }
7728 
get_tdp_intel(void)7729 double get_tdp_intel(void)
7730 {
7731 	unsigned long long msr;
7732 
7733 	if (valid_rapl_msrs & RAPL_PKG_POWER_INFO)
7734 		if (!get_msr(master_cpu, MSR_PKG_POWER_INFO, &msr))
7735 			return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
7736 	return get_quirk_tdp();
7737 }
7738 
get_tdp_amd(void)7739 double get_tdp_amd(void)
7740 {
7741 	return get_quirk_tdp();
7742 }
7743 
rapl_probe_intel(void)7744 void rapl_probe_intel(void)
7745 {
7746 	unsigned long long msr;
7747 	unsigned int time_unit;
7748 	double tdp;
7749 
7750 	if (rapl_joules) {
7751 		CLR_BIC(BIC_SysWatt, &bic_enabled);
7752 		CLR_BIC(BIC_PkgWatt, &bic_enabled);
7753 		CLR_BIC(BIC_CorWatt, &bic_enabled);
7754 		CLR_BIC(BIC_RAMWatt, &bic_enabled);
7755 		CLR_BIC(BIC_GFXWatt, &bic_enabled);
7756 	} else {
7757 		CLR_BIC(BIC_Sys_J, &bic_enabled);
7758 		CLR_BIC(BIC_Pkg_J, &bic_enabled);
7759 		CLR_BIC(BIC_Cor_J, &bic_enabled);
7760 		CLR_BIC(BIC_RAM_J, &bic_enabled);
7761 		CLR_BIC(BIC_GFX_J, &bic_enabled);
7762 	}
7763 
7764 	if (!valid_rapl_msrs || no_msr)
7765 		return;
7766 
7767 	if (!(valid_rapl_msrs & RAPL_PKG_PERF_STATUS))
7768 		CLR_BIC(BIC_PKG__, &bic_enabled);
7769 	if (!(valid_rapl_msrs & RAPL_DRAM_PERF_STATUS))
7770 		CLR_BIC(BIC_RAM__, &bic_enabled);
7771 
7772 	/* units on package 0, verify later other packages match */
7773 	if (get_msr(master_cpu, MSR_RAPL_POWER_UNIT, &msr))
7774 		return;
7775 
7776 	rapl_power_units = 1.0 / (1 << (msr & 0xF));
7777 	if (platform->has_rapl_divisor)
7778 		rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
7779 	else
7780 		rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
7781 
7782 	if (platform->has_fixed_rapl_unit)
7783 		rapl_dram_energy_units = (15.3 / 1000000);
7784 	else
7785 		rapl_dram_energy_units = rapl_energy_units;
7786 
7787 	if (platform->has_fixed_rapl_psys_unit)
7788 		rapl_psys_energy_units = 1.0;
7789 	else
7790 		rapl_psys_energy_units = rapl_energy_units;
7791 
7792 	time_unit = msr >> 16 & 0xF;
7793 	if (time_unit == 0)
7794 		time_unit = 0xA;
7795 
7796 	rapl_time_units = 1.0 / (1 << (time_unit));
7797 
7798 	tdp = get_tdp_intel();
7799 
7800 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
7801 	if (!quiet)
7802 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
7803 }
7804 
rapl_probe_amd(void)7805 void rapl_probe_amd(void)
7806 {
7807 	unsigned long long msr;
7808 	double tdp;
7809 
7810 	if (rapl_joules) {
7811 		CLR_BIC(BIC_SysWatt, &bic_enabled);
7812 		CLR_BIC(BIC_CorWatt, &bic_enabled);
7813 	} else {
7814 		CLR_BIC(BIC_Pkg_J, &bic_enabled);
7815 		CLR_BIC(BIC_Cor_J, &bic_enabled);
7816 	}
7817 
7818 	if (!valid_rapl_msrs || no_msr)
7819 		return;
7820 
7821 	if (get_msr(master_cpu, MSR_RAPL_PWR_UNIT, &msr))
7822 		return;
7823 
7824 	rapl_time_units = ldexp(1.0, -(msr >> 16 & 0xf));
7825 	rapl_energy_units = ldexp(1.0, -(msr >> 8 & 0x1f));
7826 	rapl_power_units = ldexp(1.0, -(msr & 0xf));
7827 
7828 	tdp = get_tdp_amd();
7829 
7830 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
7831 	if (!quiet)
7832 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
7833 }
7834 
print_power_limit_msr(int cpu,unsigned long long msr,char * label)7835 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
7836 {
7837 	fprintf(outf, "cpu%d: %s: %sabled (%0.3f Watts, %f sec, clamp %sabled)\n",
7838 		cpu, label,
7839 		((msr >> 15) & 1) ? "EN" : "DIS",
7840 		((msr >> 0) & 0x7FFF) * rapl_power_units,
7841 		(1.0 + (((msr >> 22) & 0x3) / 4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units, (((msr >> 16) & 1) ? "EN" : "DIS"));
7842 
7843 	return;
7844 }
7845 
fread_int(char * path,int * val)7846 static int fread_int(char *path, int *val)
7847 {
7848 	FILE *filep;
7849 	int ret;
7850 
7851 	filep = fopen(path, "r");
7852 	if (!filep)
7853 		return -1;
7854 
7855 	ret = fscanf(filep, "%d", val);
7856 	fclose(filep);
7857 	return ret;
7858 }
7859 
fread_ull(char * path,unsigned long long * val)7860 static int fread_ull(char *path, unsigned long long *val)
7861 {
7862 	FILE *filep;
7863 	int ret;
7864 
7865 	filep = fopen(path, "r");
7866 	if (!filep)
7867 		return -1;
7868 
7869 	ret = fscanf(filep, "%llu", val);
7870 	fclose(filep);
7871 	return ret;
7872 }
7873 
fread_str(char * path,char * buf,int size)7874 static int fread_str(char *path, char *buf, int size)
7875 {
7876 	FILE *filep;
7877 	int ret;
7878 	char *cp;
7879 
7880 	filep = fopen(path, "r");
7881 	if (!filep)
7882 		return -1;
7883 
7884 	ret = fread(buf, 1, size, filep);
7885 	fclose(filep);
7886 
7887 	/* replace '\n' with '\0' */
7888 	cp = strchr(buf, '\n');
7889 	if (cp != NULL)
7890 		*cp = '\0';
7891 
7892 	return ret;
7893 }
7894 
7895 #define PATH_RAPL_SYSFS	"/sys/class/powercap"
7896 
dump_one_domain(char * domain_path)7897 static int dump_one_domain(char *domain_path)
7898 {
7899 	char path[PATH_MAX];
7900 	char str[PATH_MAX];
7901 	unsigned long long val;
7902 	int constraint;
7903 	int enable;
7904 	int ret;
7905 
7906 	snprintf(path, PATH_MAX, "%s/name", domain_path);
7907 	ret = fread_str(path, str, PATH_MAX);
7908 	if (ret <= 0)
7909 		return -1;
7910 
7911 	fprintf(outf, "%s: %s", domain_path + strlen(PATH_RAPL_SYSFS) + 1, str);
7912 
7913 	snprintf(path, PATH_MAX, "%s/enabled", domain_path);
7914 	ret = fread_int(path, &enable);
7915 	if (ret <= 0)
7916 		return -1;
7917 
7918 	if (!enable) {
7919 		fputs(" disabled\n", outf);
7920 		return 0;
7921 	}
7922 
7923 	for (constraint = 0;; constraint++) {
7924 		snprintf(path, PATH_MAX, "%s/constraint_%d_time_window_us", domain_path, constraint);
7925 		ret = fread_ull(path, &val);
7926 		if (ret <= 0)
7927 			break;
7928 
7929 		if (val > 1000000)
7930 			fprintf(outf, " %0.1fs", (double)val / 1000000);
7931 		else if (val > 1000)
7932 			fprintf(outf, " %0.1fms", (double)val / 1000);
7933 		else
7934 			fprintf(outf, " %0.1fus", (double)val);
7935 
7936 		snprintf(path, PATH_MAX, "%s/constraint_%d_power_limit_uw", domain_path, constraint);
7937 		ret = fread_ull(path, &val);
7938 		if (ret > 0 && val)
7939 			fprintf(outf, ":%lluW", val / 1000000);
7940 
7941 		snprintf(path, PATH_MAX, "%s/constraint_%d_max_power_uw", domain_path, constraint);
7942 		ret = fread_ull(path, &val);
7943 		if (ret > 0 && val)
7944 			fprintf(outf, ",max:%lluW", val / 1000000);
7945 	}
7946 	fputc('\n', outf);
7947 
7948 	return 0;
7949 }
7950 
print_rapl_sysfs(void)7951 static int print_rapl_sysfs(void)
7952 {
7953 	DIR *dir, *cdir;
7954 	struct dirent *entry, *centry;
7955 	char path[PATH_MAX];
7956 	char str[PATH_MAX];
7957 
7958 	if ((dir = opendir(PATH_RAPL_SYSFS)) == NULL) {
7959 		warn("open %s failed", PATH_RAPL_SYSFS);
7960 		return 1;
7961 	}
7962 
7963 	while ((entry = readdir(dir)) != NULL) {
7964 		if (strlen(entry->d_name) > 100)
7965 			continue;
7966 
7967 		if (strncmp(entry->d_name, "intel-rapl", strlen("intel-rapl")))
7968 			continue;
7969 
7970 		snprintf(path, PATH_MAX, "%s/%s/name", PATH_RAPL_SYSFS, entry->d_name);
7971 
7972 		/* Parse top level domains first, including package and psys */
7973 		fread_str(path, str, PATH_MAX);
7974 		if (strncmp(str, "package", strlen("package")) && strncmp(str, "psys", strlen("psys")))
7975 			continue;
7976 
7977 		snprintf(path, PATH_MAX, "%s/%s", PATH_RAPL_SYSFS, entry->d_name);
7978 		if ((cdir = opendir(path)) == NULL) {
7979 			perror("opendir() error");
7980 			return 1;
7981 		}
7982 
7983 		dump_one_domain(path);
7984 
7985 		while ((centry = readdir(cdir)) != NULL) {
7986 			if (strncmp(centry->d_name, "intel-rapl", strlen("intel-rapl")))
7987 				continue;
7988 			snprintf(path, PATH_MAX, "%s/%s/%s", PATH_RAPL_SYSFS, entry->d_name, centry->d_name);
7989 			dump_one_domain(path);
7990 		}
7991 		closedir(cdir);
7992 	}
7993 
7994 	closedir(dir);
7995 	return 0;
7996 }
7997 
print_rapl(PER_THREAD_PARAMS)7998 int print_rapl(PER_THREAD_PARAMS)
7999 {
8000 	unsigned long long msr;
8001 	const char *msr_name;
8002 	int cpu;
8003 
8004 	UNUSED(c);
8005 	UNUSED(p);
8006 
8007 	if (!valid_rapl_msrs)
8008 		return 0;
8009 
8010 	/* RAPL counters are per package, so print only for 1st thread/package */
8011 	if (!is_cpu_first_thread_in_package(t, c, p))
8012 		return 0;
8013 
8014 	cpu = t->cpu_id;
8015 	if (cpu_migrate(cpu)) {
8016 		fprintf(outf, "print_rapl: Could not migrate to CPU %d\n", cpu);
8017 		return -1;
8018 	}
8019 
8020 	if (valid_rapl_msrs & RAPL_AMD_F17H) {
8021 		msr_name = "MSR_RAPL_PWR_UNIT";
8022 		if (get_msr(cpu, MSR_RAPL_PWR_UNIT, &msr))
8023 			return -1;
8024 	} else {
8025 		msr_name = "MSR_RAPL_POWER_UNIT";
8026 		if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
8027 			return -1;
8028 	}
8029 
8030 	fprintf(outf, "cpu%d: %s: 0x%08llx (%f Watts, %f Joules, %f sec.)\n", cpu, msr_name, msr, rapl_power_units, rapl_energy_units, rapl_time_units);
8031 
8032 	if (valid_rapl_msrs & RAPL_PKG_POWER_INFO) {
8033 
8034 		if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
8035 			return -5;
8036 
8037 		fprintf(outf, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
8038 			cpu, msr,
8039 			((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
8040 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
8041 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units, ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
8042 
8043 	}
8044 	if (valid_rapl_msrs & RAPL_PKG) {
8045 
8046 		if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
8047 			return -9;
8048 
8049 		fprintf(outf, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n", cpu, msr, (msr >> 63) & 1 ? "" : "UN");
8050 
8051 		print_power_limit_msr(cpu, msr, "PKG Limit #1");
8052 		fprintf(outf, "cpu%d: PKG Limit #2: %sabled (%0.3f Watts, %f* sec, clamp %sabled)\n",
8053 			cpu,
8054 			((msr >> 47) & 1) ? "EN" : "DIS",
8055 			((msr >> 32) & 0x7FFF) * rapl_power_units,
8056 			(1.0 + (((msr >> 54) & 0x3) / 4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units, ((msr >> 48) & 1) ? "EN" : "DIS");
8057 
8058 		if (get_msr(cpu, MSR_VR_CURRENT_CONFIG, &msr))
8059 			return -9;
8060 
8061 		fprintf(outf, "cpu%d: MSR_VR_CURRENT_CONFIG: 0x%08llx\n", cpu, msr);
8062 		fprintf(outf, "cpu%d: PKG Limit #4: %f Watts (%slocked)\n", cpu, ((msr >> 0) & 0x1FFF) * rapl_power_units, (msr >> 31) & 1 ? "" : "UN");
8063 	}
8064 
8065 	if (valid_rapl_msrs & RAPL_DRAM_POWER_INFO) {
8066 		if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
8067 			return -6;
8068 
8069 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
8070 			cpu, msr,
8071 			((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
8072 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
8073 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units, ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
8074 	}
8075 	if (valid_rapl_msrs & RAPL_DRAM) {
8076 		if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
8077 			return -9;
8078 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n", cpu, msr, (msr >> 31) & 1 ? "" : "UN");
8079 
8080 		print_power_limit_msr(cpu, msr, "DRAM Limit");
8081 	}
8082 	if (valid_rapl_msrs & RAPL_CORE_POLICY) {
8083 		if (get_msr(cpu, MSR_PP0_POLICY, &msr))
8084 			return -7;
8085 
8086 		fprintf(outf, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
8087 	}
8088 	if (valid_rapl_msrs & RAPL_CORE_POWER_LIMIT) {
8089 		if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
8090 			return -9;
8091 		fprintf(outf, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n", cpu, msr, (msr >> 31) & 1 ? "" : "UN");
8092 		print_power_limit_msr(cpu, msr, "Cores Limit");
8093 	}
8094 	if (valid_rapl_msrs & RAPL_GFX) {
8095 		if (get_msr(cpu, MSR_PP1_POLICY, &msr))
8096 			return -8;
8097 
8098 		fprintf(outf, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
8099 
8100 		if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
8101 			return -9;
8102 		fprintf(outf, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n", cpu, msr, (msr >> 31) & 1 ? "" : "UN");
8103 		print_power_limit_msr(cpu, msr, "GFX Limit");
8104 	}
8105 	return 0;
8106 }
8107 
8108 /*
8109  * probe_rapl_msrs
8110  *
8111  * initialize global valid_rapl_msrs to platform->plat_rapl_msrs
8112  * only if PKG_ENERGY counter is enumerated and reads non-zero
8113  */
probe_rapl_msrs(void)8114 void probe_rapl_msrs(void)
8115 {
8116 	int ret;
8117 	off_t offset;
8118 	unsigned long long msr_value;
8119 
8120 	if (no_msr)
8121 		return;
8122 
8123 	if ((platform->plat_rapl_msrs & (RAPL_PKG | RAPL_AMD_F17H)) == 0)
8124 		return;
8125 
8126 	offset = idx_to_offset(IDX_PKG_ENERGY);
8127 	if (offset < 0)
8128 		return;
8129 
8130 	ret = get_msr(master_cpu, offset, &msr_value);
8131 	if (ret) {
8132 		if (debug)
8133 			fprintf(outf, "Can not read RAPL_PKG_ENERGY MSR(0x%llx)\n", (unsigned long long)offset);
8134 		return;
8135 	}
8136 	if (msr_value == 0) {
8137 		if (debug)
8138 			fprintf(outf, "RAPL_PKG_ENERGY MSR(0x%llx) == ZERO: disabling all RAPL MSRs\n", (unsigned long long)offset);
8139 		return;
8140 	}
8141 
8142 	valid_rapl_msrs = platform->plat_rapl_msrs;	/* success */
8143 }
8144 
8145 /*
8146  * probe_rapl()
8147  *
8148  * sets rapl_power_units, rapl_energy_units, rapl_time_units
8149  */
probe_rapl(void)8150 void probe_rapl(void)
8151 {
8152 	probe_rapl_msrs();
8153 
8154 	if (genuine_intel)
8155 		rapl_probe_intel();
8156 	if (authentic_amd || hygon_genuine)
8157 		rapl_probe_amd();
8158 
8159 	if (quiet)
8160 		return;
8161 
8162 	print_rapl_sysfs();
8163 
8164 	if (!valid_rapl_msrs || no_msr)
8165 		return;
8166 
8167 	for_all_cpus(print_rapl, ODD_COUNTERS);
8168 }
8169 
8170 /*
8171  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
8172  * the Thermal Control Circuit (TCC) activates.
8173  * This is usually equal to tjMax.
8174  *
8175  * Older processors do not have this MSR, so there we guess,
8176  * but also allow cmdline over-ride with -T.
8177  *
8178  * Several MSR temperature values are in units of degrees-C
8179  * below this value, including the Digital Thermal Sensor (DTS),
8180  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
8181  */
set_temperature_target(PER_THREAD_PARAMS)8182 int set_temperature_target(PER_THREAD_PARAMS)
8183 {
8184 	unsigned long long msr;
8185 	unsigned int tcc_default, tcc_offset;
8186 	int cpu;
8187 
8188 	UNUSED(c);
8189 	UNUSED(p);
8190 
8191 	/* tj_max is used only for dts or ptm */
8192 	if (!(do_dts || do_ptm))
8193 		return 0;
8194 
8195 	/* this is a per-package concept */
8196 	if (!is_cpu_first_thread_in_package(t, c, p))
8197 		return 0;
8198 
8199 	cpu = t->cpu_id;
8200 	if (cpu_migrate(cpu)) {
8201 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
8202 		return -1;
8203 	}
8204 
8205 	if (tj_max_override != 0) {
8206 		tj_max = tj_max_override;
8207 		fprintf(outf, "cpu%d: Using cmdline TCC Target (%d C)\n", cpu, tj_max);
8208 		return 0;
8209 	}
8210 
8211 	/* Temperature Target MSR is Nehalem and newer only */
8212 	if (!platform->has_nhm_msrs || no_msr)
8213 		goto guess;
8214 
8215 	if (get_msr(master_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
8216 		goto guess;
8217 
8218 	tcc_default = (msr >> 16) & 0xFF;
8219 
8220 	if (!quiet) {
8221 		int bits = platform->tcc_offset_bits;
8222 		unsigned long long enabled = 0;
8223 
8224 		if (bits && !get_msr(master_cpu, MSR_PLATFORM_INFO, &enabled))
8225 			enabled = (enabled >> 30) & 1;
8226 
8227 		if (bits && enabled) {
8228 			tcc_offset = (msr >> 24) & GENMASK(bits - 1, 0);
8229 			fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C) (%d default - %d offset)\n",
8230 				cpu, msr, tcc_default - tcc_offset, tcc_default, tcc_offset);
8231 		} else {
8232 			fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n", cpu, msr, tcc_default);
8233 		}
8234 	}
8235 
8236 	if (!tcc_default)
8237 		goto guess;
8238 
8239 	tj_max = tcc_default;
8240 
8241 	return 0;
8242 
8243 guess:
8244 	tj_max = TJMAX_DEFAULT;
8245 	fprintf(outf, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n", cpu, tj_max);
8246 
8247 	return 0;
8248 }
8249 
print_thermal(PER_THREAD_PARAMS)8250 int print_thermal(PER_THREAD_PARAMS)
8251 {
8252 	unsigned long long msr;
8253 	unsigned int dts, dts2;
8254 	int cpu;
8255 
8256 	UNUSED(c);
8257 	UNUSED(p);
8258 
8259 	if (no_msr)
8260 		return 0;
8261 
8262 	if (!(do_dts || do_ptm))
8263 		return 0;
8264 
8265 	cpu = t->cpu_id;
8266 
8267 	/* DTS is per-core, no need to print for each thread */
8268 	if (!is_cpu_first_thread_in_core(t, c))
8269 		return 0;
8270 
8271 	if (cpu_migrate(cpu)) {
8272 		fprintf(outf, "print_thermal: Could not migrate to CPU %d\n", cpu);
8273 		return -1;
8274 	}
8275 
8276 	if (do_ptm && is_cpu_first_core_in_package(t, p)) {
8277 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
8278 			return 0;
8279 
8280 		dts = (msr >> 16) & 0x7F;
8281 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n", cpu, msr, tj_max - dts);
8282 
8283 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
8284 			return 0;
8285 
8286 		dts = (msr >> 16) & 0x7F;
8287 		dts2 = (msr >> 8) & 0x7F;
8288 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n", cpu, msr, tj_max - dts, tj_max - dts2);
8289 	}
8290 
8291 	if (do_dts && debug) {
8292 		unsigned int resolution;
8293 
8294 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
8295 			return 0;
8296 
8297 		dts = (msr >> 16) & 0x7F;
8298 		resolution = (msr >> 27) & 0xF;
8299 		fprintf(outf, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n", cpu, msr, tj_max - dts, resolution);
8300 
8301 		if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
8302 			return 0;
8303 
8304 		dts = (msr >> 16) & 0x7F;
8305 		dts2 = (msr >> 8) & 0x7F;
8306 		fprintf(outf, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n", cpu, msr, tj_max - dts, tj_max - dts2);
8307 	}
8308 
8309 	return 0;
8310 }
8311 
probe_thermal(void)8312 void probe_thermal(void)
8313 {
8314 	if (!access("/sys/devices/system/cpu/cpu0/thermal_throttle/core_throttle_count", R_OK))
8315 		BIC_PRESENT(BIC_CORE_THROT_CNT);
8316 	else
8317 		BIC_NOT_PRESENT(BIC_CORE_THROT_CNT);
8318 
8319 	for_all_cpus(set_temperature_target, ODD_COUNTERS);
8320 
8321 	if (quiet)
8322 		return;
8323 
8324 	for_all_cpus(print_thermal, ODD_COUNTERS);
8325 }
8326 
get_cpu_type(PER_THREAD_PARAMS)8327 int get_cpu_type(PER_THREAD_PARAMS)
8328 {
8329 	unsigned int eax, ebx, ecx, edx;
8330 
8331 	UNUSED(c);
8332 	UNUSED(p);
8333 
8334 	if (!genuine_intel)
8335 		return 0;
8336 
8337 	if (cpu_migrate(t->cpu_id)) {
8338 		fprintf(outf, "Could not migrate to CPU %d\n", t->cpu_id);
8339 		return -1;
8340 	}
8341 
8342 	if (max_level < 0x1a)
8343 		return 0;
8344 
8345 	__cpuid(0x1a, eax, ebx, ecx, edx);
8346 	eax = (eax >> 24) & 0xFF;
8347 	if (eax == 0x20)
8348 		t->is_atom = true;
8349 	return 0;
8350 }
8351 
decode_feature_control_msr(void)8352 void decode_feature_control_msr(void)
8353 {
8354 	unsigned long long msr;
8355 
8356 	if (no_msr)
8357 		return;
8358 
8359 	if (quiet)
8360 		return;
8361 
8362 	if (!get_msr(master_cpu, MSR_IA32_FEAT_CTL, &msr))
8363 		fprintf(outf, "cpu%d: MSR_IA32_FEATURE_CONTROL: 0x%08llx (%sLocked %s)\n",
8364 			master_cpu, msr, msr & FEAT_CTL_LOCKED ? "" : "UN-", msr & (1 << 18) ? "SGX" : "");
8365 }
8366 
decode_misc_enable_msr(void)8367 void decode_misc_enable_msr(void)
8368 {
8369 	unsigned long long msr;
8370 
8371 	if (no_msr)
8372 		return;
8373 
8374 	if (!genuine_intel)
8375 		return;
8376 
8377 	if (!get_msr(master_cpu, MSR_IA32_MISC_ENABLE, &msr))
8378 		fprintf(outf, "cpu%d: MSR_IA32_MISC_ENABLE: 0x%08llx (%sTCC %sEIST %sMWAIT %sPREFETCH %sTURBO)\n",
8379 			master_cpu, msr,
8380 			msr & MSR_IA32_MISC_ENABLE_TM1 ? "" : "No-",
8381 			msr & MSR_IA32_MISC_ENABLE_ENHANCED_SPEEDSTEP ? "" : "No-",
8382 			msr & MSR_IA32_MISC_ENABLE_MWAIT ? "" : "No-",
8383 			msr & MSR_IA32_MISC_ENABLE_PREFETCH_DISABLE ? "No-" : "", msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ? "No-" : "");
8384 }
8385 
decode_misc_feature_control(void)8386 void decode_misc_feature_control(void)
8387 {
8388 	unsigned long long msr;
8389 
8390 	if (no_msr)
8391 		return;
8392 
8393 	if (!platform->has_msr_misc_feature_control)
8394 		return;
8395 
8396 	if (!get_msr(master_cpu, MSR_MISC_FEATURE_CONTROL, &msr))
8397 		fprintf(outf,
8398 			"cpu%d: MSR_MISC_FEATURE_CONTROL: 0x%08llx (%sL2-Prefetch %sL2-Prefetch-pair %sL1-Prefetch %sL1-IP-Prefetch)\n",
8399 			master_cpu, msr, msr & (0 << 0) ? "No-" : "", msr & (1 << 0) ? "No-" : "", msr & (2 << 0) ? "No-" : "", msr & (3 << 0) ? "No-" : "");
8400 }
8401 
8402 /*
8403  * Decode MSR_MISC_PWR_MGMT
8404  *
8405  * Decode the bits according to the Nehalem documentation
8406  * bit[0] seems to continue to have same meaning going forward
8407  * bit[1] less so...
8408  */
decode_misc_pwr_mgmt_msr(void)8409 void decode_misc_pwr_mgmt_msr(void)
8410 {
8411 	unsigned long long msr;
8412 
8413 	if (no_msr)
8414 		return;
8415 
8416 	if (!platform->has_msr_misc_pwr_mgmt)
8417 		return;
8418 
8419 	if (!get_msr(master_cpu, MSR_MISC_PWR_MGMT, &msr))
8420 		fprintf(outf, "cpu%d: MSR_MISC_PWR_MGMT: 0x%08llx (%sable-EIST_Coordination %sable-EPB %sable-OOB)\n",
8421 			master_cpu, msr, msr & (1 << 0) ? "DIS" : "EN", msr & (1 << 1) ? "EN" : "DIS", msr & (1 << 8) ? "EN" : "DIS");
8422 }
8423 
8424 /*
8425  * Decode MSR_CC6_DEMOTION_POLICY_CONFIG, MSR_MC6_DEMOTION_POLICY_CONFIG
8426  *
8427  * This MSRs are present on Silvermont processors,
8428  * Intel Atom processor E3000 series (Baytrail), and friends.
8429  */
decode_c6_demotion_policy_msr(void)8430 void decode_c6_demotion_policy_msr(void)
8431 {
8432 	unsigned long long msr;
8433 
8434 	if (no_msr)
8435 		return;
8436 
8437 	if (!platform->has_msr_c6_demotion_policy_config)
8438 		return;
8439 
8440 	if (!get_msr(master_cpu, MSR_CC6_DEMOTION_POLICY_CONFIG, &msr))
8441 		fprintf(outf, "cpu%d: MSR_CC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-CC6-Demotion)\n", master_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
8442 
8443 	if (!get_msr(master_cpu, MSR_MC6_DEMOTION_POLICY_CONFIG, &msr))
8444 		fprintf(outf, "cpu%d: MSR_MC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-MC6-Demotion)\n", master_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
8445 }
8446 
print_dev_latency(void)8447 void print_dev_latency(void)
8448 {
8449 	char *path = "/dev/cpu_dma_latency";
8450 	int fd;
8451 	int value;
8452 	int retval;
8453 
8454 	fd = open(path, O_RDONLY);
8455 	if (fd < 0) {
8456 		if (debug)
8457 			warnx("Read %s failed", path);
8458 		return;
8459 	}
8460 
8461 	retval = read(fd, (void *)&value, sizeof(int));
8462 	if (retval != sizeof(int)) {
8463 		warn("read failed %s", path);
8464 		close(fd);
8465 		return;
8466 	}
8467 	fprintf(outf, "/dev/cpu_dma_latency: %d usec (%s)\n", value, value == 2000000000 ? "default" : "constrained");
8468 
8469 	close(fd);
8470 }
8471 
has_perf_instr_count_access(void)8472 static int has_perf_instr_count_access(void)
8473 {
8474 	int fd;
8475 
8476 	if (no_perf)
8477 		return 0;
8478 
8479 	fd = open_perf_counter(master_cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, -1, 0);
8480 	if (fd != -1)
8481 		close(fd);
8482 
8483 	if (fd == -1)
8484 		warnx("Failed to access %s. Some of the counters may not be available\n"
8485 		      "\tRun as root to enable them or use %s to disable the access explicitly", "perf instructions retired counter",
8486 		      "'--hide IPC' or '--no-perf'");
8487 
8488 	return (fd != -1);
8489 }
8490 
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_)8491 int 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_)
8492 {
8493 	int ret = -1;
8494 
8495 	if (no_perf)
8496 		return -1;
8497 
8498 	if (!cai->perf_name)
8499 		return -1;
8500 
8501 	const double scale = read_perf_scale(cai->perf_subsys, cai->perf_name);
8502 
8503 	if (scale == 0.0)
8504 		goto end;
8505 
8506 	const enum rapl_unit unit = read_perf_rapl_unit(cai->perf_subsys, cai->perf_name);
8507 
8508 	if (unit == RAPL_UNIT_INVALID)
8509 		goto end;
8510 
8511 	const unsigned int rapl_type = read_perf_type(cai->perf_subsys);
8512 	const unsigned int rapl_energy_pkg_config = read_perf_config(cai->perf_subsys, cai->perf_name);
8513 
8514 	ret = open_perf_counter(cpu, rapl_type, rapl_energy_pkg_config, rci->fd_perf, PERF_FORMAT_GROUP);
8515 	if (ret == -1)
8516 		goto end;
8517 
8518 	/* If it's the first counter opened, make it a group descriptor */
8519 	if (rci->fd_perf == -1)
8520 		rci->fd_perf = ret;
8521 
8522 	*scale_ = scale;
8523 	*unit_ = unit;
8524 
8525 end:
8526 	if (debug >= 2)
8527 		fprintf(stderr, "%s: %d (cpu: %d)\n", __func__, ret, cpu);
8528 
8529 	return ret;
8530 }
8531 
8532 char cpuset_buf[1024];
initialize_cpu_set_from_sysfs(cpu_set_t * cpu_set,char * sysfs_path,char * sysfs_file)8533 int initialize_cpu_set_from_sysfs(cpu_set_t *cpu_set, char *sysfs_path, char *sysfs_file)
8534 {
8535 	FILE *fp;
8536 	char path[128];
8537 
8538 	if (snprintf(path, 128, "%s/%s", sysfs_path, sysfs_file) > 128)
8539 		err(-1, "%s %s", sysfs_path, sysfs_file);
8540 
8541 	fp = fopen(path, "r");
8542 	if (!fp) {
8543 		warn("open %s", path);
8544 		return -1;
8545 	}
8546 	if (fread(cpuset_buf, sizeof(char), 1024, fp) == 0) {
8547 		warn("read %s", sysfs_path);
8548 		goto err;
8549 	}
8550 	if (parse_cpu_str(cpuset_buf, cpu_set, cpu_possible_setsize)) {
8551 		warnx("%s: cpu str malformat %s\n", sysfs_path, cpu_effective_str);
8552 		goto err;
8553 	}
8554 	return 0;
8555 
8556 err:
8557 	fclose(fp);
8558 	return -1;
8559 }
8560 
print_cpu_set(char * s,cpu_set_t * set)8561 void print_cpu_set(char *s, cpu_set_t *set)
8562 {
8563 	int i;
8564 
8565 	assert(MAX_BIC < CPU_SETSIZE);
8566 
8567 	printf("%s:", s);
8568 
8569 	for (i = 0; i <= topo.max_cpu_num; ++i)
8570 		if (CPU_ISSET(i, set))
8571 			printf(" %d", i);
8572 	putchar('\n');
8573 }
8574 
linux_perf_init_hybrid_cpus(void)8575 void linux_perf_init_hybrid_cpus(void)
8576 {
8577 	char *perf_cpu_pcore_path = "/sys/devices/cpu_core";
8578 	char *perf_cpu_ecore_path = "/sys/devices/cpu_atom";
8579 	char *perf_cpu_lcore_path = "/sys/devices/cpu_lowpower";
8580 	char path[128];
8581 
8582 	if (!access(perf_cpu_pcore_path, F_OK)) {
8583 		perf_pcore_set = CPU_ALLOC((topo.max_cpu_num + 1));
8584 		if (perf_pcore_set == NULL)
8585 			err(3, "CPU_ALLOC");
8586 		CPU_ZERO_S(cpu_possible_setsize, perf_pcore_set);
8587 		initialize_cpu_set_from_sysfs(perf_pcore_set, perf_cpu_pcore_path, "cpus");
8588 		if (debug)
8589 			print_cpu_set("perf pcores", perf_pcore_set);
8590 		sprintf(path, "%s/%s", perf_cpu_pcore_path, "type");
8591 		perf_pmu_types.pcore = snapshot_sysfs_counter(path);
8592 	}
8593 
8594 	if (!access(perf_cpu_ecore_path, F_OK)) {
8595 		perf_ecore_set = CPU_ALLOC((topo.max_cpu_num + 1));
8596 		if (perf_ecore_set == NULL)
8597 			err(3, "CPU_ALLOC");
8598 		CPU_ZERO_S(cpu_possible_setsize, perf_ecore_set);
8599 		initialize_cpu_set_from_sysfs(perf_ecore_set, perf_cpu_ecore_path, "cpus");
8600 		if (debug)
8601 			print_cpu_set("perf ecores", perf_ecore_set);
8602 		sprintf(path, "%s/%s", perf_cpu_ecore_path, "type");
8603 		perf_pmu_types.ecore = snapshot_sysfs_counter(path);
8604 	}
8605 
8606 	if (!access(perf_cpu_lcore_path, F_OK)) {
8607 		perf_lcore_set = CPU_ALLOC((topo.max_cpu_num + 1));
8608 		if (perf_lcore_set == NULL)
8609 			err(3, "CPU_ALLOC");
8610 		CPU_ZERO_S(cpu_possible_setsize, perf_lcore_set);
8611 		initialize_cpu_set_from_sysfs(perf_lcore_set, perf_cpu_lcore_path, "cpus");
8612 		if (debug)
8613 			print_cpu_set("perf lcores", perf_lcore_set);
8614 		sprintf(path, "%s/%s", perf_cpu_lcore_path, "type");
8615 		perf_pmu_types.lcore = snapshot_sysfs_counter(path);
8616 	}
8617 }
8618 
8619 /*
8620  * Linux-perf related initialization
8621  */
linux_perf_init(void)8622 void linux_perf_init(void)
8623 {
8624 	char path[128];
8625 	char *perf_cpu_path = "/sys/devices/cpu";
8626 
8627 	if (access("/proc/sys/kernel/perf_event_paranoid", F_OK))
8628 		return;
8629 
8630 	if (!access(perf_cpu_path, F_OK)) {
8631 		sprintf(path, "%s/%s", perf_cpu_path, "type");
8632 		perf_pmu_types.uniform = snapshot_sysfs_counter(path);
8633 	} else {
8634 		linux_perf_init_hybrid_cpus();
8635 	}
8636 
8637 	if (BIC_IS_ENABLED(BIC_IPC) && cpuid_has_aperf_mperf) {
8638 		fd_instr_count_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8639 		if (fd_instr_count_percpu == NULL)
8640 			err(-1, "calloc fd_instr_count_percpu");
8641 	}
8642 	if (BIC_IS_ENABLED(BIC_LLC_MRPS) || BIC_IS_ENABLED(BIC_LLC_HIT)) {
8643 		fd_llc_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8644 		if (fd_llc_percpu == NULL)
8645 			err(-1, "calloc fd_llc_percpu");
8646 	}
8647 	if (BIC_IS_ENABLED(BIC_L2_MRPS) || BIC_IS_ENABLED(BIC_L2_HIT)) {
8648 		fd_l2_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8649 		if (fd_l2_percpu == NULL)
8650 			err(-1, "calloc fd_l2_percpu");
8651 	}
8652 }
8653 
rapl_perf_init(void)8654 void rapl_perf_init(void)
8655 {
8656 	const unsigned int num_domains = get_rapl_num_domains();
8657 	bool *domain_visited = calloc(num_domains, sizeof(bool));
8658 
8659 	rapl_counter_info_perdomain = calloc(num_domains, sizeof(*rapl_counter_info_perdomain));
8660 	if (rapl_counter_info_perdomain == NULL)
8661 		err(-1, "calloc rapl_counter_info_percpu");
8662 	rapl_counter_info_perdomain_size = num_domains;
8663 
8664 	/*
8665 	 * Initialize rapl_counter_info_percpu
8666 	 */
8667 	for (unsigned int domain_id = 0; domain_id < num_domains; ++domain_id) {
8668 		struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[domain_id];
8669 
8670 		rci->fd_perf = -1;
8671 		for (size_t i = 0; i < NUM_RAPL_COUNTERS; ++i) {
8672 			rci->data[i] = 0;
8673 			rci->source[i] = COUNTER_SOURCE_NONE;
8674 		}
8675 	}
8676 
8677 	/*
8678 	 * Open/probe the counters
8679 	 * If can't get it via perf, fallback to MSR
8680 	 */
8681 	for (size_t i = 0; i < ARRAY_SIZE(rapl_counter_arch_infos); ++i) {
8682 
8683 		const struct rapl_counter_arch_info *const cai = &rapl_counter_arch_infos[i];
8684 		bool has_counter = 0;
8685 		double scale;
8686 		enum rapl_unit unit;
8687 		unsigned int next_domain;
8688 
8689 		if (!BIC_IS_ENABLED(cai->bic_number))
8690 			continue;
8691 
8692 		memset(domain_visited, 0, num_domains * sizeof(*domain_visited));
8693 
8694 		for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
8695 
8696 			if (cpu_is_not_allowed(cpu))
8697 				continue;
8698 
8699 			/* Skip already seen and handled RAPL domains */
8700 			next_domain = get_rapl_domain_id(cpu);
8701 
8702 			assert(next_domain < num_domains);
8703 
8704 			if (domain_visited[next_domain])
8705 				continue;
8706 
8707 			domain_visited[next_domain] = 1;
8708 
8709 			if ((cai->flags & RAPL_COUNTER_FLAG_PLATFORM_COUNTER) && (cpu != master_cpu))
8710 				continue;
8711 
8712 			struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[next_domain];
8713 
8714 			/*
8715 			 * rapl_counter_arch_infos[] can have multiple entries describing the same
8716 			 * counter, due to the difference from different platforms/Vendors.
8717 			 * E.g. rapl_counter_arch_infos[0] and rapl_counter_arch_infos[1] share the
8718 			 * same perf_subsys and perf_name, but with different MSR address.
8719 			 * rapl_counter_arch_infos[0] is for Intel and rapl_counter_arch_infos[1]
8720 			 * is for AMD.
8721 			 * In this case, it is possible that multiple rapl_counter_arch_infos[]
8722 			 * entries are probed just because their perf/msr is duplicate and valid.
8723 			 *
8724 			 * Thus need a check to avoid re-probe the same counters.
8725 			 */
8726 			if (rci->source[cai->rci_index] != COUNTER_SOURCE_NONE)
8727 				break;
8728 
8729 			/* Use perf API for this counter */
8730 			if (add_rapl_perf_counter(cpu, rci, cai, &scale, &unit) != -1) {
8731 				rci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
8732 				rci->scale[cai->rci_index] = scale * cai->compat_scale;
8733 				rci->unit[cai->rci_index] = unit;
8734 				rci->flags[cai->rci_index] = cai->flags;
8735 
8736 				/* Use MSR for this counter */
8737 			} else if (add_rapl_msr_counter(cpu, cai) >= 0) {
8738 				rci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
8739 				rci->msr[cai->rci_index] = cai->msr;
8740 				rci->msr_mask[cai->rci_index] = cai->msr_mask;
8741 				rci->msr_shift[cai->rci_index] = cai->msr_shift;
8742 				rci->unit[cai->rci_index] = RAPL_UNIT_JOULES;
8743 				rci->scale[cai->rci_index] = *cai->platform_rapl_msr_scale * cai->compat_scale;
8744 				rci->flags[cai->rci_index] = cai->flags;
8745 			}
8746 
8747 			if (rci->source[cai->rci_index] != COUNTER_SOURCE_NONE)
8748 				has_counter = 1;
8749 		}
8750 
8751 		/* If any CPU has access to the counter, make it present */
8752 		if (has_counter)
8753 			BIC_PRESENT(cai->bic_number);
8754 	}
8755 
8756 	free(domain_visited);
8757 }
8758 
8759 /* Assumes msr_counter_info is populated */
has_amperf_access(void)8760 static int has_amperf_access(void)
8761 {
8762 	return cpuid_has_aperf_mperf && msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].present && msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].present;
8763 }
8764 
get_cstate_perf_group_fd(struct cstate_counter_info_t * cci,const char * group_name)8765 int *get_cstate_perf_group_fd(struct cstate_counter_info_t *cci, const char *group_name)
8766 {
8767 	if (strcmp(group_name, "cstate_core") == 0)
8768 		return &cci->fd_perf_core;
8769 
8770 	if (strcmp(group_name, "cstate_pkg") == 0)
8771 		return &cci->fd_perf_pkg;
8772 
8773 	return NULL;
8774 }
8775 
add_cstate_perf_counter(int cpu,struct cstate_counter_info_t * cci,const struct cstate_counter_arch_info * cai)8776 int add_cstate_perf_counter(int cpu, struct cstate_counter_info_t *cci, const struct cstate_counter_arch_info *cai)
8777 {
8778 	int ret = -1;
8779 
8780 	if (no_perf)
8781 		return -1;
8782 
8783 	if (!cai->perf_name)
8784 		return -1;
8785 
8786 	int *pfd_group = get_cstate_perf_group_fd(cci, cai->perf_subsys);
8787 
8788 	if (pfd_group == NULL)
8789 		goto end;
8790 
8791 	const unsigned int type = read_perf_type(cai->perf_subsys);
8792 	const unsigned int config = read_perf_config(cai->perf_subsys, cai->perf_name);
8793 
8794 	ret = open_perf_counter(cpu, type, config, *pfd_group, PERF_FORMAT_GROUP);
8795 
8796 	if (ret == -1)
8797 		goto end;
8798 
8799 	/* If it's the first counter opened, make it a group descriptor */
8800 	if (*pfd_group == -1)
8801 		*pfd_group = ret;
8802 
8803 end:
8804 	if (debug >= 2)
8805 		fprintf(stderr, "%s: %d (cpu: %d)\n", __func__, ret, cpu);
8806 
8807 	return ret;
8808 }
8809 
add_msr_perf_counter(int cpu,struct msr_counter_info_t * cci,const struct msr_counter_arch_info * cai)8810 int add_msr_perf_counter(int cpu, struct msr_counter_info_t *cci, const struct msr_counter_arch_info *cai)
8811 {
8812 	int ret = -1;
8813 
8814 	if (no_perf)
8815 		return -1;
8816 
8817 	if (!cai->perf_name)
8818 		return -1;
8819 
8820 	const unsigned int type = read_perf_type(cai->perf_subsys);
8821 	const unsigned int config = read_perf_config(cai->perf_subsys, cai->perf_name);
8822 
8823 	ret = open_perf_counter(cpu, type, config, cci->fd_perf, PERF_FORMAT_GROUP);
8824 
8825 	if (ret == -1)
8826 		goto end;
8827 
8828 	/* If it's the first counter opened, make it a group descriptor */
8829 	if (cci->fd_perf == -1)
8830 		cci->fd_perf = ret;
8831 
8832 end:
8833 	if (debug)
8834 		fprintf(stderr, "%s: %s/%s: %d (cpu: %d)\n", __func__, cai->perf_subsys, cai->perf_name, ret, cpu);
8835 
8836 	return ret;
8837 }
8838 
msr_perf_init_(void)8839 void msr_perf_init_(void)
8840 {
8841 	const int mci_num = topo.max_cpu_num + 1;
8842 
8843 	msr_counter_info = calloc(mci_num, sizeof(*msr_counter_info));
8844 	if (!msr_counter_info)
8845 		err(1, "calloc msr_counter_info");
8846 	msr_counter_info_size = mci_num;
8847 
8848 	for (int cpu = 0; cpu < mci_num; ++cpu)
8849 		msr_counter_info[cpu].fd_perf = -1;
8850 
8851 	for (int cidx = 0; cidx < NUM_MSR_COUNTERS; ++cidx) {
8852 
8853 		struct msr_counter_arch_info *cai = &msr_counter_arch_infos[cidx];
8854 
8855 		cai->present = false;
8856 
8857 		for (int cpu = 0; cpu < mci_num; ++cpu) {
8858 
8859 			struct msr_counter_info_t *const cci = &msr_counter_info[cpu];
8860 
8861 			if (cpu_is_not_allowed(cpu))
8862 				continue;
8863 
8864 			if (cai->needed) {
8865 				/* Use perf API for this counter */
8866 				if (add_msr_perf_counter(cpu, cci, cai) != -1) {
8867 					cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
8868 					cai->present = true;
8869 
8870 					/* User MSR for this counter */
8871 				} else if (add_msr_counter(cpu, cai->msr) >= 0) {
8872 					cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
8873 					cci->msr[cai->rci_index] = cai->msr;
8874 					cci->msr_mask[cai->rci_index] = cai->msr_mask;
8875 					cai->present = true;
8876 				}
8877 			}
8878 		}
8879 	}
8880 }
8881 
8882 /* Initialize data for reading perf counters from the MSR group. */
msr_perf_init(void)8883 void msr_perf_init(void)
8884 {
8885 	bool need_amperf = false, need_smi = false;
8886 	const bool need_soft_c1 = (!platform->has_msr_core_c1_res) && (platform->supported_cstates & CC1);
8887 
8888 	need_amperf = BIC_IS_ENABLED(BIC_Avg_MHz) || BIC_IS_ENABLED(BIC_Busy) || BIC_IS_ENABLED(BIC_Bzy_MHz)
8889 	    || BIC_IS_ENABLED(BIC_IPC) || need_soft_c1;
8890 
8891 	if (BIC_IS_ENABLED(BIC_SMI))
8892 		need_smi = true;
8893 
8894 	/* Enable needed counters */
8895 	msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].needed = need_amperf;
8896 	msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].needed = need_amperf;
8897 	msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].needed = need_smi;
8898 
8899 	msr_perf_init_();
8900 
8901 	const bool has_amperf = has_amperf_access();
8902 	const bool has_smi = msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].present;
8903 
8904 	has_aperf_access = has_amperf;
8905 
8906 	if (has_amperf) {
8907 		BIC_PRESENT(BIC_Avg_MHz);
8908 		BIC_PRESENT(BIC_Busy);
8909 		BIC_PRESENT(BIC_Bzy_MHz);
8910 		BIC_PRESENT(BIC_SMI);
8911 	}
8912 
8913 	if (has_smi)
8914 		BIC_PRESENT(BIC_SMI);
8915 }
8916 
cstate_perf_init_(bool soft_c1)8917 void cstate_perf_init_(bool soft_c1)
8918 {
8919 	bool has_counter;
8920 	bool *cores_visited = NULL, *pkg_visited = NULL;
8921 	const int cores_visited_elems = topo.max_core_id + 1;
8922 	const int pkg_visited_elems = topo.max_package_id + 1;
8923 	const int cci_num = topo.max_cpu_num + 1;
8924 
8925 	ccstate_counter_info = calloc(cci_num, sizeof(*ccstate_counter_info));
8926 	if (!ccstate_counter_info)
8927 		err(1, "calloc ccstate_counter_arch_info");
8928 	ccstate_counter_info_size = cci_num;
8929 
8930 	cores_visited = calloc(cores_visited_elems, sizeof(*cores_visited));
8931 	if (!cores_visited)
8932 		err(1, "calloc cores_visited");
8933 
8934 	pkg_visited = calloc(pkg_visited_elems, sizeof(*pkg_visited));
8935 	if (!pkg_visited)
8936 		err(1, "calloc pkg_visited");
8937 
8938 	/* Initialize cstate_counter_info_percpu */
8939 	for (int cpu = 0; cpu < cci_num; ++cpu) {
8940 		ccstate_counter_info[cpu].fd_perf_core = -1;
8941 		ccstate_counter_info[cpu].fd_perf_pkg = -1;
8942 	}
8943 
8944 	for (int cidx = 0; cidx < NUM_CSTATE_COUNTERS; ++cidx) {
8945 		has_counter = false;
8946 		memset(cores_visited, 0, cores_visited_elems * sizeof(*cores_visited));
8947 		memset(pkg_visited, 0, pkg_visited_elems * sizeof(*pkg_visited));
8948 
8949 		const struct cstate_counter_arch_info *cai = &ccstate_counter_arch_infos[cidx];
8950 
8951 		for (int cpu = 0; cpu < cci_num; ++cpu) {
8952 
8953 			struct cstate_counter_info_t *const cci = &ccstate_counter_info[cpu];
8954 
8955 			if (cpu_is_not_allowed(cpu))
8956 				continue;
8957 
8958 			const int core_id = cpus[cpu].core_id;
8959 			const int pkg_id = cpus[cpu].package_id;
8960 
8961 			assert(core_id < cores_visited_elems);
8962 			assert(pkg_id < pkg_visited_elems);
8963 
8964 			const bool per_thread = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD;
8965 			const bool per_core = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_CORE;
8966 
8967 			if (!per_thread && cores_visited[core_id])
8968 				continue;
8969 
8970 			if (!per_core && pkg_visited[pkg_id])
8971 				continue;
8972 
8973 			const bool counter_needed = BIC_IS_ENABLED(cai->bic_number) || (soft_c1 && (cai->flags & CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY));
8974 			const bool counter_supported = (platform->supported_cstates & cai->feature_mask);
8975 
8976 			if (counter_needed && counter_supported) {
8977 				/* Use perf API for this counter */
8978 				if (add_cstate_perf_counter(cpu, cci, cai) != -1) {
8979 
8980 					cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
8981 
8982 					/* User MSR for this counter */
8983 				} else if (pkg_cstate_limit >= cai->pkg_cstate_limit && add_msr_counter(cpu, cai->msr) >= 0) {
8984 					cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
8985 					cci->msr[cai->rci_index] = cai->msr;
8986 				}
8987 			}
8988 
8989 			if (cci->source[cai->rci_index] != COUNTER_SOURCE_NONE) {
8990 				has_counter = true;
8991 				cores_visited[core_id] = true;
8992 				pkg_visited[pkg_id] = true;
8993 			}
8994 		}
8995 
8996 		/* If any CPU has access to the counter, make it present */
8997 		if (has_counter)
8998 			BIC_PRESENT(cai->bic_number);
8999 	}
9000 
9001 	free(cores_visited);
9002 	free(pkg_visited);
9003 }
9004 
cstate_perf_init(void)9005 void cstate_perf_init(void)
9006 {
9007 	/*
9008 	 * If we don't have a C1 residency MSR, we calculate it "in software",
9009 	 * but we need APERF, MPERF too.
9010 	 */
9011 	const bool soft_c1 = !platform->has_msr_core_c1_res && has_amperf_access()
9012 	    && platform->supported_cstates & CC1;
9013 
9014 	if (soft_c1)
9015 		BIC_PRESENT(BIC_CPU_c1);
9016 
9017 	cstate_perf_init_(soft_c1);
9018 }
9019 
probe_cstates(void)9020 void probe_cstates(void)
9021 {
9022 	probe_cst_limit();
9023 
9024 	if (platform->has_msr_module_c6_res_ms)
9025 		BIC_PRESENT(BIC_Mod_c6);
9026 
9027 	if (platform->has_ext_cst_msrs && !no_msr) {
9028 		BIC_PRESENT(BIC_Totl_c0);
9029 		BIC_PRESENT(BIC_Any_c0);
9030 		BIC_PRESENT(BIC_GFX_c0);
9031 		BIC_PRESENT(BIC_CPUGFX);
9032 	}
9033 
9034 	if (quiet)
9035 		return;
9036 
9037 	dump_power_ctl();
9038 	dump_cst_cfg();
9039 	decode_c6_demotion_policy_msr();
9040 	print_dev_latency();
9041 	dump_sysfs_cstate_config();
9042 	print_irtl();
9043 }
9044 
probe_lpi(void)9045 void probe_lpi(void)
9046 {
9047 	if (!access("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", R_OK))
9048 		BIC_PRESENT(BIC_CPU_LPI);
9049 	else
9050 		BIC_NOT_PRESENT(BIC_CPU_LPI);
9051 
9052 	if (!access(sys_lpi_file_sysfs, R_OK)) {
9053 		sys_lpi_file = sys_lpi_file_sysfs;
9054 		BIC_PRESENT(BIC_SYS_LPI);
9055 	} else if (!access(sys_lpi_file_debugfs, R_OK)) {
9056 		sys_lpi_file = sys_lpi_file_debugfs;
9057 		BIC_PRESENT(BIC_SYS_LPI);
9058 	} else {
9059 		sys_lpi_file_sysfs = NULL;
9060 		BIC_NOT_PRESENT(BIC_SYS_LPI);
9061 	}
9062 
9063 }
9064 
probe_pstates(void)9065 void probe_pstates(void)
9066 {
9067 	probe_bclk();
9068 
9069 	if (quiet)
9070 		return;
9071 
9072 	dump_platform_info();
9073 	dump_turbo_ratio_info();
9074 	dump_sysfs_pstate_config();
9075 	decode_misc_pwr_mgmt_msr();
9076 
9077 	for_all_cpus(print_hwp, ODD_COUNTERS);
9078 	for_all_cpus(print_epb, ODD_COUNTERS);
9079 	for_all_cpus(print_perf_limit, ODD_COUNTERS);
9080 }
9081 
dump_word_chars(unsigned int word)9082 void dump_word_chars(unsigned int word)
9083 {
9084 	int i;
9085 
9086 	for (i = 0; i < 4; ++i)
9087 		fprintf(outf, "%c", (word >> (i * 8)) & 0xFF);
9088 }
9089 
dump_cpuid_hypervisor(void)9090 void dump_cpuid_hypervisor(void)
9091 {
9092 	unsigned int ebx = 0;
9093 	unsigned int ecx = 0;
9094 	unsigned int edx = 0;
9095 
9096 	__cpuid(0x40000000, max_extended_level, ebx, ecx, edx);
9097 
9098 	fprintf(outf, "Hypervisor: ");
9099 	dump_word_chars(ebx);
9100 	dump_word_chars(ecx);
9101 	dump_word_chars(edx);
9102 	fprintf(outf, "\n");
9103 }
9104 
process_cpuid()9105 void process_cpuid()
9106 {
9107 	unsigned int eax, ebx, ecx, edx;
9108 	unsigned int fms, family, model, stepping, ecx_flags, edx_flags;
9109 	unsigned long long ucode_patch = 0;
9110 	bool ucode_patch_valid = false;
9111 
9112 	eax = ebx = ecx = edx = 0;
9113 
9114 	__cpuid(0, max_level, ebx, ecx, edx);
9115 
9116 	if (ebx == 0x756e6547 && ecx == 0x6c65746e && edx == 0x49656e69)
9117 		genuine_intel = 1;
9118 	else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65)
9119 		authentic_amd = 1;
9120 	else if (ebx == 0x6f677948 && ecx == 0x656e6975 && edx == 0x6e65476e)
9121 		hygon_genuine = 1;
9122 
9123 	if (!quiet)
9124 		fprintf(outf, "CPUID(0): %.4s%.4s%.4s 0x%x CPUID levels\n", (char *)&ebx, (char *)&edx, (char *)&ecx, max_level);
9125 
9126 	__cpuid(1, fms, ebx, ecx, edx);
9127 	family = (fms >> 8) & 0xf;
9128 	model = (fms >> 4) & 0xf;
9129 	stepping = fms & 0xf;
9130 	if (family == 0xf)
9131 		family += (fms >> 20) & 0xff;
9132 	if (family >= 6)
9133 		model += ((fms >> 16) & 0xf) << 4;
9134 	ecx_flags = ecx;
9135 	edx_flags = edx;
9136 	cpuid_has_hv = ecx_flags & (1 << 31);
9137 
9138 	if (!no_msr) {
9139 		if (get_msr(sched_getcpu(), MSR_IA32_UCODE_REV, &ucode_patch)) {
9140 			warnx("get_msr(UCODE)");
9141 		} else {
9142 			ucode_patch_valid = true;
9143 			if (!authentic_amd && !hygon_genuine)
9144 				ucode_patch >>= 32;
9145 		}
9146 	}
9147 
9148 	/*
9149 	 * check max extended function levels of CPUID.
9150 	 * This is needed to check for invariant TSC.
9151 	 * This check is valid for both Intel and AMD.
9152 	 */
9153 	ebx = ecx = edx = 0;
9154 	__cpuid(0x80000000, max_extended_level, ebx, ecx, edx);
9155 
9156 	if (!quiet) {
9157 		fprintf(outf, "CPUID(1): family:model:stepping 0x%x:%x:%x (%d:%d:%d)", family, model, stepping, family, model, stepping);
9158 		if (ucode_patch_valid)
9159 			fprintf(outf, " microcode 0x%x", (unsigned int)ucode_patch);
9160 		fputc('\n', outf);
9161 
9162 		fprintf(outf, "CPUID(0x80000000): max_extended_levels: 0x%x\n", max_extended_level);
9163 		fprintf(outf, "CPUID(1): %sSSE3 %sMONITOR %sSMX %sEIST %sTM2 %sHV %sTSC %sMSR %sACPI-TM %sHT %sTM\n",
9164 			ecx_flags & (1 << 0) ? "" : "No-",
9165 			ecx_flags & (1 << 3) ? "" : "No-",
9166 			ecx_flags & (1 << 6) ? "" : "No-",
9167 			ecx_flags & (1 << 7) ? "" : "No-",
9168 			ecx_flags & (1 << 8) ? "" : "No-",
9169 			cpuid_has_hv ? "" : "No-",
9170 			edx_flags & (1 << 4) ? "" : "No-",
9171 			edx_flags & (1 << 5) ? "" : "No-",
9172 			edx_flags & (1 << 22) ? "" : "No-", edx_flags & (1 << 28) ? "" : "No-", edx_flags & (1 << 29) ? "" : "No-");
9173 	}
9174 	if (!quiet && cpuid_has_hv)
9175 		dump_cpuid_hypervisor();
9176 
9177 	probe_platform_features(family, model);
9178 	init_perf_model_support(family, model);
9179 
9180 	if (!(edx_flags & (1 << 5)))
9181 		errx(1, "CPUID: no MSR");
9182 
9183 	if (max_extended_level >= 0x80000007) {
9184 
9185 		/*
9186 		 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
9187 		 * this check is valid for both Intel and AMD
9188 		 */
9189 		__cpuid(0x80000007, eax, ebx, ecx, edx);
9190 		has_invariant_tsc = edx & (1 << 8);
9191 	}
9192 
9193 	/*
9194 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
9195 	 * this check is valid for both Intel and AMD
9196 	 */
9197 
9198 	__cpuid(0x6, eax, ebx, ecx, edx);
9199 	cpuid_has_aperf_mperf = ecx & (1 << 0);
9200 	do_dts = eax & (1 << 0);
9201 	if (do_dts)
9202 		BIC_PRESENT(BIC_CoreTmp);
9203 	has_turbo = eax & (1 << 1);
9204 	do_ptm = eax & (1 << 6);
9205 	if (do_ptm)
9206 		BIC_PRESENT(BIC_PkgTmp);
9207 	has_hwp = eax & (1 << 7);
9208 	has_hwp_notify = eax & (1 << 8);
9209 	has_hwp_activity_window = eax & (1 << 9);
9210 	has_hwp_epp = eax & (1 << 10);
9211 	has_hwp_pkg = eax & (1 << 11);
9212 	has_epb = ecx & (1 << 3);
9213 
9214 	if (!quiet)
9215 		fprintf(outf, "CPUID(6): %sAPERF, %sTURBO, %sDTS, %sPTM, %sHWP, "
9216 			"%sHWPnotify, %sHWPwindow, %sHWPepp, %sHWPpkg, %sEPB\n",
9217 			cpuid_has_aperf_mperf ? "" : "No-",
9218 			has_turbo ? "" : "No-",
9219 			do_dts ? "" : "No-",
9220 			do_ptm ? "" : "No-",
9221 			has_hwp ? "" : "No-",
9222 			has_hwp_notify ? "" : "No-",
9223 			has_hwp_activity_window ? "" : "No-", has_hwp_epp ? "" : "No-", has_hwp_pkg ? "" : "No-", has_epb ? "" : "No-");
9224 
9225 	if (!quiet)
9226 		decode_misc_enable_msr();
9227 
9228 	if (max_level >= 0x7) {
9229 		int has_sgx;
9230 
9231 		ecx = 0;
9232 
9233 		__cpuid_count(0x7, 0, eax, ebx, ecx, edx);
9234 
9235 		has_sgx = ebx & (1 << 2);
9236 
9237 		is_hybrid = !!(edx & (1 << 15));
9238 
9239 		if (!quiet)
9240 			fprintf(outf, "CPUID(7): %sSGX %sHybrid\n", has_sgx ? "" : "No-", is_hybrid ? "" : "No-");
9241 
9242 		if (has_sgx)
9243 			decode_feature_control_msr();
9244 	}
9245 
9246 	if (max_level >= 0x15) {
9247 		unsigned int eax_crystal;
9248 		unsigned int ebx_tsc;
9249 
9250 		/*
9251 		 * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
9252 		 */
9253 		eax_crystal = ebx_tsc = crystal_hz = edx = 0;
9254 		__cpuid(0x15, eax_crystal, ebx_tsc, crystal_hz, edx);
9255 
9256 		if (ebx_tsc != 0) {
9257 			if (!quiet && (ebx != 0))
9258 				fprintf(outf, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n", eax_crystal, ebx_tsc, crystal_hz);
9259 
9260 			if (crystal_hz == 0)
9261 				crystal_hz = platform->crystal_freq;
9262 
9263 			if (crystal_hz) {
9264 				tsc_hz = (unsigned long long)crystal_hz *ebx_tsc / eax_crystal;
9265 				if (!quiet)
9266 					fprintf(outf, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n", tsc_hz / 1000000, crystal_hz, ebx_tsc, eax_crystal);
9267 			}
9268 		}
9269 	}
9270 	if (max_level >= 0x16) {
9271 		unsigned int base_mhz, max_mhz, bus_mhz, edx;
9272 
9273 		/*
9274 		 * CPUID 16H Base MHz, Max MHz, Bus MHz
9275 		 */
9276 		base_mhz = max_mhz = bus_mhz = edx = 0;
9277 
9278 		__cpuid(0x16, base_mhz, max_mhz, bus_mhz, edx);
9279 
9280 		bclk = bus_mhz;
9281 
9282 		base_hz = base_mhz * 1000000;
9283 		has_base_hz = 1;
9284 
9285 		if (platform->enable_tsc_tweak)
9286 			tsc_tweak = base_hz / tsc_hz;
9287 
9288 		if (!quiet)
9289 			fprintf(outf, "CPUID(0x16): base_mhz: %d max_mhz: %d bus_mhz: %d\n", base_mhz, max_mhz, bus_mhz);
9290 	}
9291 
9292 	if (cpuid_has_aperf_mperf)
9293 		aperf_mperf_multiplier = platform->need_perf_multiplier ? 1024 : 1;
9294 
9295 	BIC_PRESENT(BIC_IRQ);
9296 	BIC_PRESENT(BIC_NMI);
9297 	BIC_PRESENT(BIC_TSC_MHz);
9298 }
9299 
counter_info_init(void)9300 static void counter_info_init(void)
9301 {
9302 	for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i) {
9303 		struct cstate_counter_arch_info *const cai = &ccstate_counter_arch_infos[i];
9304 
9305 		if (platform->has_msr_knl_core_c6_residency && cai->msr == MSR_CORE_C6_RESIDENCY)
9306 			cai->msr = MSR_KNL_CORE_C6_RESIDENCY;
9307 
9308 		if (!platform->has_msr_core_c1_res && cai->msr == MSR_CORE_C1_RES)
9309 			cai->msr = 0;
9310 
9311 		if (platform->has_msr_atom_pkg_c6_residency && cai->msr == MSR_PKG_C6_RESIDENCY)
9312 			cai->msr = MSR_ATOM_PKG_C6_RESIDENCY;
9313 	}
9314 
9315 	for (int i = 0; i < NUM_MSR_COUNTERS; ++i) {
9316 		msr_counter_arch_infos[i].present = false;
9317 		msr_counter_arch_infos[i].needed = false;
9318 	}
9319 }
9320 
probe_pm_features(void)9321 void probe_pm_features(void)
9322 {
9323 	probe_pstates();
9324 
9325 	probe_cstates();
9326 
9327 	probe_lpi();
9328 
9329 	probe_intel_uncore_frequency();
9330 
9331 	probe_graphics();
9332 
9333 	probe_rapl();
9334 
9335 	probe_thermal();
9336 
9337 	if (platform->has_nhm_msrs && !no_msr)
9338 		BIC_PRESENT(BIC_SMI);
9339 
9340 	if (!quiet)
9341 		decode_misc_feature_control();
9342 }
9343 
9344 /*
9345  * has_perf_llc_access()
9346  *
9347  * return 1 on success, else 0
9348  */
has_perf_llc_access(void)9349 int has_perf_llc_access(void)
9350 {
9351 	int fd;
9352 
9353 	if (no_perf)
9354 		return 0;
9355 
9356 	fd = open_perf_counter(master_cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_REFERENCES, -1, PERF_FORMAT_GROUP);
9357 	if (fd != -1)
9358 		close(fd);
9359 
9360 	if (fd == -1)
9361 		warnx("Failed to access %s. Some of the counters may not be available\n"
9362 		      "\tRun as root to enable them or use %s to disable the access explicitly", "perf LLC counters", "'--hide LLC' or '--no-perf'");
9363 
9364 	return (fd != -1);
9365 }
9366 
perf_llc_init(void)9367 void perf_llc_init(void)
9368 {
9369 	int cpu;
9370 	int retval;
9371 
9372 	if (no_perf)
9373 		return;
9374 	if (!(BIC_IS_ENABLED(BIC_LLC_MRPS) || BIC_IS_ENABLED(BIC_LLC_HIT)))
9375 		return;
9376 
9377 	assert(fd_llc_percpu != 0);
9378 
9379 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
9380 
9381 		if (cpu_is_not_allowed(cpu))
9382 			continue;
9383 
9384 		fd_llc_percpu[cpu] = open_perf_counter(cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_REFERENCES, -1, PERF_FORMAT_GROUP);
9385 		if (fd_llc_percpu[cpu] == -1) {
9386 			warnx("%s: perf REFS: failed to open counter on cpu%d", __func__, cpu);
9387 			free_fd_llc_percpu();
9388 			return;
9389 		}
9390 		retval = open_perf_counter(cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES, fd_llc_percpu[cpu], PERF_FORMAT_GROUP);
9391 		if (retval == -1) {
9392 			warnx("%s: perf MISS: failed to open counter on cpu%d", __func__, cpu);
9393 			free_fd_llc_percpu();
9394 			return;
9395 		}
9396 	}
9397 	BIC_PRESENT(BIC_LLC_MRPS);
9398 	BIC_PRESENT(BIC_LLC_HIT);
9399 }
9400 
perf_l2_init(void)9401 void perf_l2_init(void)
9402 {
9403 	int cpu;
9404 	int retval;
9405 
9406 	if (no_perf)
9407 		return;
9408 	if (!(BIC_IS_ENABLED(BIC_L2_MRPS) || BIC_IS_ENABLED(BIC_L2_HIT)))
9409 		return;
9410 	if (perf_model_support == NULL)
9411 		return;
9412 
9413 	assert(fd_l2_percpu != 0);
9414 
9415 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
9416 
9417 		if (cpu_is_not_allowed(cpu))
9418 			continue;
9419 
9420 		if (!is_hybrid) {
9421 			fd_l2_percpu[cpu] = open_perf_counter(cpu, perf_pmu_types.uniform, perf_model_support->first.refs, -1, PERF_FORMAT_GROUP);
9422 			if (fd_l2_percpu[cpu] == -1) {
9423 				warnx("%s(cpu%d, 0x%x, 0x%llx) REFS", __func__, cpu, perf_pmu_types.uniform, perf_model_support->first.refs);
9424 				free_fd_l2_percpu();
9425 				return;
9426 			}
9427 			retval = open_perf_counter(cpu, perf_pmu_types.uniform, perf_model_support->first.hits, fd_l2_percpu[cpu], PERF_FORMAT_GROUP);
9428 			if (retval == -1) {
9429 				warnx("%s(cpu%d, 0x%x, 0x%llx) HITS", __func__, cpu, perf_pmu_types.uniform, perf_model_support->first.hits);
9430 				free_fd_l2_percpu();
9431 				return;
9432 			}
9433 			continue;
9434 		}
9435 		if (perf_pcore_set && CPU_ISSET_S(cpu, cpu_possible_setsize, perf_pcore_set)) {
9436 			fd_l2_percpu[cpu] = open_perf_counter(cpu, perf_pmu_types.pcore, perf_model_support->first.refs, -1, PERF_FORMAT_GROUP);
9437 			if (fd_l2_percpu[cpu] == -1) {
9438 				warnx("%s(cpu%d, 0x%x, 0x%llx) REFS", __func__, cpu, perf_pmu_types.pcore, perf_model_support->first.refs);
9439 				free_fd_l2_percpu();
9440 				return;
9441 			}
9442 			retval = open_perf_counter(cpu, perf_pmu_types.pcore, perf_model_support->first.hits, fd_l2_percpu[cpu], PERF_FORMAT_GROUP);
9443 			if (retval == -1) {
9444 				warnx("%s(cpu%d, 0x%x, 0x%llx) HITS", __func__, cpu, perf_pmu_types.pcore, perf_model_support->first.hits);
9445 				free_fd_l2_percpu();
9446 				return;
9447 			}
9448 		} else if (perf_ecore_set && CPU_ISSET_S(cpu, cpu_possible_setsize, perf_ecore_set)) {
9449 			fd_l2_percpu[cpu] = open_perf_counter(cpu, perf_pmu_types.ecore, perf_model_support->second.refs, -1, PERF_FORMAT_GROUP);
9450 			if (fd_l2_percpu[cpu] == -1) {
9451 				warnx("%s(cpu%d, 0x%x, 0x%llx) REFS", __func__, cpu, perf_pmu_types.ecore, perf_model_support->second.refs);
9452 				free_fd_l2_percpu();
9453 				return;
9454 			}
9455 			retval = open_perf_counter(cpu, perf_pmu_types.ecore, perf_model_support->second.hits, fd_l2_percpu[cpu], PERF_FORMAT_GROUP);
9456 			if (retval == -1) {
9457 				warnx("%s(cpu%d, 0x%x, 0x%llx) HITS", __func__, cpu, perf_pmu_types.ecore, perf_model_support->second.hits);
9458 				free_fd_l2_percpu();
9459 				return;
9460 			}
9461 		} else if (perf_lcore_set && CPU_ISSET_S(cpu, cpu_possible_setsize, perf_lcore_set)) {
9462 			fd_l2_percpu[cpu] = open_perf_counter(cpu, perf_pmu_types.lcore, perf_model_support->third.refs, -1, PERF_FORMAT_GROUP);
9463 			if (fd_l2_percpu[cpu] == -1) {
9464 				warnx("%s(cpu%d, 0x%x, 0x%llx) REFS", __func__, cpu, perf_pmu_types.lcore, perf_model_support->third.refs);
9465 				free_fd_l2_percpu();
9466 				return;
9467 			}
9468 			retval = open_perf_counter(cpu, perf_pmu_types.lcore, perf_model_support->third.hits, fd_l2_percpu[cpu], PERF_FORMAT_GROUP);
9469 			if (retval == -1) {
9470 				warnx("%s(cpu%d, 0x%x, 0x%llx) HITS", __func__, cpu, perf_pmu_types.lcore, perf_model_support->third.hits);
9471 				free_fd_l2_percpu();
9472 				return;
9473 			}
9474 		} else
9475 			err(-1, "%s: cpu%d: type %d", __func__, cpu, cpus[cpu].type);
9476 	}
9477 	BIC_PRESENT(BIC_L2_MRPS);
9478 	BIC_PRESENT(BIC_L2_HIT);
9479 }
9480 
9481 /*
9482  * in /dev/cpu/ return success for names that are numbers
9483  * ie. filter out ".", "..", "microcode".
9484  */
dir_filter(const struct dirent * dirp)9485 int dir_filter(const struct dirent *dirp)
9486 {
9487 	if (isdigit(dirp->d_name[0]))
9488 		return 1;
9489 	else
9490 		return 0;
9491 }
9492 
set_thread_siblings(struct cpu_topology * thiscpu)9493 int set_thread_siblings(struct cpu_topology *thiscpu)
9494 {
9495 	char path[80];
9496 	int cpu = thiscpu->cpu_id;
9497 	size_t size;
9498 	int ht_id = 0;
9499 	int i;
9500 
9501 	thiscpu->put_ids = CPU_ALLOC((topo.max_cpu_num + 1));
9502 	if (thiscpu->ht_id < 0)
9503 		thiscpu->ht_id = 0;	/* first CPU in core */
9504 	if (!thiscpu->put_ids)
9505 		return -1;
9506 
9507 	size = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9508 	CPU_ZERO_S(size, thiscpu->put_ids);
9509 
9510 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology", cpu);
9511 
9512 	initialize_cpu_set_from_sysfs(thiscpu->put_ids, path, "thread_siblings_list");
9513 
9514 	for (i = 0; i <= topo.max_cpu_num; ++i)
9515 		if (CPU_ISSET_S(i, size, thiscpu->put_ids)) {
9516 			cpus[i].ht_id = ht_id;
9517 			cpus[cpu].ht_sibling_cpu_id[ht_id] = i;
9518 			ht_id += 1;
9519 		}
9520 
9521 	return (ht_id - 1);
9522 }
9523 
topology_probe(bool startup)9524 void topology_probe(bool startup)
9525 {
9526 	int i;
9527 	int max_core_id = 0;
9528 	int max_package_id = 0;
9529 	int max_siblings = 0;
9530 
9531 	/* Initialize num_cpus, max_cpu_num */
9532 	set_max_cpu_num();
9533 	topo.num_cpus = 0;
9534 	for_all_proc_cpus(count_cpus);
9535 	if (!summary_only)
9536 		BIC_PRESENT(BIC_CPU);
9537 
9538 	if (debug > 1)
9539 		fprintf(outf, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
9540 
9541 	cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
9542 	if (cpus == NULL)
9543 		err(1, "calloc cpus");
9544 
9545 	/*
9546 	 * Allocate and initialize cpu_present_set
9547 	 */
9548 	cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
9549 	if (cpu_present_set == NULL)
9550 		err(3, "CPU_ALLOC");
9551 	cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9552 	CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
9553 	for_all_proc_cpus(mark_cpu_present);
9554 	if (debug)
9555 		print_cpu_set("present set", cpu_present_set);
9556 
9557 	/*
9558 	 * Allocate and initialize cpu_possible_set
9559 	 */
9560 	cpu_possible_set = CPU_ALLOC((topo.max_cpu_num + 1));
9561 	if (cpu_possible_set == NULL)
9562 		err(3, "CPU_ALLOC");
9563 	cpu_possible_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9564 	CPU_ZERO_S(cpu_possible_setsize, cpu_possible_set);
9565 	initialize_cpu_set_from_sysfs(cpu_possible_set, "/sys/devices/system/cpu", "possible");
9566 	if (debug)
9567 		print_cpu_set("possible set", cpu_possible_set);
9568 
9569 	/*
9570 	 * Allocate and initialize cpu_effective_set
9571 	 */
9572 	cpu_effective_set = CPU_ALLOC((topo.max_cpu_num + 1));
9573 	if (cpu_effective_set == NULL)
9574 		err(3, "CPU_ALLOC");
9575 	cpu_effective_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9576 	CPU_ZERO_S(cpu_effective_setsize, cpu_effective_set);
9577 	update_effective_set(startup);
9578 	if (debug)
9579 		print_cpu_set("effective set", cpu_effective_set);
9580 
9581 	/*
9582 	 * Allocate and initialize cpu_allowed_set
9583 	 */
9584 	cpu_allowed_set = CPU_ALLOC((topo.max_cpu_num + 1));
9585 	if (cpu_allowed_set == NULL)
9586 		err(3, "CPU_ALLOC");
9587 	cpu_allowed_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9588 	CPU_ZERO_S(cpu_allowed_setsize, cpu_allowed_set);
9589 
9590 	/*
9591 	 * Validate and update cpu_allowed_set.
9592 	 *
9593 	 * Make sure all cpus in cpu_subset are also in cpu_present_set during startup.
9594 	 * Give a warning when cpus in cpu_subset become unavailable at runtime.
9595 	 * Give a warning when cpus are not effective because of cgroup setting.
9596 	 *
9597 	 * cpu_allowed_set is the intersection of cpu_present_set/cpu_effective_set/cpu_subset.
9598 	 */
9599 	for (i = 0; i < CPU_SUBSET_MAXCPUS; ++i) {
9600 		if (cpu_subset && !CPU_ISSET_S(i, cpu_subset_size, cpu_subset))
9601 			continue;
9602 
9603 		if (!CPU_ISSET_S(i, cpu_present_setsize, cpu_present_set)) {
9604 			if (cpu_subset) {
9605 				/* cpus in cpu_subset must be in cpu_present_set during startup */
9606 				if (startup)
9607 					err(1, "cpu%d not present", i);
9608 				else
9609 					fprintf(stderr, "cpu%d not present\n", i);
9610 			}
9611 			continue;
9612 		}
9613 
9614 		if (CPU_COUNT_S(cpu_effective_setsize, cpu_effective_set)) {
9615 			if (!CPU_ISSET_S(i, cpu_effective_setsize, cpu_effective_set)) {
9616 				fprintf(stderr, "cpu%d not effective\n", i);
9617 				continue;
9618 			}
9619 		}
9620 
9621 		CPU_SET_S(i, cpu_allowed_setsize, cpu_allowed_set);
9622 	}
9623 	if (debug)
9624 		print_cpu_set("allowed set", cpu_allowed_set);
9625 
9626 	if (!CPU_COUNT_S(cpu_allowed_setsize, cpu_allowed_set))
9627 		err(-ENODEV, "No valid cpus found");
9628 	sched_setaffinity(0, cpu_allowed_setsize, cpu_allowed_set);
9629 
9630 	/*
9631 	 * Allocate and initialize cpu_affinity_set
9632 	 */
9633 	cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
9634 	if (cpu_affinity_set == NULL)
9635 		err(3, "CPU_ALLOC");
9636 	cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9637 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
9638 
9639 	for_all_proc_cpus(clear_ht_id);
9640 
9641 	for_all_proc_cpus(set_cpu_hybrid_type);
9642 
9643 	/*
9644 	 * For online cpus
9645 	 * find max_core_id, max_package_id, num_cores (per system)
9646 	 */
9647 	topo.min_module_id = 0x7FFFFFFF;
9648 	for (i = 0; i <= topo.max_cpu_num; ++i) {
9649 		int siblings;
9650 
9651 		if (cpu_is_not_present(i)) {
9652 			if (debug > 1)
9653 				fprintf(outf, "cpu%d NOT PRESENT\n", i);
9654 			continue;
9655 		}
9656 
9657 		cpus[i].cpu_id = i;
9658 
9659 		/* get package information */
9660 		cpus[i].package_id = get_package_id(i);
9661 		if (cpus[i].package_id > max_package_id)
9662 			max_package_id = cpus[i].package_id;
9663 
9664 		/* get die information */
9665 		cpus[i].die_id = get_die_id(i);
9666 		if (cpus[i].die_id > topo.max_die_id)
9667 			topo.max_die_id = cpus[i].die_id;
9668 
9669 		/* get l3 information */
9670 		cpus[i].l3_id = get_l3_id(i);
9671 		if (cpus[i].l3_id > topo.max_l3_id)
9672 			topo.max_l3_id = cpus[i].l3_id;
9673 
9674 		/* get numa node information */
9675 		cpus[i].physical_node_id = get_physical_node_id(&cpus[i]);
9676 		if (cpus[i].physical_node_id > topo.max_node_num)
9677 			topo.max_node_num = cpus[i].physical_node_id;
9678 
9679 		/* get module information */
9680 		cpus[i].module_id = get_module_id(i);
9681 		if (cpus[i].module_id > topo.max_module_id)
9682 			topo.max_module_id = cpus[i].module_id;
9683 		if (cpus[i].module_id < topo.min_module_id)
9684 			topo.min_module_id = cpus[i].module_id;
9685 
9686 		/* get core information */
9687 		cpus[i].core_id = get_core_id(i);
9688 		if (cpus[i].core_id > max_core_id)
9689 			max_core_id = cpus[i].core_id;
9690 
9691 		/* get thread information */
9692 		siblings = set_thread_siblings(&cpus[i]);
9693 		if (siblings > max_siblings)
9694 			max_siblings = siblings;
9695 		if (cpus[i].ht_id == 0)
9696 			topo.num_cores++;
9697 	}
9698 	topo.max_core_id = max_core_id;	/* within a package */
9699 	topo.max_package_id = max_package_id;
9700 
9701 	topo.cores_per_pkg = max_core_id + 1;
9702 	if (debug > 1)
9703 		fprintf(outf, "max_core_id %d, sizing for %d cores per package\n", max_core_id, topo.cores_per_pkg);
9704 	if (!summary_only)
9705 		BIC_PRESENT(BIC_Core);
9706 
9707 	if (debug > 1)
9708 		fprintf(outf, "min_module_id %d max_module_id %d\n", topo.min_module_id, topo.max_module_id);
9709 	if (!summary_only && (topo.min_module_id != topo.max_module_id))
9710 		BIC_PRESENT(BIC_Module);
9711 
9712 	topo.num_die = topo.max_die_id + 1;
9713 	if (debug > 1)
9714 		fprintf(outf, "max_die_id %d, sizing for %d die\n", topo.max_die_id, topo.num_die);
9715 	if (!summary_only && topo.num_die > 1)
9716 		BIC_PRESENT(BIC_Die);
9717 
9718 	if (!summary_only && topo.max_l3_id > 0)
9719 		BIC_PRESENT(BIC_L3);
9720 
9721 	topo.num_packages = max_package_id + 1;
9722 	if (debug > 1)
9723 		fprintf(outf, "max_package_id %d, sizing for %d packages\n", max_package_id, topo.num_packages);
9724 	if (!summary_only && topo.num_packages > 1)
9725 		BIC_PRESENT(BIC_Package);
9726 
9727 	set_node_data();
9728 	if (debug > 1)
9729 		fprintf(outf, "nodes_per_pkg %d\n", topo.nodes_per_pkg);
9730 	if (!summary_only && topo.nodes_per_pkg > 1)
9731 		BIC_PRESENT(BIC_Node);
9732 
9733 	topo.threads_per_core = max_siblings;
9734 	if (debug > 1)
9735 		fprintf(outf, "max_siblings %d\n", max_siblings);
9736 
9737 	if (debug < 1)
9738 		return;
9739 
9740 	for (i = 0; i <= topo.max_cpu_num; ++i) {
9741 		int ht_id;
9742 
9743 		if (cpu_is_not_present(i))
9744 			continue;
9745 		fprintf(outf,
9746 			"cpu %d pkg %d die %d l3 %d node %d lnode %d module 0x%x core %d ht_id %d",
9747 			i, cpus[i].package_id, cpus[i].die_id, cpus[i].l3_id,
9748 			cpus[i].physical_node_id, cpus[i].logical_node_id, cpus[i].module_id, cpus[i].core_id, cpus[i].ht_id);
9749 		fprintf(outf, " siblings");
9750 		for (ht_id = 0; ht_id <= MAX_HT_ID; ++ht_id)
9751 			fprintf(outf, " %d", cpus[i].ht_sibling_cpu_id[ht_id]);
9752 		fprintf(outf, "\n");
9753 	}
9754 
9755 }
9756 
allocate_counters_1(struct counters * counters)9757 void allocate_counters_1(struct counters *counters)
9758 {
9759 	counters->threads = calloc(1, sizeof(struct thread_data));
9760 	if (counters->threads == NULL)
9761 		goto error;
9762 
9763 	counters->cores = calloc(1, sizeof(struct core_data));
9764 	if (counters->cores == NULL)
9765 		goto error;
9766 
9767 	counters->packages = calloc(1, sizeof(struct pkg_data));
9768 	if (counters->packages == NULL)
9769 		goto error;
9770 
9771 	return;
9772 error:
9773 	err(1, "calloc counters_1");
9774 }
9775 
allocate_counters(struct counters * counters)9776 void allocate_counters(struct counters *counters)
9777 {
9778 	int i;
9779 	int num_cores = topo.cores_per_pkg * topo.num_packages;
9780 
9781 	counters->threads = calloc(topo.max_cpu_num + 1, sizeof(struct thread_data));
9782 	if (counters->threads == NULL)
9783 		goto error;
9784 
9785 	for (i = 0; i < topo.max_cpu_num + 1; i++)
9786 		(counters->threads)[i].cpu_id = -1;
9787 
9788 	counters->cores = calloc(num_cores, sizeof(struct core_data));
9789 	if (counters->cores == NULL)
9790 		goto error;
9791 
9792 	for (i = 0; i < num_cores; i++)
9793 		(counters->cores)[i].first_cpu = -1;
9794 
9795 	counters->packages = calloc(topo.num_packages, sizeof(struct pkg_data));
9796 	if (counters->packages == NULL)
9797 		goto error;
9798 
9799 	for (i = 0; i < topo.num_packages; i++)
9800 		(counters->packages)[i].first_cpu = -1;
9801 
9802 	return;
9803 error:
9804 	err(1, "calloc counters");
9805 }
9806 
9807 /*
9808  * init_counter()
9809  *
9810  * set t->cpu_id, FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
9811  */
init_counter(struct thread_data * thread_base,struct core_data * core_base,struct pkg_data * pkg_base,int cpu_id)9812 void init_counter(struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base, int cpu_id)
9813 {
9814 	int pkg_id = cpus[cpu_id].package_id;
9815 	int node_id = cpus[cpu_id].logical_node_id;
9816 	int core_id = cpus[cpu_id].core_id;
9817 	struct thread_data *t;
9818 	struct core_data *c;
9819 
9820 	/* Workaround for systems where physical_node_id==-1
9821 	 * and logical_node_id==(-1 - topo.num_cpus)
9822 	 */
9823 	if (node_id < 0)
9824 		node_id = 0;
9825 
9826 	t = &thread_base[cpu_id];
9827 	c = &core_base[GLOBAL_CORE_ID(core_id, pkg_id)];
9828 
9829 	t->cpu_id = cpu_id;
9830 	if (!cpu_is_not_allowed(cpu_id)) {
9831 
9832 		if (c->first_cpu < 0)
9833 			c->first_cpu = t->cpu_id;
9834 		if (pkg_base[pkg_id].first_cpu < 0)
9835 			pkg_base[pkg_id].first_cpu = t->cpu_id;
9836 	}
9837 }
9838 
initialize_counters(int cpu_id)9839 int initialize_counters(int cpu_id)
9840 {
9841 	init_counter(EVEN_COUNTERS, cpu_id);
9842 	init_counter(ODD_COUNTERS, cpu_id);
9843 	return 0;
9844 }
9845 
allocate_output_buffer()9846 void allocate_output_buffer()
9847 {
9848 	output_buffer = calloc(1, (1 + topo.num_cpus) * 2048);
9849 	outp = output_buffer;
9850 	if (outp == NULL)
9851 		err(-1, "calloc output buffer");
9852 }
9853 
allocate_fd_percpu(void)9854 void allocate_fd_percpu(void)
9855 {
9856 	fd_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
9857 	if (fd_percpu == NULL)
9858 		err(-1, "calloc fd_percpu");
9859 }
9860 
allocate_irq_buffers(void)9861 void allocate_irq_buffers(void)
9862 {
9863 	irq_column_2_cpu = calloc(topo.num_cpus, sizeof(int));
9864 	if (irq_column_2_cpu == NULL)
9865 		err(-1, "calloc %d", topo.num_cpus);
9866 
9867 	irqs_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
9868 	if (irqs_per_cpu == NULL)
9869 		err(-1, "calloc %d IRQ", topo.max_cpu_num + 1);
9870 
9871 	nmi_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
9872 	if (nmi_per_cpu == NULL)
9873 		err(-1, "calloc %d NMI", topo.max_cpu_num + 1);
9874 }
9875 
update_topo(PER_THREAD_PARAMS)9876 int update_topo(PER_THREAD_PARAMS)
9877 {
9878 	topo.allowed_cpus++;
9879 	if ((int)t->cpu_id == c->first_cpu)
9880 		topo.allowed_cores++;
9881 	if ((int)t->cpu_id == p->first_cpu)
9882 		topo.allowed_packages++;
9883 
9884 	return 0;
9885 }
9886 
topology_update(void)9887 void topology_update(void)
9888 {
9889 	topo.allowed_cpus = 0;
9890 	topo.allowed_cores = 0;
9891 	topo.allowed_packages = 0;
9892 	for_all_cpus(update_topo, ODD_COUNTERS);
9893 	if (debug)
9894 		fprintf(stderr, "allowed_cpus %d allowed_cores %d allowed_packages %d\n", topo.allowed_cpus, topo.allowed_cores, topo.allowed_packages);
9895 }
9896 
setup_all_buffers(bool startup)9897 void setup_all_buffers(bool startup)
9898 {
9899 	topology_probe(startup);
9900 	allocate_irq_buffers();
9901 	allocate_fd_percpu();
9902 	allocate_counters_1(&average);
9903 	allocate_counters(&even);
9904 	allocate_counters(&odd);
9905 	allocate_output_buffer();
9906 	for_all_proc_cpus(initialize_counters);
9907 	topology_update();
9908 }
9909 
set_master_cpu(void)9910 void set_master_cpu(void)
9911 {
9912 	int i;
9913 
9914 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
9915 		if (cpu_is_not_allowed(i))
9916 			continue;
9917 		master_cpu = i;
9918 		if (debug > 1)
9919 			fprintf(outf, "master_cpu = %d\n", master_cpu);
9920 		return;
9921 	}
9922 	err(-ENODEV, "No valid cpus found");
9923 }
9924 
has_added_counters(void)9925 bool has_added_counters(void)
9926 {
9927 	/*
9928 	 * It only makes sense to call this after the command line is parsed,
9929 	 * otherwise sys structure is not populated.
9930 	 */
9931 
9932 	return sys.added_core_counters | sys.added_thread_counters | sys.added_package_counters;
9933 }
9934 
check_msr_access(void)9935 void check_msr_access(void)
9936 {
9937 	check_msr_driver();
9938 	check_msr_permission();
9939 
9940 	if (no_msr)
9941 		bic_disable_msr_access();
9942 }
9943 
check_perf_access(void)9944 void check_perf_access(void)
9945 {
9946 	if (BIC_IS_ENABLED(BIC_IPC))
9947 		if (!has_perf_instr_count_access())
9948 			no_perf = 1;
9949 
9950 	if (BIC_IS_ENABLED(BIC_LLC_MRPS) || BIC_IS_ENABLED(BIC_LLC_HIT))
9951 		if (!has_perf_llc_access())
9952 			no_perf = 1;
9953 
9954 	if (no_perf)
9955 		bic_disable_perf_access();
9956 }
9957 
perf_has_hybrid_devices(void)9958 bool perf_has_hybrid_devices(void)
9959 {
9960 	/*
9961 	 *  0: unknown
9962 	 *  1: has separate perf device for p and e core
9963 	 * -1: doesn't have separate perf device for p and e core
9964 	 */
9965 	static int cached;
9966 
9967 	if (cached > 0)
9968 		return true;
9969 
9970 	if (cached < 0)
9971 		return false;
9972 
9973 	if (access("/sys/bus/event_source/devices/cpu_core", F_OK)) {
9974 		cached = -1;
9975 		return false;
9976 	}
9977 
9978 	if (access("/sys/bus/event_source/devices/cpu_atom", F_OK)) {
9979 		cached = -1;
9980 		return false;
9981 	}
9982 
9983 	cached = 1;
9984 	return true;
9985 }
9986 
added_perf_counters_init_(struct perf_counter_info * pinfo)9987 int added_perf_counters_init_(struct perf_counter_info *pinfo)
9988 {
9989 	size_t num_domains = 0;
9990 	unsigned int next_domain;
9991 	bool *domain_visited;
9992 	unsigned int perf_type, perf_config;
9993 	double perf_scale;
9994 	int fd_perf;
9995 
9996 	if (!pinfo)
9997 		return 0;
9998 
9999 	const size_t max_num_domains = MAX(topo.max_cpu_num + 1, MAX(topo.max_core_id + 1, topo.max_package_id + 1));
10000 
10001 	domain_visited = calloc(max_num_domains, sizeof(*domain_visited));
10002 
10003 	while (pinfo) {
10004 		switch (pinfo->scope) {
10005 		case SCOPE_CPU:
10006 			num_domains = topo.max_cpu_num + 1;
10007 			break;
10008 
10009 		case SCOPE_CORE:
10010 			num_domains = topo.max_core_id + 1;
10011 			break;
10012 
10013 		case SCOPE_PACKAGE:
10014 			num_domains = topo.max_package_id + 1;
10015 			break;
10016 		}
10017 
10018 		/* Allocate buffer for file descriptor for each domain. */
10019 		pinfo->fd_perf_per_domain = calloc(num_domains, sizeof(*pinfo->fd_perf_per_domain));
10020 		if (!pinfo->fd_perf_per_domain)
10021 			errx(1, "%s: alloc %s", __func__, "fd_perf_per_domain");
10022 
10023 		for (size_t i = 0; i < num_domains; ++i)
10024 			pinfo->fd_perf_per_domain[i] = -1;
10025 
10026 		pinfo->num_domains = num_domains;
10027 		pinfo->scale = 1.0;
10028 
10029 		memset(domain_visited, 0, max_num_domains * sizeof(*domain_visited));
10030 
10031 		for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
10032 
10033 			next_domain = cpu_to_domain(pinfo, cpu);
10034 
10035 			assert(next_domain < num_domains);
10036 
10037 			if (cpu_is_not_allowed(cpu))
10038 				continue;
10039 
10040 			if (domain_visited[next_domain])
10041 				continue;
10042 
10043 			/*
10044 			 * Intel hybrid platforms expose different perf devices for P and E cores.
10045 			 * Instead of one, "/sys/bus/event_source/devices/cpu" device, there are
10046 			 * "/sys/bus/event_source/devices/{cpu_core,cpu_atom}".
10047 			 *
10048 			 * This makes it more complicated to the user, because most of the counters
10049 			 * are available on both and have to be handled manually, otherwise.
10050 			 *
10051 			 * Code below, allow user to use the old "cpu" name, which is translated accordingly.
10052 			 */
10053 			const char *perf_device = pinfo->device;
10054 
10055 			if (strcmp(perf_device, "cpu") == 0 && perf_has_hybrid_devices()) {
10056 				switch (cpus[cpu].type) {
10057 				case INTEL_PCORE_TYPE:
10058 					perf_device = "cpu_core";
10059 					break;
10060 
10061 				case INTEL_ECORE_TYPE:
10062 					perf_device = "cpu_atom";
10063 					break;
10064 
10065 				default:	/* Don't change, we will probably fail and report a problem soon. */
10066 					break;
10067 				}
10068 			}
10069 
10070 			perf_type = read_perf_type(perf_device);
10071 			if (perf_type == (unsigned int)-1) {
10072 				warnx("%s: perf/%s/%s: failed to read %s", __func__, perf_device, pinfo->event, "type");
10073 				continue;
10074 			}
10075 
10076 			perf_config = read_perf_config(perf_device, pinfo->event);
10077 			if (perf_config == (unsigned int)-1) {
10078 				warnx("%s: perf/%s/%s: failed to read %s", __func__, perf_device, pinfo->event, "config");
10079 				continue;
10080 			}
10081 
10082 			/* Scale is not required, some counters just don't have it. */
10083 			perf_scale = read_perf_scale(perf_device, pinfo->event);
10084 			if (perf_scale == 0.0)
10085 				perf_scale = 1.0;
10086 
10087 			fd_perf = open_perf_counter(cpu, perf_type, perf_config, -1, 0);
10088 			if (fd_perf == -1) {
10089 				warnx("%s: perf/%s/%s: failed to open counter on cpu%d", __func__, perf_device, pinfo->event, cpu);
10090 				continue;
10091 			}
10092 
10093 			domain_visited[next_domain] = 1;
10094 			pinfo->fd_perf_per_domain[next_domain] = fd_perf;
10095 			pinfo->scale = perf_scale;
10096 
10097 			if (debug)
10098 				fprintf(stderr, "Add perf/%s/%s cpu%d: %d\n", perf_device, pinfo->event, cpu, pinfo->fd_perf_per_domain[next_domain]);
10099 		}
10100 
10101 		pinfo = pinfo->next;
10102 	}
10103 
10104 	free(domain_visited);
10105 
10106 	return 0;
10107 }
10108 
added_perf_counters_init(void)10109 void added_perf_counters_init(void)
10110 {
10111 	if (added_perf_counters_init_(sys.perf_tp))
10112 		errx(1, "%s: %s", __func__, "thread");
10113 
10114 	if (added_perf_counters_init_(sys.perf_cp))
10115 		errx(1, "%s: %s", __func__, "core");
10116 
10117 	if (added_perf_counters_init_(sys.perf_pp))
10118 		errx(1, "%s: %s", __func__, "package");
10119 }
10120 
parse_telem_info_file(int fd_dir,const char * info_filename,const char * format,unsigned long * output)10121 int parse_telem_info_file(int fd_dir, const char *info_filename, const char *format, unsigned long *output)
10122 {
10123 	int fd_telem_info;
10124 	FILE *file_telem_info;
10125 	unsigned long value;
10126 
10127 	fd_telem_info = openat(fd_dir, info_filename, O_RDONLY);
10128 	if (fd_telem_info == -1)
10129 		return -1;
10130 
10131 	file_telem_info = fdopen(fd_telem_info, "r");
10132 	if (file_telem_info == NULL) {
10133 		close(fd_telem_info);
10134 		return -1;
10135 	}
10136 
10137 	if (fscanf(file_telem_info, format, &value) != 1) {
10138 		fclose(file_telem_info);
10139 		return -1;
10140 	}
10141 
10142 	fclose(file_telem_info);
10143 
10144 	*output = value;
10145 
10146 	return 0;
10147 }
10148 
pmt_mmio_open(unsigned int target_guid)10149 struct pmt_mmio *pmt_mmio_open(unsigned int target_guid)
10150 {
10151 	struct pmt_diriter_t pmt_iter;
10152 	const struct dirent *entry;
10153 	struct stat st;
10154 	int fd_telem_dir, fd_pmt;
10155 	unsigned long guid, size, offset;
10156 	size_t mmap_size;
10157 	void *mmio;
10158 	struct pmt_mmio *head = NULL, *last = NULL;
10159 	struct pmt_mmio *new_pmt = NULL;
10160 
10161 	if (stat(SYSFS_TELEM_PATH, &st) == -1)
10162 		return NULL;
10163 
10164 	pmt_diriter_init(&pmt_iter);
10165 	entry = pmt_diriter_begin(&pmt_iter, SYSFS_TELEM_PATH);
10166 	if (!entry) {
10167 		pmt_diriter_remove(&pmt_iter);
10168 		return NULL;
10169 	}
10170 
10171 	for (; entry != NULL; entry = pmt_diriter_next(&pmt_iter)) {
10172 		if (fstatat(dirfd(pmt_iter.dir), entry->d_name, &st, 0) == -1)
10173 			break;
10174 
10175 		if (!S_ISDIR(st.st_mode))
10176 			continue;
10177 
10178 		fd_telem_dir = openat(dirfd(pmt_iter.dir), entry->d_name, O_RDONLY);
10179 		if (fd_telem_dir == -1)
10180 			break;
10181 
10182 		if (parse_telem_info_file(fd_telem_dir, "guid", "%lx", &guid)) {
10183 			close(fd_telem_dir);
10184 			break;
10185 		}
10186 
10187 		if (parse_telem_info_file(fd_telem_dir, "size", "%lu", &size)) {
10188 			close(fd_telem_dir);
10189 			break;
10190 		}
10191 
10192 		if (guid != target_guid) {
10193 			close(fd_telem_dir);
10194 			continue;
10195 		}
10196 
10197 		if (parse_telem_info_file(fd_telem_dir, "offset", "%lu", &offset)) {
10198 			close(fd_telem_dir);
10199 			break;
10200 		}
10201 
10202 		assert(offset == 0);
10203 
10204 		fd_pmt = openat(fd_telem_dir, "telem", O_RDONLY);
10205 		if (fd_pmt == -1)
10206 			goto loop_cleanup_and_break;
10207 
10208 		mmap_size = ROUND_UP_TO_PAGE_SIZE(size);
10209 		mmio = mmap(0, mmap_size, PROT_READ, MAP_SHARED, fd_pmt, 0);
10210 		if (mmio != MAP_FAILED) {
10211 			if (debug)
10212 				fprintf(stderr, "%s: 0x%lx mmaped at: %p\n", __func__, guid, mmio);
10213 
10214 			new_pmt = calloc(1, sizeof(*new_pmt));
10215 
10216 			if (!new_pmt) {
10217 				fprintf(stderr, "%s: Failed to allocate pmt_mmio\n", __func__);
10218 				exit(1);
10219 			}
10220 
10221 			/*
10222 			 * Create linked list of mmaped regions,
10223 			 * but preserve the ordering from sysfs.
10224 			 * Ordering is important for the user to
10225 			 * use the seq=%u parameter when adding a counter.
10226 			 */
10227 			new_pmt->guid = guid;
10228 			new_pmt->mmio_base = mmio;
10229 			new_pmt->pmt_offset = offset;
10230 			new_pmt->size = size;
10231 			new_pmt->next = pmt_mmios;
10232 
10233 			if (last)
10234 				last->next = new_pmt;
10235 			else
10236 				head = new_pmt;
10237 
10238 			last = new_pmt;
10239 		}
10240 
10241 loop_cleanup_and_break:
10242 		close(fd_pmt);
10243 		close(fd_telem_dir);
10244 	}
10245 
10246 	pmt_diriter_remove(&pmt_iter);
10247 
10248 	/*
10249 	 * If we found something, stick just
10250 	 * created linked list to the front.
10251 	 */
10252 	if (head)
10253 		pmt_mmios = head;
10254 
10255 	return head;
10256 }
10257 
pmt_mmio_find(unsigned int guid)10258 struct pmt_mmio *pmt_mmio_find(unsigned int guid)
10259 {
10260 	struct pmt_mmio *pmmio = pmt_mmios;
10261 
10262 	while (pmmio) {
10263 		if (pmmio->guid == guid)
10264 			return pmmio;
10265 
10266 		pmmio = pmmio->next;
10267 	}
10268 
10269 	return NULL;
10270 }
10271 
pmt_get_counter_pointer(struct pmt_mmio * pmmio,unsigned long counter_offset)10272 void *pmt_get_counter_pointer(struct pmt_mmio *pmmio, unsigned long counter_offset)
10273 {
10274 	char *ret;
10275 
10276 	/* Get base of mmaped PMT file. */
10277 	ret = (char *)pmmio->mmio_base;
10278 
10279 	/*
10280 	 * Apply PMT MMIO offset to obtain beginning of the mmaped telemetry data.
10281 	 * It's not guaranteed that the mmaped memory begins with the telemetry data
10282 	 *      - we might have to apply the offset first.
10283 	 */
10284 	ret += pmmio->pmt_offset;
10285 
10286 	/* Apply the counter offset to get the address to the mmaped counter. */
10287 	ret += counter_offset;
10288 
10289 	return ret;
10290 }
10291 
pmt_add_guid(unsigned int guid,unsigned int seq)10292 struct pmt_mmio *pmt_add_guid(unsigned int guid, unsigned int seq)
10293 {
10294 	struct pmt_mmio *ret;
10295 
10296 	ret = pmt_mmio_find(guid);
10297 	if (!ret)
10298 		ret = pmt_mmio_open(guid);
10299 
10300 	while (ret && seq) {
10301 		ret = ret->next;
10302 		--seq;
10303 	}
10304 
10305 	return ret;
10306 }
10307 
10308 enum pmt_open_mode {
10309 	PMT_OPEN_TRY,		/* Open failure is not an error. */
10310 	PMT_OPEN_REQUIRED,	/* Open failure is a fatal error. */
10311 };
10312 
pmt_find_counter(struct pmt_counter * pcounter,const char * name)10313 struct pmt_counter *pmt_find_counter(struct pmt_counter *pcounter, const char *name)
10314 {
10315 	while (pcounter) {
10316 		if (strcmp(pcounter->name, name) == 0)
10317 			break;
10318 
10319 		pcounter = pcounter->next;
10320 	}
10321 
10322 	return pcounter;
10323 }
10324 
pmt_get_scope_root(enum counter_scope scope)10325 struct pmt_counter **pmt_get_scope_root(enum counter_scope scope)
10326 {
10327 	switch (scope) {
10328 	case SCOPE_CPU:
10329 		return &sys.pmt_tp;
10330 	case SCOPE_CORE:
10331 		return &sys.pmt_cp;
10332 	case SCOPE_PACKAGE:
10333 		return &sys.pmt_pp;
10334 	}
10335 
10336 	__builtin_unreachable();
10337 }
10338 
pmt_counter_add_domain(struct pmt_counter * pcounter,unsigned long * pmmio,unsigned int domain_id)10339 void pmt_counter_add_domain(struct pmt_counter *pcounter, unsigned long *pmmio, unsigned int domain_id)
10340 {
10341 	/* Make sure the new domain fits. */
10342 	if (domain_id >= pcounter->num_domains)
10343 		pmt_counter_resize(pcounter, domain_id + 1);
10344 
10345 	assert(pcounter->domains);
10346 	assert(domain_id < pcounter->num_domains);
10347 
10348 	pcounter->domains[domain_id].pcounter = pmmio;
10349 }
10350 
pmt_add_counter(unsigned int guid,unsigned int seq,const char * name,enum pmt_datatype type,unsigned int lsb,unsigned int msb,unsigned int offset,enum counter_scope scope,enum counter_format format,unsigned int domain_id,enum pmt_open_mode mode)10351 int pmt_add_counter(unsigned int guid, unsigned int seq, const char *name, enum pmt_datatype type,
10352 		    unsigned int lsb, unsigned int msb, unsigned int offset, enum counter_scope scope,
10353 		    enum counter_format format, unsigned int domain_id, enum pmt_open_mode mode)
10354 {
10355 	struct pmt_mmio *mmio;
10356 	struct pmt_counter *pcounter;
10357 	struct pmt_counter **const pmt_root = pmt_get_scope_root(scope);
10358 	bool new_counter = false;
10359 	int conflict = 0;
10360 
10361 	if (lsb > msb) {
10362 		fprintf(stderr, "%s: %s: `%s` must be satisfied\n", __func__, "lsb <= msb", name);
10363 		exit(1);
10364 	}
10365 
10366 	if (msb >= 64) {
10367 		fprintf(stderr, "%s: %s: `%s` must be satisfied\n", __func__, "msb < 64", name);
10368 		exit(1);
10369 	}
10370 
10371 	mmio = pmt_add_guid(guid, seq);
10372 	if (!mmio) {
10373 		if (mode != PMT_OPEN_TRY) {
10374 			fprintf(stderr, "%s: failed to map PMT MMIO for guid %x, seq %u\n", __func__, guid, seq);
10375 			exit(1);
10376 		}
10377 
10378 		return 1;
10379 	}
10380 
10381 	if (offset >= mmio->size) {
10382 		if (mode != PMT_OPEN_TRY) {
10383 			fprintf(stderr, "%s: offset %u outside of PMT MMIO size %u\n", __func__, offset, mmio->size);
10384 			exit(1);
10385 		}
10386 
10387 		return 1;
10388 	}
10389 
10390 	pcounter = pmt_find_counter(*pmt_root, name);
10391 	if (!pcounter) {
10392 		pcounter = calloc(1, sizeof(*pcounter));
10393 		new_counter = true;
10394 	}
10395 
10396 	if (new_counter) {
10397 		strncpy(pcounter->name, name, ARRAY_SIZE(pcounter->name) - 1);
10398 		pcounter->type = type;
10399 		pcounter->scope = scope;
10400 		pcounter->lsb = lsb;
10401 		pcounter->msb = msb;
10402 		pcounter->format = format;
10403 	} else {
10404 		conflict += pcounter->type != type;
10405 		conflict += pcounter->scope != scope;
10406 		conflict += pcounter->lsb != lsb;
10407 		conflict += pcounter->msb != msb;
10408 		conflict += pcounter->format != format;
10409 	}
10410 
10411 	if (conflict) {
10412 		fprintf(stderr, "%s: conflicting parameters for the PMT counter with the same name %s\n", __func__, name);
10413 		exit(1);
10414 	}
10415 
10416 	pmt_counter_add_domain(pcounter, pmt_get_counter_pointer(mmio, offset), domain_id);
10417 
10418 	if (new_counter) {
10419 		pcounter->next = *pmt_root;
10420 		*pmt_root = pcounter;
10421 	}
10422 
10423 	return 0;
10424 }
10425 
pmt_init(void)10426 void pmt_init(void)
10427 {
10428 	int cpu_num;
10429 	unsigned long seq, offset, mod_num;
10430 
10431 	if (BIC_IS_ENABLED(BIC_Diec6)) {
10432 		pmt_add_counter(PMT_MTL_DC6_GUID, PMT_MTL_DC6_SEQ, "Die%c6", PMT_TYPE_XTAL_TIME,
10433 				PMT_COUNTER_MTL_DC6_LSB, PMT_COUNTER_MTL_DC6_MSB, PMT_COUNTER_MTL_DC6_OFFSET, SCOPE_PACKAGE, FORMAT_DELTA, 0, PMT_OPEN_TRY);
10434 	}
10435 
10436 	if (BIC_IS_ENABLED(BIC_CPU_c1e)) {
10437 		seq = 0;
10438 		offset = PMT_COUNTER_CWF_MC1E_OFFSET_BASE;
10439 		mod_num = 0;	/* Relative module number for current PMT file. */
10440 
10441 		/* Open the counter for each CPU. */
10442 		for (cpu_num = 0; cpu_num < topo.max_cpu_num;) {
10443 
10444 			if (cpu_is_not_allowed(cpu_num))
10445 				goto next_loop_iter;
10446 
10447 			/*
10448 			 * Set the scope to CPU, even though CWF report the counter per module.
10449 			 * CPUs inside the same module will read from the same location, instead of reporting zeros.
10450 			 *
10451 			 * CWF with newer firmware might require a PMT_TYPE_XTAL_TIME intead of PMT_TYPE_TCORE_CLOCK.
10452 			 */
10453 			pmt_add_counter(PMT_CWF_MC1E_GUID, seq, "CPU%c1e", PMT_TYPE_TCORE_CLOCK,
10454 					PMT_COUNTER_CWF_MC1E_LSB, PMT_COUNTER_CWF_MC1E_MSB, offset, SCOPE_CPU, FORMAT_DELTA, cpu_num, PMT_OPEN_TRY);
10455 
10456 			/*
10457 			 * Rather complex logic for each time we go to the next loop iteration,
10458 			 * so keep it as a label.
10459 			 */
10460 next_loop_iter:
10461 			/*
10462 			 * Advance the cpu number and check if we should also advance offset to
10463 			 * the next counter inside the PMT file.
10464 			 *
10465 			 * On Clearwater Forest platform, the counter is reported per module,
10466 			 * so open the same counter for all of the CPUs inside the module.
10467 			 * That way, reported table show the correct value for all of the CPUs inside the module,
10468 			 * instead of zeros.
10469 			 */
10470 			++cpu_num;
10471 			if (cpu_num % PMT_COUNTER_CWF_CPUS_PER_MODULE == 0) {
10472 				offset += PMT_COUNTER_CWF_MC1E_OFFSET_INCREMENT;
10473 				++mod_num;
10474 			}
10475 
10476 			/*
10477 			 * There are PMT_COUNTER_CWF_MC1E_NUM_MODULES_PER_FILE in each PMT file.
10478 			 *
10479 			 * If that number is reached, seq must be incremented to advance to the next file in a sequence.
10480 			 * Offset inside that file and a module counter has to be reset.
10481 			 */
10482 			if (mod_num == PMT_COUNTER_CWF_MC1E_NUM_MODULES_PER_FILE) {
10483 				++seq;
10484 				offset = PMT_COUNTER_CWF_MC1E_OFFSET_BASE;
10485 				mod_num = 0;
10486 			}
10487 		}
10488 	}
10489 }
10490 
turbostat_init()10491 void turbostat_init()
10492 {
10493 	setup_all_buffers(true);
10494 	set_master_cpu();
10495 	check_msr_access();
10496 	check_perf_access();
10497 	process_cpuid();
10498 	counter_info_init();
10499 	probe_pm_features();
10500 	msr_perf_init();
10501 	linux_perf_init();
10502 	rapl_perf_init();
10503 	cstate_perf_init();
10504 	perf_llc_init();
10505 	perf_l2_init();
10506 	added_perf_counters_init();
10507 	pmt_init();
10508 
10509 	for_all_cpus(get_cpu_type, ODD_COUNTERS);
10510 	for_all_cpus(get_cpu_type, EVEN_COUNTERS);
10511 
10512 	if (BIC_IS_ENABLED(BIC_IPC) && has_aperf_access && get_instr_count_fd(master_cpu) != -1)
10513 		BIC_PRESENT(BIC_IPC);
10514 
10515 	/*
10516 	 * If TSC tweak is needed, but couldn't get it,
10517 	 * disable more BICs, since it can't be reported accurately.
10518 	 */
10519 	if (platform->enable_tsc_tweak && !has_base_hz) {
10520 		CLR_BIC(BIC_Busy, &bic_enabled);
10521 		CLR_BIC(BIC_Bzy_MHz, &bic_enabled);
10522 	}
10523 }
10524 
affinitize_child(void)10525 void affinitize_child(void)
10526 {
10527 	/* Prefer cpu_possible_set, if available */
10528 	if (sched_setaffinity(0, cpu_possible_setsize, cpu_possible_set)) {
10529 		warn("sched_setaffinity cpu_possible_set");
10530 
10531 		/* Otherwise, allow child to run on same cpu set as turbostat */
10532 		if (sched_setaffinity(0, cpu_allowed_setsize, cpu_allowed_set))
10533 			warn("sched_setaffinity cpu_allowed_set");
10534 	}
10535 }
10536 
fork_it(char ** argv)10537 int fork_it(char **argv)
10538 {
10539 	pid_t child_pid;
10540 	int status;
10541 
10542 	snapshot_proc_sysfs_files();
10543 	status = for_all_cpus(get_counters, EVEN_COUNTERS);
10544 	first_counter_read = 0;
10545 	if (status)
10546 		exit(status);
10547 	gettimeofday(&tv_even, (struct timezone *)NULL);
10548 
10549 	child_pid = fork();
10550 	if (!child_pid) {
10551 		/* child */
10552 		affinitize_child();
10553 		execvp(argv[0], argv);
10554 		err(errno, "exec %s", argv[0]);
10555 	} else {
10556 
10557 		/* parent */
10558 		if (child_pid == -1)
10559 			err(1, "fork");
10560 
10561 		signal(SIGINT, SIG_IGN);
10562 		signal(SIGQUIT, SIG_IGN);
10563 		if (waitpid(child_pid, &status, 0) == -1)
10564 			err(status, "waitpid");
10565 
10566 		if (WIFEXITED(status))
10567 			status = WEXITSTATUS(status);
10568 	}
10569 	/*
10570 	 * n.b. fork_it() does not check for errors from for_all_cpus()
10571 	 * because re-starting is problematic when forking
10572 	 */
10573 	snapshot_proc_sysfs_files();
10574 	for_all_cpus(get_counters, ODD_COUNTERS);
10575 	gettimeofday(&tv_odd, (struct timezone *)NULL);
10576 	timersub(&tv_odd, &tv_even, &tv_delta);
10577 	if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS))
10578 		fprintf(outf, "%s: Counter reset detected\n", progname);
10579 	delta_platform(&platform_counters_odd, &platform_counters_even);
10580 
10581 	compute_average(EVEN_COUNTERS);
10582 	format_all_counters(EVEN_COUNTERS);
10583 
10584 	fprintf(outf, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec / 1000000.0);
10585 
10586 	flush_output_stderr();
10587 
10588 	return status;
10589 }
10590 
get_and_dump_counters(void)10591 int get_and_dump_counters(void)
10592 {
10593 	int status;
10594 
10595 	snapshot_proc_sysfs_files();
10596 	status = for_all_cpus(get_counters, ODD_COUNTERS);
10597 	if (status)
10598 		return status;
10599 
10600 	status = for_all_cpus(dump_counters, ODD_COUNTERS);
10601 	if (status)
10602 		return status;
10603 
10604 	flush_output_stdout();
10605 
10606 	return status;
10607 }
10608 
print_version()10609 void print_version()
10610 {
10611 	fprintf(outf, "turbostat version 2026.04.21 - Len Brown <lenb@kernel.org>\n");
10612 }
10613 
10614 #define COMMAND_LINE_SIZE 2048
10615 
print_bootcmd(void)10616 void print_bootcmd(void)
10617 {
10618 	char bootcmd[COMMAND_LINE_SIZE];
10619 	FILE *fp;
10620 	int ret;
10621 
10622 	memset(bootcmd, 0, COMMAND_LINE_SIZE);
10623 	fp = fopen("/proc/cmdline", "r");
10624 	if (!fp)
10625 		return;
10626 
10627 	ret = fread(bootcmd, sizeof(char), COMMAND_LINE_SIZE - 1, fp);
10628 	if (ret) {
10629 		bootcmd[ret] = '\0';
10630 		/* the last character is already '\n' */
10631 		fprintf(outf, "Kernel command line: %s", bootcmd);
10632 	}
10633 
10634 	fclose(fp);
10635 }
10636 
find_msrp_by_name(struct msr_counter * head,char * name)10637 struct msr_counter *find_msrp_by_name(struct msr_counter *head, char *name)
10638 {
10639 	struct msr_counter *mp;
10640 
10641 	for (mp = head; mp; mp = mp->next) {
10642 		if (debug)
10643 			fprintf(stderr, "%s: %s %s\n", __func__, name, mp->name);
10644 		if (!strcmp(name, mp->name))
10645 			return mp;
10646 	}
10647 	return NULL;
10648 }
10649 
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)10650 int add_counter(unsigned int msr_num, char *path, char *name,
10651 		unsigned int width, enum counter_scope scope, enum counter_type type, enum counter_format format, int flags, int id)
10652 {
10653 	struct msr_counter *msrp;
10654 
10655 	if (no_msr && msr_num)
10656 		errx(1, "Requested MSR counter 0x%x, but in --no-msr mode", msr_num);
10657 
10658 	if (debug)
10659 		fprintf(stderr, "%s(msr%d, %s, %s, width%d, scope%d, type%d, format%d, flags%x, id%d)\n",
10660 			__func__, msr_num, path, name, width, scope, type, format, flags, id);
10661 
10662 	switch (scope) {
10663 
10664 	case SCOPE_CPU:
10665 		msrp = find_msrp_by_name(sys.tp, name);
10666 		if (msrp) {
10667 			if (debug)
10668 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
10669 			break;
10670 		}
10671 		if (sys.added_thread_counters++ >= MAX_ADDED_THREAD_COUNTERS) {
10672 			warnx("ignoring thread counter %s", name);
10673 			return -1;
10674 		}
10675 		break;
10676 	case SCOPE_CORE:
10677 		msrp = find_msrp_by_name(sys.cp, name);
10678 		if (msrp) {
10679 			if (debug)
10680 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
10681 			break;
10682 		}
10683 		if (sys.added_core_counters++ >= MAX_ADDED_CORE_COUNTERS) {
10684 			warnx("ignoring core counter %s", name);
10685 			return -1;
10686 		}
10687 		break;
10688 	case SCOPE_PACKAGE:
10689 		msrp = find_msrp_by_name(sys.pp, name);
10690 		if (msrp) {
10691 			if (debug)
10692 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
10693 			break;
10694 		}
10695 		if (sys.added_package_counters++ >= MAX_ADDED_PACKAGE_COUNTERS) {
10696 			warnx("ignoring package counter %s", name);
10697 			return -1;
10698 		}
10699 		break;
10700 	default:
10701 		warnx("ignoring counter %s with unknown scope", name);
10702 		return -1;
10703 	}
10704 
10705 	if (msrp == NULL) {
10706 		msrp = calloc(1, sizeof(struct msr_counter));
10707 		if (msrp == NULL)
10708 			err(-1, "calloc msr_counter");
10709 
10710 		msrp->msr_num = msr_num;
10711 		strncpy(msrp->name, name, NAME_BYTES - 1);
10712 		msrp->width = width;
10713 		msrp->type = type;
10714 		msrp->format = format;
10715 		msrp->flags = flags;
10716 
10717 		switch (scope) {
10718 		case SCOPE_CPU:
10719 			msrp->next = sys.tp;
10720 			sys.tp = msrp;
10721 			break;
10722 		case SCOPE_CORE:
10723 			msrp->next = sys.cp;
10724 			sys.cp = msrp;
10725 			break;
10726 		case SCOPE_PACKAGE:
10727 			msrp->next = sys.pp;
10728 			sys.pp = msrp;
10729 			break;
10730 		}
10731 	}
10732 
10733 	if (path) {
10734 		struct sysfs_path *sp;
10735 
10736 		sp = calloc(1, sizeof(struct sysfs_path));
10737 		if (sp == NULL) {
10738 			perror("calloc");
10739 			exit(1);
10740 		}
10741 		strncpy(sp->path, path, PATH_BYTES - 1);
10742 		sp->id = id;
10743 		sp->next = msrp->sp;
10744 		msrp->sp = sp;
10745 	}
10746 
10747 	return 0;
10748 }
10749 
10750 /*
10751  * Initialize the fields used for identifying and opening the counter.
10752  *
10753  * Defer the initialization of any runtime buffers for actually reading
10754  * the counters for when we initialize all perf counters, so we can later
10755  * easily call re_initialize().
10756  */
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)10757 struct perf_counter_info *make_perf_counter_info(const char *perf_device,
10758 						 const char *perf_event,
10759 						 const char *name,
10760 						 unsigned int width, enum counter_scope scope, enum counter_type type, enum counter_format format)
10761 {
10762 	struct perf_counter_info *pinfo;
10763 
10764 	pinfo = calloc(1, sizeof(*pinfo));
10765 	if (!pinfo)
10766 		errx(1, "%s: Failed to allocate %s/%s\n", __func__, perf_device, perf_event);
10767 
10768 	strncpy(pinfo->device, perf_device, ARRAY_SIZE(pinfo->device) - 1);
10769 	strncpy(pinfo->event, perf_event, ARRAY_SIZE(pinfo->event) - 1);
10770 
10771 	strncpy(pinfo->name, name, ARRAY_SIZE(pinfo->name) - 1);
10772 	pinfo->width = width;
10773 	pinfo->scope = scope;
10774 	pinfo->type = type;
10775 	pinfo->format = format;
10776 
10777 	return pinfo;
10778 }
10779 
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)10780 int add_perf_counter(const char *perf_device, const char *perf_event, const char *name_buffer, unsigned int width,
10781 		     enum counter_scope scope, enum counter_type type, enum counter_format format)
10782 {
10783 	struct perf_counter_info *pinfo;
10784 
10785 	switch (scope) {
10786 	case SCOPE_CPU:
10787 		if (sys.added_thread_perf_counters >= MAX_ADDED_THREAD_COUNTERS) {
10788 			warnx("ignoring thread counter perf/%s/%s", perf_device, perf_event);
10789 			return -1;
10790 		}
10791 		break;
10792 
10793 	case SCOPE_CORE:
10794 		if (sys.added_core_perf_counters >= MAX_ADDED_CORE_COUNTERS) {
10795 			warnx("ignoring core counter perf/%s/%s", perf_device, perf_event);
10796 			return -1;
10797 		}
10798 		break;
10799 
10800 	case SCOPE_PACKAGE:
10801 		if (sys.added_package_perf_counters >= MAX_ADDED_PACKAGE_COUNTERS) {
10802 			warnx("ignoring package counter perf/%s/%s", perf_device, perf_event);
10803 			return -1;
10804 		}
10805 		break;
10806 	}
10807 
10808 	pinfo = make_perf_counter_info(perf_device, perf_event, name_buffer, width, scope, type, format);
10809 
10810 	if (!pinfo)
10811 		return -1;
10812 
10813 	switch (scope) {
10814 	case SCOPE_CPU:
10815 		pinfo->next = sys.perf_tp;
10816 		sys.perf_tp = pinfo;
10817 		++sys.added_thread_perf_counters;
10818 		break;
10819 
10820 	case SCOPE_CORE:
10821 		pinfo->next = sys.perf_cp;
10822 		sys.perf_cp = pinfo;
10823 		++sys.added_core_perf_counters;
10824 		break;
10825 
10826 	case SCOPE_PACKAGE:
10827 		pinfo->next = sys.perf_pp;
10828 		sys.perf_pp = pinfo;
10829 		++sys.added_package_perf_counters;
10830 		break;
10831 	}
10832 
10833 	// FIXME: we might not have debug here yet
10834 	if (debug)
10835 		fprintf(stderr, "%s: %s/%s, name: %s, scope%d\n", __func__, pinfo->device, pinfo->event, pinfo->name, pinfo->scope);
10836 
10837 	return 0;
10838 }
10839 
parse_add_command_msr(char * add_command)10840 void parse_add_command_msr(char *add_command)
10841 {
10842 	int msr_num = 0;
10843 	char *path = NULL;
10844 	char perf_device[PERF_DEV_NAME_BYTES] = "";
10845 	char perf_event[PERF_EVT_NAME_BYTES] = "";
10846 	char name_buffer[PERF_NAME_BYTES] = "";
10847 	int width = 64;
10848 	int fail = 0;
10849 	enum counter_scope scope = SCOPE_CPU;
10850 	enum counter_type type = COUNTER_CYCLES;
10851 	enum counter_format format = FORMAT_DELTA;
10852 
10853 	while (add_command) {
10854 
10855 		if (sscanf(add_command, "msr0x%x", &msr_num) == 1)
10856 			goto next;
10857 
10858 		if (sscanf(add_command, "msr%d", &msr_num) == 1)
10859 			goto next;
10860 
10861 		BUILD_BUG_ON(ARRAY_SIZE(perf_device) <= 31);
10862 		BUILD_BUG_ON(ARRAY_SIZE(perf_event) <= 31);
10863 		if (sscanf(add_command, "perf/%31[^/]/%31[^,]", &perf_device[0], &perf_event[0]) == 2)
10864 			goto next;
10865 
10866 		if (*add_command == '/') {
10867 			path = add_command;
10868 			goto next;
10869 		}
10870 
10871 		if (sscanf(add_command, "u%d", &width) == 1) {
10872 			if ((width == 32) || (width == 64))
10873 				goto next;
10874 			width = 64;
10875 		}
10876 		if (!strncmp(add_command, "cpu", strlen("cpu"))) {
10877 			scope = SCOPE_CPU;
10878 			goto next;
10879 		}
10880 		if (!strncmp(add_command, "core", strlen("core"))) {
10881 			scope = SCOPE_CORE;
10882 			goto next;
10883 		}
10884 		if (!strncmp(add_command, "package", strlen("package"))) {
10885 			scope = SCOPE_PACKAGE;
10886 			goto next;
10887 		}
10888 		if (!strncmp(add_command, "cycles", strlen("cycles"))) {
10889 			type = COUNTER_CYCLES;
10890 			goto next;
10891 		}
10892 		if (!strncmp(add_command, "seconds", strlen("seconds"))) {
10893 			type = COUNTER_SECONDS;
10894 			goto next;
10895 		}
10896 		if (!strncmp(add_command, "usec", strlen("usec"))) {
10897 			type = COUNTER_USEC;
10898 			goto next;
10899 		}
10900 		if (!strncmp(add_command, "raw", strlen("raw"))) {
10901 			format = FORMAT_RAW;
10902 			goto next;
10903 		}
10904 		if (!strncmp(add_command, "average", strlen("average"))) {
10905 			format = FORMAT_AVERAGE;
10906 			goto next;
10907 		}
10908 		if (!strncmp(add_command, "delta", strlen("delta"))) {
10909 			format = FORMAT_DELTA;
10910 			goto next;
10911 		}
10912 		if (!strncmp(add_command, "percent", strlen("percent"))) {
10913 			format = FORMAT_PERCENT;
10914 			goto next;
10915 		}
10916 
10917 		BUILD_BUG_ON(ARRAY_SIZE(name_buffer) <= 18);
10918 		if (sscanf(add_command, "%18s,%*s", name_buffer) == 1) {
10919 			char *eos;
10920 
10921 			eos = strchr(name_buffer, ',');
10922 			if (eos)
10923 				*eos = '\0';
10924 			goto next;
10925 		}
10926 
10927 next:
10928 		add_command = strchr(add_command, ',');
10929 		if (add_command) {
10930 			*add_command = '\0';
10931 			add_command++;
10932 		}
10933 
10934 	}
10935 	if ((msr_num == 0) && (path == NULL) && (perf_device[0] == '\0' || perf_event[0] == '\0')) {
10936 		fprintf(stderr, "--add: (msrDDD | msr0xXXX | /path_to_counter | perf/device/event) required\n");
10937 		fail++;
10938 	}
10939 
10940 	/* Test for non-empty perf_device and perf_event */
10941 	const bool is_perf_counter = perf_device[0] && perf_event[0];
10942 
10943 	/* generate default column header */
10944 	if (*name_buffer == '\0') {
10945 		if (is_perf_counter) {
10946 			snprintf(name_buffer, ARRAY_SIZE(name_buffer), "perf/%s", perf_event);
10947 		} else {
10948 			if (width == 32)
10949 				sprintf(name_buffer, "M0x%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
10950 			else
10951 				sprintf(name_buffer, "M0X%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
10952 		}
10953 	}
10954 
10955 	if (is_perf_counter) {
10956 		if (add_perf_counter(perf_device, perf_event, name_buffer, width, scope, type, format))
10957 			fail++;
10958 	} else {
10959 		if (add_counter(msr_num, path, name_buffer, width, scope, type, format, 0, 0))
10960 			fail++;
10961 	}
10962 
10963 	if (fail) {
10964 		help();
10965 		exit(1);
10966 	}
10967 }
10968 
starts_with(const char * str,const char * prefix)10969 bool starts_with(const char *str, const char *prefix)
10970 {
10971 	return strncmp(prefix, str, strlen(prefix)) == 0;
10972 }
10973 
pmt_parse_from_path(const char * target_path,unsigned int * out_guid,unsigned int * out_seq)10974 int pmt_parse_from_path(const char *target_path, unsigned int *out_guid, unsigned int *out_seq)
10975 {
10976 	struct pmt_diriter_t pmt_iter;
10977 	const struct dirent *dirname;
10978 	struct stat stat, target_stat;
10979 	int fd_telem_dir = -1;
10980 	int fd_target_dir;
10981 	unsigned int seq = 0;
10982 	unsigned long guid, target_guid;
10983 	int ret = -1;
10984 
10985 	fd_target_dir = open(target_path, O_RDONLY | O_DIRECTORY);
10986 	if (fd_target_dir == -1) {
10987 		return -1;
10988 	}
10989 
10990 	if (fstat(fd_target_dir, &target_stat) == -1) {
10991 		fprintf(stderr, "%s: Failed to stat the target: %s", __func__, strerror(errno));
10992 		exit(1);
10993 	}
10994 
10995 	if (parse_telem_info_file(fd_target_dir, "guid", "%lx", &target_guid)) {
10996 		fprintf(stderr, "%s: Failed to parse the target guid file: %s", __func__, strerror(errno));
10997 		exit(1);
10998 	}
10999 
11000 	close(fd_target_dir);
11001 
11002 	pmt_diriter_init(&pmt_iter);
11003 
11004 	for (dirname = pmt_diriter_begin(&pmt_iter, SYSFS_TELEM_PATH); dirname != NULL; dirname = pmt_diriter_next(&pmt_iter)) {
11005 
11006 		fd_telem_dir = openat(dirfd(pmt_iter.dir), dirname->d_name, O_RDONLY | O_DIRECTORY);
11007 		if (fd_telem_dir == -1)
11008 			continue;
11009 
11010 		if (parse_telem_info_file(fd_telem_dir, "guid", "%lx", &guid)) {
11011 			fprintf(stderr, "%s: Failed to parse the guid file: %s", __func__, strerror(errno));
11012 			continue;
11013 		}
11014 
11015 		if (fstat(fd_telem_dir, &stat) == -1) {
11016 			fprintf(stderr, "%s: Failed to stat %s directory: %s", __func__, dirname->d_name, strerror(errno));
11017 			continue;
11018 		}
11019 
11020 		/*
11021 		 * If reached the same directory as target, exit the loop.
11022 		 * Seq has the correct value now.
11023 		 */
11024 		if (stat.st_dev == target_stat.st_dev && stat.st_ino == target_stat.st_ino) {
11025 			ret = 0;
11026 			break;
11027 		}
11028 
11029 		/*
11030 		 * If reached directory with the same guid,
11031 		 * but it's not the target directory yet,
11032 		 * increment seq and continue the search.
11033 		 */
11034 		if (guid == target_guid)
11035 			++seq;
11036 
11037 		close(fd_telem_dir);
11038 		fd_telem_dir = -1;
11039 	}
11040 
11041 	pmt_diriter_remove(&pmt_iter);
11042 
11043 	if (fd_telem_dir != -1)
11044 		close(fd_telem_dir);
11045 
11046 	if (!ret) {
11047 		*out_guid = target_guid;
11048 		*out_seq = seq;
11049 	}
11050 
11051 	return ret;
11052 }
11053 
parse_add_command_pmt(char * add_command)11054 void parse_add_command_pmt(char *add_command)
11055 {
11056 	char *name = NULL;
11057 	char *type_name = NULL;
11058 	char *format_name = NULL;
11059 	char *direct_path = NULL;
11060 	static const char direct_path_prefix[] = "path=";
11061 	unsigned int offset;
11062 	unsigned int lsb;
11063 	unsigned int msb;
11064 	unsigned int guid;
11065 	unsigned int seq = 0;	/* By default, pick first file in a sequence with a given GUID. */
11066 	unsigned int domain_id;
11067 	enum counter_scope scope = 0;
11068 	enum pmt_datatype type = PMT_TYPE_RAW;
11069 	enum counter_format format = FORMAT_RAW;
11070 	bool has_offset = false;
11071 	bool has_lsb = false;
11072 	bool has_msb = false;
11073 	bool has_format = true;	/* Format has a default value. */
11074 	bool has_guid = false;
11075 	bool has_scope = false;
11076 	bool has_type = true;	/* Type has a default value. */
11077 
11078 	/* Consume the "pmt," prefix. */
11079 	add_command = strchr(add_command, ',');
11080 	if (!add_command) {
11081 		help();
11082 		exit(1);
11083 	}
11084 	++add_command;
11085 
11086 	while (add_command) {
11087 		if (starts_with(add_command, "name=")) {
11088 			name = add_command + strlen("name=");
11089 			goto next;
11090 		}
11091 
11092 		if (starts_with(add_command, "type=")) {
11093 			type_name = add_command + strlen("type=");
11094 			goto next;
11095 		}
11096 
11097 		if (starts_with(add_command, "domain=")) {
11098 			const size_t prefix_len = strlen("domain=");
11099 
11100 			if (sscanf(add_command + prefix_len, "cpu%u", &domain_id) == 1) {
11101 				scope = SCOPE_CPU;
11102 				has_scope = true;
11103 			} else if (sscanf(add_command + prefix_len, "core%u", &domain_id) == 1) {
11104 				scope = SCOPE_CORE;
11105 				has_scope = true;
11106 			} else if (sscanf(add_command + prefix_len, "package%u", &domain_id) == 1) {
11107 				scope = SCOPE_PACKAGE;
11108 				has_scope = true;
11109 			}
11110 
11111 			if (!has_scope) {
11112 				printf("%s: invalid value for scope. Expected cpu%%u, core%%u or package%%u.\n", __func__);
11113 				exit(1);
11114 			}
11115 
11116 			goto next;
11117 		}
11118 
11119 		if (starts_with(add_command, "format=")) {
11120 			format_name = add_command + strlen("format=");
11121 			goto next;
11122 		}
11123 
11124 		if (sscanf(add_command, "offset=%u", &offset) == 1) {
11125 			has_offset = true;
11126 			goto next;
11127 		}
11128 
11129 		if (sscanf(add_command, "lsb=%u", &lsb) == 1) {
11130 			has_lsb = true;
11131 			goto next;
11132 		}
11133 
11134 		if (sscanf(add_command, "msb=%u", &msb) == 1) {
11135 			has_msb = true;
11136 			goto next;
11137 		}
11138 
11139 		if (sscanf(add_command, "guid=%x", &guid) == 1) {
11140 			has_guid = true;
11141 			goto next;
11142 		}
11143 
11144 		if (sscanf(add_command, "seq=%x", &seq) == 1)
11145 			goto next;
11146 
11147 		if (strncmp(add_command, direct_path_prefix, strlen(direct_path_prefix)) == 0) {
11148 			direct_path = add_command + strlen(direct_path_prefix);
11149 			goto next;
11150 		}
11151 next:
11152 		add_command = strchr(add_command, ',');
11153 		if (add_command) {
11154 			*add_command = '\0';
11155 			add_command++;
11156 		}
11157 	}
11158 
11159 	if (!name) {
11160 		printf("%s: missing %s\n", __func__, "name");
11161 		exit(1);
11162 	}
11163 
11164 	if (strlen(name) >= PMT_COUNTER_NAME_SIZE_BYTES) {
11165 		printf("%s: name has to be at most %d characters long\n", __func__, PMT_COUNTER_NAME_SIZE_BYTES);
11166 		exit(1);
11167 	}
11168 
11169 	if (format_name) {
11170 		has_format = false;
11171 
11172 		if (strcmp("raw", format_name) == 0) {
11173 			format = FORMAT_RAW;
11174 			has_format = true;
11175 		}
11176 
11177 		if (strcmp("average", format_name) == 0) {
11178 			format = FORMAT_AVERAGE;
11179 			has_format = true;
11180 		}
11181 
11182 		if (strcmp("delta", format_name) == 0) {
11183 			format = FORMAT_DELTA;
11184 			has_format = true;
11185 		}
11186 
11187 		if (!has_format) {
11188 			fprintf(stderr, "%s: Invalid format %s. Expected raw, average or delta\n", __func__, format_name);
11189 			exit(1);
11190 		}
11191 	}
11192 
11193 	if (type_name) {
11194 		has_type = false;
11195 
11196 		if (strcmp("raw", type_name) == 0) {
11197 			type = PMT_TYPE_RAW;
11198 			has_type = true;
11199 		}
11200 
11201 		if (strcmp("txtal_time", type_name) == 0) {
11202 			type = PMT_TYPE_XTAL_TIME;
11203 			has_type = true;
11204 		}
11205 
11206 		if (strcmp("tcore_clock", type_name) == 0) {
11207 			type = PMT_TYPE_TCORE_CLOCK;
11208 			has_type = true;
11209 		}
11210 
11211 		if (!has_type) {
11212 			printf("%s: invalid %s: %s\n", __func__, "type", type_name);
11213 			exit(1);
11214 		}
11215 	}
11216 
11217 	if (!has_offset) {
11218 		printf("%s : missing %s\n", __func__, "offset");
11219 		exit(1);
11220 	}
11221 
11222 	if (!has_lsb) {
11223 		printf("%s: missing %s\n", __func__, "lsb");
11224 		exit(1);
11225 	}
11226 
11227 	if (!has_msb) {
11228 		printf("%s: missing %s\n", __func__, "msb");
11229 		exit(1);
11230 	}
11231 
11232 	if (direct_path && has_guid) {
11233 		printf("%s: path and guid+seq parameters are mutually exclusive\nnotice: passed guid=0x%x and path=%s\n", __func__, guid, direct_path);
11234 		exit(1);
11235 	}
11236 
11237 	if (direct_path) {
11238 		if (pmt_parse_from_path(direct_path, &guid, &seq)) {
11239 			printf("%s: failed to parse PMT file from %s\n", __func__, direct_path);
11240 			exit(1);
11241 		}
11242 
11243 		/* GUID was just infered from the direct path. */
11244 		has_guid = true;
11245 	}
11246 
11247 	if (!has_guid) {
11248 		printf("%s: missing %s\n", __func__, "guid or path");
11249 		exit(1);
11250 	}
11251 
11252 	if (!has_scope) {
11253 		printf("%s: missing %s\n", __func__, "scope");
11254 		exit(1);
11255 	}
11256 
11257 	if (lsb > msb) {
11258 		printf("%s: lsb > msb doesn't make sense\n", __func__);
11259 		exit(1);
11260 	}
11261 
11262 	pmt_add_counter(guid, seq, name, type, lsb, msb, offset, scope, format, domain_id, PMT_OPEN_REQUIRED);
11263 }
11264 
parse_add_command(char * add_command)11265 void parse_add_command(char *add_command)
11266 {
11267 	if (strncmp(add_command, "pmt", strlen("pmt")) == 0)
11268 		return parse_add_command_pmt(add_command);
11269 	return parse_add_command_msr(add_command);
11270 }
11271 
is_deferred_add(char * name)11272 int is_deferred_add(char *name)
11273 {
11274 	int i;
11275 
11276 	for (i = 0; i < deferred_add_index; ++i)
11277 		if (!strcmp(name, deferred_add_names[i])) {
11278 			deferred_add_consumed |= (1 << i);
11279 			return 1;
11280 		}
11281 	return 0;
11282 }
11283 
is_deferred_skip(char * name)11284 int is_deferred_skip(char *name)
11285 {
11286 	int i;
11287 
11288 	for (i = 0; i < deferred_skip_index; ++i)
11289 		if (!strcmp(name, deferred_skip_names[i])) {
11290 			deferred_skip_consumed |= (1 << i);
11291 			return 1;
11292 		}
11293 	return 0;
11294 }
11295 
verify_deferred_consumed(void)11296 void verify_deferred_consumed(void)
11297 {
11298 	int i;
11299 	int fail = 0;
11300 
11301 	for (i = 0; i < deferred_add_index; ++i) {
11302 		if (!(deferred_add_consumed & (1 << i))) {
11303 			warnx("Counter '%s' can not be added.", deferred_add_names[i]);
11304 			fail++;
11305 		}
11306 	}
11307 	for (i = 0; i < deferred_skip_index; ++i) {
11308 		if (!(deferred_skip_consumed & (1 << i))) {
11309 			warnx("Counter '%s' can not be skipped.", deferred_skip_names[i]);
11310 			fail++;
11311 		}
11312 	}
11313 	if (fail)
11314 		exit(-EINVAL);
11315 }
11316 
probe_cpuidle_residency(void)11317 void probe_cpuidle_residency(void)
11318 {
11319 	char path[64];
11320 	char name_buf[16];
11321 	FILE *input;
11322 	int state;
11323 	int min_state = 1024, max_state = 0;
11324 	char *sp;
11325 
11326 	for (state = 10; state >= 0; --state) {
11327 
11328 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", master_cpu, state);
11329 		input = fopen(path, "r");
11330 		if (input == NULL)
11331 			continue;
11332 		if (!fgets(name_buf, sizeof(name_buf), input))
11333 			err(1, "%s: failed to read file", path);
11334 
11335 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
11336 		sp = strchr(name_buf, '-');
11337 		if (!sp)
11338 			sp = strchrnul(name_buf, '\n');
11339 		*sp = '%';
11340 		*(sp + 1) = '\0';
11341 
11342 		remove_underbar(name_buf);
11343 
11344 		fclose(input);
11345 
11346 		sprintf(path, "cpuidle/state%d/time", state);
11347 
11348 		if (!DO_BIC(BIC_pct_idle) && !is_deferred_add(name_buf))
11349 			continue;
11350 
11351 		if (is_deferred_skip(name_buf))
11352 			continue;
11353 
11354 		add_counter(0, path, name_buf, 32, SCOPE_CPU, COUNTER_USEC, FORMAT_PERCENT, SYSFS_PERCPU, 0);
11355 
11356 		if (state > max_state)
11357 			max_state = state;
11358 		if (state < min_state)
11359 			min_state = state;
11360 	}
11361 }
11362 
cpuidle_counter_wanted(char * name)11363 static bool cpuidle_counter_wanted(char *name)
11364 {
11365 	if (is_deferred_skip(name))
11366 		return false;
11367 
11368 	return DO_BIC(BIC_cpuidle) || is_deferred_add(name);
11369 }
11370 
probe_cpuidle_counts(void)11371 void probe_cpuidle_counts(void)
11372 {
11373 	char path[64];
11374 	char name_buf[16];
11375 	FILE *input;
11376 	int state;
11377 	int min_state = 1024, max_state = 0;
11378 	char *sp;
11379 
11380 	if (!DO_BIC(BIC_cpuidle) && !deferred_add_index)
11381 		return;
11382 
11383 	for (state = 10; state >= 0; --state) {
11384 
11385 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", master_cpu, state);
11386 		input = fopen(path, "r");
11387 		if (input == NULL)
11388 			continue;
11389 		if (!fgets(name_buf, sizeof(name_buf), input))
11390 			err(1, "%s: failed to read file", path);
11391 		fclose(input);
11392 
11393 		remove_underbar(name_buf);
11394 
11395 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
11396 		sp = strchr(name_buf, '-');
11397 		if (!sp)
11398 			sp = strchrnul(name_buf, '\n');
11399 
11400 		/*
11401 		 * The 'below' sysfs file always contains 0 for the deepest state (largest index),
11402 		 * do not add it.
11403 		 */
11404 		if (state != max_state) {
11405 			/*
11406 			 * Add 'C1+' for C1, and so on. The 'below' sysfs file always contains 0 for
11407 			 * the last state, so do not add it.
11408 			 */
11409 			*sp = '+';
11410 			*(sp + 1) = '\0';
11411 			if (cpuidle_counter_wanted(name_buf)) {
11412 				sprintf(path, "cpuidle/state%d/below", state);
11413 				add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS, FORMAT_DELTA, SYSFS_PERCPU, 0);
11414 			}
11415 		}
11416 
11417 		*sp = '\0';
11418 		if (cpuidle_counter_wanted(name_buf)) {
11419 			sprintf(path, "cpuidle/state%d/usage", state);
11420 			add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS, FORMAT_DELTA, SYSFS_PERCPU, 0);
11421 		}
11422 
11423 		/*
11424 		 * The 'above' sysfs file always contains 0 for the shallowest state (smallest
11425 		 * index), do not add it.
11426 		 */
11427 		if (state != min_state) {
11428 			*sp = '-';
11429 			*(sp + 1) = '\0';
11430 			if (cpuidle_counter_wanted(name_buf)) {
11431 				sprintf(path, "cpuidle/state%d/above", state);
11432 				add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS, FORMAT_DELTA, SYSFS_PERCPU, 0);
11433 			}
11434 		}
11435 	}
11436 }
11437 
11438 /*
11439  * parse cpuset with following syntax
11440  * 1,2,4..6,8-10 and set bits in cpu_subset
11441  */
parse_cpu_command(char * optarg)11442 void parse_cpu_command(char *optarg)
11443 {
11444 	if (!strcmp(optarg, "core")) {
11445 		if (cpu_subset)
11446 			goto error;
11447 		show_core_only++;
11448 		return;
11449 	}
11450 	if (!strcmp(optarg, "package")) {
11451 		if (cpu_subset)
11452 			goto error;
11453 		show_pkg_only++;
11454 		return;
11455 	}
11456 	if (show_core_only || show_pkg_only)
11457 		goto error;
11458 
11459 	cpu_subset = CPU_ALLOC(CPU_SUBSET_MAXCPUS);
11460 	if (cpu_subset == NULL)
11461 		err(3, "CPU_ALLOC");
11462 	cpu_subset_size = CPU_ALLOC_SIZE(CPU_SUBSET_MAXCPUS);
11463 
11464 	CPU_ZERO_S(cpu_subset_size, cpu_subset);
11465 
11466 	if (parse_cpu_str(optarg, cpu_subset, cpu_subset_size))
11467 		goto error;
11468 
11469 	return;
11470 
11471 error:
11472 	fprintf(stderr, "\"--cpu %s\" malformed\n", optarg);
11473 	help();
11474 	exit(-1);
11475 }
11476 
cmdline(int argc,char ** argv)11477 void cmdline(int argc, char **argv)
11478 {
11479 	int opt;
11480 	int option_index = 0;
11481 	static struct option long_options[] = {
11482 		{ "add", required_argument, 0, 'a' },
11483 		{ "cpu", required_argument, 0, 'c' },
11484 		{ "Dump", no_argument, 0, 'D' },
11485 		{ "debug", no_argument, 0, 'd' },	/* internal, not documented */
11486 		{ "enable", required_argument, 0, 'e' },
11487 		{ "force", no_argument, 0, 'f' },
11488 		{ "interval", required_argument, 0, 'i' },
11489 		{ "IPC", no_argument, 0, 'I' },
11490 		{ "num_iterations", required_argument, 0, 'n' },
11491 		{ "header_iterations", required_argument, 0, 'N' },
11492 		{ "help", no_argument, 0, 'h' },
11493 		{ "hide", required_argument, 0, 'H' },	// meh, -h taken by --help
11494 		{ "Joules", no_argument, 0, 'J' },
11495 		{ "list", no_argument, 0, 'l' },
11496 		{ "out", required_argument, 0, 'o' },
11497 		{ "quiet", no_argument, 0, 'q' },
11498 		{ "no-msr", no_argument, 0, 'M' },
11499 		{ "no-perf", no_argument, 0, 'P' },
11500 		{ "show", required_argument, 0, 's' },
11501 		{ "Summary", no_argument, 0, 'S' },
11502 		{ "TCC", required_argument, 0, 'T' },
11503 		{ "version", no_argument, 0, 'v' },
11504 		{ 0, 0, 0, 0 }
11505 	};
11506 
11507 	progname = argv[0];
11508 
11509 	/*
11510 	 * Parse some options early, because they may make other options invalid,
11511 	 * like adding the MSR counter with --add and at the same time using --no-msr.
11512 	 */
11513 	while ((opt = getopt_long_only(argc, argv, "+:MP", long_options, &option_index)) != -1) {
11514 		switch (opt) {
11515 		case 'M':
11516 			no_msr = 1;
11517 			break;
11518 		case 'P':
11519 			no_perf = 1;
11520 			break;
11521 		default:
11522 			break;
11523 		}
11524 	}
11525 	optind = 0;
11526 
11527 	while ((opt = getopt_long_only(argc, argv, "+C:c:Dde:hi:Jn:N:o:qMPST:v", long_options, &option_index)) != -1) {
11528 		switch (opt) {
11529 		case 'a':
11530 			parse_add_command(optarg);
11531 			break;
11532 		case 'c':
11533 			parse_cpu_command(optarg);
11534 			break;
11535 		case 'D':
11536 			dump_only++;
11537 			/*
11538 			 * Force the no_perf early to prevent using it as a source.
11539 			 * User asks for raw values, but perf returns them relative
11540 			 * to the opening of the file descriptor.
11541 			 */
11542 			no_perf = 1;
11543 			break;
11544 		case 'e':
11545 			/* --enable specified counter, without clearning existing list */
11546 			bic_lookup(&bic_enabled, optarg, SHOW_LIST);
11547 			break;
11548 		case 'f':
11549 			force_load++;
11550 			break;
11551 		case 'd':
11552 			debug++;
11553 			bic_set_all(&bic_enabled);
11554 			break;
11555 		case 'H':
11556 			/*
11557 			 * --hide: do not show those specified
11558 			 *  multiple invocations simply clear more bits in enabled mask
11559 			 */
11560 			{
11561 				cpu_set_t bic_group_hide;
11562 
11563 				BIC_INIT(&bic_group_hide);
11564 
11565 				bic_lookup(&bic_group_hide, optarg, HIDE_LIST);
11566 				bic_clear_bits(&bic_enabled, &bic_group_hide);
11567 			}
11568 			break;
11569 		case 'h':
11570 			help();
11571 			exit(1);
11572 		case 'i':
11573 			{
11574 				double interval = strtod(optarg, NULL);
11575 
11576 				if (interval < 0.001) {
11577 					fprintf(outf, "interval %f seconds is too small\n", interval);
11578 					exit(2);
11579 				}
11580 
11581 				interval_tv.tv_sec = interval_ts.tv_sec = interval;
11582 				interval_tv.tv_usec = (interval - interval_tv.tv_sec) * 1000000;
11583 				interval_ts.tv_nsec = (interval - interval_ts.tv_sec) * 1000000000;
11584 			}
11585 			break;
11586 		case 'J':
11587 			rapl_joules++;
11588 			break;
11589 		case 'l':
11590 			bic_set_all(&bic_enabled);
11591 			list_header_only++;
11592 			quiet++;
11593 			break;
11594 		case 'o':
11595 			outf = fopen_or_die(optarg, "w");
11596 			break;
11597 		case 'q':
11598 			quiet = 1;
11599 			break;
11600 		case 'M':
11601 		case 'P':
11602 			/* Parsed earlier */
11603 			break;
11604 		case 'n':
11605 			num_iterations = strtoul(optarg, NULL, 0);
11606 			errno = 0;
11607 
11608 			if (errno || num_iterations == 0)
11609 				errx(-1, "invalid iteration count: %s", optarg);
11610 			break;
11611 		case 'N':
11612 			header_iterations = strtoul(optarg, NULL, 0);
11613 			errno = 0;
11614 
11615 			if (errno || header_iterations == 0)
11616 				errx(-1, "invalid header iteration count: %s", optarg);
11617 			break;
11618 		case 's':
11619 			/*
11620 			 * --show: show only those specified
11621 			 *  The 1st invocation will clear and replace the enabled mask
11622 			 *  subsequent invocations can add to it.
11623 			 */
11624 			if (shown == 0)
11625 				BIC_INIT(&bic_enabled);
11626 			bic_lookup(&bic_enabled, optarg, SHOW_LIST);
11627 			shown = 1;
11628 			break;
11629 		case 'S':
11630 			summary_only++;
11631 			break;
11632 		case 'T':
11633 			tj_max_override = atoi(optarg);
11634 			break;
11635 		case 'v':
11636 			print_version();
11637 			exit(0);
11638 			break;
11639 		default:
11640 			help();
11641 			exit(1);
11642 		}
11643 	}
11644 }
11645 
set_rlimit(void)11646 void set_rlimit(void)
11647 {
11648 	struct rlimit limit;
11649 
11650 	if (getrlimit(RLIMIT_NOFILE, &limit) < 0)
11651 		err(1, "Failed to get rlimit");
11652 
11653 	if (limit.rlim_max < MAX_NOFILE)
11654 		limit.rlim_max = MAX_NOFILE;
11655 	if (limit.rlim_cur < MAX_NOFILE)
11656 		limit.rlim_cur = MAX_NOFILE;
11657 
11658 	if (setrlimit(RLIMIT_NOFILE, &limit) < 0)
11659 		err(1, "Failed to set rlimit");
11660 }
11661 
main(int argc,char ** argv)11662 int main(int argc, char **argv)
11663 {
11664 	int fd, ret;
11665 
11666 	bic_groups_init();
11667 
11668 	fd = open("/sys/fs/cgroup/cgroup.procs", O_WRONLY);
11669 	if (fd < 0)
11670 		goto skip_cgroup_setting;
11671 
11672 	ret = write(fd, "0\n", 2);
11673 	if (ret == -1)
11674 		perror("Can't update cgroup\n");
11675 
11676 	close(fd);
11677 
11678 skip_cgroup_setting:
11679 	outf = stderr;
11680 	cmdline(argc, argv);
11681 
11682 	if (!quiet) {
11683 		print_version();
11684 		print_bootcmd();
11685 	}
11686 
11687 	probe_cpuidle_residency();
11688 	probe_cpuidle_counts();
11689 
11690 	verify_deferred_consumed();
11691 
11692 	if (!getuid())
11693 		set_rlimit();
11694 
11695 	turbostat_init();
11696 
11697 	if (!no_msr)
11698 		msr_sum_record();
11699 
11700 	/* dump counters and exit */
11701 	if (dump_only)
11702 		return get_and_dump_counters();
11703 
11704 	/* list header and exit */
11705 	if (list_header_only) {
11706 		print_header(",");
11707 		flush_output_stdout();
11708 		return 0;
11709 	}
11710 
11711 	/*
11712 	 * if any params left, it must be a command to fork
11713 	 */
11714 	if (argc - optind)
11715 		return fork_it(argv + optind);
11716 	else
11717 		turbostat_loop();
11718 
11719 	return 0;
11720 }
11721