xref: /linux-6.15/include/uapi/linux/perf_event.h (revision 01eadc8d)
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 				__reserved_1   : 30;
390 
391 	union {
392 		__u32		wakeup_events;	  /* wakeup every n events */
393 		__u32		wakeup_watermark; /* bytes before wakeup   */
394 	};
395 
396 	__u32			bp_type;
397 	union {
398 		__u64		bp_addr;
399 		__u64		kprobe_func; /* for perf_kprobe */
400 		__u64		uprobe_path; /* for perf_uprobe */
401 		__u64		config1; /* extension of config */
402 	};
403 	union {
404 		__u64		bp_len;
405 		__u64		kprobe_addr; /* when kprobe_func == NULL */
406 		__u64		probe_offset; /* for perf_[k,u]probe */
407 		__u64		config2; /* extension of config1 */
408 	};
409 	__u64	branch_sample_type; /* enum perf_branch_sample_type */
410 
411 	/*
412 	 * Defines set of user regs to dump on samples.
413 	 * See asm/perf_regs.h for details.
414 	 */
415 	__u64	sample_regs_user;
416 
417 	/*
418 	 * Defines size of the user stack to dump on samples.
419 	 */
420 	__u32	sample_stack_user;
421 
422 	__s32	clockid;
423 	/*
424 	 * Defines set of regs to dump for each sample
425 	 * state captured on:
426 	 *  - precise = 0: PMU interrupt
427 	 *  - precise > 0: sampled instruction
428 	 *
429 	 * See asm/perf_regs.h for details.
430 	 */
431 	__u64	sample_regs_intr;
432 
433 	/*
434 	 * Wakeup watermark for AUX area
435 	 */
436 	__u32	aux_watermark;
437 	__u16	sample_max_stack;
438 	__u16	__reserved_2;
439 	__u32	aux_sample_size;
440 	__u32	__reserved_3;
441 };
442 
443 /*
444  * Structure used by below PERF_EVENT_IOC_QUERY_BPF command
445  * to query bpf programs attached to the same perf tracepoint
446  * as the given perf event.
447  */
448 struct perf_event_query_bpf {
449 	/*
450 	 * The below ids array length
451 	 */
452 	__u32	ids_len;
453 	/*
454 	 * Set by the kernel to indicate the number of
455 	 * available programs
456 	 */
457 	__u32	prog_cnt;
458 	/*
459 	 * User provided buffer to store program ids
460 	 */
461 	__u32	ids[0];
462 };
463 
464 /*
465  * Ioctls that can be done on a perf event fd:
466  */
467 #define PERF_EVENT_IOC_ENABLE			_IO ('$', 0)
468 #define PERF_EVENT_IOC_DISABLE			_IO ('$', 1)
469 #define PERF_EVENT_IOC_REFRESH			_IO ('$', 2)
470 #define PERF_EVENT_IOC_RESET			_IO ('$', 3)
471 #define PERF_EVENT_IOC_PERIOD			_IOW('$', 4, __u64)
472 #define PERF_EVENT_IOC_SET_OUTPUT		_IO ('$', 5)
473 #define PERF_EVENT_IOC_SET_FILTER		_IOW('$', 6, char *)
474 #define PERF_EVENT_IOC_ID			_IOR('$', 7, __u64 *)
475 #define PERF_EVENT_IOC_SET_BPF			_IOW('$', 8, __u32)
476 #define PERF_EVENT_IOC_PAUSE_OUTPUT		_IOW('$', 9, __u32)
477 #define PERF_EVENT_IOC_QUERY_BPF		_IOWR('$', 10, struct perf_event_query_bpf *)
478 #define PERF_EVENT_IOC_MODIFY_ATTRIBUTES	_IOW('$', 11, struct perf_event_attr *)
479 
480 enum perf_event_ioc_flags {
481 	PERF_IOC_FLAG_GROUP		= 1U << 0,
482 };
483 
484 /*
485  * Structure of the page that can be mapped via mmap
486  */
487 struct perf_event_mmap_page {
488 	__u32	version;		/* version number of this structure */
489 	__u32	compat_version;		/* lowest version this is compat with */
490 
491 	/*
492 	 * Bits needed to read the hw events in user-space.
493 	 *
494 	 *   u32 seq, time_mult, time_shift, index, width;
495 	 *   u64 count, enabled, running;
496 	 *   u64 cyc, time_offset;
497 	 *   s64 pmc = 0;
498 	 *
499 	 *   do {
500 	 *     seq = pc->lock;
501 	 *     barrier()
502 	 *
503 	 *     enabled = pc->time_enabled;
504 	 *     running = pc->time_running;
505 	 *
506 	 *     if (pc->cap_usr_time && enabled != running) {
507 	 *       cyc = rdtsc();
508 	 *       time_offset = pc->time_offset;
509 	 *       time_mult   = pc->time_mult;
510 	 *       time_shift  = pc->time_shift;
511 	 *     }
512 	 *
513 	 *     index = pc->index;
514 	 *     count = pc->offset;
515 	 *     if (pc->cap_user_rdpmc && index) {
516 	 *       width = pc->pmc_width;
517 	 *       pmc = rdpmc(index - 1);
518 	 *     }
519 	 *
520 	 *     barrier();
521 	 *   } while (pc->lock != seq);
522 	 *
523 	 * NOTE: for obvious reason this only works on self-monitoring
524 	 *       processes.
525 	 */
526 	__u32	lock;			/* seqlock for synchronization */
527 	__u32	index;			/* hardware event identifier */
528 	__s64	offset;			/* add to hardware event value */
529 	__u64	time_enabled;		/* time event active */
530 	__u64	time_running;		/* time event on cpu */
531 	union {
532 		__u64	capabilities;
533 		struct {
534 			__u64	cap_bit0		: 1, /* Always 0, deprecated, see commit 860f085b74e9 */
535 				cap_bit0_is_deprecated	: 1, /* Always 1, signals that bit 0 is zero */
536 
537 				cap_user_rdpmc		: 1, /* The RDPMC instruction can be used to read counts */
538 				cap_user_time		: 1, /* The time_{shift,mult,offset} fields are used */
539 				cap_user_time_zero	: 1, /* The time_zero field is used */
540 				cap_user_time_short	: 1, /* the time_{cycle,mask} fields are used */
541 				cap_____res		: 58;
542 		};
543 	};
544 
545 	/*
546 	 * If cap_user_rdpmc this field provides the bit-width of the value
547 	 * read using the rdpmc() or equivalent instruction. This can be used
548 	 * to sign extend the result like:
549 	 *
550 	 *   pmc <<= 64 - width;
551 	 *   pmc >>= 64 - width; // signed shift right
552 	 *   count += pmc;
553 	 */
554 	__u16	pmc_width;
555 
556 	/*
557 	 * If cap_usr_time the below fields can be used to compute the time
558 	 * delta since time_enabled (in ns) using rdtsc or similar.
559 	 *
560 	 *   u64 quot, rem;
561 	 *   u64 delta;
562 	 *
563 	 *   quot = (cyc >> time_shift);
564 	 *   rem = cyc & (((u64)1 << time_shift) - 1);
565 	 *   delta = time_offset + quot * time_mult +
566 	 *              ((rem * time_mult) >> time_shift);
567 	 *
568 	 * Where time_offset,time_mult,time_shift and cyc are read in the
569 	 * seqcount loop described above. This delta can then be added to
570 	 * enabled and possible running (if index), improving the scaling:
571 	 *
572 	 *   enabled += delta;
573 	 *   if (index)
574 	 *     running += delta;
575 	 *
576 	 *   quot = count / running;
577 	 *   rem  = count % running;
578 	 *   count = quot * enabled + (rem * enabled) / running;
579 	 */
580 	__u16	time_shift;
581 	__u32	time_mult;
582 	__u64	time_offset;
583 	/*
584 	 * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated
585 	 * from sample timestamps.
586 	 *
587 	 *   time = timestamp - time_zero;
588 	 *   quot = time / time_mult;
589 	 *   rem  = time % time_mult;
590 	 *   cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
591 	 *
592 	 * And vice versa:
593 	 *
594 	 *   quot = cyc >> time_shift;
595 	 *   rem  = cyc & (((u64)1 << time_shift) - 1);
596 	 *   timestamp = time_zero + quot * time_mult +
597 	 *               ((rem * time_mult) >> time_shift);
598 	 */
599 	__u64	time_zero;
600 
601 	__u32	size;			/* Header size up to __reserved[] fields. */
602 	__u32	__reserved_1;
603 
604 	/*
605 	 * If cap_usr_time_short, the hardware clock is less than 64bit wide
606 	 * and we must compute the 'cyc' value, as used by cap_usr_time, as:
607 	 *
608 	 *   cyc = time_cycles + ((cyc - time_cycles) & time_mask)
609 	 *
610 	 * NOTE: this form is explicitly chosen such that cap_usr_time_short
611 	 *       is a correction on top of cap_usr_time, and code that doesn't
612 	 *       know about cap_usr_time_short still works under the assumption
613 	 *       the counter doesn't wrap.
614 	 */
615 	__u64	time_cycles;
616 	__u64	time_mask;
617 
618 		/*
619 		 * Hole for extension of the self monitor capabilities
620 		 */
621 
622 	__u8	__reserved[116*8];	/* align to 1k. */
623 
624 	/*
625 	 * Control data for the mmap() data buffer.
626 	 *
627 	 * User-space reading the @data_head value should issue an smp_rmb(),
628 	 * after reading this value.
629 	 *
630 	 * When the mapping is PROT_WRITE the @data_tail value should be
631 	 * written by userspace to reflect the last read data, after issueing
632 	 * an smp_mb() to separate the data read from the ->data_tail store.
633 	 * In this case the kernel will not over-write unread data.
634 	 *
635 	 * See perf_output_put_handle() for the data ordering.
636 	 *
637 	 * data_{offset,size} indicate the location and size of the perf record
638 	 * buffer within the mmapped area.
639 	 */
640 	__u64   data_head;		/* head in the data section */
641 	__u64	data_tail;		/* user-space written tail */
642 	__u64	data_offset;		/* where the buffer starts */
643 	__u64	data_size;		/* data buffer size */
644 
645 	/*
646 	 * AUX area is defined by aux_{offset,size} fields that should be set
647 	 * by the userspace, so that
648 	 *
649 	 *   aux_offset >= data_offset + data_size
650 	 *
651 	 * prior to mmap()ing it. Size of the mmap()ed area should be aux_size.
652 	 *
653 	 * Ring buffer pointers aux_{head,tail} have the same semantics as
654 	 * data_{head,tail} and same ordering rules apply.
655 	 */
656 	__u64	aux_head;
657 	__u64	aux_tail;
658 	__u64	aux_offset;
659 	__u64	aux_size;
660 };
661 
662 #define PERF_RECORD_MISC_CPUMODE_MASK		(7 << 0)
663 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN	(0 << 0)
664 #define PERF_RECORD_MISC_KERNEL			(1 << 0)
665 #define PERF_RECORD_MISC_USER			(2 << 0)
666 #define PERF_RECORD_MISC_HYPERVISOR		(3 << 0)
667 #define PERF_RECORD_MISC_GUEST_KERNEL		(4 << 0)
668 #define PERF_RECORD_MISC_GUEST_USER		(5 << 0)
669 
670 /*
671  * Indicates that /proc/PID/maps parsing are truncated by time out.
672  */
673 #define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT	(1 << 12)
674 /*
675  * Following PERF_RECORD_MISC_* are used on different
676  * events, so can reuse the same bit position:
677  *
678  *   PERF_RECORD_MISC_MMAP_DATA  - PERF_RECORD_MMAP* events
679  *   PERF_RECORD_MISC_COMM_EXEC  - PERF_RECORD_COMM event
680  *   PERF_RECORD_MISC_FORK_EXEC  - PERF_RECORD_FORK event (perf internal)
681  *   PERF_RECORD_MISC_SWITCH_OUT - PERF_RECORD_SWITCH* events
682  */
683 #define PERF_RECORD_MISC_MMAP_DATA		(1 << 13)
684 #define PERF_RECORD_MISC_COMM_EXEC		(1 << 13)
685 #define PERF_RECORD_MISC_FORK_EXEC		(1 << 13)
686 #define PERF_RECORD_MISC_SWITCH_OUT		(1 << 13)
687 /*
688  * These PERF_RECORD_MISC_* flags below are safely reused
689  * for the following events:
690  *
691  *   PERF_RECORD_MISC_EXACT_IP           - PERF_RECORD_SAMPLE of precise events
692  *   PERF_RECORD_MISC_SWITCH_OUT_PREEMPT - PERF_RECORD_SWITCH* events
693  *
694  *
695  * PERF_RECORD_MISC_EXACT_IP:
696  *   Indicates that the content of PERF_SAMPLE_IP points to
697  *   the actual instruction that triggered the event. See also
698  *   perf_event_attr::precise_ip.
699  *
700  * PERF_RECORD_MISC_SWITCH_OUT_PREEMPT:
701  *   Indicates that thread was preempted in TASK_RUNNING state.
702  */
703 #define PERF_RECORD_MISC_EXACT_IP		(1 << 14)
704 #define PERF_RECORD_MISC_SWITCH_OUT_PREEMPT	(1 << 14)
705 /*
706  * Reserve the last bit to indicate some extended misc field
707  */
708 #define PERF_RECORD_MISC_EXT_RESERVED		(1 << 15)
709 
710 struct perf_event_header {
711 	__u32	type;
712 	__u16	misc;
713 	__u16	size;
714 };
715 
716 struct perf_ns_link_info {
717 	__u64	dev;
718 	__u64	ino;
719 };
720 
721 enum {
722 	NET_NS_INDEX		= 0,
723 	UTS_NS_INDEX		= 1,
724 	IPC_NS_INDEX		= 2,
725 	PID_NS_INDEX		= 3,
726 	USER_NS_INDEX		= 4,
727 	MNT_NS_INDEX		= 5,
728 	CGROUP_NS_INDEX		= 6,
729 
730 	NR_NAMESPACES,		/* number of available namespaces */
731 };
732 
733 enum perf_event_type {
734 
735 	/*
736 	 * If perf_event_attr.sample_id_all is set then all event types will
737 	 * have the sample_type selected fields related to where/when
738 	 * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU,
739 	 * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed
740 	 * just after the perf_event_header and the fields already present for
741 	 * the existing fields, i.e. at the end of the payload. That way a newer
742 	 * perf.data file will be supported by older perf tools, with these new
743 	 * optional fields being ignored.
744 	 *
745 	 * struct sample_id {
746 	 * 	{ u32			pid, tid; } && PERF_SAMPLE_TID
747 	 * 	{ u64			time;     } && PERF_SAMPLE_TIME
748 	 * 	{ u64			id;       } && PERF_SAMPLE_ID
749 	 * 	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
750 	 * 	{ u32			cpu, res; } && PERF_SAMPLE_CPU
751 	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
752 	 * } && perf_event_attr::sample_id_all
753 	 *
754 	 * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.  The
755 	 * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed
756 	 * relative to header.size.
757 	 */
758 
759 	/*
760 	 * The MMAP events record the PROT_EXEC mappings so that we can
761 	 * correlate userspace IPs to code. They have the following structure:
762 	 *
763 	 * struct {
764 	 *	struct perf_event_header	header;
765 	 *
766 	 *	u32				pid, tid;
767 	 *	u64				addr;
768 	 *	u64				len;
769 	 *	u64				pgoff;
770 	 *	char				filename[];
771 	 * 	struct sample_id		sample_id;
772 	 * };
773 	 */
774 	PERF_RECORD_MMAP			= 1,
775 
776 	/*
777 	 * struct {
778 	 *	struct perf_event_header	header;
779 	 *	u64				id;
780 	 *	u64				lost;
781 	 * 	struct sample_id		sample_id;
782 	 * };
783 	 */
784 	PERF_RECORD_LOST			= 2,
785 
786 	/*
787 	 * struct {
788 	 *	struct perf_event_header	header;
789 	 *
790 	 *	u32				pid, tid;
791 	 *	char				comm[];
792 	 * 	struct sample_id		sample_id;
793 	 * };
794 	 */
795 	PERF_RECORD_COMM			= 3,
796 
797 	/*
798 	 * struct {
799 	 *	struct perf_event_header	header;
800 	 *	u32				pid, ppid;
801 	 *	u32				tid, ptid;
802 	 *	u64				time;
803 	 * 	struct sample_id		sample_id;
804 	 * };
805 	 */
806 	PERF_RECORD_EXIT			= 4,
807 
808 	/*
809 	 * struct {
810 	 *	struct perf_event_header	header;
811 	 *	u64				time;
812 	 *	u64				id;
813 	 *	u64				stream_id;
814 	 * 	struct sample_id		sample_id;
815 	 * };
816 	 */
817 	PERF_RECORD_THROTTLE			= 5,
818 	PERF_RECORD_UNTHROTTLE			= 6,
819 
820 	/*
821 	 * struct {
822 	 *	struct perf_event_header	header;
823 	 *	u32				pid, ppid;
824 	 *	u32				tid, ptid;
825 	 *	u64				time;
826 	 * 	struct sample_id		sample_id;
827 	 * };
828 	 */
829 	PERF_RECORD_FORK			= 7,
830 
831 	/*
832 	 * struct {
833 	 *	struct perf_event_header	header;
834 	 *	u32				pid, tid;
835 	 *
836 	 *	struct read_format		values;
837 	 * 	struct sample_id		sample_id;
838 	 * };
839 	 */
840 	PERF_RECORD_READ			= 8,
841 
842 	/*
843 	 * struct {
844 	 *	struct perf_event_header	header;
845 	 *
846 	 *	#
847 	 *	# Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.
848 	 *	# The advantage of PERF_SAMPLE_IDENTIFIER is that its position
849 	 *	# is fixed relative to header.
850 	 *	#
851 	 *
852 	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
853 	 *	{ u64			ip;	  } && PERF_SAMPLE_IP
854 	 *	{ u32			pid, tid; } && PERF_SAMPLE_TID
855 	 *	{ u64			time;     } && PERF_SAMPLE_TIME
856 	 *	{ u64			addr;     } && PERF_SAMPLE_ADDR
857 	 *	{ u64			id;	  } && PERF_SAMPLE_ID
858 	 *	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
859 	 *	{ u32			cpu, res; } && PERF_SAMPLE_CPU
860 	 *	{ u64			period;   } && PERF_SAMPLE_PERIOD
861 	 *
862 	 *	{ struct read_format	values;	  } && PERF_SAMPLE_READ
863 	 *
864 	 *	{ u64			nr,
865 	 *	  u64			ips[nr];  } && PERF_SAMPLE_CALLCHAIN
866 	 *
867 	 *	#
868 	 *	# The RAW record below is opaque data wrt the ABI
869 	 *	#
870 	 *	# That is, the ABI doesn't make any promises wrt to
871 	 *	# the stability of its content, it may vary depending
872 	 *	# on event, hardware, kernel version and phase of
873 	 *	# the moon.
874 	 *	#
875 	 *	# In other words, PERF_SAMPLE_RAW contents are not an ABI.
876 	 *	#
877 	 *
878 	 *	{ u32			size;
879 	 *	  char                  data[size];}&& PERF_SAMPLE_RAW
880 	 *
881 	 *	{ u64                   nr;
882 	 *	  { u64	hw_idx; } && PERF_SAMPLE_BRANCH_HW_INDEX
883 	 *        { u64 from, to, flags } lbr[nr];
884 	 *      } && PERF_SAMPLE_BRANCH_STACK
885 	 *
886 	 * 	{ u64			abi; # enum perf_sample_regs_abi
887 	 * 	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
888 	 *
889 	 * 	{ u64			size;
890 	 * 	  char			data[size];
891 	 * 	  u64			dyn_size; } && PERF_SAMPLE_STACK_USER
892 	 *
893 	 *	{ u64			weight;   } && PERF_SAMPLE_WEIGHT
894 	 *	{ u64			data_src; } && PERF_SAMPLE_DATA_SRC
895 	 *	{ u64			transaction; } && PERF_SAMPLE_TRANSACTION
896 	 *	{ u64			abi; # enum perf_sample_regs_abi
897 	 *	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
898 	 *	{ u64			phys_addr;} && PERF_SAMPLE_PHYS_ADDR
899 	 *	{ u64			size;
900 	 *	  char			data[size]; } && PERF_SAMPLE_AUX
901 	 *	{ u64			data_page_size;} && PERF_SAMPLE_DATA_PAGE_SIZE
902 	 *	{ u64			code_page_size;} && PERF_SAMPLE_CODE_PAGE_SIZE
903 	 * };
904 	 */
905 	PERF_RECORD_SAMPLE			= 9,
906 
907 	/*
908 	 * The MMAP2 records are an augmented version of MMAP, they add
909 	 * maj, min, ino numbers to be used to uniquely identify each mapping
910 	 *
911 	 * struct {
912 	 *	struct perf_event_header	header;
913 	 *
914 	 *	u32				pid, tid;
915 	 *	u64				addr;
916 	 *	u64				len;
917 	 *	u64				pgoff;
918 	 *	u32				maj;
919 	 *	u32				min;
920 	 *	u64				ino;
921 	 *	u64				ino_generation;
922 	 *	u32				prot, flags;
923 	 *	char				filename[];
924 	 * 	struct sample_id		sample_id;
925 	 * };
926 	 */
927 	PERF_RECORD_MMAP2			= 10,
928 
929 	/*
930 	 * Records that new data landed in the AUX buffer part.
931 	 *
932 	 * struct {
933 	 * 	struct perf_event_header	header;
934 	 *
935 	 * 	u64				aux_offset;
936 	 * 	u64				aux_size;
937 	 *	u64				flags;
938 	 * 	struct sample_id		sample_id;
939 	 * };
940 	 */
941 	PERF_RECORD_AUX				= 11,
942 
943 	/*
944 	 * Indicates that instruction trace has started
945 	 *
946 	 * struct {
947 	 *	struct perf_event_header	header;
948 	 *	u32				pid;
949 	 *	u32				tid;
950 	 *	struct sample_id		sample_id;
951 	 * };
952 	 */
953 	PERF_RECORD_ITRACE_START		= 12,
954 
955 	/*
956 	 * Records the dropped/lost sample number.
957 	 *
958 	 * struct {
959 	 *	struct perf_event_header	header;
960 	 *
961 	 *	u64				lost;
962 	 *	struct sample_id		sample_id;
963 	 * };
964 	 */
965 	PERF_RECORD_LOST_SAMPLES		= 13,
966 
967 	/*
968 	 * Records a context switch in or out (flagged by
969 	 * PERF_RECORD_MISC_SWITCH_OUT). See also
970 	 * PERF_RECORD_SWITCH_CPU_WIDE.
971 	 *
972 	 * struct {
973 	 *	struct perf_event_header	header;
974 	 *	struct sample_id		sample_id;
975 	 * };
976 	 */
977 	PERF_RECORD_SWITCH			= 14,
978 
979 	/*
980 	 * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and
981 	 * next_prev_tid that are the next (switching out) or previous
982 	 * (switching in) pid/tid.
983 	 *
984 	 * struct {
985 	 *	struct perf_event_header	header;
986 	 *	u32				next_prev_pid;
987 	 *	u32				next_prev_tid;
988 	 *	struct sample_id		sample_id;
989 	 * };
990 	 */
991 	PERF_RECORD_SWITCH_CPU_WIDE		= 15,
992 
993 	/*
994 	 * struct {
995 	 *	struct perf_event_header	header;
996 	 *	u32				pid;
997 	 *	u32				tid;
998 	 *	u64				nr_namespaces;
999 	 *	{ u64				dev, inode; } [nr_namespaces];
1000 	 *	struct sample_id		sample_id;
1001 	 * };
1002 	 */
1003 	PERF_RECORD_NAMESPACES			= 16,
1004 
1005 	/*
1006 	 * Record ksymbol register/unregister events:
1007 	 *
1008 	 * struct {
1009 	 *	struct perf_event_header	header;
1010 	 *	u64				addr;
1011 	 *	u32				len;
1012 	 *	u16				ksym_type;
1013 	 *	u16				flags;
1014 	 *	char				name[];
1015 	 *	struct sample_id		sample_id;
1016 	 * };
1017 	 */
1018 	PERF_RECORD_KSYMBOL			= 17,
1019 
1020 	/*
1021 	 * Record bpf events:
1022 	 *  enum perf_bpf_event_type {
1023 	 *	PERF_BPF_EVENT_UNKNOWN		= 0,
1024 	 *	PERF_BPF_EVENT_PROG_LOAD	= 1,
1025 	 *	PERF_BPF_EVENT_PROG_UNLOAD	= 2,
1026 	 *  };
1027 	 *
1028 	 * struct {
1029 	 *	struct perf_event_header	header;
1030 	 *	u16				type;
1031 	 *	u16				flags;
1032 	 *	u32				id;
1033 	 *	u8				tag[BPF_TAG_SIZE];
1034 	 *	struct sample_id		sample_id;
1035 	 * };
1036 	 */
1037 	PERF_RECORD_BPF_EVENT			= 18,
1038 
1039 	/*
1040 	 * struct {
1041 	 *	struct perf_event_header	header;
1042 	 *	u64				id;
1043 	 *	char				path[];
1044 	 *	struct sample_id		sample_id;
1045 	 * };
1046 	 */
1047 	PERF_RECORD_CGROUP			= 19,
1048 
1049 	/*
1050 	 * Records changes to kernel text i.e. self-modified code. 'old_len' is
1051 	 * the number of old bytes, 'new_len' is the number of new bytes. Either
1052 	 * 'old_len' or 'new_len' may be zero to indicate, for example, the
1053 	 * addition or removal of a trampoline. 'bytes' contains the old bytes
1054 	 * followed immediately by the new bytes.
1055 	 *
1056 	 * struct {
1057 	 *	struct perf_event_header	header;
1058 	 *	u64				addr;
1059 	 *	u16				old_len;
1060 	 *	u16				new_len;
1061 	 *	u8				bytes[];
1062 	 *	struct sample_id		sample_id;
1063 	 * };
1064 	 */
1065 	PERF_RECORD_TEXT_POKE			= 20,
1066 
1067 	PERF_RECORD_MAX,			/* non-ABI */
1068 };
1069 
1070 enum perf_record_ksymbol_type {
1071 	PERF_RECORD_KSYMBOL_TYPE_UNKNOWN	= 0,
1072 	PERF_RECORD_KSYMBOL_TYPE_BPF		= 1,
1073 	/*
1074 	 * Out of line code such as kprobe-replaced instructions or optimized
1075 	 * kprobes or ftrace trampolines.
1076 	 */
1077 	PERF_RECORD_KSYMBOL_TYPE_OOL		= 2,
1078 	PERF_RECORD_KSYMBOL_TYPE_MAX		/* non-ABI */
1079 };
1080 
1081 #define PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER	(1 << 0)
1082 
1083 enum perf_bpf_event_type {
1084 	PERF_BPF_EVENT_UNKNOWN		= 0,
1085 	PERF_BPF_EVENT_PROG_LOAD	= 1,
1086 	PERF_BPF_EVENT_PROG_UNLOAD	= 2,
1087 	PERF_BPF_EVENT_MAX,		/* non-ABI */
1088 };
1089 
1090 #define PERF_MAX_STACK_DEPTH		127
1091 #define PERF_MAX_CONTEXTS_PER_STACK	  8
1092 
1093 enum perf_callchain_context {
1094 	PERF_CONTEXT_HV			= (__u64)-32,
1095 	PERF_CONTEXT_KERNEL		= (__u64)-128,
1096 	PERF_CONTEXT_USER		= (__u64)-512,
1097 
1098 	PERF_CONTEXT_GUEST		= (__u64)-2048,
1099 	PERF_CONTEXT_GUEST_KERNEL	= (__u64)-2176,
1100 	PERF_CONTEXT_GUEST_USER		= (__u64)-2560,
1101 
1102 	PERF_CONTEXT_MAX		= (__u64)-4095,
1103 };
1104 
1105 /**
1106  * PERF_RECORD_AUX::flags bits
1107  */
1108 #define PERF_AUX_FLAG_TRUNCATED		0x01	/* record was truncated to fit */
1109 #define PERF_AUX_FLAG_OVERWRITE		0x02	/* snapshot from overwrite mode */
1110 #define PERF_AUX_FLAG_PARTIAL		0x04	/* record contains gaps */
1111 #define PERF_AUX_FLAG_COLLISION		0x08	/* sample collided with another */
1112 
1113 #define PERF_FLAG_FD_NO_GROUP		(1UL << 0)
1114 #define PERF_FLAG_FD_OUTPUT		(1UL << 1)
1115 #define PERF_FLAG_PID_CGROUP		(1UL << 2) /* pid=cgroup id, per-cpu mode only */
1116 #define PERF_FLAG_FD_CLOEXEC		(1UL << 3) /* O_CLOEXEC */
1117 
1118 #if defined(__LITTLE_ENDIAN_BITFIELD)
1119 union perf_mem_data_src {
1120 	__u64 val;
1121 	struct {
1122 		__u64   mem_op:5,	/* type of opcode */
1123 			mem_lvl:14,	/* memory hierarchy level */
1124 			mem_snoop:5,	/* snoop mode */
1125 			mem_lock:2,	/* lock instr */
1126 			mem_dtlb:7,	/* tlb access */
1127 			mem_lvl_num:4,	/* memory hierarchy level number */
1128 			mem_remote:1,   /* remote */
1129 			mem_snoopx:2,	/* snoop mode, ext */
1130 			mem_rsvd:24;
1131 	};
1132 };
1133 #elif defined(__BIG_ENDIAN_BITFIELD)
1134 union perf_mem_data_src {
1135 	__u64 val;
1136 	struct {
1137 		__u64	mem_rsvd:24,
1138 			mem_snoopx:2,	/* snoop mode, ext */
1139 			mem_remote:1,   /* remote */
1140 			mem_lvl_num:4,	/* memory hierarchy level number */
1141 			mem_dtlb:7,	/* tlb access */
1142 			mem_lock:2,	/* lock instr */
1143 			mem_snoop:5,	/* snoop mode */
1144 			mem_lvl:14,	/* memory hierarchy level */
1145 			mem_op:5;	/* type of opcode */
1146 	};
1147 };
1148 #else
1149 #error "Unknown endianness"
1150 #endif
1151 
1152 /* type of opcode (load/store/prefetch,code) */
1153 #define PERF_MEM_OP_NA		0x01 /* not available */
1154 #define PERF_MEM_OP_LOAD	0x02 /* load instruction */
1155 #define PERF_MEM_OP_STORE	0x04 /* store instruction */
1156 #define PERF_MEM_OP_PFETCH	0x08 /* prefetch */
1157 #define PERF_MEM_OP_EXEC	0x10 /* code (execution) */
1158 #define PERF_MEM_OP_SHIFT	0
1159 
1160 /* memory hierarchy (memory level, hit or miss) */
1161 #define PERF_MEM_LVL_NA		0x01  /* not available */
1162 #define PERF_MEM_LVL_HIT	0x02  /* hit level */
1163 #define PERF_MEM_LVL_MISS	0x04  /* miss level  */
1164 #define PERF_MEM_LVL_L1		0x08  /* L1 */
1165 #define PERF_MEM_LVL_LFB	0x10  /* Line Fill Buffer */
1166 #define PERF_MEM_LVL_L2		0x20  /* L2 */
1167 #define PERF_MEM_LVL_L3		0x40  /* L3 */
1168 #define PERF_MEM_LVL_LOC_RAM	0x80  /* Local DRAM */
1169 #define PERF_MEM_LVL_REM_RAM1	0x100 /* Remote DRAM (1 hop) */
1170 #define PERF_MEM_LVL_REM_RAM2	0x200 /* Remote DRAM (2 hops) */
1171 #define PERF_MEM_LVL_REM_CCE1	0x400 /* Remote Cache (1 hop) */
1172 #define PERF_MEM_LVL_REM_CCE2	0x800 /* Remote Cache (2 hops) */
1173 #define PERF_MEM_LVL_IO		0x1000 /* I/O memory */
1174 #define PERF_MEM_LVL_UNC	0x2000 /* Uncached memory */
1175 #define PERF_MEM_LVL_SHIFT	5
1176 
1177 #define PERF_MEM_REMOTE_REMOTE	0x01  /* Remote */
1178 #define PERF_MEM_REMOTE_SHIFT	37
1179 
1180 #define PERF_MEM_LVLNUM_L1	0x01 /* L1 */
1181 #define PERF_MEM_LVLNUM_L2	0x02 /* L2 */
1182 #define PERF_MEM_LVLNUM_L3	0x03 /* L3 */
1183 #define PERF_MEM_LVLNUM_L4	0x04 /* L4 */
1184 /* 5-0xa available */
1185 #define PERF_MEM_LVLNUM_ANY_CACHE 0x0b /* Any cache */
1186 #define PERF_MEM_LVLNUM_LFB	0x0c /* LFB */
1187 #define PERF_MEM_LVLNUM_RAM	0x0d /* RAM */
1188 #define PERF_MEM_LVLNUM_PMEM	0x0e /* PMEM */
1189 #define PERF_MEM_LVLNUM_NA	0x0f /* N/A */
1190 
1191 #define PERF_MEM_LVLNUM_SHIFT	33
1192 
1193 /* snoop mode */
1194 #define PERF_MEM_SNOOP_NA	0x01 /* not available */
1195 #define PERF_MEM_SNOOP_NONE	0x02 /* no snoop */
1196 #define PERF_MEM_SNOOP_HIT	0x04 /* snoop hit */
1197 #define PERF_MEM_SNOOP_MISS	0x08 /* snoop miss */
1198 #define PERF_MEM_SNOOP_HITM	0x10 /* snoop hit modified */
1199 #define PERF_MEM_SNOOP_SHIFT	19
1200 
1201 #define PERF_MEM_SNOOPX_FWD	0x01 /* forward */
1202 /* 1 free */
1203 #define PERF_MEM_SNOOPX_SHIFT  38
1204 
1205 /* locked instruction */
1206 #define PERF_MEM_LOCK_NA	0x01 /* not available */
1207 #define PERF_MEM_LOCK_LOCKED	0x02 /* locked transaction */
1208 #define PERF_MEM_LOCK_SHIFT	24
1209 
1210 /* TLB access */
1211 #define PERF_MEM_TLB_NA		0x01 /* not available */
1212 #define PERF_MEM_TLB_HIT	0x02 /* hit level */
1213 #define PERF_MEM_TLB_MISS	0x04 /* miss level */
1214 #define PERF_MEM_TLB_L1		0x08 /* L1 */
1215 #define PERF_MEM_TLB_L2		0x10 /* L2 */
1216 #define PERF_MEM_TLB_WK		0x20 /* Hardware Walker*/
1217 #define PERF_MEM_TLB_OS		0x40 /* OS fault handler */
1218 #define PERF_MEM_TLB_SHIFT	26
1219 
1220 #define PERF_MEM_S(a, s) \
1221 	(((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)
1222 
1223 /*
1224  * single taken branch record layout:
1225  *
1226  *      from: source instruction (may not always be a branch insn)
1227  *        to: branch target
1228  *   mispred: branch target was mispredicted
1229  * predicted: branch target was predicted
1230  *
1231  * support for mispred, predicted is optional. In case it
1232  * is not supported mispred = predicted = 0.
1233  *
1234  *     in_tx: running in a hardware transaction
1235  *     abort: aborting a hardware transaction
1236  *    cycles: cycles from last branch (or 0 if not supported)
1237  *      type: branch type
1238  */
1239 struct perf_branch_entry {
1240 	__u64	from;
1241 	__u64	to;
1242 	__u64	mispred:1,  /* target mispredicted */
1243 		predicted:1,/* target predicted */
1244 		in_tx:1,    /* in transaction */
1245 		abort:1,    /* transaction abort */
1246 		cycles:16,  /* cycle count to last branch */
1247 		type:4,     /* branch type */
1248 		reserved:40;
1249 };
1250 
1251 #endif /* _UAPI_LINUX_PERF_EVENT_H */
1252