xref: /linux/tools/power/x86/turbostat/turbostat.c (revision 7ad54bbbc9c512ba3bc90e4368264bcf15c25759)
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, "Core", NULL, 0, 0, 0, NULL, 0 },
195 	{ 0x0, "CPU", NULL, 0, 0, 0, NULL, 0 },
196 	{ 0x0, "APIC", NULL, 0, 0, 0, NULL, 0 },
197 	{ 0x0, "X2APIC", NULL, 0, 0, 0, NULL, 0 },
198 	{ 0x0, "Die", NULL, 0, 0, 0, NULL, 0 },
199 	{ 0x0, "L3", NULL, 0, 0, 0, NULL, 0 },
200 	{ 0x0, "GFXAMHz", NULL, 0, 0, 0, NULL, 0 },
201 	{ 0x0, "IPC", NULL, 0, 0, 0, NULL, 0 },
202 	{ 0x0, "CoreThr", NULL, 0, 0, 0, NULL, 0 },
203 	{ 0x0, "UncMHz", NULL, 0, 0, 0, NULL, 0 },
204 	{ 0x0, "SAM%mc6", NULL, 0, 0, 0, NULL, 0 },
205 	{ 0x0, "SAMMHz", NULL, 0, 0, 0, NULL, 0 },
206 	{ 0x0, "SAMAMHz", NULL, 0, 0, 0, NULL, 0 },
207 	{ 0x0, "Die%c6", NULL, 0, 0, 0, NULL, 0 },
208 	{ 0x0, "SysWatt", NULL, 0, 0, 0, NULL, 0 },
209 	{ 0x0, "Sys_J", NULL, 0, 0, 0, NULL, 0 },
210 	{ 0x0, "NMI", NULL, 0, 0, 0, NULL, 0 },
211 	{ 0x0, "CPU%c1e", NULL, 0, 0, 0, NULL, 0 },
212 	{ 0x0, "pct_idle", NULL, 0, 0, 0, NULL, 0 },
213 	{ 0x0, "LLCMRPS", NULL, 0, 0, 0, NULL, 0 },
214 	{ 0x0, "LLC%hit", NULL, 0, 0, 0, NULL, 0 },
215 	{ 0x0, "L2MRPS", NULL, 0, 0, 0, NULL, 0 },
216 	{ 0x0, "L2%hit", NULL, 0, 0, 0, NULL, 0 },
217 };
218 
219 /* n.b. bic_names must match the order in bic[], above */
220 enum bic_names {
221 	BIC_USEC,
222 	BIC_TOD,
223 	BIC_Package,
224 	BIC_Node,
225 	BIC_Avg_MHz,
226 	BIC_Busy,
227 	BIC_Bzy_MHz,
228 	BIC_TSC_MHz,
229 	BIC_IRQ,
230 	BIC_SMI,
231 	BIC_cpuidle,
232 	BIC_CPU_c1,
233 	BIC_CPU_c3,
234 	BIC_CPU_c6,
235 	BIC_CPU_c7,
236 	BIC_ThreadC,
237 	BIC_CoreTmp,
238 	BIC_CoreCnt,
239 	BIC_PkgTmp,
240 	BIC_GFX_rc6,
241 	BIC_GFXMHz,
242 	BIC_Pkgpc2,
243 	BIC_Pkgpc3,
244 	BIC_Pkgpc6,
245 	BIC_Pkgpc7,
246 	BIC_Pkgpc8,
247 	BIC_Pkgpc9,
248 	BIC_Pkgpc10,
249 	BIC_CPU_LPI,
250 	BIC_SYS_LPI,
251 	BIC_PkgWatt,
252 	BIC_CorWatt,
253 	BIC_GFXWatt,
254 	BIC_PkgCnt,
255 	BIC_RAMWatt,
256 	BIC_PKG__,
257 	BIC_RAM__,
258 	BIC_Pkg_J,
259 	BIC_Cor_J,
260 	BIC_GFX_J,
261 	BIC_RAM_J,
262 	BIC_Mod_c6,
263 	BIC_Totl_c0,
264 	BIC_Any_c0,
265 	BIC_GFX_c0,
266 	BIC_CPUGFX,
267 	BIC_Core,
268 	BIC_CPU,
269 	BIC_APIC,
270 	BIC_X2APIC,
271 	BIC_Die,
272 	BIC_L3,
273 	BIC_GFXACTMHz,
274 	BIC_IPC,
275 	BIC_CORE_THROT_CNT,
276 	BIC_UNCORE_MHZ,
277 	BIC_SAM_mc6,
278 	BIC_SAMMHz,
279 	BIC_SAMACTMHz,
280 	BIC_Diec6,
281 	BIC_SysWatt,
282 	BIC_Sys_J,
283 	BIC_NMI,
284 	BIC_CPU_c1e,
285 	BIC_pct_idle,
286 	BIC_LLC_MRPS,
287 	BIC_LLC_HIT,
288 	BIC_L2_MRPS,
289 	BIC_L2_HIT,
290 	MAX_BIC
291 };
292 
print_bic_set(char * s,cpu_set_t * set)293 void print_bic_set(char *s, cpu_set_t *set)
294 {
295 	int i;
296 
297 	assert(MAX_BIC < CPU_SETSIZE);
298 
299 	printf("%s:", s);
300 
301 	for (i = 0; i < MAX_BIC; ++i) {
302 
303 		if (CPU_ISSET(i, set))
304 			printf(" %s", bic[i].name);
305 	}
306 	putchar('\n');
307 }
308 
309 static cpu_set_t bic_group_topology;
310 static cpu_set_t bic_group_thermal_pwr;
311 static cpu_set_t bic_group_frequency;
312 static cpu_set_t bic_group_hw_idle;
313 static cpu_set_t bic_group_sw_idle;
314 static cpu_set_t bic_group_idle;
315 static cpu_set_t bic_group_cache;
316 static cpu_set_t bic_group_other;
317 static cpu_set_t bic_group_disabled_by_default;
318 static cpu_set_t bic_enabled;
319 static cpu_set_t bic_present;
320 
321 /* modify */
322 #define BIC_INIT(set) CPU_ZERO(set)
323 
324 #define SET_BIC(COUNTER_NUMBER, set) CPU_SET(COUNTER_NUMBER, set)
325 #define CLR_BIC(COUNTER_NUMBER, set) CPU_CLR(COUNTER_NUMBER, set)
326 
327 #define BIC_PRESENT(COUNTER_NUMBER) SET_BIC(COUNTER_NUMBER, &bic_present)
328 #define BIC_NOT_PRESENT(COUNTER_NUMBER) CPU_CLR(COUNTER_NUMBER, &bic_present)
329 
330 /* test */
331 #define BIC_IS_ENABLED(COUNTER_NUMBER) CPU_ISSET(COUNTER_NUMBER, &bic_enabled)
332 #define DO_BIC_READ(COUNTER_NUMBER) CPU_ISSET(COUNTER_NUMBER, &bic_present)
333 #define DO_BIC(COUNTER_NUMBER) (CPU_ISSET(COUNTER_NUMBER, &bic_enabled) && CPU_ISSET(COUNTER_NUMBER, &bic_present))
334 
bic_set_all(cpu_set_t * set)335 static void bic_set_all(cpu_set_t *set)
336 {
337 	int i;
338 
339 	assert(MAX_BIC < CPU_SETSIZE);
340 
341 	for (i = 0; i < MAX_BIC; ++i)
342 		SET_BIC(i, set);
343 }
344 
345 /*
346  * bic_clear_bits()
347  * clear all the bits from "clr" in "dst"
348  */
bic_clear_bits(cpu_set_t * dst,cpu_set_t * clr)349 static void bic_clear_bits(cpu_set_t *dst, cpu_set_t *clr)
350 {
351 	int i;
352 
353 	assert(MAX_BIC < CPU_SETSIZE);
354 
355 	for (i = 0; i < MAX_BIC; ++i)
356 		if (CPU_ISSET(i, clr))
357 			CLR_BIC(i, dst);
358 }
359 
bic_groups_init(void)360 static void bic_groups_init(void)
361 {
362 	BIC_INIT(&bic_group_topology);
363 	SET_BIC(BIC_Package, &bic_group_topology);
364 	SET_BIC(BIC_Node, &bic_group_topology);
365 	SET_BIC(BIC_CoreCnt, &bic_group_topology);
366 	SET_BIC(BIC_PkgCnt, &bic_group_topology);
367 	SET_BIC(BIC_Core, &bic_group_topology);
368 	SET_BIC(BIC_CPU, &bic_group_topology);
369 	SET_BIC(BIC_Die, &bic_group_topology);
370 	SET_BIC(BIC_L3, &bic_group_topology);
371 
372 	BIC_INIT(&bic_group_thermal_pwr);
373 	SET_BIC(BIC_CoreTmp, &bic_group_thermal_pwr);
374 	SET_BIC(BIC_PkgTmp, &bic_group_thermal_pwr);
375 	SET_BIC(BIC_PkgWatt, &bic_group_thermal_pwr);
376 	SET_BIC(BIC_CorWatt, &bic_group_thermal_pwr);
377 	SET_BIC(BIC_GFXWatt, &bic_group_thermal_pwr);
378 	SET_BIC(BIC_RAMWatt, &bic_group_thermal_pwr);
379 	SET_BIC(BIC_PKG__, &bic_group_thermal_pwr);
380 	SET_BIC(BIC_RAM__, &bic_group_thermal_pwr);
381 	SET_BIC(BIC_SysWatt, &bic_group_thermal_pwr);
382 
383 	BIC_INIT(&bic_group_frequency);
384 	SET_BIC(BIC_Avg_MHz, &bic_group_frequency);
385 	SET_BIC(BIC_Busy, &bic_group_frequency);
386 	SET_BIC(BIC_Bzy_MHz, &bic_group_frequency);
387 	SET_BIC(BIC_TSC_MHz, &bic_group_frequency);
388 	SET_BIC(BIC_GFXMHz, &bic_group_frequency);
389 	SET_BIC(BIC_GFXACTMHz, &bic_group_frequency);
390 	SET_BIC(BIC_SAMMHz, &bic_group_frequency);
391 	SET_BIC(BIC_SAMACTMHz, &bic_group_frequency);
392 	SET_BIC(BIC_UNCORE_MHZ, &bic_group_frequency);
393 
394 	BIC_INIT(&bic_group_hw_idle);
395 	SET_BIC(BIC_Busy, &bic_group_hw_idle);
396 	SET_BIC(BIC_CPU_c1, &bic_group_hw_idle);
397 	SET_BIC(BIC_CPU_c3, &bic_group_hw_idle);
398 	SET_BIC(BIC_CPU_c6, &bic_group_hw_idle);
399 	SET_BIC(BIC_CPU_c7, &bic_group_hw_idle);
400 	SET_BIC(BIC_GFX_rc6, &bic_group_hw_idle);
401 	SET_BIC(BIC_Pkgpc2, &bic_group_hw_idle);
402 	SET_BIC(BIC_Pkgpc3, &bic_group_hw_idle);
403 	SET_BIC(BIC_Pkgpc6, &bic_group_hw_idle);
404 	SET_BIC(BIC_Pkgpc7, &bic_group_hw_idle);
405 	SET_BIC(BIC_Pkgpc8, &bic_group_hw_idle);
406 	SET_BIC(BIC_Pkgpc9, &bic_group_hw_idle);
407 	SET_BIC(BIC_Pkgpc10, &bic_group_hw_idle);
408 	SET_BIC(BIC_CPU_LPI, &bic_group_hw_idle);
409 	SET_BIC(BIC_SYS_LPI, &bic_group_hw_idle);
410 	SET_BIC(BIC_Mod_c6, &bic_group_hw_idle);
411 	SET_BIC(BIC_Totl_c0, &bic_group_hw_idle);
412 	SET_BIC(BIC_Any_c0, &bic_group_hw_idle);
413 	SET_BIC(BIC_GFX_c0, &bic_group_hw_idle);
414 	SET_BIC(BIC_CPUGFX, &bic_group_hw_idle);
415 	SET_BIC(BIC_SAM_mc6, &bic_group_hw_idle);
416 	SET_BIC(BIC_Diec6, &bic_group_hw_idle);
417 
418 	BIC_INIT(&bic_group_sw_idle);
419 	SET_BIC(BIC_Busy, &bic_group_sw_idle);
420 	SET_BIC(BIC_cpuidle, &bic_group_sw_idle);
421 	SET_BIC(BIC_pct_idle, &bic_group_sw_idle);
422 
423 	BIC_INIT(&bic_group_idle);
424 
425 	CPU_OR(&bic_group_idle, &bic_group_idle, &bic_group_hw_idle);
426 	SET_BIC(BIC_pct_idle, &bic_group_idle);
427 
428 	BIC_INIT(&bic_group_cache);
429 	SET_BIC(BIC_LLC_MRPS, &bic_group_cache);
430 	SET_BIC(BIC_LLC_HIT, &bic_group_cache);
431 	SET_BIC(BIC_L2_MRPS, &bic_group_cache);
432 	SET_BIC(BIC_L2_HIT, &bic_group_cache);
433 
434 	BIC_INIT(&bic_group_other);
435 	SET_BIC(BIC_IRQ, &bic_group_other);
436 	SET_BIC(BIC_NMI, &bic_group_other);
437 	SET_BIC(BIC_SMI, &bic_group_other);
438 	SET_BIC(BIC_ThreadC, &bic_group_other);
439 	SET_BIC(BIC_CoreTmp, &bic_group_other);
440 	SET_BIC(BIC_IPC, &bic_group_other);
441 
442 	BIC_INIT(&bic_group_disabled_by_default);
443 	SET_BIC(BIC_USEC, &bic_group_disabled_by_default);
444 	SET_BIC(BIC_TOD, &bic_group_disabled_by_default);
445 	SET_BIC(BIC_cpuidle, &bic_group_disabled_by_default);
446 	SET_BIC(BIC_APIC, &bic_group_disabled_by_default);
447 	SET_BIC(BIC_X2APIC, &bic_group_disabled_by_default);
448 
449 	BIC_INIT(&bic_enabled);
450 	bic_set_all(&bic_enabled);
451 	bic_clear_bits(&bic_enabled, &bic_group_disabled_by_default);
452 
453 	BIC_INIT(&bic_present);
454 	SET_BIC(BIC_USEC, &bic_present);
455 	SET_BIC(BIC_TOD, &bic_present);
456 	SET_BIC(BIC_cpuidle, &bic_present);
457 	SET_BIC(BIC_APIC, &bic_present);
458 	SET_BIC(BIC_X2APIC, &bic_present);
459 	SET_BIC(BIC_pct_idle, &bic_present);
460 }
461 
462 /*
463  * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
464  * If you change the values, note they are used both in comparisons
465  * (>= PCL__7) and to index pkg_cstate_limit_strings[].
466  */
467 #define PCLUKN 0		/* Unknown */
468 #define PCLRSV 1		/* Reserved */
469 #define PCL__0 2		/* PC0 */
470 #define PCL__1 3		/* PC1 */
471 #define PCL__2 4		/* PC2 */
472 #define PCL__3 5		/* PC3 */
473 #define PCL__4 6		/* PC4 */
474 #define PCL__6 7		/* PC6 */
475 #define PCL_6N 8		/* PC6 No Retention */
476 #define PCL_6R 9		/* PC6 Retention */
477 #define PCL__7 10		/* PC7 */
478 #define PCL_7S 11		/* PC7 Shrink */
479 #define PCL__8 12		/* PC8 */
480 #define PCL__9 13		/* PC9 */
481 #define PCL_10 14		/* PC10 */
482 #define PCLUNL 15		/* Unlimited */
483 
484 char *proc_stat = "/proc/stat";
485 FILE *outf;
486 int *fd_percpu;
487 int *fd_instr_count_percpu;
488 int *fd_llc_percpu;
489 int *fd_l2_percpu;
490 struct timeval interval_tv = { 5, 0 };
491 struct timespec interval_ts = { 5, 0 };
492 
493 unsigned int num_iterations;
494 unsigned int header_iterations;
495 unsigned int debug;
496 unsigned int quiet;
497 unsigned int shown;
498 unsigned int sums_need_wide_columns;
499 unsigned int rapl_joules;
500 unsigned int valid_rapl_msrs;
501 unsigned int summary_only;
502 unsigned int list_header_only;
503 unsigned int dump_only;
504 unsigned int force_load;
505 unsigned int cpuid_has_aperf_mperf;
506 unsigned int cpuid_has_hv;
507 unsigned int has_aperf_access;
508 unsigned int has_epb;
509 unsigned int has_turbo;
510 unsigned int is_hybrid;
511 unsigned int units = 1000000;	/* MHz etc */
512 unsigned int genuine_intel;
513 unsigned int authentic_amd;
514 unsigned int hygon_genuine;
515 unsigned int max_level, max_extended_level;
516 unsigned int has_invariant_tsc;
517 unsigned int aperf_mperf_multiplier = 1;
518 double bclk;
519 double base_hz;
520 unsigned int has_base_hz;
521 double tsc_tweak = 1.0;
522 unsigned int show_pkg_only;
523 unsigned int show_core_only;
524 char *output_buffer, *outp;
525 unsigned int do_dts;
526 unsigned int do_ptm;
527 unsigned int do_ipc;
528 unsigned long long cpuidle_cur_cpu_lpi_us;
529 unsigned long long cpuidle_cur_sys_lpi_us;
530 unsigned int tj_max;
531 unsigned int tj_max_override;
532 double rapl_power_units, rapl_time_units;
533 double rapl_dram_energy_units, rapl_energy_units, rapl_psys_energy_units;
534 double rapl_joule_counter_range;
535 unsigned int crystal_hz;
536 unsigned long long tsc_hz;
537 int master_cpu;
538 unsigned int has_hwp;		/* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
539 			/* IA32_HWP_REQUEST, IA32_HWP_STATUS */
540 unsigned int has_hwp_notify;	/* IA32_HWP_INTERRUPT */
541 unsigned int has_hwp_activity_window;	/* IA32_HWP_REQUEST[bits 41:32] */
542 unsigned int has_hwp_epp;	/* IA32_HWP_REQUEST[bits 31:24] */
543 unsigned int has_hwp_pkg;	/* IA32_HWP_REQUEST_PKG */
544 unsigned int first_counter_read = 1;
545 
546 static struct timeval procsysfs_tv_begin;
547 
548 int ignore_stdin;
549 bool no_msr;
550 bool no_perf;
551 
552 enum gfx_sysfs_idx {
553 	GFX_rc6,
554 	GFX_MHz,
555 	GFX_ACTMHz,
556 	SAM_mc6,
557 	SAM_MHz,
558 	SAM_ACTMHz,
559 	GFX_MAX
560 };
561 
562 struct gfx_sysfs_info {
563 	FILE *fp;
564 	unsigned int val;
565 	unsigned long long val_ull;
566 };
567 
568 static struct gfx_sysfs_info gfx_info[GFX_MAX];
569 
570 int get_msr(int cpu, off_t offset, unsigned long long *msr);
571 int add_counter(unsigned int msr_num, char *path, char *name,
572 		unsigned int width, enum counter_scope scope, enum counter_type type, enum counter_format format, int flags, int package_num);
573 
574 /* Model specific support Start */
575 
576 /* List of features that may diverge among different platforms */
577 struct platform_features {
578 	bool has_msr_misc_feature_control;	/* MSR_MISC_FEATURE_CONTROL */
579 	bool has_msr_misc_pwr_mgmt;	/* MSR_MISC_PWR_MGMT */
580 	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 */
581 	bool has_config_tdp;	/* MSR_CONFIG_TDP_NOMINAL/LEVEL_1/LEVEL_2/CONTROL, MSR_TURBO_ACTIVATION_RATIO */
582 	int bclk_freq;		/* CPU base clock */
583 	int crystal_freq;	/* Crystal clock to use when not available from CPUID.15 */
584 	int supported_cstates;	/* Core cstates and Package cstates supported */
585 	int cst_limit;		/* MSR_PKG_CST_CONFIG_CONTROL */
586 	bool has_cst_auto_convension;	/* AUTOMATIC_CSTATE_CONVERSION bit in MSR_PKG_CST_CONFIG_CONTROL */
587 	bool has_irtl_msrs;	/* MSR_PKGC3/PKGC6/PKGC7/PKGC8/PKGC9/PKGC10_IRTL */
588 	bool has_msr_core_c1_res;	/* MSR_CORE_C1_RES */
589 	bool has_msr_module_c6_res_ms;	/* MSR_MODULE_C6_RES_MS */
590 	bool has_msr_c6_demotion_policy_config;	/* MSR_CC6_DEMOTION_POLICY_CONFIG/MSR_MC6_DEMOTION_POLICY_CONFIG */
591 	bool has_msr_atom_pkg_c6_residency;	/* MSR_ATOM_PKG_C6_RESIDENCY */
592 	bool has_msr_knl_core_c6_residency;	/* MSR_KNL_CORE_C6_RESIDENCY */
593 	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 */
594 	bool has_cst_prewake_bit;	/* Cstate prewake bit in MSR_IA32_POWER_CTL */
595 	int trl_msrs;		/* MSR_TURBO_RATIO_LIMIT/LIMIT1/LIMIT2/SECONDARY, Atom TRL MSRs */
596 	int plr_msrs;		/* MSR_CORE/GFX/RING_PERF_LIMIT_REASONS */
597 	int plat_rapl_msrs;	/* RAPL PKG/DRAM/CORE/GFX MSRs, AMD RAPL MSRs */
598 	bool has_per_core_rapl;	/* Indicates cores energy collection is per-core, not per-package. AMD specific for now */
599 	bool has_rapl_divisor;	/* Divisor for Energy unit raw value from MSR_RAPL_POWER_UNIT */
600 	bool has_fixed_rapl_unit;	/* Fixed Energy Unit used for DRAM RAPL Domain */
601 	bool has_fixed_rapl_psys_unit;	/* Fixed Energy Unit used for PSYS RAPL Domain */
602 	int rapl_quirk_tdp;	/* Hardcoded TDP value when cannot be retrieved from hardware */
603 	int tcc_offset_bits;	/* TCC Offset bits in MSR_IA32_TEMPERATURE_TARGET */
604 	bool enable_tsc_tweak;	/* Use CPU Base freq instead of TSC freq for aperf/mperf counter */
605 	bool need_perf_multiplier;	/* mperf/aperf multiplier */
606 };
607 
608 struct platform_data {
609 	unsigned int vfm;
610 	const struct platform_features *features;
611 };
612 
613 /* For BCLK */
614 enum bclk_freq {
615 	BCLK_100MHZ = 1,
616 	BCLK_133MHZ,
617 	BCLK_SLV,
618 };
619 
620 #define SLM_BCLK_FREQS 5
621 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0 };
622 
slm_bclk(void)623 double slm_bclk(void)
624 {
625 	unsigned long long msr = 3;
626 	unsigned int i;
627 	double freq;
628 
629 	if (get_msr(master_cpu, MSR_FSB_FREQ, &msr))
630 		fprintf(outf, "SLM BCLK: unknown\n");
631 
632 	i = msr & 0xf;
633 	if (i >= SLM_BCLK_FREQS) {
634 		fprintf(outf, "SLM BCLK[%d] invalid\n", i);
635 		i = 3;
636 	}
637 	freq = slm_freq_table[i];
638 
639 	if (!quiet)
640 		fprintf(outf, "SLM BCLK: %.1f Mhz\n", freq);
641 
642 	return freq;
643 }
644 
645 /* For Package cstate limit */
646 enum package_cstate_limit {
647 	CST_LIMIT_NHM = 1,
648 	CST_LIMIT_SNB,
649 	CST_LIMIT_HSW,
650 	CST_LIMIT_SKX,
651 	CST_LIMIT_ICX,
652 	CST_LIMIT_SLV,
653 	CST_LIMIT_AMT,
654 	CST_LIMIT_KNL,
655 	CST_LIMIT_GMT,
656 };
657 
658 /* For Turbo Ratio Limit MSRs */
659 enum turbo_ratio_limit_msrs {
660 	TRL_BASE = BIT(0),
661 	TRL_LIMIT1 = BIT(1),
662 	TRL_LIMIT2 = BIT(2),
663 	TRL_ATOM = BIT(3),
664 	TRL_KNL = BIT(4),
665 	TRL_CORECOUNT = BIT(5),
666 };
667 
668 /* For Perf Limit Reason MSRs */
669 enum perf_limit_reason_msrs {
670 	PLR_CORE = BIT(0),
671 	PLR_GFX = BIT(1),
672 	PLR_RING = BIT(2),
673 };
674 
675 /* For RAPL MSRs */
676 enum rapl_msrs {
677 	RAPL_PKG_POWER_LIMIT = BIT(0),	/* 0x610 MSR_PKG_POWER_LIMIT */
678 	RAPL_PKG_ENERGY_STATUS = BIT(1),	/* 0x611 MSR_PKG_ENERGY_STATUS */
679 	RAPL_PKG_PERF_STATUS = BIT(2),	/* 0x613 MSR_PKG_PERF_STATUS */
680 	RAPL_PKG_POWER_INFO = BIT(3),	/* 0x614 MSR_PKG_POWER_INFO */
681 	RAPL_DRAM_POWER_LIMIT = BIT(4),	/* 0x618 MSR_DRAM_POWER_LIMIT */
682 	RAPL_DRAM_ENERGY_STATUS = BIT(5),	/* 0x619 MSR_DRAM_ENERGY_STATUS */
683 	RAPL_DRAM_PERF_STATUS = BIT(6),	/* 0x61b MSR_DRAM_PERF_STATUS */
684 	RAPL_DRAM_POWER_INFO = BIT(7),	/* 0x61c MSR_DRAM_POWER_INFO */
685 	RAPL_CORE_POWER_LIMIT = BIT(8),	/* 0x638 MSR_PP0_POWER_LIMIT */
686 	RAPL_CORE_ENERGY_STATUS = BIT(9),	/* 0x639 MSR_PP0_ENERGY_STATUS */
687 	RAPL_CORE_POLICY = BIT(10),	/* 0x63a MSR_PP0_POLICY */
688 	RAPL_GFX_POWER_LIMIT = BIT(11),	/* 0x640 MSR_PP1_POWER_LIMIT */
689 	RAPL_GFX_ENERGY_STATUS = BIT(12),	/* 0x641 MSR_PP1_ENERGY_STATUS */
690 	RAPL_GFX_POLICY = BIT(13),	/* 0x642 MSR_PP1_POLICY */
691 	RAPL_AMD_PWR_UNIT = BIT(14),	/* 0xc0010299 MSR_AMD_RAPL_POWER_UNIT */
692 	RAPL_AMD_CORE_ENERGY_STAT = BIT(15),	/* 0xc001029a MSR_AMD_CORE_ENERGY_STATUS */
693 	RAPL_AMD_PKG_ENERGY_STAT = BIT(16),	/* 0xc001029b MSR_AMD_PKG_ENERGY_STATUS */
694 	RAPL_PLATFORM_ENERGY_LIMIT = BIT(17),	/* 0x64c MSR_PLATFORM_ENERGY_LIMIT */
695 	RAPL_PLATFORM_ENERGY_STATUS = BIT(18),	/* 0x64d MSR_PLATFORM_ENERGY_STATUS */
696 };
697 
698 #define RAPL_PKG	(RAPL_PKG_ENERGY_STATUS | RAPL_PKG_POWER_LIMIT)
699 #define RAPL_DRAM	(RAPL_DRAM_ENERGY_STATUS | RAPL_DRAM_POWER_LIMIT)
700 #define RAPL_CORE	(RAPL_CORE_ENERGY_STATUS | RAPL_CORE_POWER_LIMIT)
701 #define RAPL_GFX	(RAPL_GFX_POWER_LIMIT | RAPL_GFX_ENERGY_STATUS)
702 #define RAPL_PSYS	(RAPL_PLATFORM_ENERGY_STATUS | RAPL_PLATFORM_ENERGY_LIMIT)
703 
704 #define RAPL_PKG_ALL	(RAPL_PKG | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO)
705 #define RAPL_DRAM_ALL	(RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_DRAM_POWER_INFO)
706 #define RAPL_CORE_ALL	(RAPL_CORE | RAPL_CORE_POLICY)
707 #define RAPL_GFX_ALL	(RAPL_GFX | RAPL_GFX_POLICY)
708 
709 #define RAPL_AMD_F17H	(RAPL_AMD_PWR_UNIT | RAPL_AMD_CORE_ENERGY_STAT | RAPL_AMD_PKG_ENERGY_STAT)
710 
711 /* For Cstates */
712 enum cstates {
713 	CC1 = BIT(0),
714 	CC3 = BIT(1),
715 	CC6 = BIT(2),
716 	CC7 = BIT(3),
717 	PC2 = BIT(4),
718 	PC3 = BIT(5),
719 	PC6 = BIT(6),
720 	PC7 = BIT(7),
721 	PC8 = BIT(8),
722 	PC9 = BIT(9),
723 	PC10 = BIT(10),
724 };
725 
726 static const struct platform_features nhm_features = {
727 	.has_msr_misc_pwr_mgmt = 1,
728 	.has_nhm_msrs = 1,
729 	.bclk_freq = BCLK_133MHZ,
730 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
731 	.cst_limit = CST_LIMIT_NHM,
732 	.trl_msrs = TRL_BASE,
733 };
734 
735 static const struct platform_features nhx_features = {
736 	.has_msr_misc_pwr_mgmt = 1,
737 	.has_nhm_msrs = 1,
738 	.bclk_freq = BCLK_133MHZ,
739 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
740 	.cst_limit = CST_LIMIT_NHM,
741 };
742 
743 static const struct platform_features snb_features = {
744 	.has_msr_misc_feature_control = 1,
745 	.has_msr_misc_pwr_mgmt = 1,
746 	.has_nhm_msrs = 1,
747 	.bclk_freq = BCLK_100MHZ,
748 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
749 	.cst_limit = CST_LIMIT_SNB,
750 	.has_irtl_msrs = 1,
751 	.trl_msrs = TRL_BASE,
752 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
753 };
754 
755 static const struct platform_features snx_features = {
756 	.has_msr_misc_feature_control = 1,
757 	.has_msr_misc_pwr_mgmt = 1,
758 	.has_nhm_msrs = 1,
759 	.bclk_freq = BCLK_100MHZ,
760 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
761 	.cst_limit = CST_LIMIT_SNB,
762 	.has_irtl_msrs = 1,
763 	.trl_msrs = TRL_BASE,
764 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM_ALL,
765 };
766 
767 static const struct platform_features ivb_features = {
768 	.has_msr_misc_feature_control = 1,
769 	.has_msr_misc_pwr_mgmt = 1,
770 	.has_nhm_msrs = 1,
771 	.has_config_tdp = 1,
772 	.bclk_freq = BCLK_100MHZ,
773 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
774 	.cst_limit = CST_LIMIT_SNB,
775 	.has_irtl_msrs = 1,
776 	.trl_msrs = TRL_BASE,
777 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
778 };
779 
780 static const struct platform_features ivx_features = {
781 	.has_msr_misc_feature_control = 1,
782 	.has_msr_misc_pwr_mgmt = 1,
783 	.has_nhm_msrs = 1,
784 	.bclk_freq = BCLK_100MHZ,
785 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
786 	.cst_limit = CST_LIMIT_SNB,
787 	.has_irtl_msrs = 1,
788 	.trl_msrs = TRL_BASE | TRL_LIMIT1,
789 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM_ALL,
790 };
791 
792 static const struct platform_features hsw_features = {
793 	.has_msr_misc_feature_control = 1,
794 	.has_msr_misc_pwr_mgmt = 1,
795 	.has_nhm_msrs = 1,
796 	.has_config_tdp = 1,
797 	.bclk_freq = BCLK_100MHZ,
798 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
799 	.cst_limit = CST_LIMIT_HSW,
800 	.has_irtl_msrs = 1,
801 	.trl_msrs = TRL_BASE,
802 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
803 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
804 };
805 
806 static const struct platform_features hsx_features = {
807 	.has_msr_misc_feature_control = 1,
808 	.has_msr_misc_pwr_mgmt = 1,
809 	.has_nhm_msrs = 1,
810 	.has_config_tdp = 1,
811 	.bclk_freq = BCLK_100MHZ,
812 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
813 	.cst_limit = CST_LIMIT_HSW,
814 	.has_irtl_msrs = 1,
815 	.trl_msrs = TRL_BASE | TRL_LIMIT1 | TRL_LIMIT2,
816 	.plr_msrs = PLR_CORE | PLR_RING,
817 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
818 	.has_fixed_rapl_unit = 1,
819 };
820 
821 static const struct platform_features hswl_features = {
822 	.has_msr_misc_feature_control = 1,
823 	.has_msr_misc_pwr_mgmt = 1,
824 	.has_nhm_msrs = 1,
825 	.has_config_tdp = 1,
826 	.bclk_freq = BCLK_100MHZ,
827 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
828 	.cst_limit = CST_LIMIT_HSW,
829 	.has_irtl_msrs = 1,
830 	.trl_msrs = TRL_BASE,
831 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
832 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
833 };
834 
835 static const struct platform_features hswg_features = {
836 	.has_msr_misc_feature_control = 1,
837 	.has_msr_misc_pwr_mgmt = 1,
838 	.has_nhm_msrs = 1,
839 	.has_config_tdp = 1,
840 	.bclk_freq = BCLK_100MHZ,
841 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
842 	.cst_limit = CST_LIMIT_HSW,
843 	.has_irtl_msrs = 1,
844 	.trl_msrs = TRL_BASE,
845 	.plr_msrs = PLR_CORE | PLR_GFX | PLR_RING,
846 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
847 };
848 
849 static const struct platform_features bdw_features = {
850 	.has_msr_misc_feature_control = 1,
851 	.has_msr_misc_pwr_mgmt = 1,
852 	.has_nhm_msrs = 1,
853 	.has_config_tdp = 1,
854 	.bclk_freq = BCLK_100MHZ,
855 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
856 	.cst_limit = CST_LIMIT_HSW,
857 	.has_irtl_msrs = 1,
858 	.trl_msrs = TRL_BASE,
859 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
860 };
861 
862 static const struct platform_features bdwg_features = {
863 	.has_msr_misc_feature_control = 1,
864 	.has_msr_misc_pwr_mgmt = 1,
865 	.has_nhm_msrs = 1,
866 	.has_config_tdp = 1,
867 	.bclk_freq = BCLK_100MHZ,
868 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7,
869 	.cst_limit = CST_LIMIT_HSW,
870 	.has_irtl_msrs = 1,
871 	.trl_msrs = TRL_BASE,
872 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE_ALL | RAPL_GFX | RAPL_PKG_POWER_INFO,
873 };
874 
875 static const struct platform_features bdx_features = {
876 	.has_msr_misc_feature_control = 1,
877 	.has_msr_misc_pwr_mgmt = 1,
878 	.has_nhm_msrs = 1,
879 	.has_config_tdp = 1,
880 	.bclk_freq = BCLK_100MHZ,
881 	.supported_cstates = CC1 | CC3 | CC6 | PC2 | PC3 | PC6,
882 	.cst_limit = CST_LIMIT_HSW,
883 	.has_irtl_msrs = 1,
884 	.has_cst_auto_convension = 1,
885 	.trl_msrs = TRL_BASE,
886 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
887 	.has_fixed_rapl_unit = 1,
888 };
889 
890 static const struct platform_features skl_features = {
891 	.has_msr_misc_feature_control = 1,
892 	.has_msr_misc_pwr_mgmt = 1,
893 	.has_nhm_msrs = 1,
894 	.has_config_tdp = 1,
895 	.bclk_freq = BCLK_100MHZ,
896 	.crystal_freq = 24000000,
897 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
898 	.cst_limit = CST_LIMIT_HSW,
899 	.has_irtl_msrs = 1,
900 	.has_ext_cst_msrs = 1,
901 	.trl_msrs = TRL_BASE,
902 	.tcc_offset_bits = 6,
903 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX | RAPL_PSYS,
904 	.enable_tsc_tweak = 1,
905 };
906 
907 static const struct platform_features cnl_features = {
908 	.has_msr_misc_feature_control = 1,
909 	.has_msr_misc_pwr_mgmt = 1,
910 	.has_nhm_msrs = 1,
911 	.has_config_tdp = 1,
912 	.bclk_freq = BCLK_100MHZ,
913 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
914 	.cst_limit = CST_LIMIT_HSW,
915 	.has_irtl_msrs = 1,
916 	.has_msr_core_c1_res = 1,
917 	.has_ext_cst_msrs = 1,
918 	.trl_msrs = TRL_BASE,
919 	.tcc_offset_bits = 6,
920 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX | RAPL_PSYS,
921 	.enable_tsc_tweak = 1,
922 };
923 
924 /* Copied from cnl_features, with PC7/PC9 removed */
925 static const struct platform_features adl_features = {
926 	.has_msr_misc_feature_control	= cnl_features.has_msr_misc_feature_control,
927 	.has_msr_misc_pwr_mgmt		= cnl_features.has_msr_misc_pwr_mgmt,
928 	.has_nhm_msrs			= cnl_features.has_nhm_msrs,
929 	.has_config_tdp			= cnl_features.has_config_tdp,
930 	.bclk_freq			= cnl_features.bclk_freq,
931 	.supported_cstates		= CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC8 | PC10,
932 	.cst_limit			= cnl_features.cst_limit,
933 	.has_irtl_msrs			= cnl_features.has_irtl_msrs,
934 	.has_msr_core_c1_res		= cnl_features.has_msr_core_c1_res,
935 	.has_ext_cst_msrs		= cnl_features.has_ext_cst_msrs,
936 	.trl_msrs			= cnl_features.trl_msrs,
937 	.tcc_offset_bits		= cnl_features.tcc_offset_bits,
938 	.plat_rapl_msrs			= cnl_features.plat_rapl_msrs,
939 	.enable_tsc_tweak		= cnl_features.enable_tsc_tweak,
940 };
941 
942 /* Copied from adl_features, with PC3/PC8 removed */
943 static const struct platform_features lnl_features = {
944 	.has_msr_misc_feature_control	= adl_features.has_msr_misc_feature_control,
945 	.has_msr_misc_pwr_mgmt		= adl_features.has_msr_misc_pwr_mgmt,
946 	.has_nhm_msrs			= adl_features.has_nhm_msrs,
947 	.has_config_tdp			= adl_features.has_config_tdp,
948 	.bclk_freq			= adl_features.bclk_freq,
949 	.supported_cstates		= CC1 | CC6 | CC7 | PC2 | PC6 | PC10,
950 	.cst_limit			= adl_features.cst_limit,
951 	.has_irtl_msrs			= adl_features.has_irtl_msrs,
952 	.has_msr_core_c1_res		= adl_features.has_msr_core_c1_res,
953 	.has_ext_cst_msrs		= adl_features.has_ext_cst_msrs,
954 	.trl_msrs			= adl_features.trl_msrs,
955 	.tcc_offset_bits		= adl_features.tcc_offset_bits,
956 	.plat_rapl_msrs			= adl_features.plat_rapl_msrs,
957 	.enable_tsc_tweak		= adl_features.enable_tsc_tweak,
958 };
959 
960 static const struct platform_features skx_features = {
961 	.has_msr_misc_feature_control = 1,
962 	.has_msr_misc_pwr_mgmt = 1,
963 	.has_nhm_msrs = 1,
964 	.has_config_tdp = 1,
965 	.bclk_freq = BCLK_100MHZ,
966 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
967 	.cst_limit = CST_LIMIT_SKX,
968 	.has_irtl_msrs = 1,
969 	.has_cst_auto_convension = 1,
970 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
971 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
972 	.has_fixed_rapl_unit = 1,
973 };
974 
975 static const struct platform_features icx_features = {
976 	.has_msr_misc_feature_control = 1,
977 	.has_msr_misc_pwr_mgmt = 1,
978 	.has_nhm_msrs = 1,
979 	.has_config_tdp = 1,
980 	.bclk_freq = BCLK_100MHZ,
981 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
982 	.cst_limit = CST_LIMIT_ICX,
983 	.has_msr_core_c1_res = 1,
984 	.has_irtl_msrs = 1,
985 	.has_cst_prewake_bit = 1,
986 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
987 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
988 	.has_fixed_rapl_unit = 1,
989 };
990 
991 static const struct platform_features spr_features = {
992 	.has_msr_misc_feature_control = 1,
993 	.has_msr_misc_pwr_mgmt = 1,
994 	.has_nhm_msrs = 1,
995 	.has_config_tdp = 1,
996 	.bclk_freq = BCLK_100MHZ,
997 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
998 	.cst_limit = CST_LIMIT_SKX,
999 	.has_msr_core_c1_res = 1,
1000 	.has_irtl_msrs = 1,
1001 	.has_cst_prewake_bit = 1,
1002 	.has_fixed_rapl_psys_unit = 1,
1003 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1004 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
1005 };
1006 
1007 static const struct platform_features dmr_features = {
1008 	.has_msr_misc_feature_control	= spr_features.has_msr_misc_feature_control,
1009 	.has_msr_misc_pwr_mgmt		= spr_features.has_msr_misc_pwr_mgmt,
1010 	.has_nhm_msrs			= spr_features.has_nhm_msrs,
1011 	.bclk_freq			= spr_features.bclk_freq,
1012 	.supported_cstates		= spr_features.supported_cstates,
1013 	.cst_limit			= spr_features.cst_limit,
1014 	.has_msr_core_c1_res		= spr_features.has_msr_core_c1_res,
1015 	.has_cst_prewake_bit		= spr_features.has_cst_prewake_bit,
1016 	.has_fixed_rapl_psys_unit	= spr_features.has_fixed_rapl_psys_unit,
1017 	.trl_msrs			= spr_features.trl_msrs,
1018 	.has_msr_module_c6_res_ms	= 1,	/* DMR has Dual-Core-Module and MC6 MSR */
1019 	.plat_rapl_msrs			= 0,	/* DMR does not have RAPL MSRs */
1020 	.plr_msrs			= 0,	/* DMR does not have PLR  MSRs */
1021 	.has_irtl_msrs			= 0,	/* DMR does not have IRTL MSRs */
1022 	.has_config_tdp			= 0,	/* DMR does not have CTDP MSRs */
1023 };
1024 
1025 static const struct platform_features srf_features = {
1026 	.has_msr_misc_feature_control = 1,
1027 	.has_msr_misc_pwr_mgmt = 1,
1028 	.has_nhm_msrs = 1,
1029 	.has_config_tdp = 1,
1030 	.bclk_freq = BCLK_100MHZ,
1031 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
1032 	.cst_limit = CST_LIMIT_SKX,
1033 	.has_msr_core_c1_res = 1,
1034 	.has_msr_module_c6_res_ms = 1,
1035 	.has_irtl_msrs = 1,
1036 	.has_cst_prewake_bit = 1,
1037 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1038 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
1039 };
1040 
1041 static const struct platform_features grr_features = {
1042 	.has_msr_misc_feature_control = 1,
1043 	.has_msr_misc_pwr_mgmt = 1,
1044 	.has_nhm_msrs = 1,
1045 	.has_config_tdp = 1,
1046 	.bclk_freq = BCLK_100MHZ,
1047 	.supported_cstates = CC1 | CC6,
1048 	.cst_limit = CST_LIMIT_SKX,
1049 	.has_msr_core_c1_res = 1,
1050 	.has_msr_module_c6_res_ms = 1,
1051 	.has_irtl_msrs = 1,
1052 	.has_cst_prewake_bit = 1,
1053 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1054 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_PSYS,
1055 };
1056 
1057 static const struct platform_features slv_features = {
1058 	.has_nhm_msrs = 1,
1059 	.bclk_freq = BCLK_SLV,
1060 	.supported_cstates = CC1 | CC6 | PC6,
1061 	.cst_limit = CST_LIMIT_SLV,
1062 	.has_msr_core_c1_res = 1,
1063 	.has_msr_module_c6_res_ms = 1,
1064 	.has_msr_c6_demotion_policy_config = 1,
1065 	.has_msr_atom_pkg_c6_residency = 1,
1066 	.trl_msrs = TRL_ATOM,
1067 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE,
1068 	.has_rapl_divisor = 1,
1069 	.rapl_quirk_tdp = 30,
1070 };
1071 
1072 static const struct platform_features slvd_features = {
1073 	.has_msr_misc_pwr_mgmt = 1,
1074 	.has_nhm_msrs = 1,
1075 	.bclk_freq = BCLK_SLV,
1076 	.supported_cstates = CC1 | CC6 | PC3 | PC6,
1077 	.cst_limit = CST_LIMIT_SLV,
1078 	.has_msr_atom_pkg_c6_residency = 1,
1079 	.trl_msrs = TRL_BASE,
1080 	.plat_rapl_msrs = RAPL_PKG | RAPL_CORE,
1081 	.rapl_quirk_tdp = 30,
1082 };
1083 
1084 static const struct platform_features amt_features = {
1085 	.has_nhm_msrs = 1,
1086 	.bclk_freq = BCLK_133MHZ,
1087 	.supported_cstates = CC1 | CC3 | CC6 | PC3 | PC6,
1088 	.cst_limit = CST_LIMIT_AMT,
1089 	.trl_msrs = TRL_BASE,
1090 };
1091 
1092 static const struct platform_features gmt_features = {
1093 	.has_msr_misc_pwr_mgmt = 1,
1094 	.has_nhm_msrs = 1,
1095 	.bclk_freq = BCLK_100MHZ,
1096 	.crystal_freq = 19200000,
1097 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
1098 	.cst_limit = CST_LIMIT_GMT,
1099 	.has_irtl_msrs = 1,
1100 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1101 	.plat_rapl_msrs = RAPL_PKG | RAPL_PKG_POWER_INFO,
1102 };
1103 
1104 static const struct platform_features gmtd_features = {
1105 	.has_msr_misc_pwr_mgmt = 1,
1106 	.has_nhm_msrs = 1,
1107 	.bclk_freq = BCLK_100MHZ,
1108 	.crystal_freq = 25000000,
1109 	.supported_cstates = CC1 | CC6 | PC2 | PC6,
1110 	.cst_limit = CST_LIMIT_GMT,
1111 	.has_irtl_msrs = 1,
1112 	.has_msr_core_c1_res = 1,
1113 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1114 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL | RAPL_CORE_ENERGY_STATUS,
1115 };
1116 
1117 static const struct platform_features gmtp_features = {
1118 	.has_msr_misc_pwr_mgmt = 1,
1119 	.has_nhm_msrs = 1,
1120 	.bclk_freq = BCLK_100MHZ,
1121 	.crystal_freq = 19200000,
1122 	.supported_cstates = CC1 | CC3 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
1123 	.cst_limit = CST_LIMIT_GMT,
1124 	.has_irtl_msrs = 1,
1125 	.trl_msrs = TRL_BASE,
1126 	.plat_rapl_msrs = RAPL_PKG | RAPL_PKG_POWER_INFO,
1127 };
1128 
1129 static const struct platform_features tmt_features = {
1130 	.has_msr_misc_pwr_mgmt = 1,
1131 	.has_nhm_msrs = 1,
1132 	.bclk_freq = BCLK_100MHZ,
1133 	.supported_cstates = CC1 | CC6 | CC7 | PC2 | PC3 | PC6 | PC7 | PC8 | PC9 | PC10,
1134 	.cst_limit = CST_LIMIT_GMT,
1135 	.has_irtl_msrs = 1,
1136 	.trl_msrs = TRL_BASE,
1137 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_CORE_ALL | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_GFX,
1138 	.enable_tsc_tweak = 1,
1139 };
1140 
1141 static const struct platform_features tmtd_features = {
1142 	.has_msr_misc_pwr_mgmt = 1,
1143 	.has_nhm_msrs = 1,
1144 	.bclk_freq = BCLK_100MHZ,
1145 	.supported_cstates = CC1 | CC6,
1146 	.cst_limit = CST_LIMIT_GMT,
1147 	.has_irtl_msrs = 1,
1148 	.trl_msrs = TRL_BASE | TRL_CORECOUNT,
1149 	.plat_rapl_msrs = RAPL_PKG_ALL,
1150 };
1151 
1152 static const struct platform_features knl_features = {
1153 	.has_msr_misc_pwr_mgmt = 1,
1154 	.has_nhm_msrs = 1,
1155 	.has_config_tdp = 1,
1156 	.bclk_freq = BCLK_100MHZ,
1157 	.supported_cstates = CC1 | CC6 | PC3 | PC6,
1158 	.cst_limit = CST_LIMIT_KNL,
1159 	.has_msr_knl_core_c6_residency = 1,
1160 	.trl_msrs = TRL_KNL,
1161 	.plat_rapl_msrs = RAPL_PKG_ALL | RAPL_DRAM_ALL,
1162 	.has_fixed_rapl_unit = 1,
1163 	.need_perf_multiplier = 1,
1164 };
1165 
1166 static const struct platform_features default_features = {
1167 };
1168 
1169 static const struct platform_features amd_features_with_rapl = {
1170 	.plat_rapl_msrs = RAPL_AMD_F17H,
1171 	.has_per_core_rapl = 1,
1172 	.rapl_quirk_tdp = 280,	/* This is the max stock TDP of HEDT/Server Fam17h+ chips */
1173 };
1174 
1175 static const struct platform_data turbostat_pdata[] = {
1176 	{ INTEL_NEHALEM, &nhm_features },
1177 	{ INTEL_NEHALEM_G, &nhm_features },
1178 	{ INTEL_NEHALEM_EP, &nhm_features },
1179 	{ INTEL_NEHALEM_EX, &nhx_features },
1180 	{ INTEL_WESTMERE, &nhm_features },
1181 	{ INTEL_WESTMERE_EP, &nhm_features },
1182 	{ INTEL_WESTMERE_EX, &nhx_features },
1183 	{ INTEL_SANDYBRIDGE, &snb_features },
1184 	{ INTEL_SANDYBRIDGE_X, &snx_features },
1185 	{ INTEL_IVYBRIDGE, &ivb_features },
1186 	{ INTEL_IVYBRIDGE_X, &ivx_features },
1187 	{ INTEL_HASWELL, &hsw_features },
1188 	{ INTEL_HASWELL_X, &hsx_features },
1189 	{ INTEL_HASWELL_L, &hswl_features },
1190 	{ INTEL_HASWELL_G, &hswg_features },
1191 	{ INTEL_BROADWELL, &bdw_features },
1192 	{ INTEL_BROADWELL_G, &bdwg_features },
1193 	{ INTEL_BROADWELL_X, &bdx_features },
1194 	{ INTEL_BROADWELL_D, &bdx_features },
1195 	{ INTEL_SKYLAKE_L, &skl_features },
1196 	{ INTEL_SKYLAKE, &skl_features },
1197 	{ INTEL_SKYLAKE_X, &skx_features },
1198 	{ INTEL_KABYLAKE_L, &skl_features },
1199 	{ INTEL_KABYLAKE, &skl_features },
1200 	{ INTEL_COMETLAKE, &skl_features },
1201 	{ INTEL_COMETLAKE_L, &skl_features },
1202 	{ INTEL_CANNONLAKE_L, &cnl_features },
1203 	{ INTEL_ICELAKE_X, &icx_features },
1204 	{ INTEL_ICELAKE_D, &icx_features },
1205 	{ INTEL_ICELAKE_L, &cnl_features },
1206 	{ INTEL_ICELAKE_NNPI, &cnl_features },
1207 	{ INTEL_ROCKETLAKE, &cnl_features },
1208 	{ INTEL_TIGERLAKE_L, &cnl_features },
1209 	{ INTEL_TIGERLAKE, &cnl_features },
1210 	{ INTEL_SAPPHIRERAPIDS_X, &spr_features },
1211 	{ INTEL_EMERALDRAPIDS_X, &spr_features },
1212 	{ INTEL_GRANITERAPIDS_X, &spr_features },
1213 	{ INTEL_GRANITERAPIDS_D, &spr_features },
1214 	{ INTEL_DIAMONDRAPIDS_X, &dmr_features },
1215 	{ INTEL_LAKEFIELD, &cnl_features },
1216 	{ INTEL_ALDERLAKE, &adl_features },
1217 	{ INTEL_ALDERLAKE_L, &adl_features },
1218 	{ INTEL_RAPTORLAKE, &adl_features },
1219 	{ INTEL_RAPTORLAKE_P, &adl_features },
1220 	{ INTEL_RAPTORLAKE_S, &adl_features },
1221 	{ INTEL_BARTLETTLAKE, &adl_features },
1222 	{ INTEL_METEORLAKE, &adl_features },
1223 	{ INTEL_METEORLAKE_L, &adl_features },
1224 	{ INTEL_ARROWLAKE_H, &adl_features },
1225 	{ INTEL_ARROWLAKE_U, &adl_features },
1226 	{ INTEL_ARROWLAKE, &adl_features },
1227 	{ INTEL_LUNARLAKE_M, &lnl_features },
1228 	{ INTEL_PANTHERLAKE_L, &lnl_features },
1229 	{ INTEL_NOVALAKE, &lnl_features },
1230 	{ INTEL_NOVALAKE_L, &lnl_features },
1231 	{ INTEL_WILDCATLAKE_L, &lnl_features },
1232 	{ INTEL_ATOM_SILVERMONT, &slv_features },
1233 	{ INTEL_ATOM_SILVERMONT_D, &slvd_features },
1234 	{ INTEL_ATOM_AIRMONT, &amt_features },
1235 	{ INTEL_ATOM_GOLDMONT, &gmt_features },
1236 	{ INTEL_ATOM_GOLDMONT_D, &gmtd_features },
1237 	{ INTEL_ATOM_GOLDMONT_PLUS, &gmtp_features },
1238 	{ INTEL_ATOM_TREMONT_D, &tmtd_features },
1239 	{ INTEL_ATOM_TREMONT, &tmt_features },
1240 	{ INTEL_ATOM_TREMONT_L, &tmt_features },
1241 	{ INTEL_ATOM_GRACEMONT, &adl_features },
1242 	{ INTEL_ATOM_CRESTMONT_X, &srf_features },
1243 	{ INTEL_ATOM_CRESTMONT, &grr_features },
1244 	{ INTEL_ATOM_DARKMONT_X, &srf_features },
1245 	{ INTEL_XEON_PHI_KNL, &knl_features },
1246 	{ INTEL_XEON_PHI_KNM, &knl_features },
1247 	/*
1248 	 * Missing support for
1249 	 * INTEL_ICELAKE
1250 	 * INTEL_ATOM_SILVERMONT_MID
1251 	 * INTEL_ATOM_SILVERMONT_MID2
1252 	 * INTEL_ATOM_AIRMONT_NP
1253 	 */
1254 	{ 0, NULL },
1255 };
1256 
1257 struct {
1258 	unsigned int uniform;
1259 	unsigned int pcore;
1260 	unsigned int ecore;
1261 	unsigned int lcore;
1262 } perf_pmu_types;
1263 
1264 /*
1265  * Events are enumerated in https://github.com/intel/perfmon
1266  * and tools/perf/pmu-events/arch/x86/.../cache.json
1267  */
1268 struct perf_l2_events {
1269 	unsigned long long refs;	/* L2_REQUEST.ALL */
1270 	unsigned long long hits;	/* L2_REQUEST.HIT */
1271 };
1272 
1273 struct perf_model_support {
1274 	unsigned int vfm;
1275 	struct perf_l2_events first;
1276 	struct perf_l2_events second;
1277 	struct perf_l2_events third;
1278 } *perf_model_support;
1279 
1280 /* Perf Cache Events */
1281 #define	PCE(ext_umask, umask)	(((unsigned long long) ext_umask) << 40 | umask << 8 | 0x24)
1282 
1283 /*
1284  * Enumerate up to three perf CPU PMU's in a system.
1285  * The first, second, and third columns are populated without skipping, describing
1286  * pcore, ecore, lcore PMUs, in order, if present.  (The associated PMU "type" field is
1287  * read from sysfs in all cases.)  Eg.
1288  *
1289  * non-hybrid:
1290  *	GNR: pcore, {}, {}
1291  *	ADL-N: ecore, {}, {}
1292  * hybrid:
1293  *	MTL: pcore, ecore, {}%
1294  *	ARL-H: pcore, ecore, lcore
1295  *	LNL: ecore, ecore%%, {}
1296  *
1297  * % MTL physical lcore share architecture and PMU with ecore, and are thus not enumerated separately.
1298  * %% LNL physical lcore is enumerated by perf as ecore
1299  */
1300 static struct perf_model_support turbostat_perf_model_support[] = {
1301 	{ INTEL_SAPPHIRERAPIDS_X, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, {}, {} },
1302 	{ INTEL_EMERALDRAPIDS_X, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, {}, {} },
1303 	{ INTEL_GRANITERAPIDS_X, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, {}, {} },
1304 	{ INTEL_GRANITERAPIDS_D, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, {}, {} },
1305 	{ INTEL_DIAMONDRAPIDS_X, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, {}, {} },
1306 
1307 	{ INTEL_ATOM_GRACEMONT, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {}, {} },	/* ADL-N */
1308 	{ INTEL_ATOM_CRESTMONT_X, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {}, {} },	/* SRF */
1309 	{ INTEL_ATOM_CRESTMONT, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {}, {} },	/* GRR */
1310 	{ INTEL_ATOM_DARKMONT_X, { PCE(0x01, 0xFF), PCE(0x01, 0xBF)}, {}, {} },	/* CWF */
1311 
1312 	{ INTEL_ALDERLAKE, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1313 	{ INTEL_ALDERLAKE, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1314 	{ INTEL_ALDERLAKE_L, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1315 	{ INTEL_RAPTORLAKE, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1316 	{ INTEL_RAPTORLAKE_P, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1317 	{ INTEL_RAPTORLAKE_S, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1318 	{ INTEL_METEORLAKE_L, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1319 	{ INTEL_METEORLAKE, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1320 	{ INTEL_ARROWLAKE_U, { PCE(0x00, 0xFF), PCE(0x00, 0xDF)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)}, {} },
1321 
1322 	{ INTEL_LUNARLAKE_M, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x00, 0x07), PCE(0x00, 0x02)}, {} },
1323 	{ INTEL_ARROWLAKE_H, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x00, 0x07), PCE(0x00, 0x02)}, { PCE(0x00, 0x00), PCE(0x00, 0x02)} },
1324 	{ INTEL_ARROWLAKE, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x00, 0x07), PCE(0x00, 0x02)}, {} },
1325 
1326 	{ INTEL_PANTHERLAKE_L, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x01, 0xFF), PCE(0x01, 0xBF)}, {} },
1327 	{ INTEL_WILDCATLAKE_L, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x01, 0xFF), PCE(0x01, 0xBF)}, {} },
1328 
1329 	{ INTEL_NOVALAKE, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x01, 0xFF), PCE(0x01, 0xBF)}, {} },
1330 	{ INTEL_NOVALAKE_L, { PCE(0x00, 0xFF), PCE(0x00, 0x5F)}, { PCE(0x01, 0xFF), PCE(0x01, 0xBF)}, {} },
1331 
1332 	{ 0, {}, {}, {} }
1333 };
1334 
1335 static const struct platform_features *platform;
1336 
probe_platform_features(unsigned int family,unsigned int model)1337 void probe_platform_features(unsigned int family, unsigned int model)
1338 {
1339 	int i;
1340 
1341 	if (authentic_amd || hygon_genuine) {
1342 		/* fallback to default features on unsupported models */
1343 		force_load++;
1344 		if (max_extended_level >= 0x80000007) {
1345 			unsigned int eax, ebx, ecx, edx;
1346 
1347 			__cpuid(0x80000007, eax, ebx, ecx, edx);
1348 			/* RAPL (Fam 17h+) */
1349 			if ((edx & (1 << 14)) && family >= 0x17)
1350 				platform = &amd_features_with_rapl;
1351 		}
1352 		goto end;
1353 	}
1354 
1355 	if (!genuine_intel)
1356 		goto end;
1357 
1358 	for (i = 0; turbostat_pdata[i].features; i++) {
1359 		if (VFM_FAMILY(turbostat_pdata[i].vfm) == family && VFM_MODEL(turbostat_pdata[i].vfm) == model) {
1360 			platform = turbostat_pdata[i].features;
1361 			return;
1362 		}
1363 	}
1364 
1365 end:
1366 	if (force_load && !platform) {
1367 		fprintf(outf, "Forced to run on unsupported platform!\n");
1368 		platform = &default_features;
1369 	}
1370 
1371 	if (platform)
1372 		return;
1373 
1374 	fprintf(stderr, "Unsupported platform detected.\n\tSee RUN THE LATEST VERSION on turbostat(8)\n");
1375 	exit(1);
1376 }
1377 
init_perf_model_support(unsigned int family,unsigned int model)1378 void init_perf_model_support(unsigned int family, unsigned int model)
1379 {
1380 	int i;
1381 
1382 	if (!genuine_intel)
1383 		return;
1384 
1385 	for (i = 0; turbostat_perf_model_support[i].vfm; i++) {
1386 		if (VFM_FAMILY(turbostat_perf_model_support[i].vfm) == family && VFM_MODEL(turbostat_perf_model_support[i].vfm) == model) {
1387 			perf_model_support = &turbostat_perf_model_support[i];
1388 			return;
1389 		}
1390 	}
1391 }
1392 
1393 /* Model specific support End */
1394 
1395 #define	TJMAX_DEFAULT	100
1396 
1397 /* MSRs that are not yet in the kernel-provided header. */
1398 #define MSR_RAPL_PWR_UNIT	0xc0010299
1399 #define MSR_CORE_ENERGY_STAT	0xc001029a
1400 #define MSR_PKG_ENERGY_STAT	0xc001029b
1401 
1402 #define MAX(a, b) ((a) > (b) ? (a) : (b))
1403 
1404 int backwards_count;
1405 char *progname;
1406 
1407 #define CPU_SUBSET_MAXCPUS	8192	/* need to use before probe... */
1408 cpu_set_t *cpu_present_set, *cpu_possible_set, *cpu_effective_set, *cpu_allowed_set, *cpu_affinity_set, *cpu_subset;
1409 cpu_set_t *perf_pcore_set, *perf_ecore_set, *perf_lcore_set;
1410 size_t cpu_present_setsize, cpu_possible_setsize, cpu_effective_setsize, cpu_allowed_setsize, cpu_affinity_setsize, cpu_subset_size;
1411 #define MAX_ADDED_THREAD_COUNTERS 24
1412 #define MAX_ADDED_CORE_COUNTERS 8
1413 #define MAX_ADDED_PACKAGE_COUNTERS 16
1414 #define PMT_MAX_ADDED_THREAD_COUNTERS 24
1415 #define PMT_MAX_ADDED_CORE_COUNTERS 8
1416 #define PMT_MAX_ADDED_PACKAGE_COUNTERS 16
1417 #define BITMASK_SIZE 32
1418 
1419 #define ZERO_ARRAY(arr) (memset(arr, 0, sizeof(arr)) + __must_be_array(arr))
1420 
1421 /* Indexes used to map data read from perf and MSRs into global variables */
1422 enum rapl_rci_index {
1423 	RAPL_RCI_INDEX_ENERGY_PKG = 0,
1424 	RAPL_RCI_INDEX_ENERGY_CORES = 1,
1425 	RAPL_RCI_INDEX_DRAM = 2,
1426 	RAPL_RCI_INDEX_GFX = 3,
1427 	RAPL_RCI_INDEX_PKG_PERF_STATUS = 4,
1428 	RAPL_RCI_INDEX_DRAM_PERF_STATUS = 5,
1429 	RAPL_RCI_INDEX_CORE_ENERGY = 6,
1430 	RAPL_RCI_INDEX_ENERGY_PLATFORM = 7,
1431 	NUM_RAPL_COUNTERS,
1432 };
1433 
1434 enum rapl_unit {
1435 	RAPL_UNIT_INVALID,
1436 	RAPL_UNIT_JOULES,
1437 	RAPL_UNIT_WATTS,
1438 };
1439 
1440 struct rapl_counter_info_t {
1441 	unsigned long long data[NUM_RAPL_COUNTERS];
1442 	enum counter_source source[NUM_RAPL_COUNTERS];
1443 	unsigned long long flags[NUM_RAPL_COUNTERS];
1444 	double scale[NUM_RAPL_COUNTERS];
1445 	enum rapl_unit unit[NUM_RAPL_COUNTERS];
1446 	unsigned long long msr[NUM_RAPL_COUNTERS];
1447 	unsigned long long msr_mask[NUM_RAPL_COUNTERS];
1448 	int msr_shift[NUM_RAPL_COUNTERS];
1449 
1450 	int fd_perf;
1451 };
1452 
1453 /* struct rapl_counter_info_t for each RAPL domain */
1454 struct rapl_counter_info_t *rapl_counter_info_perdomain;
1455 unsigned int rapl_counter_info_perdomain_size;
1456 
1457 #define RAPL_COUNTER_FLAG_PLATFORM_COUNTER (1u << 0)
1458 #define RAPL_COUNTER_FLAG_USE_MSR_SUM (1u << 1)
1459 
1460 struct rapl_counter_arch_info {
1461 	int feature_mask;	/* Mask for testing if the counter is supported on host */
1462 	const char *perf_subsys;
1463 	const char *perf_name;
1464 	unsigned long long msr;
1465 	unsigned long long msr_mask;
1466 	int msr_shift;		/* Positive mean shift right, negative mean shift left */
1467 	double *platform_rapl_msr_scale;	/* Scale applied to values read by MSR (platform dependent, filled at runtime) */
1468 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1469 	unsigned int bic_number;
1470 	double compat_scale;	/* Some counters require constant scaling to be in the same range as other, similar ones */
1471 	unsigned long long flags;
1472 };
1473 
1474 static const struct rapl_counter_arch_info rapl_counter_arch_infos[] = {
1475 	{
1476 	 .feature_mask = RAPL_PKG,
1477 	 .perf_subsys = "power",
1478 	 .perf_name = "energy-pkg",
1479 	 .msr = MSR_PKG_ENERGY_STATUS,
1480 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1481 	 .msr_shift = 0,
1482 	 .platform_rapl_msr_scale = &rapl_energy_units,
1483 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1484 	 .bic_number = BIC_PkgWatt,
1485 	 .compat_scale = 1.0,
1486 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1487 	  },
1488 	{
1489 	 .feature_mask = RAPL_PKG,
1490 	 .perf_subsys = "power",
1491 	 .perf_name = "energy-pkg",
1492 	 .msr = MSR_PKG_ENERGY_STATUS,
1493 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1494 	 .msr_shift = 0,
1495 	 .platform_rapl_msr_scale = &rapl_energy_units,
1496 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1497 	 .bic_number = BIC_Pkg_J,
1498 	 .compat_scale = 1.0,
1499 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1500 	  },
1501 	{
1502 	 .feature_mask = RAPL_AMD_F17H,
1503 	 .perf_subsys = "power",
1504 	 .perf_name = "energy-pkg",
1505 	 .msr = MSR_PKG_ENERGY_STAT,
1506 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1507 	 .msr_shift = 0,
1508 	 .platform_rapl_msr_scale = &rapl_energy_units,
1509 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1510 	 .bic_number = BIC_PkgWatt,
1511 	 .compat_scale = 1.0,
1512 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1513 	  },
1514 	{
1515 	 .feature_mask = RAPL_AMD_F17H,
1516 	 .perf_subsys = "power",
1517 	 .perf_name = "energy-pkg",
1518 	 .msr = MSR_PKG_ENERGY_STAT,
1519 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1520 	 .msr_shift = 0,
1521 	 .platform_rapl_msr_scale = &rapl_energy_units,
1522 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
1523 	 .bic_number = BIC_Pkg_J,
1524 	 .compat_scale = 1.0,
1525 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1526 	  },
1527 	{
1528 	 .feature_mask = RAPL_CORE_ENERGY_STATUS,
1529 	 .perf_subsys = "power",
1530 	 .perf_name = "energy-cores",
1531 	 .msr = MSR_PP0_ENERGY_STATUS,
1532 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1533 	 .msr_shift = 0,
1534 	 .platform_rapl_msr_scale = &rapl_energy_units,
1535 	 .rci_index = RAPL_RCI_INDEX_ENERGY_CORES,
1536 	 .bic_number = BIC_CorWatt,
1537 	 .compat_scale = 1.0,
1538 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1539 	  },
1540 	{
1541 	 .feature_mask = RAPL_CORE_ENERGY_STATUS,
1542 	 .perf_subsys = "power",
1543 	 .perf_name = "energy-cores",
1544 	 .msr = MSR_PP0_ENERGY_STATUS,
1545 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1546 	 .msr_shift = 0,
1547 	 .platform_rapl_msr_scale = &rapl_energy_units,
1548 	 .rci_index = RAPL_RCI_INDEX_ENERGY_CORES,
1549 	 .bic_number = BIC_Cor_J,
1550 	 .compat_scale = 1.0,
1551 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1552 	  },
1553 	{
1554 	 .feature_mask = RAPL_DRAM,
1555 	 .perf_subsys = "power",
1556 	 .perf_name = "energy-ram",
1557 	 .msr = MSR_DRAM_ENERGY_STATUS,
1558 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1559 	 .msr_shift = 0,
1560 	 .platform_rapl_msr_scale = &rapl_dram_energy_units,
1561 	 .rci_index = RAPL_RCI_INDEX_DRAM,
1562 	 .bic_number = BIC_RAMWatt,
1563 	 .compat_scale = 1.0,
1564 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1565 	  },
1566 	{
1567 	 .feature_mask = RAPL_DRAM,
1568 	 .perf_subsys = "power",
1569 	 .perf_name = "energy-ram",
1570 	 .msr = MSR_DRAM_ENERGY_STATUS,
1571 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1572 	 .msr_shift = 0,
1573 	 .platform_rapl_msr_scale = &rapl_dram_energy_units,
1574 	 .rci_index = RAPL_RCI_INDEX_DRAM,
1575 	 .bic_number = BIC_RAM_J,
1576 	 .compat_scale = 1.0,
1577 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1578 	  },
1579 	{
1580 	 .feature_mask = RAPL_GFX,
1581 	 .perf_subsys = "power",
1582 	 .perf_name = "energy-gpu",
1583 	 .msr = MSR_PP1_ENERGY_STATUS,
1584 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1585 	 .msr_shift = 0,
1586 	 .platform_rapl_msr_scale = &rapl_energy_units,
1587 	 .rci_index = RAPL_RCI_INDEX_GFX,
1588 	 .bic_number = BIC_GFXWatt,
1589 	 .compat_scale = 1.0,
1590 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1591 	  },
1592 	{
1593 	 .feature_mask = RAPL_GFX,
1594 	 .perf_subsys = "power",
1595 	 .perf_name = "energy-gpu",
1596 	 .msr = MSR_PP1_ENERGY_STATUS,
1597 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1598 	 .msr_shift = 0,
1599 	 .platform_rapl_msr_scale = &rapl_energy_units,
1600 	 .rci_index = RAPL_RCI_INDEX_GFX,
1601 	 .bic_number = BIC_GFX_J,
1602 	 .compat_scale = 1.0,
1603 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1604 	  },
1605 	{
1606 	 .feature_mask = RAPL_PKG_PERF_STATUS,
1607 	 .perf_subsys = NULL,
1608 	 .perf_name = NULL,
1609 	 .msr = MSR_PKG_PERF_STATUS,
1610 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1611 	 .msr_shift = 0,
1612 	 .platform_rapl_msr_scale = &rapl_time_units,
1613 	 .rci_index = RAPL_RCI_INDEX_PKG_PERF_STATUS,
1614 	 .bic_number = BIC_PKG__,
1615 	 .compat_scale = 100.0,
1616 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1617 	  },
1618 	{
1619 	 .feature_mask = RAPL_DRAM_PERF_STATUS,
1620 	 .perf_subsys = NULL,
1621 	 .perf_name = NULL,
1622 	 .msr = MSR_DRAM_PERF_STATUS,
1623 	 .msr_mask = 0xFFFFFFFFFFFFFFFF,
1624 	 .msr_shift = 0,
1625 	 .platform_rapl_msr_scale = &rapl_time_units,
1626 	 .rci_index = RAPL_RCI_INDEX_DRAM_PERF_STATUS,
1627 	 .bic_number = BIC_RAM__,
1628 	 .compat_scale = 100.0,
1629 	 .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
1630 	  },
1631 	{
1632 	 .feature_mask = RAPL_AMD_F17H,
1633 	 .perf_subsys = NULL,
1634 	 .perf_name = NULL,
1635 	 .msr = MSR_CORE_ENERGY_STAT,
1636 	 .msr_mask = 0xFFFFFFFF,
1637 	 .msr_shift = 0,
1638 	 .platform_rapl_msr_scale = &rapl_energy_units,
1639 	 .rci_index = RAPL_RCI_INDEX_CORE_ENERGY,
1640 	 .bic_number = BIC_CorWatt,
1641 	 .compat_scale = 1.0,
1642 	 .flags = 0,
1643 	  },
1644 	{
1645 	 .feature_mask = RAPL_AMD_F17H,
1646 	 .perf_subsys = NULL,
1647 	 .perf_name = NULL,
1648 	 .msr = MSR_CORE_ENERGY_STAT,
1649 	 .msr_mask = 0xFFFFFFFF,
1650 	 .msr_shift = 0,
1651 	 .platform_rapl_msr_scale = &rapl_energy_units,
1652 	 .rci_index = RAPL_RCI_INDEX_CORE_ENERGY,
1653 	 .bic_number = BIC_Cor_J,
1654 	 .compat_scale = 1.0,
1655 	 .flags = 0,
1656 	  },
1657 	{
1658 	 .feature_mask = RAPL_PSYS,
1659 	 .perf_subsys = "power",
1660 	 .perf_name = "energy-psys",
1661 	 .msr = MSR_PLATFORM_ENERGY_STATUS,
1662 	 .msr_mask = 0x00000000FFFFFFFF,
1663 	 .msr_shift = 0,
1664 	 .platform_rapl_msr_scale = &rapl_psys_energy_units,
1665 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PLATFORM,
1666 	 .bic_number = BIC_SysWatt,
1667 	 .compat_scale = 1.0,
1668 	 .flags = RAPL_COUNTER_FLAG_PLATFORM_COUNTER | RAPL_COUNTER_FLAG_USE_MSR_SUM,
1669 	  },
1670 	{
1671 	 .feature_mask = RAPL_PSYS,
1672 	 .perf_subsys = "power",
1673 	 .perf_name = "energy-psys",
1674 	 .msr = MSR_PLATFORM_ENERGY_STATUS,
1675 	 .msr_mask = 0x00000000FFFFFFFF,
1676 	 .msr_shift = 0,
1677 	 .platform_rapl_msr_scale = &rapl_psys_energy_units,
1678 	 .rci_index = RAPL_RCI_INDEX_ENERGY_PLATFORM,
1679 	 .bic_number = BIC_Sys_J,
1680 	 .compat_scale = 1.0,
1681 	 .flags = RAPL_COUNTER_FLAG_PLATFORM_COUNTER | RAPL_COUNTER_FLAG_USE_MSR_SUM,
1682 	  },
1683 };
1684 
1685 struct rapl_counter {
1686 	unsigned long long raw_value;
1687 	enum rapl_unit unit;
1688 	double scale;
1689 };
1690 
1691 /* Indexes used to map data read from perf and MSRs into global variables */
1692 enum ccstate_rci_index {
1693 	CCSTATE_RCI_INDEX_C1_RESIDENCY = 0,
1694 	CCSTATE_RCI_INDEX_C3_RESIDENCY = 1,
1695 	CCSTATE_RCI_INDEX_C6_RESIDENCY = 2,
1696 	CCSTATE_RCI_INDEX_C7_RESIDENCY = 3,
1697 	PCSTATE_RCI_INDEX_C2_RESIDENCY = 4,
1698 	PCSTATE_RCI_INDEX_C3_RESIDENCY = 5,
1699 	PCSTATE_RCI_INDEX_C6_RESIDENCY = 6,
1700 	PCSTATE_RCI_INDEX_C7_RESIDENCY = 7,
1701 	PCSTATE_RCI_INDEX_C8_RESIDENCY = 8,
1702 	PCSTATE_RCI_INDEX_C9_RESIDENCY = 9,
1703 	PCSTATE_RCI_INDEX_C10_RESIDENCY = 10,
1704 	NUM_CSTATE_COUNTERS,
1705 };
1706 
1707 struct cstate_counter_info_t {
1708 	unsigned long long data[NUM_CSTATE_COUNTERS];
1709 	enum counter_source source[NUM_CSTATE_COUNTERS];
1710 	unsigned long long msr[NUM_CSTATE_COUNTERS];
1711 	int fd_perf_core;
1712 	int fd_perf_pkg;
1713 };
1714 
1715 struct cstate_counter_info_t *ccstate_counter_info;
1716 unsigned int ccstate_counter_info_size;
1717 
1718 #define CSTATE_COUNTER_FLAG_COLLECT_PER_CORE   (1u << 0)
1719 #define CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD ((1u << 1) | CSTATE_COUNTER_FLAG_COLLECT_PER_CORE)
1720 #define CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY (1u << 2)
1721 
1722 struct cstate_counter_arch_info {
1723 	int feature_mask;	/* Mask for testing if the counter is supported on host */
1724 	const char *perf_subsys;
1725 	const char *perf_name;
1726 	unsigned long long msr;
1727 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1728 	unsigned int bic_number;
1729 	unsigned long long flags;
1730 	int pkg_cstate_limit;
1731 };
1732 
1733 static struct cstate_counter_arch_info ccstate_counter_arch_infos[] = {
1734 	{
1735 	 .feature_mask = CC1,
1736 	 .perf_subsys = "cstate_core",
1737 	 .perf_name = "c1-residency",
1738 	 .msr = MSR_CORE_C1_RES,
1739 	 .rci_index = CCSTATE_RCI_INDEX_C1_RESIDENCY,
1740 	 .bic_number = BIC_CPU_c1,
1741 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD,
1742 	 .pkg_cstate_limit = 0,
1743 	  },
1744 	{
1745 	 .feature_mask = CC3,
1746 	 .perf_subsys = "cstate_core",
1747 	 .perf_name = "c3-residency",
1748 	 .msr = MSR_CORE_C3_RESIDENCY,
1749 	 .rci_index = CCSTATE_RCI_INDEX_C3_RESIDENCY,
1750 	 .bic_number = BIC_CPU_c3,
1751 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1752 	 .pkg_cstate_limit = 0,
1753 	  },
1754 	{
1755 	 .feature_mask = CC6,
1756 	 .perf_subsys = "cstate_core",
1757 	 .perf_name = "c6-residency",
1758 	 .msr = MSR_CORE_C6_RESIDENCY,
1759 	 .rci_index = CCSTATE_RCI_INDEX_C6_RESIDENCY,
1760 	 .bic_number = BIC_CPU_c6,
1761 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1762 	 .pkg_cstate_limit = 0,
1763 	  },
1764 	{
1765 	 .feature_mask = CC7,
1766 	 .perf_subsys = "cstate_core",
1767 	 .perf_name = "c7-residency",
1768 	 .msr = MSR_CORE_C7_RESIDENCY,
1769 	 .rci_index = CCSTATE_RCI_INDEX_C7_RESIDENCY,
1770 	 .bic_number = BIC_CPU_c7,
1771 	 .flags = CSTATE_COUNTER_FLAG_COLLECT_PER_CORE | CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY,
1772 	 .pkg_cstate_limit = 0,
1773 	  },
1774 	{
1775 	 .feature_mask = PC2,
1776 	 .perf_subsys = "cstate_pkg",
1777 	 .perf_name = "c2-residency",
1778 	 .msr = MSR_PKG_C2_RESIDENCY,
1779 	 .rci_index = PCSTATE_RCI_INDEX_C2_RESIDENCY,
1780 	 .bic_number = BIC_Pkgpc2,
1781 	 .flags = 0,
1782 	 .pkg_cstate_limit = PCL__2,
1783 	  },
1784 	{
1785 	 .feature_mask = PC3,
1786 	 .perf_subsys = "cstate_pkg",
1787 	 .perf_name = "c3-residency",
1788 	 .msr = MSR_PKG_C3_RESIDENCY,
1789 	 .rci_index = PCSTATE_RCI_INDEX_C3_RESIDENCY,
1790 	 .bic_number = BIC_Pkgpc3,
1791 	 .flags = 0,
1792 	 .pkg_cstate_limit = PCL__3,
1793 	  },
1794 	{
1795 	 .feature_mask = PC6,
1796 	 .perf_subsys = "cstate_pkg",
1797 	 .perf_name = "c6-residency",
1798 	 .msr = MSR_PKG_C6_RESIDENCY,
1799 	 .rci_index = PCSTATE_RCI_INDEX_C6_RESIDENCY,
1800 	 .bic_number = BIC_Pkgpc6,
1801 	 .flags = 0,
1802 	 .pkg_cstate_limit = PCL__6,
1803 	  },
1804 	{
1805 	 .feature_mask = PC7,
1806 	 .perf_subsys = "cstate_pkg",
1807 	 .perf_name = "c7-residency",
1808 	 .msr = MSR_PKG_C7_RESIDENCY,
1809 	 .rci_index = PCSTATE_RCI_INDEX_C7_RESIDENCY,
1810 	 .bic_number = BIC_Pkgpc7,
1811 	 .flags = 0,
1812 	 .pkg_cstate_limit = PCL__7,
1813 	  },
1814 	{
1815 	 .feature_mask = PC8,
1816 	 .perf_subsys = "cstate_pkg",
1817 	 .perf_name = "c8-residency",
1818 	 .msr = MSR_PKG_C8_RESIDENCY,
1819 	 .rci_index = PCSTATE_RCI_INDEX_C8_RESIDENCY,
1820 	 .bic_number = BIC_Pkgpc8,
1821 	 .flags = 0,
1822 	 .pkg_cstate_limit = PCL__8,
1823 	  },
1824 	{
1825 	 .feature_mask = PC9,
1826 	 .perf_subsys = "cstate_pkg",
1827 	 .perf_name = "c9-residency",
1828 	 .msr = MSR_PKG_C9_RESIDENCY,
1829 	 .rci_index = PCSTATE_RCI_INDEX_C9_RESIDENCY,
1830 	 .bic_number = BIC_Pkgpc9,
1831 	 .flags = 0,
1832 	 .pkg_cstate_limit = PCL__9,
1833 	  },
1834 	{
1835 	 .feature_mask = PC10,
1836 	 .perf_subsys = "cstate_pkg",
1837 	 .perf_name = "c10-residency",
1838 	 .msr = MSR_PKG_C10_RESIDENCY,
1839 	 .rci_index = PCSTATE_RCI_INDEX_C10_RESIDENCY,
1840 	 .bic_number = BIC_Pkgpc10,
1841 	 .flags = 0,
1842 	 .pkg_cstate_limit = PCL_10,
1843 	  },
1844 };
1845 
1846 /* Indexes used to map data read from perf and MSRs into global variables */
1847 enum msr_rci_index {
1848 	MSR_RCI_INDEX_APERF = 0,
1849 	MSR_RCI_INDEX_MPERF = 1,
1850 	MSR_RCI_INDEX_SMI = 2,
1851 	NUM_MSR_COUNTERS,
1852 };
1853 
1854 struct msr_counter_info_t {
1855 	unsigned long long data[NUM_MSR_COUNTERS];
1856 	enum counter_source source[NUM_MSR_COUNTERS];
1857 	unsigned long long msr[NUM_MSR_COUNTERS];
1858 	unsigned long long msr_mask[NUM_MSR_COUNTERS];
1859 	int fd_perf;
1860 };
1861 
1862 struct msr_counter_info_t *msr_counter_info;
1863 unsigned int msr_counter_info_size;
1864 
1865 struct msr_counter_arch_info {
1866 	const char *perf_subsys;
1867 	const char *perf_name;
1868 	unsigned long long msr;
1869 	unsigned long long msr_mask;
1870 	unsigned int rci_index;	/* Maps data from perf counters to global variables */
1871 	bool needed;
1872 	bool present;
1873 };
1874 
1875 enum msr_arch_info_index {
1876 	MSR_ARCH_INFO_APERF_INDEX = 0,
1877 	MSR_ARCH_INFO_MPERF_INDEX = 1,
1878 	MSR_ARCH_INFO_SMI_INDEX = 2,
1879 };
1880 
1881 static struct msr_counter_arch_info msr_counter_arch_infos[] = {
1882 	[MSR_ARCH_INFO_APERF_INDEX] = {
1883 				       .perf_subsys = "msr",
1884 				       .perf_name = "aperf",
1885 				       .msr = MSR_IA32_APERF,
1886 				       .msr_mask = 0xFFFFFFFFFFFFFFFF,
1887 				       .rci_index = MSR_RCI_INDEX_APERF,
1888 				        },
1889 
1890 	[MSR_ARCH_INFO_MPERF_INDEX] = {
1891 				       .perf_subsys = "msr",
1892 				       .perf_name = "mperf",
1893 				       .msr = MSR_IA32_MPERF,
1894 				       .msr_mask = 0xFFFFFFFFFFFFFFFF,
1895 				       .rci_index = MSR_RCI_INDEX_MPERF,
1896 				        },
1897 
1898 	[MSR_ARCH_INFO_SMI_INDEX] = {
1899 				     .perf_subsys = "msr",
1900 				     .perf_name = "smi",
1901 				     .msr = MSR_SMI_COUNT,
1902 				     .msr_mask = 0xFFFFFFFF,
1903 				     .rci_index = MSR_RCI_INDEX_SMI,
1904 				      },
1905 };
1906 
1907 /* Can be redefined when compiling, useful for testing. */
1908 #ifndef SYSFS_TELEM_PATH
1909 #define SYSFS_TELEM_PATH "/sys/class/intel_pmt"
1910 #endif
1911 
1912 #define PMT_COUNTER_MTL_DC6_OFFSET 120
1913 #define PMT_COUNTER_MTL_DC6_LSB    0
1914 #define PMT_COUNTER_MTL_DC6_MSB    63
1915 #define PMT_MTL_DC6_GUID           0x1a067102
1916 #define PMT_MTL_DC6_SEQ            0
1917 
1918 #define PMT_COUNTER_CWF_MC1E_OFFSET_BASE          20936
1919 #define PMT_COUNTER_CWF_MC1E_OFFSET_INCREMENT     24
1920 #define PMT_COUNTER_CWF_MC1E_NUM_MODULES_PER_FILE 12
1921 #define PMT_COUNTER_CWF_CPUS_PER_MODULE           4
1922 #define PMT_COUNTER_CWF_MC1E_LSB                  0
1923 #define PMT_COUNTER_CWF_MC1E_MSB                  63
1924 #define PMT_CWF_MC1E_GUID                         0x14421519
1925 
1926 unsigned long long tcore_clock_freq_hz = 800000000;
1927 
1928 #define PMT_COUNTER_NAME_SIZE_BYTES      16
1929 #define PMT_COUNTER_TYPE_NAME_SIZE_BYTES 32
1930 
1931 struct pmt_mmio {
1932 	struct pmt_mmio *next;
1933 
1934 	unsigned int guid;
1935 	unsigned int size;
1936 
1937 	/* Base pointer to the mmaped memory. */
1938 	void *mmio_base;
1939 
1940 	/*
1941 	 * Offset to be applied to the mmio_base
1942 	 * to get the beginning of the PMT counters for given GUID.
1943 	 */
1944 	unsigned long pmt_offset;
1945 } *pmt_mmios;
1946 
1947 enum pmt_datatype {
1948 	PMT_TYPE_RAW,
1949 	PMT_TYPE_XTAL_TIME,
1950 	PMT_TYPE_TCORE_CLOCK,
1951 };
1952 
1953 struct pmt_domain_info {
1954 	/*
1955 	 * Pointer to the MMIO obtained by applying a counter offset
1956 	 * to the mmio_base of the mmaped region for the given GUID.
1957 	 *
1958 	 * This is where to read the raw value of the counter from.
1959 	 */
1960 	unsigned long *pcounter;
1961 };
1962 
1963 struct pmt_counter {
1964 	struct pmt_counter *next;
1965 
1966 	/* PMT metadata */
1967 	char name[PMT_COUNTER_NAME_SIZE_BYTES];
1968 	enum pmt_datatype type;
1969 	enum counter_scope scope;
1970 	unsigned int lsb;
1971 	unsigned int msb;
1972 
1973 	/* BIC-like metadata */
1974 	enum counter_format format;
1975 
1976 	unsigned int num_domains;
1977 	struct pmt_domain_info *domains;
1978 };
1979 
1980 /*
1981  * PMT telemetry directory iterator.
1982  * Used to iterate telemetry files in sysfs in correct order.
1983  */
1984 struct pmt_diriter_t {
1985 	DIR *dir;
1986 	struct dirent **namelist;
1987 	unsigned int num_names;
1988 	unsigned int current_name_idx;
1989 };
1990 
pmt_telemdir_filter(const struct dirent * e)1991 int pmt_telemdir_filter(const struct dirent *e)
1992 {
1993 	unsigned int dummy;
1994 
1995 	return sscanf(e->d_name, "telem%u", &dummy);
1996 }
1997 
pmt_telemdir_sort(const struct dirent ** a,const struct dirent ** b)1998 int pmt_telemdir_sort(const struct dirent **a, const struct dirent **b)
1999 {
2000 	unsigned int aidx = 0, bidx = 0;
2001 
2002 	sscanf((*a)->d_name, "telem%u", &aidx);
2003 	sscanf((*b)->d_name, "telem%u", &bidx);
2004 
2005 	return (aidx > bidx) ? 1 : (aidx < bidx) ? -1 : 0;
2006 }
2007 
pmt_diriter_next(struct pmt_diriter_t * iter)2008 const struct dirent *pmt_diriter_next(struct pmt_diriter_t *iter)
2009 {
2010 	const struct dirent *ret = NULL;
2011 
2012 	if (!iter->dir)
2013 		return NULL;
2014 
2015 	if (iter->current_name_idx >= iter->num_names)
2016 		return NULL;
2017 
2018 	ret = iter->namelist[iter->current_name_idx];
2019 	++iter->current_name_idx;
2020 
2021 	return ret;
2022 }
2023 
pmt_diriter_begin(struct pmt_diriter_t * iter,const char * pmt_root_path)2024 const struct dirent *pmt_diriter_begin(struct pmt_diriter_t *iter, const char *pmt_root_path)
2025 {
2026 	int num_names = iter->num_names;
2027 
2028 	if (!iter->dir) {
2029 		iter->dir = opendir(pmt_root_path);
2030 		if (iter->dir == NULL)
2031 			return NULL;
2032 
2033 		num_names = scandir(pmt_root_path, &iter->namelist, pmt_telemdir_filter, pmt_telemdir_sort);
2034 		if (num_names == -1)
2035 			return NULL;
2036 	}
2037 
2038 	iter->current_name_idx = 0;
2039 	iter->num_names = num_names;
2040 
2041 	return pmt_diriter_next(iter);
2042 }
2043 
pmt_diriter_init(struct pmt_diriter_t * iter)2044 void pmt_diriter_init(struct pmt_diriter_t *iter)
2045 {
2046 	memset(iter, 0, sizeof(*iter));
2047 }
2048 
pmt_diriter_remove(struct pmt_diriter_t * iter)2049 void pmt_diriter_remove(struct pmt_diriter_t *iter)
2050 {
2051 	if (iter->namelist) {
2052 		for (unsigned int i = 0; i < iter->num_names; i++) {
2053 			free(iter->namelist[i]);
2054 			iter->namelist[i] = NULL;
2055 		}
2056 	}
2057 
2058 	free(iter->namelist);
2059 	iter->namelist = NULL;
2060 	iter->num_names = 0;
2061 	iter->current_name_idx = 0;
2062 
2063 	closedir(iter->dir);
2064 	iter->dir = NULL;
2065 }
2066 
pmt_counter_get_width(const struct pmt_counter * p)2067 unsigned int pmt_counter_get_width(const struct pmt_counter *p)
2068 {
2069 	return (p->msb - p->lsb) + 1;
2070 }
2071 
pmt_counter_resize_(struct pmt_counter * pcounter,unsigned int new_size)2072 void pmt_counter_resize_(struct pmt_counter *pcounter, unsigned int new_size)
2073 {
2074 	struct pmt_domain_info *new_mem;
2075 
2076 	new_mem = (struct pmt_domain_info *)reallocarray(pcounter->domains, new_size, sizeof(*pcounter->domains));
2077 	if (!new_mem) {
2078 		fprintf(stderr, "%s: failed to allocate memory for PMT counters\n", __func__);
2079 		exit(1);
2080 	}
2081 
2082 	/* Zero initialize just allocated memory. */
2083 	const size_t num_new_domains = new_size - pcounter->num_domains;
2084 
2085 	memset(&new_mem[pcounter->num_domains], 0, num_new_domains * sizeof(*pcounter->domains));
2086 
2087 	pcounter->num_domains = new_size;
2088 	pcounter->domains = new_mem;
2089 }
2090 
pmt_counter_resize(struct pmt_counter * pcounter,unsigned int new_size)2091 void pmt_counter_resize(struct pmt_counter *pcounter, unsigned int new_size)
2092 {
2093 	/*
2094 	 * Allocate more memory ahead of time.
2095 	 *
2096 	 * Always allocate space for at least 8 elements
2097 	 * and double the size when growing.
2098 	 */
2099 	if (new_size < 8)
2100 		new_size = 8;
2101 	new_size = MAX(new_size, pcounter->num_domains * 2);
2102 
2103 	pmt_counter_resize_(pcounter, new_size);
2104 }
2105 
2106 struct llc_stats {
2107 	unsigned long long references;
2108 	unsigned long long misses;
2109 };
2110 struct l2_stats {
2111 	unsigned long long references;
2112 	unsigned long long hits;
2113 };
2114 struct thread_data {
2115 	struct timeval tv_begin;
2116 	struct timeval tv_end;
2117 	struct timeval tv_delta;
2118 	unsigned long long tsc;
2119 	unsigned long long aperf;
2120 	unsigned long long mperf;
2121 	unsigned long long c1;
2122 	unsigned long long instr_count;
2123 	unsigned long long irq_count;
2124 	unsigned long long nmi_count;
2125 	unsigned int smi_count;
2126 	struct llc_stats llc;
2127 	struct l2_stats l2;
2128 	unsigned int cpu_id;
2129 	unsigned int apic_id;
2130 	unsigned int x2apic_id;
2131 	unsigned int flags;
2132 	bool is_atom;
2133 	unsigned long long counter[MAX_ADDED_THREAD_COUNTERS];
2134 	unsigned long long perf_counter[MAX_ADDED_THREAD_COUNTERS];
2135 	unsigned long long pmt_counter[PMT_MAX_ADDED_THREAD_COUNTERS];
2136 };
2137 
2138 struct core_data {
2139 	int first_cpu;
2140 	unsigned long long c3;
2141 	unsigned long long c6;
2142 	unsigned long long c7;
2143 	unsigned long long mc6_us;	/* duplicate as per-core for now, even though per module */
2144 	unsigned int core_temp_c;
2145 	struct rapl_counter core_energy;	/* MSR_CORE_ENERGY_STAT */
2146 	unsigned long long core_throt_cnt;
2147 	unsigned long long counter[MAX_ADDED_CORE_COUNTERS];
2148 	unsigned long long perf_counter[MAX_ADDED_CORE_COUNTERS];
2149 	unsigned long long pmt_counter[PMT_MAX_ADDED_CORE_COUNTERS];
2150 };
2151 
2152 struct pkg_data {
2153 	int first_cpu;
2154 	unsigned long long pc2;
2155 	unsigned long long pc3;
2156 	unsigned long long pc6;
2157 	unsigned long long pc7;
2158 	unsigned long long pc8;
2159 	unsigned long long pc9;
2160 	unsigned long long pc10;
2161 	long long cpu_lpi;
2162 	long long sys_lpi;
2163 	unsigned long long pkg_wtd_core_c0;
2164 	unsigned long long pkg_any_core_c0;
2165 	unsigned long long pkg_any_gfxe_c0;
2166 	unsigned long long pkg_both_core_gfxe_c0;
2167 	long long gfx_rc6_ms;
2168 	unsigned int gfx_mhz;
2169 	unsigned int gfx_act_mhz;
2170 	long long sam_mc6_ms;
2171 	unsigned int sam_mhz;
2172 	unsigned int sam_act_mhz;
2173 	struct rapl_counter energy_pkg;	/* MSR_PKG_ENERGY_STATUS */
2174 	struct rapl_counter energy_dram;	/* MSR_DRAM_ENERGY_STATUS */
2175 	struct rapl_counter energy_cores;	/* MSR_PP0_ENERGY_STATUS */
2176 	struct rapl_counter energy_gfx;	/* MSR_PP1_ENERGY_STATUS */
2177 	struct rapl_counter rapl_pkg_perf_status;	/* MSR_PKG_PERF_STATUS */
2178 	struct rapl_counter rapl_dram_perf_status;	/* MSR_DRAM_PERF_STATUS */
2179 	unsigned int pkg_temp_c;
2180 	unsigned int uncore_mhz;
2181 	unsigned long long die_c6;
2182 	unsigned long long counter[MAX_ADDED_PACKAGE_COUNTERS];
2183 	unsigned long long perf_counter[MAX_ADDED_PACKAGE_COUNTERS];
2184 	unsigned long long pmt_counter[PMT_MAX_ADDED_PACKAGE_COUNTERS];
2185 };
2186 
2187 #define ODD_COUNTERS odd.threads, odd.cores, odd.packages
2188 #define EVEN_COUNTERS even.threads, even.cores, even.packages
2189 
2190 /*
2191  * The accumulated sum of MSR is defined as a monotonic
2192  * increasing MSR, it will be accumulated periodically,
2193  * despite its register's bit width.
2194  */
2195 enum {
2196 	IDX_PKG_ENERGY,
2197 	IDX_DRAM_ENERGY,
2198 	IDX_PP0_ENERGY,
2199 	IDX_PP1_ENERGY,
2200 	IDX_PKG_PERF,
2201 	IDX_DRAM_PERF,
2202 	IDX_PSYS_ENERGY,
2203 	IDX_COUNT,
2204 };
2205 
2206 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr);
2207 
2208 struct msr_sum_array {
2209 	/* get_msr_sum() = sum + (get_msr() - last) */
2210 	struct {
2211 		/*The accumulated MSR value is updated by the timer */
2212 		unsigned long long sum;
2213 		/*The MSR footprint recorded in last timer */
2214 		unsigned long long last;
2215 	} entries[IDX_COUNT];
2216 };
2217 
2218 /* The percpu MSR sum array.*/
2219 struct msr_sum_array *per_cpu_msr_sum;
2220 
idx_to_offset(int idx)2221 off_t idx_to_offset(int idx)
2222 {
2223 	off_t offset;
2224 
2225 	switch (idx) {
2226 	case IDX_PKG_ENERGY:
2227 		if (platform->plat_rapl_msrs & RAPL_AMD_F17H)
2228 			offset = MSR_PKG_ENERGY_STAT;
2229 		else
2230 			offset = MSR_PKG_ENERGY_STATUS;
2231 		break;
2232 	case IDX_DRAM_ENERGY:
2233 		offset = MSR_DRAM_ENERGY_STATUS;
2234 		break;
2235 	case IDX_PP0_ENERGY:
2236 		offset = MSR_PP0_ENERGY_STATUS;
2237 		break;
2238 	case IDX_PP1_ENERGY:
2239 		offset = MSR_PP1_ENERGY_STATUS;
2240 		break;
2241 	case IDX_PKG_PERF:
2242 		offset = MSR_PKG_PERF_STATUS;
2243 		break;
2244 	case IDX_DRAM_PERF:
2245 		offset = MSR_DRAM_PERF_STATUS;
2246 		break;
2247 	case IDX_PSYS_ENERGY:
2248 		offset = MSR_PLATFORM_ENERGY_STATUS;
2249 		break;
2250 	default:
2251 		offset = -1;
2252 	}
2253 	return offset;
2254 }
2255 
offset_to_idx(off_t offset)2256 int offset_to_idx(off_t offset)
2257 {
2258 	int idx;
2259 
2260 	switch (offset) {
2261 	case MSR_PKG_ENERGY_STATUS:
2262 	case MSR_PKG_ENERGY_STAT:
2263 		idx = IDX_PKG_ENERGY;
2264 		break;
2265 	case MSR_DRAM_ENERGY_STATUS:
2266 		idx = IDX_DRAM_ENERGY;
2267 		break;
2268 	case MSR_PP0_ENERGY_STATUS:
2269 		idx = IDX_PP0_ENERGY;
2270 		break;
2271 	case MSR_PP1_ENERGY_STATUS:
2272 		idx = IDX_PP1_ENERGY;
2273 		break;
2274 	case MSR_PKG_PERF_STATUS:
2275 		idx = IDX_PKG_PERF;
2276 		break;
2277 	case MSR_DRAM_PERF_STATUS:
2278 		idx = IDX_DRAM_PERF;
2279 		break;
2280 	case MSR_PLATFORM_ENERGY_STATUS:
2281 		idx = IDX_PSYS_ENERGY;
2282 		break;
2283 	default:
2284 		idx = -1;
2285 	}
2286 	return idx;
2287 }
2288 
idx_valid(int idx)2289 int idx_valid(int idx)
2290 {
2291 	switch (idx) {
2292 	case IDX_PKG_ENERGY:
2293 		return valid_rapl_msrs & (RAPL_PKG | RAPL_AMD_F17H);
2294 	case IDX_DRAM_ENERGY:
2295 		return valid_rapl_msrs & RAPL_DRAM;
2296 	case IDX_PP0_ENERGY:
2297 		return valid_rapl_msrs & RAPL_CORE_ENERGY_STATUS;
2298 	case IDX_PP1_ENERGY:
2299 		return valid_rapl_msrs & RAPL_GFX;
2300 	case IDX_PKG_PERF:
2301 		return valid_rapl_msrs & RAPL_PKG_PERF_STATUS;
2302 	case IDX_DRAM_PERF:
2303 		return valid_rapl_msrs & RAPL_DRAM_PERF_STATUS;
2304 	case IDX_PSYS_ENERGY:
2305 		return valid_rapl_msrs & RAPL_PSYS;
2306 	default:
2307 		return 0;
2308 	}
2309 }
2310 
2311 struct sys_counters {
2312 	/* MSR added counters */
2313 	unsigned int added_thread_counters;
2314 	unsigned int added_core_counters;
2315 	unsigned int added_package_counters;
2316 	struct msr_counter *tp;
2317 	struct msr_counter *cp;
2318 	struct msr_counter *pp;
2319 
2320 	/* perf added counters */
2321 	unsigned int added_thread_perf_counters;
2322 	unsigned int added_core_perf_counters;
2323 	unsigned int added_package_perf_counters;
2324 	struct perf_counter_info *perf_tp;
2325 	struct perf_counter_info *perf_cp;
2326 	struct perf_counter_info *perf_pp;
2327 
2328 	struct pmt_counter *pmt_tp;
2329 	struct pmt_counter *pmt_cp;
2330 	struct pmt_counter *pmt_pp;
2331 } sys;
2332 
free_msr_counters_(struct msr_counter ** pp)2333 static size_t free_msr_counters_(struct msr_counter **pp)
2334 {
2335 	struct msr_counter *p = NULL;
2336 	size_t num_freed = 0;
2337 
2338 	while (*pp) {
2339 		p = *pp;
2340 
2341 		if (p->msr_num != 0) {
2342 			*pp = p->next;
2343 
2344 			free(p);
2345 			++num_freed;
2346 
2347 			continue;
2348 		}
2349 
2350 		pp = &p->next;
2351 	}
2352 
2353 	return num_freed;
2354 }
2355 
2356 /*
2357  * Free all added counters accessed via msr.
2358  */
free_sys_msr_counters(void)2359 static void free_sys_msr_counters(void)
2360 {
2361 	/* Thread counters */
2362 	sys.added_thread_counters -= free_msr_counters_(&sys.tp);
2363 
2364 	/* Core counters */
2365 	sys.added_core_counters -= free_msr_counters_(&sys.cp);
2366 
2367 	/* Package counters */
2368 	sys.added_package_counters -= free_msr_counters_(&sys.pp);
2369 }
2370 
2371 struct counters {
2372 	struct thread_data *threads;
2373 	struct core_data *cores;
2374 	struct pkg_data *packages;
2375 } average, even, odd;
2376 
2377 struct platform_counters {
2378 	struct rapl_counter energy_psys;	/* MSR_PLATFORM_ENERGY_STATUS */
2379 } platform_counters_odd, platform_counters_even;
2380 
2381 #define	MAX_HT_ID	3	/* support SMT-4 */
2382 
2383 struct cpu_topology {
2384 	int cpu_id;
2385 	int core_id;		/* unique within a package */
2386 	int package_id;
2387 	int die_id;
2388 	int l3_id;
2389 	int physical_node_id;
2390 	int logical_node_id;	/* 0-based count within the package */
2391 	int ht_id;		/* unique within a core */
2392 	int ht_sibling_cpu_id[MAX_HT_ID + 1];
2393 	int type;
2394 	cpu_set_t *put_ids;	/* Processing Unit/Thread IDs */
2395 } *cpus;
2396 
2397 struct topo_params {
2398 	int num_packages;
2399 	int num_die;
2400 	int num_cpus;
2401 	int num_cores;		/* system wide */
2402 	int allowed_packages;
2403 	int allowed_cpus;
2404 	int allowed_cores;
2405 	int max_cpu_num;
2406 	int max_core_id;	/* within a package */
2407 	int max_package_id;
2408 	int max_die_id;
2409 	int max_l3_id;
2410 	int max_node_num;
2411 	int nodes_per_pkg;
2412 	int cores_per_node;
2413 	int threads_per_core;
2414 } topo;
2415 
2416 struct timeval tv_even, tv_odd, tv_delta;
2417 
2418 int *irq_column_2_cpu;		/* /proc/interrupts column numbers */
2419 int *irqs_per_cpu;		/* indexed by cpu_num */
2420 int *nmi_per_cpu;		/* indexed by cpu_num */
2421 
2422 void setup_all_buffers(bool startup);
2423 
2424 char *sys_lpi_file;
2425 char *sys_lpi_file_sysfs = "/sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us";
2426 char *sys_lpi_file_debugfs = "/sys/kernel/debug/pmc_core/slp_s0_residency_usec";
2427 
cpu_is_not_present(int cpu)2428 int cpu_is_not_present(int cpu)
2429 {
2430 	return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
2431 }
2432 
cpu_is_not_allowed(int cpu)2433 int cpu_is_not_allowed(int cpu)
2434 {
2435 	return !CPU_ISSET_S(cpu, cpu_allowed_setsize, cpu_allowed_set);
2436 }
2437 
2438 #define GLOBAL_CORE_ID(core_id, pkg_id)	(core_id + pkg_id * (topo.max_core_id + 1))
2439 /*
2440  * run func(thread, core, package) in topology order
2441  * skip non-present cpus
2442  */
2443 
2444 #define PER_THREAD_PARAMS  struct thread_data *t, struct core_data *c, struct pkg_data *p
2445 
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)2446 int for_all_cpus(int (func) (struct thread_data *, struct core_data *, struct pkg_data *),
2447 		 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
2448 {
2449 	int cpu, retval;
2450 
2451 	retval = 0;
2452 
2453 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
2454 		struct thread_data *t;
2455 		struct core_data *c;
2456 		struct pkg_data *p;
2457 
2458 		int pkg_id = cpus[cpu].package_id;
2459 
2460 		if (cpu_is_not_allowed(cpu))
2461 			continue;
2462 
2463 		if (cpus[cpu].ht_id > 0)	/* skip HT sibling */
2464 			continue;
2465 
2466 		t = &thread_base[cpu];
2467 		c = &core_base[GLOBAL_CORE_ID(cpus[cpu].core_id, pkg_id)];
2468 		p = &pkg_base[pkg_id];
2469 
2470 		retval |= func(t, c, p);
2471 
2472 		/* Handle HT sibling now */
2473 		int i;
2474 
2475 		for (i = MAX_HT_ID; i > 0; --i) {	/* ht_id 0 is self */
2476 			if (cpus[cpu].ht_sibling_cpu_id[i] <= 0)
2477 				continue;
2478 			t = &thread_base[cpus[cpu].ht_sibling_cpu_id[i]];
2479 
2480 			retval |= func(t, c, p);
2481 		}
2482 	}
2483 	return retval;
2484 }
2485 
is_cpu_first_thread_in_core(struct thread_data * t,struct core_data * c)2486 int is_cpu_first_thread_in_core(struct thread_data *t, struct core_data *c)
2487 {
2488 	return ((int)t->cpu_id == c->first_cpu || c->first_cpu < 0);
2489 }
2490 
is_cpu_first_core_in_package(struct thread_data * t,struct pkg_data * p)2491 int is_cpu_first_core_in_package(struct thread_data *t, struct pkg_data *p)
2492 {
2493 	return ((int)t->cpu_id == p->first_cpu || p->first_cpu < 0);
2494 }
2495 
is_cpu_first_thread_in_package(struct thread_data * t,struct core_data * c,struct pkg_data * p)2496 int is_cpu_first_thread_in_package(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2497 {
2498 	return is_cpu_first_thread_in_core(t, c) && is_cpu_first_core_in_package(t, p);
2499 }
2500 
cpu_migrate(int cpu)2501 int cpu_migrate(int cpu)
2502 {
2503 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2504 	CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
2505 	if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
2506 		return -1;
2507 	else
2508 		return 0;
2509 }
2510 
get_msr_fd(int cpu)2511 int get_msr_fd(int cpu)
2512 {
2513 	char pathname[32];
2514 	int fd;
2515 
2516 	fd = fd_percpu[cpu];
2517 
2518 	if (fd)
2519 		return fd;
2520 	sprintf(pathname, use_android_msr_path ? "/dev/msr%d" : "/dev/cpu/%d/msr", cpu);
2521 	fd = open(pathname, O_RDONLY);
2522 	if (fd < 0)
2523 		err(-1, "%s open failed, try chown or chmod +r %s, "
2524 		    "or run with --no-msr, or run as root", pathname, use_android_msr_path ? "/dev/msr*" : "/dev/cpu/*/msr");
2525 	fd_percpu[cpu] = fd;
2526 
2527 	return fd;
2528 }
2529 
bic_disable_msr_access(void)2530 static void bic_disable_msr_access(void)
2531 {
2532 	CLR_BIC(BIC_Mod_c6, &bic_enabled);
2533 	CLR_BIC(BIC_CoreTmp, &bic_enabled);
2534 	CLR_BIC(BIC_Totl_c0, &bic_enabled);
2535 	CLR_BIC(BIC_Any_c0, &bic_enabled);
2536 	CLR_BIC(BIC_GFX_c0, &bic_enabled);
2537 	CLR_BIC(BIC_CPUGFX, &bic_enabled);
2538 	CLR_BIC(BIC_PkgTmp, &bic_enabled);
2539 
2540 	free_sys_msr_counters();
2541 }
2542 
bic_disable_perf_access(void)2543 static void bic_disable_perf_access(void)
2544 {
2545 	CLR_BIC(BIC_IPC, &bic_enabled);
2546 	CLR_BIC(BIC_LLC_MRPS, &bic_enabled);
2547 	CLR_BIC(BIC_LLC_HIT, &bic_enabled);
2548 	CLR_BIC(BIC_L2_MRPS, &bic_enabled);
2549 	CLR_BIC(BIC_L2_HIT, &bic_enabled);
2550 }
2551 
perf_event_open(struct perf_event_attr * hw_event,pid_t pid,int cpu,int group_fd,unsigned long flags)2552 static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid, int cpu, int group_fd, unsigned long flags)
2553 {
2554 	assert(!no_perf);
2555 
2556 	return syscall(__NR_perf_event_open, hw_event, pid, cpu, group_fd, flags);
2557 }
2558 
open_perf_counter(int cpu,unsigned int type,unsigned int config,int group_fd,__u64 read_format)2559 static long open_perf_counter(int cpu, unsigned int type, unsigned int config, int group_fd, __u64 read_format)
2560 {
2561 	struct perf_event_attr attr;
2562 	const pid_t pid = -1;
2563 	const unsigned long flags = 0;
2564 
2565 	assert(!no_perf);
2566 
2567 	memset(&attr, 0, sizeof(struct perf_event_attr));
2568 
2569 	attr.type = type;
2570 	attr.size = sizeof(struct perf_event_attr);
2571 	attr.config = config;
2572 	attr.disabled = 0;
2573 	attr.sample_type = PERF_SAMPLE_IDENTIFIER;
2574 	attr.read_format = read_format;
2575 
2576 	const int fd = perf_event_open(&attr, pid, cpu, group_fd, flags);
2577 
2578 	return fd;
2579 }
2580 
get_instr_count_fd(int cpu)2581 int get_instr_count_fd(int cpu)
2582 {
2583 	if (fd_instr_count_percpu[cpu])
2584 		return fd_instr_count_percpu[cpu];
2585 
2586 	fd_instr_count_percpu[cpu] = open_perf_counter(cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, -1, 0);
2587 
2588 	return fd_instr_count_percpu[cpu];
2589 }
2590 
get_msr(int cpu,off_t offset,unsigned long long * msr)2591 int get_msr(int cpu, off_t offset, unsigned long long *msr)
2592 {
2593 	ssize_t retval;
2594 
2595 	assert(!no_msr);
2596 
2597 	retval = pread(get_msr_fd(cpu), msr, sizeof(*msr), offset);
2598 
2599 	if (retval != sizeof *msr)
2600 		err(-1, "cpu%d: msr offset 0x%llx read failed", cpu, (unsigned long long)offset);
2601 
2602 	return 0;
2603 }
2604 
add_msr_counter(int cpu,off_t offset)2605 int add_msr_counter(int cpu, off_t offset)
2606 {
2607 	ssize_t retval;
2608 	unsigned long long value;
2609 
2610 	if (no_msr)
2611 		return -1;
2612 
2613 	if (!offset)
2614 		return -1;
2615 
2616 	retval = pread(get_msr_fd(cpu), &value, sizeof(value), offset);
2617 
2618 	/* if the read failed, the probe fails */
2619 	if (retval != sizeof(value))
2620 		return -1;
2621 
2622 	if (value == 0)
2623 		return 0;
2624 
2625 	return 1;
2626 }
2627 
add_rapl_msr_counter(int cpu,const struct rapl_counter_arch_info * cai)2628 int add_rapl_msr_counter(int cpu, const struct rapl_counter_arch_info *cai)
2629 {
2630 	int ret;
2631 
2632 	if (!(valid_rapl_msrs & cai->feature_mask))
2633 		return -1;
2634 
2635 	ret = add_msr_counter(cpu, cai->msr);
2636 	if (ret < 0)
2637 		return -1;
2638 
2639 	switch (cai->rci_index) {
2640 	case RAPL_RCI_INDEX_ENERGY_PKG:
2641 	case RAPL_RCI_INDEX_ENERGY_CORES:
2642 	case RAPL_RCI_INDEX_DRAM:
2643 	case RAPL_RCI_INDEX_GFX:
2644 	case RAPL_RCI_INDEX_ENERGY_PLATFORM:
2645 		if (ret == 0)
2646 			return 1;
2647 	}
2648 
2649 	/* PKG,DRAM_PERF_STATUS MSRs, can return any value */
2650 	return 1;
2651 }
2652 
2653 /* Convert CPU ID to domain ID for given added perf counter. */
cpu_to_domain(const struct perf_counter_info * pc,int cpu)2654 unsigned int cpu_to_domain(const struct perf_counter_info *pc, int cpu)
2655 {
2656 	switch (pc->scope) {
2657 	case SCOPE_CPU:
2658 		return cpu;
2659 
2660 	case SCOPE_CORE:
2661 		return cpus[cpu].core_id;
2662 
2663 	case SCOPE_PACKAGE:
2664 		return cpus[cpu].package_id;
2665 	}
2666 
2667 	__builtin_unreachable();
2668 }
2669 
2670 #define MAX_DEFERRED 16
2671 char *deferred_add_names[MAX_DEFERRED];
2672 char *deferred_skip_names[MAX_DEFERRED];
2673 int deferred_add_index;
2674 int deferred_skip_index;
2675 unsigned int deferred_add_consumed;
2676 unsigned int deferred_skip_consumed;
2677 
2678 /*
2679  * HIDE_LIST - hide this list of counters, show the rest [default]
2680  * SHOW_LIST - show this list of counters, hide the rest
2681  */
2682 enum show_hide_mode { SHOW_LIST, HIDE_LIST } global_show_hide_mode = HIDE_LIST;
2683 
help(void)2684 void help(void)
2685 {
2686 	fprintf(outf,
2687 		"Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
2688 		"\n"
2689 		"Turbostat forks the specified COMMAND and prints statistics\n"
2690 		"when COMMAND completes.\n"
2691 		"If no COMMAND is specified, turbostat wakes every 5-seconds\n"
2692 		"to print statistics, until interrupted.\n"
2693 		"  -a, --add counter\n"
2694 		"		add a counter\n"
2695 		"		  eg. --add msr0x10,u64,cpu,delta,MY_TSC\n"
2696 		"		  eg. --add perf/cstate_pkg/c2-residency,package,delta,percent,perfPC2\n"
2697 		"		  eg. --add pmt,name=XTAL,type=raw,domain=package0,offset=0,lsb=0,msb=63,guid=0x1a067102\n"
2698 		"  -c, --cpu cpu-set\n"
2699 		"		limit output to summary plus cpu-set:\n"
2700 		"		  {core | package | j,k,l..m,n-p }\n"
2701 		"  -d, --debug\n"
2702 		"		displays usec, Time_Of_Day_Seconds and more debugging\n"
2703 		"		debug messages are printed to stderr\n"
2704 		"  -D, --Dump\n"
2705 		"		displays the raw counter values\n"
2706 		"  -e, --enable [all | column]\n"
2707 		"		shows all or the specified disabled column\n"
2708 		"  -f, --force\n"
2709 		"		force load turbostat with minimum default features on unsupported platforms.\n"
2710 		"  -H, --hide [column | column,column,...]\n"
2711 		"		hide the specified column(s)\n"
2712 		"  -i, --interval sec.subsec\n"
2713 		"		override default 5-second measurement interval\n"
2714 		"  -J, --Joules\n"
2715 		"		displays energy in Joules instead of Watts\n"
2716 		"  -l, --list\n"
2717 		"		list column headers only\n"
2718 		"  -M, --no-msr\n"
2719 		"		disable all uses of the MSR driver\n"
2720 		"  -P, --no-perf\n"
2721 		"		disable all uses of the perf API\n"
2722 		"  -n, --num_iterations num\n"
2723 		"		number of the measurement iterations\n"
2724 		"  -N, --header_iterations num\n"
2725 		"		print header every num iterations\n"
2726 		"  -o, --out file\n"
2727 		"		create or truncate \"file\" for all output\n"
2728 		"  -q, --quiet\n"
2729 		"		skip decoding system configuration header\n"
2730 		"  -s, --show [column | column,column,...]\n"
2731 		"		show only the specified column(s)\n"
2732 		"  -S, --Summary\n"
2733 		"		limits output to 1-line system summary per interval\n"
2734 		"  -T, --TCC temperature\n"
2735 		"		sets the Thermal Control Circuit temperature in\n"
2736 		"		  degrees Celsius\n"
2737 		"  -h, --help\n"
2738 		"		print this help message\n  -v, --version\n\t\tprint version information\n\nFor more help, run \"man turbostat\"\n");
2739 }
2740 
2741 /*
2742  * bic_lookup
2743  * for all the strings in comma separate name_list,
2744  * set the approprate bit in return value.
2745  */
bic_lookup(cpu_set_t * ret_set,char * name_list,enum show_hide_mode mode)2746 void bic_lookup(cpu_set_t *ret_set, char *name_list, enum show_hide_mode mode)
2747 {
2748 	unsigned int i;
2749 
2750 	while (name_list) {
2751 		char *comma;
2752 
2753 		comma = strchr(name_list, ',');
2754 
2755 		if (comma)
2756 			*comma = '\0';
2757 
2758 		for (i = 0; i < MAX_BIC; ++i) {
2759 			if (!strcmp(name_list, bic[i].name)) {
2760 				SET_BIC(i, ret_set);
2761 				break;
2762 			}
2763 			if (!strcmp(name_list, "all")) {
2764 				bic_set_all(ret_set);
2765 				break;
2766 			} else if (!strcmp(name_list, "topology")) {
2767 				CPU_OR(ret_set, ret_set, &bic_group_topology);
2768 				break;
2769 			} else if (!strcmp(name_list, "power")) {
2770 				CPU_OR(ret_set, ret_set, &bic_group_thermal_pwr);
2771 				break;
2772 			} else if (!strcmp(name_list, "idle")) {
2773 				CPU_OR(ret_set, ret_set, &bic_group_idle);
2774 				break;
2775 			} else if (!strcmp(name_list, "cache")) {
2776 				CPU_OR(ret_set, ret_set, &bic_group_cache);
2777 				break;
2778 			} else if (!strcmp(name_list, "llc")) {
2779 				CPU_OR(ret_set, ret_set, &bic_group_cache);
2780 				break;
2781 			} else if (!strcmp(name_list, "swidle")) {
2782 				CPU_OR(ret_set, ret_set, &bic_group_sw_idle);
2783 				break;
2784 			} else if (!strcmp(name_list, "sysfs")) {	/* legacy compatibility */
2785 				CPU_OR(ret_set, ret_set, &bic_group_sw_idle);
2786 				break;
2787 			} else if (!strcmp(name_list, "hwidle")) {
2788 				CPU_OR(ret_set, ret_set, &bic_group_hw_idle);
2789 				break;
2790 			} else if (!strcmp(name_list, "frequency")) {
2791 				CPU_OR(ret_set, ret_set, &bic_group_frequency);
2792 				break;
2793 			} else if (!strcmp(name_list, "other")) {
2794 				CPU_OR(ret_set, ret_set, &bic_group_other);
2795 				break;
2796 			}
2797 		}
2798 		if (i == MAX_BIC) {
2799 			if (mode == SHOW_LIST) {
2800 				deferred_add_names[deferred_add_index++] = name_list;
2801 				if (deferred_add_index >= MAX_DEFERRED) {
2802 					fprintf(stderr, "More than max %d un-recognized --add options '%s'\n", MAX_DEFERRED, name_list);
2803 					help();
2804 					exit(1);
2805 				}
2806 			} else {
2807 				deferred_skip_names[deferred_skip_index++] = name_list;
2808 				if (debug)
2809 					fprintf(stderr, "deferred \"%s\"\n", name_list);
2810 				if (deferred_skip_index >= MAX_DEFERRED) {
2811 					fprintf(stderr, "More than max %d un-recognized --skip options '%s'\n", MAX_DEFERRED, name_list);
2812 					help();
2813 					exit(1);
2814 				}
2815 			}
2816 		}
2817 
2818 		name_list = comma;
2819 		if (name_list)
2820 			name_list++;
2821 
2822 	}
2823 }
2824 
2825 /*
2826  * print_name()
2827  * Print column header name for raw 64-bit counter in 16 columns (at least 8-char plus a tab)
2828  * Otherwise, allow the name + tab to fit within 8-coumn tab-stop.
2829  * In both cases, left justififed, just like other turbostat columns,
2830  * to allow the column values to consume the tab.
2831  *
2832  * Yes, 32-bit counters can overflow 8-columns, and
2833  * 64-bit counters can overflow 16-columns, but that is uncommon.
2834  */
print_name(int width,int * printed,char * delim,char * name,enum counter_type type,enum counter_format format)2835 static inline int print_name(int width, int *printed, char *delim, char *name, enum counter_type type, enum counter_format format)
2836 {
2837 	UNUSED(type);
2838 
2839 	if (format == FORMAT_RAW && width >= 64)
2840 		return (sprintf(outp, "%s%-8s", (*printed++ ? delim : ""), name));
2841 	else
2842 		return (sprintf(outp, "%s%s", (*printed++ ? delim : ""), name));
2843 }
2844 
print_hex_value(int width,int * printed,char * delim,unsigned long long value)2845 static inline int print_hex_value(int width, int *printed, char *delim, unsigned long long value)
2846 {
2847 	if (width <= 32)
2848 		return (sprintf(outp, "%s%08x", (*printed++ ? delim : ""), (unsigned int)value));
2849 	else
2850 		return (sprintf(outp, "%s%016llx", (*printed++ ? delim : ""), value));
2851 }
2852 
print_decimal_value(int width,int * printed,char * delim,unsigned long long value)2853 static inline int print_decimal_value(int width, int *printed, char *delim, unsigned long long value)
2854 {
2855 	if (width <= 32)
2856 		return (sprintf(outp, "%s%d", (*printed++ ? delim : ""), (unsigned int)value));
2857 	else
2858 		return (sprintf(outp, "%s%-8lld", (*printed++ ? delim : ""), value));
2859 }
2860 
print_float_value(int * printed,char * delim,double value)2861 static inline int print_float_value(int *printed, char *delim, double value)
2862 {
2863 	return (sprintf(outp, "%s%0.2f", (*printed++ ? delim : ""), value));
2864 }
2865 
print_header(char * delim)2866 void print_header(char *delim)
2867 {
2868 	struct msr_counter *mp;
2869 	struct perf_counter_info *pp;
2870 	struct pmt_counter *ppmt;
2871 	int printed = 0;
2872 
2873 	if (DO_BIC(BIC_USEC))
2874 		outp += sprintf(outp, "%susec", (printed++ ? delim : ""));
2875 	if (DO_BIC(BIC_TOD))
2876 		outp += sprintf(outp, "%sTime_Of_Day_Seconds", (printed++ ? delim : ""));
2877 	if (DO_BIC(BIC_Package))
2878 		outp += sprintf(outp, "%sPackage", (printed++ ? delim : ""));
2879 	if (DO_BIC(BIC_Die))
2880 		outp += sprintf(outp, "%sDie", (printed++ ? delim : ""));
2881 	if (DO_BIC(BIC_L3))
2882 		outp += sprintf(outp, "%sL3", (printed++ ? delim : ""));
2883 	if (DO_BIC(BIC_Node))
2884 		outp += sprintf(outp, "%sNode", (printed++ ? delim : ""));
2885 	if (DO_BIC(BIC_Core))
2886 		outp += sprintf(outp, "%sCore", (printed++ ? delim : ""));
2887 	if (DO_BIC(BIC_CPU))
2888 		outp += sprintf(outp, "%sCPU", (printed++ ? delim : ""));
2889 	if (DO_BIC(BIC_APIC))
2890 		outp += sprintf(outp, "%sAPIC", (printed++ ? delim : ""));
2891 	if (DO_BIC(BIC_X2APIC))
2892 		outp += sprintf(outp, "%sX2APIC", (printed++ ? delim : ""));
2893 	if (DO_BIC(BIC_Avg_MHz))
2894 		outp += sprintf(outp, "%sAvg_MHz", (printed++ ? delim : ""));
2895 	if (DO_BIC(BIC_Busy))
2896 		outp += sprintf(outp, "%sBusy%%", (printed++ ? delim : ""));
2897 	if (DO_BIC(BIC_Bzy_MHz))
2898 		outp += sprintf(outp, "%sBzy_MHz", (printed++ ? delim : ""));
2899 	if (DO_BIC(BIC_TSC_MHz))
2900 		outp += sprintf(outp, "%sTSC_MHz", (printed++ ? delim : ""));
2901 
2902 	if (DO_BIC(BIC_IPC))
2903 		outp += sprintf(outp, "%sIPC", (printed++ ? delim : ""));
2904 
2905 	if (DO_BIC(BIC_IRQ)) {
2906 		if (sums_need_wide_columns)
2907 			outp += sprintf(outp, "%s     IRQ", (printed++ ? delim : ""));
2908 		else
2909 			outp += sprintf(outp, "%sIRQ", (printed++ ? delim : ""));
2910 	}
2911 	if (DO_BIC(BIC_NMI)) {
2912 		if (sums_need_wide_columns)
2913 			outp += sprintf(outp, "%s     NMI", (printed++ ? delim : ""));
2914 		else
2915 			outp += sprintf(outp, "%sNMI", (printed++ ? delim : ""));
2916 	}
2917 
2918 	if (DO_BIC(BIC_SMI))
2919 		outp += sprintf(outp, "%sSMI", (printed++ ? delim : ""));
2920 
2921 	if (DO_BIC(BIC_LLC_MRPS))
2922 		outp += sprintf(outp, "%sLLCMRPS", (printed++ ? delim : ""));
2923 
2924 	if (DO_BIC(BIC_LLC_HIT))
2925 		outp += sprintf(outp, "%sLLC%%hit", (printed++ ? delim : ""));
2926 
2927 	if (DO_BIC(BIC_L2_MRPS))
2928 		outp += sprintf(outp, "%sL2MRPS", (printed++ ? delim : ""));
2929 
2930 	if (DO_BIC(BIC_L2_HIT))
2931 		outp += sprintf(outp, "%sL2%%hit", (printed++ ? delim : ""));
2932 
2933 	for (mp = sys.tp; mp; mp = mp->next)
2934 		outp += print_name(mp->width, &printed, delim, mp->name, mp->type, mp->format);
2935 
2936 	for (pp = sys.perf_tp; pp; pp = pp->next)
2937 		outp += print_name(pp->width, &printed, delim, pp->name, pp->type, pp->format);
2938 
2939 	ppmt = sys.pmt_tp;
2940 	while (ppmt) {
2941 		switch (ppmt->type) {
2942 		case PMT_TYPE_RAW:
2943 			outp += print_name(pmt_counter_get_width(ppmt), &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
2944 			break;
2945 
2946 		case PMT_TYPE_XTAL_TIME:
2947 		case PMT_TYPE_TCORE_CLOCK:
2948 			outp += print_name(32, &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
2949 			break;
2950 		}
2951 
2952 		ppmt = ppmt->next;
2953 	}
2954 
2955 	if (DO_BIC(BIC_CPU_c1))
2956 		outp += sprintf(outp, "%sCPU%%c1", (printed++ ? delim : ""));
2957 	if (DO_BIC(BIC_CPU_c3))
2958 		outp += sprintf(outp, "%sCPU%%c3", (printed++ ? delim : ""));
2959 	if (DO_BIC(BIC_CPU_c6))
2960 		outp += sprintf(outp, "%sCPU%%c6", (printed++ ? delim : ""));
2961 	if (DO_BIC(BIC_CPU_c7))
2962 		outp += sprintf(outp, "%sCPU%%c7", (printed++ ? delim : ""));
2963 
2964 	if (DO_BIC(BIC_Mod_c6))
2965 		outp += sprintf(outp, "%sMod%%c6", (printed++ ? delim : ""));
2966 
2967 	if (DO_BIC(BIC_CoreTmp))
2968 		outp += sprintf(outp, "%sCoreTmp", (printed++ ? delim : ""));
2969 
2970 	if (DO_BIC(BIC_CORE_THROT_CNT))
2971 		outp += sprintf(outp, "%sCoreThr", (printed++ ? delim : ""));
2972 
2973 	if (valid_rapl_msrs && !rapl_joules) {
2974 		if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
2975 			outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
2976 	} else if (valid_rapl_msrs && rapl_joules) {
2977 		if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
2978 			outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
2979 	}
2980 
2981 	for (mp = sys.cp; mp; mp = mp->next)
2982 		outp += print_name(mp->width, &printed, delim, mp->name, mp->type, mp->format);
2983 
2984 	for (pp = sys.perf_cp; pp; pp = pp->next)
2985 		outp += print_name(pp->width, &printed, delim, pp->name, pp->type, pp->format);
2986 
2987 	ppmt = sys.pmt_cp;
2988 	while (ppmt) {
2989 		switch (ppmt->type) {
2990 		case PMT_TYPE_RAW:
2991 			outp += print_name(pmt_counter_get_width(ppmt), &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
2992 
2993 			break;
2994 
2995 		case PMT_TYPE_XTAL_TIME:
2996 		case PMT_TYPE_TCORE_CLOCK:
2997 			outp += print_name(32, &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
2998 			break;
2999 		}
3000 
3001 		ppmt = ppmt->next;
3002 	}
3003 	if (DO_BIC(BIC_PkgTmp))
3004 		outp += sprintf(outp, "%sPkgTmp", (printed++ ? delim : ""));
3005 
3006 	if (DO_BIC(BIC_GFX_rc6))
3007 		outp += sprintf(outp, "%sGFX%%rc6", (printed++ ? delim : ""));
3008 
3009 	if (DO_BIC(BIC_GFXMHz))
3010 		outp += sprintf(outp, "%sGFXMHz", (printed++ ? delim : ""));
3011 
3012 	if (DO_BIC(BIC_GFXACTMHz))
3013 		outp += sprintf(outp, "%sGFXAMHz", (printed++ ? delim : ""));
3014 
3015 	if (DO_BIC(BIC_SAM_mc6))
3016 		outp += sprintf(outp, "%sSAM%%mc6", (printed++ ? delim : ""));
3017 
3018 	if (DO_BIC(BIC_SAMMHz))
3019 		outp += sprintf(outp, "%sSAMMHz", (printed++ ? delim : ""));
3020 
3021 	if (DO_BIC(BIC_SAMACTMHz))
3022 		outp += sprintf(outp, "%sSAMAMHz", (printed++ ? delim : ""));
3023 
3024 	if (DO_BIC(BIC_Totl_c0))
3025 		outp += sprintf(outp, "%sTotl%%C0", (printed++ ? delim : ""));
3026 	if (DO_BIC(BIC_Any_c0))
3027 		outp += sprintf(outp, "%sAny%%C0", (printed++ ? delim : ""));
3028 	if (DO_BIC(BIC_GFX_c0))
3029 		outp += sprintf(outp, "%sGFX%%C0", (printed++ ? delim : ""));
3030 	if (DO_BIC(BIC_CPUGFX))
3031 		outp += sprintf(outp, "%sCPUGFX%%", (printed++ ? delim : ""));
3032 
3033 	if (DO_BIC(BIC_Pkgpc2))
3034 		outp += sprintf(outp, "%sPkg%%pc2", (printed++ ? delim : ""));
3035 	if (DO_BIC(BIC_Pkgpc3))
3036 		outp += sprintf(outp, "%sPkg%%pc3", (printed++ ? delim : ""));
3037 	if (DO_BIC(BIC_Pkgpc6))
3038 		outp += sprintf(outp, "%sPkg%%pc6", (printed++ ? delim : ""));
3039 	if (DO_BIC(BIC_Pkgpc7))
3040 		outp += sprintf(outp, "%sPkg%%pc7", (printed++ ? delim : ""));
3041 	if (DO_BIC(BIC_Pkgpc8))
3042 		outp += sprintf(outp, "%sPkg%%pc8", (printed++ ? delim : ""));
3043 	if (DO_BIC(BIC_Pkgpc9))
3044 		outp += sprintf(outp, "%sPkg%%pc9", (printed++ ? delim : ""));
3045 	if (DO_BIC(BIC_Pkgpc10))
3046 		outp += sprintf(outp, "%sPk%%pc10", (printed++ ? delim : ""));
3047 	if (DO_BIC(BIC_Diec6))
3048 		outp += sprintf(outp, "%sDie%%c6", (printed++ ? delim : ""));
3049 	if (DO_BIC(BIC_CPU_LPI))
3050 		outp += sprintf(outp, "%sCPU%%LPI", (printed++ ? delim : ""));
3051 	if (DO_BIC(BIC_SYS_LPI))
3052 		outp += sprintf(outp, "%sSYS%%LPI", (printed++ ? delim : ""));
3053 
3054 	if (!rapl_joules) {
3055 		if (DO_BIC(BIC_PkgWatt))
3056 			outp += sprintf(outp, "%sPkgWatt", (printed++ ? delim : ""));
3057 		if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
3058 			outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
3059 		if (DO_BIC(BIC_GFXWatt))
3060 			outp += sprintf(outp, "%sGFXWatt", (printed++ ? delim : ""));
3061 		if (DO_BIC(BIC_RAMWatt))
3062 			outp += sprintf(outp, "%sRAMWatt", (printed++ ? delim : ""));
3063 		if (DO_BIC(BIC_PKG__))
3064 			outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
3065 		if (DO_BIC(BIC_RAM__))
3066 			outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
3067 	} else {
3068 		if (DO_BIC(BIC_Pkg_J))
3069 			outp += sprintf(outp, "%sPkg_J", (printed++ ? delim : ""));
3070 		if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
3071 			outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
3072 		if (DO_BIC(BIC_GFX_J))
3073 			outp += sprintf(outp, "%sGFX_J", (printed++ ? delim : ""));
3074 		if (DO_BIC(BIC_RAM_J))
3075 			outp += sprintf(outp, "%sRAM_J", (printed++ ? delim : ""));
3076 		if (DO_BIC(BIC_PKG__))
3077 			outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
3078 		if (DO_BIC(BIC_RAM__))
3079 			outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
3080 	}
3081 	if (DO_BIC(BIC_UNCORE_MHZ))
3082 		outp += sprintf(outp, "%sUncMHz", (printed++ ? delim : ""));
3083 
3084 	for (mp = sys.pp; mp; mp = mp->next)
3085 		outp += print_name(mp->width, &printed, delim, mp->name, mp->type, mp->format);
3086 
3087 	for (pp = sys.perf_pp; pp; pp = pp->next)
3088 		outp += print_name(pp->width, &printed, delim, pp->name, pp->type, pp->format);
3089 
3090 	ppmt = sys.pmt_pp;
3091 	while (ppmt) {
3092 		switch (ppmt->type) {
3093 		case PMT_TYPE_RAW:
3094 			outp += print_name(pmt_counter_get_width(ppmt), &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
3095 			break;
3096 
3097 		case PMT_TYPE_XTAL_TIME:
3098 		case PMT_TYPE_TCORE_CLOCK:
3099 			outp += print_name(32, &printed, delim, ppmt->name, COUNTER_ITEMS, ppmt->format);
3100 			break;
3101 		}
3102 
3103 		ppmt = ppmt->next;
3104 	}
3105 
3106 	if (DO_BIC(BIC_SysWatt))
3107 		outp += sprintf(outp, "%sSysWatt", (printed++ ? delim : ""));
3108 	if (DO_BIC(BIC_Sys_J))
3109 		outp += sprintf(outp, "%sSys_J", (printed++ ? delim : ""));
3110 
3111 	outp += sprintf(outp, "\n");
3112 }
3113 
3114 /*
3115  * pct(numerator, denominator)
3116  *
3117  * Return sanity checked percentage (100.0 * numerator/denominotor)
3118  *
3119  * n < 0: nan
3120  * d <= 0: nan
3121  * n/d > 1.1: nan
3122  */
pct(double numerator,double denominator)3123 double pct(double numerator, double denominator)
3124 {
3125 	double retval;
3126 
3127 	if (numerator < 0)
3128 		return nan("");
3129 
3130 	if (denominator <= 0)
3131 		return nan("");
3132 
3133 	retval = 100.0 * numerator / denominator;
3134 
3135 	if (retval > 110.0)
3136 		return nan("");
3137 
3138 	return retval;
3139 }
3140 
dump_counters(PER_THREAD_PARAMS)3141 int dump_counters(PER_THREAD_PARAMS)
3142 {
3143 	int i;
3144 	struct msr_counter *mp;
3145 	struct platform_counters *pplat_cnt = p == odd.packages ? &platform_counters_odd : &platform_counters_even;
3146 
3147 	outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
3148 
3149 	if (t) {
3150 		outp += sprintf(outp, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags);
3151 		outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
3152 		outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
3153 		outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
3154 		outp += sprintf(outp, "c1: %016llX\n", t->c1);
3155 
3156 		if (DO_BIC(BIC_IPC))
3157 			outp += sprintf(outp, "IPC: %lld\n", t->instr_count);
3158 
3159 		if (DO_BIC(BIC_IRQ))
3160 			outp += sprintf(outp, "IRQ: %lld\n", t->irq_count);
3161 		if (DO_BIC(BIC_NMI))
3162 			outp += sprintf(outp, "IRQ: %lld\n", t->nmi_count);
3163 		if (DO_BIC(BIC_SMI))
3164 			outp += sprintf(outp, "SMI: %d\n", t->smi_count);
3165 
3166 		outp += sprintf(outp, "LLC refs: %lld", t->llc.references);
3167 		outp += sprintf(outp, "LLC miss: %lld", t->llc.misses);
3168 		outp += sprintf(outp, "LLC Hit%%: %.2f", pct((t->llc.references - t->llc.misses), t->llc.references));
3169 
3170 		outp += sprintf(outp, "L2 refs: %lld", t->l2.references);
3171 		outp += sprintf(outp, "L2 hits: %lld", t->l2.hits);
3172 		outp += sprintf(outp, "L2 Hit%%: %.2f", pct(t->l2.hits, t->l2.references));
3173 
3174 		for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3175 			outp += sprintf(outp, "tADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num, t->counter[i], mp->sp->path);
3176 		}
3177 	}
3178 
3179 	if (c && is_cpu_first_thread_in_core(t, c)) {
3180 		outp += sprintf(outp, "core: %d\n", cpus[t->cpu_id].core_id);
3181 		outp += sprintf(outp, "c3: %016llX\n", c->c3);
3182 		outp += sprintf(outp, "c6: %016llX\n", c->c6);
3183 		outp += sprintf(outp, "c7: %016llX\n", c->c7);
3184 		outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
3185 		outp += sprintf(outp, "cpu_throt_count: %016llX\n", c->core_throt_cnt);
3186 
3187 		const unsigned long long energy_value = c->core_energy.raw_value * c->core_energy.scale;
3188 		const double energy_scale = c->core_energy.scale;
3189 
3190 		if (c->core_energy.unit == RAPL_UNIT_JOULES)
3191 			outp += sprintf(outp, "Joules: %0llX (scale: %lf)\n", energy_value, energy_scale);
3192 
3193 		for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3194 			outp += sprintf(outp, "cADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num, c->counter[i], mp->sp->path);
3195 		}
3196 		outp += sprintf(outp, "mc6_us: %016llX\n", c->mc6_us);
3197 	}
3198 
3199 	if (p && is_cpu_first_core_in_package(t, p)) {
3200 		outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
3201 		outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
3202 		outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
3203 		outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
3204 
3205 		outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
3206 		if (DO_BIC(BIC_Pkgpc3))
3207 			outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
3208 		if (DO_BIC(BIC_Pkgpc6))
3209 			outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
3210 		if (DO_BIC(BIC_Pkgpc7))
3211 			outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
3212 		outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
3213 		outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
3214 		outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
3215 		outp += sprintf(outp, "cpu_lpi: %016llX\n", p->cpu_lpi);
3216 		outp += sprintf(outp, "sys_lpi: %016llX\n", p->sys_lpi);
3217 		outp += sprintf(outp, "Joules PKG: %0llX\n", p->energy_pkg.raw_value);
3218 		outp += sprintf(outp, "Joules COR: %0llX\n", p->energy_cores.raw_value);
3219 		outp += sprintf(outp, "Joules GFX: %0llX\n", p->energy_gfx.raw_value);
3220 		outp += sprintf(outp, "Joules RAM: %0llX\n", p->energy_dram.raw_value);
3221 		outp += sprintf(outp, "Joules PSYS: %0llX\n", pplat_cnt->energy_psys.raw_value);
3222 		outp += sprintf(outp, "Throttle PKG: %0llX\n", p->rapl_pkg_perf_status.raw_value);
3223 		outp += sprintf(outp, "Throttle RAM: %0llX\n", p->rapl_dram_perf_status.raw_value);
3224 		outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
3225 
3226 		for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3227 			outp += sprintf(outp, "pADDED [%d] %8s msr0x%x: %08llX %s\n", i, mp->name, mp->msr_num, p->counter[i], mp->sp->path);
3228 		}
3229 	}
3230 
3231 	outp += sprintf(outp, "\n");
3232 
3233 	return 0;
3234 }
3235 
rapl_counter_get_value(const struct rapl_counter * c,enum rapl_unit desired_unit,double interval)3236 double rapl_counter_get_value(const struct rapl_counter *c, enum rapl_unit desired_unit, double interval)
3237 {
3238 	assert(desired_unit != RAPL_UNIT_INVALID);
3239 
3240 	/*
3241 	 * For now we don't expect anything other than joules,
3242 	 * so just simplify the logic.
3243 	 */
3244 	assert(c->unit == RAPL_UNIT_JOULES);
3245 
3246 	const double scaled = c->raw_value * c->scale;
3247 
3248 	if (desired_unit == RAPL_UNIT_WATTS)
3249 		return scaled / interval;
3250 	return scaled;
3251 }
3252 
get_perf_llc_stats(int cpu,struct llc_stats * llc)3253 void get_perf_llc_stats(int cpu, struct llc_stats *llc)
3254 {
3255 	struct read_format {
3256 		unsigned long long num_read;
3257 		struct llc_stats llc;
3258 	} r;
3259 	const ssize_t expected_read_size = sizeof(r);
3260 	ssize_t actual_read_size;
3261 
3262 	actual_read_size = read(fd_llc_percpu[cpu], &r, expected_read_size);
3263 
3264 	if (actual_read_size == -1)
3265 		err(-1, "%s(cpu%d,) %d,,%ld", __func__, cpu, fd_llc_percpu[cpu], expected_read_size);
3266 
3267 	llc->references = r.llc.references;
3268 	llc->misses = r.llc.misses;
3269 	if (actual_read_size != expected_read_size)
3270 		warn("%s: failed to read perf_data (req %zu act %zu)", __func__, expected_read_size, actual_read_size);
3271 }
3272 
get_perf_l2_stats(int cpu,struct l2_stats * l2)3273 void get_perf_l2_stats(int cpu, struct l2_stats *l2)
3274 {
3275 	struct read_format {
3276 		unsigned long long num_read;
3277 		struct l2_stats l2;
3278 	} r;
3279 	const ssize_t expected_read_size = sizeof(r);
3280 	ssize_t actual_read_size;
3281 
3282 	actual_read_size = read(fd_l2_percpu[cpu], &r, expected_read_size);
3283 
3284 	if (actual_read_size == -1)
3285 		err(-1, "%s(cpu%d,) %d,,%ld", __func__, cpu, fd_l2_percpu[cpu], expected_read_size);
3286 
3287 	l2->references = r.l2.references;
3288 	l2->hits = r.l2.hits;
3289 	if (actual_read_size != expected_read_size)
3290 		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);
3291 }
3292 
3293 /*
3294  * column formatting convention & formats
3295  */
format_counters(PER_THREAD_PARAMS)3296 int format_counters(PER_THREAD_PARAMS)
3297 {
3298 	static int count;
3299 
3300 	struct platform_counters *pplat_cnt = NULL;
3301 	double interval_float, tsc;
3302 	char *fmt8 = "%s%.2f";
3303 
3304 	int i;
3305 	struct msr_counter *mp;
3306 	struct perf_counter_info *pp;
3307 	struct pmt_counter *ppmt;
3308 	char *delim = "\t";
3309 	int printed = 0;
3310 
3311 	if (t == average.threads) {
3312 		pplat_cnt = count & 1 ? &platform_counters_odd : &platform_counters_even;
3313 		++count;
3314 	}
3315 
3316 	/* if showing only 1st thread in core and this isn't one, bail out */
3317 	if (show_core_only && !is_cpu_first_thread_in_core(t, c))
3318 		return 0;
3319 
3320 	/* if showing only 1st thread in pkg and this isn't one, bail out */
3321 	if (show_pkg_only && !is_cpu_first_core_in_package(t, p))
3322 		return 0;
3323 
3324 	/*if not summary line and --cpu is used */
3325 	if ((t != average.threads) && (cpu_subset && !CPU_ISSET_S(t->cpu_id, cpu_subset_size, cpu_subset)))
3326 		return 0;
3327 
3328 	if (DO_BIC(BIC_USEC)) {
3329 		/* on each row, print how many usec each timestamp took to gather */
3330 		struct timeval tv;
3331 
3332 		timersub(&t->tv_end, &t->tv_begin, &tv);
3333 		outp += sprintf(outp, "%5ld\t", tv.tv_sec * 1000000 + tv.tv_usec);
3334 	}
3335 
3336 	/* Time_Of_Day_Seconds: on each row, print sec.usec last timestamp taken */
3337 	if (DO_BIC(BIC_TOD))
3338 		outp += sprintf(outp, "%10ld.%06ld\t", t->tv_end.tv_sec, t->tv_end.tv_usec);
3339 
3340 	interval_float = t->tv_delta.tv_sec + t->tv_delta.tv_usec / 1000000.0;
3341 
3342 	tsc = t->tsc * tsc_tweak;
3343 
3344 	/* topo columns, print blanks on 1st (average) line */
3345 	if (t == average.threads) {
3346 		if (DO_BIC(BIC_Package))
3347 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3348 		if (DO_BIC(BIC_Die))
3349 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3350 		if (DO_BIC(BIC_L3))
3351 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3352 		if (DO_BIC(BIC_Node))
3353 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3354 		if (DO_BIC(BIC_Core))
3355 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3356 		if (DO_BIC(BIC_CPU))
3357 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3358 		if (DO_BIC(BIC_APIC))
3359 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3360 		if (DO_BIC(BIC_X2APIC))
3361 			outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3362 	} else {
3363 		if (DO_BIC(BIC_Package)) {
3364 			if (p)
3365 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].package_id);
3366 			else
3367 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3368 		}
3369 		if (DO_BIC(BIC_Die)) {
3370 			if (c)
3371 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].die_id);
3372 			else
3373 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3374 		}
3375 		if (DO_BIC(BIC_L3)) {
3376 			if (c)
3377 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].l3_id);
3378 			else
3379 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3380 		}
3381 		if (DO_BIC(BIC_Node)) {
3382 			if (t)
3383 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].physical_node_id);
3384 			else
3385 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3386 		}
3387 		if (DO_BIC(BIC_Core)) {
3388 			if (c)
3389 				outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].core_id);
3390 			else
3391 				outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
3392 		}
3393 		if (DO_BIC(BIC_CPU))
3394 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->cpu_id);
3395 		if (DO_BIC(BIC_APIC))
3396 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->apic_id);
3397 		if (DO_BIC(BIC_X2APIC))
3398 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->x2apic_id);
3399 	}
3400 
3401 	if (DO_BIC(BIC_Avg_MHz))
3402 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 / units * t->aperf / interval_float);
3403 
3404 	if (DO_BIC(BIC_Busy))
3405 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(t->mperf, tsc));
3406 
3407 	if (DO_BIC(BIC_Bzy_MHz)) {
3408 		if (has_base_hz)
3409 			outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), base_hz / units * t->aperf / t->mperf);
3410 		else
3411 			outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), tsc / units * t->aperf / t->mperf / interval_float);
3412 	}
3413 
3414 	if (DO_BIC(BIC_TSC_MHz))
3415 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 * t->tsc / units / interval_float);
3416 
3417 	if (DO_BIC(BIC_IPC))
3418 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 1.0 * t->instr_count / t->aperf);
3419 
3420 	/* IRQ */
3421 	if (DO_BIC(BIC_IRQ)) {
3422 		if (sums_need_wide_columns)
3423 			outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->irq_count);
3424 		else
3425 			outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->irq_count);
3426 	}
3427 
3428 	/* NMI */
3429 	if (DO_BIC(BIC_NMI)) {
3430 		if (sums_need_wide_columns)
3431 			outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->nmi_count);
3432 		else
3433 			outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->nmi_count);
3434 	}
3435 
3436 	/* SMI */
3437 	if (DO_BIC(BIC_SMI))
3438 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->smi_count);
3439 
3440 	/* LLC Stats */
3441 	if (DO_BIC(BIC_LLC_MRPS))
3442 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), t->llc.references / interval_float / 1000000);
3443 
3444 	if (DO_BIC(BIC_LLC_HIT))
3445 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), pct((t->llc.references - t->llc.misses), t->llc.references));
3446 
3447 	/* L2 Stats */
3448 	if (DO_BIC(BIC_L2_MRPS))
3449 		outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), t->l2.references / interval_float / 1000000);
3450 
3451 	if (DO_BIC(BIC_L2_HIT))
3452 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), pct(t->l2.hits, t->l2.references));
3453 
3454 	/* Added Thread Counters */
3455 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
3456 		if (mp->format == FORMAT_RAW)
3457 			outp += print_hex_value(mp->width, &printed, delim, t->counter[i]);
3458 		else if (mp->format == FORMAT_DELTA || mp->format == FORMAT_AVERAGE)
3459 			outp += print_decimal_value(mp->width, &printed, delim, t->counter[i]);
3460 		else if (mp->format == FORMAT_PERCENT) {
3461 			if (mp->type == COUNTER_USEC)
3462 				outp += print_float_value(&printed, delim, t->counter[i] / interval_float / 10000);
3463 			else
3464 				outp += print_float_value(&printed, delim, pct(t->counter[i], tsc));
3465 		}
3466 	}
3467 
3468 	/* Added perf Thread Counters */
3469 	for (i = 0, pp = sys.perf_tp; pp; ++i, pp = pp->next) {
3470 		if (pp->format == FORMAT_RAW)
3471 			outp += print_hex_value(pp->width, &printed, delim, t->perf_counter[i]);
3472 		else if (pp->format == FORMAT_DELTA || mp->format == FORMAT_AVERAGE)
3473 			outp += print_decimal_value(pp->width, &printed, delim, t->perf_counter[i]);
3474 		else if (pp->format == FORMAT_PERCENT) {
3475 			if (pp->type == COUNTER_USEC)
3476 				outp += print_float_value(&printed, delim, t->perf_counter[i] / interval_float / 10000);
3477 			else
3478 				outp += print_float_value(&printed, delim, pct(t->perf_counter[i], tsc));
3479 		}
3480 	}
3481 
3482 	/* Added PMT Thread Counters */
3483 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
3484 		const unsigned long value_raw = t->pmt_counter[i];
3485 		double value_converted;
3486 		switch (ppmt->type) {
3487 		case PMT_TYPE_RAW:
3488 			outp += print_hex_value(pmt_counter_get_width(ppmt), &printed, delim, t->pmt_counter[i]);
3489 			break;
3490 
3491 		case PMT_TYPE_XTAL_TIME:
3492 			value_converted = pct(value_raw / crystal_hz, interval_float);
3493 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
3494 			break;
3495 
3496 		case PMT_TYPE_TCORE_CLOCK:
3497 			value_converted = pct(value_raw / tcore_clock_freq_hz, interval_float);
3498 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), value_converted);
3499 		}
3500 	}
3501 
3502 	/* C1 */
3503 	if (DO_BIC(BIC_CPU_c1))
3504 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(t->c1, tsc));
3505 
3506 	/* print per-core data only for 1st thread in core */
3507 	if (!is_cpu_first_thread_in_core(t, c))
3508 		goto done;
3509 
3510 	if (DO_BIC(BIC_CPU_c3))
3511 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(c->c3, tsc));
3512 	if (DO_BIC(BIC_CPU_c6))
3513 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(c->c6, tsc));
3514 	if (DO_BIC(BIC_CPU_c7))
3515 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(c->c7, tsc));
3516 
3517 	/* Mod%c6 */
3518 	if (DO_BIC(BIC_Mod_c6))
3519 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(c->mc6_us, tsc));
3520 
3521 	if (DO_BIC(BIC_CoreTmp))
3522 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_temp_c);
3523 
3524 	/* Core throttle count */
3525 	if (DO_BIC(BIC_CORE_THROT_CNT))
3526 		outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->core_throt_cnt);
3527 
3528 	/* Added Core Counters */
3529 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3530 		if (mp->format == FORMAT_RAW)
3531 			outp += print_hex_value(mp->width, &printed, delim, c->counter[i]);
3532 		else if (mp->format == FORMAT_DELTA || mp->format == FORMAT_AVERAGE)
3533 			outp += print_decimal_value(mp->width, &printed, delim, c->counter[i]);
3534 		else if (mp->format == FORMAT_PERCENT)
3535 			outp += print_float_value(&printed, delim, pct(c->counter[i], tsc));
3536 	}
3537 
3538 	/* Added perf Core counters */
3539 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3540 		if (pp->format == FORMAT_RAW)
3541 			outp += print_hex_value(pp->width, &printed, delim, c->perf_counter[i]);
3542 		else if (pp->format == FORMAT_DELTA || mp->format == FORMAT_AVERAGE)
3543 			outp += print_decimal_value(pp->width, &printed, delim, c->perf_counter[i]);
3544 		else if (pp->format == FORMAT_PERCENT)
3545 			outp += print_float_value(&printed, delim, pct(c->perf_counter[i], tsc));
3546 	}
3547 
3548 	/* Added PMT Core counters */
3549 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3550 		const unsigned long value_raw = c->pmt_counter[i];
3551 		double value_converted;
3552 		switch (ppmt->type) {
3553 		case PMT_TYPE_RAW:
3554 			outp += print_hex_value(pmt_counter_get_width(ppmt), &printed, delim, c->pmt_counter[i]);
3555 			break;
3556 
3557 		case PMT_TYPE_XTAL_TIME:
3558 			value_converted = pct(value_raw / crystal_hz, interval_float);
3559 			outp += print_float_value(&printed, delim, value_converted);
3560 			break;
3561 
3562 		case PMT_TYPE_TCORE_CLOCK:
3563 			value_converted = pct(value_raw / tcore_clock_freq_hz, interval_float);
3564 			outp += print_float_value(&printed, delim, value_converted);
3565 		}
3566 	}
3567 
3568 	if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
3569 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&c->core_energy, RAPL_UNIT_WATTS, interval_float));
3570 	if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
3571 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&c->core_energy, RAPL_UNIT_JOULES, interval_float));
3572 
3573 	/* print per-package data only for 1st core in package */
3574 	if (!is_cpu_first_core_in_package(t, p))
3575 		goto done;
3576 
3577 	/* PkgTmp */
3578 	if (DO_BIC(BIC_PkgTmp))
3579 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->pkg_temp_c);
3580 
3581 	/* GFXrc6 */
3582 	if (DO_BIC(BIC_GFX_rc6)) {
3583 		if (p->gfx_rc6_ms == -1) {	/* detect GFX counter reset */
3584 			outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
3585 		} else {
3586 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), p->gfx_rc6_ms / 10.0 / interval_float);
3587 		}
3588 	}
3589 
3590 	/* GFXMHz */
3591 	if (DO_BIC(BIC_GFXMHz))
3592 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_mhz);
3593 
3594 	/* GFXACTMHz */
3595 	if (DO_BIC(BIC_GFXACTMHz))
3596 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_act_mhz);
3597 
3598 	/* SAMmc6 */
3599 	if (DO_BIC(BIC_SAM_mc6)) {
3600 		if (p->sam_mc6_ms == -1) {	/* detect GFX counter reset */
3601 			outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
3602 		} else {
3603 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), p->sam_mc6_ms / 10.0 / interval_float);
3604 		}
3605 	}
3606 
3607 	/* SAMMHz */
3608 	if (DO_BIC(BIC_SAMMHz))
3609 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->sam_mhz);
3610 
3611 	/* SAMACTMHz */
3612 	if (DO_BIC(BIC_SAMACTMHz))
3613 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->sam_act_mhz);
3614 
3615 	/* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
3616 	if (DO_BIC(BIC_Totl_c0))
3617 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100 * p->pkg_wtd_core_c0 / tsc);	/* can exceed 100% */
3618 	if (DO_BIC(BIC_Any_c0))
3619 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pkg_any_core_c0, tsc));
3620 	if (DO_BIC(BIC_GFX_c0))
3621 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pkg_any_gfxe_c0, tsc));
3622 	if (DO_BIC(BIC_CPUGFX))
3623 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pkg_both_core_gfxe_c0, tsc));
3624 
3625 	if (DO_BIC(BIC_Pkgpc2))
3626 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc2, tsc));
3627 	if (DO_BIC(BIC_Pkgpc3))
3628 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc3, tsc));
3629 	if (DO_BIC(BIC_Pkgpc6))
3630 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc6, tsc));
3631 	if (DO_BIC(BIC_Pkgpc7))
3632 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc7, tsc));
3633 	if (DO_BIC(BIC_Pkgpc8))
3634 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc8, tsc));
3635 	if (DO_BIC(BIC_Pkgpc9))
3636 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc9, tsc));
3637 	if (DO_BIC(BIC_Pkgpc10))
3638 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->pc10, tsc));
3639 
3640 	if (DO_BIC(BIC_Diec6))
3641 		outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->die_c6 / crystal_hz, interval_float));
3642 
3643 	if (DO_BIC(BIC_CPU_LPI)) {
3644 		if (p->cpu_lpi >= 0)
3645 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->cpu_lpi / 1000000.0, interval_float));
3646 		else
3647 			outp += sprintf(outp, "%s(neg)", (printed++ ? delim : ""));
3648 	}
3649 	if (DO_BIC(BIC_SYS_LPI)) {
3650 		if (p->sys_lpi >= 0)
3651 			outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), pct(p->sys_lpi / 1000000.0, interval_float));
3652 		else
3653 			outp += sprintf(outp, "%s(neg)", (printed++ ? delim : ""));
3654 	}
3655 
3656 	if (DO_BIC(BIC_PkgWatt))
3657 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_WATTS, interval_float));
3658 	if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
3659 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_WATTS, interval_float));
3660 	if (DO_BIC(BIC_GFXWatt))
3661 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_WATTS, interval_float));
3662 	if (DO_BIC(BIC_RAMWatt))
3663 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_WATTS, interval_float));
3664 	if (DO_BIC(BIC_Pkg_J))
3665 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_JOULES, interval_float));
3666 	if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
3667 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_JOULES, interval_float));
3668 	if (DO_BIC(BIC_GFX_J))
3669 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_JOULES, interval_float));
3670 	if (DO_BIC(BIC_RAM_J))
3671 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_JOULES, interval_float));
3672 	if (DO_BIC(BIC_PKG__))
3673 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->rapl_pkg_perf_status, RAPL_UNIT_WATTS, interval_float));
3674 	if (DO_BIC(BIC_RAM__))
3675 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&p->rapl_dram_perf_status, RAPL_UNIT_WATTS, interval_float));
3676 	/* UncMHz */
3677 	if (DO_BIC(BIC_UNCORE_MHZ))
3678 		outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->uncore_mhz);
3679 
3680 	/* Added Package Counters */
3681 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3682 		if (mp->format == FORMAT_RAW)
3683 			outp += print_hex_value(mp->width, &printed, delim, p->counter[i]);
3684 		else if (mp->type == COUNTER_K2M)
3685 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), (unsigned int)p->counter[i] / 1000);
3686 		else if (mp->format == FORMAT_DELTA || mp->format == FORMAT_AVERAGE)
3687 			outp += print_decimal_value(mp->width, &printed, delim, p->counter[i]);
3688 		else if (mp->format == FORMAT_PERCENT)
3689 			outp += print_float_value(&printed, delim, pct(p->counter[i], tsc));
3690 	}
3691 
3692 	/* Added perf Package Counters */
3693 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3694 		if (pp->format == FORMAT_RAW)
3695 			outp += print_hex_value(pp->width, &printed, delim, p->perf_counter[i]);
3696 		else if (pp->type == COUNTER_K2M)
3697 			outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), (unsigned int)p->perf_counter[i] / 1000);
3698 		else if (pp->format == FORMAT_DELTA || mp->format == FORMAT_AVERAGE)
3699 			outp += print_decimal_value(pp->width, &printed, delim, p->perf_counter[i]);
3700 		else if (pp->format == FORMAT_PERCENT)
3701 			outp += print_float_value(&printed, delim, pct(p->perf_counter[i], tsc));
3702 	}
3703 
3704 	/* Added PMT Package Counters */
3705 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3706 		const unsigned long value_raw = p->pmt_counter[i];
3707 		double value_converted;
3708 		switch (ppmt->type) {
3709 		case PMT_TYPE_RAW:
3710 			outp += print_hex_value(pmt_counter_get_width(ppmt), &printed, delim, p->pmt_counter[i]);
3711 			break;
3712 
3713 		case PMT_TYPE_XTAL_TIME:
3714 			value_converted = pct(value_raw / crystal_hz, interval_float);
3715 			outp += print_float_value(&printed, delim, value_converted);
3716 			break;
3717 
3718 		case PMT_TYPE_TCORE_CLOCK:
3719 			value_converted = pct(value_raw / tcore_clock_freq_hz, interval_float);
3720 			outp += print_float_value(&printed, delim, value_converted);
3721 		}
3722 	}
3723 
3724 	if (DO_BIC(BIC_SysWatt) && (t == average.threads))
3725 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&pplat_cnt->energy_psys, RAPL_UNIT_WATTS, interval_float));
3726 	if (DO_BIC(BIC_Sys_J) && (t == average.threads))
3727 		outp += sprintf(outp, fmt8, (printed++ ? delim : ""), rapl_counter_get_value(&pplat_cnt->energy_psys, RAPL_UNIT_JOULES, interval_float));
3728 
3729 done:
3730 	if (*(outp - 1) != '\n')
3731 		outp += sprintf(outp, "\n");
3732 
3733 	return 0;
3734 }
3735 
flush_output_stdout(void)3736 void flush_output_stdout(void)
3737 {
3738 	FILE *filep;
3739 
3740 	if (outf == stderr)
3741 		filep = stdout;
3742 	else
3743 		filep = outf;
3744 
3745 	fputs(output_buffer, filep);
3746 	fflush(filep);
3747 
3748 	outp = output_buffer;
3749 }
3750 
flush_output_stderr(void)3751 void flush_output_stderr(void)
3752 {
3753 	fputs(output_buffer, outf);
3754 	fflush(outf);
3755 	outp = output_buffer;
3756 }
3757 
format_all_counters(PER_THREAD_PARAMS)3758 void format_all_counters(PER_THREAD_PARAMS)
3759 {
3760 	static int count;
3761 
3762 	if ((!count || (header_iterations && !(count % header_iterations))) || !summary_only)
3763 		print_header("\t");
3764 
3765 	format_counters(average.threads, average.cores, average.packages);
3766 
3767 	count++;
3768 
3769 	if (summary_only)
3770 		return;
3771 
3772 	for_all_cpus(format_counters, t, c, p);
3773 }
3774 
3775 #define DELTA_WRAP32(new, old)			\
3776 	old = ((((unsigned long long)new << 32) - ((unsigned long long)old << 32)) >> 32);
3777 
delta_package(struct pkg_data * new,struct pkg_data * old)3778 int delta_package(struct pkg_data *new, struct pkg_data *old)
3779 {
3780 	int i;
3781 	struct msr_counter *mp;
3782 	struct perf_counter_info *pp;
3783 	struct pmt_counter *ppmt;
3784 
3785 	if (DO_BIC(BIC_Totl_c0))
3786 		old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
3787 	if (DO_BIC(BIC_Any_c0))
3788 		old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
3789 	if (DO_BIC(BIC_GFX_c0))
3790 		old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
3791 	if (DO_BIC(BIC_CPUGFX))
3792 		old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
3793 
3794 	old->pc2 = new->pc2 - old->pc2;
3795 	if (DO_BIC(BIC_Pkgpc3))
3796 		old->pc3 = new->pc3 - old->pc3;
3797 	if (DO_BIC(BIC_Pkgpc6))
3798 		old->pc6 = new->pc6 - old->pc6;
3799 	if (DO_BIC(BIC_Pkgpc7))
3800 		old->pc7 = new->pc7 - old->pc7;
3801 	old->pc8 = new->pc8 - old->pc8;
3802 	old->pc9 = new->pc9 - old->pc9;
3803 	old->pc10 = new->pc10 - old->pc10;
3804 	old->die_c6 = new->die_c6 - old->die_c6;
3805 	old->cpu_lpi = new->cpu_lpi - old->cpu_lpi;
3806 	old->sys_lpi = new->sys_lpi - old->sys_lpi;
3807 	old->pkg_temp_c = new->pkg_temp_c;
3808 
3809 	/* flag an error when rc6 counter resets/wraps */
3810 	if (old->gfx_rc6_ms > new->gfx_rc6_ms)
3811 		old->gfx_rc6_ms = -1;
3812 	else
3813 		old->gfx_rc6_ms = new->gfx_rc6_ms - old->gfx_rc6_ms;
3814 
3815 	old->uncore_mhz = new->uncore_mhz;
3816 	old->gfx_mhz = new->gfx_mhz;
3817 	old->gfx_act_mhz = new->gfx_act_mhz;
3818 
3819 	/* flag an error when mc6 counter resets/wraps */
3820 	if (old->sam_mc6_ms > new->sam_mc6_ms)
3821 		old->sam_mc6_ms = -1;
3822 	else
3823 		old->sam_mc6_ms = new->sam_mc6_ms - old->sam_mc6_ms;
3824 
3825 	old->sam_mhz = new->sam_mhz;
3826 	old->sam_act_mhz = new->sam_act_mhz;
3827 
3828 	old->energy_pkg.raw_value = new->energy_pkg.raw_value - old->energy_pkg.raw_value;
3829 	old->energy_cores.raw_value = new->energy_cores.raw_value - old->energy_cores.raw_value;
3830 	old->energy_gfx.raw_value = new->energy_gfx.raw_value - old->energy_gfx.raw_value;
3831 	old->energy_dram.raw_value = new->energy_dram.raw_value - old->energy_dram.raw_value;
3832 	old->rapl_pkg_perf_status.raw_value = new->rapl_pkg_perf_status.raw_value - old->rapl_pkg_perf_status.raw_value;
3833 	old->rapl_dram_perf_status.raw_value = new->rapl_dram_perf_status.raw_value - old->rapl_dram_perf_status.raw_value;
3834 
3835 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
3836 		if (mp->format == FORMAT_RAW)
3837 			old->counter[i] = new->counter[i];
3838 		else if (mp->format == FORMAT_AVERAGE)
3839 			old->counter[i] = new->counter[i];
3840 		else
3841 			old->counter[i] = new->counter[i] - old->counter[i];
3842 	}
3843 
3844 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
3845 		if (pp->format == FORMAT_RAW)
3846 			old->perf_counter[i] = new->perf_counter[i];
3847 		else if (pp->format == FORMAT_AVERAGE)
3848 			old->perf_counter[i] = new->perf_counter[i];
3849 		else
3850 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3851 	}
3852 
3853 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
3854 		if (ppmt->format == FORMAT_RAW)
3855 			old->pmt_counter[i] = new->pmt_counter[i];
3856 		else
3857 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3858 	}
3859 
3860 	return 0;
3861 }
3862 
delta_core(struct core_data * new,struct core_data * old)3863 void delta_core(struct core_data *new, struct core_data *old)
3864 {
3865 	int i;
3866 	struct msr_counter *mp;
3867 	struct perf_counter_info *pp;
3868 	struct pmt_counter *ppmt;
3869 
3870 	old->c3 = new->c3 - old->c3;
3871 	old->c6 = new->c6 - old->c6;
3872 	old->c7 = new->c7 - old->c7;
3873 	old->core_temp_c = new->core_temp_c;
3874 	old->core_throt_cnt = new->core_throt_cnt - old->core_throt_cnt;
3875 	old->mc6_us = new->mc6_us - old->mc6_us;
3876 
3877 	DELTA_WRAP32(new->core_energy.raw_value, old->core_energy.raw_value);
3878 
3879 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
3880 		if (mp->format == FORMAT_RAW || mp->format == FORMAT_AVERAGE)
3881 			old->counter[i] = new->counter[i];
3882 		else
3883 			old->counter[i] = new->counter[i] - old->counter[i];
3884 	}
3885 
3886 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
3887 		if (pp->format == FORMAT_RAW)
3888 			old->perf_counter[i] = new->perf_counter[i];
3889 		else
3890 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
3891 	}
3892 
3893 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
3894 		if (ppmt->format == FORMAT_RAW)
3895 			old->pmt_counter[i] = new->pmt_counter[i];
3896 		else
3897 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
3898 	}
3899 }
3900 
soft_c1_residency_display(int bic)3901 int soft_c1_residency_display(int bic)
3902 {
3903 	if (!DO_BIC(BIC_CPU_c1) || platform->has_msr_core_c1_res)
3904 		return 0;
3905 
3906 	return DO_BIC_READ(bic);
3907 }
3908 
3909 /*
3910  * old = new - old
3911  */
delta_thread(struct thread_data * new,struct thread_data * old,struct core_data * core_delta)3912 int delta_thread(struct thread_data *new, struct thread_data *old, struct core_data *core_delta)
3913 {
3914 	int i;
3915 	struct msr_counter *mp;
3916 	struct perf_counter_info *pp;
3917 	struct pmt_counter *ppmt;
3918 
3919 	/* we run cpuid just the 1st time, copy the results */
3920 	if (DO_BIC(BIC_APIC))
3921 		new->apic_id = old->apic_id;
3922 	if (DO_BIC(BIC_X2APIC))
3923 		new->x2apic_id = old->x2apic_id;
3924 
3925 	/*
3926 	 * the timestamps from start of measurement interval are in "old"
3927 	 * the timestamp from end of measurement interval are in "new"
3928 	 * over-write old w/ new so we can print end of interval values
3929 	 */
3930 
3931 	timersub(&new->tv_begin, &old->tv_begin, &old->tv_delta);
3932 	old->tv_begin = new->tv_begin;
3933 	old->tv_end = new->tv_end;
3934 
3935 	old->tsc = new->tsc - old->tsc;
3936 
3937 	/* check for TSC < 1 Mcycles over interval */
3938 	if (old->tsc < (1000 * 1000))
3939 		errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
3940 		     "You can disable all c-states by booting with \"idle=poll\"\nor just the deep ones with \"processor.max_cstate=1\"");
3941 
3942 	old->c1 = new->c1 - old->c1;
3943 
3944 	if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz) || DO_BIC(BIC_IPC)
3945 	    || soft_c1_residency_display(BIC_Avg_MHz)) {
3946 		if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
3947 			old->aperf = new->aperf - old->aperf;
3948 			old->mperf = new->mperf - old->mperf;
3949 		} else {
3950 			return -1;
3951 		}
3952 	}
3953 
3954 	if (platform->has_msr_core_c1_res) {
3955 		/*
3956 		 * Some models have a dedicated C1 residency MSR,
3957 		 * which should be more accurate than the derivation below.
3958 		 */
3959 	} else {
3960 		/*
3961 		 * As counter collection is not atomic,
3962 		 * it is possible for mperf's non-halted cycles + idle states
3963 		 * to exceed TSC's all cycles: show c1 = 0% in that case.
3964 		 */
3965 		if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > (old->tsc * tsc_tweak))
3966 			old->c1 = 0;
3967 		else {
3968 			/* normal case, derive c1 */
3969 			old->c1 = (old->tsc * tsc_tweak) - old->mperf - core_delta->c3 - core_delta->c6 - core_delta->c7;
3970 		}
3971 	}
3972 
3973 	if (old->mperf == 0) {
3974 		if (debug > 1)
3975 			fprintf(outf, "cpu%d MPERF 0!\n", old->cpu_id);
3976 		old->mperf = 1;	/* divide by 0 protection */
3977 	}
3978 
3979 	if (DO_BIC(BIC_IPC))
3980 		old->instr_count = new->instr_count - old->instr_count;
3981 
3982 	if (DO_BIC(BIC_IRQ))
3983 		old->irq_count = new->irq_count - old->irq_count;
3984 
3985 	if (DO_BIC(BIC_NMI))
3986 		old->nmi_count = new->nmi_count - old->nmi_count;
3987 
3988 	if (DO_BIC(BIC_SMI))
3989 		old->smi_count = new->smi_count - old->smi_count;
3990 
3991 	if (DO_BIC(BIC_LLC_MRPS) || DO_BIC(BIC_LLC_HIT))
3992 		old->llc.references = new->llc.references - old->llc.references;
3993 
3994 	if (DO_BIC(BIC_LLC_HIT))
3995 		old->llc.misses = new->llc.misses - old->llc.misses;
3996 
3997 	if (DO_BIC(BIC_L2_MRPS) || DO_BIC(BIC_L2_HIT))
3998 		old->l2.references = new->l2.references - old->l2.references;
3999 
4000 	if (DO_BIC(BIC_L2_HIT))
4001 		old->l2.hits = new->l2.hits - old->l2.hits;
4002 
4003 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
4004 		if (mp->format == FORMAT_RAW || mp->format == FORMAT_AVERAGE)
4005 			old->counter[i] = new->counter[i];
4006 		else
4007 			old->counter[i] = new->counter[i] - old->counter[i];
4008 	}
4009 
4010 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
4011 		if (pp->format == FORMAT_RAW)
4012 			old->perf_counter[i] = new->perf_counter[i];
4013 		else
4014 			old->perf_counter[i] = new->perf_counter[i] - old->perf_counter[i];
4015 	}
4016 
4017 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
4018 		if (ppmt->format == FORMAT_RAW)
4019 			old->pmt_counter[i] = new->pmt_counter[i];
4020 		else
4021 			old->pmt_counter[i] = new->pmt_counter[i] - old->pmt_counter[i];
4022 	}
4023 
4024 	return 0;
4025 }
4026 
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)4027 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)
4028 {
4029 	int retval = 0;
4030 
4031 	/* calculate core delta only for 1st thread in core */
4032 	if (is_cpu_first_thread_in_core(t, c))
4033 		delta_core(c, c2);
4034 
4035 	/* always calculate thread delta */
4036 	retval = delta_thread(t, t2, c2);	/* c2 is core delta */
4037 
4038 	/* calculate package delta only for 1st core in package */
4039 	if (is_cpu_first_core_in_package(t, p))
4040 		retval |= delta_package(p, p2);
4041 
4042 	return retval;
4043 }
4044 
delta_platform(struct platform_counters * new,struct platform_counters * old)4045 void delta_platform(struct platform_counters *new, struct platform_counters *old)
4046 {
4047 	old->energy_psys.raw_value = new->energy_psys.raw_value - old->energy_psys.raw_value;
4048 }
4049 
rapl_counter_clear(struct rapl_counter * c)4050 void rapl_counter_clear(struct rapl_counter *c)
4051 {
4052 	c->raw_value = 0;
4053 	c->scale = 0.0;
4054 	c->unit = RAPL_UNIT_INVALID;
4055 }
4056 
clear_counters(PER_THREAD_PARAMS)4057 void clear_counters(PER_THREAD_PARAMS)
4058 {
4059 	int i;
4060 	struct msr_counter *mp;
4061 
4062 	t->tv_begin.tv_sec = 0;
4063 	t->tv_begin.tv_usec = 0;
4064 	t->tv_end.tv_sec = 0;
4065 	t->tv_end.tv_usec = 0;
4066 	t->tv_delta.tv_sec = 0;
4067 	t->tv_delta.tv_usec = 0;
4068 
4069 	t->tsc = 0;
4070 	t->aperf = 0;
4071 	t->mperf = 0;
4072 	t->c1 = 0;
4073 
4074 	t->instr_count = 0;
4075 
4076 	t->irq_count = 0;
4077 	t->nmi_count = 0;
4078 	t->smi_count = 0;
4079 
4080 	t->llc.references = 0;
4081 	t->llc.misses = 0;
4082 
4083 	t->l2.references = 0;
4084 	t->l2.hits = 0;
4085 
4086 	c->c3 = 0;
4087 	c->c6 = 0;
4088 	c->c7 = 0;
4089 	c->mc6_us = 0;
4090 	c->core_temp_c = 0;
4091 	rapl_counter_clear(&c->core_energy);
4092 	c->core_throt_cnt = 0;
4093 
4094 	p->pkg_wtd_core_c0 = 0;
4095 	p->pkg_any_core_c0 = 0;
4096 	p->pkg_any_gfxe_c0 = 0;
4097 	p->pkg_both_core_gfxe_c0 = 0;
4098 
4099 	p->pc2 = 0;
4100 	if (DO_BIC(BIC_Pkgpc3))
4101 		p->pc3 = 0;
4102 	if (DO_BIC(BIC_Pkgpc6))
4103 		p->pc6 = 0;
4104 	if (DO_BIC(BIC_Pkgpc7))
4105 		p->pc7 = 0;
4106 	p->pc8 = 0;
4107 	p->pc9 = 0;
4108 	p->pc10 = 0;
4109 	p->die_c6 = 0;
4110 	p->cpu_lpi = 0;
4111 	p->sys_lpi = 0;
4112 
4113 	rapl_counter_clear(&p->energy_pkg);
4114 	rapl_counter_clear(&p->energy_dram);
4115 	rapl_counter_clear(&p->energy_cores);
4116 	rapl_counter_clear(&p->energy_gfx);
4117 	rapl_counter_clear(&p->rapl_pkg_perf_status);
4118 	rapl_counter_clear(&p->rapl_dram_perf_status);
4119 	p->pkg_temp_c = 0;
4120 
4121 	p->gfx_rc6_ms = 0;
4122 	p->uncore_mhz = 0;
4123 	p->gfx_mhz = 0;
4124 	p->gfx_act_mhz = 0;
4125 	p->sam_mc6_ms = 0;
4126 	p->sam_mhz = 0;
4127 	p->sam_act_mhz = 0;
4128 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next)
4129 		t->counter[i] = 0;
4130 
4131 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next)
4132 		c->counter[i] = 0;
4133 
4134 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next)
4135 		p->counter[i] = 0;
4136 
4137 	memset(&t->perf_counter[0], 0, sizeof(t->perf_counter));
4138 	memset(&c->perf_counter[0], 0, sizeof(c->perf_counter));
4139 	memset(&p->perf_counter[0], 0, sizeof(p->perf_counter));
4140 
4141 	memset(&t->pmt_counter[0], 0, ARRAY_SIZE(t->pmt_counter));
4142 	memset(&c->pmt_counter[0], 0, ARRAY_SIZE(c->pmt_counter));
4143 	memset(&p->pmt_counter[0], 0, ARRAY_SIZE(p->pmt_counter));
4144 }
4145 
rapl_counter_accumulate(struct rapl_counter * dst,const struct rapl_counter * src)4146 void rapl_counter_accumulate(struct rapl_counter *dst, const struct rapl_counter *src)
4147 {
4148 	/* Copy unit and scale from src if dst is not initialized */
4149 	if (dst->unit == RAPL_UNIT_INVALID) {
4150 		dst->unit = src->unit;
4151 		dst->scale = src->scale;
4152 	}
4153 
4154 	assert(dst->unit == src->unit);
4155 	assert(dst->scale == src->scale);
4156 
4157 	dst->raw_value += src->raw_value;
4158 }
4159 
sum_counters(PER_THREAD_PARAMS)4160 int sum_counters(PER_THREAD_PARAMS)
4161 {
4162 	int i;
4163 	struct msr_counter *mp;
4164 	struct perf_counter_info *pp;
4165 	struct pmt_counter *ppmt;
4166 
4167 	/* copy un-changing apic_id's */
4168 	if (DO_BIC(BIC_APIC))
4169 		average.threads->apic_id = t->apic_id;
4170 	if (DO_BIC(BIC_X2APIC))
4171 		average.threads->x2apic_id = t->x2apic_id;
4172 
4173 	/* remember first tv_begin */
4174 	if (average.threads->tv_begin.tv_sec == 0)
4175 		average.threads->tv_begin = procsysfs_tv_begin;
4176 
4177 	/* remember last tv_end */
4178 	average.threads->tv_end = t->tv_end;
4179 
4180 	average.threads->tsc += t->tsc;
4181 	average.threads->aperf += t->aperf;
4182 	average.threads->mperf += t->mperf;
4183 	average.threads->c1 += t->c1;
4184 
4185 	average.threads->instr_count += t->instr_count;
4186 
4187 	average.threads->irq_count += t->irq_count;
4188 	average.threads->nmi_count += t->nmi_count;
4189 	average.threads->smi_count += t->smi_count;
4190 
4191 	average.threads->llc.references += t->llc.references;
4192 	average.threads->llc.misses += t->llc.misses;
4193 
4194 	average.threads->l2.references += t->l2.references;
4195 	average.threads->l2.hits += t->l2.hits;
4196 
4197 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
4198 		if (mp->format == FORMAT_RAW)
4199 			continue;
4200 		average.threads->counter[i] += t->counter[i];
4201 	}
4202 
4203 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
4204 		if (pp->format == FORMAT_RAW)
4205 			continue;
4206 		average.threads->perf_counter[i] += t->perf_counter[i];
4207 	}
4208 
4209 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
4210 		average.threads->pmt_counter[i] += t->pmt_counter[i];
4211 	}
4212 
4213 	/* sum per-core values only for 1st thread in core */
4214 	if (!is_cpu_first_thread_in_core(t, c))
4215 		return 0;
4216 
4217 	average.cores->c3 += c->c3;
4218 	average.cores->c6 += c->c6;
4219 	average.cores->c7 += c->c7;
4220 	average.cores->mc6_us += c->mc6_us;
4221 
4222 	average.cores->core_temp_c = MAX(average.cores->core_temp_c, c->core_temp_c);
4223 	average.cores->core_throt_cnt = MAX(average.cores->core_throt_cnt, c->core_throt_cnt);
4224 
4225 	rapl_counter_accumulate(&average.cores->core_energy, &c->core_energy);
4226 
4227 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
4228 		if (mp->format == FORMAT_RAW)
4229 			continue;
4230 		average.cores->counter[i] += c->counter[i];
4231 	}
4232 
4233 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
4234 		if (pp->format == FORMAT_RAW)
4235 			continue;
4236 		average.cores->perf_counter[i] += c->perf_counter[i];
4237 	}
4238 
4239 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
4240 		average.cores->pmt_counter[i] += c->pmt_counter[i];
4241 	}
4242 
4243 	/* sum per-pkg values only for 1st core in pkg */
4244 	if (!is_cpu_first_core_in_package(t, p))
4245 		return 0;
4246 
4247 	if (DO_BIC(BIC_Totl_c0))
4248 		average.packages->pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
4249 	if (DO_BIC(BIC_Any_c0))
4250 		average.packages->pkg_any_core_c0 += p->pkg_any_core_c0;
4251 	if (DO_BIC(BIC_GFX_c0))
4252 		average.packages->pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
4253 	if (DO_BIC(BIC_CPUGFX))
4254 		average.packages->pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
4255 
4256 	average.packages->pc2 += p->pc2;
4257 	if (DO_BIC(BIC_Pkgpc3))
4258 		average.packages->pc3 += p->pc3;
4259 	if (DO_BIC(BIC_Pkgpc6))
4260 		average.packages->pc6 += p->pc6;
4261 	if (DO_BIC(BIC_Pkgpc7))
4262 		average.packages->pc7 += p->pc7;
4263 	average.packages->pc8 += p->pc8;
4264 	average.packages->pc9 += p->pc9;
4265 	average.packages->pc10 += p->pc10;
4266 	average.packages->die_c6 += p->die_c6;
4267 
4268 	average.packages->cpu_lpi = p->cpu_lpi;
4269 	average.packages->sys_lpi = p->sys_lpi;
4270 
4271 	rapl_counter_accumulate(&average.packages->energy_pkg, &p->energy_pkg);
4272 	rapl_counter_accumulate(&average.packages->energy_dram, &p->energy_dram);
4273 	rapl_counter_accumulate(&average.packages->energy_cores, &p->energy_cores);
4274 	rapl_counter_accumulate(&average.packages->energy_gfx, &p->energy_gfx);
4275 
4276 	average.packages->gfx_rc6_ms = p->gfx_rc6_ms;
4277 	average.packages->uncore_mhz = p->uncore_mhz;
4278 	average.packages->gfx_mhz = p->gfx_mhz;
4279 	average.packages->gfx_act_mhz = p->gfx_act_mhz;
4280 	average.packages->sam_mc6_ms = p->sam_mc6_ms;
4281 	average.packages->sam_mhz = p->sam_mhz;
4282 	average.packages->sam_act_mhz = p->sam_act_mhz;
4283 
4284 	average.packages->pkg_temp_c = MAX(average.packages->pkg_temp_c, p->pkg_temp_c);
4285 
4286 	rapl_counter_accumulate(&average.packages->rapl_pkg_perf_status, &p->rapl_pkg_perf_status);
4287 	rapl_counter_accumulate(&average.packages->rapl_dram_perf_status, &p->rapl_dram_perf_status);
4288 
4289 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
4290 		if ((mp->format == FORMAT_RAW) && (topo.num_packages == 0))
4291 			average.packages->counter[i] = p->counter[i];
4292 		else
4293 			average.packages->counter[i] += p->counter[i];
4294 	}
4295 
4296 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
4297 		if ((pp->format == FORMAT_RAW) && (topo.num_packages == 0))
4298 			average.packages->perf_counter[i] = p->perf_counter[i];
4299 		else
4300 			average.packages->perf_counter[i] += p->perf_counter[i];
4301 	}
4302 
4303 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
4304 		average.packages->pmt_counter[i] += p->pmt_counter[i];
4305 	}
4306 
4307 	return 0;
4308 }
4309 
4310 /*
4311  * sum the counters for all cpus in the system
4312  * compute the weighted average
4313  */
compute_average(PER_THREAD_PARAMS)4314 void compute_average(PER_THREAD_PARAMS)
4315 {
4316 	int i;
4317 	struct msr_counter *mp;
4318 	struct perf_counter_info *pp;
4319 	struct pmt_counter *ppmt;
4320 
4321 	clear_counters(average.threads, average.cores, average.packages);
4322 
4323 	for_all_cpus(sum_counters, t, c, p);
4324 
4325 	/* Use the global time delta for the average. */
4326 	average.threads->tv_delta = tv_delta;
4327 
4328 	average.threads->tsc /= topo.allowed_cpus;
4329 	average.threads->aperf /= topo.allowed_cpus;
4330 	average.threads->mperf /= topo.allowed_cpus;
4331 	average.threads->instr_count /= topo.allowed_cpus;
4332 	average.threads->c1 /= topo.allowed_cpus;
4333 
4334 	if (average.threads->irq_count > 9999999)
4335 		sums_need_wide_columns = 1;
4336 	if (average.threads->nmi_count > 9999999)
4337 		sums_need_wide_columns = 1;
4338 
4339 	average.cores->c3 /= topo.allowed_cores;
4340 	average.cores->c6 /= topo.allowed_cores;
4341 	average.cores->c7 /= topo.allowed_cores;
4342 	average.cores->mc6_us /= topo.allowed_cores;
4343 
4344 	if (DO_BIC(BIC_Totl_c0))
4345 		average.packages->pkg_wtd_core_c0 /= topo.allowed_packages;
4346 	if (DO_BIC(BIC_Any_c0))
4347 		average.packages->pkg_any_core_c0 /= topo.allowed_packages;
4348 	if (DO_BIC(BIC_GFX_c0))
4349 		average.packages->pkg_any_gfxe_c0 /= topo.allowed_packages;
4350 	if (DO_BIC(BIC_CPUGFX))
4351 		average.packages->pkg_both_core_gfxe_c0 /= topo.allowed_packages;
4352 
4353 	average.packages->pc2 /= topo.allowed_packages;
4354 	if (DO_BIC(BIC_Pkgpc3))
4355 		average.packages->pc3 /= topo.allowed_packages;
4356 	if (DO_BIC(BIC_Pkgpc6))
4357 		average.packages->pc6 /= topo.allowed_packages;
4358 	if (DO_BIC(BIC_Pkgpc7))
4359 		average.packages->pc7 /= topo.allowed_packages;
4360 
4361 	average.packages->pc8 /= topo.allowed_packages;
4362 	average.packages->pc9 /= topo.allowed_packages;
4363 	average.packages->pc10 /= topo.allowed_packages;
4364 	average.packages->die_c6 /= topo.allowed_packages;
4365 
4366 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
4367 		if (mp->format == FORMAT_RAW)
4368 			continue;
4369 		if (mp->type == COUNTER_ITEMS) {
4370 			if (average.threads->counter[i] > 9999999)
4371 				sums_need_wide_columns = 1;
4372 			continue;
4373 		}
4374 		average.threads->counter[i] /= topo.allowed_cpus;
4375 	}
4376 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
4377 		if (mp->format == FORMAT_RAW)
4378 			continue;
4379 		if (mp->type == COUNTER_ITEMS) {
4380 			if (average.cores->counter[i] > 9999999)
4381 				sums_need_wide_columns = 1;
4382 		}
4383 		average.cores->counter[i] /= topo.allowed_cores;
4384 	}
4385 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
4386 		if (mp->format == FORMAT_RAW)
4387 			continue;
4388 		if (mp->type == COUNTER_ITEMS) {
4389 			if (average.packages->counter[i] > 9999999)
4390 				sums_need_wide_columns = 1;
4391 		}
4392 		average.packages->counter[i] /= topo.allowed_packages;
4393 	}
4394 
4395 	for (i = 0, pp = sys.perf_tp; pp; i++, pp = pp->next) {
4396 		if (pp->format == FORMAT_RAW)
4397 			continue;
4398 		if (pp->type == COUNTER_ITEMS) {
4399 			if (average.threads->perf_counter[i] > 9999999)
4400 				sums_need_wide_columns = 1;
4401 			continue;
4402 		}
4403 		average.threads->perf_counter[i] /= topo.allowed_cpus;
4404 	}
4405 	for (i = 0, pp = sys.perf_cp; pp; i++, pp = pp->next) {
4406 		if (pp->format == FORMAT_RAW)
4407 			continue;
4408 		if (pp->type == COUNTER_ITEMS) {
4409 			if (average.cores->perf_counter[i] > 9999999)
4410 				sums_need_wide_columns = 1;
4411 		}
4412 		average.cores->perf_counter[i] /= topo.allowed_cores;
4413 	}
4414 	for (i = 0, pp = sys.perf_pp; pp; i++, pp = pp->next) {
4415 		if (pp->format == FORMAT_RAW)
4416 			continue;
4417 		if (pp->type == COUNTER_ITEMS) {
4418 			if (average.packages->perf_counter[i] > 9999999)
4419 				sums_need_wide_columns = 1;
4420 		}
4421 		average.packages->perf_counter[i] /= topo.allowed_packages;
4422 	}
4423 
4424 	for (i = 0, ppmt = sys.pmt_tp; ppmt; i++, ppmt = ppmt->next) {
4425 		average.threads->pmt_counter[i] /= topo.allowed_cpus;
4426 	}
4427 	for (i = 0, ppmt = sys.pmt_cp; ppmt; i++, ppmt = ppmt->next) {
4428 		average.cores->pmt_counter[i] /= topo.allowed_cores;
4429 	}
4430 	for (i = 0, ppmt = sys.pmt_pp; ppmt; i++, ppmt = ppmt->next) {
4431 		average.packages->pmt_counter[i] /= topo.allowed_packages;
4432 	}
4433 }
4434 
rdtsc(void)4435 static unsigned long long rdtsc(void)
4436 {
4437 	unsigned int low, high;
4438 
4439 	asm volatile ("rdtsc":"=a" (low), "=d"(high));
4440 
4441 	return low | ((unsigned long long)high) << 32;
4442 }
4443 
4444 /*
4445  * Open a file, and exit on failure
4446  */
fopen_or_die(const char * path,const char * mode)4447 FILE *fopen_or_die(const char *path, const char *mode)
4448 {
4449 	FILE *filep = fopen(path, mode);
4450 
4451 	if (!filep)
4452 		err(1, "%s: open failed", path);
4453 	return filep;
4454 }
4455 
4456 /*
4457  * snapshot_sysfs_counter()
4458  *
4459  * return snapshot of given counter
4460  */
snapshot_sysfs_counter(char * path)4461 unsigned long long snapshot_sysfs_counter(char *path)
4462 {
4463 	FILE *fp;
4464 	int retval;
4465 	unsigned long long counter;
4466 
4467 	fp = fopen_or_die(path, "r");
4468 
4469 	retval = fscanf(fp, "%lld", &counter);
4470 	if (retval != 1)
4471 		err(1, "snapshot_sysfs_counter(%s)", path);
4472 
4473 	fclose(fp);
4474 
4475 	return counter;
4476 }
4477 
get_mp(int cpu,struct msr_counter * mp,unsigned long long * counterp,char * counter_path)4478 int get_mp(int cpu, struct msr_counter *mp, unsigned long long *counterp, char *counter_path)
4479 {
4480 	if (mp->msr_num != 0) {
4481 		assert(!no_msr);
4482 		if (get_msr(cpu, mp->msr_num, counterp))
4483 			return -1;
4484 	} else {
4485 		char path[128 + PATH_BYTES];
4486 
4487 		if (mp->flags & SYSFS_PERCPU) {
4488 			sprintf(path, "/sys/devices/system/cpu/cpu%d/%s", cpu, mp->sp->path);
4489 
4490 			*counterp = snapshot_sysfs_counter(path);
4491 		} else {
4492 			*counterp = snapshot_sysfs_counter(counter_path);
4493 		}
4494 	}
4495 
4496 	return 0;
4497 }
4498 
get_legacy_uncore_mhz(int package)4499 unsigned long long get_legacy_uncore_mhz(int package)
4500 {
4501 	char path[128];
4502 	int die;
4503 	static int warn_once;
4504 
4505 	/*
4506 	 * for this package, use the first die_id that exists
4507 	 */
4508 	for (die = 0; die <= topo.max_die_id; ++die) {
4509 
4510 		sprintf(path, "/sys/devices/system/cpu/intel_uncore_frequency/package_%02d_die_%02d/current_freq_khz", package, die);
4511 
4512 		if (access(path, R_OK) == 0)
4513 			return (snapshot_sysfs_counter(path) / 1000);
4514 	}
4515 	if (!warn_once) {
4516 		warnx("BUG: %s: No %s", __func__, path);
4517 		warn_once = 1;
4518 	}
4519 
4520 	return 0;
4521 }
4522 
get_epb(int cpu)4523 int get_epb(int cpu)
4524 {
4525 	char path[128 + PATH_BYTES];
4526 	unsigned long long msr;
4527 	int ret, epb = -1;
4528 	FILE *fp;
4529 
4530 	sprintf(path, "/sys/devices/system/cpu/cpu%d/power/energy_perf_bias", cpu);
4531 
4532 	fp = fopen(path, "r");
4533 	if (!fp)
4534 		goto msr_fallback;
4535 
4536 	ret = fscanf(fp, "%d", &epb);
4537 	if (ret != 1)
4538 		err(1, "%s(%s)", __func__, path);
4539 
4540 	fclose(fp);
4541 
4542 	return epb;
4543 
4544 msr_fallback:
4545 	if (no_msr)
4546 		return -1;
4547 
4548 	get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr);
4549 
4550 	return msr & 0xf;
4551 }
4552 
get_apic_id(struct thread_data * t)4553 void get_apic_id(struct thread_data *t)
4554 {
4555 	unsigned int eax, ebx, ecx, edx;
4556 
4557 	if (DO_BIC(BIC_APIC)) {
4558 		eax = ebx = ecx = edx = 0;
4559 		__cpuid(1, eax, ebx, ecx, edx);
4560 
4561 		t->apic_id = (ebx >> 24) & 0xff;
4562 	}
4563 
4564 	if (!DO_BIC(BIC_X2APIC))
4565 		return;
4566 
4567 	if (authentic_amd || hygon_genuine) {
4568 		unsigned int topology_extensions;
4569 
4570 		if (max_extended_level < 0x8000001e)
4571 			return;
4572 
4573 		eax = ebx = ecx = edx = 0;
4574 		__cpuid(0x80000001, eax, ebx, ecx, edx);
4575 		topology_extensions = ecx & (1 << 22);
4576 
4577 		if (topology_extensions == 0)
4578 			return;
4579 
4580 		eax = ebx = ecx = edx = 0;
4581 		__cpuid(0x8000001e, eax, ebx, ecx, edx);
4582 
4583 		t->x2apic_id = eax;
4584 		return;
4585 	}
4586 
4587 	if (!genuine_intel)
4588 		return;
4589 
4590 	if (max_level < 0xb)
4591 		return;
4592 
4593 	ecx = 0;
4594 	__cpuid(0xb, eax, ebx, ecx, edx);
4595 	t->x2apic_id = edx;
4596 
4597 	if (debug && (t->apic_id != (t->x2apic_id & 0xff)))
4598 		fprintf(outf, "cpu%d: BIOS BUG: apic 0x%x x2apic 0x%x\n", t->cpu_id, t->apic_id, t->x2apic_id);
4599 }
4600 
get_core_throt_cnt(int cpu,unsigned long long * cnt)4601 int get_core_throt_cnt(int cpu, unsigned long long *cnt)
4602 {
4603 	char path[128 + PATH_BYTES];
4604 	unsigned long long tmp;
4605 	FILE *fp;
4606 	int ret;
4607 
4608 	sprintf(path, "/sys/devices/system/cpu/cpu%d/thermal_throttle/core_throttle_count", cpu);
4609 	fp = fopen(path, "r");
4610 	if (!fp)
4611 		return -1;
4612 	ret = fscanf(fp, "%lld", &tmp);
4613 	fclose(fp);
4614 	if (ret != 1)
4615 		return -1;
4616 	*cnt = tmp;
4617 
4618 	return 0;
4619 }
4620 
read_perf_counter_info(const char * const path,const char * const parse_format,void * value_ptr)4621 static int read_perf_counter_info(const char *const path, const char *const parse_format, void *value_ptr)
4622 {
4623 	int fdmt;
4624 	int bytes_read;
4625 	char buf[64];
4626 	int ret = -1;
4627 
4628 	fdmt = open(path, O_RDONLY, 0);
4629 	if (fdmt == -1) {
4630 		if (debug)
4631 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4632 		ret = -1;
4633 		goto cleanup_and_exit;
4634 	}
4635 
4636 	bytes_read = read(fdmt, buf, sizeof(buf) - 1);
4637 	if (bytes_read <= 0 || bytes_read >= (int)sizeof(buf)) {
4638 		if (debug)
4639 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4640 		ret = -1;
4641 		goto cleanup_and_exit;
4642 	}
4643 
4644 	buf[bytes_read] = '\0';
4645 
4646 	if (sscanf(buf, parse_format, value_ptr) != 1) {
4647 		if (debug)
4648 			fprintf(stderr, "Failed to parse perf counter info %s\n", path);
4649 		ret = -1;
4650 		goto cleanup_and_exit;
4651 	}
4652 
4653 	ret = 0;
4654 
4655 cleanup_and_exit:
4656 	close(fdmt);
4657 	return ret;
4658 }
4659 
read_perf_counter_info_n(const char * const path,const char * const parse_format)4660 static unsigned int read_perf_counter_info_n(const char *const path, const char *const parse_format)
4661 {
4662 	unsigned int v;
4663 	int status;
4664 
4665 	status = read_perf_counter_info(path, parse_format, &v);
4666 	if (status)
4667 		v = -1;
4668 
4669 	return v;
4670 }
4671 
read_perf_type(const char * subsys)4672 static unsigned int read_perf_type(const char *subsys)
4673 {
4674 	const char *const path_format = "/sys/bus/event_source/devices/%s/type";
4675 	const char *const format = "%u";
4676 	char path[128];
4677 
4678 	snprintf(path, sizeof(path), path_format, subsys);
4679 
4680 	return read_perf_counter_info_n(path, format);
4681 }
4682 
read_perf_config(const char * subsys,const char * event_name)4683 static unsigned int read_perf_config(const char *subsys, const char *event_name)
4684 {
4685 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s";
4686 	FILE *fconfig = NULL;
4687 	char path[128];
4688 	char config_str[64];
4689 	unsigned int config;
4690 	unsigned int umask;
4691 	bool has_config = false;
4692 	bool has_umask = false;
4693 	unsigned int ret = -1;
4694 
4695 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4696 
4697 	fconfig = fopen(path, "r");
4698 	if (!fconfig)
4699 		return -1;
4700 
4701 	if (fgets(config_str, ARRAY_SIZE(config_str), fconfig) != config_str)
4702 		goto cleanup_and_exit;
4703 
4704 	for (char *pconfig_str = &config_str[0]; pconfig_str;) {
4705 		if (sscanf(pconfig_str, "event=%x", &config) == 1) {
4706 			has_config = true;
4707 			goto next;
4708 		}
4709 
4710 		if (sscanf(pconfig_str, "umask=%x", &umask) == 1) {
4711 			has_umask = true;
4712 			goto next;
4713 		}
4714 
4715 next:
4716 		pconfig_str = strchr(pconfig_str, ',');
4717 		if (pconfig_str) {
4718 			*pconfig_str = '\0';
4719 			++pconfig_str;
4720 		}
4721 	}
4722 
4723 	if (!has_umask)
4724 		umask = 0;
4725 
4726 	if (has_config)
4727 		ret = (umask << 8) | config;
4728 
4729 cleanup_and_exit:
4730 	fclose(fconfig);
4731 	return ret;
4732 }
4733 
read_perf_rapl_unit(const char * subsys,const char * event_name)4734 static unsigned int read_perf_rapl_unit(const char *subsys, const char *event_name)
4735 {
4736 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.unit";
4737 	const char *const format = "%s";
4738 	char path[128];
4739 	char unit_buffer[16];
4740 
4741 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4742 
4743 	read_perf_counter_info(path, format, &unit_buffer);
4744 	if (strcmp("Joules", unit_buffer) == 0)
4745 		return RAPL_UNIT_JOULES;
4746 
4747 	return RAPL_UNIT_INVALID;
4748 }
4749 
read_perf_scale(const char * subsys,const char * event_name)4750 static double read_perf_scale(const char *subsys, const char *event_name)
4751 {
4752 	const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.scale";
4753 	const char *const format = "%lf";
4754 	char path[128];
4755 	double scale;
4756 
4757 	snprintf(path, sizeof(path), path_format, subsys, event_name);
4758 
4759 	if (read_perf_counter_info(path, format, &scale))
4760 		return 0.0;
4761 
4762 	return scale;
4763 }
4764 
rapl_counter_info_count_perf(const struct rapl_counter_info_t * rci)4765 size_t rapl_counter_info_count_perf(const struct rapl_counter_info_t *rci)
4766 {
4767 	size_t ret = 0;
4768 
4769 	for (int i = 0; i < NUM_RAPL_COUNTERS; ++i)
4770 		if (rci->source[i] == COUNTER_SOURCE_PERF)
4771 			++ret;
4772 
4773 	return ret;
4774 }
4775 
cstate_counter_info_count_perf(const struct cstate_counter_info_t * cci)4776 static size_t cstate_counter_info_count_perf(const struct cstate_counter_info_t *cci)
4777 {
4778 	size_t ret = 0;
4779 
4780 	for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i)
4781 		if (cci->source[i] == COUNTER_SOURCE_PERF)
4782 			++ret;
4783 
4784 	return ret;
4785 }
4786 
write_rapl_counter(struct rapl_counter * rc,struct rapl_counter_info_t * rci,unsigned int idx)4787 void write_rapl_counter(struct rapl_counter *rc, struct rapl_counter_info_t *rci, unsigned int idx)
4788 {
4789 	if (rci->source[idx] == COUNTER_SOURCE_NONE)
4790 		return;
4791 
4792 	rc->raw_value = rci->data[idx];
4793 	rc->unit = rci->unit[idx];
4794 	rc->scale = rci->scale[idx];
4795 }
4796 
get_rapl_counters(int cpu,unsigned int domain,struct core_data * c,struct pkg_data * p)4797 int get_rapl_counters(int cpu, unsigned int domain, struct core_data *c, struct pkg_data *p)
4798 {
4799 	struct platform_counters *pplat_cnt = p == odd.packages ? &platform_counters_odd : &platform_counters_even;
4800 	unsigned long long perf_data[NUM_RAPL_COUNTERS + 1];
4801 	struct rapl_counter_info_t *rci;
4802 
4803 	if (debug >= 2)
4804 		fprintf(stderr, "%s: cpu%d domain%d\n", __func__, cpu, domain);
4805 
4806 	assert(rapl_counter_info_perdomain);
4807 	assert(domain < rapl_counter_info_perdomain_size);
4808 
4809 	rci = &rapl_counter_info_perdomain[domain];
4810 
4811 	/*
4812 	 * If we have any perf counters to read, read them all now, in bulk
4813 	 */
4814 	if (rci->fd_perf != -1) {
4815 		size_t num_perf_counters = rapl_counter_info_count_perf(rci);
4816 		const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
4817 		const ssize_t actual_read_size = read(rci->fd_perf, &perf_data[0], sizeof(perf_data));
4818 
4819 		if (actual_read_size != expected_read_size)
4820 			err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size, actual_read_size);
4821 	}
4822 
4823 	for (unsigned int i = 0, pi = 1; i < NUM_RAPL_COUNTERS; ++i) {
4824 		switch (rci->source[i]) {
4825 		case COUNTER_SOURCE_NONE:
4826 			rci->data[i] = 0;
4827 			break;
4828 
4829 		case COUNTER_SOURCE_PERF:
4830 			assert(pi < ARRAY_SIZE(perf_data));
4831 			assert(rci->fd_perf != -1);
4832 
4833 			if (debug >= 2)
4834 				fprintf(stderr, "Reading rapl counter via perf at %u (%llu %e %lf)\n",
4835 					i, perf_data[pi], rci->scale[i], perf_data[pi] * rci->scale[i]);
4836 
4837 			rci->data[i] = perf_data[pi];
4838 
4839 			++pi;
4840 			break;
4841 
4842 		case COUNTER_SOURCE_MSR:
4843 			if (debug >= 2)
4844 				fprintf(stderr, "Reading rapl counter via msr at %u\n", i);
4845 
4846 			assert(!no_msr);
4847 			if (rci->flags[i] & RAPL_COUNTER_FLAG_USE_MSR_SUM) {
4848 				if (get_msr_sum(cpu, rci->msr[i], &rci->data[i]))
4849 					return -13 - i;
4850 			} else {
4851 				if (get_msr(cpu, rci->msr[i], &rci->data[i]))
4852 					return -13 - i;
4853 			}
4854 
4855 			rci->data[i] &= rci->msr_mask[i];
4856 			if (rci->msr_shift[i] >= 0)
4857 				rci->data[i] >>= abs(rci->msr_shift[i]);
4858 			else
4859 				rci->data[i] <<= abs(rci->msr_shift[i]);
4860 
4861 			break;
4862 		}
4863 	}
4864 
4865 	BUILD_BUG_ON(NUM_RAPL_COUNTERS != 8);
4866 	write_rapl_counter(&p->energy_pkg, rci, RAPL_RCI_INDEX_ENERGY_PKG);
4867 	write_rapl_counter(&p->energy_cores, rci, RAPL_RCI_INDEX_ENERGY_CORES);
4868 	write_rapl_counter(&p->energy_dram, rci, RAPL_RCI_INDEX_DRAM);
4869 	write_rapl_counter(&p->energy_gfx, rci, RAPL_RCI_INDEX_GFX);
4870 	write_rapl_counter(&p->rapl_pkg_perf_status, rci, RAPL_RCI_INDEX_PKG_PERF_STATUS);
4871 	write_rapl_counter(&p->rapl_dram_perf_status, rci, RAPL_RCI_INDEX_DRAM_PERF_STATUS);
4872 	write_rapl_counter(&c->core_energy, rci, RAPL_RCI_INDEX_CORE_ENERGY);
4873 	write_rapl_counter(&pplat_cnt->energy_psys, rci, RAPL_RCI_INDEX_ENERGY_PLATFORM);
4874 
4875 	return 0;
4876 }
4877 
find_sysfs_path_by_id(struct sysfs_path * sp,int id)4878 char *find_sysfs_path_by_id(struct sysfs_path *sp, int id)
4879 {
4880 	while (sp) {
4881 		if (sp->id == id)
4882 			return (sp->path);
4883 		sp = sp->next;
4884 	}
4885 	if (debug)
4886 		warnx("%s: id%d not found", __func__, id);
4887 	return NULL;
4888 }
4889 
get_cstate_counters(unsigned int cpu,PER_THREAD_PARAMS)4890 int get_cstate_counters(unsigned int cpu, PER_THREAD_PARAMS)
4891 {
4892 	/*
4893 	 * Overcommit memory a little bit here,
4894 	 * but skip calculating exact sizes for the buffers.
4895 	 */
4896 	unsigned long long perf_data[NUM_CSTATE_COUNTERS];
4897 	unsigned long long perf_data_core[NUM_CSTATE_COUNTERS + 1];
4898 	unsigned long long perf_data_pkg[NUM_CSTATE_COUNTERS + 1];
4899 
4900 	struct cstate_counter_info_t *cci;
4901 
4902 	if (debug >= 2)
4903 		fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
4904 
4905 	assert(ccstate_counter_info);
4906 	assert(cpu <= ccstate_counter_info_size);
4907 
4908 	ZERO_ARRAY(perf_data);
4909 	ZERO_ARRAY(perf_data_core);
4910 	ZERO_ARRAY(perf_data_pkg);
4911 
4912 	cci = &ccstate_counter_info[cpu];
4913 
4914 	/*
4915 	 * If we have any perf counters to read, read them all now, in bulk
4916 	 */
4917 	const size_t num_perf_counters = cstate_counter_info_count_perf(cci);
4918 	ssize_t expected_read_size = num_perf_counters * sizeof(unsigned long long);
4919 	ssize_t actual_read_size_core = 0, actual_read_size_pkg = 0;
4920 
4921 	if (cci->fd_perf_core != -1) {
4922 		/* Each descriptor read begins with number of counters read. */
4923 		expected_read_size += sizeof(unsigned long long);
4924 
4925 		actual_read_size_core = read(cci->fd_perf_core, &perf_data_core[0], sizeof(perf_data_core));
4926 
4927 		if (actual_read_size_core <= 0)
4928 			err(-1, "%s: read perf %s: %ld", __func__, "core", actual_read_size_core);
4929 	}
4930 
4931 	if (cci->fd_perf_pkg != -1) {
4932 		/* Each descriptor read begins with number of counters read. */
4933 		expected_read_size += sizeof(unsigned long long);
4934 
4935 		actual_read_size_pkg = read(cci->fd_perf_pkg, &perf_data_pkg[0], sizeof(perf_data_pkg));
4936 
4937 		if (actual_read_size_pkg <= 0)
4938 			err(-1, "%s: read perf %s: %ld", __func__, "pkg", actual_read_size_pkg);
4939 	}
4940 
4941 	const ssize_t actual_read_size_total = actual_read_size_core + actual_read_size_pkg;
4942 
4943 	if (actual_read_size_total != expected_read_size)
4944 		err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size, actual_read_size_total);
4945 
4946 	/*
4947 	 * Copy ccstate and pcstate data into unified buffer.
4948 	 *
4949 	 * Skip first element from core and pkg buffers.
4950 	 * Kernel puts there how many counters were read.
4951 	 */
4952 	const size_t num_core_counters = perf_data_core[0];
4953 	const size_t num_pkg_counters = perf_data_pkg[0];
4954 
4955 	assert(num_perf_counters == num_core_counters + num_pkg_counters);
4956 
4957 	/* Copy ccstate perf data */
4958 	memcpy(&perf_data[0], &perf_data_core[1], num_core_counters * sizeof(unsigned long long));
4959 
4960 	/* Copy pcstate perf data */
4961 	memcpy(&perf_data[num_core_counters], &perf_data_pkg[1], num_pkg_counters * sizeof(unsigned long long));
4962 
4963 	for (unsigned int i = 0, pi = 0; i < NUM_CSTATE_COUNTERS; ++i) {
4964 		switch (cci->source[i]) {
4965 		case COUNTER_SOURCE_NONE:
4966 			break;
4967 
4968 		case COUNTER_SOURCE_PERF:
4969 			assert(pi < ARRAY_SIZE(perf_data));
4970 			assert(cci->fd_perf_core != -1 || cci->fd_perf_pkg != -1);
4971 
4972 			if (debug >= 2)
4973 				fprintf(stderr, "cstate via %s %u: %llu\n", "perf", i, perf_data[pi]);
4974 
4975 			cci->data[i] = perf_data[pi];
4976 
4977 			++pi;
4978 			break;
4979 
4980 		case COUNTER_SOURCE_MSR:
4981 			assert(!no_msr);
4982 			if (get_msr(cpu, cci->msr[i], &cci->data[i]))
4983 				return -13 - i;
4984 
4985 			if (debug >= 2)
4986 				fprintf(stderr, "cstate via %s0x%llx %u: %llu\n", "msr", cci->msr[i], i, cci->data[i]);
4987 
4988 			break;
4989 		}
4990 	}
4991 
4992 	/*
4993 	 * Helper to write the data only if the source of
4994 	 * the counter for the current cpu is not none.
4995 	 *
4996 	 * Otherwise we would overwrite core data with 0 (default value),
4997 	 * when invoked for the thread sibling.
4998 	 */
4999 #define PERF_COUNTER_WRITE_DATA(out_counter, index) do {	\
5000 	if (cci->source[index] != COUNTER_SOURCE_NONE)		\
5001 		out_counter = cci->data[index];			\
5002 } while (0)
5003 
5004 	BUILD_BUG_ON(NUM_CSTATE_COUNTERS != 11);
5005 
5006 	PERF_COUNTER_WRITE_DATA(t->c1, CCSTATE_RCI_INDEX_C1_RESIDENCY);
5007 	PERF_COUNTER_WRITE_DATA(c->c3, CCSTATE_RCI_INDEX_C3_RESIDENCY);
5008 	PERF_COUNTER_WRITE_DATA(c->c6, CCSTATE_RCI_INDEX_C6_RESIDENCY);
5009 	PERF_COUNTER_WRITE_DATA(c->c7, CCSTATE_RCI_INDEX_C7_RESIDENCY);
5010 
5011 	PERF_COUNTER_WRITE_DATA(p->pc2, PCSTATE_RCI_INDEX_C2_RESIDENCY);
5012 	PERF_COUNTER_WRITE_DATA(p->pc3, PCSTATE_RCI_INDEX_C3_RESIDENCY);
5013 	PERF_COUNTER_WRITE_DATA(p->pc6, PCSTATE_RCI_INDEX_C6_RESIDENCY);
5014 	PERF_COUNTER_WRITE_DATA(p->pc7, PCSTATE_RCI_INDEX_C7_RESIDENCY);
5015 	PERF_COUNTER_WRITE_DATA(p->pc8, PCSTATE_RCI_INDEX_C8_RESIDENCY);
5016 	PERF_COUNTER_WRITE_DATA(p->pc9, PCSTATE_RCI_INDEX_C9_RESIDENCY);
5017 	PERF_COUNTER_WRITE_DATA(p->pc10, PCSTATE_RCI_INDEX_C10_RESIDENCY);
5018 
5019 #undef PERF_COUNTER_WRITE_DATA
5020 
5021 	return 0;
5022 }
5023 
msr_counter_info_count_perf(const struct msr_counter_info_t * mci)5024 size_t msr_counter_info_count_perf(const struct msr_counter_info_t *mci)
5025 {
5026 	size_t ret = 0;
5027 
5028 	for (int i = 0; i < NUM_MSR_COUNTERS; ++i)
5029 		if (mci->source[i] == COUNTER_SOURCE_PERF)
5030 			++ret;
5031 
5032 	return ret;
5033 }
5034 
get_smi_aperf_mperf(unsigned int cpu,struct thread_data * t)5035 int get_smi_aperf_mperf(unsigned int cpu, struct thread_data *t)
5036 {
5037 	unsigned long long perf_data[NUM_MSR_COUNTERS + 1];
5038 
5039 	struct msr_counter_info_t *mci;
5040 
5041 	if (debug >= 2)
5042 		fprintf(stderr, "%s: cpu%d\n", __func__, cpu);
5043 
5044 	assert(msr_counter_info);
5045 	assert(cpu <= msr_counter_info_size);
5046 
5047 	mci = &msr_counter_info[cpu];
5048 
5049 	ZERO_ARRAY(perf_data);
5050 	ZERO_ARRAY(mci->data);
5051 
5052 	if (mci->fd_perf != -1) {
5053 		const size_t num_perf_counters = msr_counter_info_count_perf(mci);
5054 		const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
5055 		const ssize_t actual_read_size = read(mci->fd_perf, &perf_data[0], sizeof(perf_data));
5056 
5057 		if (actual_read_size != expected_read_size)
5058 			err(-1, "%s: failed to read perf_data (%zu %zu)", __func__, expected_read_size, actual_read_size);
5059 	}
5060 
5061 	for (unsigned int i = 0, pi = 1; i < NUM_MSR_COUNTERS; ++i) {
5062 		switch (mci->source[i]) {
5063 		case COUNTER_SOURCE_NONE:
5064 			break;
5065 
5066 		case COUNTER_SOURCE_PERF:
5067 			assert(pi < ARRAY_SIZE(perf_data));
5068 			assert(mci->fd_perf != -1);
5069 
5070 			if (debug >= 2)
5071 				fprintf(stderr, "Reading msr counter via perf at %u: %llu\n", i, perf_data[pi]);
5072 
5073 			mci->data[i] = perf_data[pi];
5074 
5075 			++pi;
5076 			break;
5077 
5078 		case COUNTER_SOURCE_MSR:
5079 			assert(!no_msr);
5080 
5081 			if (get_msr(cpu, mci->msr[i], &mci->data[i]))
5082 				return -2 - i;
5083 
5084 			mci->data[i] &= mci->msr_mask[i];
5085 
5086 			if (debug >= 2)
5087 				fprintf(stderr, "Reading msr counter via msr at %u: %llu\n", i, mci->data[i]);
5088 
5089 			break;
5090 		}
5091 	}
5092 
5093 	BUILD_BUG_ON(NUM_MSR_COUNTERS != 3);
5094 	t->aperf = mci->data[MSR_RCI_INDEX_APERF];
5095 	t->mperf = mci->data[MSR_RCI_INDEX_MPERF];
5096 	t->smi_count = mci->data[MSR_RCI_INDEX_SMI];
5097 
5098 	return 0;
5099 }
5100 
perf_counter_info_read_values(struct perf_counter_info * pp,int cpu,unsigned long long * out,size_t out_size)5101 int perf_counter_info_read_values(struct perf_counter_info *pp, int cpu, unsigned long long *out, size_t out_size)
5102 {
5103 	unsigned int domain;
5104 	unsigned long long value;
5105 	int fd_counter;
5106 
5107 	for (size_t i = 0; pp; ++i, pp = pp->next) {
5108 		domain = cpu_to_domain(pp, cpu);
5109 		assert(domain < pp->num_domains);
5110 
5111 		fd_counter = pp->fd_perf_per_domain[domain];
5112 
5113 		if (fd_counter == -1)
5114 			continue;
5115 
5116 		if (read(fd_counter, &value, sizeof(value)) != sizeof(value))
5117 			return 1;
5118 
5119 		assert(i < out_size);
5120 		out[i] = value * pp->scale;
5121 	}
5122 
5123 	return 0;
5124 }
5125 
pmt_gen_value_mask(unsigned int lsb,unsigned int msb)5126 unsigned long pmt_gen_value_mask(unsigned int lsb, unsigned int msb)
5127 {
5128 	unsigned long mask;
5129 
5130 	if (msb == 63)
5131 		mask = 0xffffffffffffffff;
5132 	else
5133 		mask = ((1 << (msb + 1)) - 1);
5134 
5135 	mask -= (1 << lsb) - 1;
5136 
5137 	return mask;
5138 }
5139 
pmt_read_counter(struct pmt_counter * ppmt,unsigned int domain_id)5140 unsigned long pmt_read_counter(struct pmt_counter *ppmt, unsigned int domain_id)
5141 {
5142 	if (domain_id >= ppmt->num_domains)
5143 		return 0;
5144 
5145 	const unsigned long *pmmio = ppmt->domains[domain_id].pcounter;
5146 	const unsigned long value = pmmio ? *pmmio : 0;
5147 	const unsigned long value_mask = pmt_gen_value_mask(ppmt->lsb, ppmt->msb);
5148 	const unsigned long value_shift = ppmt->lsb;
5149 
5150 	return (value & value_mask) >> value_shift;
5151 }
5152 
5153 /* Rapl domain enumeration helpers */
get_rapl_num_domains(void)5154 static inline int get_rapl_num_domains(void)
5155 {
5156 	if (!platform->has_per_core_rapl)
5157 		return topo.num_packages;
5158 
5159 	return topo.num_cores;
5160 }
5161 
get_rapl_domain_id(int cpu)5162 static inline int get_rapl_domain_id(int cpu)
5163 {
5164 	if (!platform->has_per_core_rapl)
5165 		return cpus[cpu].package_id;
5166 
5167 	return GLOBAL_CORE_ID(cpus[cpu].core_id, cpus[cpu].package_id);
5168 }
5169 
5170 /*
5171  * get_counters(...)
5172  * migrate to cpu
5173  * acquire and record local counters for that cpu
5174  */
get_counters(PER_THREAD_PARAMS)5175 int get_counters(PER_THREAD_PARAMS)
5176 {
5177 	int cpu = t->cpu_id;
5178 	unsigned long long msr;
5179 	struct msr_counter *mp;
5180 	struct pmt_counter *pp;
5181 	int i;
5182 	int status;
5183 
5184 	if (cpu_migrate(cpu)) {
5185 		fprintf(outf, "%s: Could not migrate to CPU %d\n", __func__, cpu);
5186 		return -1;
5187 	}
5188 
5189 	gettimeofday(&t->tv_begin, (struct timezone *)NULL);
5190 
5191 	if (first_counter_read)
5192 		get_apic_id(t);
5193 
5194 	t->tsc = rdtsc();	/* we are running on local CPU of interest */
5195 
5196 	get_smi_aperf_mperf(cpu, t);
5197 
5198 	if (DO_BIC(BIC_LLC_MRPS) || DO_BIC(BIC_LLC_HIT))
5199 		get_perf_llc_stats(cpu, &t->llc);
5200 
5201 	if (DO_BIC(BIC_L2_MRPS) || DO_BIC(BIC_L2_HIT))
5202 		get_perf_l2_stats(cpu, &t->l2);
5203 
5204 	if (DO_BIC(BIC_IPC))
5205 		if (read(get_instr_count_fd(cpu), &t->instr_count, sizeof(long long)) != sizeof(long long))
5206 			return -4;
5207 
5208 	if (DO_BIC(BIC_IRQ))
5209 		t->irq_count = irqs_per_cpu[cpu];
5210 	if (DO_BIC(BIC_NMI))
5211 		t->nmi_count = nmi_per_cpu[cpu];
5212 
5213 	get_cstate_counters(cpu, t, c, p);
5214 
5215 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
5216 		if (get_mp(cpu, mp, &t->counter[i], mp->sp->path))
5217 			return -10;
5218 	}
5219 
5220 	if (perf_counter_info_read_values(sys.perf_tp, cpu, t->perf_counter, MAX_ADDED_THREAD_COUNTERS))
5221 		return -10;
5222 
5223 	for (i = 0, pp = sys.pmt_tp; pp; i++, pp = pp->next)
5224 		t->pmt_counter[i] = pmt_read_counter(pp, t->cpu_id);
5225 
5226 	/* collect core counters only for 1st thread in core */
5227 	if (!is_cpu_first_thread_in_core(t, c))
5228 		goto done;
5229 
5230 	if (platform->has_per_core_rapl) {
5231 		status = get_rapl_counters(cpu, get_rapl_domain_id(cpu), c, p);
5232 		if (status != 0)
5233 			return status;
5234 	}
5235 
5236 	if (DO_BIC(BIC_CPU_c7) && t->is_atom) {
5237 		/*
5238 		 * For Atom CPUs that has core cstate deeper than c6,
5239 		 * MSR_CORE_C6_RESIDENCY returns residency of cc6 and deeper.
5240 		 * Minus CC7 (and deeper cstates) residency to get
5241 		 * accturate cc6 residency.
5242 		 */
5243 		c->c6 -= c->c7;
5244 	}
5245 
5246 	if (DO_BIC(BIC_Mod_c6))
5247 		if (get_msr(cpu, MSR_MODULE_C6_RES_MS, &c->mc6_us))
5248 			return -8;
5249 
5250 	if (DO_BIC(BIC_CoreTmp)) {
5251 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
5252 			return -9;
5253 		c->core_temp_c = tj_max - ((msr >> 16) & 0x7F);
5254 	}
5255 
5256 	if (DO_BIC(BIC_CORE_THROT_CNT))
5257 		get_core_throt_cnt(cpu, &c->core_throt_cnt);
5258 
5259 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
5260 		if (get_mp(cpu, mp, &c->counter[i], mp->sp->path))
5261 			return -10;
5262 	}
5263 
5264 	if (perf_counter_info_read_values(sys.perf_cp, cpu, c->perf_counter, MAX_ADDED_CORE_COUNTERS))
5265 		return -10;
5266 
5267 	for (i = 0, pp = sys.pmt_cp; pp; i++, pp = pp->next)
5268 		c->pmt_counter[i] = pmt_read_counter(pp, cpus[t->cpu_id].core_id);
5269 
5270 	/* collect package counters only for 1st core in package */
5271 	if (!is_cpu_first_core_in_package(t, p))
5272 		goto done;
5273 
5274 	if (DO_BIC(BIC_Totl_c0)) {
5275 		if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
5276 			return -10;
5277 	}
5278 	if (DO_BIC(BIC_Any_c0)) {
5279 		if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
5280 			return -11;
5281 	}
5282 	if (DO_BIC(BIC_GFX_c0)) {
5283 		if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
5284 			return -12;
5285 	}
5286 	if (DO_BIC(BIC_CPUGFX)) {
5287 		if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
5288 			return -13;
5289 	}
5290 
5291 	if (DO_BIC(BIC_CPU_LPI))
5292 		p->cpu_lpi = cpuidle_cur_cpu_lpi_us;
5293 	if (DO_BIC(BIC_SYS_LPI))
5294 		p->sys_lpi = cpuidle_cur_sys_lpi_us;
5295 
5296 	if (!platform->has_per_core_rapl) {
5297 		status = get_rapl_counters(cpu, get_rapl_domain_id(cpu), c, p);
5298 		if (status != 0)
5299 			return status;
5300 	}
5301 
5302 	if (DO_BIC(BIC_PkgTmp)) {
5303 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
5304 			return -17;
5305 		p->pkg_temp_c = tj_max - ((msr >> 16) & 0x7F);
5306 	}
5307 
5308 	if (DO_BIC(BIC_UNCORE_MHZ))
5309 		p->uncore_mhz = get_legacy_uncore_mhz(cpus[t->cpu_id].package_id);
5310 
5311 	if (DO_BIC(BIC_GFX_rc6))
5312 		p->gfx_rc6_ms = gfx_info[GFX_rc6].val_ull;
5313 
5314 	if (DO_BIC(BIC_GFXMHz))
5315 		p->gfx_mhz = gfx_info[GFX_MHz].val;
5316 
5317 	if (DO_BIC(BIC_GFXACTMHz))
5318 		p->gfx_act_mhz = gfx_info[GFX_ACTMHz].val;
5319 
5320 	if (DO_BIC(BIC_SAM_mc6))
5321 		p->sam_mc6_ms = gfx_info[SAM_mc6].val_ull;
5322 
5323 	if (DO_BIC(BIC_SAMMHz))
5324 		p->sam_mhz = gfx_info[SAM_MHz].val;
5325 
5326 	if (DO_BIC(BIC_SAMACTMHz))
5327 		p->sam_act_mhz = gfx_info[SAM_ACTMHz].val;
5328 
5329 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
5330 		char *path = NULL;
5331 
5332 		if (mp->msr_num == 0) {
5333 			path = find_sysfs_path_by_id(mp->sp, cpus[t->cpu_id].package_id);
5334 			if (path == NULL) {
5335 				warnx("%s: package_id %d not found", __func__, cpus[t->cpu_id].package_id);
5336 				return -10;
5337 			}
5338 		}
5339 		if (get_mp(cpu, mp, &p->counter[i], path))
5340 			return -10;
5341 	}
5342 
5343 	if (perf_counter_info_read_values(sys.perf_pp, cpu, p->perf_counter, MAX_ADDED_PACKAGE_COUNTERS))
5344 		return -10;
5345 
5346 	for (i = 0, pp = sys.pmt_pp; pp; i++, pp = pp->next)
5347 		p->pmt_counter[i] = pmt_read_counter(pp, cpus[t->cpu_id].package_id);
5348 
5349 done:
5350 	gettimeofday(&t->tv_end, (struct timezone *)NULL);
5351 
5352 	return 0;
5353 }
5354 
5355 int pkg_cstate_limit = PCLUKN;
5356 char *pkg_cstate_limit_strings[] = { "unknown", "reserved", "pc0", "pc1", "pc2",
5357 	"pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "pc10", "unlimited"
5358 };
5359 
5360 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,
5361 	PCLRSV, PCLRSV
5362 };
5363 
5364 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,
5365 	PCLRSV, PCLRSV
5366 };
5367 
5368 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,
5369 	PCLRSV, PCLRSV
5370 };
5371 
5372 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,
5373 	PCL__6, PCL__7
5374 };
5375 
5376 int amt_pkg_cstate_limits[16] = { PCLUNL, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5377 	PCLRSV, PCLRSV
5378 };
5379 
5380 int phi_pkg_cstate_limits[16] = { PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5381 	PCLRSV, PCLRSV
5382 };
5383 
5384 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,
5385 	PCLRSV, PCLRSV
5386 };
5387 
5388 int skx_pkg_cstate_limits[16] = { PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5389 	PCLRSV, PCLRSV
5390 };
5391 
5392 int icx_pkg_cstate_limits[16] = { PCL__0, PCL__2, PCL__6, PCL__6, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV,
5393 	PCLRSV, PCLRSV
5394 };
5395 
probe_cst_limit(void)5396 void probe_cst_limit(void)
5397 {
5398 	unsigned long long msr;
5399 	int *pkg_cstate_limits;
5400 
5401 	if (!platform->has_nhm_msrs || no_msr)
5402 		return;
5403 
5404 	switch (platform->cst_limit) {
5405 	case CST_LIMIT_NHM:
5406 		pkg_cstate_limits = nhm_pkg_cstate_limits;
5407 		break;
5408 	case CST_LIMIT_SNB:
5409 		pkg_cstate_limits = snb_pkg_cstate_limits;
5410 		break;
5411 	case CST_LIMIT_HSW:
5412 		pkg_cstate_limits = hsw_pkg_cstate_limits;
5413 		break;
5414 	case CST_LIMIT_SKX:
5415 		pkg_cstate_limits = skx_pkg_cstate_limits;
5416 		break;
5417 	case CST_LIMIT_ICX:
5418 		pkg_cstate_limits = icx_pkg_cstate_limits;
5419 		break;
5420 	case CST_LIMIT_SLV:
5421 		pkg_cstate_limits = slv_pkg_cstate_limits;
5422 		break;
5423 	case CST_LIMIT_AMT:
5424 		pkg_cstate_limits = amt_pkg_cstate_limits;
5425 		break;
5426 	case CST_LIMIT_KNL:
5427 		pkg_cstate_limits = phi_pkg_cstate_limits;
5428 		break;
5429 	case CST_LIMIT_GMT:
5430 		pkg_cstate_limits = glm_pkg_cstate_limits;
5431 		break;
5432 	default:
5433 		return;
5434 	}
5435 
5436 	get_msr(master_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
5437 	pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
5438 }
5439 
dump_platform_info(void)5440 static void dump_platform_info(void)
5441 {
5442 	unsigned long long msr;
5443 	unsigned int ratio;
5444 
5445 	if (!platform->has_nhm_msrs || no_msr)
5446 		return;
5447 
5448 	get_msr(master_cpu, MSR_PLATFORM_INFO, &msr);
5449 
5450 	fprintf(outf, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", master_cpu, msr);
5451 
5452 	ratio = (msr >> 40) & 0xFF;
5453 	fprintf(outf, "%d * %.1f = %.1f MHz max efficiency frequency\n", ratio, bclk, ratio * bclk);
5454 
5455 	ratio = (msr >> 8) & 0xFF;
5456 	fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n", ratio, bclk, ratio * bclk);
5457 }
5458 
dump_power_ctl(void)5459 static void dump_power_ctl(void)
5460 {
5461 	unsigned long long msr;
5462 
5463 	if (!platform->has_nhm_msrs || no_msr)
5464 		return;
5465 
5466 	get_msr(master_cpu, MSR_IA32_POWER_CTL, &msr);
5467 	fprintf(outf, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n", master_cpu, msr, msr & 0x2 ? "EN" : "DIS");
5468 
5469 	/* C-state Pre-wake Disable (CSTATE_PREWAKE_DISABLE) */
5470 	if (platform->has_cst_prewake_bit)
5471 		fprintf(outf, "C-state Pre-wake: %sabled\n", msr & 0x40000000 ? "DIS" : "EN");
5472 
5473 	return;
5474 }
5475 
dump_turbo_ratio_limit2(void)5476 static void dump_turbo_ratio_limit2(void)
5477 {
5478 	unsigned long long msr;
5479 	unsigned int ratio;
5480 
5481 	get_msr(master_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
5482 
5483 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", master_cpu, msr);
5484 
5485 	ratio = (msr >> 8) & 0xFF;
5486 	if (ratio)
5487 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 18 active cores\n", ratio, bclk, ratio * bclk);
5488 
5489 	ratio = (msr >> 0) & 0xFF;
5490 	if (ratio)
5491 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 17 active cores\n", ratio, bclk, ratio * bclk);
5492 	return;
5493 }
5494 
dump_turbo_ratio_limit1(void)5495 static void dump_turbo_ratio_limit1(void)
5496 {
5497 	unsigned long long msr;
5498 	unsigned int ratio;
5499 
5500 	get_msr(master_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
5501 
5502 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", master_cpu, msr);
5503 
5504 	ratio = (msr >> 56) & 0xFF;
5505 	if (ratio)
5506 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 16 active cores\n", ratio, bclk, ratio * bclk);
5507 
5508 	ratio = (msr >> 48) & 0xFF;
5509 	if (ratio)
5510 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 15 active cores\n", ratio, bclk, ratio * bclk);
5511 
5512 	ratio = (msr >> 40) & 0xFF;
5513 	if (ratio)
5514 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 14 active cores\n", ratio, bclk, ratio * bclk);
5515 
5516 	ratio = (msr >> 32) & 0xFF;
5517 	if (ratio)
5518 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 13 active cores\n", ratio, bclk, ratio * bclk);
5519 
5520 	ratio = (msr >> 24) & 0xFF;
5521 	if (ratio)
5522 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 12 active cores\n", ratio, bclk, ratio * bclk);
5523 
5524 	ratio = (msr >> 16) & 0xFF;
5525 	if (ratio)
5526 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 11 active cores\n", ratio, bclk, ratio * bclk);
5527 
5528 	ratio = (msr >> 8) & 0xFF;
5529 	if (ratio)
5530 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 10 active cores\n", ratio, bclk, ratio * bclk);
5531 
5532 	ratio = (msr >> 0) & 0xFF;
5533 	if (ratio)
5534 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 9 active cores\n", ratio, bclk, ratio * bclk);
5535 	return;
5536 }
5537 
dump_turbo_ratio_limits(int trl_msr_offset)5538 static void dump_turbo_ratio_limits(int trl_msr_offset)
5539 {
5540 	unsigned long long msr, core_counts;
5541 	int shift;
5542 
5543 	get_msr(master_cpu, trl_msr_offset, &msr);
5544 	fprintf(outf, "cpu%d: MSR_%sTURBO_RATIO_LIMIT: 0x%08llx\n", master_cpu, trl_msr_offset == MSR_SECONDARY_TURBO_RATIO_LIMIT ? "SECONDARY_" : "", msr);
5545 
5546 	if (platform->trl_msrs & TRL_CORECOUNT) {
5547 		get_msr(master_cpu, MSR_TURBO_RATIO_LIMIT1, &core_counts);
5548 		fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", master_cpu, core_counts);
5549 	} else {
5550 		core_counts = 0x0807060504030201;
5551 	}
5552 
5553 	for (shift = 56; shift >= 0; shift -= 8) {
5554 		unsigned int ratio, group_size;
5555 
5556 		ratio = (msr >> shift) & 0xFF;
5557 		group_size = (core_counts >> shift) & 0xFF;
5558 		if (ratio)
5559 			fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n", ratio, bclk, ratio * bclk, group_size);
5560 	}
5561 
5562 	return;
5563 }
5564 
dump_atom_turbo_ratio_limits(void)5565 static void dump_atom_turbo_ratio_limits(void)
5566 {
5567 	unsigned long long msr;
5568 	unsigned int ratio;
5569 
5570 	get_msr(master_cpu, MSR_ATOM_CORE_RATIOS, &msr);
5571 	fprintf(outf, "cpu%d: MSR_ATOM_CORE_RATIOS: 0x%08llx\n", master_cpu, msr & 0xFFFFFFFF);
5572 
5573 	ratio = (msr >> 0) & 0x3F;
5574 	if (ratio)
5575 		fprintf(outf, "%d * %.1f = %.1f MHz minimum operating frequency\n", ratio, bclk, ratio * bclk);
5576 
5577 	ratio = (msr >> 8) & 0x3F;
5578 	if (ratio)
5579 		fprintf(outf, "%d * %.1f = %.1f MHz low frequency mode (LFM)\n", ratio, bclk, ratio * bclk);
5580 
5581 	ratio = (msr >> 16) & 0x3F;
5582 	if (ratio)
5583 		fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n", ratio, bclk, ratio * bclk);
5584 
5585 	get_msr(master_cpu, MSR_ATOM_CORE_TURBO_RATIOS, &msr);
5586 	fprintf(outf, "cpu%d: MSR_ATOM_CORE_TURBO_RATIOS: 0x%08llx\n", master_cpu, msr & 0xFFFFFFFF);
5587 
5588 	ratio = (msr >> 24) & 0x3F;
5589 	if (ratio)
5590 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 4 active cores\n", ratio, bclk, ratio * bclk);
5591 
5592 	ratio = (msr >> 16) & 0x3F;
5593 	if (ratio)
5594 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 3 active cores\n", ratio, bclk, ratio * bclk);
5595 
5596 	ratio = (msr >> 8) & 0x3F;
5597 	if (ratio)
5598 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 2 active cores\n", ratio, bclk, ratio * bclk);
5599 
5600 	ratio = (msr >> 0) & 0x3F;
5601 	if (ratio)
5602 		fprintf(outf, "%d * %.1f = %.1f MHz max turbo 1 active core\n", ratio, bclk, ratio * bclk);
5603 }
5604 
dump_knl_turbo_ratio_limits(void)5605 static void dump_knl_turbo_ratio_limits(void)
5606 {
5607 	const unsigned int buckets_no = 7;
5608 
5609 	unsigned long long msr;
5610 	int delta_cores, delta_ratio;
5611 	int i, b_nr;
5612 	unsigned int cores[buckets_no];
5613 	unsigned int ratio[buckets_no];
5614 
5615 	get_msr(master_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
5616 
5617 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", master_cpu, msr);
5618 
5619 	/*
5620 	 * Turbo encoding in KNL is as follows:
5621 	 * [0] -- Reserved
5622 	 * [7:1] -- Base value of number of active cores of bucket 1.
5623 	 * [15:8] -- Base value of freq ratio of bucket 1.
5624 	 * [20:16] -- +ve delta of number of active cores of bucket 2.
5625 	 * i.e. active cores of bucket 2 =
5626 	 * active cores of bucket 1 + delta
5627 	 * [23:21] -- Negative delta of freq ratio of bucket 2.
5628 	 * i.e. freq ratio of bucket 2 =
5629 	 * freq ratio of bucket 1 - delta
5630 	 * [28:24]-- +ve delta of number of active cores of bucket 3.
5631 	 * [31:29]-- -ve delta of freq ratio of bucket 3.
5632 	 * [36:32]-- +ve delta of number of active cores of bucket 4.
5633 	 * [39:37]-- -ve delta of freq ratio of bucket 4.
5634 	 * [44:40]-- +ve delta of number of active cores of bucket 5.
5635 	 * [47:45]-- -ve delta of freq ratio of bucket 5.
5636 	 * [52:48]-- +ve delta of number of active cores of bucket 6.
5637 	 * [55:53]-- -ve delta of freq ratio of bucket 6.
5638 	 * [60:56]-- +ve delta of number of active cores of bucket 7.
5639 	 * [63:61]-- -ve delta of freq ratio of bucket 7.
5640 	 */
5641 
5642 	b_nr = 0;
5643 	cores[b_nr] = (msr & 0xFF) >> 1;
5644 	ratio[b_nr] = (msr >> 8) & 0xFF;
5645 
5646 	for (i = 16; i < 64; i += 8) {
5647 		delta_cores = (msr >> i) & 0x1F;
5648 		delta_ratio = (msr >> (i + 5)) & 0x7;
5649 
5650 		cores[b_nr + 1] = cores[b_nr] + delta_cores;
5651 		ratio[b_nr + 1] = ratio[b_nr] - delta_ratio;
5652 		b_nr++;
5653 	}
5654 
5655 	for (i = buckets_no - 1; i >= 0; i--)
5656 		if (i > 0 ? ratio[i] != ratio[i - 1] : 1)
5657 			fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n", ratio[i], bclk, ratio[i] * bclk, cores[i]);
5658 }
5659 
dump_cst_cfg(void)5660 static void dump_cst_cfg(void)
5661 {
5662 	unsigned long long msr;
5663 
5664 	if (!platform->has_nhm_msrs || no_msr)
5665 		return;
5666 
5667 	get_msr(master_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
5668 
5669 	fprintf(outf, "cpu%d: MSR_PKG_CST_CONFIG_CONTROL: 0x%08llx", master_cpu, msr);
5670 
5671 	fprintf(outf, " (%s%s%s%s%slocked, pkg-cstate-limit=%d (%s)",
5672 		(msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
5673 		(msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
5674 		(msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
5675 		(msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
5676 		(msr & (1 << 15)) ? "" : "UN", (unsigned int)msr & 0xF, pkg_cstate_limit_strings[pkg_cstate_limit]);
5677 
5678 #define AUTOMATIC_CSTATE_CONVERSION		(1UL << 16)
5679 	if (platform->has_cst_auto_convension) {
5680 		fprintf(outf, ", automatic c-state conversion=%s", (msr & AUTOMATIC_CSTATE_CONVERSION) ? "on" : "off");
5681 	}
5682 
5683 	fprintf(outf, ")\n");
5684 
5685 	return;
5686 }
5687 
dump_config_tdp(void)5688 static void dump_config_tdp(void)
5689 {
5690 	unsigned long long msr;
5691 
5692 	get_msr(master_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
5693 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", master_cpu, msr);
5694 	fprintf(outf, " (base_ratio=%d)\n", (unsigned int)msr & 0xFF);
5695 
5696 	get_msr(master_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
5697 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", master_cpu, msr);
5698 	if (msr) {
5699 		fprintf(outf, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
5700 		fprintf(outf, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
5701 		fprintf(outf, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
5702 		fprintf(outf, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0x7FFF);
5703 	}
5704 	fprintf(outf, ")\n");
5705 
5706 	get_msr(master_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
5707 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", master_cpu, msr);
5708 	if (msr) {
5709 		fprintf(outf, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
5710 		fprintf(outf, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
5711 		fprintf(outf, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
5712 		fprintf(outf, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0x7FFF);
5713 	}
5714 	fprintf(outf, ")\n");
5715 
5716 	get_msr(master_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
5717 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", master_cpu, msr);
5718 	if ((msr) & 0x3)
5719 		fprintf(outf, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
5720 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
5721 	fprintf(outf, ")\n");
5722 
5723 	get_msr(master_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
5724 	fprintf(outf, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", master_cpu, msr);
5725 	fprintf(outf, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xFF);
5726 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
5727 	fprintf(outf, ")\n");
5728 }
5729 
5730 unsigned int irtl_time_units[] = { 1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
5731 
print_irtl(void)5732 void print_irtl(void)
5733 {
5734 	unsigned long long msr;
5735 
5736 	if (!platform->has_irtl_msrs || no_msr)
5737 		return;
5738 
5739 	if (platform->supported_cstates & PC3) {
5740 		get_msr(master_cpu, MSR_PKGC3_IRTL, &msr);
5741 		fprintf(outf, "cpu%d: MSR_PKGC3_IRTL: 0x%08llx (", master_cpu, msr);
5742 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5743 	}
5744 
5745 	if (platform->supported_cstates & PC6) {
5746 		get_msr(master_cpu, MSR_PKGC6_IRTL, &msr);
5747 		fprintf(outf, "cpu%d: MSR_PKGC6_IRTL: 0x%08llx (", master_cpu, msr);
5748 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5749 	}
5750 
5751 	if (platform->supported_cstates & PC7) {
5752 		get_msr(master_cpu, MSR_PKGC7_IRTL, &msr);
5753 		fprintf(outf, "cpu%d: MSR_PKGC7_IRTL: 0x%08llx (", master_cpu, msr);
5754 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5755 	}
5756 
5757 	if (platform->supported_cstates & PC8) {
5758 		get_msr(master_cpu, MSR_PKGC8_IRTL, &msr);
5759 		fprintf(outf, "cpu%d: MSR_PKGC8_IRTL: 0x%08llx (", master_cpu, msr);
5760 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5761 	}
5762 
5763 	if (platform->supported_cstates & PC9) {
5764 		get_msr(master_cpu, MSR_PKGC9_IRTL, &msr);
5765 		fprintf(outf, "cpu%d: MSR_PKGC9_IRTL: 0x%08llx (", master_cpu, msr);
5766 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5767 	}
5768 
5769 	if (platform->supported_cstates & PC10) {
5770 		get_msr(master_cpu, MSR_PKGC10_IRTL, &msr);
5771 		fprintf(outf, "cpu%d: MSR_PKGC10_IRTL: 0x%08llx (", master_cpu, msr);
5772 		fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT", (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
5773 	}
5774 }
5775 
free_fd_percpu(void)5776 void free_fd_percpu(void)
5777 {
5778 	int i;
5779 
5780 	if (!fd_percpu)
5781 		return;
5782 
5783 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
5784 		if (fd_percpu[i] != 0)
5785 			close(fd_percpu[i]);
5786 	}
5787 
5788 	free(fd_percpu);
5789 	fd_percpu = NULL;
5790 }
5791 
free_fd_instr_count_percpu(void)5792 void free_fd_instr_count_percpu(void)
5793 {
5794 	if (!fd_instr_count_percpu)
5795 		return;
5796 
5797 	for (int i = 0; i < topo.max_cpu_num + 1; ++i) {
5798 		if (fd_instr_count_percpu[i] != 0)
5799 			close(fd_instr_count_percpu[i]);
5800 	}
5801 
5802 	free(fd_instr_count_percpu);
5803 	fd_instr_count_percpu = NULL;
5804 }
5805 
free_fd_llc_percpu(void)5806 void free_fd_llc_percpu(void)
5807 {
5808 	if (!fd_llc_percpu)
5809 		return;
5810 
5811 	for (int i = 0; i < topo.max_cpu_num + 1; ++i) {
5812 		if (fd_llc_percpu[i] != 0)
5813 			close(fd_llc_percpu[i]);
5814 	}
5815 
5816 	free(fd_llc_percpu);
5817 	fd_llc_percpu = NULL;
5818 
5819 	BIC_NOT_PRESENT(BIC_LLC_MRPS);
5820 	BIC_NOT_PRESENT(BIC_LLC_HIT);
5821 }
5822 
free_fd_l2_percpu(void)5823 void free_fd_l2_percpu(void)
5824 {
5825 	if (!fd_l2_percpu)
5826 		return;
5827 
5828 	for (int i = 0; i < topo.max_cpu_num + 1; ++i) {
5829 		if (fd_l2_percpu[i] != 0)
5830 			close(fd_l2_percpu[i]);
5831 	}
5832 
5833 	free(fd_l2_percpu);
5834 	fd_l2_percpu = NULL;
5835 
5836 	BIC_NOT_PRESENT(BIC_L2_MRPS);
5837 	BIC_NOT_PRESENT(BIC_L2_HIT);
5838 }
5839 
free_fd_cstate(void)5840 void free_fd_cstate(void)
5841 {
5842 	if (!ccstate_counter_info)
5843 		return;
5844 
5845 	const int counter_info_num = ccstate_counter_info_size;
5846 
5847 	for (int counter_id = 0; counter_id < counter_info_num; ++counter_id) {
5848 		if (ccstate_counter_info[counter_id].fd_perf_core != -1)
5849 			close(ccstate_counter_info[counter_id].fd_perf_core);
5850 
5851 		if (ccstate_counter_info[counter_id].fd_perf_pkg != -1)
5852 			close(ccstate_counter_info[counter_id].fd_perf_pkg);
5853 	}
5854 
5855 	free(ccstate_counter_info);
5856 	ccstate_counter_info = NULL;
5857 	ccstate_counter_info_size = 0;
5858 }
5859 
free_fd_msr(void)5860 void free_fd_msr(void)
5861 {
5862 	if (!msr_counter_info)
5863 		return;
5864 
5865 	for (int cpu = 0; cpu < topo.max_cpu_num; ++cpu) {
5866 		if (msr_counter_info[cpu].fd_perf != -1)
5867 			close(msr_counter_info[cpu].fd_perf);
5868 	}
5869 
5870 	free(msr_counter_info);
5871 	msr_counter_info = NULL;
5872 	msr_counter_info_size = 0;
5873 }
5874 
free_fd_rapl_percpu(void)5875 void free_fd_rapl_percpu(void)
5876 {
5877 	if (!rapl_counter_info_perdomain)
5878 		return;
5879 
5880 	const int num_domains = rapl_counter_info_perdomain_size;
5881 
5882 	for (int domain_id = 0; domain_id < num_domains; ++domain_id) {
5883 		if (rapl_counter_info_perdomain[domain_id].fd_perf != -1)
5884 			close(rapl_counter_info_perdomain[domain_id].fd_perf);
5885 	}
5886 
5887 	free(rapl_counter_info_perdomain);
5888 	rapl_counter_info_perdomain = NULL;
5889 	rapl_counter_info_perdomain_size = 0;
5890 }
5891 
free_fd_added_perf_counters_(struct perf_counter_info * pp)5892 void free_fd_added_perf_counters_(struct perf_counter_info *pp)
5893 {
5894 	if (!pp)
5895 		return;
5896 
5897 	if (!pp->fd_perf_per_domain)
5898 		return;
5899 
5900 	while (pp) {
5901 		for (size_t domain = 0; domain < pp->num_domains; ++domain) {
5902 			if (pp->fd_perf_per_domain[domain] != -1) {
5903 				close(pp->fd_perf_per_domain[domain]);
5904 				pp->fd_perf_per_domain[domain] = -1;
5905 			}
5906 		}
5907 
5908 		free(pp->fd_perf_per_domain);
5909 		pp->fd_perf_per_domain = NULL;
5910 
5911 		pp = pp->next;
5912 	}
5913 }
5914 
free_fd_added_perf_counters(void)5915 void free_fd_added_perf_counters(void)
5916 {
5917 	free_fd_added_perf_counters_(sys.perf_tp);
5918 	free_fd_added_perf_counters_(sys.perf_cp);
5919 	free_fd_added_perf_counters_(sys.perf_pp);
5920 }
5921 
free_all_buffers(void)5922 void free_all_buffers(void)
5923 {
5924 	int i;
5925 
5926 	CPU_FREE(cpu_present_set);
5927 	cpu_present_set = NULL;
5928 	cpu_present_setsize = 0;
5929 
5930 	CPU_FREE(cpu_effective_set);
5931 	cpu_effective_set = NULL;
5932 	cpu_effective_setsize = 0;
5933 
5934 	CPU_FREE(cpu_allowed_set);
5935 	cpu_allowed_set = NULL;
5936 	cpu_allowed_setsize = 0;
5937 
5938 	CPU_FREE(cpu_affinity_set);
5939 	cpu_affinity_set = NULL;
5940 	cpu_affinity_setsize = 0;
5941 
5942 	if (perf_pcore_set) {
5943 		CPU_FREE(perf_pcore_set);
5944 		perf_pcore_set = NULL;
5945 	}
5946 
5947 	if (perf_ecore_set) {
5948 		CPU_FREE(perf_ecore_set);
5949 		perf_ecore_set = NULL;
5950 	}
5951 
5952 	if (perf_lcore_set) {
5953 		CPU_FREE(perf_lcore_set);
5954 		perf_lcore_set = NULL;
5955 	}
5956 
5957 	free(even.threads);
5958 	free(even.cores);
5959 	free(even.packages);
5960 
5961 	even.threads = NULL;
5962 	even.cores = NULL;
5963 	even.packages = NULL;
5964 
5965 	free(odd.threads);
5966 	free(odd.cores);
5967 	free(odd.packages);
5968 
5969 	odd.threads = NULL;
5970 	odd.cores = NULL;
5971 	odd.packages = NULL;
5972 
5973 	free(output_buffer);
5974 	output_buffer = NULL;
5975 	outp = NULL;
5976 
5977 	free_fd_percpu();
5978 	free_fd_instr_count_percpu();
5979 	free_fd_llc_percpu();
5980 	free_fd_l2_percpu();
5981 	free_fd_msr();
5982 	free_fd_rapl_percpu();
5983 	free_fd_cstate();
5984 	free_fd_added_perf_counters();
5985 
5986 	free(irq_column_2_cpu);
5987 	free(irqs_per_cpu);
5988 	free(nmi_per_cpu);
5989 
5990 	for (i = 0; i <= topo.max_cpu_num; ++i) {
5991 		if (cpus[i].put_ids)
5992 			CPU_FREE(cpus[i].put_ids);
5993 	}
5994 	free(cpus);
5995 }
5996 
5997 /*
5998  * Parse a file containing a single int.
5999  * Return 0 if file can not be opened
6000  * Exit if file can be opened, but can not be parsed
6001  */
parse_int_file(const char * fmt,...)6002 int parse_int_file(const char *fmt, ...)
6003 {
6004 	va_list args;
6005 	char path[PATH_MAX];
6006 	FILE *filep;
6007 	int value;
6008 
6009 	va_start(args, fmt);
6010 	vsnprintf(path, sizeof(path), fmt, args);
6011 	va_end(args);
6012 	filep = fopen(path, "r");
6013 	if (!filep)
6014 		return 0;
6015 	if (fscanf(filep, "%d", &value) != 1)
6016 		err(1, "%s: failed to parse number from file", path);
6017 	fclose(filep);
6018 	return value;
6019 }
6020 
6021 /*
6022  * cpu_is_first_core_in_package(cpu)
6023  * return 1 if given CPU is 1st core in package
6024  */
cpu_is_first_core_in_package(int cpu)6025 int cpu_is_first_core_in_package(int cpu)
6026 {
6027 	return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
6028 }
6029 
get_package_id(int cpu)6030 int get_package_id(int cpu)
6031 {
6032 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
6033 }
6034 
get_die_id(int cpu)6035 int get_die_id(int cpu)
6036 {
6037 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/die_id", cpu);
6038 }
6039 
get_l3_id(int cpu)6040 int get_l3_id(int cpu)
6041 {
6042 	return parse_int_file("/sys/devices/system/cpu/cpu%d/cache/index3/id", cpu);
6043 }
6044 
get_core_id(int cpu)6045 int get_core_id(int cpu)
6046 {
6047 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
6048 }
6049 
set_node_data(void)6050 void set_node_data(void)
6051 {
6052 	int pkg, node, lnode, cpu, cpux;
6053 	int cpu_count;
6054 
6055 	/* initialize logical_node_id */
6056 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu)
6057 		cpus[cpu].logical_node_id = -1;
6058 
6059 	cpu_count = 0;
6060 	for (pkg = 0; pkg < topo.num_packages; pkg++) {
6061 		lnode = 0;
6062 		for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
6063 			if (cpus[cpu].package_id != pkg)
6064 				continue;
6065 			/* find a cpu with an unset logical_node_id */
6066 			if (cpus[cpu].logical_node_id != -1)
6067 				continue;
6068 			cpus[cpu].logical_node_id = lnode;
6069 			node = cpus[cpu].physical_node_id;
6070 			cpu_count++;
6071 			/*
6072 			 * find all matching cpus on this pkg and set
6073 			 * the logical_node_id
6074 			 */
6075 			for (cpux = cpu; cpux <= topo.max_cpu_num; cpux++) {
6076 				if ((cpus[cpux].package_id == pkg) && (cpus[cpux].physical_node_id == node)) {
6077 					cpus[cpux].logical_node_id = lnode;
6078 					cpu_count++;
6079 				}
6080 			}
6081 			lnode++;
6082 			if (lnode > topo.nodes_per_pkg)
6083 				topo.nodes_per_pkg = lnode;
6084 		}
6085 		if (cpu_count >= topo.max_cpu_num)
6086 			break;
6087 	}
6088 }
6089 
get_physical_node_id(struct cpu_topology * thiscpu)6090 int get_physical_node_id(struct cpu_topology *thiscpu)
6091 {
6092 	char path[80];
6093 	FILE *filep;
6094 	int i;
6095 	int cpu = thiscpu->cpu_id;
6096 
6097 	for (i = 0; i <= topo.max_cpu_num; i++) {
6098 		sprintf(path, "/sys/devices/system/cpu/cpu%d/node%i/cpulist", cpu, i);
6099 		filep = fopen(path, "r");
6100 		if (!filep)
6101 			continue;
6102 		fclose(filep);
6103 		return i;
6104 	}
6105 	return -1;
6106 }
6107 
parse_cpu_str(char * cpu_str,cpu_set_t * cpu_set,int cpu_set_size)6108 static int parse_cpu_str(char *cpu_str, cpu_set_t *cpu_set, int cpu_set_size)
6109 {
6110 	unsigned int start, end;
6111 	char *next = cpu_str;
6112 
6113 	while (next && *next) {
6114 
6115 		if (*next == '-')	/* no negative cpu numbers */
6116 			return 1;
6117 
6118 		if (*next == '\0' || *next == '\n')
6119 			break;
6120 
6121 		start = strtoul(next, &next, 10);
6122 
6123 		if (start >= CPU_SUBSET_MAXCPUS)
6124 			return 1;
6125 		CPU_SET_S(start, cpu_set_size, cpu_set);
6126 
6127 		if (*next == '\0' || *next == '\n')
6128 			break;
6129 
6130 		if (*next == ',') {
6131 			next += 1;
6132 			continue;
6133 		}
6134 
6135 		if (*next == '-') {
6136 			next += 1;	/* start range */
6137 		} else if (*next == '.') {
6138 			next += 1;
6139 			if (*next == '.')
6140 				next += 1;	/* start range */
6141 			else
6142 				return 1;
6143 		}
6144 
6145 		end = strtoul(next, &next, 10);
6146 		if (end <= start)
6147 			return 1;
6148 
6149 		while (++start <= end) {
6150 			if (start >= CPU_SUBSET_MAXCPUS)
6151 				return 1;
6152 			CPU_SET_S(start, cpu_set_size, cpu_set);
6153 		}
6154 
6155 		if (*next == ',')
6156 			next += 1;
6157 		else if (*next != '\0' && *next != '\n')
6158 			return 1;
6159 	}
6160 
6161 	return 0;
6162 }
6163 
set_thread_siblings(struct cpu_topology * thiscpu)6164 int set_thread_siblings(struct cpu_topology *thiscpu)
6165 {
6166 	char path[80], character;
6167 	FILE *filep;
6168 	unsigned long map;
6169 	int so, shift, sib_core;
6170 	int cpu = thiscpu->cpu_id;
6171 	int offset = topo.max_cpu_num + 1;
6172 	size_t size;
6173 	int thread_id = 0;
6174 
6175 	thiscpu->put_ids = CPU_ALLOC((topo.max_cpu_num + 1));
6176 	if (thiscpu->ht_id < 0)
6177 		thiscpu->ht_id = thread_id++;
6178 	if (!thiscpu->put_ids)
6179 		return -1;
6180 
6181 	size = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
6182 	CPU_ZERO_S(size, thiscpu->put_ids);
6183 
6184 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", cpu);
6185 	filep = fopen(path, "r");
6186 
6187 	if (!filep) {
6188 		warnx("%s: open failed", path);
6189 		return -1;
6190 	}
6191 	do {
6192 		offset -= BITMASK_SIZE;
6193 		if (fscanf(filep, "%lx%c", &map, &character) != 2)
6194 			err(1, "%s: failed to parse file", path);
6195 		for (shift = 0; shift < BITMASK_SIZE; shift++) {
6196 			if ((map >> shift) & 0x1) {
6197 				so = shift + offset;
6198 				sib_core = get_core_id(so);
6199 				if (sib_core == thiscpu->core_id) {
6200 					CPU_SET_S(so, size, thiscpu->put_ids);
6201 					if ((so != cpu) && (cpus[so].ht_id < 0)) {
6202 						cpus[so].ht_id = thread_id;
6203 						cpus[cpu].ht_sibling_cpu_id[thread_id] = so;
6204 						if (debug)
6205 							fprintf(stderr, "%s: cpu%d.ht_sibling_cpu_id[%d] = %d\n", __func__, cpu, thread_id, so);
6206 						thread_id += 1;
6207 					}
6208 				}
6209 			}
6210 		}
6211 	} while (character == ',');
6212 	fclose(filep);
6213 
6214 	return CPU_COUNT_S(size, thiscpu->put_ids);
6215 }
6216 
6217 /*
6218  * run func(thread, core, package) in topology order
6219  * skip non-present cpus
6220  */
6221 
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)6222 int for_all_cpus_2(int (func) (struct thread_data *, struct core_data *,
6223 			       struct pkg_data *, struct thread_data *, struct core_data *,
6224 			       struct pkg_data *), struct thread_data *thread_base,
6225 		   struct core_data *core_base, struct pkg_data *pkg_base,
6226 		   struct thread_data *thread_base2, struct core_data *core_base2, struct pkg_data *pkg_base2)
6227 {
6228 	int cpu, retval;
6229 
6230 	retval = 0;
6231 
6232 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
6233 		struct thread_data *t, *t2;
6234 		struct core_data *c, *c2;
6235 		struct pkg_data *p, *p2;
6236 
6237 		if (cpu_is_not_allowed(cpu))
6238 			continue;
6239 
6240 		if (cpus[cpu].ht_id > 0)	/* skip HT sibling */
6241 			continue;
6242 
6243 		t = &thread_base[cpu];
6244 		t2 = &thread_base2[cpu];
6245 		c = &core_base[GLOBAL_CORE_ID(cpus[cpu].core_id, cpus[cpu].package_id)];
6246 		c2 = &core_base2[GLOBAL_CORE_ID(cpus[cpu].core_id, cpus[cpu].package_id)];
6247 		p = &pkg_base[cpus[cpu].package_id];
6248 		p2 = &pkg_base2[cpus[cpu].package_id];
6249 
6250 		retval |= func(t, c, p, t2, c2, p2);
6251 
6252 		/* Handle HT sibling now */
6253 		int i;
6254 
6255 		for (i = MAX_HT_ID; i > 0; --i) {	/* ht_id 0 is self */
6256 			if (cpus[cpu].ht_sibling_cpu_id[i] <= 0)
6257 				continue;
6258 			t = &thread_base[cpus[cpu].ht_sibling_cpu_id[i]];
6259 			t2 = &thread_base2[cpus[cpu].ht_sibling_cpu_id[i]];
6260 
6261 			retval |= func(t, c, p, t2, c2, p2);
6262 		}
6263 	}
6264 	return retval;
6265 }
6266 
6267 /*
6268  * run func(cpu) on every cpu in /proc/stat
6269  * return max_cpu number
6270  */
for_all_proc_cpus(int (func)(int))6271 int for_all_proc_cpus(int (func) (int))
6272 {
6273 	FILE *fp;
6274 	int cpu_num;
6275 	int retval;
6276 
6277 	fp = fopen_or_die(proc_stat, "r");
6278 
6279 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
6280 	if (retval != 0)
6281 		err(1, "%s: failed to parse format", proc_stat);
6282 
6283 	while (1) {
6284 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
6285 		if (retval != 1)
6286 			break;
6287 
6288 		retval = func(cpu_num);
6289 		if (retval) {
6290 			fclose(fp);
6291 			return (retval);
6292 		}
6293 	}
6294 	fclose(fp);
6295 	return 0;
6296 }
6297 
6298 #define PATH_EFFECTIVE_CPUS	"/sys/fs/cgroup/cpuset.cpus.effective"
6299 
6300 static char cpu_effective_str[1024];
6301 
update_effective_str(bool startup)6302 static int update_effective_str(bool startup)
6303 {
6304 	FILE *fp;
6305 	char *pos;
6306 	char buf[1024];
6307 	int ret;
6308 
6309 	if (cpu_effective_str[0] == '\0' && !startup)
6310 		return 0;
6311 
6312 	fp = fopen(PATH_EFFECTIVE_CPUS, "r");
6313 	if (!fp)
6314 		return 0;
6315 
6316 	pos = fgets(buf, 1024, fp);
6317 	if (!pos)
6318 		err(1, "%s: file read failed", PATH_EFFECTIVE_CPUS);
6319 
6320 	fclose(fp);
6321 
6322 	ret = strncmp(cpu_effective_str, buf, 1024);
6323 	if (!ret)
6324 		return 0;
6325 
6326 	strncpy(cpu_effective_str, buf, 1024);
6327 	return 1;
6328 }
6329 
update_effective_set(bool startup)6330 static void update_effective_set(bool startup)
6331 {
6332 	update_effective_str(startup);
6333 
6334 	if (parse_cpu_str(cpu_effective_str, cpu_effective_set, cpu_effective_setsize))
6335 		err(1, "%s: cpu str malformat %s", PATH_EFFECTIVE_CPUS, cpu_effective_str);
6336 }
6337 
6338 void linux_perf_init(void);
6339 void msr_perf_init(void);
6340 void rapl_perf_init(void);
6341 void cstate_perf_init(void);
6342 void perf_llc_init(void);
6343 void perf_l2_init(void);
6344 void added_perf_counters_init(void);
6345 void pmt_init(void);
6346 
re_initialize(void)6347 void re_initialize(void)
6348 {
6349 	free_all_buffers();
6350 	setup_all_buffers(false);
6351 	linux_perf_init();
6352 	msr_perf_init();
6353 	rapl_perf_init();
6354 	cstate_perf_init();
6355 	perf_llc_init();
6356 	perf_l2_init();
6357 	added_perf_counters_init();
6358 	pmt_init();
6359 	fprintf(outf, "turbostat: re-initialized with num_cpus %d, allowed_cpus %d\n", topo.num_cpus, topo.allowed_cpus);
6360 }
6361 
set_max_cpu_num(void)6362 void set_max_cpu_num(void)
6363 {
6364 	FILE *filep;
6365 	int current_cpu;
6366 	unsigned long dummy;
6367 	char pathname[64];
6368 
6369 	current_cpu = sched_getcpu();
6370 	if (current_cpu < 0)
6371 		err(1, "cannot find calling cpu ID");
6372 	sprintf(pathname, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", current_cpu);
6373 
6374 	filep = fopen_or_die(pathname, "r");
6375 	topo.max_cpu_num = 0;
6376 	while (fscanf(filep, "%lx,", &dummy) == 1)
6377 		topo.max_cpu_num += BITMASK_SIZE;
6378 	fclose(filep);
6379 	topo.max_cpu_num--;	/* 0 based */
6380 }
6381 
6382 /*
6383  * count_cpus()
6384  * remember the last one seen, it will be the max
6385  */
count_cpus(int cpu)6386 int count_cpus(int cpu)
6387 {
6388 	UNUSED(cpu);
6389 
6390 	topo.num_cpus++;
6391 	return 0;
6392 }
6393 
mark_cpu_present(int cpu)6394 int mark_cpu_present(int cpu)
6395 {
6396 	CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
6397 	return 0;
6398 }
6399 
clear_ht_id(int cpu)6400 int clear_ht_id(int cpu)
6401 {
6402 	int i;
6403 
6404 	cpus[cpu].ht_id = -1;
6405 	for (i = 0; i <= MAX_HT_ID; ++i)
6406 		cpus[cpu].ht_sibling_cpu_id[i] = -1;
6407 	return 0;
6408 }
6409 
set_my_cpu_type(void)6410 int set_my_cpu_type(void)
6411 {
6412 	unsigned int eax, ebx, ecx, edx;
6413 	unsigned int max_level;
6414 
6415 	__cpuid(0, max_level, ebx, ecx, edx);
6416 
6417 	if (max_level < CPUID_LEAF_MODEL_ID)
6418 		return 0;
6419 
6420 	__cpuid(CPUID_LEAF_MODEL_ID, eax, ebx, ecx, edx);
6421 
6422 	return (eax >> CPUID_LEAF_MODEL_ID_CORE_TYPE_SHIFT);
6423 }
6424 
set_cpu_hybrid_type(int cpu)6425 int set_cpu_hybrid_type(int cpu)
6426 {
6427 	if (cpu_migrate(cpu))
6428 		return -1;
6429 
6430 	int type = set_my_cpu_type();
6431 
6432 	cpus[cpu].type = type;
6433 	return 0;
6434 }
6435 
6436 /*
6437  * snapshot_proc_interrupts()
6438  *
6439  * read and record summary of /proc/interrupts
6440  *
6441  * return 1 if config change requires a restart, else return 0
6442  */
snapshot_proc_interrupts(void)6443 int snapshot_proc_interrupts(void)
6444 {
6445 	static FILE *fp;
6446 	int column, retval;
6447 
6448 	if (fp == NULL)
6449 		fp = fopen_or_die("/proc/interrupts", "r");
6450 	else
6451 		rewind(fp);
6452 
6453 	/* read 1st line of /proc/interrupts to get cpu* name for each column */
6454 	for (column = 0; column < topo.num_cpus; ++column) {
6455 		int cpu_number;
6456 
6457 		retval = fscanf(fp, " CPU%d", &cpu_number);
6458 		if (retval != 1)
6459 			break;
6460 
6461 		if (cpu_number > topo.max_cpu_num) {
6462 			warn("/proc/interrupts: cpu%d: > %d", cpu_number, topo.max_cpu_num);
6463 			return 1;
6464 		}
6465 
6466 		irq_column_2_cpu[column] = cpu_number;
6467 		irqs_per_cpu[cpu_number] = 0;
6468 		nmi_per_cpu[cpu_number] = 0;
6469 	}
6470 
6471 	/* read /proc/interrupt count lines and sum up irqs per cpu */
6472 	while (1) {
6473 		int column;
6474 		char buf[64];
6475 		int this_row_is_nmi = 0;
6476 
6477 		retval = fscanf(fp, " %s:", buf);	/* irq# "N:" */
6478 		if (retval != 1)
6479 			break;
6480 
6481 		if (strncmp(buf, "NMI", strlen("NMI")) == 0)
6482 			this_row_is_nmi = 1;
6483 
6484 		/* read the count per cpu */
6485 		for (column = 0; column < topo.num_cpus; ++column) {
6486 
6487 			int cpu_number, irq_count;
6488 
6489 			retval = fscanf(fp, " %d", &irq_count);
6490 
6491 			if (retval != 1)
6492 				break;
6493 
6494 			cpu_number = irq_column_2_cpu[column];
6495 			irqs_per_cpu[cpu_number] += irq_count;
6496 			if (this_row_is_nmi)
6497 				nmi_per_cpu[cpu_number] += irq_count;
6498 		}
6499 		while (getc(fp) != '\n') ;	/* flush interrupt description */
6500 
6501 	}
6502 	return 0;
6503 }
6504 
6505 /*
6506  * snapshot_graphics()
6507  *
6508  * record snapshot of specified graphics sysfs knob
6509  *
6510  * return 1 if config change requires a restart, else return 0
6511  */
snapshot_graphics(int idx)6512 int snapshot_graphics(int idx)
6513 {
6514 	int retval;
6515 
6516 	rewind(gfx_info[idx].fp);
6517 	fflush(gfx_info[idx].fp);
6518 
6519 	switch (idx) {
6520 	case GFX_rc6:
6521 	case SAM_mc6:
6522 		retval = fscanf(gfx_info[idx].fp, "%lld", &gfx_info[idx].val_ull);
6523 		if (retval != 1)
6524 			err(1, "rc6");
6525 		return 0;
6526 	case GFX_MHz:
6527 	case GFX_ACTMHz:
6528 	case SAM_MHz:
6529 	case SAM_ACTMHz:
6530 		retval = fscanf(gfx_info[idx].fp, "%d", &gfx_info[idx].val);
6531 		if (retval != 1)
6532 			err(1, "MHz");
6533 		return 0;
6534 	default:
6535 		return -EINVAL;
6536 	}
6537 }
6538 
6539 /*
6540  * snapshot_cpu_lpi()
6541  *
6542  * record snapshot of
6543  * /sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us
6544  */
snapshot_cpu_lpi_us(void)6545 int snapshot_cpu_lpi_us(void)
6546 {
6547 	FILE *fp;
6548 	int retval;
6549 
6550 	fp = fopen_or_die("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", "r");
6551 
6552 	retval = fscanf(fp, "%lld", &cpuidle_cur_cpu_lpi_us);
6553 	if (retval != 1) {
6554 		fprintf(stderr, "Disabling Low Power Idle CPU output\n");
6555 		BIC_NOT_PRESENT(BIC_CPU_LPI);
6556 		fclose(fp);
6557 		return -1;
6558 	}
6559 
6560 	fclose(fp);
6561 
6562 	return 0;
6563 }
6564 
6565 /*
6566  * snapshot_sys_lpi()
6567  *
6568  * record snapshot of sys_lpi_file
6569  */
snapshot_sys_lpi_us(void)6570 int snapshot_sys_lpi_us(void)
6571 {
6572 	FILE *fp;
6573 	int retval;
6574 
6575 	fp = fopen_or_die(sys_lpi_file, "r");
6576 
6577 	retval = fscanf(fp, "%lld", &cpuidle_cur_sys_lpi_us);
6578 	if (retval != 1) {
6579 		fprintf(stderr, "Disabling Low Power Idle System output\n");
6580 		BIC_NOT_PRESENT(BIC_SYS_LPI);
6581 		fclose(fp);
6582 		return -1;
6583 	}
6584 	fclose(fp);
6585 
6586 	return 0;
6587 }
6588 
6589 /*
6590  * snapshot /proc and /sys files
6591  *
6592  * return 1 if configuration restart needed, else return 0
6593  */
snapshot_proc_sysfs_files(void)6594 int snapshot_proc_sysfs_files(void)
6595 {
6596 	gettimeofday(&procsysfs_tv_begin, (struct timezone *)NULL);
6597 
6598 	if (DO_BIC(BIC_IRQ) || DO_BIC(BIC_NMI))
6599 		if (snapshot_proc_interrupts())
6600 			return 1;
6601 
6602 	if (DO_BIC(BIC_GFX_rc6))
6603 		snapshot_graphics(GFX_rc6);
6604 
6605 	if (DO_BIC(BIC_GFXMHz))
6606 		snapshot_graphics(GFX_MHz);
6607 
6608 	if (DO_BIC(BIC_GFXACTMHz))
6609 		snapshot_graphics(GFX_ACTMHz);
6610 
6611 	if (DO_BIC(BIC_SAM_mc6))
6612 		snapshot_graphics(SAM_mc6);
6613 
6614 	if (DO_BIC(BIC_SAMMHz))
6615 		snapshot_graphics(SAM_MHz);
6616 
6617 	if (DO_BIC(BIC_SAMACTMHz))
6618 		snapshot_graphics(SAM_ACTMHz);
6619 
6620 	if (DO_BIC(BIC_CPU_LPI))
6621 		snapshot_cpu_lpi_us();
6622 
6623 	if (DO_BIC(BIC_SYS_LPI))
6624 		snapshot_sys_lpi_us();
6625 
6626 	return 0;
6627 }
6628 
6629 int exit_requested;
6630 
signal_handler(int signal)6631 static void signal_handler(int signal)
6632 {
6633 	switch (signal) {
6634 	case SIGINT:
6635 		exit_requested = 1;
6636 		if (debug)
6637 			fprintf(stderr, " SIGINT\n");
6638 		break;
6639 	case SIGUSR1:
6640 		if (debug > 1)
6641 			fprintf(stderr, "SIGUSR1\n");
6642 		break;
6643 	}
6644 }
6645 
setup_signal_handler(void)6646 void setup_signal_handler(void)
6647 {
6648 	struct sigaction sa;
6649 
6650 	memset(&sa, 0, sizeof(sa));
6651 
6652 	sa.sa_handler = &signal_handler;
6653 
6654 	if (sigaction(SIGINT, &sa, NULL) < 0)
6655 		err(1, "sigaction SIGINT");
6656 	if (sigaction(SIGUSR1, &sa, NULL) < 0)
6657 		err(1, "sigaction SIGUSR1");
6658 }
6659 
do_sleep(void)6660 void do_sleep(void)
6661 {
6662 	struct timeval tout;
6663 	struct timespec rest;
6664 	fd_set readfds;
6665 	int retval;
6666 
6667 	FD_ZERO(&readfds);
6668 	FD_SET(0, &readfds);
6669 
6670 	if (ignore_stdin) {
6671 		nanosleep(&interval_ts, NULL);
6672 		return;
6673 	}
6674 
6675 	tout = interval_tv;
6676 	retval = select(1, &readfds, NULL, NULL, &tout);
6677 
6678 	if (retval == 1) {
6679 		switch (getc(stdin)) {
6680 		case 'q':
6681 			exit_requested = 1;
6682 			break;
6683 		case EOF:
6684 			/*
6685 			 * 'stdin' is a pipe closed on the other end. There
6686 			 * won't be any further input.
6687 			 */
6688 			ignore_stdin = 1;
6689 			/* Sleep the rest of the time */
6690 			rest.tv_sec = (tout.tv_sec + tout.tv_usec / 1000000);
6691 			rest.tv_nsec = (tout.tv_usec % 1000000) * 1000;
6692 			nanosleep(&rest, NULL);
6693 		}
6694 	}
6695 }
6696 
get_msr_sum(int cpu,off_t offset,unsigned long long * msr)6697 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr)
6698 {
6699 	int ret, idx;
6700 	unsigned long long msr_cur, msr_last;
6701 
6702 	assert(!no_msr);
6703 
6704 	if (!per_cpu_msr_sum)
6705 		return 1;
6706 
6707 	idx = offset_to_idx(offset);
6708 	if (idx < 0)
6709 		return idx;
6710 	/* get_msr_sum() = sum + (get_msr() - last) */
6711 	ret = get_msr(cpu, offset, &msr_cur);
6712 	if (ret)
6713 		return ret;
6714 	msr_last = per_cpu_msr_sum[cpu].entries[idx].last;
6715 	DELTA_WRAP32(msr_cur, msr_last);
6716 	*msr = msr_last + per_cpu_msr_sum[cpu].entries[idx].sum;
6717 
6718 	return 0;
6719 }
6720 
6721 timer_t timerid;
6722 
6723 /* Timer callback, update the sum of MSRs periodically. */
update_msr_sum(PER_THREAD_PARAMS)6724 static int update_msr_sum(PER_THREAD_PARAMS)
6725 {
6726 	int i, ret;
6727 	int cpu = t->cpu_id;
6728 
6729 	UNUSED(c);
6730 	UNUSED(p);
6731 
6732 	assert(!no_msr);
6733 
6734 	for (i = IDX_PKG_ENERGY; i < IDX_COUNT; i++) {
6735 		unsigned long long msr_cur, msr_last;
6736 		off_t offset;
6737 
6738 		if (!idx_valid(i))
6739 			continue;
6740 		offset = idx_to_offset(i);
6741 		if (offset < 0)
6742 			continue;
6743 		ret = get_msr(cpu, offset, &msr_cur);
6744 		if (ret) {
6745 			fprintf(outf, "Can not update msr(0x%llx)\n", (unsigned long long)offset);
6746 			continue;
6747 		}
6748 
6749 		msr_last = per_cpu_msr_sum[cpu].entries[i].last;
6750 		per_cpu_msr_sum[cpu].entries[i].last = msr_cur & 0xffffffff;
6751 
6752 		DELTA_WRAP32(msr_cur, msr_last);
6753 		per_cpu_msr_sum[cpu].entries[i].sum += msr_last;
6754 	}
6755 	return 0;
6756 }
6757 
msr_record_handler(union sigval v)6758 static void msr_record_handler(union sigval v)
6759 {
6760 	UNUSED(v);
6761 
6762 	for_all_cpus(update_msr_sum, EVEN_COUNTERS);
6763 }
6764 
msr_sum_record(void)6765 void msr_sum_record(void)
6766 {
6767 	struct itimerspec its;
6768 	struct sigevent sev;
6769 
6770 	per_cpu_msr_sum = calloc(topo.max_cpu_num + 1, sizeof(struct msr_sum_array));
6771 	if (!per_cpu_msr_sum) {
6772 		fprintf(outf, "Can not allocate memory for long time MSR.\n");
6773 		return;
6774 	}
6775 	/*
6776 	 * Signal handler might be restricted, so use thread notifier instead.
6777 	 */
6778 	memset(&sev, 0, sizeof(struct sigevent));
6779 	sev.sigev_notify = SIGEV_THREAD;
6780 	sev.sigev_notify_function = msr_record_handler;
6781 
6782 	sev.sigev_value.sival_ptr = &timerid;
6783 	if (timer_create(CLOCK_REALTIME, &sev, &timerid) == -1) {
6784 		fprintf(outf, "Can not create timer.\n");
6785 		goto release_msr;
6786 	}
6787 
6788 	its.it_value.tv_sec = 0;
6789 	its.it_value.tv_nsec = 1;
6790 	/*
6791 	 * A wraparound time has been calculated early.
6792 	 * Some sources state that the peak power for a
6793 	 * microprocessor is usually 1.5 times the TDP rating,
6794 	 * use 2 * TDP for safety.
6795 	 */
6796 	its.it_interval.tv_sec = rapl_joule_counter_range / 2;
6797 	its.it_interval.tv_nsec = 0;
6798 
6799 	if (timer_settime(timerid, 0, &its, NULL) == -1) {
6800 		fprintf(outf, "Can not set timer.\n");
6801 		goto release_timer;
6802 	}
6803 	return;
6804 
6805 release_timer:
6806 	timer_delete(timerid);
6807 release_msr:
6808 	free(per_cpu_msr_sum);
6809 	per_cpu_msr_sum = NULL;
6810 }
6811 
6812 /*
6813  * set_my_sched_priority(pri)
6814  * return previous priority on success
6815  * return value < -20 on failure
6816  */
set_my_sched_priority(int priority)6817 int set_my_sched_priority(int priority)
6818 {
6819 	int retval;
6820 	int original_priority;
6821 
6822 	errno = 0;
6823 	original_priority = getpriority(PRIO_PROCESS, 0);
6824 	if (errno && (original_priority == -1))
6825 		return -21;
6826 
6827 	retval = setpriority(PRIO_PROCESS, 0, priority);
6828 	if (retval)
6829 		return -21;
6830 
6831 	errno = 0;
6832 	retval = getpriority(PRIO_PROCESS, 0);
6833 	if (retval != priority)
6834 		return -21;
6835 
6836 	return original_priority;
6837 }
6838 
turbostat_loop()6839 void turbostat_loop()
6840 {
6841 	int retval;
6842 	int restarted = 0;
6843 	unsigned int done_iters = 0;
6844 
6845 	setup_signal_handler();
6846 
6847 	/*
6848 	 * elevate own priority for interval mode
6849 	 *
6850 	 * ignore on error - we probably don't have permission to set it, but
6851 	 * it's not a big deal
6852 	 */
6853 	set_my_sched_priority(-20);
6854 
6855 restart:
6856 	restarted++;
6857 
6858 	snapshot_proc_sysfs_files();
6859 	retval = for_all_cpus(get_counters, EVEN_COUNTERS);
6860 	first_counter_read = 0;
6861 	if (retval < -1) {
6862 		exit(retval);
6863 	} else if (retval == -1) {
6864 		if (restarted > 10) {
6865 			exit(retval);
6866 		}
6867 		re_initialize();
6868 		goto restart;
6869 	}
6870 	restarted = 0;
6871 	done_iters = 0;
6872 	gettimeofday(&tv_even, (struct timezone *)NULL);
6873 
6874 	while (1) {
6875 		if (for_all_proc_cpus(cpu_is_not_present)) {
6876 			re_initialize();
6877 			goto restart;
6878 		}
6879 		if (update_effective_str(false)) {
6880 			re_initialize();
6881 			goto restart;
6882 		}
6883 		do_sleep();
6884 		if (snapshot_proc_sysfs_files())
6885 			goto restart;
6886 		retval = for_all_cpus(get_counters, ODD_COUNTERS);
6887 		if (retval < -1) {
6888 			exit(retval);
6889 		} else if (retval == -1) {
6890 			re_initialize();
6891 			goto restart;
6892 		}
6893 		gettimeofday(&tv_odd, (struct timezone *)NULL);
6894 		timersub(&tv_odd, &tv_even, &tv_delta);
6895 		if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) {
6896 			re_initialize();
6897 			goto restart;
6898 		}
6899 		delta_platform(&platform_counters_odd, &platform_counters_even);
6900 		compute_average(EVEN_COUNTERS);
6901 		format_all_counters(EVEN_COUNTERS);
6902 		flush_output_stdout();
6903 		if (exit_requested)
6904 			break;
6905 		if (num_iterations && ++done_iters >= num_iterations)
6906 			break;
6907 		do_sleep();
6908 		if (snapshot_proc_sysfs_files())
6909 			goto restart;
6910 		retval = for_all_cpus(get_counters, EVEN_COUNTERS);
6911 		if (retval < -1) {
6912 			exit(retval);
6913 		} else if (retval == -1) {
6914 			re_initialize();
6915 			goto restart;
6916 		}
6917 		gettimeofday(&tv_even, (struct timezone *)NULL);
6918 		timersub(&tv_even, &tv_odd, &tv_delta);
6919 		if (for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) {
6920 			re_initialize();
6921 			goto restart;
6922 		}
6923 		delta_platform(&platform_counters_even, &platform_counters_odd);
6924 		compute_average(ODD_COUNTERS);
6925 		format_all_counters(ODD_COUNTERS);
6926 		flush_output_stdout();
6927 		if (exit_requested)
6928 			break;
6929 		if (num_iterations && ++done_iters >= num_iterations)
6930 			break;
6931 	}
6932 }
6933 
probe_dev_msr(void)6934 int probe_dev_msr(void)
6935 {
6936 	struct stat sb;
6937 	char pathname[32];
6938 
6939 	sprintf(pathname, "/dev/msr%d", master_cpu);
6940 	return !stat(pathname, &sb);
6941 }
6942 
probe_dev_cpu_msr(void)6943 int probe_dev_cpu_msr(void)
6944 {
6945 	struct stat sb;
6946 	char pathname[32];
6947 
6948 	sprintf(pathname, "/dev/cpu/%d/msr", master_cpu);
6949 	return !stat(pathname, &sb);
6950 }
6951 
probe_msr_driver(void)6952 int probe_msr_driver(void)
6953 {
6954 	if (probe_dev_msr()) {
6955 		use_android_msr_path = 1;
6956 		return 1;
6957 	}
6958 	return probe_dev_cpu_msr();
6959 }
6960 
check_msr_driver(void)6961 void check_msr_driver(void)
6962 {
6963 	if (probe_msr_driver())
6964 		return;
6965 
6966 	if (system("/sbin/modprobe msr > /dev/null 2>&1"))
6967 		no_msr = 1;
6968 
6969 	if (!probe_msr_driver())
6970 		no_msr = 1;
6971 }
6972 
6973 /*
6974  * check for CAP_SYS_RAWIO
6975  * return 0 on success
6976  * return 1 on fail
6977  */
check_for_cap_sys_rawio(void)6978 int check_for_cap_sys_rawio(void)
6979 {
6980 	cap_t caps;
6981 	cap_flag_value_t cap_flag_value;
6982 	int ret = 0;
6983 
6984 	caps = cap_get_proc();
6985 	if (caps == NULL) {
6986 		/*
6987 		 * CONFIG_MULTIUSER=n kernels have no cap_get_proc()
6988 		 * Allow them to continue and attempt to access MSRs
6989 		 */
6990 		if (errno == ENOSYS)
6991 			return 0;
6992 
6993 		return 1;
6994 	}
6995 
6996 	if (cap_get_flag(caps, CAP_SYS_RAWIO, CAP_EFFECTIVE, &cap_flag_value)) {
6997 		ret = 1;
6998 		goto free_and_exit;
6999 	}
7000 
7001 	if (cap_flag_value != CAP_SET) {
7002 		ret = 1;
7003 		goto free_and_exit;
7004 	}
7005 
7006 free_and_exit:
7007 	if (cap_free(caps) == -1)
7008 		err(-6, "cap_free");
7009 
7010 	return ret;
7011 }
7012 
check_msr_permission(void)7013 void check_msr_permission(void)
7014 {
7015 	int failed = 0;
7016 	char pathname[32];
7017 
7018 	if (no_msr)
7019 		return;
7020 
7021 	/* check for CAP_SYS_RAWIO */
7022 	failed += check_for_cap_sys_rawio();
7023 
7024 	/* test file permissions */
7025 	sprintf(pathname, use_android_msr_path ? "/dev/msr%d" : "/dev/cpu/%d/msr", master_cpu);
7026 	if (euidaccess(pathname, R_OK)) {
7027 		failed++;
7028 	}
7029 
7030 	if (failed) {
7031 		warnx("Failed to access %s. Some of the counters may not be available\n"
7032 		      "\tRun as root to enable them or use %s to disable the access explicitly", pathname, "--no-msr");
7033 		no_msr = 1;
7034 	}
7035 }
7036 
probe_bclk(void)7037 void probe_bclk(void)
7038 {
7039 	unsigned long long msr;
7040 	unsigned int base_ratio;
7041 
7042 	if (!platform->has_nhm_msrs || no_msr)
7043 		return;
7044 
7045 	if (platform->bclk_freq == BCLK_100MHZ)
7046 		bclk = 100.00;
7047 	else if (platform->bclk_freq == BCLK_133MHZ)
7048 		bclk = 133.33;
7049 	else if (platform->bclk_freq == BCLK_SLV)
7050 		bclk = slm_bclk();
7051 	else
7052 		return;
7053 
7054 	get_msr(master_cpu, MSR_PLATFORM_INFO, &msr);
7055 	base_ratio = (msr >> 8) & 0xFF;
7056 
7057 	base_hz = base_ratio * bclk * 1000000;
7058 	has_base_hz = 1;
7059 
7060 	if (platform->enable_tsc_tweak)
7061 		tsc_tweak = base_hz / tsc_hz;
7062 }
7063 
remove_underbar(char * s)7064 static void remove_underbar(char *s)
7065 {
7066 	char *to = s;
7067 
7068 	while (*s) {
7069 		if (*s != '_')
7070 			*to++ = *s;
7071 		s++;
7072 	}
7073 
7074 	*to = 0;
7075 }
7076 
dump_turbo_ratio_info(void)7077 static void dump_turbo_ratio_info(void)
7078 {
7079 	if (!has_turbo)
7080 		return;
7081 
7082 	if (!platform->has_nhm_msrs || no_msr)
7083 		return;
7084 
7085 	if (platform->trl_msrs & TRL_LIMIT2)
7086 		dump_turbo_ratio_limit2();
7087 
7088 	if (platform->trl_msrs & TRL_LIMIT1)
7089 		dump_turbo_ratio_limit1();
7090 
7091 	if (platform->trl_msrs & TRL_BASE) {
7092 		dump_turbo_ratio_limits(MSR_TURBO_RATIO_LIMIT);
7093 
7094 		if (is_hybrid)
7095 			dump_turbo_ratio_limits(MSR_SECONDARY_TURBO_RATIO_LIMIT);
7096 	}
7097 
7098 	if (platform->trl_msrs & TRL_ATOM)
7099 		dump_atom_turbo_ratio_limits();
7100 
7101 	if (platform->trl_msrs & TRL_KNL)
7102 		dump_knl_turbo_ratio_limits();
7103 
7104 	if (platform->has_config_tdp)
7105 		dump_config_tdp();
7106 }
7107 
read_sysfs_int(char * path)7108 static int read_sysfs_int(char *path)
7109 {
7110 	FILE *input;
7111 	int retval = -1;
7112 
7113 	input = fopen(path, "r");
7114 	if (input == NULL) {
7115 		if (debug)
7116 			fprintf(outf, "NSFOD %s\n", path);
7117 		return (-1);
7118 	}
7119 	if (fscanf(input, "%d", &retval) != 1)
7120 		err(1, "%s: failed to read int from file", path);
7121 	fclose(input);
7122 
7123 	return (retval);
7124 }
7125 
dump_sysfs_file(char * path)7126 static void dump_sysfs_file(char *path)
7127 {
7128 	FILE *input;
7129 	char cpuidle_buf[64];
7130 
7131 	input = fopen(path, "r");
7132 	if (input == NULL) {
7133 		if (debug)
7134 			fprintf(outf, "NSFOD %s\n", path);
7135 		return;
7136 	}
7137 	if (!fgets(cpuidle_buf, sizeof(cpuidle_buf), input))
7138 		err(1, "%s: failed to read file", path);
7139 	fclose(input);
7140 
7141 	fprintf(outf, "%s: %s", strrchr(path, '/') + 1, cpuidle_buf);
7142 }
7143 
probe_intel_uncore_frequency_legacy(void)7144 static void probe_intel_uncore_frequency_legacy(void)
7145 {
7146 	int i, j;
7147 	char path[256];
7148 
7149 	for (i = 0; i < topo.num_packages; ++i) {
7150 		for (j = 0; j <= topo.max_die_id; ++j) {
7151 			int k, l;
7152 			char path_base[128];
7153 
7154 			sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/package_%02d_die_%02d", i, j);
7155 
7156 			sprintf(path, "%s/current_freq_khz", path_base);
7157 			if (access(path, R_OK))
7158 				continue;
7159 
7160 			BIC_PRESENT(BIC_UNCORE_MHZ);
7161 
7162 			if (quiet)
7163 				return;
7164 
7165 			sprintf(path, "%s/min_freq_khz", path_base);
7166 			k = read_sysfs_int(path);
7167 			sprintf(path, "%s/max_freq_khz", path_base);
7168 			l = read_sysfs_int(path);
7169 			fprintf(outf, "Uncore Frequency package%d die%d: %d - %d MHz ", i, j, k / 1000, l / 1000);
7170 
7171 			sprintf(path, "%s/initial_min_freq_khz", path_base);
7172 			k = read_sysfs_int(path);
7173 			sprintf(path, "%s/initial_max_freq_khz", path_base);
7174 			l = read_sysfs_int(path);
7175 			fprintf(outf, "(%d - %d MHz)", k / 1000, l / 1000);
7176 
7177 			sprintf(path, "%s/current_freq_khz", path_base);
7178 			k = read_sysfs_int(path);
7179 			fprintf(outf, " %d MHz\n", k / 1000);
7180 		}
7181 	}
7182 }
7183 
probe_intel_uncore_frequency_cluster(void)7184 static void probe_intel_uncore_frequency_cluster(void)
7185 {
7186 	int i, uncore_max_id;
7187 	char path[256];
7188 	char path_base[128];
7189 
7190 	if (access("/sys/devices/system/cpu/intel_uncore_frequency/uncore00/current_freq_khz", R_OK))
7191 		return;
7192 
7193 	for (uncore_max_id = 0;; ++uncore_max_id) {
7194 
7195 		sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/uncore%02d", uncore_max_id);
7196 
7197 		/* uncore## start at 00 and skips no numbers, so stop upon first missing */
7198 		if (access(path_base, R_OK)) {
7199 			uncore_max_id -= 1;
7200 			break;
7201 		}
7202 	}
7203 	for (i = uncore_max_id; i >= 0; --i) {
7204 		int k, l;
7205 		int unc_pkg_id, domain_id, cluster_id;
7206 		char name_buf[16];
7207 
7208 		sprintf(path_base, "/sys/devices/system/cpu/intel_uncore_frequency/uncore%02d", i);
7209 
7210 		if (access(path_base, R_OK))
7211 			err(1, "%s: %s", __func__, path_base);
7212 
7213 		sprintf(path, "%s/package_id", path_base);
7214 		unc_pkg_id = read_sysfs_int(path);
7215 
7216 		sprintf(path, "%s/domain_id", path_base);
7217 		domain_id = read_sysfs_int(path);
7218 
7219 		sprintf(path, "%s/fabric_cluster_id", path_base);
7220 		cluster_id = read_sysfs_int(path);
7221 
7222 		sprintf(path, "%s/current_freq_khz", path_base);
7223 		sprintf(name_buf, "UMHz%d.%d", domain_id, cluster_id);
7224 
7225 		/*
7226 		 * Once add_couter() is called, that counter is always read
7227 		 * and reported -- So it is effectively (enabled & present).
7228 		 * Only call add_counter() here if legacy BIC_UNCORE_MHZ (UncMHz)
7229 		 * is (enabled).  Since we are in this routine, we
7230 		 * know we will not probe and set (present) the legacy counter.
7231 		 *
7232 		 * This allows "--show/--hide UncMHz" to be effective for
7233 		 * the clustered MHz counters, as a group.
7234 		 */
7235 		if BIC_IS_ENABLED
7236 			(BIC_UNCORE_MHZ)
7237 			    add_counter(0, path, name_buf, 0, SCOPE_PACKAGE, COUNTER_K2M, FORMAT_AVERAGE, 0, unc_pkg_id);
7238 
7239 		if (quiet)
7240 			continue;
7241 
7242 		sprintf(path, "%s/min_freq_khz", path_base);
7243 		k = read_sysfs_int(path);
7244 		sprintf(path, "%s/max_freq_khz", path_base);
7245 		l = read_sysfs_int(path);
7246 		fprintf(outf, "Uncore Frequency package%d domain%d cluster%d: %d - %d MHz ", unc_pkg_id, domain_id, cluster_id, k / 1000, l / 1000);
7247 
7248 		sprintf(path, "%s/initial_min_freq_khz", path_base);
7249 		k = read_sysfs_int(path);
7250 		sprintf(path, "%s/initial_max_freq_khz", path_base);
7251 		l = read_sysfs_int(path);
7252 		fprintf(outf, "(%d - %d MHz)", k / 1000, l / 1000);
7253 
7254 		sprintf(path, "%s/current_freq_khz", path_base);
7255 		k = read_sysfs_int(path);
7256 		fprintf(outf, " %d MHz\n", k / 1000);
7257 	}
7258 }
7259 
probe_intel_uncore_frequency(void)7260 static void probe_intel_uncore_frequency(void)
7261 {
7262 	if (!genuine_intel)
7263 		return;
7264 
7265 	if (access("/sys/devices/system/cpu/intel_uncore_frequency/uncore00", R_OK) == 0)
7266 		probe_intel_uncore_frequency_cluster();
7267 	else
7268 		probe_intel_uncore_frequency_legacy();
7269 }
7270 
set_graphics_fp(char * path,int idx)7271 static void set_graphics_fp(char *path, int idx)
7272 {
7273 	if (!access(path, R_OK))
7274 		gfx_info[idx].fp = fopen_or_die(path, "r");
7275 }
7276 
7277 /* Enlarge this if there are /sys/class/drm/card2 ... */
7278 #define GFX_MAX_CARDS	2
7279 
probe_graphics(void)7280 static void probe_graphics(void)
7281 {
7282 	char path[PATH_MAX];
7283 	int i;
7284 
7285 	/* Xe graphics sysfs knobs */
7286 	if (!access("/sys/class/drm/card0/device/tile0/gt0/gtidle/idle_residency_ms", R_OK)) {
7287 		FILE *fp;
7288 		char buf[8];
7289 		bool gt0_is_gt;
7290 
7291 		fp = fopen("/sys/class/drm/card0/device/tile0/gt0/gtidle/name", "r");
7292 		if (!fp)
7293 			goto next;
7294 
7295 		if (!fread(buf, sizeof(char), 7, fp)) {
7296 			fclose(fp);
7297 			goto next;
7298 		}
7299 		fclose(fp);
7300 
7301 		if (!strncmp(buf, "gt0-rc", strlen("gt0-rc")))
7302 			gt0_is_gt = true;
7303 		else if (!strncmp(buf, "gt0-mc", strlen("gt0-mc")))
7304 			gt0_is_gt = false;
7305 		else
7306 			goto next;
7307 
7308 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/gtidle/idle_residency_ms", gt0_is_gt ? GFX_rc6 : SAM_mc6);
7309 
7310 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/freq0/cur_freq", gt0_is_gt ? GFX_MHz : SAM_MHz);
7311 
7312 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt0/freq0/act_freq", gt0_is_gt ? GFX_ACTMHz : SAM_ACTMHz);
7313 
7314 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/gtidle/idle_residency_ms", gt0_is_gt ? SAM_mc6 : GFX_rc6);
7315 
7316 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/freq0/cur_freq", gt0_is_gt ? SAM_MHz : GFX_MHz);
7317 
7318 		set_graphics_fp("/sys/class/drm/card0/device/tile0/gt1/freq0/act_freq", gt0_is_gt ? SAM_ACTMHz : GFX_ACTMHz);
7319 
7320 		goto end;
7321 	}
7322 
7323 next:
7324 	/* New i915 graphics sysfs knobs */
7325 	for (i = 0; i < GFX_MAX_CARDS; i++) {
7326 		snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rc6_residency_ms", i);
7327 		if (!access(path, R_OK))
7328 			break;
7329 	}
7330 
7331 	if (i == GFX_MAX_CARDS)
7332 		goto legacy_i915;
7333 
7334 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rc6_residency_ms", i);
7335 	set_graphics_fp(path, GFX_rc6);
7336 
7337 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rps_cur_freq_mhz", i);
7338 	set_graphics_fp(path, GFX_MHz);
7339 
7340 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt0/rps_act_freq_mhz", i);
7341 	set_graphics_fp(path, GFX_ACTMHz);
7342 
7343 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rc6_residency_ms", i);
7344 	set_graphics_fp(path, SAM_mc6);
7345 
7346 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rps_cur_freq_mhz", i);
7347 	set_graphics_fp(path, SAM_MHz);
7348 
7349 	snprintf(path, PATH_MAX, "/sys/class/drm/card%d/gt/gt1/rps_act_freq_mhz", i);
7350 	set_graphics_fp(path, SAM_ACTMHz);
7351 
7352 	goto end;
7353 
7354 legacy_i915:
7355 	/* Fall back to traditional i915 graphics sysfs knobs */
7356 	set_graphics_fp("/sys/class/drm/card0/power/rc6_residency_ms", GFX_rc6);
7357 
7358 	set_graphics_fp("/sys/class/drm/card0/gt_cur_freq_mhz", GFX_MHz);
7359 	if (!gfx_info[GFX_MHz].fp)
7360 		set_graphics_fp("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", GFX_MHz);
7361 
7362 	set_graphics_fp("/sys/class/drm/card0/gt_act_freq_mhz", GFX_ACTMHz);
7363 	if (!gfx_info[GFX_ACTMHz].fp)
7364 		set_graphics_fp("/sys/class/graphics/fb0/device/drm/card0/gt_act_freq_mhz", GFX_ACTMHz);
7365 
7366 end:
7367 	if (gfx_info[GFX_rc6].fp)
7368 		BIC_PRESENT(BIC_GFX_rc6);
7369 	if (gfx_info[GFX_MHz].fp)
7370 		BIC_PRESENT(BIC_GFXMHz);
7371 	if (gfx_info[GFX_ACTMHz].fp)
7372 		BIC_PRESENT(BIC_GFXACTMHz);
7373 	if (gfx_info[SAM_mc6].fp)
7374 		BIC_PRESENT(BIC_SAM_mc6);
7375 	if (gfx_info[SAM_MHz].fp)
7376 		BIC_PRESENT(BIC_SAMMHz);
7377 	if (gfx_info[SAM_ACTMHz].fp)
7378 		BIC_PRESENT(BIC_SAMACTMHz);
7379 }
7380 
dump_sysfs_cstate_config(void)7381 static void dump_sysfs_cstate_config(void)
7382 {
7383 	char path[64];
7384 	char name_buf[16];
7385 	char desc[64];
7386 	FILE *input;
7387 	int state;
7388 	char *sp;
7389 
7390 	if (access("/sys/devices/system/cpu/cpuidle", R_OK)) {
7391 		fprintf(outf, "cpuidle not loaded\n");
7392 		return;
7393 	}
7394 
7395 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_driver");
7396 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor");
7397 	dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor_ro");
7398 
7399 	for (state = 0; state < 10; ++state) {
7400 
7401 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", master_cpu, state);
7402 		input = fopen(path, "r");
7403 		if (input == NULL)
7404 			continue;
7405 		if (!fgets(name_buf, sizeof(name_buf), input))
7406 			err(1, "%s: failed to read file", path);
7407 
7408 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
7409 		sp = strchr(name_buf, '-');
7410 		if (!sp)
7411 			sp = strchrnul(name_buf, '\n');
7412 		*sp = '\0';
7413 		fclose(input);
7414 
7415 		remove_underbar(name_buf);
7416 
7417 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/desc", master_cpu, state);
7418 		input = fopen(path, "r");
7419 		if (input == NULL)
7420 			continue;
7421 		if (!fgets(desc, sizeof(desc), input))
7422 			err(1, "%s: failed to read file", path);
7423 
7424 		fprintf(outf, "cpu%d: %s: %s", master_cpu, name_buf, desc);
7425 		fclose(input);
7426 	}
7427 }
7428 
dump_sysfs_pstate_config(void)7429 static void dump_sysfs_pstate_config(void)
7430 {
7431 	char path[64];
7432 	char driver_buf[64];
7433 	char governor_buf[64];
7434 	FILE *input;
7435 	int turbo;
7436 
7437 	sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_driver", master_cpu);
7438 	input = fopen(path, "r");
7439 	if (input == NULL) {
7440 		fprintf(outf, "NSFOD %s\n", path);
7441 		return;
7442 	}
7443 	if (!fgets(driver_buf, sizeof(driver_buf), input))
7444 		err(1, "%s: failed to read file", path);
7445 	fclose(input);
7446 
7447 	sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_governor", master_cpu);
7448 	input = fopen(path, "r");
7449 	if (input == NULL) {
7450 		fprintf(outf, "NSFOD %s\n", path);
7451 		return;
7452 	}
7453 	if (!fgets(governor_buf, sizeof(governor_buf), input))
7454 		err(1, "%s: failed to read file", path);
7455 	fclose(input);
7456 
7457 	fprintf(outf, "cpu%d: cpufreq driver: %s", master_cpu, driver_buf);
7458 	fprintf(outf, "cpu%d: cpufreq governor: %s", master_cpu, governor_buf);
7459 
7460 	sprintf(path, "/sys/devices/system/cpu/cpufreq/boost");
7461 	input = fopen(path, "r");
7462 	if (input != NULL) {
7463 		if (fscanf(input, "%d", &turbo) != 1)
7464 			err(1, "%s: failed to parse number from file", path);
7465 		fprintf(outf, "cpufreq boost: %d\n", turbo);
7466 		fclose(input);
7467 	}
7468 
7469 	sprintf(path, "/sys/devices/system/cpu/intel_pstate/no_turbo");
7470 	input = fopen(path, "r");
7471 	if (input != NULL) {
7472 		if (fscanf(input, "%d", &turbo) != 1)
7473 			err(1, "%s: failed to parse number from file", path);
7474 		fprintf(outf, "cpufreq intel_pstate no_turbo: %d\n", turbo);
7475 		fclose(input);
7476 	}
7477 }
7478 
7479 /*
7480  * print_epb()
7481  * Decode the ENERGY_PERF_BIAS MSR
7482  */
print_epb(PER_THREAD_PARAMS)7483 int print_epb(PER_THREAD_PARAMS)
7484 {
7485 	char *epb_string;
7486 	int cpu, epb;
7487 
7488 	UNUSED(c);
7489 	UNUSED(p);
7490 
7491 	if (!has_epb)
7492 		return 0;
7493 
7494 	cpu = t->cpu_id;
7495 
7496 	/* EPB is per-package */
7497 	if (!is_cpu_first_thread_in_package(t, c, p))
7498 		return 0;
7499 
7500 	if (cpu_migrate(cpu)) {
7501 		fprintf(outf, "print_epb: Could not migrate to CPU %d\n", cpu);
7502 		return -1;
7503 	}
7504 
7505 	epb = get_epb(cpu);
7506 	if (epb < 0)
7507 		return 0;
7508 
7509 	switch (epb) {
7510 	case ENERGY_PERF_BIAS_PERFORMANCE:
7511 		epb_string = "performance";
7512 		break;
7513 	case ENERGY_PERF_BIAS_NORMAL:
7514 		epb_string = "balanced";
7515 		break;
7516 	case ENERGY_PERF_BIAS_POWERSAVE:
7517 		epb_string = "powersave";
7518 		break;
7519 	default:
7520 		epb_string = "custom";
7521 		break;
7522 	}
7523 	fprintf(outf, "cpu%d: EPB: %d (%s)\n", cpu, epb, epb_string);
7524 
7525 	return 0;
7526 }
7527 
7528 /*
7529  * print_hwp()
7530  * Decode the MSR_HWP_CAPABILITIES
7531  */
print_hwp(PER_THREAD_PARAMS)7532 int print_hwp(PER_THREAD_PARAMS)
7533 {
7534 	unsigned long long msr;
7535 	int cpu;
7536 
7537 	UNUSED(c);
7538 	UNUSED(p);
7539 
7540 	if (no_msr)
7541 		return 0;
7542 
7543 	if (!has_hwp)
7544 		return 0;
7545 
7546 	cpu = t->cpu_id;
7547 
7548 	/* MSR_HWP_CAPABILITIES is per-package */
7549 	if (!is_cpu_first_thread_in_package(t, c, p))
7550 		return 0;
7551 
7552 	if (cpu_migrate(cpu)) {
7553 		fprintf(outf, "print_hwp: Could not migrate to CPU %d\n", cpu);
7554 		return -1;
7555 	}
7556 
7557 	if (get_msr(cpu, MSR_PM_ENABLE, &msr))
7558 		return 0;
7559 
7560 	fprintf(outf, "cpu%d: MSR_PM_ENABLE: 0x%08llx (%sHWP)\n", cpu, msr, (msr & (1 << 0)) ? "" : "No-");
7561 
7562 	/* MSR_PM_ENABLE[1] == 1 if HWP is enabled and MSRs visible */
7563 	if ((msr & (1 << 0)) == 0)
7564 		return 0;
7565 
7566 	if (get_msr(cpu, MSR_HWP_CAPABILITIES, &msr))
7567 		return 0;
7568 
7569 	fprintf(outf, "cpu%d: MSR_HWP_CAPABILITIES: 0x%08llx "
7570 		"(high %d guar %d eff %d low %d)\n",
7571 		cpu, msr,
7572 		(unsigned int)HWP_HIGHEST_PERF(msr),
7573 		(unsigned int)HWP_GUARANTEED_PERF(msr), (unsigned int)HWP_MOSTEFFICIENT_PERF(msr), (unsigned int)HWP_LOWEST_PERF(msr));
7574 
7575 	if (get_msr(cpu, MSR_HWP_REQUEST, &msr))
7576 		return 0;
7577 
7578 	fprintf(outf, "cpu%d: MSR_HWP_REQUEST: 0x%08llx "
7579 		"(min %d max %d des %d epp 0x%x window 0x%x pkg 0x%x)\n",
7580 		cpu, msr,
7581 		(unsigned int)(((msr) >> 0) & 0xff),
7582 		(unsigned int)(((msr) >> 8) & 0xff),
7583 		(unsigned int)(((msr) >> 16) & 0xff),
7584 		(unsigned int)(((msr) >> 24) & 0xff), (unsigned int)(((msr) >> 32) & 0xff3), (unsigned int)(((msr) >> 42) & 0x1));
7585 
7586 	if (has_hwp_pkg) {
7587 		if (get_msr(cpu, MSR_HWP_REQUEST_PKG, &msr))
7588 			return 0;
7589 
7590 		fprintf(outf, "cpu%d: MSR_HWP_REQUEST_PKG: 0x%08llx "
7591 			"(min %d max %d des %d epp 0x%x window 0x%x)\n",
7592 			cpu, msr,
7593 			(unsigned int)(((msr) >> 0) & 0xff),
7594 			(unsigned int)(((msr) >> 8) & 0xff),
7595 			(unsigned int)(((msr) >> 16) & 0xff), (unsigned int)(((msr) >> 24) & 0xff), (unsigned int)(((msr) >> 32) & 0xff3));
7596 	}
7597 	if (has_hwp_notify) {
7598 		if (get_msr(cpu, MSR_HWP_INTERRUPT, &msr))
7599 			return 0;
7600 
7601 		fprintf(outf, "cpu%d: MSR_HWP_INTERRUPT: 0x%08llx "
7602 			"(%s_Guaranteed_Perf_Change, %s_Excursion_Min)\n", cpu, msr, ((msr) & 0x1) ? "EN" : "Dis", ((msr) & 0x2) ? "EN" : "Dis");
7603 	}
7604 	if (get_msr(cpu, MSR_HWP_STATUS, &msr))
7605 		return 0;
7606 
7607 	fprintf(outf, "cpu%d: MSR_HWP_STATUS: 0x%08llx "
7608 		"(%sGuaranteed_Perf_Change, %sExcursion_Min)\n", cpu, msr, ((msr) & 0x1) ? "" : "No-", ((msr) & 0x4) ? "" : "No-");
7609 
7610 	return 0;
7611 }
7612 
7613 /*
7614  * print_perf_limit()
7615  */
print_perf_limit(PER_THREAD_PARAMS)7616 int print_perf_limit(PER_THREAD_PARAMS)
7617 {
7618 	unsigned long long msr;
7619 	int cpu;
7620 
7621 	UNUSED(c);
7622 	UNUSED(p);
7623 
7624 	if (no_msr)
7625 		return 0;
7626 
7627 	cpu = t->cpu_id;
7628 
7629 	/* per-package */
7630 	if (!is_cpu_first_thread_in_package(t, c, p))
7631 		return 0;
7632 
7633 	if (cpu_migrate(cpu)) {
7634 		fprintf(outf, "print_perf_limit: Could not migrate to CPU %d\n", cpu);
7635 		return -1;
7636 	}
7637 
7638 	if (platform->plr_msrs & PLR_CORE) {
7639 		get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
7640 		fprintf(outf, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
7641 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
7642 			(msr & 1 << 15) ? "bit15, " : "",
7643 			(msr & 1 << 14) ? "bit14, " : "",
7644 			(msr & 1 << 13) ? "Transitions, " : "",
7645 			(msr & 1 << 12) ? "MultiCoreTurbo, " : "",
7646 			(msr & 1 << 11) ? "PkgPwrL2, " : "",
7647 			(msr & 1 << 10) ? "PkgPwrL1, " : "",
7648 			(msr & 1 << 9) ? "CorePwr, " : "",
7649 			(msr & 1 << 8) ? "Amps, " : "",
7650 			(msr & 1 << 6) ? "VR-Therm, " : "",
7651 			(msr & 1 << 5) ? "Auto-HWP, " : "",
7652 			(msr & 1 << 4) ? "Graphics, " : "",
7653 			(msr & 1 << 2) ? "bit2, " : "", (msr & 1 << 1) ? "ThermStatus, " : "", (msr & 1 << 0) ? "PROCHOT, " : "");
7654 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
7655 			(msr & 1 << 31) ? "bit31, " : "",
7656 			(msr & 1 << 30) ? "bit30, " : "",
7657 			(msr & 1 << 29) ? "Transitions, " : "",
7658 			(msr & 1 << 28) ? "MultiCoreTurbo, " : "",
7659 			(msr & 1 << 27) ? "PkgPwrL2, " : "",
7660 			(msr & 1 << 26) ? "PkgPwrL1, " : "",
7661 			(msr & 1 << 25) ? "CorePwr, " : "",
7662 			(msr & 1 << 24) ? "Amps, " : "",
7663 			(msr & 1 << 22) ? "VR-Therm, " : "",
7664 			(msr & 1 << 21) ? "Auto-HWP, " : "",
7665 			(msr & 1 << 20) ? "Graphics, " : "",
7666 			(msr & 1 << 18) ? "bit18, " : "", (msr & 1 << 17) ? "ThermStatus, " : "", (msr & 1 << 16) ? "PROCHOT, " : "");
7667 
7668 	}
7669 	if (platform->plr_msrs & PLR_GFX) {
7670 		get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
7671 		fprintf(outf, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
7672 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s)",
7673 			(msr & 1 << 0) ? "PROCHOT, " : "",
7674 			(msr & 1 << 1) ? "ThermStatus, " : "",
7675 			(msr & 1 << 4) ? "Graphics, " : "",
7676 			(msr & 1 << 6) ? "VR-Therm, " : "",
7677 			(msr & 1 << 8) ? "Amps, " : "",
7678 			(msr & 1 << 9) ? "GFXPwr, " : "", (msr & 1 << 10) ? "PkgPwrL1, " : "", (msr & 1 << 11) ? "PkgPwrL2, " : "");
7679 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s)\n",
7680 			(msr & 1 << 16) ? "PROCHOT, " : "",
7681 			(msr & 1 << 17) ? "ThermStatus, " : "",
7682 			(msr & 1 << 20) ? "Graphics, " : "",
7683 			(msr & 1 << 22) ? "VR-Therm, " : "",
7684 			(msr & 1 << 24) ? "Amps, " : "",
7685 			(msr & 1 << 25) ? "GFXPwr, " : "", (msr & 1 << 26) ? "PkgPwrL1, " : "", (msr & 1 << 27) ? "PkgPwrL2, " : "");
7686 	}
7687 	if (platform->plr_msrs & PLR_RING) {
7688 		get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
7689 		fprintf(outf, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
7690 		fprintf(outf, " (Active: %s%s%s%s%s%s)",
7691 			(msr & 1 << 0) ? "PROCHOT, " : "",
7692 			(msr & 1 << 1) ? "ThermStatus, " : "",
7693 			(msr & 1 << 6) ? "VR-Therm, " : "",
7694 			(msr & 1 << 8) ? "Amps, " : "", (msr & 1 << 10) ? "PkgPwrL1, " : "", (msr & 1 << 11) ? "PkgPwrL2, " : "");
7695 		fprintf(outf, " (Logged: %s%s%s%s%s%s)\n",
7696 			(msr & 1 << 16) ? "PROCHOT, " : "",
7697 			(msr & 1 << 17) ? "ThermStatus, " : "",
7698 			(msr & 1 << 22) ? "VR-Therm, " : "",
7699 			(msr & 1 << 24) ? "Amps, " : "", (msr & 1 << 26) ? "PkgPwrL1, " : "", (msr & 1 << 27) ? "PkgPwrL2, " : "");
7700 	}
7701 	return 0;
7702 }
7703 
7704 #define	RAPL_POWER_GRANULARITY	0x7FFF	/* 15 bit power granularity */
7705 #define	RAPL_TIME_GRANULARITY	0x3F	/* 6 bit time granularity */
7706 
get_quirk_tdp(void)7707 double get_quirk_tdp(void)
7708 {
7709 	if (platform->rapl_quirk_tdp)
7710 		return platform->rapl_quirk_tdp;
7711 
7712 	return 135.0;
7713 }
7714 
get_tdp_intel(void)7715 double get_tdp_intel(void)
7716 {
7717 	unsigned long long msr;
7718 
7719 	if (valid_rapl_msrs & RAPL_PKG_POWER_INFO)
7720 		if (!get_msr(master_cpu, MSR_PKG_POWER_INFO, &msr))
7721 			return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
7722 	return get_quirk_tdp();
7723 }
7724 
get_tdp_amd(void)7725 double get_tdp_amd(void)
7726 {
7727 	return get_quirk_tdp();
7728 }
7729 
rapl_probe_intel(void)7730 void rapl_probe_intel(void)
7731 {
7732 	unsigned long long msr;
7733 	unsigned int time_unit;
7734 	double tdp;
7735 
7736 	if (rapl_joules) {
7737 		CLR_BIC(BIC_SysWatt, &bic_enabled);
7738 		CLR_BIC(BIC_PkgWatt, &bic_enabled);
7739 		CLR_BIC(BIC_CorWatt, &bic_enabled);
7740 		CLR_BIC(BIC_RAMWatt, &bic_enabled);
7741 		CLR_BIC(BIC_GFXWatt, &bic_enabled);
7742 	} else {
7743 		CLR_BIC(BIC_Sys_J, &bic_enabled);
7744 		CLR_BIC(BIC_Pkg_J, &bic_enabled);
7745 		CLR_BIC(BIC_Cor_J, &bic_enabled);
7746 		CLR_BIC(BIC_RAM_J, &bic_enabled);
7747 		CLR_BIC(BIC_GFX_J, &bic_enabled);
7748 	}
7749 
7750 	if (!valid_rapl_msrs || no_msr)
7751 		return;
7752 
7753 	if (!(valid_rapl_msrs & RAPL_PKG_PERF_STATUS))
7754 		CLR_BIC(BIC_PKG__, &bic_enabled);
7755 	if (!(valid_rapl_msrs & RAPL_DRAM_PERF_STATUS))
7756 		CLR_BIC(BIC_RAM__, &bic_enabled);
7757 
7758 	/* units on package 0, verify later other packages match */
7759 	if (get_msr(master_cpu, MSR_RAPL_POWER_UNIT, &msr))
7760 		return;
7761 
7762 	rapl_power_units = 1.0 / (1 << (msr & 0xF));
7763 	if (platform->has_rapl_divisor)
7764 		rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
7765 	else
7766 		rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
7767 
7768 	if (platform->has_fixed_rapl_unit)
7769 		rapl_dram_energy_units = (15.3 / 1000000);
7770 	else
7771 		rapl_dram_energy_units = rapl_energy_units;
7772 
7773 	if (platform->has_fixed_rapl_psys_unit)
7774 		rapl_psys_energy_units = 1.0;
7775 	else
7776 		rapl_psys_energy_units = rapl_energy_units;
7777 
7778 	time_unit = msr >> 16 & 0xF;
7779 	if (time_unit == 0)
7780 		time_unit = 0xA;
7781 
7782 	rapl_time_units = 1.0 / (1 << (time_unit));
7783 
7784 	tdp = get_tdp_intel();
7785 
7786 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
7787 	if (!quiet)
7788 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
7789 }
7790 
rapl_probe_amd(void)7791 void rapl_probe_amd(void)
7792 {
7793 	unsigned long long msr;
7794 	double tdp;
7795 
7796 	if (rapl_joules) {
7797 		CLR_BIC(BIC_SysWatt, &bic_enabled);
7798 		CLR_BIC(BIC_CorWatt, &bic_enabled);
7799 	} else {
7800 		CLR_BIC(BIC_Pkg_J, &bic_enabled);
7801 		CLR_BIC(BIC_Cor_J, &bic_enabled);
7802 	}
7803 
7804 	if (!valid_rapl_msrs || no_msr)
7805 		return;
7806 
7807 	if (get_msr(master_cpu, MSR_RAPL_PWR_UNIT, &msr))
7808 		return;
7809 
7810 	rapl_time_units = ldexp(1.0, -(msr >> 16 & 0xf));
7811 	rapl_energy_units = ldexp(1.0, -(msr >> 8 & 0x1f));
7812 	rapl_power_units = ldexp(1.0, -(msr & 0xf));
7813 
7814 	tdp = get_tdp_amd();
7815 
7816 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
7817 	if (!quiet)
7818 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
7819 }
7820 
print_power_limit_msr(int cpu,unsigned long long msr,char * label)7821 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
7822 {
7823 	fprintf(outf, "cpu%d: %s: %sabled (%0.3f Watts, %f sec, clamp %sabled)\n",
7824 		cpu, label,
7825 		((msr >> 15) & 1) ? "EN" : "DIS",
7826 		((msr >> 0) & 0x7FFF) * rapl_power_units,
7827 		(1.0 + (((msr >> 22) & 0x3) / 4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units, (((msr >> 16) & 1) ? "EN" : "DIS"));
7828 
7829 	return;
7830 }
7831 
fread_int(char * path,int * val)7832 static int fread_int(char *path, int *val)
7833 {
7834 	FILE *filep;
7835 	int ret;
7836 
7837 	filep = fopen(path, "r");
7838 	if (!filep)
7839 		return -1;
7840 
7841 	ret = fscanf(filep, "%d", val);
7842 	fclose(filep);
7843 	return ret;
7844 }
7845 
fread_ull(char * path,unsigned long long * val)7846 static int fread_ull(char *path, unsigned long long *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, "%llu", val);
7856 	fclose(filep);
7857 	return ret;
7858 }
7859 
fread_str(char * path,char * buf,int size)7860 static int fread_str(char *path, char *buf, int size)
7861 {
7862 	FILE *filep;
7863 	int ret;
7864 	char *cp;
7865 
7866 	filep = fopen(path, "r");
7867 	if (!filep)
7868 		return -1;
7869 
7870 	ret = fread(buf, 1, size, filep);
7871 	fclose(filep);
7872 
7873 	/* replace '\n' with '\0' */
7874 	cp = strchr(buf, '\n');
7875 	if (cp != NULL)
7876 		*cp = '\0';
7877 
7878 	return ret;
7879 }
7880 
7881 #define PATH_RAPL_SYSFS	"/sys/class/powercap"
7882 
dump_one_domain(char * domain_path)7883 static int dump_one_domain(char *domain_path)
7884 {
7885 	char path[PATH_MAX];
7886 	char str[PATH_MAX];
7887 	unsigned long long val;
7888 	int constraint;
7889 	int enable;
7890 	int ret;
7891 
7892 	snprintf(path, PATH_MAX, "%s/name", domain_path);
7893 	ret = fread_str(path, str, PATH_MAX);
7894 	if (ret <= 0)
7895 		return -1;
7896 
7897 	fprintf(outf, "%s: %s", domain_path + strlen(PATH_RAPL_SYSFS) + 1, str);
7898 
7899 	snprintf(path, PATH_MAX, "%s/enabled", domain_path);
7900 	ret = fread_int(path, &enable);
7901 	if (ret <= 0)
7902 		return -1;
7903 
7904 	if (!enable) {
7905 		fputs(" disabled\n", outf);
7906 		return 0;
7907 	}
7908 
7909 	for (constraint = 0;; constraint++) {
7910 		snprintf(path, PATH_MAX, "%s/constraint_%d_time_window_us", domain_path, constraint);
7911 		ret = fread_ull(path, &val);
7912 		if (ret <= 0)
7913 			break;
7914 
7915 		if (val > 1000000)
7916 			fprintf(outf, " %0.1fs", (double)val / 1000000);
7917 		else if (val > 1000)
7918 			fprintf(outf, " %0.1fms", (double)val / 1000);
7919 		else
7920 			fprintf(outf, " %0.1fus", (double)val);
7921 
7922 		snprintf(path, PATH_MAX, "%s/constraint_%d_power_limit_uw", domain_path, constraint);
7923 		ret = fread_ull(path, &val);
7924 		if (ret > 0 && val)
7925 			fprintf(outf, ":%lluW", val / 1000000);
7926 
7927 		snprintf(path, PATH_MAX, "%s/constraint_%d_max_power_uw", domain_path, constraint);
7928 		ret = fread_ull(path, &val);
7929 		if (ret > 0 && val)
7930 			fprintf(outf, ",max:%lluW", val / 1000000);
7931 	}
7932 	fputc('\n', outf);
7933 
7934 	return 0;
7935 }
7936 
print_rapl_sysfs(void)7937 static int print_rapl_sysfs(void)
7938 {
7939 	DIR *dir, *cdir;
7940 	struct dirent *entry, *centry;
7941 	char path[PATH_MAX];
7942 	char str[PATH_MAX];
7943 
7944 	if ((dir = opendir(PATH_RAPL_SYSFS)) == NULL) {
7945 		warn("open %s failed", PATH_RAPL_SYSFS);
7946 		return 1;
7947 	}
7948 
7949 	while ((entry = readdir(dir)) != NULL) {
7950 		if (strlen(entry->d_name) > 100)
7951 			continue;
7952 
7953 		if (strncmp(entry->d_name, "intel-rapl", strlen("intel-rapl")))
7954 			continue;
7955 
7956 		snprintf(path, PATH_MAX, "%s/%s/name", PATH_RAPL_SYSFS, entry->d_name);
7957 
7958 		/* Parse top level domains first, including package and psys */
7959 		fread_str(path, str, PATH_MAX);
7960 		if (strncmp(str, "package", strlen("package")) && strncmp(str, "psys", strlen("psys")))
7961 			continue;
7962 
7963 		snprintf(path, PATH_MAX, "%s/%s", PATH_RAPL_SYSFS, entry->d_name);
7964 		if ((cdir = opendir(path)) == NULL) {
7965 			perror("opendir() error");
7966 			return 1;
7967 		}
7968 
7969 		dump_one_domain(path);
7970 
7971 		while ((centry = readdir(cdir)) != NULL) {
7972 			if (strncmp(centry->d_name, "intel-rapl", strlen("intel-rapl")))
7973 				continue;
7974 			snprintf(path, PATH_MAX, "%s/%s/%s", PATH_RAPL_SYSFS, entry->d_name, centry->d_name);
7975 			dump_one_domain(path);
7976 		}
7977 		closedir(cdir);
7978 	}
7979 
7980 	closedir(dir);
7981 	return 0;
7982 }
7983 
print_rapl(PER_THREAD_PARAMS)7984 int print_rapl(PER_THREAD_PARAMS)
7985 {
7986 	unsigned long long msr;
7987 	const char *msr_name;
7988 	int cpu;
7989 
7990 	UNUSED(c);
7991 	UNUSED(p);
7992 
7993 	if (!valid_rapl_msrs)
7994 		return 0;
7995 
7996 	/* RAPL counters are per package, so print only for 1st thread/package */
7997 	if (!is_cpu_first_thread_in_package(t, c, p))
7998 		return 0;
7999 
8000 	cpu = t->cpu_id;
8001 	if (cpu_migrate(cpu)) {
8002 		fprintf(outf, "print_rapl: Could not migrate to CPU %d\n", cpu);
8003 		return -1;
8004 	}
8005 
8006 	if (valid_rapl_msrs & RAPL_AMD_F17H) {
8007 		msr_name = "MSR_RAPL_PWR_UNIT";
8008 		if (get_msr(cpu, MSR_RAPL_PWR_UNIT, &msr))
8009 			return -1;
8010 	} else {
8011 		msr_name = "MSR_RAPL_POWER_UNIT";
8012 		if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
8013 			return -1;
8014 	}
8015 
8016 	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);
8017 
8018 	if (valid_rapl_msrs & RAPL_PKG_POWER_INFO) {
8019 
8020 		if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
8021 			return -5;
8022 
8023 		fprintf(outf, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
8024 			cpu, msr,
8025 			((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
8026 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
8027 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units, ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
8028 
8029 	}
8030 	if (valid_rapl_msrs & RAPL_PKG) {
8031 
8032 		if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
8033 			return -9;
8034 
8035 		fprintf(outf, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n", cpu, msr, (msr >> 63) & 1 ? "" : "UN");
8036 
8037 		print_power_limit_msr(cpu, msr, "PKG Limit #1");
8038 		fprintf(outf, "cpu%d: PKG Limit #2: %sabled (%0.3f Watts, %f* sec, clamp %sabled)\n",
8039 			cpu,
8040 			((msr >> 47) & 1) ? "EN" : "DIS",
8041 			((msr >> 32) & 0x7FFF) * rapl_power_units,
8042 			(1.0 + (((msr >> 54) & 0x3) / 4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units, ((msr >> 48) & 1) ? "EN" : "DIS");
8043 
8044 		if (get_msr(cpu, MSR_VR_CURRENT_CONFIG, &msr))
8045 			return -9;
8046 
8047 		fprintf(outf, "cpu%d: MSR_VR_CURRENT_CONFIG: 0x%08llx\n", cpu, msr);
8048 		fprintf(outf, "cpu%d: PKG Limit #4: %f Watts (%slocked)\n", cpu, ((msr >> 0) & 0x1FFF) * rapl_power_units, (msr >> 31) & 1 ? "" : "UN");
8049 	}
8050 
8051 	if (valid_rapl_msrs & RAPL_DRAM_POWER_INFO) {
8052 		if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
8053 			return -6;
8054 
8055 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
8056 			cpu, msr,
8057 			((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
8058 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
8059 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units, ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
8060 	}
8061 	if (valid_rapl_msrs & RAPL_DRAM) {
8062 		if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
8063 			return -9;
8064 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n", cpu, msr, (msr >> 31) & 1 ? "" : "UN");
8065 
8066 		print_power_limit_msr(cpu, msr, "DRAM Limit");
8067 	}
8068 	if (valid_rapl_msrs & RAPL_CORE_POLICY) {
8069 		if (get_msr(cpu, MSR_PP0_POLICY, &msr))
8070 			return -7;
8071 
8072 		fprintf(outf, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
8073 	}
8074 	if (valid_rapl_msrs & RAPL_CORE_POWER_LIMIT) {
8075 		if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
8076 			return -9;
8077 		fprintf(outf, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n", cpu, msr, (msr >> 31) & 1 ? "" : "UN");
8078 		print_power_limit_msr(cpu, msr, "Cores Limit");
8079 	}
8080 	if (valid_rapl_msrs & RAPL_GFX) {
8081 		if (get_msr(cpu, MSR_PP1_POLICY, &msr))
8082 			return -8;
8083 
8084 		fprintf(outf, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
8085 
8086 		if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
8087 			return -9;
8088 		fprintf(outf, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n", cpu, msr, (msr >> 31) & 1 ? "" : "UN");
8089 		print_power_limit_msr(cpu, msr, "GFX Limit");
8090 	}
8091 	return 0;
8092 }
8093 
8094 /*
8095  * probe_rapl_msrs
8096  *
8097  * initialize global valid_rapl_msrs to platform->plat_rapl_msrs
8098  * only if PKG_ENERGY counter is enumerated and reads non-zero
8099  */
probe_rapl_msrs(void)8100 void probe_rapl_msrs(void)
8101 {
8102 	int ret;
8103 	off_t offset;
8104 	unsigned long long msr_value;
8105 
8106 	if (no_msr)
8107 		return;
8108 
8109 	if ((platform->plat_rapl_msrs & (RAPL_PKG | RAPL_AMD_F17H)) == 0)
8110 		return;
8111 
8112 	offset = idx_to_offset(IDX_PKG_ENERGY);
8113 	if (offset < 0)
8114 		return;
8115 
8116 	ret = get_msr(master_cpu, offset, &msr_value);
8117 	if (ret) {
8118 		if (debug)
8119 			fprintf(outf, "Can not read RAPL_PKG_ENERGY MSR(0x%llx)\n", (unsigned long long)offset);
8120 		return;
8121 	}
8122 	if (msr_value == 0) {
8123 		if (debug)
8124 			fprintf(outf, "RAPL_PKG_ENERGY MSR(0x%llx) == ZERO: disabling all RAPL MSRs\n", (unsigned long long)offset);
8125 		return;
8126 	}
8127 
8128 	valid_rapl_msrs = platform->plat_rapl_msrs;	/* success */
8129 }
8130 
8131 /*
8132  * probe_rapl()
8133  *
8134  * sets rapl_power_units, rapl_energy_units, rapl_time_units
8135  */
probe_rapl(void)8136 void probe_rapl(void)
8137 {
8138 	probe_rapl_msrs();
8139 
8140 	if (genuine_intel)
8141 		rapl_probe_intel();
8142 	if (authentic_amd || hygon_genuine)
8143 		rapl_probe_amd();
8144 
8145 	if (quiet)
8146 		return;
8147 
8148 	print_rapl_sysfs();
8149 
8150 	if (!valid_rapl_msrs || no_msr)
8151 		return;
8152 
8153 	for_all_cpus(print_rapl, ODD_COUNTERS);
8154 }
8155 
8156 /*
8157  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
8158  * the Thermal Control Circuit (TCC) activates.
8159  * This is usually equal to tjMax.
8160  *
8161  * Older processors do not have this MSR, so there we guess,
8162  * but also allow cmdline over-ride with -T.
8163  *
8164  * Several MSR temperature values are in units of degrees-C
8165  * below this value, including the Digital Thermal Sensor (DTS),
8166  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
8167  */
set_temperature_target(PER_THREAD_PARAMS)8168 int set_temperature_target(PER_THREAD_PARAMS)
8169 {
8170 	unsigned long long msr;
8171 	unsigned int tcc_default, tcc_offset;
8172 	int cpu;
8173 
8174 	UNUSED(c);
8175 	UNUSED(p);
8176 
8177 	/* tj_max is used only for dts or ptm */
8178 	if (!(do_dts || do_ptm))
8179 		return 0;
8180 
8181 	/* this is a per-package concept */
8182 	if (!is_cpu_first_thread_in_package(t, c, p))
8183 		return 0;
8184 
8185 	cpu = t->cpu_id;
8186 	if (cpu_migrate(cpu)) {
8187 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
8188 		return -1;
8189 	}
8190 
8191 	if (tj_max_override != 0) {
8192 		tj_max = tj_max_override;
8193 		fprintf(outf, "cpu%d: Using cmdline TCC Target (%d C)\n", cpu, tj_max);
8194 		return 0;
8195 	}
8196 
8197 	/* Temperature Target MSR is Nehalem and newer only */
8198 	if (!platform->has_nhm_msrs || no_msr)
8199 		goto guess;
8200 
8201 	if (get_msr(master_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
8202 		goto guess;
8203 
8204 	tcc_default = (msr >> 16) & 0xFF;
8205 
8206 	if (!quiet) {
8207 		int bits = platform->tcc_offset_bits;
8208 		unsigned long long enabled = 0;
8209 
8210 		if (bits && !get_msr(master_cpu, MSR_PLATFORM_INFO, &enabled))
8211 			enabled = (enabled >> 30) & 1;
8212 
8213 		if (bits && enabled) {
8214 			tcc_offset = (msr >> 24) & GENMASK(bits - 1, 0);
8215 			fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C) (%d default - %d offset)\n",
8216 				cpu, msr, tcc_default - tcc_offset, tcc_default, tcc_offset);
8217 		} else {
8218 			fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n", cpu, msr, tcc_default);
8219 		}
8220 	}
8221 
8222 	if (!tcc_default)
8223 		goto guess;
8224 
8225 	tj_max = tcc_default;
8226 
8227 	return 0;
8228 
8229 guess:
8230 	tj_max = TJMAX_DEFAULT;
8231 	fprintf(outf, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n", cpu, tj_max);
8232 
8233 	return 0;
8234 }
8235 
print_thermal(PER_THREAD_PARAMS)8236 int print_thermal(PER_THREAD_PARAMS)
8237 {
8238 	unsigned long long msr;
8239 	unsigned int dts, dts2;
8240 	int cpu;
8241 
8242 	UNUSED(c);
8243 	UNUSED(p);
8244 
8245 	if (no_msr)
8246 		return 0;
8247 
8248 	if (!(do_dts || do_ptm))
8249 		return 0;
8250 
8251 	cpu = t->cpu_id;
8252 
8253 	/* DTS is per-core, no need to print for each thread */
8254 	if (!is_cpu_first_thread_in_core(t, c))
8255 		return 0;
8256 
8257 	if (cpu_migrate(cpu)) {
8258 		fprintf(outf, "print_thermal: Could not migrate to CPU %d\n", cpu);
8259 		return -1;
8260 	}
8261 
8262 	if (do_ptm && is_cpu_first_core_in_package(t, p)) {
8263 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
8264 			return 0;
8265 
8266 		dts = (msr >> 16) & 0x7F;
8267 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n", cpu, msr, tj_max - dts);
8268 
8269 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
8270 			return 0;
8271 
8272 		dts = (msr >> 16) & 0x7F;
8273 		dts2 = (msr >> 8) & 0x7F;
8274 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n", cpu, msr, tj_max - dts, tj_max - dts2);
8275 	}
8276 
8277 	if (do_dts && debug) {
8278 		unsigned int resolution;
8279 
8280 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
8281 			return 0;
8282 
8283 		dts = (msr >> 16) & 0x7F;
8284 		resolution = (msr >> 27) & 0xF;
8285 		fprintf(outf, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n", cpu, msr, tj_max - dts, resolution);
8286 
8287 		if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
8288 			return 0;
8289 
8290 		dts = (msr >> 16) & 0x7F;
8291 		dts2 = (msr >> 8) & 0x7F;
8292 		fprintf(outf, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n", cpu, msr, tj_max - dts, tj_max - dts2);
8293 	}
8294 
8295 	return 0;
8296 }
8297 
probe_thermal(void)8298 void probe_thermal(void)
8299 {
8300 	if (!access("/sys/devices/system/cpu/cpu0/thermal_throttle/core_throttle_count", R_OK))
8301 		BIC_PRESENT(BIC_CORE_THROT_CNT);
8302 	else
8303 		BIC_NOT_PRESENT(BIC_CORE_THROT_CNT);
8304 
8305 	for_all_cpus(set_temperature_target, ODD_COUNTERS);
8306 
8307 	if (quiet)
8308 		return;
8309 
8310 	for_all_cpus(print_thermal, ODD_COUNTERS);
8311 }
8312 
get_cpu_type(PER_THREAD_PARAMS)8313 int get_cpu_type(PER_THREAD_PARAMS)
8314 {
8315 	unsigned int eax, ebx, ecx, edx;
8316 
8317 	UNUSED(c);
8318 	UNUSED(p);
8319 
8320 	if (!genuine_intel)
8321 		return 0;
8322 
8323 	if (cpu_migrate(t->cpu_id)) {
8324 		fprintf(outf, "Could not migrate to CPU %d\n", t->cpu_id);
8325 		return -1;
8326 	}
8327 
8328 	if (max_level < 0x1a)
8329 		return 0;
8330 
8331 	__cpuid(0x1a, eax, ebx, ecx, edx);
8332 	eax = (eax >> 24) & 0xFF;
8333 	if (eax == 0x20)
8334 		t->is_atom = true;
8335 	return 0;
8336 }
8337 
decode_feature_control_msr(void)8338 void decode_feature_control_msr(void)
8339 {
8340 	unsigned long long msr;
8341 
8342 	if (no_msr)
8343 		return;
8344 
8345 	if (quiet)
8346 		return;
8347 
8348 	if (!get_msr(master_cpu, MSR_IA32_FEAT_CTL, &msr))
8349 		fprintf(outf, "cpu%d: MSR_IA32_FEATURE_CONTROL: 0x%08llx (%sLocked %s)\n",
8350 			master_cpu, msr, msr & FEAT_CTL_LOCKED ? "" : "UN-", msr & (1 << 18) ? "SGX" : "");
8351 }
8352 
decode_misc_enable_msr(void)8353 void decode_misc_enable_msr(void)
8354 {
8355 	unsigned long long msr;
8356 
8357 	if (no_msr)
8358 		return;
8359 
8360 	if (!genuine_intel)
8361 		return;
8362 
8363 	if (!get_msr(master_cpu, MSR_IA32_MISC_ENABLE, &msr))
8364 		fprintf(outf, "cpu%d: MSR_IA32_MISC_ENABLE: 0x%08llx (%sTCC %sEIST %sMWAIT %sPREFETCH %sTURBO)\n",
8365 			master_cpu, msr,
8366 			msr & MSR_IA32_MISC_ENABLE_TM1 ? "" : "No-",
8367 			msr & MSR_IA32_MISC_ENABLE_ENHANCED_SPEEDSTEP ? "" : "No-",
8368 			msr & MSR_IA32_MISC_ENABLE_MWAIT ? "" : "No-",
8369 			msr & MSR_IA32_MISC_ENABLE_PREFETCH_DISABLE ? "No-" : "", msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ? "No-" : "");
8370 }
8371 
decode_misc_feature_control(void)8372 void decode_misc_feature_control(void)
8373 {
8374 	unsigned long long msr;
8375 
8376 	if (no_msr)
8377 		return;
8378 
8379 	if (!platform->has_msr_misc_feature_control)
8380 		return;
8381 
8382 	if (!get_msr(master_cpu, MSR_MISC_FEATURE_CONTROL, &msr))
8383 		fprintf(outf,
8384 			"cpu%d: MSR_MISC_FEATURE_CONTROL: 0x%08llx (%sL2-Prefetch %sL2-Prefetch-pair %sL1-Prefetch %sL1-IP-Prefetch)\n",
8385 			master_cpu, msr, msr & (0 << 0) ? "No-" : "", msr & (1 << 0) ? "No-" : "", msr & (2 << 0) ? "No-" : "", msr & (3 << 0) ? "No-" : "");
8386 }
8387 
8388 /*
8389  * Decode MSR_MISC_PWR_MGMT
8390  *
8391  * Decode the bits according to the Nehalem documentation
8392  * bit[0] seems to continue to have same meaning going forward
8393  * bit[1] less so...
8394  */
decode_misc_pwr_mgmt_msr(void)8395 void decode_misc_pwr_mgmt_msr(void)
8396 {
8397 	unsigned long long msr;
8398 
8399 	if (no_msr)
8400 		return;
8401 
8402 	if (!platform->has_msr_misc_pwr_mgmt)
8403 		return;
8404 
8405 	if (!get_msr(master_cpu, MSR_MISC_PWR_MGMT, &msr))
8406 		fprintf(outf, "cpu%d: MSR_MISC_PWR_MGMT: 0x%08llx (%sable-EIST_Coordination %sable-EPB %sable-OOB)\n",
8407 			master_cpu, msr, msr & (1 << 0) ? "DIS" : "EN", msr & (1 << 1) ? "EN" : "DIS", msr & (1 << 8) ? "EN" : "DIS");
8408 }
8409 
8410 /*
8411  * Decode MSR_CC6_DEMOTION_POLICY_CONFIG, MSR_MC6_DEMOTION_POLICY_CONFIG
8412  *
8413  * This MSRs are present on Silvermont processors,
8414  * Intel Atom processor E3000 series (Baytrail), and friends.
8415  */
decode_c6_demotion_policy_msr(void)8416 void decode_c6_demotion_policy_msr(void)
8417 {
8418 	unsigned long long msr;
8419 
8420 	if (no_msr)
8421 		return;
8422 
8423 	if (!platform->has_msr_c6_demotion_policy_config)
8424 		return;
8425 
8426 	if (!get_msr(master_cpu, MSR_CC6_DEMOTION_POLICY_CONFIG, &msr))
8427 		fprintf(outf, "cpu%d: MSR_CC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-CC6-Demotion)\n", master_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
8428 
8429 	if (!get_msr(master_cpu, MSR_MC6_DEMOTION_POLICY_CONFIG, &msr))
8430 		fprintf(outf, "cpu%d: MSR_MC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-MC6-Demotion)\n", master_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
8431 }
8432 
print_dev_latency(void)8433 void print_dev_latency(void)
8434 {
8435 	char *path = "/dev/cpu_dma_latency";
8436 	int fd;
8437 	int value;
8438 	int retval;
8439 
8440 	fd = open(path, O_RDONLY);
8441 	if (fd < 0) {
8442 		if (debug)
8443 			warnx("Read %s failed", path);
8444 		return;
8445 	}
8446 
8447 	retval = read(fd, (void *)&value, sizeof(int));
8448 	if (retval != sizeof(int)) {
8449 		warn("read failed %s", path);
8450 		close(fd);
8451 		return;
8452 	}
8453 	fprintf(outf, "/dev/cpu_dma_latency: %d usec (%s)\n", value, value == 2000000000 ? "default" : "constrained");
8454 
8455 	close(fd);
8456 }
8457 
has_perf_instr_count_access(void)8458 static int has_perf_instr_count_access(void)
8459 {
8460 	int fd;
8461 
8462 	if (no_perf)
8463 		return 0;
8464 
8465 	fd = open_perf_counter(master_cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS, -1, 0);
8466 	if (fd != -1)
8467 		close(fd);
8468 
8469 	if (fd == -1)
8470 		warnx("Failed to access %s. Some of the counters may not be available\n"
8471 		      "\tRun as root to enable them or use %s to disable the access explicitly", "perf instructions retired counter",
8472 		      "'--hide IPC' or '--no-perf'");
8473 
8474 	return (fd != -1);
8475 }
8476 
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_)8477 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_)
8478 {
8479 	int ret = -1;
8480 
8481 	if (no_perf)
8482 		return -1;
8483 
8484 	if (!cai->perf_name)
8485 		return -1;
8486 
8487 	const double scale = read_perf_scale(cai->perf_subsys, cai->perf_name);
8488 
8489 	if (scale == 0.0)
8490 		goto end;
8491 
8492 	const enum rapl_unit unit = read_perf_rapl_unit(cai->perf_subsys, cai->perf_name);
8493 
8494 	if (unit == RAPL_UNIT_INVALID)
8495 		goto end;
8496 
8497 	const unsigned int rapl_type = read_perf_type(cai->perf_subsys);
8498 	const unsigned int rapl_energy_pkg_config = read_perf_config(cai->perf_subsys, cai->perf_name);
8499 
8500 	ret = open_perf_counter(cpu, rapl_type, rapl_energy_pkg_config, rci->fd_perf, PERF_FORMAT_GROUP);
8501 	if (ret == -1)
8502 		goto end;
8503 
8504 	/* If it's the first counter opened, make it a group descriptor */
8505 	if (rci->fd_perf == -1)
8506 		rci->fd_perf = ret;
8507 
8508 	*scale_ = scale;
8509 	*unit_ = unit;
8510 
8511 end:
8512 	if (debug >= 2)
8513 		fprintf(stderr, "%s: %d (cpu: %d)\n", __func__, ret, cpu);
8514 
8515 	return ret;
8516 }
8517 
8518 char cpuset_buf[1024];
initialize_cpu_set_from_sysfs(cpu_set_t * cpu_set,char * sysfs_path,char * sysfs_file)8519 int initialize_cpu_set_from_sysfs(cpu_set_t *cpu_set, char *sysfs_path, char *sysfs_file)
8520 {
8521 	FILE *fp;
8522 	char path[128];
8523 
8524 	if (snprintf(path, 128, "%s/%s", sysfs_path, sysfs_file) > 128)
8525 		err(-1, "%s %s", sysfs_path, sysfs_file);
8526 
8527 	fp = fopen(path, "r");
8528 	if (!fp) {
8529 		warn("open %s", path);
8530 		return -1;
8531 	}
8532 	if (fread(cpuset_buf, sizeof(char), 1024, fp) == 0) {
8533 		warn("read %s", sysfs_path);
8534 		goto err;
8535 	}
8536 	if (parse_cpu_str(cpuset_buf, cpu_set, cpu_possible_setsize)) {
8537 		warnx("%s: cpu str malformat %s\n", sysfs_path, cpu_effective_str);
8538 		goto err;
8539 	}
8540 	return 0;
8541 
8542 err:
8543 	fclose(fp);
8544 	return -1;
8545 }
8546 
print_cpu_set(char * s,cpu_set_t * set)8547 void print_cpu_set(char *s, cpu_set_t *set)
8548 {
8549 	int i;
8550 
8551 	assert(MAX_BIC < CPU_SETSIZE);
8552 
8553 	printf("%s:", s);
8554 
8555 	for (i = 0; i <= topo.max_cpu_num; ++i)
8556 		if (CPU_ISSET(i, set))
8557 			printf(" %d", i);
8558 	putchar('\n');
8559 }
8560 
linux_perf_init_hybrid_cpus(void)8561 void linux_perf_init_hybrid_cpus(void)
8562 {
8563 	char *perf_cpu_pcore_path = "/sys/devices/cpu_core";
8564 	char *perf_cpu_ecore_path = "/sys/devices/cpu_atom";
8565 	char *perf_cpu_lcore_path = "/sys/devices/cpu_lowpower";
8566 	char path[128];
8567 
8568 	if (!access(perf_cpu_pcore_path, F_OK)) {
8569 		perf_pcore_set = CPU_ALLOC((topo.max_cpu_num + 1));
8570 		if (perf_pcore_set == NULL)
8571 			err(3, "CPU_ALLOC");
8572 		CPU_ZERO_S(cpu_possible_setsize, perf_pcore_set);
8573 		initialize_cpu_set_from_sysfs(perf_pcore_set, perf_cpu_pcore_path, "cpus");
8574 		if (debug)
8575 			print_cpu_set("perf pcores", perf_pcore_set);
8576 		sprintf(path, "%s/%s", perf_cpu_pcore_path, "type");
8577 		perf_pmu_types.pcore = snapshot_sysfs_counter(path);
8578 	}
8579 
8580 	if (!access(perf_cpu_ecore_path, F_OK)) {
8581 		perf_ecore_set = CPU_ALLOC((topo.max_cpu_num + 1));
8582 		if (perf_ecore_set == NULL)
8583 			err(3, "CPU_ALLOC");
8584 		CPU_ZERO_S(cpu_possible_setsize, perf_ecore_set);
8585 		initialize_cpu_set_from_sysfs(perf_ecore_set, perf_cpu_ecore_path, "cpus");
8586 		if (debug)
8587 			print_cpu_set("perf ecores", perf_ecore_set);
8588 		sprintf(path, "%s/%s", perf_cpu_ecore_path, "type");
8589 		perf_pmu_types.ecore = snapshot_sysfs_counter(path);
8590 	}
8591 
8592 	if (!access(perf_cpu_lcore_path, F_OK)) {
8593 		perf_lcore_set = CPU_ALLOC((topo.max_cpu_num + 1));
8594 		if (perf_lcore_set == NULL)
8595 			err(3, "CPU_ALLOC");
8596 		CPU_ZERO_S(cpu_possible_setsize, perf_lcore_set);
8597 		initialize_cpu_set_from_sysfs(perf_lcore_set, perf_cpu_lcore_path, "cpus");
8598 		if (debug)
8599 			print_cpu_set("perf lcores", perf_lcore_set);
8600 		sprintf(path, "%s/%s", perf_cpu_lcore_path, "type");
8601 		perf_pmu_types.lcore = snapshot_sysfs_counter(path);
8602 	}
8603 }
8604 
8605 /*
8606  * Linux-perf related initialization
8607  */
linux_perf_init(void)8608 void linux_perf_init(void)
8609 {
8610 	char path[128];
8611 	char *perf_cpu_path = "/sys/devices/cpu";
8612 
8613 	if (access("/proc/sys/kernel/perf_event_paranoid", F_OK))
8614 		return;
8615 
8616 	if (!access(perf_cpu_path, F_OK)) {
8617 		sprintf(path, "%s/%s", perf_cpu_path, "type");
8618 		perf_pmu_types.uniform = snapshot_sysfs_counter(path);
8619 	} else {
8620 		linux_perf_init_hybrid_cpus();
8621 	}
8622 
8623 	if (BIC_IS_ENABLED(BIC_IPC) && cpuid_has_aperf_mperf) {
8624 		fd_instr_count_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8625 		if (fd_instr_count_percpu == NULL)
8626 			err(-1, "calloc fd_instr_count_percpu");
8627 	}
8628 	if (BIC_IS_ENABLED(BIC_LLC_MRPS) || BIC_IS_ENABLED(BIC_LLC_HIT)) {
8629 		fd_llc_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8630 		if (fd_llc_percpu == NULL)
8631 			err(-1, "calloc fd_llc_percpu");
8632 	}
8633 	if (BIC_IS_ENABLED(BIC_L2_MRPS) || BIC_IS_ENABLED(BIC_L2_HIT)) {
8634 		fd_l2_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
8635 		if (fd_l2_percpu == NULL)
8636 			err(-1, "calloc fd_l2_percpu");
8637 	}
8638 }
8639 
rapl_perf_init(void)8640 void rapl_perf_init(void)
8641 {
8642 	const unsigned int num_domains = get_rapl_num_domains();
8643 	bool *domain_visited = calloc(num_domains, sizeof(bool));
8644 
8645 	rapl_counter_info_perdomain = calloc(num_domains, sizeof(*rapl_counter_info_perdomain));
8646 	if (rapl_counter_info_perdomain == NULL)
8647 		err(-1, "calloc rapl_counter_info_percpu");
8648 	rapl_counter_info_perdomain_size = num_domains;
8649 
8650 	/*
8651 	 * Initialize rapl_counter_info_percpu
8652 	 */
8653 	for (unsigned int domain_id = 0; domain_id < num_domains; ++domain_id) {
8654 		struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[domain_id];
8655 
8656 		rci->fd_perf = -1;
8657 		for (size_t i = 0; i < NUM_RAPL_COUNTERS; ++i) {
8658 			rci->data[i] = 0;
8659 			rci->source[i] = COUNTER_SOURCE_NONE;
8660 		}
8661 	}
8662 
8663 	/*
8664 	 * Open/probe the counters
8665 	 * If can't get it via perf, fallback to MSR
8666 	 */
8667 	for (size_t i = 0; i < ARRAY_SIZE(rapl_counter_arch_infos); ++i) {
8668 
8669 		const struct rapl_counter_arch_info *const cai = &rapl_counter_arch_infos[i];
8670 		bool has_counter = 0;
8671 		double scale;
8672 		enum rapl_unit unit;
8673 		unsigned int next_domain;
8674 
8675 		if (!BIC_IS_ENABLED(cai->bic_number))
8676 			continue;
8677 
8678 		memset(domain_visited, 0, num_domains * sizeof(*domain_visited));
8679 
8680 		for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
8681 
8682 			if (cpu_is_not_allowed(cpu))
8683 				continue;
8684 
8685 			/* Skip already seen and handled RAPL domains */
8686 			next_domain = get_rapl_domain_id(cpu);
8687 
8688 			assert(next_domain < num_domains);
8689 
8690 			if (domain_visited[next_domain])
8691 				continue;
8692 
8693 			domain_visited[next_domain] = 1;
8694 
8695 			if ((cai->flags & RAPL_COUNTER_FLAG_PLATFORM_COUNTER) && (cpu != master_cpu))
8696 				continue;
8697 
8698 			struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[next_domain];
8699 
8700 			/*
8701 			 * rapl_counter_arch_infos[] can have multiple entries describing the same
8702 			 * counter, due to the difference from different platforms/Vendors.
8703 			 * E.g. rapl_counter_arch_infos[0] and rapl_counter_arch_infos[1] share the
8704 			 * same perf_subsys and perf_name, but with different MSR address.
8705 			 * rapl_counter_arch_infos[0] is for Intel and rapl_counter_arch_infos[1]
8706 			 * is for AMD.
8707 			 * In this case, it is possible that multiple rapl_counter_arch_infos[]
8708 			 * entries are probed just because their perf/msr is duplicate and valid.
8709 			 *
8710 			 * Thus need a check to avoid re-probe the same counters.
8711 			 */
8712 			if (rci->source[cai->rci_index] != COUNTER_SOURCE_NONE)
8713 				break;
8714 
8715 			/* Use perf API for this counter */
8716 			if (add_rapl_perf_counter(cpu, rci, cai, &scale, &unit) != -1) {
8717 				rci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
8718 				rci->scale[cai->rci_index] = scale * cai->compat_scale;
8719 				rci->unit[cai->rci_index] = unit;
8720 				rci->flags[cai->rci_index] = cai->flags;
8721 
8722 				/* Use MSR for this counter */
8723 			} else if (add_rapl_msr_counter(cpu, cai) >= 0) {
8724 				rci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
8725 				rci->msr[cai->rci_index] = cai->msr;
8726 				rci->msr_mask[cai->rci_index] = cai->msr_mask;
8727 				rci->msr_shift[cai->rci_index] = cai->msr_shift;
8728 				rci->unit[cai->rci_index] = RAPL_UNIT_JOULES;
8729 				rci->scale[cai->rci_index] = *cai->platform_rapl_msr_scale * cai->compat_scale;
8730 				rci->flags[cai->rci_index] = cai->flags;
8731 			}
8732 
8733 			if (rci->source[cai->rci_index] != COUNTER_SOURCE_NONE)
8734 				has_counter = 1;
8735 		}
8736 
8737 		/* If any CPU has access to the counter, make it present */
8738 		if (has_counter)
8739 			BIC_PRESENT(cai->bic_number);
8740 	}
8741 
8742 	free(domain_visited);
8743 }
8744 
8745 /* Assumes msr_counter_info is populated */
has_amperf_access(void)8746 static int has_amperf_access(void)
8747 {
8748 	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;
8749 }
8750 
get_cstate_perf_group_fd(struct cstate_counter_info_t * cci,const char * group_name)8751 int *get_cstate_perf_group_fd(struct cstate_counter_info_t *cci, const char *group_name)
8752 {
8753 	if (strcmp(group_name, "cstate_core") == 0)
8754 		return &cci->fd_perf_core;
8755 
8756 	if (strcmp(group_name, "cstate_pkg") == 0)
8757 		return &cci->fd_perf_pkg;
8758 
8759 	return NULL;
8760 }
8761 
add_cstate_perf_counter(int cpu,struct cstate_counter_info_t * cci,const struct cstate_counter_arch_info * cai)8762 int add_cstate_perf_counter(int cpu, struct cstate_counter_info_t *cci, const struct cstate_counter_arch_info *cai)
8763 {
8764 	int ret = -1;
8765 
8766 	if (no_perf)
8767 		return -1;
8768 
8769 	if (!cai->perf_name)
8770 		return -1;
8771 
8772 	int *pfd_group = get_cstate_perf_group_fd(cci, cai->perf_subsys);
8773 
8774 	if (pfd_group == NULL)
8775 		goto end;
8776 
8777 	const unsigned int type = read_perf_type(cai->perf_subsys);
8778 	const unsigned int config = read_perf_config(cai->perf_subsys, cai->perf_name);
8779 
8780 	ret = open_perf_counter(cpu, type, config, *pfd_group, PERF_FORMAT_GROUP);
8781 
8782 	if (ret == -1)
8783 		goto end;
8784 
8785 	/* If it's the first counter opened, make it a group descriptor */
8786 	if (*pfd_group == -1)
8787 		*pfd_group = ret;
8788 
8789 end:
8790 	if (debug >= 2)
8791 		fprintf(stderr, "%s: %d (cpu: %d)\n", __func__, ret, cpu);
8792 
8793 	return ret;
8794 }
8795 
add_msr_perf_counter(int cpu,struct msr_counter_info_t * cci,const struct msr_counter_arch_info * cai)8796 int add_msr_perf_counter(int cpu, struct msr_counter_info_t *cci, const struct msr_counter_arch_info *cai)
8797 {
8798 	int ret = -1;
8799 
8800 	if (no_perf)
8801 		return -1;
8802 
8803 	if (!cai->perf_name)
8804 		return -1;
8805 
8806 	const unsigned int type = read_perf_type(cai->perf_subsys);
8807 	const unsigned int config = read_perf_config(cai->perf_subsys, cai->perf_name);
8808 
8809 	ret = open_perf_counter(cpu, type, config, cci->fd_perf, PERF_FORMAT_GROUP);
8810 
8811 	if (ret == -1)
8812 		goto end;
8813 
8814 	/* If it's the first counter opened, make it a group descriptor */
8815 	if (cci->fd_perf == -1)
8816 		cci->fd_perf = ret;
8817 
8818 end:
8819 	if (debug)
8820 		fprintf(stderr, "%s: %s/%s: %d (cpu: %d)\n", __func__, cai->perf_subsys, cai->perf_name, ret, cpu);
8821 
8822 	return ret;
8823 }
8824 
msr_perf_init_(void)8825 void msr_perf_init_(void)
8826 {
8827 	const int mci_num = topo.max_cpu_num + 1;
8828 
8829 	msr_counter_info = calloc(mci_num, sizeof(*msr_counter_info));
8830 	if (!msr_counter_info)
8831 		err(1, "calloc msr_counter_info");
8832 	msr_counter_info_size = mci_num;
8833 
8834 	for (int cpu = 0; cpu < mci_num; ++cpu)
8835 		msr_counter_info[cpu].fd_perf = -1;
8836 
8837 	for (int cidx = 0; cidx < NUM_MSR_COUNTERS; ++cidx) {
8838 
8839 		struct msr_counter_arch_info *cai = &msr_counter_arch_infos[cidx];
8840 
8841 		cai->present = false;
8842 
8843 		for (int cpu = 0; cpu < mci_num; ++cpu) {
8844 
8845 			struct msr_counter_info_t *const cci = &msr_counter_info[cpu];
8846 
8847 			if (cpu_is_not_allowed(cpu))
8848 				continue;
8849 
8850 			if (cai->needed) {
8851 				/* Use perf API for this counter */
8852 				if (add_msr_perf_counter(cpu, cci, cai) != -1) {
8853 					cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
8854 					cai->present = true;
8855 
8856 					/* User MSR for this counter */
8857 				} else if (add_msr_counter(cpu, cai->msr) >= 0) {
8858 					cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
8859 					cci->msr[cai->rci_index] = cai->msr;
8860 					cci->msr_mask[cai->rci_index] = cai->msr_mask;
8861 					cai->present = true;
8862 				}
8863 			}
8864 		}
8865 	}
8866 }
8867 
8868 /* Initialize data for reading perf counters from the MSR group. */
msr_perf_init(void)8869 void msr_perf_init(void)
8870 {
8871 	bool need_amperf = false, need_smi = false;
8872 	const bool need_soft_c1 = (!platform->has_msr_core_c1_res) && (platform->supported_cstates & CC1);
8873 
8874 	need_amperf = BIC_IS_ENABLED(BIC_Avg_MHz) || BIC_IS_ENABLED(BIC_Busy) || BIC_IS_ENABLED(BIC_Bzy_MHz)
8875 	    || BIC_IS_ENABLED(BIC_IPC) || need_soft_c1;
8876 
8877 	if (BIC_IS_ENABLED(BIC_SMI))
8878 		need_smi = true;
8879 
8880 	/* Enable needed counters */
8881 	msr_counter_arch_infos[MSR_ARCH_INFO_APERF_INDEX].needed = need_amperf;
8882 	msr_counter_arch_infos[MSR_ARCH_INFO_MPERF_INDEX].needed = need_amperf;
8883 	msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].needed = need_smi;
8884 
8885 	msr_perf_init_();
8886 
8887 	const bool has_amperf = has_amperf_access();
8888 	const bool has_smi = msr_counter_arch_infos[MSR_ARCH_INFO_SMI_INDEX].present;
8889 
8890 	has_aperf_access = has_amperf;
8891 
8892 	if (has_amperf) {
8893 		BIC_PRESENT(BIC_Avg_MHz);
8894 		BIC_PRESENT(BIC_Busy);
8895 		BIC_PRESENT(BIC_Bzy_MHz);
8896 		BIC_PRESENT(BIC_SMI);
8897 	}
8898 
8899 	if (has_smi)
8900 		BIC_PRESENT(BIC_SMI);
8901 }
8902 
cstate_perf_init_(bool soft_c1)8903 void cstate_perf_init_(bool soft_c1)
8904 {
8905 	bool has_counter;
8906 	bool *cores_visited = NULL, *pkg_visited = NULL;
8907 	const int cores_visited_elems = topo.max_core_id + 1;
8908 	const int pkg_visited_elems = topo.max_package_id + 1;
8909 	const int cci_num = topo.max_cpu_num + 1;
8910 
8911 	ccstate_counter_info = calloc(cci_num, sizeof(*ccstate_counter_info));
8912 	if (!ccstate_counter_info)
8913 		err(1, "calloc ccstate_counter_arch_info");
8914 	ccstate_counter_info_size = cci_num;
8915 
8916 	cores_visited = calloc(cores_visited_elems, sizeof(*cores_visited));
8917 	if (!cores_visited)
8918 		err(1, "calloc cores_visited");
8919 
8920 	pkg_visited = calloc(pkg_visited_elems, sizeof(*pkg_visited));
8921 	if (!pkg_visited)
8922 		err(1, "calloc pkg_visited");
8923 
8924 	/* Initialize cstate_counter_info_percpu */
8925 	for (int cpu = 0; cpu < cci_num; ++cpu) {
8926 		ccstate_counter_info[cpu].fd_perf_core = -1;
8927 		ccstate_counter_info[cpu].fd_perf_pkg = -1;
8928 	}
8929 
8930 	for (int cidx = 0; cidx < NUM_CSTATE_COUNTERS; ++cidx) {
8931 		has_counter = false;
8932 		memset(cores_visited, 0, cores_visited_elems * sizeof(*cores_visited));
8933 		memset(pkg_visited, 0, pkg_visited_elems * sizeof(*pkg_visited));
8934 
8935 		const struct cstate_counter_arch_info *cai = &ccstate_counter_arch_infos[cidx];
8936 
8937 		for (int cpu = 0; cpu < cci_num; ++cpu) {
8938 
8939 			struct cstate_counter_info_t *const cci = &ccstate_counter_info[cpu];
8940 
8941 			if (cpu_is_not_allowed(cpu))
8942 				continue;
8943 
8944 			const int core_id = cpus[cpu].core_id;
8945 			const int pkg_id = cpus[cpu].package_id;
8946 
8947 			assert(core_id < cores_visited_elems);
8948 			assert(pkg_id < pkg_visited_elems);
8949 
8950 			const bool per_thread = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_THREAD;
8951 			const bool per_core = cai->flags & CSTATE_COUNTER_FLAG_COLLECT_PER_CORE;
8952 
8953 			if (!per_thread && cores_visited[core_id])
8954 				continue;
8955 
8956 			if (!per_core && pkg_visited[pkg_id])
8957 				continue;
8958 
8959 			const bool counter_needed = BIC_IS_ENABLED(cai->bic_number) || (soft_c1 && (cai->flags & CSTATE_COUNTER_FLAG_SOFT_C1_DEPENDENCY));
8960 			const bool counter_supported = (platform->supported_cstates & cai->feature_mask);
8961 
8962 			if (counter_needed && counter_supported) {
8963 				/* Use perf API for this counter */
8964 				if (add_cstate_perf_counter(cpu, cci, cai) != -1) {
8965 
8966 					cci->source[cai->rci_index] = COUNTER_SOURCE_PERF;
8967 
8968 					/* User MSR for this counter */
8969 				} else if (pkg_cstate_limit >= cai->pkg_cstate_limit && add_msr_counter(cpu, cai->msr) >= 0) {
8970 					cci->source[cai->rci_index] = COUNTER_SOURCE_MSR;
8971 					cci->msr[cai->rci_index] = cai->msr;
8972 				}
8973 			}
8974 
8975 			if (cci->source[cai->rci_index] != COUNTER_SOURCE_NONE) {
8976 				has_counter = true;
8977 				cores_visited[core_id] = true;
8978 				pkg_visited[pkg_id] = true;
8979 			}
8980 		}
8981 
8982 		/* If any CPU has access to the counter, make it present */
8983 		if (has_counter)
8984 			BIC_PRESENT(cai->bic_number);
8985 	}
8986 
8987 	free(cores_visited);
8988 	free(pkg_visited);
8989 }
8990 
cstate_perf_init(void)8991 void cstate_perf_init(void)
8992 {
8993 	/*
8994 	 * If we don't have a C1 residency MSR, we calculate it "in software",
8995 	 * but we need APERF, MPERF too.
8996 	 */
8997 	const bool soft_c1 = !platform->has_msr_core_c1_res && has_amperf_access()
8998 	    && platform->supported_cstates & CC1;
8999 
9000 	if (soft_c1)
9001 		BIC_PRESENT(BIC_CPU_c1);
9002 
9003 	cstate_perf_init_(soft_c1);
9004 }
9005 
probe_cstates(void)9006 void probe_cstates(void)
9007 {
9008 	probe_cst_limit();
9009 
9010 	if (platform->has_msr_module_c6_res_ms)
9011 		BIC_PRESENT(BIC_Mod_c6);
9012 
9013 	if (platform->has_ext_cst_msrs && !no_msr) {
9014 		BIC_PRESENT(BIC_Totl_c0);
9015 		BIC_PRESENT(BIC_Any_c0);
9016 		BIC_PRESENT(BIC_GFX_c0);
9017 		BIC_PRESENT(BIC_CPUGFX);
9018 	}
9019 
9020 	if (quiet)
9021 		return;
9022 
9023 	dump_power_ctl();
9024 	dump_cst_cfg();
9025 	decode_c6_demotion_policy_msr();
9026 	print_dev_latency();
9027 	dump_sysfs_cstate_config();
9028 	print_irtl();
9029 }
9030 
probe_lpi(void)9031 void probe_lpi(void)
9032 {
9033 	if (!access("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", R_OK))
9034 		BIC_PRESENT(BIC_CPU_LPI);
9035 	else
9036 		BIC_NOT_PRESENT(BIC_CPU_LPI);
9037 
9038 	if (!access(sys_lpi_file_sysfs, R_OK)) {
9039 		sys_lpi_file = sys_lpi_file_sysfs;
9040 		BIC_PRESENT(BIC_SYS_LPI);
9041 	} else if (!access(sys_lpi_file_debugfs, R_OK)) {
9042 		sys_lpi_file = sys_lpi_file_debugfs;
9043 		BIC_PRESENT(BIC_SYS_LPI);
9044 	} else {
9045 		sys_lpi_file_sysfs = NULL;
9046 		BIC_NOT_PRESENT(BIC_SYS_LPI);
9047 	}
9048 
9049 }
9050 
probe_pstates(void)9051 void probe_pstates(void)
9052 {
9053 	probe_bclk();
9054 
9055 	if (quiet)
9056 		return;
9057 
9058 	dump_platform_info();
9059 	dump_turbo_ratio_info();
9060 	dump_sysfs_pstate_config();
9061 	decode_misc_pwr_mgmt_msr();
9062 
9063 	for_all_cpus(print_hwp, ODD_COUNTERS);
9064 	for_all_cpus(print_epb, ODD_COUNTERS);
9065 	for_all_cpus(print_perf_limit, ODD_COUNTERS);
9066 }
9067 
dump_word_chars(unsigned int word)9068 void dump_word_chars(unsigned int word)
9069 {
9070 	int i;
9071 
9072 	for (i = 0; i < 4; ++i)
9073 		fprintf(outf, "%c", (word >> (i * 8)) & 0xFF);
9074 }
9075 
dump_cpuid_hypervisor(void)9076 void dump_cpuid_hypervisor(void)
9077 {
9078 	unsigned int ebx = 0;
9079 	unsigned int ecx = 0;
9080 	unsigned int edx = 0;
9081 
9082 	__cpuid(0x40000000, max_extended_level, ebx, ecx, edx);
9083 
9084 	fprintf(outf, "Hypervisor: ");
9085 	dump_word_chars(ebx);
9086 	dump_word_chars(ecx);
9087 	dump_word_chars(edx);
9088 	fprintf(outf, "\n");
9089 }
9090 
process_cpuid()9091 void process_cpuid()
9092 {
9093 	unsigned int eax, ebx, ecx, edx;
9094 	unsigned int fms, family, model, stepping, ecx_flags, edx_flags;
9095 	unsigned long long ucode_patch = 0;
9096 	bool ucode_patch_valid = false;
9097 
9098 	eax = ebx = ecx = edx = 0;
9099 
9100 	__cpuid(0, max_level, ebx, ecx, edx);
9101 
9102 	if (ebx == 0x756e6547 && ecx == 0x6c65746e && edx == 0x49656e69)
9103 		genuine_intel = 1;
9104 	else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65)
9105 		authentic_amd = 1;
9106 	else if (ebx == 0x6f677948 && ecx == 0x656e6975 && edx == 0x6e65476e)
9107 		hygon_genuine = 1;
9108 
9109 	if (!quiet)
9110 		fprintf(outf, "CPUID(0): %.4s%.4s%.4s 0x%x CPUID levels\n", (char *)&ebx, (char *)&edx, (char *)&ecx, max_level);
9111 
9112 	__cpuid(1, fms, ebx, ecx, edx);
9113 	family = (fms >> 8) & 0xf;
9114 	model = (fms >> 4) & 0xf;
9115 	stepping = fms & 0xf;
9116 	if (family == 0xf)
9117 		family += (fms >> 20) & 0xff;
9118 	if (family >= 6)
9119 		model += ((fms >> 16) & 0xf) << 4;
9120 	ecx_flags = ecx;
9121 	edx_flags = edx;
9122 	cpuid_has_hv = ecx_flags & (1 << 31);
9123 
9124 	if (!no_msr) {
9125 		if (get_msr(sched_getcpu(), MSR_IA32_UCODE_REV, &ucode_patch))
9126 			warnx("get_msr(UCODE)");
9127 		else
9128 			ucode_patch_valid = true;
9129 	}
9130 
9131 	/*
9132 	 * check max extended function levels of CPUID.
9133 	 * This is needed to check for invariant TSC.
9134 	 * This check is valid for both Intel and AMD.
9135 	 */
9136 	ebx = ecx = edx = 0;
9137 	__cpuid(0x80000000, max_extended_level, ebx, ecx, edx);
9138 
9139 	if (!quiet) {
9140 		fprintf(outf, "CPUID(1): family:model:stepping 0x%x:%x:%x (%d:%d:%d)", family, model, stepping, family, model, stepping);
9141 		if (ucode_patch_valid)
9142 			fprintf(outf, " microcode 0x%x", (unsigned int)((ucode_patch >> 32) & 0xFFFFFFFF));
9143 		fputc('\n', outf);
9144 
9145 		fprintf(outf, "CPUID(0x80000000): max_extended_levels: 0x%x\n", max_extended_level);
9146 		fprintf(outf, "CPUID(1): %sSSE3 %sMONITOR %sSMX %sEIST %sTM2 %sHV %sTSC %sMSR %sACPI-TM %sHT %sTM\n",
9147 			ecx_flags & (1 << 0) ? "" : "No-",
9148 			ecx_flags & (1 << 3) ? "" : "No-",
9149 			ecx_flags & (1 << 6) ? "" : "No-",
9150 			ecx_flags & (1 << 7) ? "" : "No-",
9151 			ecx_flags & (1 << 8) ? "" : "No-",
9152 			cpuid_has_hv ? "" : "No-",
9153 			edx_flags & (1 << 4) ? "" : "No-",
9154 			edx_flags & (1 << 5) ? "" : "No-",
9155 			edx_flags & (1 << 22) ? "" : "No-", edx_flags & (1 << 28) ? "" : "No-", edx_flags & (1 << 29) ? "" : "No-");
9156 	}
9157 	if (!quiet && cpuid_has_hv)
9158 		dump_cpuid_hypervisor();
9159 
9160 	probe_platform_features(family, model);
9161 	init_perf_model_support(family, model);
9162 
9163 	if (!(edx_flags & (1 << 5)))
9164 		errx(1, "CPUID: no MSR");
9165 
9166 	if (max_extended_level >= 0x80000007) {
9167 
9168 		/*
9169 		 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
9170 		 * this check is valid for both Intel and AMD
9171 		 */
9172 		__cpuid(0x80000007, eax, ebx, ecx, edx);
9173 		has_invariant_tsc = edx & (1 << 8);
9174 	}
9175 
9176 	/*
9177 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
9178 	 * this check is valid for both Intel and AMD
9179 	 */
9180 
9181 	__cpuid(0x6, eax, ebx, ecx, edx);
9182 	cpuid_has_aperf_mperf = ecx & (1 << 0);
9183 	do_dts = eax & (1 << 0);
9184 	if (do_dts)
9185 		BIC_PRESENT(BIC_CoreTmp);
9186 	has_turbo = eax & (1 << 1);
9187 	do_ptm = eax & (1 << 6);
9188 	if (do_ptm)
9189 		BIC_PRESENT(BIC_PkgTmp);
9190 	has_hwp = eax & (1 << 7);
9191 	has_hwp_notify = eax & (1 << 8);
9192 	has_hwp_activity_window = eax & (1 << 9);
9193 	has_hwp_epp = eax & (1 << 10);
9194 	has_hwp_pkg = eax & (1 << 11);
9195 	has_epb = ecx & (1 << 3);
9196 
9197 	if (!quiet)
9198 		fprintf(outf, "CPUID(6): %sAPERF, %sTURBO, %sDTS, %sPTM, %sHWP, "
9199 			"%sHWPnotify, %sHWPwindow, %sHWPepp, %sHWPpkg, %sEPB\n",
9200 			cpuid_has_aperf_mperf ? "" : "No-",
9201 			has_turbo ? "" : "No-",
9202 			do_dts ? "" : "No-",
9203 			do_ptm ? "" : "No-",
9204 			has_hwp ? "" : "No-",
9205 			has_hwp_notify ? "" : "No-",
9206 			has_hwp_activity_window ? "" : "No-", has_hwp_epp ? "" : "No-", has_hwp_pkg ? "" : "No-", has_epb ? "" : "No-");
9207 
9208 	if (!quiet)
9209 		decode_misc_enable_msr();
9210 
9211 	if (max_level >= 0x7) {
9212 		int has_sgx;
9213 
9214 		ecx = 0;
9215 
9216 		__cpuid_count(0x7, 0, eax, ebx, ecx, edx);
9217 
9218 		has_sgx = ebx & (1 << 2);
9219 
9220 		is_hybrid = !!(edx & (1 << 15));
9221 
9222 		if (!quiet)
9223 			fprintf(outf, "CPUID(7): %sSGX %sHybrid\n", has_sgx ? "" : "No-", is_hybrid ? "" : "No-");
9224 
9225 		if (has_sgx)
9226 			decode_feature_control_msr();
9227 	}
9228 
9229 	if (max_level >= 0x15) {
9230 		unsigned int eax_crystal;
9231 		unsigned int ebx_tsc;
9232 
9233 		/*
9234 		 * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
9235 		 */
9236 		eax_crystal = ebx_tsc = crystal_hz = edx = 0;
9237 		__cpuid(0x15, eax_crystal, ebx_tsc, crystal_hz, edx);
9238 
9239 		if (ebx_tsc != 0) {
9240 			if (!quiet && (ebx != 0))
9241 				fprintf(outf, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n", eax_crystal, ebx_tsc, crystal_hz);
9242 
9243 			if (crystal_hz == 0)
9244 				crystal_hz = platform->crystal_freq;
9245 
9246 			if (crystal_hz) {
9247 				tsc_hz = (unsigned long long)crystal_hz *ebx_tsc / eax_crystal;
9248 				if (!quiet)
9249 					fprintf(outf, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n", tsc_hz / 1000000, crystal_hz, ebx_tsc, eax_crystal);
9250 			}
9251 		}
9252 	}
9253 	if (max_level >= 0x16) {
9254 		unsigned int base_mhz, max_mhz, bus_mhz, edx;
9255 
9256 		/*
9257 		 * CPUID 16H Base MHz, Max MHz, Bus MHz
9258 		 */
9259 		base_mhz = max_mhz = bus_mhz = edx = 0;
9260 
9261 		__cpuid(0x16, base_mhz, max_mhz, bus_mhz, edx);
9262 
9263 		bclk = bus_mhz;
9264 
9265 		base_hz = base_mhz * 1000000;
9266 		has_base_hz = 1;
9267 
9268 		if (platform->enable_tsc_tweak)
9269 			tsc_tweak = base_hz / tsc_hz;
9270 
9271 		if (!quiet)
9272 			fprintf(outf, "CPUID(0x16): base_mhz: %d max_mhz: %d bus_mhz: %d\n", base_mhz, max_mhz, bus_mhz);
9273 	}
9274 
9275 	if (cpuid_has_aperf_mperf)
9276 		aperf_mperf_multiplier = platform->need_perf_multiplier ? 1024 : 1;
9277 
9278 	BIC_PRESENT(BIC_IRQ);
9279 	BIC_PRESENT(BIC_NMI);
9280 	BIC_PRESENT(BIC_TSC_MHz);
9281 }
9282 
counter_info_init(void)9283 static void counter_info_init(void)
9284 {
9285 	for (int i = 0; i < NUM_CSTATE_COUNTERS; ++i) {
9286 		struct cstate_counter_arch_info *const cai = &ccstate_counter_arch_infos[i];
9287 
9288 		if (platform->has_msr_knl_core_c6_residency && cai->msr == MSR_CORE_C6_RESIDENCY)
9289 			cai->msr = MSR_KNL_CORE_C6_RESIDENCY;
9290 
9291 		if (!platform->has_msr_core_c1_res && cai->msr == MSR_CORE_C1_RES)
9292 			cai->msr = 0;
9293 
9294 		if (platform->has_msr_atom_pkg_c6_residency && cai->msr == MSR_PKG_C6_RESIDENCY)
9295 			cai->msr = MSR_ATOM_PKG_C6_RESIDENCY;
9296 	}
9297 
9298 	for (int i = 0; i < NUM_MSR_COUNTERS; ++i) {
9299 		msr_counter_arch_infos[i].present = false;
9300 		msr_counter_arch_infos[i].needed = false;
9301 	}
9302 }
9303 
probe_pm_features(void)9304 void probe_pm_features(void)
9305 {
9306 	probe_pstates();
9307 
9308 	probe_cstates();
9309 
9310 	probe_lpi();
9311 
9312 	probe_intel_uncore_frequency();
9313 
9314 	probe_graphics();
9315 
9316 	probe_rapl();
9317 
9318 	probe_thermal();
9319 
9320 	if (platform->has_nhm_msrs && !no_msr)
9321 		BIC_PRESENT(BIC_SMI);
9322 
9323 	if (!quiet)
9324 		decode_misc_feature_control();
9325 }
9326 
9327 /*
9328  * has_perf_llc_access()
9329  *
9330  * return 1 on success, else 0
9331  */
has_perf_llc_access(void)9332 int has_perf_llc_access(void)
9333 {
9334 	int fd;
9335 
9336 	if (no_perf)
9337 		return 0;
9338 
9339 	fd = open_perf_counter(master_cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_REFERENCES, -1, PERF_FORMAT_GROUP);
9340 	if (fd != -1)
9341 		close(fd);
9342 
9343 	if (fd == -1)
9344 		warnx("Failed to access %s. Some of the counters may not be available\n"
9345 		      "\tRun as root to enable them or use %s to disable the access explicitly", "perf LLC counters", "'--hide LLC' or '--no-perf'");
9346 
9347 	return (fd != -1);
9348 }
9349 
perf_llc_init(void)9350 void perf_llc_init(void)
9351 {
9352 	int cpu;
9353 	int retval;
9354 
9355 	if (no_perf)
9356 		return;
9357 	if (!(BIC_IS_ENABLED(BIC_LLC_MRPS) || BIC_IS_ENABLED(BIC_LLC_HIT)))
9358 		return;
9359 
9360 	assert(fd_llc_percpu != 0);
9361 
9362 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
9363 
9364 		if (cpu_is_not_allowed(cpu))
9365 			continue;
9366 
9367 		fd_llc_percpu[cpu] = open_perf_counter(cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_REFERENCES, -1, PERF_FORMAT_GROUP);
9368 		if (fd_llc_percpu[cpu] == -1) {
9369 			warnx("%s: perf REFS: failed to open counter on cpu%d", __func__, cpu);
9370 			free_fd_llc_percpu();
9371 			return;
9372 		}
9373 		retval = open_perf_counter(cpu, PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES, fd_llc_percpu[cpu], PERF_FORMAT_GROUP);
9374 		if (retval == -1) {
9375 			warnx("%s: perf MISS: failed to open counter on cpu%d", __func__, cpu);
9376 			free_fd_llc_percpu();
9377 			return;
9378 		}
9379 	}
9380 	BIC_PRESENT(BIC_LLC_MRPS);
9381 	BIC_PRESENT(BIC_LLC_HIT);
9382 }
9383 
perf_l2_init(void)9384 void perf_l2_init(void)
9385 {
9386 	int cpu;
9387 	int retval;
9388 
9389 	if (no_perf)
9390 		return;
9391 	if (!(BIC_IS_ENABLED(BIC_L2_MRPS) || BIC_IS_ENABLED(BIC_L2_HIT)))
9392 		return;
9393 	if (perf_model_support == NULL)
9394 		return;
9395 
9396 	assert(fd_l2_percpu != 0);
9397 
9398 	for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
9399 
9400 		if (cpu_is_not_allowed(cpu))
9401 			continue;
9402 
9403 		if (!is_hybrid) {
9404 			fd_l2_percpu[cpu] = open_perf_counter(cpu, perf_pmu_types.uniform, perf_model_support->first.refs, -1, PERF_FORMAT_GROUP);
9405 			if (fd_l2_percpu[cpu] == -1) {
9406 				err(-1, "%s(cpu%d, 0x%x, 0x%llx) REFS", __func__, cpu, perf_pmu_types.uniform, perf_model_support->first.refs);
9407 				free_fd_l2_percpu();
9408 				return;
9409 			}
9410 			retval = open_perf_counter(cpu, perf_pmu_types.uniform, perf_model_support->first.hits, fd_l2_percpu[cpu], PERF_FORMAT_GROUP);
9411 			if (retval == -1) {
9412 				err(-1, "%s(cpu%d, 0x%x, 0x%llx) HITS", __func__, cpu, perf_pmu_types.uniform, perf_model_support->first.hits);
9413 				free_fd_l2_percpu();
9414 				return;
9415 			}
9416 			continue;
9417 		}
9418 		if (perf_pcore_set && CPU_ISSET_S(cpu, cpu_possible_setsize, perf_pcore_set)) {
9419 			fd_l2_percpu[cpu] = open_perf_counter(cpu, perf_pmu_types.pcore, perf_model_support->first.refs, -1, PERF_FORMAT_GROUP);
9420 			if (fd_l2_percpu[cpu] == -1) {
9421 				err(-1, "%s(cpu%d, 0x%x, 0x%llx) REFS", __func__, cpu, perf_pmu_types.pcore, perf_model_support->first.refs);
9422 				free_fd_l2_percpu();
9423 				return;
9424 			}
9425 			retval = open_perf_counter(cpu, perf_pmu_types.pcore, perf_model_support->first.hits, fd_l2_percpu[cpu], PERF_FORMAT_GROUP);
9426 			if (retval == -1) {
9427 				err(-1, "%s(cpu%d, 0x%x, 0x%llx) HITS", __func__, cpu, perf_pmu_types.pcore, perf_model_support->first.hits);
9428 				free_fd_l2_percpu();
9429 				return;
9430 			}
9431 		} else if (perf_ecore_set && CPU_ISSET_S(cpu, cpu_possible_setsize, perf_ecore_set)) {
9432 			fd_l2_percpu[cpu] = open_perf_counter(cpu, perf_pmu_types.ecore, perf_model_support->second.refs, -1, PERF_FORMAT_GROUP);
9433 			if (fd_l2_percpu[cpu] == -1) {
9434 				err(-1, "%s(cpu%d, 0x%x, 0x%llx) REFS", __func__, cpu, perf_pmu_types.pcore, perf_model_support->second.refs);
9435 				free_fd_l2_percpu();
9436 				return;
9437 			}
9438 			retval = open_perf_counter(cpu, perf_pmu_types.ecore, perf_model_support->second.hits, fd_l2_percpu[cpu], PERF_FORMAT_GROUP);
9439 			if (retval == -1) {
9440 				err(-1, "%s(cpu%d, 0x%x, 0x%llx) HITS", __func__, cpu, perf_pmu_types.pcore, perf_model_support->second.hits);
9441 				free_fd_l2_percpu();
9442 				return;
9443 			}
9444 		} else if (perf_lcore_set && CPU_ISSET_S(cpu, cpu_possible_setsize, perf_lcore_set)) {
9445 			fd_l2_percpu[cpu] = open_perf_counter(cpu, perf_pmu_types.lcore, perf_model_support->third.refs, -1, PERF_FORMAT_GROUP);
9446 			if (fd_l2_percpu[cpu] == -1) {
9447 				err(-1, "%s(cpu%d, 0x%x, 0x%llx) REFS", __func__, cpu, perf_pmu_types.pcore, perf_model_support->third.refs);
9448 				free_fd_l2_percpu();
9449 				return;
9450 			}
9451 			retval = open_perf_counter(cpu, perf_pmu_types.lcore, perf_model_support->third.hits, fd_l2_percpu[cpu], PERF_FORMAT_GROUP);
9452 			if (retval == -1) {
9453 				err(-1, "%s(cpu%d, 0x%x, 0x%llx) HITS", __func__, cpu, perf_pmu_types.pcore, perf_model_support->third.hits);
9454 				free_fd_l2_percpu();
9455 				return;
9456 			}
9457 		} else
9458 			err(-1, "%s: cpu%d: type %d", __func__, cpu, cpus[cpu].type);
9459 	}
9460 	BIC_PRESENT(BIC_L2_MRPS);
9461 	BIC_PRESENT(BIC_L2_HIT);
9462 }
9463 
9464 /*
9465  * in /dev/cpu/ return success for names that are numbers
9466  * ie. filter out ".", "..", "microcode".
9467  */
dir_filter(const struct dirent * dirp)9468 int dir_filter(const struct dirent *dirp)
9469 {
9470 	if (isdigit(dirp->d_name[0]))
9471 		return 1;
9472 	else
9473 		return 0;
9474 }
9475 
topology_probe(bool startup)9476 void topology_probe(bool startup)
9477 {
9478 	int i;
9479 	int max_core_id = 0;
9480 	int max_package_id = 0;
9481 	int max_siblings = 0;
9482 
9483 	/* Initialize num_cpus, max_cpu_num */
9484 	set_max_cpu_num();
9485 	topo.num_cpus = 0;
9486 	for_all_proc_cpus(count_cpus);
9487 	if (!summary_only)
9488 		BIC_PRESENT(BIC_CPU);
9489 
9490 	if (debug > 1)
9491 		fprintf(outf, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
9492 
9493 	cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
9494 	if (cpus == NULL)
9495 		err(1, "calloc cpus");
9496 
9497 	/*
9498 	 * Allocate and initialize cpu_present_set
9499 	 */
9500 	cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
9501 	if (cpu_present_set == NULL)
9502 		err(3, "CPU_ALLOC");
9503 	cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9504 	CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
9505 	for_all_proc_cpus(mark_cpu_present);
9506 
9507 	/*
9508 	 * Allocate and initialize cpu_possible_set
9509 	 */
9510 	cpu_possible_set = CPU_ALLOC((topo.max_cpu_num + 1));
9511 	if (cpu_possible_set == NULL)
9512 		err(3, "CPU_ALLOC");
9513 	cpu_possible_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9514 	CPU_ZERO_S(cpu_possible_setsize, cpu_possible_set);
9515 	initialize_cpu_set_from_sysfs(cpu_possible_set, "/sys/devices/system/cpu", "possible");
9516 
9517 	/*
9518 	 * Allocate and initialize cpu_effective_set
9519 	 */
9520 	cpu_effective_set = CPU_ALLOC((topo.max_cpu_num + 1));
9521 	if (cpu_effective_set == NULL)
9522 		err(3, "CPU_ALLOC");
9523 	cpu_effective_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9524 	CPU_ZERO_S(cpu_effective_setsize, cpu_effective_set);
9525 	update_effective_set(startup);
9526 
9527 	/*
9528 	 * Allocate and initialize cpu_allowed_set
9529 	 */
9530 	cpu_allowed_set = CPU_ALLOC((topo.max_cpu_num + 1));
9531 	if (cpu_allowed_set == NULL)
9532 		err(3, "CPU_ALLOC");
9533 	cpu_allowed_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9534 	CPU_ZERO_S(cpu_allowed_setsize, cpu_allowed_set);
9535 
9536 	/*
9537 	 * Validate and update cpu_allowed_set.
9538 	 *
9539 	 * Make sure all cpus in cpu_subset are also in cpu_present_set during startup.
9540 	 * Give a warning when cpus in cpu_subset become unavailable at runtime.
9541 	 * Give a warning when cpus are not effective because of cgroup setting.
9542 	 *
9543 	 * cpu_allowed_set is the intersection of cpu_present_set/cpu_effective_set/cpu_subset.
9544 	 */
9545 	for (i = 0; i < CPU_SUBSET_MAXCPUS; ++i) {
9546 		if (cpu_subset && !CPU_ISSET_S(i, cpu_subset_size, cpu_subset))
9547 			continue;
9548 
9549 		if (!CPU_ISSET_S(i, cpu_present_setsize, cpu_present_set)) {
9550 			if (cpu_subset) {
9551 				/* cpus in cpu_subset must be in cpu_present_set during startup */
9552 				if (startup)
9553 					err(1, "cpu%d not present", i);
9554 				else
9555 					fprintf(stderr, "cpu%d not present\n", i);
9556 			}
9557 			continue;
9558 		}
9559 
9560 		if (CPU_COUNT_S(cpu_effective_setsize, cpu_effective_set)) {
9561 			if (!CPU_ISSET_S(i, cpu_effective_setsize, cpu_effective_set)) {
9562 				fprintf(stderr, "cpu%d not effective\n", i);
9563 				continue;
9564 			}
9565 		}
9566 
9567 		CPU_SET_S(i, cpu_allowed_setsize, cpu_allowed_set);
9568 	}
9569 
9570 	if (!CPU_COUNT_S(cpu_allowed_setsize, cpu_allowed_set))
9571 		err(-ENODEV, "No valid cpus found");
9572 	sched_setaffinity(0, cpu_allowed_setsize, cpu_allowed_set);
9573 
9574 	/*
9575 	 * Allocate and initialize cpu_affinity_set
9576 	 */
9577 	cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
9578 	if (cpu_affinity_set == NULL)
9579 		err(3, "CPU_ALLOC");
9580 	cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
9581 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
9582 
9583 	for_all_proc_cpus(clear_ht_id);
9584 
9585 	for_all_proc_cpus(set_cpu_hybrid_type);
9586 
9587 	/*
9588 	 * For online cpus
9589 	 * find max_core_id, max_package_id, num_cores (per system)
9590 	 */
9591 	for (i = 0; i <= topo.max_cpu_num; ++i) {
9592 		int siblings;
9593 
9594 		if (cpu_is_not_present(i)) {
9595 			if (debug > 1)
9596 				fprintf(outf, "cpu%d NOT PRESENT\n", i);
9597 			continue;
9598 		}
9599 
9600 		cpus[i].cpu_id = i;
9601 
9602 		/* get package information */
9603 		cpus[i].package_id = get_package_id(i);
9604 		if (cpus[i].package_id > max_package_id)
9605 			max_package_id = cpus[i].package_id;
9606 
9607 		/* get die information */
9608 		cpus[i].die_id = get_die_id(i);
9609 		if (cpus[i].die_id > topo.max_die_id)
9610 			topo.max_die_id = cpus[i].die_id;
9611 
9612 		/* get l3 information */
9613 		cpus[i].l3_id = get_l3_id(i);
9614 		if (cpus[i].l3_id > topo.max_l3_id)
9615 			topo.max_l3_id = cpus[i].l3_id;
9616 
9617 		/* get numa node information */
9618 		cpus[i].physical_node_id = get_physical_node_id(&cpus[i]);
9619 		if (cpus[i].physical_node_id > topo.max_node_num)
9620 			topo.max_node_num = cpus[i].physical_node_id;
9621 
9622 		/* get core information */
9623 		cpus[i].core_id = get_core_id(i);
9624 		if (cpus[i].core_id > max_core_id)
9625 			max_core_id = cpus[i].core_id;
9626 
9627 		/* get thread information */
9628 		siblings = set_thread_siblings(&cpus[i]);
9629 		if (siblings > max_siblings)
9630 			max_siblings = siblings;
9631 		if (cpus[i].ht_id == 0)
9632 			topo.num_cores++;
9633 	}
9634 	topo.max_core_id = max_core_id;	/* within a package */
9635 	topo.max_package_id = max_package_id;
9636 
9637 	topo.cores_per_node = max_core_id + 1;
9638 	if (debug > 1)
9639 		fprintf(outf, "max_core_id %d, sizing for %d cores per package\n", max_core_id, topo.cores_per_node);
9640 	if (!summary_only)
9641 		BIC_PRESENT(BIC_Core);
9642 
9643 	topo.num_die = topo.max_die_id + 1;
9644 	if (debug > 1)
9645 		fprintf(outf, "max_die_id %d, sizing for %d die\n", topo.max_die_id, topo.num_die);
9646 	if (!summary_only && topo.num_die > 1)
9647 		BIC_PRESENT(BIC_Die);
9648 
9649 	if (!summary_only && topo.max_l3_id > 0)
9650 		BIC_PRESENT(BIC_L3);
9651 
9652 	topo.num_packages = max_package_id + 1;
9653 	if (debug > 1)
9654 		fprintf(outf, "max_package_id %d, sizing for %d packages\n", max_package_id, topo.num_packages);
9655 	if (!summary_only && topo.num_packages > 1)
9656 		BIC_PRESENT(BIC_Package);
9657 
9658 	set_node_data();
9659 	if (debug > 1)
9660 		fprintf(outf, "nodes_per_pkg %d\n", topo.nodes_per_pkg);
9661 	if (!summary_only && topo.nodes_per_pkg > 1)
9662 		BIC_PRESENT(BIC_Node);
9663 
9664 	topo.threads_per_core = max_siblings;
9665 	if (debug > 1)
9666 		fprintf(outf, "max_siblings %d\n", max_siblings);
9667 
9668 	if (debug < 1)
9669 		return;
9670 
9671 	for (i = 0; i <= topo.max_cpu_num; ++i) {
9672 		if (cpu_is_not_present(i))
9673 			continue;
9674 		fprintf(outf,
9675 			"cpu %d pkg %d die %d l3 %d node %d lnode %d core %d thread %d\n",
9676 			i, cpus[i].package_id, cpus[i].die_id, cpus[i].l3_id,
9677 			cpus[i].physical_node_id, cpus[i].logical_node_id, cpus[i].core_id, cpus[i].ht_id);
9678 	}
9679 
9680 }
9681 
allocate_counters_1(struct counters * counters)9682 void allocate_counters_1(struct counters *counters)
9683 {
9684 	counters->threads = calloc(1, sizeof(struct thread_data));
9685 	if (counters->threads == NULL)
9686 		goto error;
9687 
9688 	counters->cores = calloc(1, sizeof(struct core_data));
9689 	if (counters->cores == NULL)
9690 		goto error;
9691 
9692 	counters->packages = calloc(1, sizeof(struct pkg_data));
9693 	if (counters->packages == NULL)
9694 		goto error;
9695 
9696 	return;
9697 error:
9698 	err(1, "calloc counters_1");
9699 }
9700 
allocate_counters(struct counters * counters)9701 void allocate_counters(struct counters *counters)
9702 {
9703 	int i;
9704 	int num_cores = topo.cores_per_node * topo.nodes_per_pkg * topo.num_packages;
9705 	int num_threads = topo.threads_per_core * num_cores;
9706 
9707 	counters->threads = calloc(num_threads, sizeof(struct thread_data));
9708 	if (counters->threads == NULL)
9709 		goto error;
9710 
9711 	for (i = 0; i < num_threads; i++)
9712 		(counters->threads)[i].cpu_id = -1;
9713 
9714 	counters->cores = calloc(num_cores, sizeof(struct core_data));
9715 	if (counters->cores == NULL)
9716 		goto error;
9717 
9718 	for (i = 0; i < num_cores; i++)
9719 		(counters->cores)[i].first_cpu = -1;
9720 
9721 	counters->packages = calloc(topo.num_packages, sizeof(struct pkg_data));
9722 	if (counters->packages == NULL)
9723 		goto error;
9724 
9725 	for (i = 0; i < topo.num_packages; i++)
9726 		(counters->packages)[i].first_cpu = -1;
9727 
9728 	return;
9729 error:
9730 	err(1, "calloc counters");
9731 }
9732 
9733 /*
9734  * init_counter()
9735  *
9736  * set t->cpu_id, FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
9737  */
init_counter(struct thread_data * thread_base,struct core_data * core_base,struct pkg_data * pkg_base,int cpu_id)9738 void init_counter(struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base, int cpu_id)
9739 {
9740 	int pkg_id = cpus[cpu_id].package_id;
9741 	int node_id = cpus[cpu_id].logical_node_id;
9742 	int core_id = cpus[cpu_id].core_id;
9743 	struct thread_data *t;
9744 	struct core_data *c;
9745 
9746 	/* Workaround for systems where physical_node_id==-1
9747 	 * and logical_node_id==(-1 - topo.num_cpus)
9748 	 */
9749 	if (node_id < 0)
9750 		node_id = 0;
9751 
9752 	t = &thread_base[cpu_id];
9753 	c = &core_base[GLOBAL_CORE_ID(core_id, pkg_id)];
9754 
9755 	t->cpu_id = cpu_id;
9756 	if (!cpu_is_not_allowed(cpu_id)) {
9757 
9758 		if (c->first_cpu < 0)
9759 			c->first_cpu = t->cpu_id;
9760 		if (pkg_base[pkg_id].first_cpu < 0)
9761 			pkg_base[pkg_id].first_cpu = t->cpu_id;
9762 	}
9763 }
9764 
initialize_counters(int cpu_id)9765 int initialize_counters(int cpu_id)
9766 {
9767 	init_counter(EVEN_COUNTERS, cpu_id);
9768 	init_counter(ODD_COUNTERS, cpu_id);
9769 	return 0;
9770 }
9771 
allocate_output_buffer()9772 void allocate_output_buffer()
9773 {
9774 	output_buffer = calloc(1, (1 + topo.num_cpus) * 2048);
9775 	outp = output_buffer;
9776 	if (outp == NULL)
9777 		err(-1, "calloc output buffer");
9778 }
9779 
allocate_fd_percpu(void)9780 void allocate_fd_percpu(void)
9781 {
9782 	fd_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
9783 	if (fd_percpu == NULL)
9784 		err(-1, "calloc fd_percpu");
9785 }
9786 
allocate_irq_buffers(void)9787 void allocate_irq_buffers(void)
9788 {
9789 	irq_column_2_cpu = calloc(topo.num_cpus, sizeof(int));
9790 	if (irq_column_2_cpu == NULL)
9791 		err(-1, "calloc %d", topo.num_cpus);
9792 
9793 	irqs_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
9794 	if (irqs_per_cpu == NULL)
9795 		err(-1, "calloc %d IRQ", topo.max_cpu_num + 1);
9796 
9797 	nmi_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
9798 	if (nmi_per_cpu == NULL)
9799 		err(-1, "calloc %d NMI", topo.max_cpu_num + 1);
9800 }
9801 
update_topo(PER_THREAD_PARAMS)9802 int update_topo(PER_THREAD_PARAMS)
9803 {
9804 	topo.allowed_cpus++;
9805 	if ((int)t->cpu_id == c->first_cpu)
9806 		topo.allowed_cores++;
9807 	if ((int)t->cpu_id == p->first_cpu)
9808 		topo.allowed_packages++;
9809 
9810 	return 0;
9811 }
9812 
topology_update(void)9813 void topology_update(void)
9814 {
9815 	topo.allowed_cpus = 0;
9816 	topo.allowed_cores = 0;
9817 	topo.allowed_packages = 0;
9818 	for_all_cpus(update_topo, ODD_COUNTERS);
9819 }
9820 
setup_all_buffers(bool startup)9821 void setup_all_buffers(bool startup)
9822 {
9823 	topology_probe(startup);
9824 	allocate_irq_buffers();
9825 	allocate_fd_percpu();
9826 	allocate_counters_1(&average);
9827 	allocate_counters(&even);
9828 	allocate_counters(&odd);
9829 	allocate_output_buffer();
9830 	for_all_proc_cpus(initialize_counters);
9831 	topology_update();
9832 }
9833 
set_master_cpu(void)9834 void set_master_cpu(void)
9835 {
9836 	int i;
9837 
9838 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
9839 		if (cpu_is_not_allowed(i))
9840 			continue;
9841 		master_cpu = i;
9842 		if (debug > 1)
9843 			fprintf(outf, "master_cpu = %d\n", master_cpu);
9844 		return;
9845 	}
9846 	err(-ENODEV, "No valid cpus found");
9847 }
9848 
has_added_counters(void)9849 bool has_added_counters(void)
9850 {
9851 	/*
9852 	 * It only makes sense to call this after the command line is parsed,
9853 	 * otherwise sys structure is not populated.
9854 	 */
9855 
9856 	return sys.added_core_counters | sys.added_thread_counters | sys.added_package_counters;
9857 }
9858 
check_msr_access(void)9859 void check_msr_access(void)
9860 {
9861 	check_msr_driver();
9862 	check_msr_permission();
9863 
9864 	if (no_msr)
9865 		bic_disable_msr_access();
9866 }
9867 
check_perf_access(void)9868 void check_perf_access(void)
9869 {
9870 	if (BIC_IS_ENABLED(BIC_IPC))
9871 		if (!has_perf_instr_count_access())
9872 			no_perf = 1;
9873 
9874 	if (BIC_IS_ENABLED(BIC_LLC_MRPS) || BIC_IS_ENABLED(BIC_LLC_HIT))
9875 		if (!has_perf_llc_access())
9876 			no_perf = 1;
9877 
9878 	if (no_perf)
9879 		bic_disable_perf_access();
9880 }
9881 
perf_has_hybrid_devices(void)9882 bool perf_has_hybrid_devices(void)
9883 {
9884 	/*
9885 	 *  0: unknown
9886 	 *  1: has separate perf device for p and e core
9887 	 * -1: doesn't have separate perf device for p and e core
9888 	 */
9889 	static int cached;
9890 
9891 	if (cached > 0)
9892 		return true;
9893 
9894 	if (cached < 0)
9895 		return false;
9896 
9897 	if (access("/sys/bus/event_source/devices/cpu_core", F_OK)) {
9898 		cached = -1;
9899 		return false;
9900 	}
9901 
9902 	if (access("/sys/bus/event_source/devices/cpu_atom", F_OK)) {
9903 		cached = -1;
9904 		return false;
9905 	}
9906 
9907 	cached = 1;
9908 	return true;
9909 }
9910 
added_perf_counters_init_(struct perf_counter_info * pinfo)9911 int added_perf_counters_init_(struct perf_counter_info *pinfo)
9912 {
9913 	size_t num_domains = 0;
9914 	unsigned int next_domain;
9915 	bool *domain_visited;
9916 	unsigned int perf_type, perf_config;
9917 	double perf_scale;
9918 	int fd_perf;
9919 
9920 	if (!pinfo)
9921 		return 0;
9922 
9923 	const size_t max_num_domains = MAX(topo.max_cpu_num + 1, MAX(topo.max_core_id + 1, topo.max_package_id + 1));
9924 
9925 	domain_visited = calloc(max_num_domains, sizeof(*domain_visited));
9926 
9927 	while (pinfo) {
9928 		switch (pinfo->scope) {
9929 		case SCOPE_CPU:
9930 			num_domains = topo.max_cpu_num + 1;
9931 			break;
9932 
9933 		case SCOPE_CORE:
9934 			num_domains = topo.max_core_id + 1;
9935 			break;
9936 
9937 		case SCOPE_PACKAGE:
9938 			num_domains = topo.max_package_id + 1;
9939 			break;
9940 		}
9941 
9942 		/* Allocate buffer for file descriptor for each domain. */
9943 		pinfo->fd_perf_per_domain = calloc(num_domains, sizeof(*pinfo->fd_perf_per_domain));
9944 		if (!pinfo->fd_perf_per_domain)
9945 			errx(1, "%s: alloc %s", __func__, "fd_perf_per_domain");
9946 
9947 		for (size_t i = 0; i < num_domains; ++i)
9948 			pinfo->fd_perf_per_domain[i] = -1;
9949 
9950 		pinfo->num_domains = num_domains;
9951 		pinfo->scale = 1.0;
9952 
9953 		memset(domain_visited, 0, max_num_domains * sizeof(*domain_visited));
9954 
9955 		for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
9956 
9957 			next_domain = cpu_to_domain(pinfo, cpu);
9958 
9959 			assert(next_domain < num_domains);
9960 
9961 			if (cpu_is_not_allowed(cpu))
9962 				continue;
9963 
9964 			if (domain_visited[next_domain])
9965 				continue;
9966 
9967 			/*
9968 			 * Intel hybrid platforms expose different perf devices for P and E cores.
9969 			 * Instead of one, "/sys/bus/event_source/devices/cpu" device, there are
9970 			 * "/sys/bus/event_source/devices/{cpu_core,cpu_atom}".
9971 			 *
9972 			 * This makes it more complicated to the user, because most of the counters
9973 			 * are available on both and have to be handled manually, otherwise.
9974 			 *
9975 			 * Code below, allow user to use the old "cpu" name, which is translated accordingly.
9976 			 */
9977 			const char *perf_device = pinfo->device;
9978 
9979 			if (strcmp(perf_device, "cpu") == 0 && perf_has_hybrid_devices()) {
9980 				switch (cpus[cpu].type) {
9981 				case INTEL_PCORE_TYPE:
9982 					perf_device = "cpu_core";
9983 					break;
9984 
9985 				case INTEL_ECORE_TYPE:
9986 					perf_device = "cpu_atom";
9987 					break;
9988 
9989 				default:	/* Don't change, we will probably fail and report a problem soon. */
9990 					break;
9991 				}
9992 			}
9993 
9994 			perf_type = read_perf_type(perf_device);
9995 			if (perf_type == (unsigned int)-1) {
9996 				warnx("%s: perf/%s/%s: failed to read %s", __func__, perf_device, pinfo->event, "type");
9997 				continue;
9998 			}
9999 
10000 			perf_config = read_perf_config(perf_device, pinfo->event);
10001 			if (perf_config == (unsigned int)-1) {
10002 				warnx("%s: perf/%s/%s: failed to read %s", __func__, perf_device, pinfo->event, "config");
10003 				continue;
10004 			}
10005 
10006 			/* Scale is not required, some counters just don't have it. */
10007 			perf_scale = read_perf_scale(perf_device, pinfo->event);
10008 			if (perf_scale == 0.0)
10009 				perf_scale = 1.0;
10010 
10011 			fd_perf = open_perf_counter(cpu, perf_type, perf_config, -1, 0);
10012 			if (fd_perf == -1) {
10013 				warnx("%s: perf/%s/%s: failed to open counter on cpu%d", __func__, perf_device, pinfo->event, cpu);
10014 				continue;
10015 			}
10016 
10017 			domain_visited[next_domain] = 1;
10018 			pinfo->fd_perf_per_domain[next_domain] = fd_perf;
10019 			pinfo->scale = perf_scale;
10020 
10021 			if (debug)
10022 				fprintf(stderr, "Add perf/%s/%s cpu%d: %d\n", perf_device, pinfo->event, cpu, pinfo->fd_perf_per_domain[next_domain]);
10023 		}
10024 
10025 		pinfo = pinfo->next;
10026 	}
10027 
10028 	free(domain_visited);
10029 
10030 	return 0;
10031 }
10032 
added_perf_counters_init(void)10033 void added_perf_counters_init(void)
10034 {
10035 	if (added_perf_counters_init_(sys.perf_tp))
10036 		errx(1, "%s: %s", __func__, "thread");
10037 
10038 	if (added_perf_counters_init_(sys.perf_cp))
10039 		errx(1, "%s: %s", __func__, "core");
10040 
10041 	if (added_perf_counters_init_(sys.perf_pp))
10042 		errx(1, "%s: %s", __func__, "package");
10043 }
10044 
parse_telem_info_file(int fd_dir,const char * info_filename,const char * format,unsigned long * output)10045 int parse_telem_info_file(int fd_dir, const char *info_filename, const char *format, unsigned long *output)
10046 {
10047 	int fd_telem_info;
10048 	FILE *file_telem_info;
10049 	unsigned long value;
10050 
10051 	fd_telem_info = openat(fd_dir, info_filename, O_RDONLY);
10052 	if (fd_telem_info == -1)
10053 		return -1;
10054 
10055 	file_telem_info = fdopen(fd_telem_info, "r");
10056 	if (file_telem_info == NULL) {
10057 		close(fd_telem_info);
10058 		return -1;
10059 	}
10060 
10061 	if (fscanf(file_telem_info, format, &value) != 1) {
10062 		fclose(file_telem_info);
10063 		return -1;
10064 	}
10065 
10066 	fclose(file_telem_info);
10067 
10068 	*output = value;
10069 
10070 	return 0;
10071 }
10072 
pmt_mmio_open(unsigned int target_guid)10073 struct pmt_mmio *pmt_mmio_open(unsigned int target_guid)
10074 {
10075 	struct pmt_diriter_t pmt_iter;
10076 	const struct dirent *entry;
10077 	struct stat st;
10078 	int fd_telem_dir, fd_pmt;
10079 	unsigned long guid, size, offset;
10080 	size_t mmap_size;
10081 	void *mmio;
10082 	struct pmt_mmio *head = NULL, *last = NULL;
10083 	struct pmt_mmio *new_pmt = NULL;
10084 
10085 	if (stat(SYSFS_TELEM_PATH, &st) == -1)
10086 		return NULL;
10087 
10088 	pmt_diriter_init(&pmt_iter);
10089 	entry = pmt_diriter_begin(&pmt_iter, SYSFS_TELEM_PATH);
10090 	if (!entry) {
10091 		pmt_diriter_remove(&pmt_iter);
10092 		return NULL;
10093 	}
10094 
10095 	for (; entry != NULL; entry = pmt_diriter_next(&pmt_iter)) {
10096 		if (fstatat(dirfd(pmt_iter.dir), entry->d_name, &st, 0) == -1)
10097 			break;
10098 
10099 		if (!S_ISDIR(st.st_mode))
10100 			continue;
10101 
10102 		fd_telem_dir = openat(dirfd(pmt_iter.dir), entry->d_name, O_RDONLY);
10103 		if (fd_telem_dir == -1)
10104 			break;
10105 
10106 		if (parse_telem_info_file(fd_telem_dir, "guid", "%lx", &guid)) {
10107 			close(fd_telem_dir);
10108 			break;
10109 		}
10110 
10111 		if (parse_telem_info_file(fd_telem_dir, "size", "%lu", &size)) {
10112 			close(fd_telem_dir);
10113 			break;
10114 		}
10115 
10116 		if (guid != target_guid) {
10117 			close(fd_telem_dir);
10118 			continue;
10119 		}
10120 
10121 		if (parse_telem_info_file(fd_telem_dir, "offset", "%lu", &offset)) {
10122 			close(fd_telem_dir);
10123 			break;
10124 		}
10125 
10126 		assert(offset == 0);
10127 
10128 		fd_pmt = openat(fd_telem_dir, "telem", O_RDONLY);
10129 		if (fd_pmt == -1)
10130 			goto loop_cleanup_and_break;
10131 
10132 		mmap_size = ROUND_UP_TO_PAGE_SIZE(size);
10133 		mmio = mmap(0, mmap_size, PROT_READ, MAP_SHARED, fd_pmt, 0);
10134 		if (mmio != MAP_FAILED) {
10135 			if (debug)
10136 				fprintf(stderr, "%s: 0x%lx mmaped at: %p\n", __func__, guid, mmio);
10137 
10138 			new_pmt = calloc(1, sizeof(*new_pmt));
10139 
10140 			if (!new_pmt) {
10141 				fprintf(stderr, "%s: Failed to allocate pmt_mmio\n", __func__);
10142 				exit(1);
10143 			}
10144 
10145 			/*
10146 			 * Create linked list of mmaped regions,
10147 			 * but preserve the ordering from sysfs.
10148 			 * Ordering is important for the user to
10149 			 * use the seq=%u parameter when adding a counter.
10150 			 */
10151 			new_pmt->guid = guid;
10152 			new_pmt->mmio_base = mmio;
10153 			new_pmt->pmt_offset = offset;
10154 			new_pmt->size = size;
10155 			new_pmt->next = pmt_mmios;
10156 
10157 			if (last)
10158 				last->next = new_pmt;
10159 			else
10160 				head = new_pmt;
10161 
10162 			last = new_pmt;
10163 		}
10164 
10165 loop_cleanup_and_break:
10166 		close(fd_pmt);
10167 		close(fd_telem_dir);
10168 	}
10169 
10170 	pmt_diriter_remove(&pmt_iter);
10171 
10172 	/*
10173 	 * If we found something, stick just
10174 	 * created linked list to the front.
10175 	 */
10176 	if (head)
10177 		pmt_mmios = head;
10178 
10179 	return head;
10180 }
10181 
pmt_mmio_find(unsigned int guid)10182 struct pmt_mmio *pmt_mmio_find(unsigned int guid)
10183 {
10184 	struct pmt_mmio *pmmio = pmt_mmios;
10185 
10186 	while (pmmio) {
10187 		if (pmmio->guid == guid)
10188 			return pmmio;
10189 
10190 		pmmio = pmmio->next;
10191 	}
10192 
10193 	return NULL;
10194 }
10195 
pmt_get_counter_pointer(struct pmt_mmio * pmmio,unsigned long counter_offset)10196 void *pmt_get_counter_pointer(struct pmt_mmio *pmmio, unsigned long counter_offset)
10197 {
10198 	char *ret;
10199 
10200 	/* Get base of mmaped PMT file. */
10201 	ret = (char *)pmmio->mmio_base;
10202 
10203 	/*
10204 	 * Apply PMT MMIO offset to obtain beginning of the mmaped telemetry data.
10205 	 * It's not guaranteed that the mmaped memory begins with the telemetry data
10206 	 *      - we might have to apply the offset first.
10207 	 */
10208 	ret += pmmio->pmt_offset;
10209 
10210 	/* Apply the counter offset to get the address to the mmaped counter. */
10211 	ret += counter_offset;
10212 
10213 	return ret;
10214 }
10215 
pmt_add_guid(unsigned int guid,unsigned int seq)10216 struct pmt_mmio *pmt_add_guid(unsigned int guid, unsigned int seq)
10217 {
10218 	struct pmt_mmio *ret;
10219 
10220 	ret = pmt_mmio_find(guid);
10221 	if (!ret)
10222 		ret = pmt_mmio_open(guid);
10223 
10224 	while (ret && seq) {
10225 		ret = ret->next;
10226 		--seq;
10227 	}
10228 
10229 	return ret;
10230 }
10231 
10232 enum pmt_open_mode {
10233 	PMT_OPEN_TRY,		/* Open failure is not an error. */
10234 	PMT_OPEN_REQUIRED,	/* Open failure is a fatal error. */
10235 };
10236 
pmt_find_counter(struct pmt_counter * pcounter,const char * name)10237 struct pmt_counter *pmt_find_counter(struct pmt_counter *pcounter, const char *name)
10238 {
10239 	while (pcounter) {
10240 		if (strcmp(pcounter->name, name) == 0)
10241 			break;
10242 
10243 		pcounter = pcounter->next;
10244 	}
10245 
10246 	return pcounter;
10247 }
10248 
pmt_get_scope_root(enum counter_scope scope)10249 struct pmt_counter **pmt_get_scope_root(enum counter_scope scope)
10250 {
10251 	switch (scope) {
10252 	case SCOPE_CPU:
10253 		return &sys.pmt_tp;
10254 	case SCOPE_CORE:
10255 		return &sys.pmt_cp;
10256 	case SCOPE_PACKAGE:
10257 		return &sys.pmt_pp;
10258 	}
10259 
10260 	__builtin_unreachable();
10261 }
10262 
pmt_counter_add_domain(struct pmt_counter * pcounter,unsigned long * pmmio,unsigned int domain_id)10263 void pmt_counter_add_domain(struct pmt_counter *pcounter, unsigned long *pmmio, unsigned int domain_id)
10264 {
10265 	/* Make sure the new domain fits. */
10266 	if (domain_id >= pcounter->num_domains)
10267 		pmt_counter_resize(pcounter, domain_id + 1);
10268 
10269 	assert(pcounter->domains);
10270 	assert(domain_id < pcounter->num_domains);
10271 
10272 	pcounter->domains[domain_id].pcounter = pmmio;
10273 }
10274 
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)10275 int pmt_add_counter(unsigned int guid, unsigned int seq, const char *name, enum pmt_datatype type,
10276 		    unsigned int lsb, unsigned int msb, unsigned int offset, enum counter_scope scope,
10277 		    enum counter_format format, unsigned int domain_id, enum pmt_open_mode mode)
10278 {
10279 	struct pmt_mmio *mmio;
10280 	struct pmt_counter *pcounter;
10281 	struct pmt_counter **const pmt_root = pmt_get_scope_root(scope);
10282 	bool new_counter = false;
10283 	int conflict = 0;
10284 
10285 	if (lsb > msb) {
10286 		fprintf(stderr, "%s: %s: `%s` must be satisfied\n", __func__, "lsb <= msb", name);
10287 		exit(1);
10288 	}
10289 
10290 	if (msb >= 64) {
10291 		fprintf(stderr, "%s: %s: `%s` must be satisfied\n", __func__, "msb < 64", name);
10292 		exit(1);
10293 	}
10294 
10295 	mmio = pmt_add_guid(guid, seq);
10296 	if (!mmio) {
10297 		if (mode != PMT_OPEN_TRY) {
10298 			fprintf(stderr, "%s: failed to map PMT MMIO for guid %x, seq %u\n", __func__, guid, seq);
10299 			exit(1);
10300 		}
10301 
10302 		return 1;
10303 	}
10304 
10305 	if (offset >= mmio->size) {
10306 		if (mode != PMT_OPEN_TRY) {
10307 			fprintf(stderr, "%s: offset %u outside of PMT MMIO size %u\n", __func__, offset, mmio->size);
10308 			exit(1);
10309 		}
10310 
10311 		return 1;
10312 	}
10313 
10314 	pcounter = pmt_find_counter(*pmt_root, name);
10315 	if (!pcounter) {
10316 		pcounter = calloc(1, sizeof(*pcounter));
10317 		new_counter = true;
10318 	}
10319 
10320 	if (new_counter) {
10321 		strncpy(pcounter->name, name, ARRAY_SIZE(pcounter->name) - 1);
10322 		pcounter->type = type;
10323 		pcounter->scope = scope;
10324 		pcounter->lsb = lsb;
10325 		pcounter->msb = msb;
10326 		pcounter->format = format;
10327 	} else {
10328 		conflict += pcounter->type != type;
10329 		conflict += pcounter->scope != scope;
10330 		conflict += pcounter->lsb != lsb;
10331 		conflict += pcounter->msb != msb;
10332 		conflict += pcounter->format != format;
10333 	}
10334 
10335 	if (conflict) {
10336 		fprintf(stderr, "%s: conflicting parameters for the PMT counter with the same name %s\n", __func__, name);
10337 		exit(1);
10338 	}
10339 
10340 	pmt_counter_add_domain(pcounter, pmt_get_counter_pointer(mmio, offset), domain_id);
10341 
10342 	if (new_counter) {
10343 		pcounter->next = *pmt_root;
10344 		*pmt_root = pcounter;
10345 	}
10346 
10347 	return 0;
10348 }
10349 
pmt_init(void)10350 void pmt_init(void)
10351 {
10352 	int cpu_num;
10353 	unsigned long seq, offset, mod_num;
10354 
10355 	if (BIC_IS_ENABLED(BIC_Diec6)) {
10356 		pmt_add_counter(PMT_MTL_DC6_GUID, PMT_MTL_DC6_SEQ, "Die%c6", PMT_TYPE_XTAL_TIME,
10357 				PMT_COUNTER_MTL_DC6_LSB, PMT_COUNTER_MTL_DC6_MSB, PMT_COUNTER_MTL_DC6_OFFSET, SCOPE_PACKAGE, FORMAT_DELTA, 0, PMT_OPEN_TRY);
10358 	}
10359 
10360 	if (BIC_IS_ENABLED(BIC_CPU_c1e)) {
10361 		seq = 0;
10362 		offset = PMT_COUNTER_CWF_MC1E_OFFSET_BASE;
10363 		mod_num = 0;	/* Relative module number for current PMT file. */
10364 
10365 		/* Open the counter for each CPU. */
10366 		for (cpu_num = 0; cpu_num < topo.max_cpu_num;) {
10367 
10368 			if (cpu_is_not_allowed(cpu_num))
10369 				goto next_loop_iter;
10370 
10371 			/*
10372 			 * Set the scope to CPU, even though CWF report the counter per module.
10373 			 * CPUs inside the same module will read from the same location, instead of reporting zeros.
10374 			 *
10375 			 * CWF with newer firmware might require a PMT_TYPE_XTAL_TIME intead of PMT_TYPE_TCORE_CLOCK.
10376 			 */
10377 			pmt_add_counter(PMT_CWF_MC1E_GUID, seq, "CPU%c1e", PMT_TYPE_TCORE_CLOCK,
10378 					PMT_COUNTER_CWF_MC1E_LSB, PMT_COUNTER_CWF_MC1E_MSB, offset, SCOPE_CPU, FORMAT_DELTA, cpu_num, PMT_OPEN_TRY);
10379 
10380 			/*
10381 			 * Rather complex logic for each time we go to the next loop iteration,
10382 			 * so keep it as a label.
10383 			 */
10384 next_loop_iter:
10385 			/*
10386 			 * Advance the cpu number and check if we should also advance offset to
10387 			 * the next counter inside the PMT file.
10388 			 *
10389 			 * On Clearwater Forest platform, the counter is reported per module,
10390 			 * so open the same counter for all of the CPUs inside the module.
10391 			 * That way, reported table show the correct value for all of the CPUs inside the module,
10392 			 * instead of zeros.
10393 			 */
10394 			++cpu_num;
10395 			if (cpu_num % PMT_COUNTER_CWF_CPUS_PER_MODULE == 0) {
10396 				offset += PMT_COUNTER_CWF_MC1E_OFFSET_INCREMENT;
10397 				++mod_num;
10398 			}
10399 
10400 			/*
10401 			 * There are PMT_COUNTER_CWF_MC1E_NUM_MODULES_PER_FILE in each PMT file.
10402 			 *
10403 			 * If that number is reached, seq must be incremented to advance to the next file in a sequence.
10404 			 * Offset inside that file and a module counter has to be reset.
10405 			 */
10406 			if (mod_num == PMT_COUNTER_CWF_MC1E_NUM_MODULES_PER_FILE) {
10407 				++seq;
10408 				offset = PMT_COUNTER_CWF_MC1E_OFFSET_BASE;
10409 				mod_num = 0;
10410 			}
10411 		}
10412 	}
10413 }
10414 
turbostat_init()10415 void turbostat_init()
10416 {
10417 	setup_all_buffers(true);
10418 	set_master_cpu();
10419 	check_msr_access();
10420 	check_perf_access();
10421 	process_cpuid();
10422 	counter_info_init();
10423 	probe_pm_features();
10424 	msr_perf_init();
10425 	linux_perf_init();
10426 	rapl_perf_init();
10427 	cstate_perf_init();
10428 	perf_llc_init();
10429 	perf_l2_init();
10430 	added_perf_counters_init();
10431 	pmt_init();
10432 
10433 	for_all_cpus(get_cpu_type, ODD_COUNTERS);
10434 	for_all_cpus(get_cpu_type, EVEN_COUNTERS);
10435 
10436 	if (BIC_IS_ENABLED(BIC_IPC) && has_aperf_access && get_instr_count_fd(master_cpu) != -1)
10437 		BIC_PRESENT(BIC_IPC);
10438 
10439 	/*
10440 	 * If TSC tweak is needed, but couldn't get it,
10441 	 * disable more BICs, since it can't be reported accurately.
10442 	 */
10443 	if (platform->enable_tsc_tweak && !has_base_hz) {
10444 		CLR_BIC(BIC_Busy, &bic_enabled);
10445 		CLR_BIC(BIC_Bzy_MHz, &bic_enabled);
10446 	}
10447 }
10448 
affinitize_child(void)10449 void affinitize_child(void)
10450 {
10451 	/* Prefer cpu_possible_set, if available */
10452 	if (sched_setaffinity(0, cpu_possible_setsize, cpu_possible_set)) {
10453 		warn("sched_setaffinity cpu_possible_set");
10454 
10455 		/* Otherwise, allow child to run on same cpu set as turbostat */
10456 		if (sched_setaffinity(0, cpu_allowed_setsize, cpu_allowed_set))
10457 			warn("sched_setaffinity cpu_allowed_set");
10458 	}
10459 }
10460 
fork_it(char ** argv)10461 int fork_it(char **argv)
10462 {
10463 	pid_t child_pid;
10464 	int status;
10465 
10466 	snapshot_proc_sysfs_files();
10467 	status = for_all_cpus(get_counters, EVEN_COUNTERS);
10468 	first_counter_read = 0;
10469 	if (status)
10470 		exit(status);
10471 	gettimeofday(&tv_even, (struct timezone *)NULL);
10472 
10473 	child_pid = fork();
10474 	if (!child_pid) {
10475 		/* child */
10476 		affinitize_child();
10477 		execvp(argv[0], argv);
10478 		err(errno, "exec %s", argv[0]);
10479 	} else {
10480 
10481 		/* parent */
10482 		if (child_pid == -1)
10483 			err(1, "fork");
10484 
10485 		signal(SIGINT, SIG_IGN);
10486 		signal(SIGQUIT, SIG_IGN);
10487 		if (waitpid(child_pid, &status, 0) == -1)
10488 			err(status, "waitpid");
10489 
10490 		if (WIFEXITED(status))
10491 			status = WEXITSTATUS(status);
10492 	}
10493 	/*
10494 	 * n.b. fork_it() does not check for errors from for_all_cpus()
10495 	 * because re-starting is problematic when forking
10496 	 */
10497 	snapshot_proc_sysfs_files();
10498 	for_all_cpus(get_counters, ODD_COUNTERS);
10499 	gettimeofday(&tv_odd, (struct timezone *)NULL);
10500 	timersub(&tv_odd, &tv_even, &tv_delta);
10501 	if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS))
10502 		fprintf(outf, "%s: Counter reset detected\n", progname);
10503 	delta_platform(&platform_counters_odd, &platform_counters_even);
10504 
10505 	compute_average(EVEN_COUNTERS);
10506 	format_all_counters(EVEN_COUNTERS);
10507 
10508 	fprintf(outf, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec / 1000000.0);
10509 
10510 	flush_output_stderr();
10511 
10512 	return status;
10513 }
10514 
get_and_dump_counters(void)10515 int get_and_dump_counters(void)
10516 {
10517 	int status;
10518 
10519 	snapshot_proc_sysfs_files();
10520 	status = for_all_cpus(get_counters, ODD_COUNTERS);
10521 	if (status)
10522 		return status;
10523 
10524 	status = for_all_cpus(dump_counters, ODD_COUNTERS);
10525 	if (status)
10526 		return status;
10527 
10528 	flush_output_stdout();
10529 
10530 	return status;
10531 }
10532 
print_version()10533 void print_version()
10534 {
10535 	fprintf(outf, "turbostat version 2026.02.14 - Len Brown <lenb@kernel.org>\n");
10536 }
10537 
10538 #define COMMAND_LINE_SIZE 2048
10539 
print_bootcmd(void)10540 void print_bootcmd(void)
10541 {
10542 	char bootcmd[COMMAND_LINE_SIZE];
10543 	FILE *fp;
10544 	int ret;
10545 
10546 	memset(bootcmd, 0, COMMAND_LINE_SIZE);
10547 	fp = fopen("/proc/cmdline", "r");
10548 	if (!fp)
10549 		return;
10550 
10551 	ret = fread(bootcmd, sizeof(char), COMMAND_LINE_SIZE - 1, fp);
10552 	if (ret) {
10553 		bootcmd[ret] = '\0';
10554 		/* the last character is already '\n' */
10555 		fprintf(outf, "Kernel command line: %s", bootcmd);
10556 	}
10557 
10558 	fclose(fp);
10559 }
10560 
find_msrp_by_name(struct msr_counter * head,char * name)10561 struct msr_counter *find_msrp_by_name(struct msr_counter *head, char *name)
10562 {
10563 	struct msr_counter *mp;
10564 
10565 	for (mp = head; mp; mp = mp->next) {
10566 		if (debug)
10567 			fprintf(stderr, "%s: %s %s\n", __func__, name, mp->name);
10568 		if (!strcmp(name, mp->name))
10569 			return mp;
10570 	}
10571 	return NULL;
10572 }
10573 
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)10574 int add_counter(unsigned int msr_num, char *path, char *name,
10575 		unsigned int width, enum counter_scope scope, enum counter_type type, enum counter_format format, int flags, int id)
10576 {
10577 	struct msr_counter *msrp;
10578 
10579 	if (no_msr && msr_num)
10580 		errx(1, "Requested MSR counter 0x%x, but in --no-msr mode", msr_num);
10581 
10582 	if (debug)
10583 		fprintf(stderr, "%s(msr%d, %s, %s, width%d, scope%d, type%d, format%d, flags%x, id%d)\n",
10584 			__func__, msr_num, path, name, width, scope, type, format, flags, id);
10585 
10586 	switch (scope) {
10587 
10588 	case SCOPE_CPU:
10589 		msrp = find_msrp_by_name(sys.tp, name);
10590 		if (msrp) {
10591 			if (debug)
10592 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
10593 			break;
10594 		}
10595 		if (sys.added_thread_counters++ >= MAX_ADDED_THREAD_COUNTERS) {
10596 			warnx("ignoring thread counter %s", name);
10597 			return -1;
10598 		}
10599 		break;
10600 	case SCOPE_CORE:
10601 		msrp = find_msrp_by_name(sys.cp, name);
10602 		if (msrp) {
10603 			if (debug)
10604 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
10605 			break;
10606 		}
10607 		if (sys.added_core_counters++ >= MAX_ADDED_CORE_COUNTERS) {
10608 			warnx("ignoring core counter %s", name);
10609 			return -1;
10610 		}
10611 		break;
10612 	case SCOPE_PACKAGE:
10613 		msrp = find_msrp_by_name(sys.pp, name);
10614 		if (msrp) {
10615 			if (debug)
10616 				fprintf(stderr, "%s: %s FOUND\n", __func__, name);
10617 			break;
10618 		}
10619 		if (sys.added_package_counters++ >= MAX_ADDED_PACKAGE_COUNTERS) {
10620 			warnx("ignoring package counter %s", name);
10621 			return -1;
10622 		}
10623 		break;
10624 	default:
10625 		warnx("ignoring counter %s with unknown scope", name);
10626 		return -1;
10627 	}
10628 
10629 	if (msrp == NULL) {
10630 		msrp = calloc(1, sizeof(struct msr_counter));
10631 		if (msrp == NULL)
10632 			err(-1, "calloc msr_counter");
10633 
10634 		msrp->msr_num = msr_num;
10635 		strncpy(msrp->name, name, NAME_BYTES - 1);
10636 		msrp->width = width;
10637 		msrp->type = type;
10638 		msrp->format = format;
10639 		msrp->flags = flags;
10640 
10641 		switch (scope) {
10642 		case SCOPE_CPU:
10643 			msrp->next = sys.tp;
10644 			sys.tp = msrp;
10645 			break;
10646 		case SCOPE_CORE:
10647 			msrp->next = sys.cp;
10648 			sys.cp = msrp;
10649 			break;
10650 		case SCOPE_PACKAGE:
10651 			msrp->next = sys.pp;
10652 			sys.pp = msrp;
10653 			break;
10654 		}
10655 	}
10656 
10657 	if (path) {
10658 		struct sysfs_path *sp;
10659 
10660 		sp = calloc(1, sizeof(struct sysfs_path));
10661 		if (sp == NULL) {
10662 			perror("calloc");
10663 			exit(1);
10664 		}
10665 		strncpy(sp->path, path, PATH_BYTES - 1);
10666 		sp->id = id;
10667 		sp->next = msrp->sp;
10668 		msrp->sp = sp;
10669 	}
10670 
10671 	return 0;
10672 }
10673 
10674 /*
10675  * Initialize the fields used for identifying and opening the counter.
10676  *
10677  * Defer the initialization of any runtime buffers for actually reading
10678  * the counters for when we initialize all perf counters, so we can later
10679  * easily call re_initialize().
10680  */
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)10681 struct perf_counter_info *make_perf_counter_info(const char *perf_device,
10682 						 const char *perf_event,
10683 						 const char *name,
10684 						 unsigned int width, enum counter_scope scope, enum counter_type type, enum counter_format format)
10685 {
10686 	struct perf_counter_info *pinfo;
10687 
10688 	pinfo = calloc(1, sizeof(*pinfo));
10689 	if (!pinfo)
10690 		errx(1, "%s: Failed to allocate %s/%s\n", __func__, perf_device, perf_event);
10691 
10692 	strncpy(pinfo->device, perf_device, ARRAY_SIZE(pinfo->device) - 1);
10693 	strncpy(pinfo->event, perf_event, ARRAY_SIZE(pinfo->event) - 1);
10694 
10695 	strncpy(pinfo->name, name, ARRAY_SIZE(pinfo->name) - 1);
10696 	pinfo->width = width;
10697 	pinfo->scope = scope;
10698 	pinfo->type = type;
10699 	pinfo->format = format;
10700 
10701 	return pinfo;
10702 }
10703 
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)10704 int add_perf_counter(const char *perf_device, const char *perf_event, const char *name_buffer, unsigned int width,
10705 		     enum counter_scope scope, enum counter_type type, enum counter_format format)
10706 {
10707 	struct perf_counter_info *pinfo;
10708 
10709 	switch (scope) {
10710 	case SCOPE_CPU:
10711 		if (sys.added_thread_perf_counters >= MAX_ADDED_THREAD_COUNTERS) {
10712 			warnx("ignoring thread counter perf/%s/%s", perf_device, perf_event);
10713 			return -1;
10714 		}
10715 		break;
10716 
10717 	case SCOPE_CORE:
10718 		if (sys.added_core_perf_counters >= MAX_ADDED_CORE_COUNTERS) {
10719 			warnx("ignoring core counter perf/%s/%s", perf_device, perf_event);
10720 			return -1;
10721 		}
10722 		break;
10723 
10724 	case SCOPE_PACKAGE:
10725 		if (sys.added_package_perf_counters >= MAX_ADDED_PACKAGE_COUNTERS) {
10726 			warnx("ignoring package counter perf/%s/%s", perf_device, perf_event);
10727 			return -1;
10728 		}
10729 		break;
10730 	}
10731 
10732 	pinfo = make_perf_counter_info(perf_device, perf_event, name_buffer, width, scope, type, format);
10733 
10734 	if (!pinfo)
10735 		return -1;
10736 
10737 	switch (scope) {
10738 	case SCOPE_CPU:
10739 		pinfo->next = sys.perf_tp;
10740 		sys.perf_tp = pinfo;
10741 		++sys.added_thread_perf_counters;
10742 		break;
10743 
10744 	case SCOPE_CORE:
10745 		pinfo->next = sys.perf_cp;
10746 		sys.perf_cp = pinfo;
10747 		++sys.added_core_perf_counters;
10748 		break;
10749 
10750 	case SCOPE_PACKAGE:
10751 		pinfo->next = sys.perf_pp;
10752 		sys.perf_pp = pinfo;
10753 		++sys.added_package_perf_counters;
10754 		break;
10755 	}
10756 
10757 	// FIXME: we might not have debug here yet
10758 	if (debug)
10759 		fprintf(stderr, "%s: %s/%s, name: %s, scope%d\n", __func__, pinfo->device, pinfo->event, pinfo->name, pinfo->scope);
10760 
10761 	return 0;
10762 }
10763 
parse_add_command_msr(char * add_command)10764 void parse_add_command_msr(char *add_command)
10765 {
10766 	int msr_num = 0;
10767 	char *path = NULL;
10768 	char perf_device[PERF_DEV_NAME_BYTES] = "";
10769 	char perf_event[PERF_EVT_NAME_BYTES] = "";
10770 	char name_buffer[PERF_NAME_BYTES] = "";
10771 	int width = 64;
10772 	int fail = 0;
10773 	enum counter_scope scope = SCOPE_CPU;
10774 	enum counter_type type = COUNTER_CYCLES;
10775 	enum counter_format format = FORMAT_DELTA;
10776 
10777 	while (add_command) {
10778 
10779 		if (sscanf(add_command, "msr0x%x", &msr_num) == 1)
10780 			goto next;
10781 
10782 		if (sscanf(add_command, "msr%d", &msr_num) == 1)
10783 			goto next;
10784 
10785 		BUILD_BUG_ON(ARRAY_SIZE(perf_device) <= 31);
10786 		BUILD_BUG_ON(ARRAY_SIZE(perf_event) <= 31);
10787 		if (sscanf(add_command, "perf/%31[^/]/%31[^,]", &perf_device[0], &perf_event[0]) == 2)
10788 			goto next;
10789 
10790 		if (*add_command == '/') {
10791 			path = add_command;
10792 			goto next;
10793 		}
10794 
10795 		if (sscanf(add_command, "u%d", &width) == 1) {
10796 			if ((width == 32) || (width == 64))
10797 				goto next;
10798 			width = 64;
10799 		}
10800 		if (!strncmp(add_command, "cpu", strlen("cpu"))) {
10801 			scope = SCOPE_CPU;
10802 			goto next;
10803 		}
10804 		if (!strncmp(add_command, "core", strlen("core"))) {
10805 			scope = SCOPE_CORE;
10806 			goto next;
10807 		}
10808 		if (!strncmp(add_command, "package", strlen("package"))) {
10809 			scope = SCOPE_PACKAGE;
10810 			goto next;
10811 		}
10812 		if (!strncmp(add_command, "cycles", strlen("cycles"))) {
10813 			type = COUNTER_CYCLES;
10814 			goto next;
10815 		}
10816 		if (!strncmp(add_command, "seconds", strlen("seconds"))) {
10817 			type = COUNTER_SECONDS;
10818 			goto next;
10819 		}
10820 		if (!strncmp(add_command, "usec", strlen("usec"))) {
10821 			type = COUNTER_USEC;
10822 			goto next;
10823 		}
10824 		if (!strncmp(add_command, "raw", strlen("raw"))) {
10825 			format = FORMAT_RAW;
10826 			goto next;
10827 		}
10828 		if (!strncmp(add_command, "average", strlen("average"))) {
10829 			format = FORMAT_AVERAGE;
10830 			goto next;
10831 		}
10832 		if (!strncmp(add_command, "delta", strlen("delta"))) {
10833 			format = FORMAT_DELTA;
10834 			goto next;
10835 		}
10836 		if (!strncmp(add_command, "percent", strlen("percent"))) {
10837 			format = FORMAT_PERCENT;
10838 			goto next;
10839 		}
10840 
10841 		BUILD_BUG_ON(ARRAY_SIZE(name_buffer) <= 18);
10842 		if (sscanf(add_command, "%18s,%*s", name_buffer) == 1) {
10843 			char *eos;
10844 
10845 			eos = strchr(name_buffer, ',');
10846 			if (eos)
10847 				*eos = '\0';
10848 			goto next;
10849 		}
10850 
10851 next:
10852 		add_command = strchr(add_command, ',');
10853 		if (add_command) {
10854 			*add_command = '\0';
10855 			add_command++;
10856 		}
10857 
10858 	}
10859 	if ((msr_num == 0) && (path == NULL) && (perf_device[0] == '\0' || perf_event[0] == '\0')) {
10860 		fprintf(stderr, "--add: (msrDDD | msr0xXXX | /path_to_counter | perf/device/event) required\n");
10861 		fail++;
10862 	}
10863 
10864 	/* Test for non-empty perf_device and perf_event */
10865 	const bool is_perf_counter = perf_device[0] && perf_event[0];
10866 
10867 	/* generate default column header */
10868 	if (*name_buffer == '\0') {
10869 		if (is_perf_counter) {
10870 			snprintf(name_buffer, ARRAY_SIZE(name_buffer), "perf/%s", perf_event);
10871 		} else {
10872 			if (width == 32)
10873 				sprintf(name_buffer, "M0x%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
10874 			else
10875 				sprintf(name_buffer, "M0X%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
10876 		}
10877 	}
10878 
10879 	if (is_perf_counter) {
10880 		if (add_perf_counter(perf_device, perf_event, name_buffer, width, scope, type, format))
10881 			fail++;
10882 	} else {
10883 		if (add_counter(msr_num, path, name_buffer, width, scope, type, format, 0, 0))
10884 			fail++;
10885 	}
10886 
10887 	if (fail) {
10888 		help();
10889 		exit(1);
10890 	}
10891 }
10892 
starts_with(const char * str,const char * prefix)10893 bool starts_with(const char *str, const char *prefix)
10894 {
10895 	return strncmp(prefix, str, strlen(prefix)) == 0;
10896 }
10897 
pmt_parse_from_path(const char * target_path,unsigned int * out_guid,unsigned int * out_seq)10898 int pmt_parse_from_path(const char *target_path, unsigned int *out_guid, unsigned int *out_seq)
10899 {
10900 	struct pmt_diriter_t pmt_iter;
10901 	const struct dirent *dirname;
10902 	struct stat stat, target_stat;
10903 	int fd_telem_dir = -1;
10904 	int fd_target_dir;
10905 	unsigned int seq = 0;
10906 	unsigned long guid, target_guid;
10907 	int ret = -1;
10908 
10909 	fd_target_dir = open(target_path, O_RDONLY | O_DIRECTORY);
10910 	if (fd_target_dir == -1) {
10911 		return -1;
10912 	}
10913 
10914 	if (fstat(fd_target_dir, &target_stat) == -1) {
10915 		fprintf(stderr, "%s: Failed to stat the target: %s", __func__, strerror(errno));
10916 		exit(1);
10917 	}
10918 
10919 	if (parse_telem_info_file(fd_target_dir, "guid", "%lx", &target_guid)) {
10920 		fprintf(stderr, "%s: Failed to parse the target guid file: %s", __func__, strerror(errno));
10921 		exit(1);
10922 	}
10923 
10924 	close(fd_target_dir);
10925 
10926 	pmt_diriter_init(&pmt_iter);
10927 
10928 	for (dirname = pmt_diriter_begin(&pmt_iter, SYSFS_TELEM_PATH); dirname != NULL; dirname = pmt_diriter_next(&pmt_iter)) {
10929 
10930 		fd_telem_dir = openat(dirfd(pmt_iter.dir), dirname->d_name, O_RDONLY | O_DIRECTORY);
10931 		if (fd_telem_dir == -1)
10932 			continue;
10933 
10934 		if (parse_telem_info_file(fd_telem_dir, "guid", "%lx", &guid)) {
10935 			fprintf(stderr, "%s: Failed to parse the guid file: %s", __func__, strerror(errno));
10936 			continue;
10937 		}
10938 
10939 		if (fstat(fd_telem_dir, &stat) == -1) {
10940 			fprintf(stderr, "%s: Failed to stat %s directory: %s", __func__, dirname->d_name, strerror(errno));
10941 			continue;
10942 		}
10943 
10944 		/*
10945 		 * If reached the same directory as target, exit the loop.
10946 		 * Seq has the correct value now.
10947 		 */
10948 		if (stat.st_dev == target_stat.st_dev && stat.st_ino == target_stat.st_ino) {
10949 			ret = 0;
10950 			break;
10951 		}
10952 
10953 		/*
10954 		 * If reached directory with the same guid,
10955 		 * but it's not the target directory yet,
10956 		 * increment seq and continue the search.
10957 		 */
10958 		if (guid == target_guid)
10959 			++seq;
10960 
10961 		close(fd_telem_dir);
10962 		fd_telem_dir = -1;
10963 	}
10964 
10965 	pmt_diriter_remove(&pmt_iter);
10966 
10967 	if (fd_telem_dir != -1)
10968 		close(fd_telem_dir);
10969 
10970 	if (!ret) {
10971 		*out_guid = target_guid;
10972 		*out_seq = seq;
10973 	}
10974 
10975 	return ret;
10976 }
10977 
parse_add_command_pmt(char * add_command)10978 void parse_add_command_pmt(char *add_command)
10979 {
10980 	char *name = NULL;
10981 	char *type_name = NULL;
10982 	char *format_name = NULL;
10983 	char *direct_path = NULL;
10984 	static const char direct_path_prefix[] = "path=";
10985 	unsigned int offset;
10986 	unsigned int lsb;
10987 	unsigned int msb;
10988 	unsigned int guid;
10989 	unsigned int seq = 0;	/* By default, pick first file in a sequence with a given GUID. */
10990 	unsigned int domain_id;
10991 	enum counter_scope scope = 0;
10992 	enum pmt_datatype type = PMT_TYPE_RAW;
10993 	enum counter_format format = FORMAT_RAW;
10994 	bool has_offset = false;
10995 	bool has_lsb = false;
10996 	bool has_msb = false;
10997 	bool has_format = true;	/* Format has a default value. */
10998 	bool has_guid = false;
10999 	bool has_scope = false;
11000 	bool has_type = true;	/* Type has a default value. */
11001 
11002 	/* Consume the "pmt," prefix. */
11003 	add_command = strchr(add_command, ',');
11004 	if (!add_command) {
11005 		help();
11006 		exit(1);
11007 	}
11008 	++add_command;
11009 
11010 	while (add_command) {
11011 		if (starts_with(add_command, "name=")) {
11012 			name = add_command + strlen("name=");
11013 			goto next;
11014 		}
11015 
11016 		if (starts_with(add_command, "type=")) {
11017 			type_name = add_command + strlen("type=");
11018 			goto next;
11019 		}
11020 
11021 		if (starts_with(add_command, "domain=")) {
11022 			const size_t prefix_len = strlen("domain=");
11023 
11024 			if (sscanf(add_command + prefix_len, "cpu%u", &domain_id) == 1) {
11025 				scope = SCOPE_CPU;
11026 				has_scope = true;
11027 			} else if (sscanf(add_command + prefix_len, "core%u", &domain_id) == 1) {
11028 				scope = SCOPE_CORE;
11029 				has_scope = true;
11030 			} else if (sscanf(add_command + prefix_len, "package%u", &domain_id) == 1) {
11031 				scope = SCOPE_PACKAGE;
11032 				has_scope = true;
11033 			}
11034 
11035 			if (!has_scope) {
11036 				printf("%s: invalid value for scope. Expected cpu%%u, core%%u or package%%u.\n", __func__);
11037 				exit(1);
11038 			}
11039 
11040 			goto next;
11041 		}
11042 
11043 		if (starts_with(add_command, "format=")) {
11044 			format_name = add_command + strlen("format=");
11045 			goto next;
11046 		}
11047 
11048 		if (sscanf(add_command, "offset=%u", &offset) == 1) {
11049 			has_offset = true;
11050 			goto next;
11051 		}
11052 
11053 		if (sscanf(add_command, "lsb=%u", &lsb) == 1) {
11054 			has_lsb = true;
11055 			goto next;
11056 		}
11057 
11058 		if (sscanf(add_command, "msb=%u", &msb) == 1) {
11059 			has_msb = true;
11060 			goto next;
11061 		}
11062 
11063 		if (sscanf(add_command, "guid=%x", &guid) == 1) {
11064 			has_guid = true;
11065 			goto next;
11066 		}
11067 
11068 		if (sscanf(add_command, "seq=%x", &seq) == 1)
11069 			goto next;
11070 
11071 		if (strncmp(add_command, direct_path_prefix, strlen(direct_path_prefix)) == 0) {
11072 			direct_path = add_command + strlen(direct_path_prefix);
11073 			goto next;
11074 		}
11075 next:
11076 		add_command = strchr(add_command, ',');
11077 		if (add_command) {
11078 			*add_command = '\0';
11079 			add_command++;
11080 		}
11081 	}
11082 
11083 	if (!name) {
11084 		printf("%s: missing %s\n", __func__, "name");
11085 		exit(1);
11086 	}
11087 
11088 	if (strlen(name) >= PMT_COUNTER_NAME_SIZE_BYTES) {
11089 		printf("%s: name has to be at most %d characters long\n", __func__, PMT_COUNTER_NAME_SIZE_BYTES);
11090 		exit(1);
11091 	}
11092 
11093 	if (format_name) {
11094 		has_format = false;
11095 
11096 		if (strcmp("raw", format_name) == 0) {
11097 			format = FORMAT_RAW;
11098 			has_format = true;
11099 		}
11100 
11101 		if (strcmp("average", format_name) == 0) {
11102 			format = FORMAT_AVERAGE;
11103 			has_format = true;
11104 		}
11105 
11106 		if (strcmp("delta", format_name) == 0) {
11107 			format = FORMAT_DELTA;
11108 			has_format = true;
11109 		}
11110 
11111 		if (!has_format) {
11112 			fprintf(stderr, "%s: Invalid format %s. Expected raw, average or delta\n", __func__, format_name);
11113 			exit(1);
11114 		}
11115 	}
11116 
11117 	if (type_name) {
11118 		has_type = false;
11119 
11120 		if (strcmp("raw", type_name) == 0) {
11121 			type = PMT_TYPE_RAW;
11122 			has_type = true;
11123 		}
11124 
11125 		if (strcmp("txtal_time", type_name) == 0) {
11126 			type = PMT_TYPE_XTAL_TIME;
11127 			has_type = true;
11128 		}
11129 
11130 		if (strcmp("tcore_clock", type_name) == 0) {
11131 			type = PMT_TYPE_TCORE_CLOCK;
11132 			has_type = true;
11133 		}
11134 
11135 		if (!has_type) {
11136 			printf("%s: invalid %s: %s\n", __func__, "type", type_name);
11137 			exit(1);
11138 		}
11139 	}
11140 
11141 	if (!has_offset) {
11142 		printf("%s : missing %s\n", __func__, "offset");
11143 		exit(1);
11144 	}
11145 
11146 	if (!has_lsb) {
11147 		printf("%s: missing %s\n", __func__, "lsb");
11148 		exit(1);
11149 	}
11150 
11151 	if (!has_msb) {
11152 		printf("%s: missing %s\n", __func__, "msb");
11153 		exit(1);
11154 	}
11155 
11156 	if (direct_path && has_guid) {
11157 		printf("%s: path and guid+seq parameters are mutually exclusive\nnotice: passed guid=0x%x and path=%s\n", __func__, guid, direct_path);
11158 		exit(1);
11159 	}
11160 
11161 	if (direct_path) {
11162 		if (pmt_parse_from_path(direct_path, &guid, &seq)) {
11163 			printf("%s: failed to parse PMT file from %s\n", __func__, direct_path);
11164 			exit(1);
11165 		}
11166 
11167 		/* GUID was just infered from the direct path. */
11168 		has_guid = true;
11169 	}
11170 
11171 	if (!has_guid) {
11172 		printf("%s: missing %s\n", __func__, "guid or path");
11173 		exit(1);
11174 	}
11175 
11176 	if (!has_scope) {
11177 		printf("%s: missing %s\n", __func__, "scope");
11178 		exit(1);
11179 	}
11180 
11181 	if (lsb > msb) {
11182 		printf("%s: lsb > msb doesn't make sense\n", __func__);
11183 		exit(1);
11184 	}
11185 
11186 	pmt_add_counter(guid, seq, name, type, lsb, msb, offset, scope, format, domain_id, PMT_OPEN_REQUIRED);
11187 }
11188 
parse_add_command(char * add_command)11189 void parse_add_command(char *add_command)
11190 {
11191 	if (strncmp(add_command, "pmt", strlen("pmt")) == 0)
11192 		return parse_add_command_pmt(add_command);
11193 	return parse_add_command_msr(add_command);
11194 }
11195 
is_deferred_add(char * name)11196 int is_deferred_add(char *name)
11197 {
11198 	int i;
11199 
11200 	for (i = 0; i < deferred_add_index; ++i)
11201 		if (!strcmp(name, deferred_add_names[i])) {
11202 			deferred_add_consumed |= (1 << i);
11203 			return 1;
11204 		}
11205 	return 0;
11206 }
11207 
is_deferred_skip(char * name)11208 int is_deferred_skip(char *name)
11209 {
11210 	int i;
11211 
11212 	for (i = 0; i < deferred_skip_index; ++i)
11213 		if (!strcmp(name, deferred_skip_names[i])) {
11214 			deferred_skip_consumed |= (1 << i);
11215 			return 1;
11216 		}
11217 	return 0;
11218 }
11219 
verify_deferred_consumed(void)11220 void verify_deferred_consumed(void)
11221 {
11222 	int i;
11223 	int fail = 0;
11224 
11225 	for (i = 0; i < deferred_add_index; ++i) {
11226 		if (!(deferred_add_consumed & (1 << i))) {
11227 			warnx("Counter '%s' can not be added.", deferred_add_names[i]);
11228 			fail++;
11229 		}
11230 	}
11231 	for (i = 0; i < deferred_skip_index; ++i) {
11232 		if (!(deferred_skip_consumed & (1 << i))) {
11233 			warnx("Counter '%s' can not be skipped.", deferred_skip_names[i]);
11234 			fail++;
11235 		}
11236 	}
11237 	if (fail)
11238 		exit(-EINVAL);
11239 }
11240 
probe_cpuidle_residency(void)11241 void probe_cpuidle_residency(void)
11242 {
11243 	char path[64];
11244 	char name_buf[16];
11245 	FILE *input;
11246 	int state;
11247 	int min_state = 1024, max_state = 0;
11248 	char *sp;
11249 
11250 	for (state = 10; state >= 0; --state) {
11251 
11252 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", master_cpu, state);
11253 		input = fopen(path, "r");
11254 		if (input == NULL)
11255 			continue;
11256 		if (!fgets(name_buf, sizeof(name_buf), input))
11257 			err(1, "%s: failed to read file", path);
11258 
11259 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
11260 		sp = strchr(name_buf, '-');
11261 		if (!sp)
11262 			sp = strchrnul(name_buf, '\n');
11263 		*sp = '%';
11264 		*(sp + 1) = '\0';
11265 
11266 		remove_underbar(name_buf);
11267 
11268 		fclose(input);
11269 
11270 		sprintf(path, "cpuidle/state%d/time", state);
11271 
11272 		if (!DO_BIC(BIC_pct_idle) && !is_deferred_add(name_buf))
11273 			continue;
11274 
11275 		if (is_deferred_skip(name_buf))
11276 			continue;
11277 
11278 		add_counter(0, path, name_buf, 32, SCOPE_CPU, COUNTER_USEC, FORMAT_PERCENT, SYSFS_PERCPU, 0);
11279 
11280 		if (state > max_state)
11281 			max_state = state;
11282 		if (state < min_state)
11283 			min_state = state;
11284 	}
11285 }
11286 
probe_cpuidle_counts(void)11287 void probe_cpuidle_counts(void)
11288 {
11289 	char path[64];
11290 	char name_buf[16];
11291 	FILE *input;
11292 	int state;
11293 	int min_state = 1024, max_state = 0;
11294 	char *sp;
11295 
11296 	if (!DO_BIC(BIC_cpuidle))
11297 		return;
11298 
11299 	for (state = 10; state >= 0; --state) {
11300 
11301 		sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name", master_cpu, state);
11302 		input = fopen(path, "r");
11303 		if (input == NULL)
11304 			continue;
11305 		if (!fgets(name_buf, sizeof(name_buf), input))
11306 			err(1, "%s: failed to read file", path);
11307 		fclose(input);
11308 
11309 		remove_underbar(name_buf);
11310 
11311 		if (!DO_BIC(BIC_cpuidle) && !is_deferred_add(name_buf))
11312 			continue;
11313 
11314 		if (is_deferred_skip(name_buf))
11315 			continue;
11316 
11317 		/* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
11318 		sp = strchr(name_buf, '-');
11319 		if (!sp)
11320 			sp = strchrnul(name_buf, '\n');
11321 
11322 		/*
11323 		 * The 'below' sysfs file always contains 0 for the deepest state (largest index),
11324 		 * do not add it.
11325 		 */
11326 		if (state != max_state) {
11327 			/*
11328 			 * Add 'C1+' for C1, and so on. The 'below' sysfs file always contains 0 for
11329 			 * the last state, so do not add it.
11330 			 */
11331 
11332 			*sp = '+';
11333 			*(sp + 1) = '\0';
11334 			sprintf(path, "cpuidle/state%d/below", state);
11335 			add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS, FORMAT_DELTA, SYSFS_PERCPU, 0);
11336 		}
11337 
11338 		*sp = '\0';
11339 		sprintf(path, "cpuidle/state%d/usage", state);
11340 		add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS, FORMAT_DELTA, SYSFS_PERCPU, 0);
11341 
11342 		/*
11343 		 * The 'above' sysfs file always contains 0 for the shallowest state (smallest
11344 		 * index), do not add it.
11345 		 */
11346 		if (state != min_state) {
11347 			*sp = '-';
11348 			*(sp + 1) = '\0';
11349 			sprintf(path, "cpuidle/state%d/above", state);
11350 			add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS, FORMAT_DELTA, SYSFS_PERCPU, 0);
11351 		}
11352 	}
11353 }
11354 
11355 /*
11356  * parse cpuset with following syntax
11357  * 1,2,4..6,8-10 and set bits in cpu_subset
11358  */
parse_cpu_command(char * optarg)11359 void parse_cpu_command(char *optarg)
11360 {
11361 	if (!strcmp(optarg, "core")) {
11362 		if (cpu_subset)
11363 			goto error;
11364 		show_core_only++;
11365 		return;
11366 	}
11367 	if (!strcmp(optarg, "package")) {
11368 		if (cpu_subset)
11369 			goto error;
11370 		show_pkg_only++;
11371 		return;
11372 	}
11373 	if (show_core_only || show_pkg_only)
11374 		goto error;
11375 
11376 	cpu_subset = CPU_ALLOC(CPU_SUBSET_MAXCPUS);
11377 	if (cpu_subset == NULL)
11378 		err(3, "CPU_ALLOC");
11379 	cpu_subset_size = CPU_ALLOC_SIZE(CPU_SUBSET_MAXCPUS);
11380 
11381 	CPU_ZERO_S(cpu_subset_size, cpu_subset);
11382 
11383 	if (parse_cpu_str(optarg, cpu_subset, cpu_subset_size))
11384 		goto error;
11385 
11386 	return;
11387 
11388 error:
11389 	fprintf(stderr, "\"--cpu %s\" malformed\n", optarg);
11390 	help();
11391 	exit(-1);
11392 }
11393 
cmdline(int argc,char ** argv)11394 void cmdline(int argc, char **argv)
11395 {
11396 	int opt;
11397 	int option_index = 0;
11398 	static struct option long_options[] = {
11399 		{ "add", required_argument, 0, 'a' },
11400 		{ "cpu", required_argument, 0, 'c' },
11401 		{ "Dump", no_argument, 0, 'D' },
11402 		{ "debug", no_argument, 0, 'd' },	/* internal, not documented */
11403 		{ "enable", required_argument, 0, 'e' },
11404 		{ "force", no_argument, 0, 'f' },
11405 		{ "interval", required_argument, 0, 'i' },
11406 		{ "IPC", no_argument, 0, 'I' },
11407 		{ "num_iterations", required_argument, 0, 'n' },
11408 		{ "header_iterations", required_argument, 0, 'N' },
11409 		{ "help", no_argument, 0, 'h' },
11410 		{ "hide", required_argument, 0, 'H' },	// meh, -h taken by --help
11411 		{ "Joules", no_argument, 0, 'J' },
11412 		{ "list", no_argument, 0, 'l' },
11413 		{ "out", required_argument, 0, 'o' },
11414 		{ "quiet", no_argument, 0, 'q' },
11415 		{ "no-msr", no_argument, 0, 'M' },
11416 		{ "no-perf", no_argument, 0, 'P' },
11417 		{ "show", required_argument, 0, 's' },
11418 		{ "Summary", no_argument, 0, 'S' },
11419 		{ "TCC", required_argument, 0, 'T' },
11420 		{ "version", no_argument, 0, 'v' },
11421 		{ 0, 0, 0, 0 }
11422 	};
11423 
11424 	progname = argv[0];
11425 
11426 	/*
11427 	 * Parse some options early, because they may make other options invalid,
11428 	 * like adding the MSR counter with --add and at the same time using --no-msr.
11429 	 */
11430 	while ((opt = getopt_long_only(argc, argv, "+:MP", long_options, &option_index)) != -1) {
11431 		switch (opt) {
11432 		case 'M':
11433 			no_msr = 1;
11434 			break;
11435 		case 'P':
11436 			no_perf = 1;
11437 			break;
11438 		default:
11439 			break;
11440 		}
11441 	}
11442 	optind = 0;
11443 
11444 	while ((opt = getopt_long_only(argc, argv, "+C:c:Dde:hi:Jn:N:o:qMST:v", long_options, &option_index)) != -1) {
11445 		switch (opt) {
11446 		case 'a':
11447 			parse_add_command(optarg);
11448 			break;
11449 		case 'c':
11450 			parse_cpu_command(optarg);
11451 			break;
11452 		case 'D':
11453 			dump_only++;
11454 			/*
11455 			 * Force the no_perf early to prevent using it as a source.
11456 			 * User asks for raw values, but perf returns them relative
11457 			 * to the opening of the file descriptor.
11458 			 */
11459 			no_perf = 1;
11460 			break;
11461 		case 'e':
11462 			/* --enable specified counter, without clearning existing list */
11463 			bic_lookup(&bic_enabled, optarg, SHOW_LIST);
11464 			break;
11465 		case 'f':
11466 			force_load++;
11467 			break;
11468 		case 'd':
11469 			debug++;
11470 			bic_set_all(&bic_enabled);
11471 			break;
11472 		case 'H':
11473 			/*
11474 			 * --hide: do not show those specified
11475 			 *  multiple invocations simply clear more bits in enabled mask
11476 			 */
11477 			{
11478 				cpu_set_t bic_group_hide;
11479 
11480 				BIC_INIT(&bic_group_hide);
11481 
11482 				bic_lookup(&bic_group_hide, optarg, HIDE_LIST);
11483 				bic_clear_bits(&bic_enabled, &bic_group_hide);
11484 			}
11485 			break;
11486 		case 'h':
11487 			help();
11488 			exit(1);
11489 		case 'i':
11490 			{
11491 				double interval = strtod(optarg, NULL);
11492 
11493 				if (interval < 0.001) {
11494 					fprintf(outf, "interval %f seconds is too small\n", interval);
11495 					exit(2);
11496 				}
11497 
11498 				interval_tv.tv_sec = interval_ts.tv_sec = interval;
11499 				interval_tv.tv_usec = (interval - interval_tv.tv_sec) * 1000000;
11500 				interval_ts.tv_nsec = (interval - interval_ts.tv_sec) * 1000000000;
11501 			}
11502 			break;
11503 		case 'J':
11504 			rapl_joules++;
11505 			break;
11506 		case 'l':
11507 			bic_set_all(&bic_enabled);
11508 			list_header_only++;
11509 			quiet++;
11510 			break;
11511 		case 'o':
11512 			outf = fopen_or_die(optarg, "w");
11513 			break;
11514 		case 'q':
11515 			quiet = 1;
11516 			break;
11517 		case 'M':
11518 		case 'P':
11519 			/* Parsed earlier */
11520 			break;
11521 		case 'n':
11522 			num_iterations = strtoul(optarg, NULL, 0);
11523 			errno = 0;
11524 
11525 			if (errno || num_iterations == 0)
11526 				errx(-1, "invalid iteration count: %s", optarg);
11527 			break;
11528 		case 'N':
11529 			header_iterations = strtoul(optarg, NULL, 0);
11530 			errno = 0;
11531 
11532 			if (errno || header_iterations == 0)
11533 				errx(-1, "invalid header iteration count: %s", optarg);
11534 			break;
11535 		case 's':
11536 			/*
11537 			 * --show: show only those specified
11538 			 *  The 1st invocation will clear and replace the enabled mask
11539 			 *  subsequent invocations can add to it.
11540 			 */
11541 			if (shown == 0)
11542 				BIC_INIT(&bic_enabled);
11543 			bic_lookup(&bic_enabled, optarg, SHOW_LIST);
11544 			shown = 1;
11545 			break;
11546 		case 'S':
11547 			summary_only++;
11548 			break;
11549 		case 'T':
11550 			tj_max_override = atoi(optarg);
11551 			break;
11552 		case 'v':
11553 			print_version();
11554 			exit(0);
11555 			break;
11556 		default:
11557 			help();
11558 			exit(1);
11559 		}
11560 	}
11561 }
11562 
set_rlimit(void)11563 void set_rlimit(void)
11564 {
11565 	struct rlimit limit;
11566 
11567 	if (getrlimit(RLIMIT_NOFILE, &limit) < 0)
11568 		err(1, "Failed to get rlimit");
11569 
11570 	if (limit.rlim_max < MAX_NOFILE)
11571 		limit.rlim_max = MAX_NOFILE;
11572 	if (limit.rlim_cur < MAX_NOFILE)
11573 		limit.rlim_cur = MAX_NOFILE;
11574 
11575 	if (setrlimit(RLIMIT_NOFILE, &limit) < 0)
11576 		err(1, "Failed to set rlimit");
11577 }
11578 
main(int argc,char ** argv)11579 int main(int argc, char **argv)
11580 {
11581 	int fd, ret;
11582 
11583 	bic_groups_init();
11584 
11585 	fd = open("/sys/fs/cgroup/cgroup.procs", O_WRONLY);
11586 	if (fd < 0)
11587 		goto skip_cgroup_setting;
11588 
11589 	ret = write(fd, "0\n", 2);
11590 	if (ret == -1)
11591 		perror("Can't update cgroup\n");
11592 
11593 	close(fd);
11594 
11595 skip_cgroup_setting:
11596 	outf = stderr;
11597 	cmdline(argc, argv);
11598 
11599 	if (!quiet) {
11600 		print_version();
11601 		print_bootcmd();
11602 	}
11603 
11604 	probe_cpuidle_residency();
11605 	probe_cpuidle_counts();
11606 
11607 	verify_deferred_consumed();
11608 
11609 	if (!getuid())
11610 		set_rlimit();
11611 
11612 	turbostat_init();
11613 
11614 	if (!no_msr)
11615 		msr_sum_record();
11616 
11617 	/* dump counters and exit */
11618 	if (dump_only)
11619 		return get_and_dump_counters();
11620 
11621 	/* list header and exit */
11622 	if (list_header_only) {
11623 		print_header(",");
11624 		flush_output_stdout();
11625 		return 0;
11626 	}
11627 
11628 	/*
11629 	 * if any params left, it must be a command to fork
11630 	 */
11631 	if (argc - optind)
11632 		return fork_it(argv + optind);
11633 	else
11634 		turbostat_loop();
11635 
11636 	return 0;
11637 }
11638