xref: /linux-6.15/include/uapi/linux/perf_event.h (revision b190bc4a)
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  * attr.config layout for type PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE
42  * PERF_TYPE_HARDWARE:			0xEEEEEEEE000000AA
43  *					AA: hardware event ID
44  *					EEEEEEEE: PMU type ID
45  * PERF_TYPE_HW_CACHE:			0xEEEEEEEE00DDCCBB
46  *					BB: hardware cache ID
47  *					CC: hardware cache op ID
48  *					DD: hardware cache op result ID
49  *					EEEEEEEE: PMU type ID
50  * If the PMU type ID is 0, the PERF_TYPE_RAW will be applied.
51  */
52 #define PERF_PMU_TYPE_SHIFT		32
53 #define PERF_HW_EVENT_MASK		0xffffffff
54 
55 /*
56  * Generalized performance event event_id types, used by the
57  * attr.event_id parameter of the sys_perf_event_open()
58  * syscall:
59  */
60 enum perf_hw_id {
61 	/*
62 	 * Common hardware events, generalized by the kernel:
63 	 */
64 	PERF_COUNT_HW_CPU_CYCLES		= 0,
65 	PERF_COUNT_HW_INSTRUCTIONS		= 1,
66 	PERF_COUNT_HW_CACHE_REFERENCES		= 2,
67 	PERF_COUNT_HW_CACHE_MISSES		= 3,
68 	PERF_COUNT_HW_BRANCH_INSTRUCTIONS	= 4,
69 	PERF_COUNT_HW_BRANCH_MISSES		= 5,
70 	PERF_COUNT_HW_BUS_CYCLES		= 6,
71 	PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	= 7,
72 	PERF_COUNT_HW_STALLED_CYCLES_BACKEND	= 8,
73 	PERF_COUNT_HW_REF_CPU_CYCLES		= 9,
74 
75 	PERF_COUNT_HW_MAX,			/* non-ABI */
76 };
77 
78 /*
79  * Generalized hardware cache events:
80  *
81  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
82  *       { read, write, prefetch } x
83  *       { accesses, misses }
84  */
85 enum perf_hw_cache_id {
86 	PERF_COUNT_HW_CACHE_L1D			= 0,
87 	PERF_COUNT_HW_CACHE_L1I			= 1,
88 	PERF_COUNT_HW_CACHE_LL			= 2,
89 	PERF_COUNT_HW_CACHE_DTLB		= 3,
90 	PERF_COUNT_HW_CACHE_ITLB		= 4,
91 	PERF_COUNT_HW_CACHE_BPU			= 5,
92 	PERF_COUNT_HW_CACHE_NODE		= 6,
93 
94 	PERF_COUNT_HW_CACHE_MAX,		/* non-ABI */
95 };
96 
97 enum perf_hw_cache_op_id {
98 	PERF_COUNT_HW_CACHE_OP_READ		= 0,
99 	PERF_COUNT_HW_CACHE_OP_WRITE		= 1,
100 	PERF_COUNT_HW_CACHE_OP_PREFETCH		= 2,
101 
102 	PERF_COUNT_HW_CACHE_OP_MAX,		/* non-ABI */
103 };
104 
105 enum perf_hw_cache_op_result_id {
106 	PERF_COUNT_HW_CACHE_RESULT_ACCESS	= 0,
107 	PERF_COUNT_HW_CACHE_RESULT_MISS		= 1,
108 
109 	PERF_COUNT_HW_CACHE_RESULT_MAX,		/* non-ABI */
110 };
111 
112 /*
113  * Special "software" events provided by the kernel, even if the hardware
114  * does not support performance events. These events measure various
115  * physical and sw events of the kernel (and allow the profiling of them as
116  * well):
117  */
118 enum perf_sw_ids {
119 	PERF_COUNT_SW_CPU_CLOCK			= 0,
120 	PERF_COUNT_SW_TASK_CLOCK		= 1,
121 	PERF_COUNT_SW_PAGE_FAULTS		= 2,
122 	PERF_COUNT_SW_CONTEXT_SWITCHES		= 3,
123 	PERF_COUNT_SW_CPU_MIGRATIONS		= 4,
124 	PERF_COUNT_SW_PAGE_FAULTS_MIN		= 5,
125 	PERF_COUNT_SW_PAGE_FAULTS_MAJ		= 6,
126 	PERF_COUNT_SW_ALIGNMENT_FAULTS		= 7,
127 	PERF_COUNT_SW_EMULATION_FAULTS		= 8,
128 	PERF_COUNT_SW_DUMMY			= 9,
129 	PERF_COUNT_SW_BPF_OUTPUT		= 10,
130 	PERF_COUNT_SW_CGROUP_SWITCHES		= 11,
131 
132 	PERF_COUNT_SW_MAX,			/* non-ABI */
133 };
134 
135 /*
136  * Bits that can be set in attr.sample_type to request information
137  * in the overflow packets.
138  */
139 enum perf_event_sample_format {
140 	PERF_SAMPLE_IP				= 1U << 0,
141 	PERF_SAMPLE_TID				= 1U << 1,
142 	PERF_SAMPLE_TIME			= 1U << 2,
143 	PERF_SAMPLE_ADDR			= 1U << 3,
144 	PERF_SAMPLE_READ			= 1U << 4,
145 	PERF_SAMPLE_CALLCHAIN			= 1U << 5,
146 	PERF_SAMPLE_ID				= 1U << 6,
147 	PERF_SAMPLE_CPU				= 1U << 7,
148 	PERF_SAMPLE_PERIOD			= 1U << 8,
149 	PERF_SAMPLE_STREAM_ID			= 1U << 9,
150 	PERF_SAMPLE_RAW				= 1U << 10,
151 	PERF_SAMPLE_BRANCH_STACK		= 1U << 11,
152 	PERF_SAMPLE_REGS_USER			= 1U << 12,
153 	PERF_SAMPLE_STACK_USER			= 1U << 13,
154 	PERF_SAMPLE_WEIGHT			= 1U << 14,
155 	PERF_SAMPLE_DATA_SRC			= 1U << 15,
156 	PERF_SAMPLE_IDENTIFIER			= 1U << 16,
157 	PERF_SAMPLE_TRANSACTION			= 1U << 17,
158 	PERF_SAMPLE_REGS_INTR			= 1U << 18,
159 	PERF_SAMPLE_PHYS_ADDR			= 1U << 19,
160 	PERF_SAMPLE_AUX				= 1U << 20,
161 	PERF_SAMPLE_CGROUP			= 1U << 21,
162 	PERF_SAMPLE_DATA_PAGE_SIZE		= 1U << 22,
163 	PERF_SAMPLE_CODE_PAGE_SIZE		= 1U << 23,
164 	PERF_SAMPLE_WEIGHT_STRUCT		= 1U << 24,
165 
166 	PERF_SAMPLE_MAX = 1U << 25,		/* non-ABI */
167 
168 	__PERF_SAMPLE_CALLCHAIN_EARLY		= 1ULL << 63, /* non-ABI; internal use */
169 };
170 
171 #define PERF_SAMPLE_WEIGHT_TYPE	(PERF_SAMPLE_WEIGHT | PERF_SAMPLE_WEIGHT_STRUCT)
172 /*
173  * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
174  *
175  * If the user does not pass priv level information via branch_sample_type,
176  * the kernel uses the event's priv level. Branch and event priv levels do
177  * not have to match. Branch priv level is checked for permissions.
178  *
179  * The branch types can be combined, however BRANCH_ANY covers all types
180  * of branches and therefore it supersedes all the other types.
181  */
182 enum perf_branch_sample_type_shift {
183 	PERF_SAMPLE_BRANCH_USER_SHIFT		= 0, /* user branches */
184 	PERF_SAMPLE_BRANCH_KERNEL_SHIFT		= 1, /* kernel branches */
185 	PERF_SAMPLE_BRANCH_HV_SHIFT		= 2, /* hypervisor branches */
186 
187 	PERF_SAMPLE_BRANCH_ANY_SHIFT		= 3, /* any branch types */
188 	PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT	= 4, /* any call branch */
189 	PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT	= 5, /* any return branch */
190 	PERF_SAMPLE_BRANCH_IND_CALL_SHIFT	= 6, /* indirect calls */
191 	PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT	= 7, /* transaction aborts */
192 	PERF_SAMPLE_BRANCH_IN_TX_SHIFT		= 8, /* in transaction */
193 	PERF_SAMPLE_BRANCH_NO_TX_SHIFT		= 9, /* not in transaction */
194 	PERF_SAMPLE_BRANCH_COND_SHIFT		= 10, /* conditional branches */
195 
196 	PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT	= 11, /* call/ret stack */
197 	PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT	= 12, /* indirect jumps */
198 	PERF_SAMPLE_BRANCH_CALL_SHIFT		= 13, /* direct call */
199 
200 	PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT	= 14, /* no flags */
201 	PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT	= 15, /* no cycles */
202 
203 	PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT	= 16, /* save branch type */
204 
205 	PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT	= 17, /* save low level index of raw branch records */
206 
207 	PERF_SAMPLE_BRANCH_MAX_SHIFT		/* non-ABI */
208 };
209 
210 enum perf_branch_sample_type {
211 	PERF_SAMPLE_BRANCH_USER		= 1U << PERF_SAMPLE_BRANCH_USER_SHIFT,
212 	PERF_SAMPLE_BRANCH_KERNEL	= 1U << PERF_SAMPLE_BRANCH_KERNEL_SHIFT,
213 	PERF_SAMPLE_BRANCH_HV		= 1U << PERF_SAMPLE_BRANCH_HV_SHIFT,
214 
215 	PERF_SAMPLE_BRANCH_ANY		= 1U << PERF_SAMPLE_BRANCH_ANY_SHIFT,
216 	PERF_SAMPLE_BRANCH_ANY_CALL	= 1U << PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT,
217 	PERF_SAMPLE_BRANCH_ANY_RETURN	= 1U << PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT,
218 	PERF_SAMPLE_BRANCH_IND_CALL	= 1U << PERF_SAMPLE_BRANCH_IND_CALL_SHIFT,
219 	PERF_SAMPLE_BRANCH_ABORT_TX	= 1U << PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT,
220 	PERF_SAMPLE_BRANCH_IN_TX	= 1U << PERF_SAMPLE_BRANCH_IN_TX_SHIFT,
221 	PERF_SAMPLE_BRANCH_NO_TX	= 1U << PERF_SAMPLE_BRANCH_NO_TX_SHIFT,
222 	PERF_SAMPLE_BRANCH_COND		= 1U << PERF_SAMPLE_BRANCH_COND_SHIFT,
223 
224 	PERF_SAMPLE_BRANCH_CALL_STACK	= 1U << PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT,
225 	PERF_SAMPLE_BRANCH_IND_JUMP	= 1U << PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT,
226 	PERF_SAMPLE_BRANCH_CALL		= 1U << PERF_SAMPLE_BRANCH_CALL_SHIFT,
227 
228 	PERF_SAMPLE_BRANCH_NO_FLAGS	= 1U << PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT,
229 	PERF_SAMPLE_BRANCH_NO_CYCLES	= 1U << PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT,
230 
231 	PERF_SAMPLE_BRANCH_TYPE_SAVE	=
232 		1U << PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT,
233 
234 	PERF_SAMPLE_BRANCH_HW_INDEX	= 1U << PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT,
235 
236 	PERF_SAMPLE_BRANCH_MAX		= 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT,
237 };
238 
239 /*
240  * Common flow change classification
241  */
242 enum {
243 	PERF_BR_UNKNOWN		= 0,	/* unknown */
244 	PERF_BR_COND		= 1,	/* conditional */
245 	PERF_BR_UNCOND		= 2,	/* unconditional  */
246 	PERF_BR_IND		= 3,	/* indirect */
247 	PERF_BR_CALL		= 4,	/* function call */
248 	PERF_BR_IND_CALL	= 5,	/* indirect function call */
249 	PERF_BR_RET		= 6,	/* function return */
250 	PERF_BR_SYSCALL		= 7,	/* syscall */
251 	PERF_BR_SYSRET		= 8,	/* syscall return */
252 	PERF_BR_COND_CALL	= 9,	/* conditional function call */
253 	PERF_BR_COND_RET	= 10,	/* conditional function return */
254 	PERF_BR_ERET		= 11,	/* exception return */
255 	PERF_BR_IRQ		= 12,	/* irq */
256 	PERF_BR_SERROR		= 13,	/* system error */
257 	PERF_BR_NO_TX		= 14,	/* not in transaction */
258 	PERF_BR_EXTEND_ABI	= 15,	/* extend ABI */
259 	PERF_BR_MAX,
260 };
261 
262 /*
263  * Common branch speculation outcome classification
264  */
265 enum {
266 	PERF_BR_SPEC_NA			= 0,	/* Not available */
267 	PERF_BR_SPEC_WRONG_PATH		= 1,	/* Speculative but on wrong path */
268 	PERF_BR_NON_SPEC_CORRECT_PATH	= 2,	/* Non-speculative but on correct path */
269 	PERF_BR_SPEC_CORRECT_PATH	= 3,	/* Speculative and on correct path */
270 	PERF_BR_SPEC_MAX,
271 };
272 
273 enum {
274 	PERF_BR_NEW_FAULT_ALGN		= 0,    /* Alignment fault */
275 	PERF_BR_NEW_FAULT_DATA		= 1,    /* Data fault */
276 	PERF_BR_NEW_FAULT_INST		= 2,    /* Inst fault */
277 	PERF_BR_NEW_ARCH_1		= 3,    /* Architecture specific */
278 	PERF_BR_NEW_ARCH_2		= 4,    /* Architecture specific */
279 	PERF_BR_NEW_ARCH_3		= 5,    /* Architecture specific */
280 	PERF_BR_NEW_ARCH_4		= 6,    /* Architecture specific */
281 	PERF_BR_NEW_ARCH_5		= 7,    /* Architecture specific */
282 	PERF_BR_NEW_MAX,
283 };
284 
285 #define PERF_SAMPLE_BRANCH_PLM_ALL \
286 	(PERF_SAMPLE_BRANCH_USER|\
287 	 PERF_SAMPLE_BRANCH_KERNEL|\
288 	 PERF_SAMPLE_BRANCH_HV)
289 
290 /*
291  * Values to determine ABI of the registers dump.
292  */
293 enum perf_sample_regs_abi {
294 	PERF_SAMPLE_REGS_ABI_NONE	= 0,
295 	PERF_SAMPLE_REGS_ABI_32		= 1,
296 	PERF_SAMPLE_REGS_ABI_64		= 2,
297 };
298 
299 /*
300  * Values for the memory transaction event qualifier, mostly for
301  * abort events. Multiple bits can be set.
302  */
303 enum {
304 	PERF_TXN_ELISION        = (1 << 0), /* From elision */
305 	PERF_TXN_TRANSACTION    = (1 << 1), /* From transaction */
306 	PERF_TXN_SYNC           = (1 << 2), /* Instruction is related */
307 	PERF_TXN_ASYNC          = (1 << 3), /* Instruction not related */
308 	PERF_TXN_RETRY          = (1 << 4), /* Retry possible */
309 	PERF_TXN_CONFLICT       = (1 << 5), /* Conflict abort */
310 	PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */
311 	PERF_TXN_CAPACITY_READ  = (1 << 7), /* Capacity read abort */
312 
313 	PERF_TXN_MAX	        = (1 << 8), /* non-ABI */
314 
315 	/* bits 32..63 are reserved for the abort code */
316 
317 	PERF_TXN_ABORT_MASK  = (0xffffffffULL << 32),
318 	PERF_TXN_ABORT_SHIFT = 32,
319 };
320 
321 /*
322  * The format of the data returned by read() on a perf event fd,
323  * as specified by attr.read_format:
324  *
325  * struct read_format {
326  *	{ u64		value;
327  *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
328  *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
329  *	  { u64		id;           } && PERF_FORMAT_ID
330  *	  { u64		lost;         } && PERF_FORMAT_LOST
331  *	} && !PERF_FORMAT_GROUP
332  *
333  *	{ u64		nr;
334  *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
335  *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
336  *	  { u64		value;
337  *	    { u64	id;           } && PERF_FORMAT_ID
338  *	    { u64	lost;         } && PERF_FORMAT_LOST
339  *	  }		cntr[nr];
340  *	} && PERF_FORMAT_GROUP
341  * };
342  */
343 enum perf_event_read_format {
344 	PERF_FORMAT_TOTAL_TIME_ENABLED		= 1U << 0,
345 	PERF_FORMAT_TOTAL_TIME_RUNNING		= 1U << 1,
346 	PERF_FORMAT_ID				= 1U << 2,
347 	PERF_FORMAT_GROUP			= 1U << 3,
348 	PERF_FORMAT_LOST			= 1U << 4,
349 
350 	PERF_FORMAT_MAX = 1U << 5,		/* non-ABI */
351 };
352 
353 #define PERF_ATTR_SIZE_VER0	64	/* sizeof first published struct */
354 #define PERF_ATTR_SIZE_VER1	72	/* add: config2 */
355 #define PERF_ATTR_SIZE_VER2	80	/* add: branch_sample_type */
356 #define PERF_ATTR_SIZE_VER3	96	/* add: sample_regs_user */
357 					/* add: sample_stack_user */
358 #define PERF_ATTR_SIZE_VER4	104	/* add: sample_regs_intr */
359 #define PERF_ATTR_SIZE_VER5	112	/* add: aux_watermark */
360 #define PERF_ATTR_SIZE_VER6	120	/* add: aux_sample_size */
361 #define PERF_ATTR_SIZE_VER7	128	/* add: sig_data */
362 
363 /*
364  * Hardware event_id to monitor via a performance monitoring event:
365  *
366  * @sample_max_stack: Max number of frame pointers in a callchain,
367  *		      should be < /proc/sys/kernel/perf_event_max_stack
368  */
369 struct perf_event_attr {
370 
371 	/*
372 	 * Major type: hardware/software/tracepoint/etc.
373 	 */
374 	__u32			type;
375 
376 	/*
377 	 * Size of the attr structure, for fwd/bwd compat.
378 	 */
379 	__u32			size;
380 
381 	/*
382 	 * Type specific configuration information.
383 	 */
384 	__u64			config;
385 
386 	union {
387 		__u64		sample_period;
388 		__u64		sample_freq;
389 	};
390 
391 	__u64			sample_type;
392 	__u64			read_format;
393 
394 	__u64			disabled       :  1, /* off by default        */
395 				inherit	       :  1, /* children inherit it   */
396 				pinned	       :  1, /* must always be on PMU */
397 				exclusive      :  1, /* only group on PMU     */
398 				exclude_user   :  1, /* don't count user      */
399 				exclude_kernel :  1, /* ditto kernel          */
400 				exclude_hv     :  1, /* ditto hypervisor      */
401 				exclude_idle   :  1, /* don't count when idle */
402 				mmap           :  1, /* include mmap data     */
403 				comm	       :  1, /* include comm data     */
404 				freq           :  1, /* use freq, not period  */
405 				inherit_stat   :  1, /* per task counts       */
406 				enable_on_exec :  1, /* next exec enables     */
407 				task           :  1, /* trace fork/exit       */
408 				watermark      :  1, /* wakeup_watermark      */
409 				/*
410 				 * precise_ip:
411 				 *
412 				 *  0 - SAMPLE_IP can have arbitrary skid
413 				 *  1 - SAMPLE_IP must have constant skid
414 				 *  2 - SAMPLE_IP requested to have 0 skid
415 				 *  3 - SAMPLE_IP must have 0 skid
416 				 *
417 				 *  See also PERF_RECORD_MISC_EXACT_IP
418 				 */
419 				precise_ip     :  2, /* skid constraint       */
420 				mmap_data      :  1, /* non-exec mmap data    */
421 				sample_id_all  :  1, /* sample_type all events */
422 
423 				exclude_host   :  1, /* don't count in host   */
424 				exclude_guest  :  1, /* don't count in guest  */
425 
426 				exclude_callchain_kernel : 1, /* exclude kernel callchains */
427 				exclude_callchain_user   : 1, /* exclude user callchains */
428 				mmap2          :  1, /* include mmap with inode data     */
429 				comm_exec      :  1, /* flag comm events that are due to an exec */
430 				use_clockid    :  1, /* use @clockid for time fields */
431 				context_switch :  1, /* context switch data */
432 				write_backward :  1, /* Write ring buffer from end to beginning */
433 				namespaces     :  1, /* include namespaces data */
434 				ksymbol        :  1, /* include ksymbol events */
435 				bpf_event      :  1, /* include bpf events */
436 				aux_output     :  1, /* generate AUX records instead of events */
437 				cgroup         :  1, /* include cgroup events */
438 				text_poke      :  1, /* include text poke events */
439 				build_id       :  1, /* use build id in mmap2 events */
440 				inherit_thread :  1, /* children only inherit if cloned with CLONE_THREAD */
441 				remove_on_exec :  1, /* event is removed from task on exec */
442 				sigtrap        :  1, /* send synchronous SIGTRAP on event */
443 				__reserved_1   : 26;
444 
445 	union {
446 		__u32		wakeup_events;	  /* wakeup every n events */
447 		__u32		wakeup_watermark; /* bytes before wakeup   */
448 	};
449 
450 	__u32			bp_type;
451 	union {
452 		__u64		bp_addr;
453 		__u64		kprobe_func; /* for perf_kprobe */
454 		__u64		uprobe_path; /* for perf_uprobe */
455 		__u64		config1; /* extension of config */
456 	};
457 	union {
458 		__u64		bp_len;
459 		__u64		kprobe_addr; /* when kprobe_func == NULL */
460 		__u64		probe_offset; /* for perf_[k,u]probe */
461 		__u64		config2; /* extension of config1 */
462 	};
463 	__u64	branch_sample_type; /* enum perf_branch_sample_type */
464 
465 	/*
466 	 * Defines set of user regs to dump on samples.
467 	 * See asm/perf_regs.h for details.
468 	 */
469 	__u64	sample_regs_user;
470 
471 	/*
472 	 * Defines size of the user stack to dump on samples.
473 	 */
474 	__u32	sample_stack_user;
475 
476 	__s32	clockid;
477 	/*
478 	 * Defines set of regs to dump for each sample
479 	 * state captured on:
480 	 *  - precise = 0: PMU interrupt
481 	 *  - precise > 0: sampled instruction
482 	 *
483 	 * See asm/perf_regs.h for details.
484 	 */
485 	__u64	sample_regs_intr;
486 
487 	/*
488 	 * Wakeup watermark for AUX area
489 	 */
490 	__u32	aux_watermark;
491 	__u16	sample_max_stack;
492 	__u16	__reserved_2;
493 	__u32	aux_sample_size;
494 	__u32	__reserved_3;
495 
496 	/*
497 	 * User provided data if sigtrap=1, passed back to user via
498 	 * siginfo_t::si_perf_data, e.g. to permit user to identify the event.
499 	 * Note, siginfo_t::si_perf_data is long-sized, and sig_data will be
500 	 * truncated accordingly on 32 bit architectures.
501 	 */
502 	__u64	sig_data;
503 };
504 
505 /*
506  * Structure used by below PERF_EVENT_IOC_QUERY_BPF command
507  * to query bpf programs attached to the same perf tracepoint
508  * as the given perf event.
509  */
510 struct perf_event_query_bpf {
511 	/*
512 	 * The below ids array length
513 	 */
514 	__u32	ids_len;
515 	/*
516 	 * Set by the kernel to indicate the number of
517 	 * available programs
518 	 */
519 	__u32	prog_cnt;
520 	/*
521 	 * User provided buffer to store program ids
522 	 */
523 	__u32	ids[];
524 };
525 
526 /*
527  * Ioctls that can be done on a perf event fd:
528  */
529 #define PERF_EVENT_IOC_ENABLE			_IO ('$', 0)
530 #define PERF_EVENT_IOC_DISABLE			_IO ('$', 1)
531 #define PERF_EVENT_IOC_REFRESH			_IO ('$', 2)
532 #define PERF_EVENT_IOC_RESET			_IO ('$', 3)
533 #define PERF_EVENT_IOC_PERIOD			_IOW('$', 4, __u64)
534 #define PERF_EVENT_IOC_SET_OUTPUT		_IO ('$', 5)
535 #define PERF_EVENT_IOC_SET_FILTER		_IOW('$', 6, char *)
536 #define PERF_EVENT_IOC_ID			_IOR('$', 7, __u64 *)
537 #define PERF_EVENT_IOC_SET_BPF			_IOW('$', 8, __u32)
538 #define PERF_EVENT_IOC_PAUSE_OUTPUT		_IOW('$', 9, __u32)
539 #define PERF_EVENT_IOC_QUERY_BPF		_IOWR('$', 10, struct perf_event_query_bpf *)
540 #define PERF_EVENT_IOC_MODIFY_ATTRIBUTES	_IOW('$', 11, struct perf_event_attr *)
541 
542 enum perf_event_ioc_flags {
543 	PERF_IOC_FLAG_GROUP		= 1U << 0,
544 };
545 
546 /*
547  * Structure of the page that can be mapped via mmap
548  */
549 struct perf_event_mmap_page {
550 	__u32	version;		/* version number of this structure */
551 	__u32	compat_version;		/* lowest version this is compat with */
552 
553 	/*
554 	 * Bits needed to read the hw events in user-space.
555 	 *
556 	 *   u32 seq, time_mult, time_shift, index, width;
557 	 *   u64 count, enabled, running;
558 	 *   u64 cyc, time_offset;
559 	 *   s64 pmc = 0;
560 	 *
561 	 *   do {
562 	 *     seq = pc->lock;
563 	 *     barrier()
564 	 *
565 	 *     enabled = pc->time_enabled;
566 	 *     running = pc->time_running;
567 	 *
568 	 *     if (pc->cap_usr_time && enabled != running) {
569 	 *       cyc = rdtsc();
570 	 *       time_offset = pc->time_offset;
571 	 *       time_mult   = pc->time_mult;
572 	 *       time_shift  = pc->time_shift;
573 	 *     }
574 	 *
575 	 *     index = pc->index;
576 	 *     count = pc->offset;
577 	 *     if (pc->cap_user_rdpmc && index) {
578 	 *       width = pc->pmc_width;
579 	 *       pmc = rdpmc(index - 1);
580 	 *     }
581 	 *
582 	 *     barrier();
583 	 *   } while (pc->lock != seq);
584 	 *
585 	 * NOTE: for obvious reason this only works on self-monitoring
586 	 *       processes.
587 	 */
588 	__u32	lock;			/* seqlock for synchronization */
589 	__u32	index;			/* hardware event identifier */
590 	__s64	offset;			/* add to hardware event value */
591 	__u64	time_enabled;		/* time event active */
592 	__u64	time_running;		/* time event on cpu */
593 	union {
594 		__u64	capabilities;
595 		struct {
596 			__u64	cap_bit0		: 1, /* Always 0, deprecated, see commit 860f085b74e9 */
597 				cap_bit0_is_deprecated	: 1, /* Always 1, signals that bit 0 is zero */
598 
599 				cap_user_rdpmc		: 1, /* The RDPMC instruction can be used to read counts */
600 				cap_user_time		: 1, /* The time_{shift,mult,offset} fields are used */
601 				cap_user_time_zero	: 1, /* The time_zero field is used */
602 				cap_user_time_short	: 1, /* the time_{cycle,mask} fields are used */
603 				cap_____res		: 58;
604 		};
605 	};
606 
607 	/*
608 	 * If cap_user_rdpmc this field provides the bit-width of the value
609 	 * read using the rdpmc() or equivalent instruction. This can be used
610 	 * to sign extend the result like:
611 	 *
612 	 *   pmc <<= 64 - width;
613 	 *   pmc >>= 64 - width; // signed shift right
614 	 *   count += pmc;
615 	 */
616 	__u16	pmc_width;
617 
618 	/*
619 	 * If cap_usr_time the below fields can be used to compute the time
620 	 * delta since time_enabled (in ns) using rdtsc or similar.
621 	 *
622 	 *   u64 quot, rem;
623 	 *   u64 delta;
624 	 *
625 	 *   quot = (cyc >> time_shift);
626 	 *   rem = cyc & (((u64)1 << time_shift) - 1);
627 	 *   delta = time_offset + quot * time_mult +
628 	 *              ((rem * time_mult) >> time_shift);
629 	 *
630 	 * Where time_offset,time_mult,time_shift and cyc are read in the
631 	 * seqcount loop described above. This delta can then be added to
632 	 * enabled and possible running (if index), improving the scaling:
633 	 *
634 	 *   enabled += delta;
635 	 *   if (index)
636 	 *     running += delta;
637 	 *
638 	 *   quot = count / running;
639 	 *   rem  = count % running;
640 	 *   count = quot * enabled + (rem * enabled) / running;
641 	 */
642 	__u16	time_shift;
643 	__u32	time_mult;
644 	__u64	time_offset;
645 	/*
646 	 * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated
647 	 * from sample timestamps.
648 	 *
649 	 *   time = timestamp - time_zero;
650 	 *   quot = time / time_mult;
651 	 *   rem  = time % time_mult;
652 	 *   cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
653 	 *
654 	 * And vice versa:
655 	 *
656 	 *   quot = cyc >> time_shift;
657 	 *   rem  = cyc & (((u64)1 << time_shift) - 1);
658 	 *   timestamp = time_zero + quot * time_mult +
659 	 *               ((rem * time_mult) >> time_shift);
660 	 */
661 	__u64	time_zero;
662 
663 	__u32	size;			/* Header size up to __reserved[] fields. */
664 	__u32	__reserved_1;
665 
666 	/*
667 	 * If cap_usr_time_short, the hardware clock is less than 64bit wide
668 	 * and we must compute the 'cyc' value, as used by cap_usr_time, as:
669 	 *
670 	 *   cyc = time_cycles + ((cyc - time_cycles) & time_mask)
671 	 *
672 	 * NOTE: this form is explicitly chosen such that cap_usr_time_short
673 	 *       is a correction on top of cap_usr_time, and code that doesn't
674 	 *       know about cap_usr_time_short still works under the assumption
675 	 *       the counter doesn't wrap.
676 	 */
677 	__u64	time_cycles;
678 	__u64	time_mask;
679 
680 		/*
681 		 * Hole for extension of the self monitor capabilities
682 		 */
683 
684 	__u8	__reserved[116*8];	/* align to 1k. */
685 
686 	/*
687 	 * Control data for the mmap() data buffer.
688 	 *
689 	 * User-space reading the @data_head value should issue an smp_rmb(),
690 	 * after reading this value.
691 	 *
692 	 * When the mapping is PROT_WRITE the @data_tail value should be
693 	 * written by userspace to reflect the last read data, after issueing
694 	 * an smp_mb() to separate the data read from the ->data_tail store.
695 	 * In this case the kernel will not over-write unread data.
696 	 *
697 	 * See perf_output_put_handle() for the data ordering.
698 	 *
699 	 * data_{offset,size} indicate the location and size of the perf record
700 	 * buffer within the mmapped area.
701 	 */
702 	__u64   data_head;		/* head in the data section */
703 	__u64	data_tail;		/* user-space written tail */
704 	__u64	data_offset;		/* where the buffer starts */
705 	__u64	data_size;		/* data buffer size */
706 
707 	/*
708 	 * AUX area is defined by aux_{offset,size} fields that should be set
709 	 * by the userspace, so that
710 	 *
711 	 *   aux_offset >= data_offset + data_size
712 	 *
713 	 * prior to mmap()ing it. Size of the mmap()ed area should be aux_size.
714 	 *
715 	 * Ring buffer pointers aux_{head,tail} have the same semantics as
716 	 * data_{head,tail} and same ordering rules apply.
717 	 */
718 	__u64	aux_head;
719 	__u64	aux_tail;
720 	__u64	aux_offset;
721 	__u64	aux_size;
722 };
723 
724 /*
725  * The current state of perf_event_header::misc bits usage:
726  * ('|' used bit, '-' unused bit)
727  *
728  *  012         CDEF
729  *  |||---------||||
730  *
731  *  Where:
732  *    0-2     CPUMODE_MASK
733  *
734  *    C       PROC_MAP_PARSE_TIMEOUT
735  *    D       MMAP_DATA / COMM_EXEC / FORK_EXEC / SWITCH_OUT
736  *    E       MMAP_BUILD_ID / EXACT_IP / SCHED_OUT_PREEMPT
737  *    F       (reserved)
738  */
739 
740 #define PERF_RECORD_MISC_CPUMODE_MASK		(7 << 0)
741 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN	(0 << 0)
742 #define PERF_RECORD_MISC_KERNEL			(1 << 0)
743 #define PERF_RECORD_MISC_USER			(2 << 0)
744 #define PERF_RECORD_MISC_HYPERVISOR		(3 << 0)
745 #define PERF_RECORD_MISC_GUEST_KERNEL		(4 << 0)
746 #define PERF_RECORD_MISC_GUEST_USER		(5 << 0)
747 
748 /*
749  * Indicates that /proc/PID/maps parsing are truncated by time out.
750  */
751 #define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT	(1 << 12)
752 /*
753  * Following PERF_RECORD_MISC_* are used on different
754  * events, so can reuse the same bit position:
755  *
756  *   PERF_RECORD_MISC_MMAP_DATA  - PERF_RECORD_MMAP* events
757  *   PERF_RECORD_MISC_COMM_EXEC  - PERF_RECORD_COMM event
758  *   PERF_RECORD_MISC_FORK_EXEC  - PERF_RECORD_FORK event (perf internal)
759  *   PERF_RECORD_MISC_SWITCH_OUT - PERF_RECORD_SWITCH* events
760  */
761 #define PERF_RECORD_MISC_MMAP_DATA		(1 << 13)
762 #define PERF_RECORD_MISC_COMM_EXEC		(1 << 13)
763 #define PERF_RECORD_MISC_FORK_EXEC		(1 << 13)
764 #define PERF_RECORD_MISC_SWITCH_OUT		(1 << 13)
765 /*
766  * These PERF_RECORD_MISC_* flags below are safely reused
767  * for the following events:
768  *
769  *   PERF_RECORD_MISC_EXACT_IP           - PERF_RECORD_SAMPLE of precise events
770  *   PERF_RECORD_MISC_SWITCH_OUT_PREEMPT - PERF_RECORD_SWITCH* events
771  *   PERF_RECORD_MISC_MMAP_BUILD_ID      - PERF_RECORD_MMAP2 event
772  *
773  *
774  * PERF_RECORD_MISC_EXACT_IP:
775  *   Indicates that the content of PERF_SAMPLE_IP points to
776  *   the actual instruction that triggered the event. See also
777  *   perf_event_attr::precise_ip.
778  *
779  * PERF_RECORD_MISC_SWITCH_OUT_PREEMPT:
780  *   Indicates that thread was preempted in TASK_RUNNING state.
781  *
782  * PERF_RECORD_MISC_MMAP_BUILD_ID:
783  *   Indicates that mmap2 event carries build id data.
784  */
785 #define PERF_RECORD_MISC_EXACT_IP		(1 << 14)
786 #define PERF_RECORD_MISC_SWITCH_OUT_PREEMPT	(1 << 14)
787 #define PERF_RECORD_MISC_MMAP_BUILD_ID		(1 << 14)
788 /*
789  * Reserve the last bit to indicate some extended misc field
790  */
791 #define PERF_RECORD_MISC_EXT_RESERVED		(1 << 15)
792 
793 struct perf_event_header {
794 	__u32	type;
795 	__u16	misc;
796 	__u16	size;
797 };
798 
799 struct perf_ns_link_info {
800 	__u64	dev;
801 	__u64	ino;
802 };
803 
804 enum {
805 	NET_NS_INDEX		= 0,
806 	UTS_NS_INDEX		= 1,
807 	IPC_NS_INDEX		= 2,
808 	PID_NS_INDEX		= 3,
809 	USER_NS_INDEX		= 4,
810 	MNT_NS_INDEX		= 5,
811 	CGROUP_NS_INDEX		= 6,
812 
813 	NR_NAMESPACES,		/* number of available namespaces */
814 };
815 
816 enum perf_event_type {
817 
818 	/*
819 	 * If perf_event_attr.sample_id_all is set then all event types will
820 	 * have the sample_type selected fields related to where/when
821 	 * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU,
822 	 * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed
823 	 * just after the perf_event_header and the fields already present for
824 	 * the existing fields, i.e. at the end of the payload. That way a newer
825 	 * perf.data file will be supported by older perf tools, with these new
826 	 * optional fields being ignored.
827 	 *
828 	 * struct sample_id {
829 	 * 	{ u32			pid, tid; } && PERF_SAMPLE_TID
830 	 * 	{ u64			time;     } && PERF_SAMPLE_TIME
831 	 * 	{ u64			id;       } && PERF_SAMPLE_ID
832 	 * 	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
833 	 * 	{ u32			cpu, res; } && PERF_SAMPLE_CPU
834 	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
835 	 * } && perf_event_attr::sample_id_all
836 	 *
837 	 * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.  The
838 	 * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed
839 	 * relative to header.size.
840 	 */
841 
842 	/*
843 	 * The MMAP events record the PROT_EXEC mappings so that we can
844 	 * correlate userspace IPs to code. They have the following structure:
845 	 *
846 	 * struct {
847 	 *	struct perf_event_header	header;
848 	 *
849 	 *	u32				pid, tid;
850 	 *	u64				addr;
851 	 *	u64				len;
852 	 *	u64				pgoff;
853 	 *	char				filename[];
854 	 * 	struct sample_id		sample_id;
855 	 * };
856 	 */
857 	PERF_RECORD_MMAP			= 1,
858 
859 	/*
860 	 * struct {
861 	 *	struct perf_event_header	header;
862 	 *	u64				id;
863 	 *	u64				lost;
864 	 * 	struct sample_id		sample_id;
865 	 * };
866 	 */
867 	PERF_RECORD_LOST			= 2,
868 
869 	/*
870 	 * struct {
871 	 *	struct perf_event_header	header;
872 	 *
873 	 *	u32				pid, tid;
874 	 *	char				comm[];
875 	 * 	struct sample_id		sample_id;
876 	 * };
877 	 */
878 	PERF_RECORD_COMM			= 3,
879 
880 	/*
881 	 * struct {
882 	 *	struct perf_event_header	header;
883 	 *	u32				pid, ppid;
884 	 *	u32				tid, ptid;
885 	 *	u64				time;
886 	 * 	struct sample_id		sample_id;
887 	 * };
888 	 */
889 	PERF_RECORD_EXIT			= 4,
890 
891 	/*
892 	 * struct {
893 	 *	struct perf_event_header	header;
894 	 *	u64				time;
895 	 *	u64				id;
896 	 *	u64				stream_id;
897 	 * 	struct sample_id		sample_id;
898 	 * };
899 	 */
900 	PERF_RECORD_THROTTLE			= 5,
901 	PERF_RECORD_UNTHROTTLE			= 6,
902 
903 	/*
904 	 * struct {
905 	 *	struct perf_event_header	header;
906 	 *	u32				pid, ppid;
907 	 *	u32				tid, ptid;
908 	 *	u64				time;
909 	 * 	struct sample_id		sample_id;
910 	 * };
911 	 */
912 	PERF_RECORD_FORK			= 7,
913 
914 	/*
915 	 * struct {
916 	 *	struct perf_event_header	header;
917 	 *	u32				pid, tid;
918 	 *
919 	 *	struct read_format		values;
920 	 * 	struct sample_id		sample_id;
921 	 * };
922 	 */
923 	PERF_RECORD_READ			= 8,
924 
925 	/*
926 	 * struct {
927 	 *	struct perf_event_header	header;
928 	 *
929 	 *	#
930 	 *	# Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.
931 	 *	# The advantage of PERF_SAMPLE_IDENTIFIER is that its position
932 	 *	# is fixed relative to header.
933 	 *	#
934 	 *
935 	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
936 	 *	{ u64			ip;	  } && PERF_SAMPLE_IP
937 	 *	{ u32			pid, tid; } && PERF_SAMPLE_TID
938 	 *	{ u64			time;     } && PERF_SAMPLE_TIME
939 	 *	{ u64			addr;     } && PERF_SAMPLE_ADDR
940 	 *	{ u64			id;	  } && PERF_SAMPLE_ID
941 	 *	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
942 	 *	{ u32			cpu, res; } && PERF_SAMPLE_CPU
943 	 *	{ u64			period;   } && PERF_SAMPLE_PERIOD
944 	 *
945 	 *	{ struct read_format	values;	  } && PERF_SAMPLE_READ
946 	 *
947 	 *	{ u64			nr,
948 	 *	  u64			ips[nr];  } && PERF_SAMPLE_CALLCHAIN
949 	 *
950 	 *	#
951 	 *	# The RAW record below is opaque data wrt the ABI
952 	 *	#
953 	 *	# That is, the ABI doesn't make any promises wrt to
954 	 *	# the stability of its content, it may vary depending
955 	 *	# on event, hardware, kernel version and phase of
956 	 *	# the moon.
957 	 *	#
958 	 *	# In other words, PERF_SAMPLE_RAW contents are not an ABI.
959 	 *	#
960 	 *
961 	 *	{ u32			size;
962 	 *	  char                  data[size];}&& PERF_SAMPLE_RAW
963 	 *
964 	 *	{ u64                   nr;
965 	 *	  { u64	hw_idx; } && PERF_SAMPLE_BRANCH_HW_INDEX
966 	 *        { u64 from, to, flags } lbr[nr];
967 	 *      } && PERF_SAMPLE_BRANCH_STACK
968 	 *
969 	 * 	{ u64			abi; # enum perf_sample_regs_abi
970 	 * 	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
971 	 *
972 	 * 	{ u64			size;
973 	 * 	  char			data[size];
974 	 * 	  u64			dyn_size; } && PERF_SAMPLE_STACK_USER
975 	 *
976 	 *	{ union perf_sample_weight
977 	 *	 {
978 	 *		u64		full; && PERF_SAMPLE_WEIGHT
979 	 *	#if defined(__LITTLE_ENDIAN_BITFIELD)
980 	 *		struct {
981 	 *			u32	var1_dw;
982 	 *			u16	var2_w;
983 	 *			u16	var3_w;
984 	 *		} && PERF_SAMPLE_WEIGHT_STRUCT
985 	 *	#elif defined(__BIG_ENDIAN_BITFIELD)
986 	 *		struct {
987 	 *			u16	var3_w;
988 	 *			u16	var2_w;
989 	 *			u32	var1_dw;
990 	 *		} && PERF_SAMPLE_WEIGHT_STRUCT
991 	 *	#endif
992 	 *	 }
993 	 *	}
994 	 *	{ u64			data_src; } && PERF_SAMPLE_DATA_SRC
995 	 *	{ u64			transaction; } && PERF_SAMPLE_TRANSACTION
996 	 *	{ u64			abi; # enum perf_sample_regs_abi
997 	 *	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
998 	 *	{ u64			phys_addr;} && PERF_SAMPLE_PHYS_ADDR
999 	 *	{ u64			size;
1000 	 *	  char			data[size]; } && PERF_SAMPLE_AUX
1001 	 *	{ u64			data_page_size;} && PERF_SAMPLE_DATA_PAGE_SIZE
1002 	 *	{ u64			code_page_size;} && PERF_SAMPLE_CODE_PAGE_SIZE
1003 	 * };
1004 	 */
1005 	PERF_RECORD_SAMPLE			= 9,
1006 
1007 	/*
1008 	 * The MMAP2 records are an augmented version of MMAP, they add
1009 	 * maj, min, ino numbers to be used to uniquely identify each mapping
1010 	 *
1011 	 * struct {
1012 	 *	struct perf_event_header	header;
1013 	 *
1014 	 *	u32				pid, tid;
1015 	 *	u64				addr;
1016 	 *	u64				len;
1017 	 *	u64				pgoff;
1018 	 *	union {
1019 	 *		struct {
1020 	 *			u32		maj;
1021 	 *			u32		min;
1022 	 *			u64		ino;
1023 	 *			u64		ino_generation;
1024 	 *		};
1025 	 *		struct {
1026 	 *			u8		build_id_size;
1027 	 *			u8		__reserved_1;
1028 	 *			u16		__reserved_2;
1029 	 *			u8		build_id[20];
1030 	 *		};
1031 	 *	};
1032 	 *	u32				prot, flags;
1033 	 *	char				filename[];
1034 	 * 	struct sample_id		sample_id;
1035 	 * };
1036 	 */
1037 	PERF_RECORD_MMAP2			= 10,
1038 
1039 	/*
1040 	 * Records that new data landed in the AUX buffer part.
1041 	 *
1042 	 * struct {
1043 	 * 	struct perf_event_header	header;
1044 	 *
1045 	 * 	u64				aux_offset;
1046 	 * 	u64				aux_size;
1047 	 *	u64				flags;
1048 	 * 	struct sample_id		sample_id;
1049 	 * };
1050 	 */
1051 	PERF_RECORD_AUX				= 11,
1052 
1053 	/*
1054 	 * Indicates that instruction trace has started
1055 	 *
1056 	 * struct {
1057 	 *	struct perf_event_header	header;
1058 	 *	u32				pid;
1059 	 *	u32				tid;
1060 	 *	struct sample_id		sample_id;
1061 	 * };
1062 	 */
1063 	PERF_RECORD_ITRACE_START		= 12,
1064 
1065 	/*
1066 	 * Records the dropped/lost sample number.
1067 	 *
1068 	 * struct {
1069 	 *	struct perf_event_header	header;
1070 	 *
1071 	 *	u64				lost;
1072 	 *	struct sample_id		sample_id;
1073 	 * };
1074 	 */
1075 	PERF_RECORD_LOST_SAMPLES		= 13,
1076 
1077 	/*
1078 	 * Records a context switch in or out (flagged by
1079 	 * PERF_RECORD_MISC_SWITCH_OUT). See also
1080 	 * PERF_RECORD_SWITCH_CPU_WIDE.
1081 	 *
1082 	 * struct {
1083 	 *	struct perf_event_header	header;
1084 	 *	struct sample_id		sample_id;
1085 	 * };
1086 	 */
1087 	PERF_RECORD_SWITCH			= 14,
1088 
1089 	/*
1090 	 * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and
1091 	 * next_prev_tid that are the next (switching out) or previous
1092 	 * (switching in) pid/tid.
1093 	 *
1094 	 * struct {
1095 	 *	struct perf_event_header	header;
1096 	 *	u32				next_prev_pid;
1097 	 *	u32				next_prev_tid;
1098 	 *	struct sample_id		sample_id;
1099 	 * };
1100 	 */
1101 	PERF_RECORD_SWITCH_CPU_WIDE		= 15,
1102 
1103 	/*
1104 	 * struct {
1105 	 *	struct perf_event_header	header;
1106 	 *	u32				pid;
1107 	 *	u32				tid;
1108 	 *	u64				nr_namespaces;
1109 	 *	{ u64				dev, inode; } [nr_namespaces];
1110 	 *	struct sample_id		sample_id;
1111 	 * };
1112 	 */
1113 	PERF_RECORD_NAMESPACES			= 16,
1114 
1115 	/*
1116 	 * Record ksymbol register/unregister events:
1117 	 *
1118 	 * struct {
1119 	 *	struct perf_event_header	header;
1120 	 *	u64				addr;
1121 	 *	u32				len;
1122 	 *	u16				ksym_type;
1123 	 *	u16				flags;
1124 	 *	char				name[];
1125 	 *	struct sample_id		sample_id;
1126 	 * };
1127 	 */
1128 	PERF_RECORD_KSYMBOL			= 17,
1129 
1130 	/*
1131 	 * Record bpf events:
1132 	 *  enum perf_bpf_event_type {
1133 	 *	PERF_BPF_EVENT_UNKNOWN		= 0,
1134 	 *	PERF_BPF_EVENT_PROG_LOAD	= 1,
1135 	 *	PERF_BPF_EVENT_PROG_UNLOAD	= 2,
1136 	 *  };
1137 	 *
1138 	 * struct {
1139 	 *	struct perf_event_header	header;
1140 	 *	u16				type;
1141 	 *	u16				flags;
1142 	 *	u32				id;
1143 	 *	u8				tag[BPF_TAG_SIZE];
1144 	 *	struct sample_id		sample_id;
1145 	 * };
1146 	 */
1147 	PERF_RECORD_BPF_EVENT			= 18,
1148 
1149 	/*
1150 	 * struct {
1151 	 *	struct perf_event_header	header;
1152 	 *	u64				id;
1153 	 *	char				path[];
1154 	 *	struct sample_id		sample_id;
1155 	 * };
1156 	 */
1157 	PERF_RECORD_CGROUP			= 19,
1158 
1159 	/*
1160 	 * Records changes to kernel text i.e. self-modified code. 'old_len' is
1161 	 * the number of old bytes, 'new_len' is the number of new bytes. Either
1162 	 * 'old_len' or 'new_len' may be zero to indicate, for example, the
1163 	 * addition or removal of a trampoline. 'bytes' contains the old bytes
1164 	 * followed immediately by the new bytes.
1165 	 *
1166 	 * struct {
1167 	 *	struct perf_event_header	header;
1168 	 *	u64				addr;
1169 	 *	u16				old_len;
1170 	 *	u16				new_len;
1171 	 *	u8				bytes[];
1172 	 *	struct sample_id		sample_id;
1173 	 * };
1174 	 */
1175 	PERF_RECORD_TEXT_POKE			= 20,
1176 
1177 	/*
1178 	 * Data written to the AUX area by hardware due to aux_output, may need
1179 	 * to be matched to the event by an architecture-specific hardware ID.
1180 	 * This records the hardware ID, but requires sample_id to provide the
1181 	 * event ID. e.g. Intel PT uses this record to disambiguate PEBS-via-PT
1182 	 * records from multiple events.
1183 	 *
1184 	 * struct {
1185 	 *	struct perf_event_header	header;
1186 	 *	u64				hw_id;
1187 	 *	struct sample_id		sample_id;
1188 	 * };
1189 	 */
1190 	PERF_RECORD_AUX_OUTPUT_HW_ID		= 21,
1191 
1192 	PERF_RECORD_MAX,			/* non-ABI */
1193 };
1194 
1195 enum perf_record_ksymbol_type {
1196 	PERF_RECORD_KSYMBOL_TYPE_UNKNOWN	= 0,
1197 	PERF_RECORD_KSYMBOL_TYPE_BPF		= 1,
1198 	/*
1199 	 * Out of line code such as kprobe-replaced instructions or optimized
1200 	 * kprobes or ftrace trampolines.
1201 	 */
1202 	PERF_RECORD_KSYMBOL_TYPE_OOL		= 2,
1203 	PERF_RECORD_KSYMBOL_TYPE_MAX		/* non-ABI */
1204 };
1205 
1206 #define PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER	(1 << 0)
1207 
1208 enum perf_bpf_event_type {
1209 	PERF_BPF_EVENT_UNKNOWN		= 0,
1210 	PERF_BPF_EVENT_PROG_LOAD	= 1,
1211 	PERF_BPF_EVENT_PROG_UNLOAD	= 2,
1212 	PERF_BPF_EVENT_MAX,		/* non-ABI */
1213 };
1214 
1215 #define PERF_MAX_STACK_DEPTH		127
1216 #define PERF_MAX_CONTEXTS_PER_STACK	  8
1217 
1218 enum perf_callchain_context {
1219 	PERF_CONTEXT_HV			= (__u64)-32,
1220 	PERF_CONTEXT_KERNEL		= (__u64)-128,
1221 	PERF_CONTEXT_USER		= (__u64)-512,
1222 
1223 	PERF_CONTEXT_GUEST		= (__u64)-2048,
1224 	PERF_CONTEXT_GUEST_KERNEL	= (__u64)-2176,
1225 	PERF_CONTEXT_GUEST_USER		= (__u64)-2560,
1226 
1227 	PERF_CONTEXT_MAX		= (__u64)-4095,
1228 };
1229 
1230 /**
1231  * PERF_RECORD_AUX::flags bits
1232  */
1233 #define PERF_AUX_FLAG_TRUNCATED			0x01	/* record was truncated to fit */
1234 #define PERF_AUX_FLAG_OVERWRITE			0x02	/* snapshot from overwrite mode */
1235 #define PERF_AUX_FLAG_PARTIAL			0x04	/* record contains gaps */
1236 #define PERF_AUX_FLAG_COLLISION			0x08	/* sample collided with another */
1237 #define PERF_AUX_FLAG_PMU_FORMAT_TYPE_MASK	0xff00	/* PMU specific trace format type */
1238 
1239 /* CoreSight PMU AUX buffer formats */
1240 #define PERF_AUX_FLAG_CORESIGHT_FORMAT_CORESIGHT	0x0000 /* Default for backward compatibility */
1241 #define PERF_AUX_FLAG_CORESIGHT_FORMAT_RAW		0x0100 /* Raw format of the source */
1242 
1243 #define PERF_FLAG_FD_NO_GROUP		(1UL << 0)
1244 #define PERF_FLAG_FD_OUTPUT		(1UL << 1)
1245 #define PERF_FLAG_PID_CGROUP		(1UL << 2) /* pid=cgroup id, per-cpu mode only */
1246 #define PERF_FLAG_FD_CLOEXEC		(1UL << 3) /* O_CLOEXEC */
1247 
1248 #if defined(__LITTLE_ENDIAN_BITFIELD)
1249 union perf_mem_data_src {
1250 	__u64 val;
1251 	struct {
1252 		__u64   mem_op:5,	/* type of opcode */
1253 			mem_lvl:14,	/* memory hierarchy level */
1254 			mem_snoop:5,	/* snoop mode */
1255 			mem_lock:2,	/* lock instr */
1256 			mem_dtlb:7,	/* tlb access */
1257 			mem_lvl_num:4,	/* memory hierarchy level number */
1258 			mem_remote:1,   /* remote */
1259 			mem_snoopx:2,	/* snoop mode, ext */
1260 			mem_blk:3,	/* access blocked */
1261 			mem_hops:3,	/* hop level */
1262 			mem_rsvd:18;
1263 	};
1264 };
1265 #elif defined(__BIG_ENDIAN_BITFIELD)
1266 union perf_mem_data_src {
1267 	__u64 val;
1268 	struct {
1269 		__u64	mem_rsvd:18,
1270 			mem_hops:3,	/* hop level */
1271 			mem_blk:3,	/* access blocked */
1272 			mem_snoopx:2,	/* snoop mode, ext */
1273 			mem_remote:1,   /* remote */
1274 			mem_lvl_num:4,	/* memory hierarchy level number */
1275 			mem_dtlb:7,	/* tlb access */
1276 			mem_lock:2,	/* lock instr */
1277 			mem_snoop:5,	/* snoop mode */
1278 			mem_lvl:14,	/* memory hierarchy level */
1279 			mem_op:5;	/* type of opcode */
1280 	};
1281 };
1282 #else
1283 #error "Unknown endianness"
1284 #endif
1285 
1286 /* type of opcode (load/store/prefetch,code) */
1287 #define PERF_MEM_OP_NA		0x01 /* not available */
1288 #define PERF_MEM_OP_LOAD	0x02 /* load instruction */
1289 #define PERF_MEM_OP_STORE	0x04 /* store instruction */
1290 #define PERF_MEM_OP_PFETCH	0x08 /* prefetch */
1291 #define PERF_MEM_OP_EXEC	0x10 /* code (execution) */
1292 #define PERF_MEM_OP_SHIFT	0
1293 
1294 /*
1295  * PERF_MEM_LVL_* namespace being depricated to some extent in the
1296  * favour of newer composite PERF_MEM_{LVLNUM_,REMOTE_,SNOOPX_} fields.
1297  * Supporting this namespace inorder to not break defined ABIs.
1298  *
1299  * memory hierarchy (memory level, hit or miss)
1300  */
1301 #define PERF_MEM_LVL_NA		0x01  /* not available */
1302 #define PERF_MEM_LVL_HIT	0x02  /* hit level */
1303 #define PERF_MEM_LVL_MISS	0x04  /* miss level  */
1304 #define PERF_MEM_LVL_L1		0x08  /* L1 */
1305 #define PERF_MEM_LVL_LFB	0x10  /* Line Fill Buffer */
1306 #define PERF_MEM_LVL_L2		0x20  /* L2 */
1307 #define PERF_MEM_LVL_L3		0x40  /* L3 */
1308 #define PERF_MEM_LVL_LOC_RAM	0x80  /* Local DRAM */
1309 #define PERF_MEM_LVL_REM_RAM1	0x100 /* Remote DRAM (1 hop) */
1310 #define PERF_MEM_LVL_REM_RAM2	0x200 /* Remote DRAM (2 hops) */
1311 #define PERF_MEM_LVL_REM_CCE1	0x400 /* Remote Cache (1 hop) */
1312 #define PERF_MEM_LVL_REM_CCE2	0x800 /* Remote Cache (2 hops) */
1313 #define PERF_MEM_LVL_IO		0x1000 /* I/O memory */
1314 #define PERF_MEM_LVL_UNC	0x2000 /* Uncached memory */
1315 #define PERF_MEM_LVL_SHIFT	5
1316 
1317 #define PERF_MEM_REMOTE_REMOTE	0x01  /* Remote */
1318 #define PERF_MEM_REMOTE_SHIFT	37
1319 
1320 #define PERF_MEM_LVLNUM_L1	0x01 /* L1 */
1321 #define PERF_MEM_LVLNUM_L2	0x02 /* L2 */
1322 #define PERF_MEM_LVLNUM_L3	0x03 /* L3 */
1323 #define PERF_MEM_LVLNUM_L4	0x04 /* L4 */
1324 /* 5-0xa available */
1325 #define PERF_MEM_LVLNUM_ANY_CACHE 0x0b /* Any cache */
1326 #define PERF_MEM_LVLNUM_LFB	0x0c /* LFB */
1327 #define PERF_MEM_LVLNUM_RAM	0x0d /* RAM */
1328 #define PERF_MEM_LVLNUM_PMEM	0x0e /* PMEM */
1329 #define PERF_MEM_LVLNUM_NA	0x0f /* N/A */
1330 
1331 #define PERF_MEM_LVLNUM_SHIFT	33
1332 
1333 /* snoop mode */
1334 #define PERF_MEM_SNOOP_NA	0x01 /* not available */
1335 #define PERF_MEM_SNOOP_NONE	0x02 /* no snoop */
1336 #define PERF_MEM_SNOOP_HIT	0x04 /* snoop hit */
1337 #define PERF_MEM_SNOOP_MISS	0x08 /* snoop miss */
1338 #define PERF_MEM_SNOOP_HITM	0x10 /* snoop hit modified */
1339 #define PERF_MEM_SNOOP_SHIFT	19
1340 
1341 #define PERF_MEM_SNOOPX_FWD	0x01 /* forward */
1342 /* 1 free */
1343 #define PERF_MEM_SNOOPX_SHIFT  38
1344 
1345 /* locked instruction */
1346 #define PERF_MEM_LOCK_NA	0x01 /* not available */
1347 #define PERF_MEM_LOCK_LOCKED	0x02 /* locked transaction */
1348 #define PERF_MEM_LOCK_SHIFT	24
1349 
1350 /* TLB access */
1351 #define PERF_MEM_TLB_NA		0x01 /* not available */
1352 #define PERF_MEM_TLB_HIT	0x02 /* hit level */
1353 #define PERF_MEM_TLB_MISS	0x04 /* miss level */
1354 #define PERF_MEM_TLB_L1		0x08 /* L1 */
1355 #define PERF_MEM_TLB_L2		0x10 /* L2 */
1356 #define PERF_MEM_TLB_WK		0x20 /* Hardware Walker*/
1357 #define PERF_MEM_TLB_OS		0x40 /* OS fault handler */
1358 #define PERF_MEM_TLB_SHIFT	26
1359 
1360 /* Access blocked */
1361 #define PERF_MEM_BLK_NA		0x01 /* not available */
1362 #define PERF_MEM_BLK_DATA	0x02 /* data could not be forwarded */
1363 #define PERF_MEM_BLK_ADDR	0x04 /* address conflict */
1364 #define PERF_MEM_BLK_SHIFT	40
1365 
1366 /* hop level */
1367 #define PERF_MEM_HOPS_0		0x01 /* remote core, same node */
1368 #define PERF_MEM_HOPS_1		0x02 /* remote node, same socket */
1369 #define PERF_MEM_HOPS_2		0x03 /* remote socket, same board */
1370 #define PERF_MEM_HOPS_3		0x04 /* remote board */
1371 /* 5-7 available */
1372 #define PERF_MEM_HOPS_SHIFT	43
1373 
1374 #define PERF_MEM_S(a, s) \
1375 	(((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)
1376 
1377 /*
1378  * single taken branch record layout:
1379  *
1380  *      from: source instruction (may not always be a branch insn)
1381  *        to: branch target
1382  *   mispred: branch target was mispredicted
1383  * predicted: branch target was predicted
1384  *
1385  * support for mispred, predicted is optional. In case it
1386  * is not supported mispred = predicted = 0.
1387  *
1388  *     in_tx: running in a hardware transaction
1389  *     abort: aborting a hardware transaction
1390  *    cycles: cycles from last branch (or 0 if not supported)
1391  *      type: branch type
1392  *      spec: branch speculation info (or 0 if not supported)
1393  */
1394 struct perf_branch_entry {
1395 	__u64	from;
1396 	__u64	to;
1397 	__u64	mispred:1,  /* target mispredicted */
1398 		predicted:1,/* target predicted */
1399 		in_tx:1,    /* in transaction */
1400 		abort:1,    /* transaction abort */
1401 		cycles:16,  /* cycle count to last branch */
1402 		type:4,     /* branch type */
1403 		spec:2,     /* branch speculation info */
1404 		new_type:4, /* additional branch type */
1405 		reserved:34;
1406 };
1407 
1408 union perf_sample_weight {
1409 	__u64		full;
1410 #if defined(__LITTLE_ENDIAN_BITFIELD)
1411 	struct {
1412 		__u32	var1_dw;
1413 		__u16	var2_w;
1414 		__u16	var3_w;
1415 	};
1416 #elif defined(__BIG_ENDIAN_BITFIELD)
1417 	struct {
1418 		__u16	var3_w;
1419 		__u16	var2_w;
1420 		__u32	var1_dw;
1421 	};
1422 #else
1423 #error "Unknown endianness"
1424 #endif
1425 };
1426 
1427 #endif /* _UAPI_LINUX_PERF_EVENT_H */
1428