1 /* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
2 /*
3  * Performance events:
4  *
5  *    Copyright (C) 2008-2009, Thomas Gleixner <[email protected]>
6  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
7  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
8  *
9  * Data type definitions, declarations, prototypes.
10  *
11  *    Started by: Thomas Gleixner and Ingo Molnar
12  *
13  * For licencing details see kernel-base/COPYING
14  */
15 #ifndef _UAPI_LINUX_PERF_EVENT_H
16 #define _UAPI_LINUX_PERF_EVENT_H
17 
18 #include <linux/types.h>
19 #include <linux/ioctl.h>
20 #include <asm/byteorder.h>
21 
22 /*
23  * User-space ABI bits:
24  */
25 
26 /*
27  * attr.type
28  */
29 enum perf_type_id {
30 	PERF_TYPE_HARDWARE			= 0,
31 	PERF_TYPE_SOFTWARE			= 1,
32 	PERF_TYPE_TRACEPOINT			= 2,
33 	PERF_TYPE_HW_CACHE			= 3,
34 	PERF_TYPE_RAW				= 4,
35 	PERF_TYPE_BREAKPOINT			= 5,
36 
37 	PERF_TYPE_MAX,				/* non-ABI */
38 };
39 
40 /*
41  * Generalized performance event event_id types, used by the
42  * attr.event_id parameter of the sys_perf_event_open()
43  * syscall:
44  */
45 enum perf_hw_id {
46 	/*
47 	 * Common hardware events, generalized by the kernel:
48 	 */
49 	PERF_COUNT_HW_CPU_CYCLES		= 0,
50 	PERF_COUNT_HW_INSTRUCTIONS		= 1,
51 	PERF_COUNT_HW_CACHE_REFERENCES		= 2,
52 	PERF_COUNT_HW_CACHE_MISSES		= 3,
53 	PERF_COUNT_HW_BRANCH_INSTRUCTIONS	= 4,
54 	PERF_COUNT_HW_BRANCH_MISSES		= 5,
55 	PERF_COUNT_HW_BUS_CYCLES		= 6,
56 	PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	= 7,
57 	PERF_COUNT_HW_STALLED_CYCLES_BACKEND	= 8,
58 	PERF_COUNT_HW_REF_CPU_CYCLES		= 9,
59 
60 	PERF_COUNT_HW_MAX,			/* non-ABI */
61 };
62 
63 /*
64  * Generalized hardware cache events:
65  *
66  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
67  *       { read, write, prefetch } x
68  *       { accesses, misses }
69  */
70 enum perf_hw_cache_id {
71 	PERF_COUNT_HW_CACHE_L1D			= 0,
72 	PERF_COUNT_HW_CACHE_L1I			= 1,
73 	PERF_COUNT_HW_CACHE_LL			= 2,
74 	PERF_COUNT_HW_CACHE_DTLB		= 3,
75 	PERF_COUNT_HW_CACHE_ITLB		= 4,
76 	PERF_COUNT_HW_CACHE_BPU			= 5,
77 	PERF_COUNT_HW_CACHE_NODE		= 6,
78 
79 	PERF_COUNT_HW_CACHE_MAX,		/* non-ABI */
80 };
81 
82 enum perf_hw_cache_op_id {
83 	PERF_COUNT_HW_CACHE_OP_READ		= 0,
84 	PERF_COUNT_HW_CACHE_OP_WRITE		= 1,
85 	PERF_COUNT_HW_CACHE_OP_PREFETCH		= 2,
86 
87 	PERF_COUNT_HW_CACHE_OP_MAX,		/* non-ABI */
88 };
89 
90 enum perf_hw_cache_op_result_id {
91 	PERF_COUNT_HW_CACHE_RESULT_ACCESS	= 0,
92 	PERF_COUNT_HW_CACHE_RESULT_MISS		= 1,
93 
94 	PERF_COUNT_HW_CACHE_RESULT_MAX,		/* non-ABI */
95 };
96 
97 /*
98  * Special "software" events provided by the kernel, even if the hardware
99  * does not support performance events. These events measure various
100  * physical and sw events of the kernel (and allow the profiling of them as
101  * well):
102  */
103 enum perf_sw_ids {
104 	PERF_COUNT_SW_CPU_CLOCK			= 0,
105 	PERF_COUNT_SW_TASK_CLOCK		= 1,
106 	PERF_COUNT_SW_PAGE_FAULTS		= 2,
107 	PERF_COUNT_SW_CONTEXT_SWITCHES		= 3,
108 	PERF_COUNT_SW_CPU_MIGRATIONS		= 4,
109 	PERF_COUNT_SW_PAGE_FAULTS_MIN		= 5,
110 	PERF_COUNT_SW_PAGE_FAULTS_MAJ		= 6,
111 	PERF_COUNT_SW_ALIGNMENT_FAULTS		= 7,
112 	PERF_COUNT_SW_EMULATION_FAULTS		= 8,
113 	PERF_COUNT_SW_DUMMY			= 9,
114 	PERF_COUNT_SW_BPF_OUTPUT		= 10,
115 
116 	PERF_COUNT_SW_MAX,			/* non-ABI */
117 };
118 
119 /*
120  * Bits that can be set in attr.sample_type to request information
121  * in the overflow packets.
122  */
123 enum perf_event_sample_format {
124 	PERF_SAMPLE_IP				= 1U << 0,
125 	PERF_SAMPLE_TID				= 1U << 1,
126 	PERF_SAMPLE_TIME			= 1U << 2,
127 	PERF_SAMPLE_ADDR			= 1U << 3,
128 	PERF_SAMPLE_READ			= 1U << 4,
129 	PERF_SAMPLE_CALLCHAIN			= 1U << 5,
130 	PERF_SAMPLE_ID				= 1U << 6,
131 	PERF_SAMPLE_CPU				= 1U << 7,
132 	PERF_SAMPLE_PERIOD			= 1U << 8,
133 	PERF_SAMPLE_STREAM_ID			= 1U << 9,
134 	PERF_SAMPLE_RAW				= 1U << 10,
135 	PERF_SAMPLE_BRANCH_STACK		= 1U << 11,
136 	PERF_SAMPLE_REGS_USER			= 1U << 12,
137 	PERF_SAMPLE_STACK_USER			= 1U << 13,
138 	PERF_SAMPLE_WEIGHT			= 1U << 14,
139 	PERF_SAMPLE_DATA_SRC			= 1U << 15,
140 	PERF_SAMPLE_IDENTIFIER			= 1U << 16,
141 	PERF_SAMPLE_TRANSACTION			= 1U << 17,
142 	PERF_SAMPLE_REGS_INTR			= 1U << 18,
143 	PERF_SAMPLE_PHYS_ADDR			= 1U << 19,
144 	PERF_SAMPLE_AUX				= 1U << 20,
145 	PERF_SAMPLE_CGROUP			= 1U << 21,
146 	PERF_SAMPLE_DATA_PAGE_SIZE		= 1U << 22,
147 	PERF_SAMPLE_CODE_PAGE_SIZE		= 1U << 23,
148 
149 	PERF_SAMPLE_MAX = 1U << 24,		/* non-ABI */
150 
151 	__PERF_SAMPLE_CALLCHAIN_EARLY		= 1ULL << 63, /* non-ABI; internal use */
152 };
153 
154 /*
155  * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
156  *
157  * If the user does not pass priv level information via branch_sample_type,
158  * the kernel uses the event's priv level. Branch and event priv levels do
159  * not have to match. Branch priv level is checked for permissions.
160  *
161  * The branch types can be combined, however BRANCH_ANY covers all types
162  * of branches and therefore it supersedes all the other types.
163  */
164 enum perf_branch_sample_type_shift {
165 	PERF_SAMPLE_BRANCH_USER_SHIFT		= 0, /* user branches */
166 	PERF_SAMPLE_BRANCH_KERNEL_SHIFT		= 1, /* kernel branches */
167 	PERF_SAMPLE_BRANCH_HV_SHIFT		= 2, /* hypervisor branches */
168 
169 	PERF_SAMPLE_BRANCH_ANY_SHIFT		= 3, /* any branch types */
170 	PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT	= 4, /* any call branch */
171 	PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT	= 5, /* any return branch */
172 	PERF_SAMPLE_BRANCH_IND_CALL_SHIFT	= 6, /* indirect calls */
173 	PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT	= 7, /* transaction aborts */
174 	PERF_SAMPLE_BRANCH_IN_TX_SHIFT		= 8, /* in transaction */
175 	PERF_SAMPLE_BRANCH_NO_TX_SHIFT		= 9, /* not in transaction */
176 	PERF_SAMPLE_BRANCH_COND_SHIFT		= 10, /* conditional branches */
177 
178 	PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT	= 11, /* call/ret stack */
179 	PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT	= 12, /* indirect jumps */
180 	PERF_SAMPLE_BRANCH_CALL_SHIFT		= 13, /* direct call */
181 
182 	PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT	= 14, /* no flags */
183 	PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT	= 15, /* no cycles */
184 
185 	PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT	= 16, /* save branch type */
186 
187 	PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT	= 17, /* save low level index of raw branch records */
188 
189 	PERF_SAMPLE_BRANCH_MAX_SHIFT		/* non-ABI */
190 };
191 
192 enum perf_branch_sample_type {
193 	PERF_SAMPLE_BRANCH_USER		= 1U << PERF_SAMPLE_BRANCH_USER_SHIFT,
194 	PERF_SAMPLE_BRANCH_KERNEL	= 1U << PERF_SAMPLE_BRANCH_KERNEL_SHIFT,
195 	PERF_SAMPLE_BRANCH_HV		= 1U << PERF_SAMPLE_BRANCH_HV_SHIFT,
196 
197 	PERF_SAMPLE_BRANCH_ANY		= 1U << PERF_SAMPLE_BRANCH_ANY_SHIFT,
198 	PERF_SAMPLE_BRANCH_ANY_CALL	= 1U << PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT,
199 	PERF_SAMPLE_BRANCH_ANY_RETURN	= 1U << PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT,
200 	PERF_SAMPLE_BRANCH_IND_CALL	= 1U << PERF_SAMPLE_BRANCH_IND_CALL_SHIFT,
201 	PERF_SAMPLE_BRANCH_ABORT_TX	= 1U << PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT,
202 	PERF_SAMPLE_BRANCH_IN_TX	= 1U << PERF_SAMPLE_BRANCH_IN_TX_SHIFT,
203 	PERF_SAMPLE_BRANCH_NO_TX	= 1U << PERF_SAMPLE_BRANCH_NO_TX_SHIFT,
204 	PERF_SAMPLE_BRANCH_COND		= 1U << PERF_SAMPLE_BRANCH_COND_SHIFT,
205 
206 	PERF_SAMPLE_BRANCH_CALL_STACK	= 1U << PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT,
207 	PERF_SAMPLE_BRANCH_IND_JUMP	= 1U << PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT,
208 	PERF_SAMPLE_BRANCH_CALL		= 1U << PERF_SAMPLE_BRANCH_CALL_SHIFT,
209 
210 	PERF_SAMPLE_BRANCH_NO_FLAGS	= 1U << PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT,
211 	PERF_SAMPLE_BRANCH_NO_CYCLES	= 1U << PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT,
212 
213 	PERF_SAMPLE_BRANCH_TYPE_SAVE	=
214 		1U << PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT,
215 
216 	PERF_SAMPLE_BRANCH_HW_INDEX	= 1U << PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT,
217 
218 	PERF_SAMPLE_BRANCH_MAX		= 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT,
219 };
220 
221 /*
222  * Common flow change classification
223  */
224 enum {
225 	PERF_BR_UNKNOWN		= 0,	/* unknown */
226 	PERF_BR_COND		= 1,	/* conditional */
227 	PERF_BR_UNCOND		= 2,	/* unconditional  */
228 	PERF_BR_IND		= 3,	/* indirect */
229 	PERF_BR_CALL		= 4,	/* function call */
230 	PERF_BR_IND_CALL	= 5,	/* indirect function call */
231 	PERF_BR_RET		= 6,	/* function return */
232 	PERF_BR_SYSCALL		= 7,	/* syscall */
233 	PERF_BR_SYSRET		= 8,	/* syscall return */
234 	PERF_BR_COND_CALL	= 9,	/* conditional function call */
235 	PERF_BR_COND_RET	= 10,	/* conditional function return */
236 	PERF_BR_MAX,
237 };
238 
239 #define PERF_SAMPLE_BRANCH_PLM_ALL \
240 	(PERF_SAMPLE_BRANCH_USER|\
241 	 PERF_SAMPLE_BRANCH_KERNEL|\
242 	 PERF_SAMPLE_BRANCH_HV)
243 
244 /*
245  * Values to determine ABI of the registers dump.
246  */
247 enum perf_sample_regs_abi {
248 	PERF_SAMPLE_REGS_ABI_NONE	= 0,
249 	PERF_SAMPLE_REGS_ABI_32		= 1,
250 	PERF_SAMPLE_REGS_ABI_64		= 2,
251 };
252 
253 /*
254  * Values for the memory transaction event qualifier, mostly for
255  * abort events. Multiple bits can be set.
256  */
257 enum {
258 	PERF_TXN_ELISION        = (1 << 0), /* From elision */
259 	PERF_TXN_TRANSACTION    = (1 << 1), /* From transaction */
260 	PERF_TXN_SYNC           = (1 << 2), /* Instruction is related */
261 	PERF_TXN_ASYNC          = (1 << 3), /* Instruction not related */
262 	PERF_TXN_RETRY          = (1 << 4), /* Retry possible */
263 	PERF_TXN_CONFLICT       = (1 << 5), /* Conflict abort */
264 	PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */
265 	PERF_TXN_CAPACITY_READ  = (1 << 7), /* Capacity read abort */
266 
267 	PERF_TXN_MAX	        = (1 << 8), /* non-ABI */
268 
269 	/* bits 32..63 are reserved for the abort code */
270 
271 	PERF_TXN_ABORT_MASK  = (0xffffffffULL << 32),
272 	PERF_TXN_ABORT_SHIFT = 32,
273 };
274 
275 /*
276  * The format of the data returned by read() on a perf event fd,
277  * as specified by attr.read_format:
278  *
279  * struct read_format {
280  *	{ u64		value;
281  *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
282  *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
283  *	  { u64		id;           } && PERF_FORMAT_ID
284  *	} && !PERF_FORMAT_GROUP
285  *
286  *	{ u64		nr;
287  *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
288  *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
289  *	  { u64		value;
290  *	    { u64	id;           } && PERF_FORMAT_ID
291  *	  }		cntr[nr];
292  *	} && PERF_FORMAT_GROUP
293  * };
294  */
295 enum perf_event_read_format {
296 	PERF_FORMAT_TOTAL_TIME_ENABLED		= 1U << 0,
297 	PERF_FORMAT_TOTAL_TIME_RUNNING		= 1U << 1,
298 	PERF_FORMAT_ID				= 1U << 2,
299 	PERF_FORMAT_GROUP			= 1U << 3,
300 
301 	PERF_FORMAT_MAX = 1U << 4,		/* non-ABI */
302 };
303 
304 #define PERF_ATTR_SIZE_VER0	64	/* sizeof first published struct */
305 #define PERF_ATTR_SIZE_VER1	72	/* add: config2 */
306 #define PERF_ATTR_SIZE_VER2	80	/* add: branch_sample_type */
307 #define PERF_ATTR_SIZE_VER3	96	/* add: sample_regs_user */
308 					/* add: sample_stack_user */
309 #define PERF_ATTR_SIZE_VER4	104	/* add: sample_regs_intr */
310 #define PERF_ATTR_SIZE_VER5	112	/* add: aux_watermark */
311 #define PERF_ATTR_SIZE_VER6	120	/* add: aux_sample_size */
312 
313 /*
314  * Hardware event_id to monitor via a performance monitoring event:
315  *
316  * @sample_max_stack: Max number of frame pointers in a callchain,
317  *		      should be < /proc/sys/kernel/perf_event_max_stack
318  */
319 struct perf_event_attr {
320 
321 	/*
322 	 * Major type: hardware/software/tracepoint/etc.
323 	 */
324 	__u32			type;
325 
326 	/*
327 	 * Size of the attr structure, for fwd/bwd compat.
328 	 */
329 	__u32			size;
330 
331 	/*
332 	 * Type specific configuration information.
333 	 */
334 	__u64			config;
335 
336 	union {
337 		__u64		sample_period;
338 		__u64		sample_freq;
339 	};
340 
341 	__u64			sample_type;
342 	__u64			read_format;
343 
344 	__u64			disabled       :  1, /* off by default        */
345 				inherit	       :  1, /* children inherit it   */
346 				pinned	       :  1, /* must always be on PMU */
347 				exclusive      :  1, /* only group on PMU     */
348 				exclude_user   :  1, /* don't count user      */
349 				exclude_kernel :  1, /* ditto kernel          */
350 				exclude_hv     :  1, /* ditto hypervisor      */
351 				exclude_idle   :  1, /* don't count when idle */
352 				mmap           :  1, /* include mmap data     */
353 				comm	       :  1, /* include comm data     */
354 				freq           :  1, /* use freq, not period  */
355 				inherit_stat   :  1, /* per task counts       */
356 				enable_on_exec :  1, /* next exec enables     */
357 				task           :  1, /* trace fork/exit       */
358 				watermark      :  1, /* wakeup_watermark      */
359 				/*
360 				 * precise_ip:
361 				 *
362 				 *  0 - SAMPLE_IP can have arbitrary skid
363 				 *  1 - SAMPLE_IP must have constant skid
364 				 *  2 - SAMPLE_IP requested to have 0 skid
365 				 *  3 - SAMPLE_IP must have 0 skid
366 				 *
367 				 *  See also PERF_RECORD_MISC_EXACT_IP
368 				 */
369 				precise_ip     :  2, /* skid constraint       */
370 				mmap_data      :  1, /* non-exec mmap data    */
371 				sample_id_all  :  1, /* sample_type all events */
372 
373 				exclude_host   :  1, /* don't count in host   */
374 				exclude_guest  :  1, /* don't count in guest  */
375 
376 				exclude_callchain_kernel : 1, /* exclude kernel callchains */
377 				exclude_callchain_user   : 1, /* exclude user callchains */
378 				mmap2          :  1, /* include mmap with inode data     */
379 				comm_exec      :  1, /* flag comm events that are due to an exec */
380 				use_clockid    :  1, /* use @clockid for time fields */
381 				context_switch :  1, /* context switch data */
382 				write_backward :  1, /* Write ring buffer from end to beginning */
383 				namespaces     :  1, /* include namespaces data */
384 				ksymbol        :  1, /* include ksymbol events */
385 				bpf_event      :  1, /* include bpf events */
386 				aux_output     :  1, /* generate AUX records instead of events */
387 				cgroup         :  1, /* include cgroup events */
388 				text_poke      :  1, /* include text poke events */
389 				build_id       :  1, /* use build id in mmap2 events */
390 				__reserved_1   : 29;
391 
392 	union {
393 		__u32		wakeup_events;	  /* wakeup every n events */
394 		__u32		wakeup_watermark; /* bytes before wakeup   */
395 	};
396 
397 	__u32			bp_type;
398 	union {
399 		__u64		bp_addr;
400 		__u64		kprobe_func; /* for perf_kprobe */
401 		__u64		uprobe_path; /* for perf_uprobe */
402 		__u64		config1; /* extension of config */
403 	};
404 	union {
405 		__u64		bp_len;
406 		__u64		kprobe_addr; /* when kprobe_func == NULL */
407 		__u64		probe_offset; /* for perf_[k,u]probe */
408 		__u64		config2; /* extension of config1 */
409 	};
410 	__u64	branch_sample_type; /* enum perf_branch_sample_type */
411 
412 	/*
413 	 * Defines set of user regs to dump on samples.
414 	 * See asm/perf_regs.h for details.
415 	 */
416 	__u64	sample_regs_user;
417 
418 	/*
419 	 * Defines size of the user stack to dump on samples.
420 	 */
421 	__u32	sample_stack_user;
422 
423 	__s32	clockid;
424 	/*
425 	 * Defines set of regs to dump for each sample
426 	 * state captured on:
427 	 *  - precise = 0: PMU interrupt
428 	 *  - precise > 0: sampled instruction
429 	 *
430 	 * See asm/perf_regs.h for details.
431 	 */
432 	__u64	sample_regs_intr;
433 
434 	/*
435 	 * Wakeup watermark for AUX area
436 	 */
437 	__u32	aux_watermark;
438 	__u16	sample_max_stack;
439 	__u16	__reserved_2;
440 	__u32	aux_sample_size;
441 	__u32	__reserved_3;
442 };
443 
444 /*
445  * Structure used by below PERF_EVENT_IOC_QUERY_BPF command
446  * to query bpf programs attached to the same perf tracepoint
447  * as the given perf event.
448  */
449 struct perf_event_query_bpf {
450 	/*
451 	 * The below ids array length
452 	 */
453 	__u32	ids_len;
454 	/*
455 	 * Set by the kernel to indicate the number of
456 	 * available programs
457 	 */
458 	__u32	prog_cnt;
459 	/*
460 	 * User provided buffer to store program ids
461 	 */
462 	__u32	ids[0];
463 };
464 
465 /*
466  * Ioctls that can be done on a perf event fd:
467  */
468 #define PERF_EVENT_IOC_ENABLE			_IO ('$', 0)
469 #define PERF_EVENT_IOC_DISABLE			_IO ('$', 1)
470 #define PERF_EVENT_IOC_REFRESH			_IO ('$', 2)
471 #define PERF_EVENT_IOC_RESET			_IO ('$', 3)
472 #define PERF_EVENT_IOC_PERIOD			_IOW('$', 4, __u64)
473 #define PERF_EVENT_IOC_SET_OUTPUT		_IO ('$', 5)
474 #define PERF_EVENT_IOC_SET_FILTER		_IOW('$', 6, char *)
475 #define PERF_EVENT_IOC_ID			_IOR('$', 7, __u64 *)
476 #define PERF_EVENT_IOC_SET_BPF			_IOW('$', 8, __u32)
477 #define PERF_EVENT_IOC_PAUSE_OUTPUT		_IOW('$', 9, __u32)
478 #define PERF_EVENT_IOC_QUERY_BPF		_IOWR('$', 10, struct perf_event_query_bpf *)
479 #define PERF_EVENT_IOC_MODIFY_ATTRIBUTES	_IOW('$', 11, struct perf_event_attr *)
480 
481 enum perf_event_ioc_flags {
482 	PERF_IOC_FLAG_GROUP		= 1U << 0,
483 };
484 
485 /*
486  * Structure of the page that can be mapped via mmap
487  */
488 struct perf_event_mmap_page {
489 	__u32	version;		/* version number of this structure */
490 	__u32	compat_version;		/* lowest version this is compat with */
491 
492 	/*
493 	 * Bits needed to read the hw events in user-space.
494 	 *
495 	 *   u32 seq, time_mult, time_shift, index, width;
496 	 *   u64 count, enabled, running;
497 	 *   u64 cyc, time_offset;
498 	 *   s64 pmc = 0;
499 	 *
500 	 *   do {
501 	 *     seq = pc->lock;
502 	 *     barrier()
503 	 *
504 	 *     enabled = pc->time_enabled;
505 	 *     running = pc->time_running;
506 	 *
507 	 *     if (pc->cap_usr_time && enabled != running) {
508 	 *       cyc = rdtsc();
509 	 *       time_offset = pc->time_offset;
510 	 *       time_mult   = pc->time_mult;
511 	 *       time_shift  = pc->time_shift;
512 	 *     }
513 	 *
514 	 *     index = pc->index;
515 	 *     count = pc->offset;
516 	 *     if (pc->cap_user_rdpmc && index) {
517 	 *       width = pc->pmc_width;
518 	 *       pmc = rdpmc(index - 1);
519 	 *     }
520 	 *
521 	 *     barrier();
522 	 *   } while (pc->lock != seq);
523 	 *
524 	 * NOTE: for obvious reason this only works on self-monitoring
525 	 *       processes.
526 	 */
527 	__u32	lock;			/* seqlock for synchronization */
528 	__u32	index;			/* hardware event identifier */
529 	__s64	offset;			/* add to hardware event value */
530 	__u64	time_enabled;		/* time event active */
531 	__u64	time_running;		/* time event on cpu */
532 	union {
533 		__u64	capabilities;
534 		struct {
535 			__u64	cap_bit0		: 1, /* Always 0, deprecated, see commit 860f085b74e9 */
536 				cap_bit0_is_deprecated	: 1, /* Always 1, signals that bit 0 is zero */
537 
538 				cap_user_rdpmc		: 1, /* The RDPMC instruction can be used to read counts */
539 				cap_user_time		: 1, /* The time_{shift,mult,offset} fields are used */
540 				cap_user_time_zero	: 1, /* The time_zero field is used */
541 				cap_user_time_short	: 1, /* the time_{cycle,mask} fields are used */
542 				cap_____res		: 58;
543 		};
544 	};
545 
546 	/*
547 	 * If cap_user_rdpmc this field provides the bit-width of the value
548 	 * read using the rdpmc() or equivalent instruction. This can be used
549 	 * to sign extend the result like:
550 	 *
551 	 *   pmc <<= 64 - width;
552 	 *   pmc >>= 64 - width; // signed shift right
553 	 *   count += pmc;
554 	 */
555 	__u16	pmc_width;
556 
557 	/*
558 	 * If cap_usr_time the below fields can be used to compute the time
559 	 * delta since time_enabled (in ns) using rdtsc or similar.
560 	 *
561 	 *   u64 quot, rem;
562 	 *   u64 delta;
563 	 *
564 	 *   quot = (cyc >> time_shift);
565 	 *   rem = cyc & (((u64)1 << time_shift) - 1);
566 	 *   delta = time_offset + quot * time_mult +
567 	 *              ((rem * time_mult) >> time_shift);
568 	 *
569 	 * Where time_offset,time_mult,time_shift and cyc are read in the
570 	 * seqcount loop described above. This delta can then be added to
571 	 * enabled and possible running (if index), improving the scaling:
572 	 *
573 	 *   enabled += delta;
574 	 *   if (index)
575 	 *     running += delta;
576 	 *
577 	 *   quot = count / running;
578 	 *   rem  = count % running;
579 	 *   count = quot * enabled + (rem * enabled) / running;
580 	 */
581 	__u16	time_shift;
582 	__u32	time_mult;
583 	__u64	time_offset;
584 	/*
585 	 * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated
586 	 * from sample timestamps.
587 	 *
588 	 *   time = timestamp - time_zero;
589 	 *   quot = time / time_mult;
590 	 *   rem  = time % time_mult;
591 	 *   cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
592 	 *
593 	 * And vice versa:
594 	 *
595 	 *   quot = cyc >> time_shift;
596 	 *   rem  = cyc & (((u64)1 << time_shift) - 1);
597 	 *   timestamp = time_zero + quot * time_mult +
598 	 *               ((rem * time_mult) >> time_shift);
599 	 */
600 	__u64	time_zero;
601 
602 	__u32	size;			/* Header size up to __reserved[] fields. */
603 	__u32	__reserved_1;
604 
605 	/*
606 	 * If cap_usr_time_short, the hardware clock is less than 64bit wide
607 	 * and we must compute the 'cyc' value, as used by cap_usr_time, as:
608 	 *
609 	 *   cyc = time_cycles + ((cyc - time_cycles) & time_mask)
610 	 *
611 	 * NOTE: this form is explicitly chosen such that cap_usr_time_short
612 	 *       is a correction on top of cap_usr_time, and code that doesn't
613 	 *       know about cap_usr_time_short still works under the assumption
614 	 *       the counter doesn't wrap.
615 	 */
616 	__u64	time_cycles;
617 	__u64	time_mask;
618 
619 		/*
620 		 * Hole for extension of the self monitor capabilities
621 		 */
622 
623 	__u8	__reserved[116*8];	/* align to 1k. */
624 
625 	/*
626 	 * Control data for the mmap() data buffer.
627 	 *
628 	 * User-space reading the @data_head value should issue an smp_rmb(),
629 	 * after reading this value.
630 	 *
631 	 * When the mapping is PROT_WRITE the @data_tail value should be
632 	 * written by userspace to reflect the last read data, after issueing
633 	 * an smp_mb() to separate the data read from the ->data_tail store.
634 	 * In this case the kernel will not over-write unread data.
635 	 *
636 	 * See perf_output_put_handle() for the data ordering.
637 	 *
638 	 * data_{offset,size} indicate the location and size of the perf record
639 	 * buffer within the mmapped area.
640 	 */
641 	__u64   data_head;		/* head in the data section */
642 	__u64	data_tail;		/* user-space written tail */
643 	__u64	data_offset;		/* where the buffer starts */
644 	__u64	data_size;		/* data buffer size */
645 
646 	/*
647 	 * AUX area is defined by aux_{offset,size} fields that should be set
648 	 * by the userspace, so that
649 	 *
650 	 *   aux_offset >= data_offset + data_size
651 	 *
652 	 * prior to mmap()ing it. Size of the mmap()ed area should be aux_size.
653 	 *
654 	 * Ring buffer pointers aux_{head,tail} have the same semantics as
655 	 * data_{head,tail} and same ordering rules apply.
656 	 */
657 	__u64	aux_head;
658 	__u64	aux_tail;
659 	__u64	aux_offset;
660 	__u64	aux_size;
661 };
662 
663 /*
664  * The current state of perf_event_header::misc bits usage:
665  * ('|' used bit, '-' unused bit)
666  *
667  *  012         CDEF
668  *  |||---------||||
669  *
670  *  Where:
671  *    0-2     CPUMODE_MASK
672  *
673  *    C       PROC_MAP_PARSE_TIMEOUT
674  *    D       MMAP_DATA / COMM_EXEC / FORK_EXEC / SWITCH_OUT
675  *    E       MMAP_BUILD_ID / EXACT_IP / SCHED_OUT_PREEMPT
676  *    F       (reserved)
677  */
678 
679 #define PERF_RECORD_MISC_CPUMODE_MASK		(7 << 0)
680 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN	(0 << 0)
681 #define PERF_RECORD_MISC_KERNEL			(1 << 0)
682 #define PERF_RECORD_MISC_USER			(2 << 0)
683 #define PERF_RECORD_MISC_HYPERVISOR		(3 << 0)
684 #define PERF_RECORD_MISC_GUEST_KERNEL		(4 << 0)
685 #define PERF_RECORD_MISC_GUEST_USER		(5 << 0)
686 
687 /*
688  * Indicates that /proc/PID/maps parsing are truncated by time out.
689  */
690 #define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT	(1 << 12)
691 /*
692  * Following PERF_RECORD_MISC_* are used on different
693  * events, so can reuse the same bit position:
694  *
695  *   PERF_RECORD_MISC_MMAP_DATA  - PERF_RECORD_MMAP* events
696  *   PERF_RECORD_MISC_COMM_EXEC  - PERF_RECORD_COMM event
697  *   PERF_RECORD_MISC_FORK_EXEC  - PERF_RECORD_FORK event (perf internal)
698  *   PERF_RECORD_MISC_SWITCH_OUT - PERF_RECORD_SWITCH* events
699  */
700 #define PERF_RECORD_MISC_MMAP_DATA		(1 << 13)
701 #define PERF_RECORD_MISC_COMM_EXEC		(1 << 13)
702 #define PERF_RECORD_MISC_FORK_EXEC		(1 << 13)
703 #define PERF_RECORD_MISC_SWITCH_OUT		(1 << 13)
704 /*
705  * These PERF_RECORD_MISC_* flags below are safely reused
706  * for the following events:
707  *
708  *   PERF_RECORD_MISC_EXACT_IP           - PERF_RECORD_SAMPLE of precise events
709  *   PERF_RECORD_MISC_SWITCH_OUT_PREEMPT - PERF_RECORD_SWITCH* events
710  *   PERF_RECORD_MISC_MMAP_BUILD_ID      - PERF_RECORD_MMAP2 event
711  *
712  *
713  * PERF_RECORD_MISC_EXACT_IP:
714  *   Indicates that the content of PERF_SAMPLE_IP points to
715  *   the actual instruction that triggered the event. See also
716  *   perf_event_attr::precise_ip.
717  *
718  * PERF_RECORD_MISC_SWITCH_OUT_PREEMPT:
719  *   Indicates that thread was preempted in TASK_RUNNING state.
720  *
721  * PERF_RECORD_MISC_MMAP_BUILD_ID:
722  *   Indicates that mmap2 event carries build id data.
723  */
724 #define PERF_RECORD_MISC_EXACT_IP		(1 << 14)
725 #define PERF_RECORD_MISC_SWITCH_OUT_PREEMPT	(1 << 14)
726 #define PERF_RECORD_MISC_MMAP_BUILD_ID		(1 << 14)
727 /*
728  * Reserve the last bit to indicate some extended misc field
729  */
730 #define PERF_RECORD_MISC_EXT_RESERVED		(1 << 15)
731 
732 struct perf_event_header {
733 	__u32	type;
734 	__u16	misc;
735 	__u16	size;
736 };
737 
738 struct perf_ns_link_info {
739 	__u64	dev;
740 	__u64	ino;
741 };
742 
743 enum {
744 	NET_NS_INDEX		= 0,
745 	UTS_NS_INDEX		= 1,
746 	IPC_NS_INDEX		= 2,
747 	PID_NS_INDEX		= 3,
748 	USER_NS_INDEX		= 4,
749 	MNT_NS_INDEX		= 5,
750 	CGROUP_NS_INDEX		= 6,
751 
752 	NR_NAMESPACES,		/* number of available namespaces */
753 };
754 
755 enum perf_event_type {
756 
757 	/*
758 	 * If perf_event_attr.sample_id_all is set then all event types will
759 	 * have the sample_type selected fields related to where/when
760 	 * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU,
761 	 * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed
762 	 * just after the perf_event_header and the fields already present for
763 	 * the existing fields, i.e. at the end of the payload. That way a newer
764 	 * perf.data file will be supported by older perf tools, with these new
765 	 * optional fields being ignored.
766 	 *
767 	 * struct sample_id {
768 	 * 	{ u32			pid, tid; } && PERF_SAMPLE_TID
769 	 * 	{ u64			time;     } && PERF_SAMPLE_TIME
770 	 * 	{ u64			id;       } && PERF_SAMPLE_ID
771 	 * 	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
772 	 * 	{ u32			cpu, res; } && PERF_SAMPLE_CPU
773 	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
774 	 * } && perf_event_attr::sample_id_all
775 	 *
776 	 * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.  The
777 	 * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed
778 	 * relative to header.size.
779 	 */
780 
781 	/*
782 	 * The MMAP events record the PROT_EXEC mappings so that we can
783 	 * correlate userspace IPs to code. They have the following structure:
784 	 *
785 	 * struct {
786 	 *	struct perf_event_header	header;
787 	 *
788 	 *	u32				pid, tid;
789 	 *	u64				addr;
790 	 *	u64				len;
791 	 *	u64				pgoff;
792 	 *	char				filename[];
793 	 * 	struct sample_id		sample_id;
794 	 * };
795 	 */
796 	PERF_RECORD_MMAP			= 1,
797 
798 	/*
799 	 * struct {
800 	 *	struct perf_event_header	header;
801 	 *	u64				id;
802 	 *	u64				lost;
803 	 * 	struct sample_id		sample_id;
804 	 * };
805 	 */
806 	PERF_RECORD_LOST			= 2,
807 
808 	/*
809 	 * struct {
810 	 *	struct perf_event_header	header;
811 	 *
812 	 *	u32				pid, tid;
813 	 *	char				comm[];
814 	 * 	struct sample_id		sample_id;
815 	 * };
816 	 */
817 	PERF_RECORD_COMM			= 3,
818 
819 	/*
820 	 * struct {
821 	 *	struct perf_event_header	header;
822 	 *	u32				pid, ppid;
823 	 *	u32				tid, ptid;
824 	 *	u64				time;
825 	 * 	struct sample_id		sample_id;
826 	 * };
827 	 */
828 	PERF_RECORD_EXIT			= 4,
829 
830 	/*
831 	 * struct {
832 	 *	struct perf_event_header	header;
833 	 *	u64				time;
834 	 *	u64				id;
835 	 *	u64				stream_id;
836 	 * 	struct sample_id		sample_id;
837 	 * };
838 	 */
839 	PERF_RECORD_THROTTLE			= 5,
840 	PERF_RECORD_UNTHROTTLE			= 6,
841 
842 	/*
843 	 * struct {
844 	 *	struct perf_event_header	header;
845 	 *	u32				pid, ppid;
846 	 *	u32				tid, ptid;
847 	 *	u64				time;
848 	 * 	struct sample_id		sample_id;
849 	 * };
850 	 */
851 	PERF_RECORD_FORK			= 7,
852 
853 	/*
854 	 * struct {
855 	 *	struct perf_event_header	header;
856 	 *	u32				pid, tid;
857 	 *
858 	 *	struct read_format		values;
859 	 * 	struct sample_id		sample_id;
860 	 * };
861 	 */
862 	PERF_RECORD_READ			= 8,
863 
864 	/*
865 	 * struct {
866 	 *	struct perf_event_header	header;
867 	 *
868 	 *	#
869 	 *	# Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.
870 	 *	# The advantage of PERF_SAMPLE_IDENTIFIER is that its position
871 	 *	# is fixed relative to header.
872 	 *	#
873 	 *
874 	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
875 	 *	{ u64			ip;	  } && PERF_SAMPLE_IP
876 	 *	{ u32			pid, tid; } && PERF_SAMPLE_TID
877 	 *	{ u64			time;     } && PERF_SAMPLE_TIME
878 	 *	{ u64			addr;     } && PERF_SAMPLE_ADDR
879 	 *	{ u64			id;	  } && PERF_SAMPLE_ID
880 	 *	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
881 	 *	{ u32			cpu, res; } && PERF_SAMPLE_CPU
882 	 *	{ u64			period;   } && PERF_SAMPLE_PERIOD
883 	 *
884 	 *	{ struct read_format	values;	  } && PERF_SAMPLE_READ
885 	 *
886 	 *	{ u64			nr,
887 	 *	  u64			ips[nr];  } && PERF_SAMPLE_CALLCHAIN
888 	 *
889 	 *	#
890 	 *	# The RAW record below is opaque data wrt the ABI
891 	 *	#
892 	 *	# That is, the ABI doesn't make any promises wrt to
893 	 *	# the stability of its content, it may vary depending
894 	 *	# on event, hardware, kernel version and phase of
895 	 *	# the moon.
896 	 *	#
897 	 *	# In other words, PERF_SAMPLE_RAW contents are not an ABI.
898 	 *	#
899 	 *
900 	 *	{ u32			size;
901 	 *	  char                  data[size];}&& PERF_SAMPLE_RAW
902 	 *
903 	 *	{ u64                   nr;
904 	 *	  { u64	hw_idx; } && PERF_SAMPLE_BRANCH_HW_INDEX
905 	 *        { u64 from, to, flags } lbr[nr];
906 	 *      } && PERF_SAMPLE_BRANCH_STACK
907 	 *
908 	 * 	{ u64			abi; # enum perf_sample_regs_abi
909 	 * 	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
910 	 *
911 	 * 	{ u64			size;
912 	 * 	  char			data[size];
913 	 * 	  u64			dyn_size; } && PERF_SAMPLE_STACK_USER
914 	 *
915 	 *	{ u64			weight;   } && PERF_SAMPLE_WEIGHT
916 	 *	{ u64			data_src; } && PERF_SAMPLE_DATA_SRC
917 	 *	{ u64			transaction; } && PERF_SAMPLE_TRANSACTION
918 	 *	{ u64			abi; # enum perf_sample_regs_abi
919 	 *	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
920 	 *	{ u64			phys_addr;} && PERF_SAMPLE_PHYS_ADDR
921 	 *	{ u64			size;
922 	 *	  char			data[size]; } && PERF_SAMPLE_AUX
923 	 *	{ u64			data_page_size;} && PERF_SAMPLE_DATA_PAGE_SIZE
924 	 *	{ u64			code_page_size;} && PERF_SAMPLE_CODE_PAGE_SIZE
925 	 * };
926 	 */
927 	PERF_RECORD_SAMPLE			= 9,
928 
929 	/*
930 	 * The MMAP2 records are an augmented version of MMAP, they add
931 	 * maj, min, ino numbers to be used to uniquely identify each mapping
932 	 *
933 	 * struct {
934 	 *	struct perf_event_header	header;
935 	 *
936 	 *	u32				pid, tid;
937 	 *	u64				addr;
938 	 *	u64				len;
939 	 *	u64				pgoff;
940 	 *	union {
941 	 *		struct {
942 	 *			u32		maj;
943 	 *			u32		min;
944 	 *			u64		ino;
945 	 *			u64		ino_generation;
946 	 *		};
947 	 *		struct {
948 	 *			u8		build_id_size;
949 	 *			u8		__reserved_1;
950 	 *			u16		__reserved_2;
951 	 *			u8		build_id[20];
952 	 *		};
953 	 *	};
954 	 *	u32				prot, flags;
955 	 *	char				filename[];
956 	 * 	struct sample_id		sample_id;
957 	 * };
958 	 */
959 	PERF_RECORD_MMAP2			= 10,
960 
961 	/*
962 	 * Records that new data landed in the AUX buffer part.
963 	 *
964 	 * struct {
965 	 * 	struct perf_event_header	header;
966 	 *
967 	 * 	u64				aux_offset;
968 	 * 	u64				aux_size;
969 	 *	u64				flags;
970 	 * 	struct sample_id		sample_id;
971 	 * };
972 	 */
973 	PERF_RECORD_AUX				= 11,
974 
975 	/*
976 	 * Indicates that instruction trace has started
977 	 *
978 	 * struct {
979 	 *	struct perf_event_header	header;
980 	 *	u32				pid;
981 	 *	u32				tid;
982 	 *	struct sample_id		sample_id;
983 	 * };
984 	 */
985 	PERF_RECORD_ITRACE_START		= 12,
986 
987 	/*
988 	 * Records the dropped/lost sample number.
989 	 *
990 	 * struct {
991 	 *	struct perf_event_header	header;
992 	 *
993 	 *	u64				lost;
994 	 *	struct sample_id		sample_id;
995 	 * };
996 	 */
997 	PERF_RECORD_LOST_SAMPLES		= 13,
998 
999 	/*
1000 	 * Records a context switch in or out (flagged by
1001 	 * PERF_RECORD_MISC_SWITCH_OUT). See also
1002 	 * PERF_RECORD_SWITCH_CPU_WIDE.
1003 	 *
1004 	 * struct {
1005 	 *	struct perf_event_header	header;
1006 	 *	struct sample_id		sample_id;
1007 	 * };
1008 	 */
1009 	PERF_RECORD_SWITCH			= 14,
1010 
1011 	/*
1012 	 * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and
1013 	 * next_prev_tid that are the next (switching out) or previous
1014 	 * (switching in) pid/tid.
1015 	 *
1016 	 * struct {
1017 	 *	struct perf_event_header	header;
1018 	 *	u32				next_prev_pid;
1019 	 *	u32				next_prev_tid;
1020 	 *	struct sample_id		sample_id;
1021 	 * };
1022 	 */
1023 	PERF_RECORD_SWITCH_CPU_WIDE		= 15,
1024 
1025 	/*
1026 	 * struct {
1027 	 *	struct perf_event_header	header;
1028 	 *	u32				pid;
1029 	 *	u32				tid;
1030 	 *	u64				nr_namespaces;
1031 	 *	{ u64				dev, inode; } [nr_namespaces];
1032 	 *	struct sample_id		sample_id;
1033 	 * };
1034 	 */
1035 	PERF_RECORD_NAMESPACES			= 16,
1036 
1037 	/*
1038 	 * Record ksymbol register/unregister events:
1039 	 *
1040 	 * struct {
1041 	 *	struct perf_event_header	header;
1042 	 *	u64				addr;
1043 	 *	u32				len;
1044 	 *	u16				ksym_type;
1045 	 *	u16				flags;
1046 	 *	char				name[];
1047 	 *	struct sample_id		sample_id;
1048 	 * };
1049 	 */
1050 	PERF_RECORD_KSYMBOL			= 17,
1051 
1052 	/*
1053 	 * Record bpf events:
1054 	 *  enum perf_bpf_event_type {
1055 	 *	PERF_BPF_EVENT_UNKNOWN		= 0,
1056 	 *	PERF_BPF_EVENT_PROG_LOAD	= 1,
1057 	 *	PERF_BPF_EVENT_PROG_UNLOAD	= 2,
1058 	 *  };
1059 	 *
1060 	 * struct {
1061 	 *	struct perf_event_header	header;
1062 	 *	u16				type;
1063 	 *	u16				flags;
1064 	 *	u32				id;
1065 	 *	u8				tag[BPF_TAG_SIZE];
1066 	 *	struct sample_id		sample_id;
1067 	 * };
1068 	 */
1069 	PERF_RECORD_BPF_EVENT			= 18,
1070 
1071 	/*
1072 	 * struct {
1073 	 *	struct perf_event_header	header;
1074 	 *	u64				id;
1075 	 *	char				path[];
1076 	 *	struct sample_id		sample_id;
1077 	 * };
1078 	 */
1079 	PERF_RECORD_CGROUP			= 19,
1080 
1081 	/*
1082 	 * Records changes to kernel text i.e. self-modified code. 'old_len' is
1083 	 * the number of old bytes, 'new_len' is the number of new bytes. Either
1084 	 * 'old_len' or 'new_len' may be zero to indicate, for example, the
1085 	 * addition or removal of a trampoline. 'bytes' contains the old bytes
1086 	 * followed immediately by the new bytes.
1087 	 *
1088 	 * struct {
1089 	 *	struct perf_event_header	header;
1090 	 *	u64				addr;
1091 	 *	u16				old_len;
1092 	 *	u16				new_len;
1093 	 *	u8				bytes[];
1094 	 *	struct sample_id		sample_id;
1095 	 * };
1096 	 */
1097 	PERF_RECORD_TEXT_POKE			= 20,
1098 
1099 	PERF_RECORD_MAX,			/* non-ABI */
1100 };
1101 
1102 enum perf_record_ksymbol_type {
1103 	PERF_RECORD_KSYMBOL_TYPE_UNKNOWN	= 0,
1104 	PERF_RECORD_KSYMBOL_TYPE_BPF		= 1,
1105 	/*
1106 	 * Out of line code such as kprobe-replaced instructions or optimized
1107 	 * kprobes or ftrace trampolines.
1108 	 */
1109 	PERF_RECORD_KSYMBOL_TYPE_OOL		= 2,
1110 	PERF_RECORD_KSYMBOL_TYPE_MAX		/* non-ABI */
1111 };
1112 
1113 #define PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER	(1 << 0)
1114 
1115 enum perf_bpf_event_type {
1116 	PERF_BPF_EVENT_UNKNOWN		= 0,
1117 	PERF_BPF_EVENT_PROG_LOAD	= 1,
1118 	PERF_BPF_EVENT_PROG_UNLOAD	= 2,
1119 	PERF_BPF_EVENT_MAX,		/* non-ABI */
1120 };
1121 
1122 #define PERF_MAX_STACK_DEPTH		127
1123 #define PERF_MAX_CONTEXTS_PER_STACK	  8
1124 
1125 enum perf_callchain_context {
1126 	PERF_CONTEXT_HV			= (__u64)-32,
1127 	PERF_CONTEXT_KERNEL		= (__u64)-128,
1128 	PERF_CONTEXT_USER		= (__u64)-512,
1129 
1130 	PERF_CONTEXT_GUEST		= (__u64)-2048,
1131 	PERF_CONTEXT_GUEST_KERNEL	= (__u64)-2176,
1132 	PERF_CONTEXT_GUEST_USER		= (__u64)-2560,
1133 
1134 	PERF_CONTEXT_MAX		= (__u64)-4095,
1135 };
1136 
1137 /**
1138  * PERF_RECORD_AUX::flags bits
1139  */
1140 #define PERF_AUX_FLAG_TRUNCATED		0x01	/* record was truncated to fit */
1141 #define PERF_AUX_FLAG_OVERWRITE		0x02	/* snapshot from overwrite mode */
1142 #define PERF_AUX_FLAG_PARTIAL		0x04	/* record contains gaps */
1143 #define PERF_AUX_FLAG_COLLISION		0x08	/* sample collided with another */
1144 
1145 #define PERF_FLAG_FD_NO_GROUP		(1UL << 0)
1146 #define PERF_FLAG_FD_OUTPUT		(1UL << 1)
1147 #define PERF_FLAG_PID_CGROUP		(1UL << 2) /* pid=cgroup id, per-cpu mode only */
1148 #define PERF_FLAG_FD_CLOEXEC		(1UL << 3) /* O_CLOEXEC */
1149 
1150 #if defined(__LITTLE_ENDIAN_BITFIELD)
1151 union perf_mem_data_src {
1152 	__u64 val;
1153 	struct {
1154 		__u64   mem_op:5,	/* type of opcode */
1155 			mem_lvl:14,	/* memory hierarchy level */
1156 			mem_snoop:5,	/* snoop mode */
1157 			mem_lock:2,	/* lock instr */
1158 			mem_dtlb:7,	/* tlb access */
1159 			mem_lvl_num:4,	/* memory hierarchy level number */
1160 			mem_remote:1,   /* remote */
1161 			mem_snoopx:2,	/* snoop mode, ext */
1162 			mem_rsvd:24;
1163 	};
1164 };
1165 #elif defined(__BIG_ENDIAN_BITFIELD)
1166 union perf_mem_data_src {
1167 	__u64 val;
1168 	struct {
1169 		__u64	mem_rsvd:24,
1170 			mem_snoopx:2,	/* snoop mode, ext */
1171 			mem_remote:1,   /* remote */
1172 			mem_lvl_num:4,	/* memory hierarchy level number */
1173 			mem_dtlb:7,	/* tlb access */
1174 			mem_lock:2,	/* lock instr */
1175 			mem_snoop:5,	/* snoop mode */
1176 			mem_lvl:14,	/* memory hierarchy level */
1177 			mem_op:5;	/* type of opcode */
1178 	};
1179 };
1180 #else
1181 #error "Unknown endianness"
1182 #endif
1183 
1184 /* type of opcode (load/store/prefetch,code) */
1185 #define PERF_MEM_OP_NA		0x01 /* not available */
1186 #define PERF_MEM_OP_LOAD	0x02 /* load instruction */
1187 #define PERF_MEM_OP_STORE	0x04 /* store instruction */
1188 #define PERF_MEM_OP_PFETCH	0x08 /* prefetch */
1189 #define PERF_MEM_OP_EXEC	0x10 /* code (execution) */
1190 #define PERF_MEM_OP_SHIFT	0
1191 
1192 /* memory hierarchy (memory level, hit or miss) */
1193 #define PERF_MEM_LVL_NA		0x01  /* not available */
1194 #define PERF_MEM_LVL_HIT	0x02  /* hit level */
1195 #define PERF_MEM_LVL_MISS	0x04  /* miss level  */
1196 #define PERF_MEM_LVL_L1		0x08  /* L1 */
1197 #define PERF_MEM_LVL_LFB	0x10  /* Line Fill Buffer */
1198 #define PERF_MEM_LVL_L2		0x20  /* L2 */
1199 #define PERF_MEM_LVL_L3		0x40  /* L3 */
1200 #define PERF_MEM_LVL_LOC_RAM	0x80  /* Local DRAM */
1201 #define PERF_MEM_LVL_REM_RAM1	0x100 /* Remote DRAM (1 hop) */
1202 #define PERF_MEM_LVL_REM_RAM2	0x200 /* Remote DRAM (2 hops) */
1203 #define PERF_MEM_LVL_REM_CCE1	0x400 /* Remote Cache (1 hop) */
1204 #define PERF_MEM_LVL_REM_CCE2	0x800 /* Remote Cache (2 hops) */
1205 #define PERF_MEM_LVL_IO		0x1000 /* I/O memory */
1206 #define PERF_MEM_LVL_UNC	0x2000 /* Uncached memory */
1207 #define PERF_MEM_LVL_SHIFT	5
1208 
1209 #define PERF_MEM_REMOTE_REMOTE	0x01  /* Remote */
1210 #define PERF_MEM_REMOTE_SHIFT	37
1211 
1212 #define PERF_MEM_LVLNUM_L1	0x01 /* L1 */
1213 #define PERF_MEM_LVLNUM_L2	0x02 /* L2 */
1214 #define PERF_MEM_LVLNUM_L3	0x03 /* L3 */
1215 #define PERF_MEM_LVLNUM_L4	0x04 /* L4 */
1216 /* 5-0xa available */
1217 #define PERF_MEM_LVLNUM_ANY_CACHE 0x0b /* Any cache */
1218 #define PERF_MEM_LVLNUM_LFB	0x0c /* LFB */
1219 #define PERF_MEM_LVLNUM_RAM	0x0d /* RAM */
1220 #define PERF_MEM_LVLNUM_PMEM	0x0e /* PMEM */
1221 #define PERF_MEM_LVLNUM_NA	0x0f /* N/A */
1222 
1223 #define PERF_MEM_LVLNUM_SHIFT	33
1224 
1225 /* snoop mode */
1226 #define PERF_MEM_SNOOP_NA	0x01 /* not available */
1227 #define PERF_MEM_SNOOP_NONE	0x02 /* no snoop */
1228 #define PERF_MEM_SNOOP_HIT	0x04 /* snoop hit */
1229 #define PERF_MEM_SNOOP_MISS	0x08 /* snoop miss */
1230 #define PERF_MEM_SNOOP_HITM	0x10 /* snoop hit modified */
1231 #define PERF_MEM_SNOOP_SHIFT	19
1232 
1233 #define PERF_MEM_SNOOPX_FWD	0x01 /* forward */
1234 /* 1 free */
1235 #define PERF_MEM_SNOOPX_SHIFT  38
1236 
1237 /* locked instruction */
1238 #define PERF_MEM_LOCK_NA	0x01 /* not available */
1239 #define PERF_MEM_LOCK_LOCKED	0x02 /* locked transaction */
1240 #define PERF_MEM_LOCK_SHIFT	24
1241 
1242 /* TLB access */
1243 #define PERF_MEM_TLB_NA		0x01 /* not available */
1244 #define PERF_MEM_TLB_HIT	0x02 /* hit level */
1245 #define PERF_MEM_TLB_MISS	0x04 /* miss level */
1246 #define PERF_MEM_TLB_L1		0x08 /* L1 */
1247 #define PERF_MEM_TLB_L2		0x10 /* L2 */
1248 #define PERF_MEM_TLB_WK		0x20 /* Hardware Walker*/
1249 #define PERF_MEM_TLB_OS		0x40 /* OS fault handler */
1250 #define PERF_MEM_TLB_SHIFT	26
1251 
1252 #define PERF_MEM_S(a, s) \
1253 	(((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)
1254 
1255 /*
1256  * single taken branch record layout:
1257  *
1258  *      from: source instruction (may not always be a branch insn)
1259  *        to: branch target
1260  *   mispred: branch target was mispredicted
1261  * predicted: branch target was predicted
1262  *
1263  * support for mispred, predicted is optional. In case it
1264  * is not supported mispred = predicted = 0.
1265  *
1266  *     in_tx: running in a hardware transaction
1267  *     abort: aborting a hardware transaction
1268  *    cycles: cycles from last branch (or 0 if not supported)
1269  *      type: branch type
1270  */
1271 struct perf_branch_entry {
1272 	__u64	from;
1273 	__u64	to;
1274 	__u64	mispred:1,  /* target mispredicted */
1275 		predicted:1,/* target predicted */
1276 		in_tx:1,    /* in transaction */
1277 		abort:1,    /* transaction abort */
1278 		cycles:16,  /* cycle count to last branch */
1279 		type:4,     /* branch type */
1280 		reserved:40;
1281 };
1282 
1283 #endif /* _UAPI_LINUX_PERF_EVENT_H */
1284