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[0];
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 perf_sample {
118 u64 ip;
119 u32 pid, tid;
120 u64 time;
121 u64 addr;
122 u64 id;
123 u64 stream_id;
124 u64 period;
125 u64 weight;
126 u64 transaction;
127 u64 insn_cnt;
128 u64 cyc_cnt;
129 u32 cpu;
130 u32 raw_size;
131 u64 data_src;
132 u64 phys_addr;
133 u32 flags;
134 u16 insn_len;
135 u8 cpumode;
136 u16 misc;
137 char insn[MAX_INSN];
138 void *raw_data;
139 struct ip_callchain *callchain;
140 struct branch_stack *branch_stack;
141 struct regs_dump user_regs;
142 struct regs_dump intr_regs;
143 struct stack_dump user_stack;
144 struct sample_read read;
145 };
146
147 #define PERF_MEM_DATA_SRC_NONE \
148 (PERF_MEM_S(OP, NA) |\
149 PERF_MEM_S(LVL, NA) |\
150 PERF_MEM_S(SNOOP, NA) |\
151 PERF_MEM_S(LOCK, NA) |\
152 PERF_MEM_S(TLB, NA))
153
154 /* Attribute type for custom synthesized events */
155 #define PERF_TYPE_SYNTH (INT_MAX + 1U)
156
157 /* Attribute config for custom synthesized events */
158 enum perf_synth_id {
159 PERF_SYNTH_INTEL_PTWRITE,
160 PERF_SYNTH_INTEL_MWAIT,
161 PERF_SYNTH_INTEL_PWRE,
162 PERF_SYNTH_INTEL_EXSTOP,
163 PERF_SYNTH_INTEL_PWRX,
164 PERF_SYNTH_INTEL_CBR,
165 };
166
167 /*
168 * Raw data formats for synthesized events. Note that 4 bytes of padding are
169 * present to match the 'size' member of PERF_SAMPLE_RAW data which is always
170 * 8-byte aligned. That means we must dereference raw_data with an offset of 4.
171 * Refer perf_sample__synth_ptr() and perf_synth__raw_data(). It also means the
172 * structure sizes are 4 bytes bigger than the raw_size, refer
173 * perf_synth__raw_size().
174 */
175
176 struct perf_synth_intel_ptwrite {
177 u32 padding;
178 union {
179 struct {
180 u32 ip : 1,
181 reserved : 31;
182 };
183 u32 flags;
184 };
185 u64 payload;
186 };
187
188 struct perf_synth_intel_mwait {
189 u32 padding;
190 u32 reserved;
191 union {
192 struct {
193 u64 hints : 8,
194 reserved1 : 24,
195 extensions : 2,
196 reserved2 : 30;
197 };
198 u64 payload;
199 };
200 };
201
202 struct perf_synth_intel_pwre {
203 u32 padding;
204 u32 reserved;
205 union {
206 struct {
207 u64 reserved1 : 7,
208 hw : 1,
209 subcstate : 4,
210 cstate : 4,
211 reserved2 : 48;
212 };
213 u64 payload;
214 };
215 };
216
217 struct perf_synth_intel_exstop {
218 u32 padding;
219 union {
220 struct {
221 u32 ip : 1,
222 reserved : 31;
223 };
224 u32 flags;
225 };
226 };
227
228 struct perf_synth_intel_pwrx {
229 u32 padding;
230 u32 reserved;
231 union {
232 struct {
233 u64 deepest_cstate : 4,
234 last_cstate : 4,
235 wake_reason : 4,
236 reserved1 : 52;
237 };
238 u64 payload;
239 };
240 };
241
242 struct perf_synth_intel_cbr {
243 u32 padding;
244 union {
245 struct {
246 u32 cbr : 8,
247 reserved1 : 8,
248 max_nonturbo : 8,
249 reserved2 : 8;
250 };
251 u32 flags;
252 };
253 u32 freq;
254 u32 reserved3;
255 };
256
257 /*
258 * raw_data is always 4 bytes from an 8-byte boundary, so subtract 4 to get
259 * 8-byte alignment.
260 */
perf_sample__synth_ptr(struct perf_sample * sample)261 static inline void *perf_sample__synth_ptr(struct perf_sample *sample)
262 {
263 return sample->raw_data - 4;
264 }
265
perf_synth__raw_data(void * p)266 static inline void *perf_synth__raw_data(void *p)
267 {
268 return p + 4;
269 }
270
271 #define perf_synth__raw_size(d) (sizeof(d) - 4)
272
273 #define perf_sample__bad_synth_size(s, d) ((s)->raw_size < sizeof(d) - 4)
274
275 enum {
276 PERF_STAT_ROUND_TYPE__INTERVAL = 0,
277 PERF_STAT_ROUND_TYPE__FINAL = 1,
278 };
279
280 void perf_event__print_totals(void);
281
282 struct perf_cpu_map;
283 struct perf_record_stat_config;
284 struct perf_stat_config;
285 struct perf_tool;
286
287 void perf_event__read_stat_config(struct perf_stat_config *config,
288 struct perf_record_stat_config *event);
289
290 int perf_event__process_comm(struct perf_tool *tool,
291 union perf_event *event,
292 struct perf_sample *sample,
293 struct machine *machine);
294 int perf_event__process_lost(struct perf_tool *tool,
295 union perf_event *event,
296 struct perf_sample *sample,
297 struct machine *machine);
298 int perf_event__process_lost_samples(struct perf_tool *tool,
299 union perf_event *event,
300 struct perf_sample *sample,
301 struct machine *machine);
302 int perf_event__process_aux(struct perf_tool *tool,
303 union perf_event *event,
304 struct perf_sample *sample,
305 struct machine *machine);
306 int perf_event__process_itrace_start(struct perf_tool *tool,
307 union perf_event *event,
308 struct perf_sample *sample,
309 struct machine *machine);
310 int perf_event__process_switch(struct perf_tool *tool,
311 union perf_event *event,
312 struct perf_sample *sample,
313 struct machine *machine);
314 int perf_event__process_namespaces(struct perf_tool *tool,
315 union perf_event *event,
316 struct perf_sample *sample,
317 struct machine *machine);
318 int perf_event__process_mmap(struct perf_tool *tool,
319 union perf_event *event,
320 struct perf_sample *sample,
321 struct machine *machine);
322 int perf_event__process_mmap2(struct perf_tool *tool,
323 union perf_event *event,
324 struct perf_sample *sample,
325 struct machine *machine);
326 int perf_event__process_fork(struct perf_tool *tool,
327 union perf_event *event,
328 struct perf_sample *sample,
329 struct machine *machine);
330 int perf_event__process_exit(struct perf_tool *tool,
331 union perf_event *event,
332 struct perf_sample *sample,
333 struct machine *machine);
334 int perf_event__process_ksymbol(struct perf_tool *tool,
335 union perf_event *event,
336 struct perf_sample *sample,
337 struct machine *machine);
338 int perf_event__process_bpf(struct perf_tool *tool,
339 union perf_event *event,
340 struct perf_sample *sample,
341 struct machine *machine);
342 int perf_event__process(struct perf_tool *tool,
343 union perf_event *event,
344 struct perf_sample *sample,
345 struct machine *machine);
346
347 struct addr_location;
348
349 int machine__resolve(struct machine *machine, struct addr_location *al,
350 struct perf_sample *sample);
351
352 void addr_location__put(struct addr_location *al);
353
354 struct thread;
355
356 bool is_bts_event(struct perf_event_attr *attr);
357 bool sample_addr_correlates_sym(struct perf_event_attr *attr);
358 void thread__resolve(struct thread *thread, struct addr_location *al,
359 struct perf_sample *sample);
360
361 const char *perf_event__name(unsigned int id);
362
363 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp);
364 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp);
365 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp);
366 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp);
367 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp);
368 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp);
369 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp);
370 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp);
371 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp);
372 size_t perf_event__fprintf_namespaces(union perf_event *event, FILE *fp);
373 size_t perf_event__fprintf_ksymbol(union perf_event *event, FILE *fp);
374 size_t perf_event__fprintf_bpf(union perf_event *event, FILE *fp);
375 size_t perf_event__fprintf(union perf_event *event, FILE *fp);
376
377 int kallsyms__get_function_start(const char *kallsyms_filename,
378 const char *symbol_name, u64 *addr);
379
380 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max);
381 void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map,
382 u16 type, int max);
383
384 void event_attr_init(struct perf_event_attr *attr);
385
386 int perf_event_paranoid(void);
387 bool perf_event_paranoid_check(int max_level);
388
389 extern int sysctl_perf_event_max_stack;
390 extern int sysctl_perf_event_max_contexts_per_stack;
391 extern unsigned int proc_map_timeout;
392
393 #endif /* __PERF_RECORD_H */
394
