xref: /openbmc/linux/tools/perf/util/event.h (revision 305c8388)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef __PERF_RECORD_H
3 #define __PERF_RECORD_H
4 /*
5  * The linux/stddef.h isn't need here, but is needed for __always_inline used
6  * in files included from uapi/linux/perf_event.h such as
7  * /usr/include/linux/swab.h and /usr/include/linux/byteorder/little_endian.h,
8  * detected in at least musl libc, used in Alpine Linux. -acme
9  */
10 #include <stdio.h>
11 #include <linux/stddef.h>
12 #include <perf/event.h>
13 #include <linux/types.h>
14 
15 #include "perf_regs.h"
16 
17 struct dso;
18 struct machine;
19 struct perf_event_attr;
20 
21 #ifdef __LP64__
22 /*
23  * /usr/include/inttypes.h uses just 'lu' for PRIu64, but we end up defining
24  * __u64 as long long unsigned int, and then -Werror=format= kicks in and
25  * complains of the mismatched types, so use these two special extra PRI
26  * macros to overcome that.
27  */
28 #define PRI_lu64 "l" PRIu64
29 #define PRI_lx64 "l" PRIx64
30 #define PRI_ld64 "l" PRId64
31 #else
32 #define PRI_lu64 PRIu64
33 #define PRI_lx64 PRIx64
34 #define PRI_ld64 PRId64
35 #endif
36 
37 #define PERF_SAMPLE_MASK				\
38 	(PERF_SAMPLE_IP | PERF_SAMPLE_TID |		\
39 	 PERF_SAMPLE_TIME | PERF_SAMPLE_ADDR |		\
40 	PERF_SAMPLE_ID | PERF_SAMPLE_STREAM_ID |	\
41 	 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD |		\
42 	 PERF_SAMPLE_IDENTIFIER)
43 
44 /* perf sample has 16 bits size limit */
45 #define PERF_SAMPLE_MAX_SIZE (1 << 16)
46 
47 struct regs_dump {
48 	u64 abi;
49 	u64 mask;
50 	u64 *regs;
51 
52 	/* Cached values/mask filled by first register access. */
53 	u64 cache_regs[PERF_REGS_MAX];
54 	u64 cache_mask;
55 };
56 
57 struct stack_dump {
58 	u16 offset;
59 	u64 size;
60 	char *data;
61 };
62 
63 struct sample_read_value {
64 	u64 value;
65 	u64 id;
66 };
67 
68 struct sample_read {
69 	u64 time_enabled;
70 	u64 time_running;
71 	union {
72 		struct {
73 			u64 nr;
74 			struct sample_read_value *values;
75 		} group;
76 		struct sample_read_value one;
77 	};
78 };
79 
80 struct ip_callchain {
81 	u64 nr;
82 	u64 ips[];
83 };
84 
85 struct branch_stack;
86 
87 enum {
88 	PERF_IP_FLAG_BRANCH		= 1ULL << 0,
89 	PERF_IP_FLAG_CALL		= 1ULL << 1,
90 	PERF_IP_FLAG_RETURN		= 1ULL << 2,
91 	PERF_IP_FLAG_CONDITIONAL	= 1ULL << 3,
92 	PERF_IP_FLAG_SYSCALLRET		= 1ULL << 4,
93 	PERF_IP_FLAG_ASYNC		= 1ULL << 5,
94 	PERF_IP_FLAG_INTERRUPT		= 1ULL << 6,
95 	PERF_IP_FLAG_TX_ABORT		= 1ULL << 7,
96 	PERF_IP_FLAG_TRACE_BEGIN	= 1ULL << 8,
97 	PERF_IP_FLAG_TRACE_END		= 1ULL << 9,
98 	PERF_IP_FLAG_IN_TX		= 1ULL << 10,
99 };
100 
101 #define PERF_IP_FLAG_CHARS "bcrosyiABEx"
102 
103 #define PERF_BRANCH_MASK		(\
104 	PERF_IP_FLAG_BRANCH		|\
105 	PERF_IP_FLAG_CALL		|\
106 	PERF_IP_FLAG_RETURN		|\
107 	PERF_IP_FLAG_CONDITIONAL	|\
108 	PERF_IP_FLAG_SYSCALLRET		|\
109 	PERF_IP_FLAG_ASYNC		|\
110 	PERF_IP_FLAG_INTERRUPT		|\
111 	PERF_IP_FLAG_TX_ABORT		|\
112 	PERF_IP_FLAG_TRACE_BEGIN	|\
113 	PERF_IP_FLAG_TRACE_END)
114 
115 #define MAX_INSN 16
116 
117 struct aux_sample {
118 	u64 size;
119 	void *data;
120 };
121 
122 struct perf_sample {
123 	u64 ip;
124 	u32 pid, tid;
125 	u64 time;
126 	u64 addr;
127 	u64 id;
128 	u64 stream_id;
129 	u64 period;
130 	u64 weight;
131 	u64 transaction;
132 	u64 insn_cnt;
133 	u64 cyc_cnt;
134 	u32 cpu;
135 	u32 raw_size;
136 	u64 data_src;
137 	u64 phys_addr;
138 	u64 cgroup;
139 	u32 flags;
140 	u16 insn_len;
141 	u8  cpumode;
142 	u16 misc;
143 	bool no_hw_idx;		/* No hw_idx collected in branch_stack */
144 	char insn[MAX_INSN];
145 	void *raw_data;
146 	struct ip_callchain *callchain;
147 	struct branch_stack *branch_stack;
148 	struct regs_dump  user_regs;
149 	struct regs_dump  intr_regs;
150 	struct stack_dump user_stack;
151 	struct sample_read read;
152 	struct aux_sample aux_sample;
153 };
154 
155 #define PERF_MEM_DATA_SRC_NONE \
156 	(PERF_MEM_S(OP, NA) |\
157 	 PERF_MEM_S(LVL, NA) |\
158 	 PERF_MEM_S(SNOOP, NA) |\
159 	 PERF_MEM_S(LOCK, NA) |\
160 	 PERF_MEM_S(TLB, NA))
161 
162 /* Attribute type for custom synthesized events */
163 #define PERF_TYPE_SYNTH		(INT_MAX + 1U)
164 
165 /* Attribute config for custom synthesized events */
166 enum perf_synth_id {
167 	PERF_SYNTH_INTEL_PTWRITE,
168 	PERF_SYNTH_INTEL_MWAIT,
169 	PERF_SYNTH_INTEL_PWRE,
170 	PERF_SYNTH_INTEL_EXSTOP,
171 	PERF_SYNTH_INTEL_PWRX,
172 	PERF_SYNTH_INTEL_CBR,
173 };
174 
175 /*
176  * Raw data formats for synthesized events. Note that 4 bytes of padding are
177  * present to match the 'size' member of PERF_SAMPLE_RAW data which is always
178  * 8-byte aligned. That means we must dereference raw_data with an offset of 4.
179  * Refer perf_sample__synth_ptr() and perf_synth__raw_data().  It also means the
180  * structure sizes are 4 bytes bigger than the raw_size, refer
181  * perf_synth__raw_size().
182  */
183 
184 struct perf_synth_intel_ptwrite {
185 	u32 padding;
186 	union {
187 		struct {
188 			u32	ip		:  1,
189 				reserved	: 31;
190 		};
191 		u32	flags;
192 	};
193 	u64	payload;
194 };
195 
196 struct perf_synth_intel_mwait {
197 	u32 padding;
198 	u32 reserved;
199 	union {
200 		struct {
201 			u64	hints		:  8,
202 				reserved1	: 24,
203 				extensions	:  2,
204 				reserved2	: 30;
205 		};
206 		u64	payload;
207 	};
208 };
209 
210 struct perf_synth_intel_pwre {
211 	u32 padding;
212 	u32 reserved;
213 	union {
214 		struct {
215 			u64	reserved1	:  7,
216 				hw		:  1,
217 				subcstate	:  4,
218 				cstate		:  4,
219 				reserved2	: 48;
220 		};
221 		u64	payload;
222 	};
223 };
224 
225 struct perf_synth_intel_exstop {
226 	u32 padding;
227 	union {
228 		struct {
229 			u32	ip		:  1,
230 				reserved	: 31;
231 		};
232 		u32	flags;
233 	};
234 };
235 
236 struct perf_synth_intel_pwrx {
237 	u32 padding;
238 	u32 reserved;
239 	union {
240 		struct {
241 			u64	deepest_cstate	:  4,
242 				last_cstate	:  4,
243 				wake_reason	:  4,
244 				reserved1	: 52;
245 		};
246 		u64	payload;
247 	};
248 };
249 
250 struct perf_synth_intel_cbr {
251 	u32 padding;
252 	union {
253 		struct {
254 			u32	cbr		:  8,
255 				reserved1	:  8,
256 				max_nonturbo	:  8,
257 				reserved2	:  8;
258 		};
259 		u32	flags;
260 	};
261 	u32 freq;
262 	u32 reserved3;
263 };
264 
265 /*
266  * raw_data is always 4 bytes from an 8-byte boundary, so subtract 4 to get
267  * 8-byte alignment.
268  */
269 static inline void *perf_sample__synth_ptr(struct perf_sample *sample)
270 {
271 	return sample->raw_data - 4;
272 }
273 
274 static inline void *perf_synth__raw_data(void *p)
275 {
276 	return p + 4;
277 }
278 
279 #define perf_synth__raw_size(d) (sizeof(d) - 4)
280 
281 #define perf_sample__bad_synth_size(s, d) ((s)->raw_size < sizeof(d) - 4)
282 
283 enum {
284 	PERF_STAT_ROUND_TYPE__INTERVAL	= 0,
285 	PERF_STAT_ROUND_TYPE__FINAL	= 1,
286 };
287 
288 void perf_event__print_totals(void);
289 
290 struct perf_cpu_map;
291 struct perf_record_stat_config;
292 struct perf_stat_config;
293 struct perf_tool;
294 
295 void perf_event__read_stat_config(struct perf_stat_config *config,
296 				  struct perf_record_stat_config *event);
297 
298 int perf_event__process_comm(struct perf_tool *tool,
299 			     union perf_event *event,
300 			     struct perf_sample *sample,
301 			     struct machine *machine);
302 int perf_event__process_lost(struct perf_tool *tool,
303 			     union perf_event *event,
304 			     struct perf_sample *sample,
305 			     struct machine *machine);
306 int perf_event__process_lost_samples(struct perf_tool *tool,
307 				     union perf_event *event,
308 				     struct perf_sample *sample,
309 				     struct machine *machine);
310 int perf_event__process_aux(struct perf_tool *tool,
311 			    union perf_event *event,
312 			    struct perf_sample *sample,
313 			    struct machine *machine);
314 int perf_event__process_itrace_start(struct perf_tool *tool,
315 				     union perf_event *event,
316 				     struct perf_sample *sample,
317 				     struct machine *machine);
318 int perf_event__process_switch(struct perf_tool *tool,
319 			       union perf_event *event,
320 			       struct perf_sample *sample,
321 			       struct machine *machine);
322 int perf_event__process_namespaces(struct perf_tool *tool,
323 				   union perf_event *event,
324 				   struct perf_sample *sample,
325 				   struct machine *machine);
326 int perf_event__process_cgroup(struct perf_tool *tool,
327 			       union perf_event *event,
328 			       struct perf_sample *sample,
329 			       struct machine *machine);
330 int perf_event__process_mmap(struct perf_tool *tool,
331 			     union perf_event *event,
332 			     struct perf_sample *sample,
333 			     struct machine *machine);
334 int perf_event__process_mmap2(struct perf_tool *tool,
335 			     union perf_event *event,
336 			     struct perf_sample *sample,
337 			     struct machine *machine);
338 int perf_event__process_fork(struct perf_tool *tool,
339 			     union perf_event *event,
340 			     struct perf_sample *sample,
341 			     struct machine *machine);
342 int perf_event__process_exit(struct perf_tool *tool,
343 			     union perf_event *event,
344 			     struct perf_sample *sample,
345 			     struct machine *machine);
346 int perf_event__process_ksymbol(struct perf_tool *tool,
347 				union perf_event *event,
348 				struct perf_sample *sample,
349 				struct machine *machine);
350 int perf_event__process_bpf(struct perf_tool *tool,
351 			    union perf_event *event,
352 			    struct perf_sample *sample,
353 			    struct machine *machine);
354 int perf_event__process(struct perf_tool *tool,
355 			union perf_event *event,
356 			struct perf_sample *sample,
357 			struct machine *machine);
358 
359 struct addr_location;
360 
361 int machine__resolve(struct machine *machine, struct addr_location *al,
362 		     struct perf_sample *sample);
363 
364 void addr_location__put(struct addr_location *al);
365 
366 struct thread;
367 
368 bool is_bts_event(struct perf_event_attr *attr);
369 bool sample_addr_correlates_sym(struct perf_event_attr *attr);
370 void thread__resolve(struct thread *thread, struct addr_location *al,
371 		     struct perf_sample *sample);
372 
373 const char *perf_event__name(unsigned int id);
374 
375 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp);
376 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp);
377 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp);
378 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp);
379 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp);
380 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp);
381 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp);
382 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp);
383 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp);
384 size_t perf_event__fprintf_namespaces(union perf_event *event, FILE *fp);
385 size_t perf_event__fprintf_cgroup(union perf_event *event, FILE *fp);
386 size_t perf_event__fprintf_ksymbol(union perf_event *event, FILE *fp);
387 size_t perf_event__fprintf_bpf(union perf_event *event, FILE *fp);
388 size_t perf_event__fprintf(union perf_event *event, FILE *fp);
389 
390 int kallsyms__get_function_start(const char *kallsyms_filename,
391 				 const char *symbol_name, u64 *addr);
392 
393 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max);
394 void  cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map,
395 			       u16 type, int max);
396 
397 void event_attr_init(struct perf_event_attr *attr);
398 
399 int perf_event_paranoid(void);
400 bool perf_event_paranoid_check(int max_level);
401 
402 extern int sysctl_perf_event_max_stack;
403 extern int sysctl_perf_event_max_contexts_per_stack;
404 extern unsigned int proc_map_timeout;
405 
406 #endif /* __PERF_RECORD_H */
407