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 /* number of register is bound by the number of bits in regs_dump::mask (64) */
48 #define PERF_SAMPLE_REGS_CACHE_SIZE (8 * sizeof(u64))
49
50 struct regs_dump {
51 u64 abi;
52 u64 mask;
53 u64 *regs;
54
55 /* Cached values/mask filled by first register access. */
56 u64 cache_regs[PERF_SAMPLE_REGS_CACHE_SIZE];
57 u64 cache_mask;
58 };
59
60 struct stack_dump {
61 u16 offset;
62 u64 size;
63 char *data;
64 };
65
66 struct sample_read_value {
67 u64 value;
68 u64 id;
69 };
70
71 struct sample_read {
72 u64 time_enabled;
73 u64 time_running;
74 union {
75 struct {
76 u64 nr;
77 struct sample_read_value *values;
78 } group;
79 struct sample_read_value one;
80 };
81 };
82
83 struct ip_callchain {
84 u64 nr;
85 u64 ips[];
86 };
87
88 struct branch_stack;
89
90 enum {
91 PERF_IP_FLAG_BRANCH = 1ULL << 0,
92 PERF_IP_FLAG_CALL = 1ULL << 1,
93 PERF_IP_FLAG_RETURN = 1ULL << 2,
94 PERF_IP_FLAG_CONDITIONAL = 1ULL << 3,
95 PERF_IP_FLAG_SYSCALLRET = 1ULL << 4,
96 PERF_IP_FLAG_ASYNC = 1ULL << 5,
97 PERF_IP_FLAG_INTERRUPT = 1ULL << 6,
98 PERF_IP_FLAG_TX_ABORT = 1ULL << 7,
99 PERF_IP_FLAG_TRACE_BEGIN = 1ULL << 8,
100 PERF_IP_FLAG_TRACE_END = 1ULL << 9,
101 PERF_IP_FLAG_IN_TX = 1ULL << 10,
102 PERF_IP_FLAG_VMENTRY = 1ULL << 11,
103 PERF_IP_FLAG_VMEXIT = 1ULL << 12,
104 };
105
106 #define PERF_IP_FLAG_CHARS "bcrosyiABExgh"
107
108 #define PERF_BRANCH_MASK (\
109 PERF_IP_FLAG_BRANCH |\
110 PERF_IP_FLAG_CALL |\
111 PERF_IP_FLAG_RETURN |\
112 PERF_IP_FLAG_CONDITIONAL |\
113 PERF_IP_FLAG_SYSCALLRET |\
114 PERF_IP_FLAG_ASYNC |\
115 PERF_IP_FLAG_INTERRUPT |\
116 PERF_IP_FLAG_TX_ABORT |\
117 PERF_IP_FLAG_TRACE_BEGIN |\
118 PERF_IP_FLAG_TRACE_END |\
119 PERF_IP_FLAG_VMENTRY |\
120 PERF_IP_FLAG_VMEXIT)
121
122 #define MAX_INSN 16
123
124 struct aux_sample {
125 u64 size;
126 void *data;
127 };
128
129 struct perf_sample {
130 u64 ip;
131 u32 pid, tid;
132 u64 time;
133 u64 addr;
134 u64 id;
135 u64 stream_id;
136 u64 period;
137 u64 weight;
138 u64 transaction;
139 u64 insn_cnt;
140 u64 cyc_cnt;
141 u32 cpu;
142 u32 raw_size;
143 u64 data_src;
144 u64 phys_addr;
145 u64 data_page_size;
146 u64 code_page_size;
147 u64 cgroup;
148 u32 flags;
149 u16 insn_len;
150 u8 cpumode;
151 u16 misc;
152 u16 ins_lat;
153 u16 p_stage_cyc;
154 bool no_hw_idx; /* No hw_idx collected in branch_stack */
155 char insn[MAX_INSN];
156 void *raw_data;
157 struct ip_callchain *callchain;
158 struct branch_stack *branch_stack;
159 struct regs_dump user_regs;
160 struct regs_dump intr_regs;
161 struct stack_dump user_stack;
162 struct sample_read read;
163 struct aux_sample aux_sample;
164 };
165
166 #define PERF_MEM_DATA_SRC_NONE \
167 (PERF_MEM_S(OP, NA) |\
168 PERF_MEM_S(LVL, NA) |\
169 PERF_MEM_S(SNOOP, NA) |\
170 PERF_MEM_S(LOCK, NA) |\
171 PERF_MEM_S(TLB, NA))
172
173 /* Attribute type for custom synthesized events */
174 #define PERF_TYPE_SYNTH (INT_MAX + 1U)
175
176 /* Attribute config for custom synthesized events */
177 enum perf_synth_id {
178 PERF_SYNTH_INTEL_PTWRITE,
179 PERF_SYNTH_INTEL_MWAIT,
180 PERF_SYNTH_INTEL_PWRE,
181 PERF_SYNTH_INTEL_EXSTOP,
182 PERF_SYNTH_INTEL_PWRX,
183 PERF_SYNTH_INTEL_CBR,
184 PERF_SYNTH_INTEL_PSB,
185 };
186
187 /*
188 * Raw data formats for synthesized events. Note that 4 bytes of padding are
189 * present to match the 'size' member of PERF_SAMPLE_RAW data which is always
190 * 8-byte aligned. That means we must dereference raw_data with an offset of 4.
191 * Refer perf_sample__synth_ptr() and perf_synth__raw_data(). It also means the
192 * structure sizes are 4 bytes bigger than the raw_size, refer
193 * perf_synth__raw_size().
194 */
195
196 struct perf_synth_intel_ptwrite {
197 u32 padding;
198 union {
199 struct {
200 u32 ip : 1,
201 reserved : 31;
202 };
203 u32 flags;
204 };
205 u64 payload;
206 };
207
208 struct perf_synth_intel_mwait {
209 u32 padding;
210 u32 reserved;
211 union {
212 struct {
213 u64 hints : 8,
214 reserved1 : 24,
215 extensions : 2,
216 reserved2 : 30;
217 };
218 u64 payload;
219 };
220 };
221
222 struct perf_synth_intel_pwre {
223 u32 padding;
224 u32 reserved;
225 union {
226 struct {
227 u64 reserved1 : 7,
228 hw : 1,
229 subcstate : 4,
230 cstate : 4,
231 reserved2 : 48;
232 };
233 u64 payload;
234 };
235 };
236
237 struct perf_synth_intel_exstop {
238 u32 padding;
239 union {
240 struct {
241 u32 ip : 1,
242 reserved : 31;
243 };
244 u32 flags;
245 };
246 };
247
248 struct perf_synth_intel_pwrx {
249 u32 padding;
250 u32 reserved;
251 union {
252 struct {
253 u64 deepest_cstate : 4,
254 last_cstate : 4,
255 wake_reason : 4,
256 reserved1 : 52;
257 };
258 u64 payload;
259 };
260 };
261
262 struct perf_synth_intel_cbr {
263 u32 padding;
264 union {
265 struct {
266 u32 cbr : 8,
267 reserved1 : 8,
268 max_nonturbo : 8,
269 reserved2 : 8;
270 };
271 u32 flags;
272 };
273 u32 freq;
274 u32 reserved3;
275 };
276
277 struct perf_synth_intel_psb {
278 u32 padding;
279 u32 reserved;
280 u64 offset;
281 };
282
283 /*
284 * raw_data is always 4 bytes from an 8-byte boundary, so subtract 4 to get
285 * 8-byte alignment.
286 */
perf_sample__synth_ptr(struct perf_sample * sample)287 static inline void *perf_sample__synth_ptr(struct perf_sample *sample)
288 {
289 return sample->raw_data - 4;
290 }
291
perf_synth__raw_data(void * p)292 static inline void *perf_synth__raw_data(void *p)
293 {
294 return p + 4;
295 }
296
297 #define perf_synth__raw_size(d) (sizeof(d) - 4)
298
299 #define perf_sample__bad_synth_size(s, d) ((s)->raw_size < sizeof(d) - 4)
300
301 enum {
302 PERF_STAT_ROUND_TYPE__INTERVAL = 0,
303 PERF_STAT_ROUND_TYPE__FINAL = 1,
304 };
305
306 void perf_event__print_totals(void);
307
308 struct perf_cpu_map;
309 struct perf_record_stat_config;
310 struct perf_stat_config;
311 struct perf_tool;
312
313 void perf_event__read_stat_config(struct perf_stat_config *config,
314 struct perf_record_stat_config *event);
315
316 int perf_event__process_comm(struct perf_tool *tool,
317 union perf_event *event,
318 struct perf_sample *sample,
319 struct machine *machine);
320 int perf_event__process_lost(struct perf_tool *tool,
321 union perf_event *event,
322 struct perf_sample *sample,
323 struct machine *machine);
324 int perf_event__process_lost_samples(struct perf_tool *tool,
325 union perf_event *event,
326 struct perf_sample *sample,
327 struct machine *machine);
328 int perf_event__process_aux(struct perf_tool *tool,
329 union perf_event *event,
330 struct perf_sample *sample,
331 struct machine *machine);
332 int perf_event__process_itrace_start(struct perf_tool *tool,
333 union perf_event *event,
334 struct perf_sample *sample,
335 struct machine *machine);
336 int perf_event__process_aux_output_hw_id(struct perf_tool *tool,
337 union perf_event *event,
338 struct perf_sample *sample,
339 struct machine *machine);
340 int perf_event__process_switch(struct perf_tool *tool,
341 union perf_event *event,
342 struct perf_sample *sample,
343 struct machine *machine);
344 int perf_event__process_namespaces(struct perf_tool *tool,
345 union perf_event *event,
346 struct perf_sample *sample,
347 struct machine *machine);
348 int perf_event__process_cgroup(struct perf_tool *tool,
349 union perf_event *event,
350 struct perf_sample *sample,
351 struct machine *machine);
352 int perf_event__process_mmap(struct perf_tool *tool,
353 union perf_event *event,
354 struct perf_sample *sample,
355 struct machine *machine);
356 int perf_event__process_mmap2(struct perf_tool *tool,
357 union perf_event *event,
358 struct perf_sample *sample,
359 struct machine *machine);
360 int perf_event__process_fork(struct perf_tool *tool,
361 union perf_event *event,
362 struct perf_sample *sample,
363 struct machine *machine);
364 int perf_event__process_exit(struct perf_tool *tool,
365 union perf_event *event,
366 struct perf_sample *sample,
367 struct machine *machine);
368 int perf_event__process_ksymbol(struct perf_tool *tool,
369 union perf_event *event,
370 struct perf_sample *sample,
371 struct machine *machine);
372 int perf_event__process_bpf(struct perf_tool *tool,
373 union perf_event *event,
374 struct perf_sample *sample,
375 struct machine *machine);
376 int perf_event__process_text_poke(struct perf_tool *tool,
377 union perf_event *event,
378 struct perf_sample *sample,
379 struct machine *machine);
380 int perf_event__process(struct perf_tool *tool,
381 union perf_event *event,
382 struct perf_sample *sample,
383 struct machine *machine);
384
385 struct addr_location;
386
387 int machine__resolve(struct machine *machine, struct addr_location *al,
388 struct perf_sample *sample);
389
390 void addr_location__put(struct addr_location *al);
391
392 struct thread;
393
394 bool is_bts_event(struct perf_event_attr *attr);
395 bool sample_addr_correlates_sym(struct perf_event_attr *attr);
396 void thread__resolve(struct thread *thread, struct addr_location *al,
397 struct perf_sample *sample);
398
399 const char *perf_event__name(unsigned int id);
400
401 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp);
402 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp);
403 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp);
404 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp);
405 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp);
406 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp);
407 size_t perf_event__fprintf_aux_output_hw_id(union perf_event *event, FILE *fp);
408 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp);
409 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp);
410 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp);
411 size_t perf_event__fprintf_namespaces(union perf_event *event, FILE *fp);
412 size_t perf_event__fprintf_cgroup(union perf_event *event, FILE *fp);
413 size_t perf_event__fprintf_ksymbol(union perf_event *event, FILE *fp);
414 size_t perf_event__fprintf_bpf(union perf_event *event, FILE *fp);
415 size_t perf_event__fprintf_text_poke(union perf_event *event, struct machine *machine,FILE *fp);
416 size_t perf_event__fprintf(union perf_event *event, struct machine *machine, FILE *fp);
417
418 int kallsyms__get_function_start(const char *kallsyms_filename,
419 const char *symbol_name, u64 *addr);
420
421 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max);
422 void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map,
423 u16 type, int max);
424
425 void event_attr_init(struct perf_event_attr *attr);
426
427 int perf_event_paranoid(void);
428 bool perf_event_paranoid_check(int max_level);
429
430 extern int sysctl_perf_event_max_stack;
431 extern int sysctl_perf_event_max_contexts_per_stack;
432 extern unsigned int proc_map_timeout;
433
434 #define PAGE_SIZE_NAME_LEN 32
435 char *get_page_size_name(u64 size, char *str);
436
437 void arch_perf_parse_sample_weight(struct perf_sample *data, const __u64 *array, u64 type);
438 void arch_perf_synthesize_sample_weight(const struct perf_sample *data, __u64 *array, u64 type);
439 const char *arch_perf_header_entry(const char *se_header);
440 int arch_support_sort_key(const char *sort_key);
441
442 #endif /* __PERF_RECORD_H */
443