xref: /linux/tools/include/uapi/linux/perf_event.h (revision e5a52fd2b8cdb700b3c07b030e050a49ef3156b9)
1 /* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
2 /*
3  * Performance events:
4  *
5  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
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 
147 	PERF_SAMPLE_MAX = 1U << 22,		/* non-ABI */
148 
149 	__PERF_SAMPLE_CALLCHAIN_EARLY		= 1ULL << 63, /* non-ABI; internal use */
150 };
151 
152 /*
153  * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
154  *
155  * If the user does not pass priv level information via branch_sample_type,
156  * the kernel uses the event's priv level. Branch and event priv levels do
157  * not have to match. Branch priv level is checked for permissions.
158  *
159  * The branch types can be combined, however BRANCH_ANY covers all types
160  * of branches and therefore it supersedes all the other types.
161  */
162 enum perf_branch_sample_type_shift {
163 	PERF_SAMPLE_BRANCH_USER_SHIFT		= 0, /* user branches */
164 	PERF_SAMPLE_BRANCH_KERNEL_SHIFT		= 1, /* kernel branches */
165 	PERF_SAMPLE_BRANCH_HV_SHIFT		= 2, /* hypervisor branches */
166 
167 	PERF_SAMPLE_BRANCH_ANY_SHIFT		= 3, /* any branch types */
168 	PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT	= 4, /* any call branch */
169 	PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT	= 5, /* any return branch */
170 	PERF_SAMPLE_BRANCH_IND_CALL_SHIFT	= 6, /* indirect calls */
171 	PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT	= 7, /* transaction aborts */
172 	PERF_SAMPLE_BRANCH_IN_TX_SHIFT		= 8, /* in transaction */
173 	PERF_SAMPLE_BRANCH_NO_TX_SHIFT		= 9, /* not in transaction */
174 	PERF_SAMPLE_BRANCH_COND_SHIFT		= 10, /* conditional branches */
175 
176 	PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT	= 11, /* call/ret stack */
177 	PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT	= 12, /* indirect jumps */
178 	PERF_SAMPLE_BRANCH_CALL_SHIFT		= 13, /* direct call */
179 
180 	PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT	= 14, /* no flags */
181 	PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT	= 15, /* no cycles */
182 
183 	PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT	= 16, /* save branch type */
184 
185 	PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT	= 17, /* save low level index of raw branch records */
186 
187 	PERF_SAMPLE_BRANCH_MAX_SHIFT		/* non-ABI */
188 };
189 
190 enum perf_branch_sample_type {
191 	PERF_SAMPLE_BRANCH_USER		= 1U << PERF_SAMPLE_BRANCH_USER_SHIFT,
192 	PERF_SAMPLE_BRANCH_KERNEL	= 1U << PERF_SAMPLE_BRANCH_KERNEL_SHIFT,
193 	PERF_SAMPLE_BRANCH_HV		= 1U << PERF_SAMPLE_BRANCH_HV_SHIFT,
194 
195 	PERF_SAMPLE_BRANCH_ANY		= 1U << PERF_SAMPLE_BRANCH_ANY_SHIFT,
196 	PERF_SAMPLE_BRANCH_ANY_CALL	= 1U << PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT,
197 	PERF_SAMPLE_BRANCH_ANY_RETURN	= 1U << PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT,
198 	PERF_SAMPLE_BRANCH_IND_CALL	= 1U << PERF_SAMPLE_BRANCH_IND_CALL_SHIFT,
199 	PERF_SAMPLE_BRANCH_ABORT_TX	= 1U << PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT,
200 	PERF_SAMPLE_BRANCH_IN_TX	= 1U << PERF_SAMPLE_BRANCH_IN_TX_SHIFT,
201 	PERF_SAMPLE_BRANCH_NO_TX	= 1U << PERF_SAMPLE_BRANCH_NO_TX_SHIFT,
202 	PERF_SAMPLE_BRANCH_COND		= 1U << PERF_SAMPLE_BRANCH_COND_SHIFT,
203 
204 	PERF_SAMPLE_BRANCH_CALL_STACK	= 1U << PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT,
205 	PERF_SAMPLE_BRANCH_IND_JUMP	= 1U << PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT,
206 	PERF_SAMPLE_BRANCH_CALL		= 1U << PERF_SAMPLE_BRANCH_CALL_SHIFT,
207 
208 	PERF_SAMPLE_BRANCH_NO_FLAGS	= 1U << PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT,
209 	PERF_SAMPLE_BRANCH_NO_CYCLES	= 1U << PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT,
210 
211 	PERF_SAMPLE_BRANCH_TYPE_SAVE	=
212 		1U << PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT,
213 
214 	PERF_SAMPLE_BRANCH_HW_INDEX	= 1U << PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT,
215 
216 	PERF_SAMPLE_BRANCH_MAX		= 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT,
217 };
218 
219 /*
220  * Common flow change classification
221  */
222 enum {
223 	PERF_BR_UNKNOWN		= 0,	/* unknown */
224 	PERF_BR_COND		= 1,	/* conditional */
225 	PERF_BR_UNCOND		= 2,	/* unconditional  */
226 	PERF_BR_IND		= 3,	/* indirect */
227 	PERF_BR_CALL		= 4,	/* function call */
228 	PERF_BR_IND_CALL	= 5,	/* indirect function call */
229 	PERF_BR_RET		= 6,	/* function return */
230 	PERF_BR_SYSCALL		= 7,	/* syscall */
231 	PERF_BR_SYSRET		= 8,	/* syscall return */
232 	PERF_BR_COND_CALL	= 9,	/* conditional function call */
233 	PERF_BR_COND_RET	= 10,	/* conditional function return */
234 	PERF_BR_MAX,
235 };
236 
237 #define PERF_SAMPLE_BRANCH_PLM_ALL \
238 	(PERF_SAMPLE_BRANCH_USER|\
239 	 PERF_SAMPLE_BRANCH_KERNEL|\
240 	 PERF_SAMPLE_BRANCH_HV)
241 
242 /*
243  * Values to determine ABI of the registers dump.
244  */
245 enum perf_sample_regs_abi {
246 	PERF_SAMPLE_REGS_ABI_NONE	= 0,
247 	PERF_SAMPLE_REGS_ABI_32		= 1,
248 	PERF_SAMPLE_REGS_ABI_64		= 2,
249 };
250 
251 /*
252  * Values for the memory transaction event qualifier, mostly for
253  * abort events. Multiple bits can be set.
254  */
255 enum {
256 	PERF_TXN_ELISION        = (1 << 0), /* From elision */
257 	PERF_TXN_TRANSACTION    = (1 << 1), /* From transaction */
258 	PERF_TXN_SYNC           = (1 << 2), /* Instruction is related */
259 	PERF_TXN_ASYNC          = (1 << 3), /* Instruction not related */
260 	PERF_TXN_RETRY          = (1 << 4), /* Retry possible */
261 	PERF_TXN_CONFLICT       = (1 << 5), /* Conflict abort */
262 	PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */
263 	PERF_TXN_CAPACITY_READ  = (1 << 7), /* Capacity read abort */
264 
265 	PERF_TXN_MAX	        = (1 << 8), /* non-ABI */
266 
267 	/* bits 32..63 are reserved for the abort code */
268 
269 	PERF_TXN_ABORT_MASK  = (0xffffffffULL << 32),
270 	PERF_TXN_ABORT_SHIFT = 32,
271 };
272 
273 /*
274  * The format of the data returned by read() on a perf event fd,
275  * as specified by attr.read_format:
276  *
277  * struct read_format {
278  *	{ u64		value;
279  *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
280  *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
281  *	  { u64		id;           } && PERF_FORMAT_ID
282  *	} && !PERF_FORMAT_GROUP
283  *
284  *	{ u64		nr;
285  *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
286  *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
287  *	  { u64		value;
288  *	    { u64	id;           } && PERF_FORMAT_ID
289  *	  }		cntr[nr];
290  *	} && PERF_FORMAT_GROUP
291  * };
292  */
293 enum perf_event_read_format {
294 	PERF_FORMAT_TOTAL_TIME_ENABLED		= 1U << 0,
295 	PERF_FORMAT_TOTAL_TIME_RUNNING		= 1U << 1,
296 	PERF_FORMAT_ID				= 1U << 2,
297 	PERF_FORMAT_GROUP			= 1U << 3,
298 
299 	PERF_FORMAT_MAX = 1U << 4,		/* non-ABI */
300 };
301 
302 #define PERF_ATTR_SIZE_VER0	64	/* sizeof first published struct */
303 #define PERF_ATTR_SIZE_VER1	72	/* add: config2 */
304 #define PERF_ATTR_SIZE_VER2	80	/* add: branch_sample_type */
305 #define PERF_ATTR_SIZE_VER3	96	/* add: sample_regs_user */
306 					/* add: sample_stack_user */
307 #define PERF_ATTR_SIZE_VER4	104	/* add: sample_regs_intr */
308 #define PERF_ATTR_SIZE_VER5	112	/* add: aux_watermark */
309 #define PERF_ATTR_SIZE_VER6	120	/* add: aux_sample_size */
310 
311 /*
312  * Hardware event_id to monitor via a performance monitoring event:
313  *
314  * @sample_max_stack: Max number of frame pointers in a callchain,
315  *		      should be < /proc/sys/kernel/perf_event_max_stack
316  */
317 struct perf_event_attr {
318 
319 	/*
320 	 * Major type: hardware/software/tracepoint/etc.
321 	 */
322 	__u32			type;
323 
324 	/*
325 	 * Size of the attr structure, for fwd/bwd compat.
326 	 */
327 	__u32			size;
328 
329 	/*
330 	 * Type specific configuration information.
331 	 */
332 	__u64			config;
333 
334 	union {
335 		__u64		sample_period;
336 		__u64		sample_freq;
337 	};
338 
339 	__u64			sample_type;
340 	__u64			read_format;
341 
342 	__u64			disabled       :  1, /* off by default        */
343 				inherit	       :  1, /* children inherit it   */
344 				pinned	       :  1, /* must always be on PMU */
345 				exclusive      :  1, /* only group on PMU     */
346 				exclude_user   :  1, /* don't count user      */
347 				exclude_kernel :  1, /* ditto kernel          */
348 				exclude_hv     :  1, /* ditto hypervisor      */
349 				exclude_idle   :  1, /* don't count when idle */
350 				mmap           :  1, /* include mmap data     */
351 				comm	       :  1, /* include comm data     */
352 				freq           :  1, /* use freq, not period  */
353 				inherit_stat   :  1, /* per task counts       */
354 				enable_on_exec :  1, /* next exec enables     */
355 				task           :  1, /* trace fork/exit       */
356 				watermark      :  1, /* wakeup_watermark      */
357 				/*
358 				 * precise_ip:
359 				 *
360 				 *  0 - SAMPLE_IP can have arbitrary skid
361 				 *  1 - SAMPLE_IP must have constant skid
362 				 *  2 - SAMPLE_IP requested to have 0 skid
363 				 *  3 - SAMPLE_IP must have 0 skid
364 				 *
365 				 *  See also PERF_RECORD_MISC_EXACT_IP
366 				 */
367 				precise_ip     :  2, /* skid constraint       */
368 				mmap_data      :  1, /* non-exec mmap data    */
369 				sample_id_all  :  1, /* sample_type all events */
370 
371 				exclude_host   :  1, /* don't count in host   */
372 				exclude_guest  :  1, /* don't count in guest  */
373 
374 				exclude_callchain_kernel : 1, /* exclude kernel callchains */
375 				exclude_callchain_user   : 1, /* exclude user callchains */
376 				mmap2          :  1, /* include mmap with inode data     */
377 				comm_exec      :  1, /* flag comm events that are due to an exec */
378 				use_clockid    :  1, /* use @clockid for time fields */
379 				context_switch :  1, /* context switch data */
380 				write_backward :  1, /* Write ring buffer from end to beginning */
381 				namespaces     :  1, /* include namespaces data */
382 				ksymbol        :  1, /* include ksymbol events */
383 				bpf_event      :  1, /* include bpf events */
384 				aux_output     :  1, /* generate AUX records instead of events */
385 				cgroup         :  1, /* include cgroup events */
386 				__reserved_1   : 31;
387 
388 	union {
389 		__u32		wakeup_events;	  /* wakeup every n events */
390 		__u32		wakeup_watermark; /* bytes before wakeup   */
391 	};
392 
393 	__u32			bp_type;
394 	union {
395 		__u64		bp_addr;
396 		__u64		kprobe_func; /* for perf_kprobe */
397 		__u64		uprobe_path; /* for perf_uprobe */
398 		__u64		config1; /* extension of config */
399 	};
400 	union {
401 		__u64		bp_len;
402 		__u64		kprobe_addr; /* when kprobe_func == NULL */
403 		__u64		probe_offset; /* for perf_[k,u]probe */
404 		__u64		config2; /* extension of config1 */
405 	};
406 	__u64	branch_sample_type; /* enum perf_branch_sample_type */
407 
408 	/*
409 	 * Defines set of user regs to dump on samples.
410 	 * See asm/perf_regs.h for details.
411 	 */
412 	__u64	sample_regs_user;
413 
414 	/*
415 	 * Defines size of the user stack to dump on samples.
416 	 */
417 	__u32	sample_stack_user;
418 
419 	__s32	clockid;
420 	/*
421 	 * Defines set of regs to dump for each sample
422 	 * state captured on:
423 	 *  - precise = 0: PMU interrupt
424 	 *  - precise > 0: sampled instruction
425 	 *
426 	 * See asm/perf_regs.h for details.
427 	 */
428 	__u64	sample_regs_intr;
429 
430 	/*
431 	 * Wakeup watermark for AUX area
432 	 */
433 	__u32	aux_watermark;
434 	__u16	sample_max_stack;
435 	__u16	__reserved_2;
436 	__u32	aux_sample_size;
437 	__u32	__reserved_3;
438 };
439 
440 /*
441  * Structure used by below PERF_EVENT_IOC_QUERY_BPF command
442  * to query bpf programs attached to the same perf tracepoint
443  * as the given perf event.
444  */
445 struct perf_event_query_bpf {
446 	/*
447 	 * The below ids array length
448 	 */
449 	__u32	ids_len;
450 	/*
451 	 * Set by the kernel to indicate the number of
452 	 * available programs
453 	 */
454 	__u32	prog_cnt;
455 	/*
456 	 * User provided buffer to store program ids
457 	 */
458 	__u32	ids[0];
459 };
460 
461 /*
462  * Ioctls that can be done on a perf event fd:
463  */
464 #define PERF_EVENT_IOC_ENABLE			_IO ('$', 0)
465 #define PERF_EVENT_IOC_DISABLE			_IO ('$', 1)
466 #define PERF_EVENT_IOC_REFRESH			_IO ('$', 2)
467 #define PERF_EVENT_IOC_RESET			_IO ('$', 3)
468 #define PERF_EVENT_IOC_PERIOD			_IOW('$', 4, __u64)
469 #define PERF_EVENT_IOC_SET_OUTPUT		_IO ('$', 5)
470 #define PERF_EVENT_IOC_SET_FILTER		_IOW('$', 6, char *)
471 #define PERF_EVENT_IOC_ID			_IOR('$', 7, __u64 *)
472 #define PERF_EVENT_IOC_SET_BPF			_IOW('$', 8, __u32)
473 #define PERF_EVENT_IOC_PAUSE_OUTPUT		_IOW('$', 9, __u32)
474 #define PERF_EVENT_IOC_QUERY_BPF		_IOWR('$', 10, struct perf_event_query_bpf *)
475 #define PERF_EVENT_IOC_MODIFY_ATTRIBUTES	_IOW('$', 11, struct perf_event_attr *)
476 
477 enum perf_event_ioc_flags {
478 	PERF_IOC_FLAG_GROUP		= 1U << 0,
479 };
480 
481 /*
482  * Structure of the page that can be mapped via mmap
483  */
484 struct perf_event_mmap_page {
485 	__u32	version;		/* version number of this structure */
486 	__u32	compat_version;		/* lowest version this is compat with */
487 
488 	/*
489 	 * Bits needed to read the hw events in user-space.
490 	 *
491 	 *   u32 seq, time_mult, time_shift, index, width;
492 	 *   u64 count, enabled, running;
493 	 *   u64 cyc, time_offset;
494 	 *   s64 pmc = 0;
495 	 *
496 	 *   do {
497 	 *     seq = pc->lock;
498 	 *     barrier()
499 	 *
500 	 *     enabled = pc->time_enabled;
501 	 *     running = pc->time_running;
502 	 *
503 	 *     if (pc->cap_usr_time && enabled != running) {
504 	 *       cyc = rdtsc();
505 	 *       time_offset = pc->time_offset;
506 	 *       time_mult   = pc->time_mult;
507 	 *       time_shift  = pc->time_shift;
508 	 *     }
509 	 *
510 	 *     index = pc->index;
511 	 *     count = pc->offset;
512 	 *     if (pc->cap_user_rdpmc && index) {
513 	 *       width = pc->pmc_width;
514 	 *       pmc = rdpmc(index - 1);
515 	 *     }
516 	 *
517 	 *     barrier();
518 	 *   } while (pc->lock != seq);
519 	 *
520 	 * NOTE: for obvious reason this only works on self-monitoring
521 	 *       processes.
522 	 */
523 	__u32	lock;			/* seqlock for synchronization */
524 	__u32	index;			/* hardware event identifier */
525 	__s64	offset;			/* add to hardware event value */
526 	__u64	time_enabled;		/* time event active */
527 	__u64	time_running;		/* time event on cpu */
528 	union {
529 		__u64	capabilities;
530 		struct {
531 			__u64	cap_bit0		: 1, /* Always 0, deprecated, see commit 860f085b74e9 */
532 				cap_bit0_is_deprecated	: 1, /* Always 1, signals that bit 0 is zero */
533 
534 				cap_user_rdpmc		: 1, /* The RDPMC instruction can be used to read counts */
535 				cap_user_time		: 1, /* The time_* fields are used */
536 				cap_user_time_zero	: 1, /* The time_zero field is used */
537 				cap_____res		: 59;
538 		};
539 	};
540 
541 	/*
542 	 * If cap_user_rdpmc this field provides the bit-width of the value
543 	 * read using the rdpmc() or equivalent instruction. This can be used
544 	 * to sign extend the result like:
545 	 *
546 	 *   pmc <<= 64 - width;
547 	 *   pmc >>= 64 - width; // signed shift right
548 	 *   count += pmc;
549 	 */
550 	__u16	pmc_width;
551 
552 	/*
553 	 * If cap_usr_time the below fields can be used to compute the time
554 	 * delta since time_enabled (in ns) using rdtsc or similar.
555 	 *
556 	 *   u64 quot, rem;
557 	 *   u64 delta;
558 	 *
559 	 *   quot = (cyc >> time_shift);
560 	 *   rem = cyc & (((u64)1 << time_shift) - 1);
561 	 *   delta = time_offset + quot * time_mult +
562 	 *              ((rem * time_mult) >> time_shift);
563 	 *
564 	 * Where time_offset,time_mult,time_shift and cyc are read in the
565 	 * seqcount loop described above. This delta can then be added to
566 	 * enabled and possible running (if index), improving the scaling:
567 	 *
568 	 *   enabled += delta;
569 	 *   if (index)
570 	 *     running += delta;
571 	 *
572 	 *   quot = count / running;
573 	 *   rem  = count % running;
574 	 *   count = quot * enabled + (rem * enabled) / running;
575 	 */
576 	__u16	time_shift;
577 	__u32	time_mult;
578 	__u64	time_offset;
579 	/*
580 	 * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated
581 	 * from sample timestamps.
582 	 *
583 	 *   time = timestamp - time_zero;
584 	 *   quot = time / time_mult;
585 	 *   rem  = time % time_mult;
586 	 *   cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
587 	 *
588 	 * And vice versa:
589 	 *
590 	 *   quot = cyc >> time_shift;
591 	 *   rem  = cyc & (((u64)1 << time_shift) - 1);
592 	 *   timestamp = time_zero + quot * time_mult +
593 	 *               ((rem * time_mult) >> time_shift);
594 	 */
595 	__u64	time_zero;
596 	__u32	size;			/* Header size up to __reserved[] fields. */
597 
598 		/*
599 		 * Hole for extension of the self monitor capabilities
600 		 */
601 
602 	__u8	__reserved[118*8+4];	/* align to 1k. */
603 
604 	/*
605 	 * Control data for the mmap() data buffer.
606 	 *
607 	 * User-space reading the @data_head value should issue an smp_rmb(),
608 	 * after reading this value.
609 	 *
610 	 * When the mapping is PROT_WRITE the @data_tail value should be
611 	 * written by userspace to reflect the last read data, after issueing
612 	 * an smp_mb() to separate the data read from the ->data_tail store.
613 	 * In this case the kernel will not over-write unread data.
614 	 *
615 	 * See perf_output_put_handle() for the data ordering.
616 	 *
617 	 * data_{offset,size} indicate the location and size of the perf record
618 	 * buffer within the mmapped area.
619 	 */
620 	__u64   data_head;		/* head in the data section */
621 	__u64	data_tail;		/* user-space written tail */
622 	__u64	data_offset;		/* where the buffer starts */
623 	__u64	data_size;		/* data buffer size */
624 
625 	/*
626 	 * AUX area is defined by aux_{offset,size} fields that should be set
627 	 * by the userspace, so that
628 	 *
629 	 *   aux_offset >= data_offset + data_size
630 	 *
631 	 * prior to mmap()ing it. Size of the mmap()ed area should be aux_size.
632 	 *
633 	 * Ring buffer pointers aux_{head,tail} have the same semantics as
634 	 * data_{head,tail} and same ordering rules apply.
635 	 */
636 	__u64	aux_head;
637 	__u64	aux_tail;
638 	__u64	aux_offset;
639 	__u64	aux_size;
640 };
641 
642 #define PERF_RECORD_MISC_CPUMODE_MASK		(7 << 0)
643 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN	(0 << 0)
644 #define PERF_RECORD_MISC_KERNEL			(1 << 0)
645 #define PERF_RECORD_MISC_USER			(2 << 0)
646 #define PERF_RECORD_MISC_HYPERVISOR		(3 << 0)
647 #define PERF_RECORD_MISC_GUEST_KERNEL		(4 << 0)
648 #define PERF_RECORD_MISC_GUEST_USER		(5 << 0)
649 
650 /*
651  * Indicates that /proc/PID/maps parsing are truncated by time out.
652  */
653 #define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT	(1 << 12)
654 /*
655  * Following PERF_RECORD_MISC_* are used on different
656  * events, so can reuse the same bit position:
657  *
658  *   PERF_RECORD_MISC_MMAP_DATA  - PERF_RECORD_MMAP* events
659  *   PERF_RECORD_MISC_COMM_EXEC  - PERF_RECORD_COMM event
660  *   PERF_RECORD_MISC_FORK_EXEC  - PERF_RECORD_FORK event (perf internal)
661  *   PERF_RECORD_MISC_SWITCH_OUT - PERF_RECORD_SWITCH* events
662  */
663 #define PERF_RECORD_MISC_MMAP_DATA		(1 << 13)
664 #define PERF_RECORD_MISC_COMM_EXEC		(1 << 13)
665 #define PERF_RECORD_MISC_FORK_EXEC		(1 << 13)
666 #define PERF_RECORD_MISC_SWITCH_OUT		(1 << 13)
667 /*
668  * These PERF_RECORD_MISC_* flags below are safely reused
669  * for the following events:
670  *
671  *   PERF_RECORD_MISC_EXACT_IP           - PERF_RECORD_SAMPLE of precise events
672  *   PERF_RECORD_MISC_SWITCH_OUT_PREEMPT - PERF_RECORD_SWITCH* events
673  *
674  *
675  * PERF_RECORD_MISC_EXACT_IP:
676  *   Indicates that the content of PERF_SAMPLE_IP points to
677  *   the actual instruction that triggered the event. See also
678  *   perf_event_attr::precise_ip.
679  *
680  * PERF_RECORD_MISC_SWITCH_OUT_PREEMPT:
681  *   Indicates that thread was preempted in TASK_RUNNING state.
682  */
683 #define PERF_RECORD_MISC_EXACT_IP		(1 << 14)
684 #define PERF_RECORD_MISC_SWITCH_OUT_PREEMPT	(1 << 14)
685 /*
686  * Reserve the last bit to indicate some extended misc field
687  */
688 #define PERF_RECORD_MISC_EXT_RESERVED		(1 << 15)
689 
690 struct perf_event_header {
691 	__u32	type;
692 	__u16	misc;
693 	__u16	size;
694 };
695 
696 struct perf_ns_link_info {
697 	__u64	dev;
698 	__u64	ino;
699 };
700 
701 enum {
702 	NET_NS_INDEX		= 0,
703 	UTS_NS_INDEX		= 1,
704 	IPC_NS_INDEX		= 2,
705 	PID_NS_INDEX		= 3,
706 	USER_NS_INDEX		= 4,
707 	MNT_NS_INDEX		= 5,
708 	CGROUP_NS_INDEX		= 6,
709 
710 	NR_NAMESPACES,		/* number of available namespaces */
711 };
712 
713 enum perf_event_type {
714 
715 	/*
716 	 * If perf_event_attr.sample_id_all is set then all event types will
717 	 * have the sample_type selected fields related to where/when
718 	 * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU,
719 	 * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed
720 	 * just after the perf_event_header and the fields already present for
721 	 * the existing fields, i.e. at the end of the payload. That way a newer
722 	 * perf.data file will be supported by older perf tools, with these new
723 	 * optional fields being ignored.
724 	 *
725 	 * struct sample_id {
726 	 * 	{ u32			pid, tid; } && PERF_SAMPLE_TID
727 	 * 	{ u64			time;     } && PERF_SAMPLE_TIME
728 	 * 	{ u64			id;       } && PERF_SAMPLE_ID
729 	 * 	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
730 	 * 	{ u32			cpu, res; } && PERF_SAMPLE_CPU
731 	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
732 	 * } && perf_event_attr::sample_id_all
733 	 *
734 	 * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.  The
735 	 * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed
736 	 * relative to header.size.
737 	 */
738 
739 	/*
740 	 * The MMAP events record the PROT_EXEC mappings so that we can
741 	 * correlate userspace IPs to code. They have the following structure:
742 	 *
743 	 * struct {
744 	 *	struct perf_event_header	header;
745 	 *
746 	 *	u32				pid, tid;
747 	 *	u64				addr;
748 	 *	u64				len;
749 	 *	u64				pgoff;
750 	 *	char				filename[];
751 	 * 	struct sample_id		sample_id;
752 	 * };
753 	 */
754 	PERF_RECORD_MMAP			= 1,
755 
756 	/*
757 	 * struct {
758 	 *	struct perf_event_header	header;
759 	 *	u64				id;
760 	 *	u64				lost;
761 	 * 	struct sample_id		sample_id;
762 	 * };
763 	 */
764 	PERF_RECORD_LOST			= 2,
765 
766 	/*
767 	 * struct {
768 	 *	struct perf_event_header	header;
769 	 *
770 	 *	u32				pid, tid;
771 	 *	char				comm[];
772 	 * 	struct sample_id		sample_id;
773 	 * };
774 	 */
775 	PERF_RECORD_COMM			= 3,
776 
777 	/*
778 	 * struct {
779 	 *	struct perf_event_header	header;
780 	 *	u32				pid, ppid;
781 	 *	u32				tid, ptid;
782 	 *	u64				time;
783 	 * 	struct sample_id		sample_id;
784 	 * };
785 	 */
786 	PERF_RECORD_EXIT			= 4,
787 
788 	/*
789 	 * struct {
790 	 *	struct perf_event_header	header;
791 	 *	u64				time;
792 	 *	u64				id;
793 	 *	u64				stream_id;
794 	 * 	struct sample_id		sample_id;
795 	 * };
796 	 */
797 	PERF_RECORD_THROTTLE			= 5,
798 	PERF_RECORD_UNTHROTTLE			= 6,
799 
800 	/*
801 	 * struct {
802 	 *	struct perf_event_header	header;
803 	 *	u32				pid, ppid;
804 	 *	u32				tid, ptid;
805 	 *	u64				time;
806 	 * 	struct sample_id		sample_id;
807 	 * };
808 	 */
809 	PERF_RECORD_FORK			= 7,
810 
811 	/*
812 	 * struct {
813 	 *	struct perf_event_header	header;
814 	 *	u32				pid, tid;
815 	 *
816 	 *	struct read_format		values;
817 	 * 	struct sample_id		sample_id;
818 	 * };
819 	 */
820 	PERF_RECORD_READ			= 8,
821 
822 	/*
823 	 * struct {
824 	 *	struct perf_event_header	header;
825 	 *
826 	 *	#
827 	 *	# Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.
828 	 *	# The advantage of PERF_SAMPLE_IDENTIFIER is that its position
829 	 *	# is fixed relative to header.
830 	 *	#
831 	 *
832 	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
833 	 *	{ u64			ip;	  } && PERF_SAMPLE_IP
834 	 *	{ u32			pid, tid; } && PERF_SAMPLE_TID
835 	 *	{ u64			time;     } && PERF_SAMPLE_TIME
836 	 *	{ u64			addr;     } && PERF_SAMPLE_ADDR
837 	 *	{ u64			id;	  } && PERF_SAMPLE_ID
838 	 *	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
839 	 *	{ u32			cpu, res; } && PERF_SAMPLE_CPU
840 	 *	{ u64			period;   } && PERF_SAMPLE_PERIOD
841 	 *
842 	 *	{ struct read_format	values;	  } && PERF_SAMPLE_READ
843 	 *
844 	 *	{ u64			nr,
845 	 *	  u64			ips[nr];  } && PERF_SAMPLE_CALLCHAIN
846 	 *
847 	 *	#
848 	 *	# The RAW record below is opaque data wrt the ABI
849 	 *	#
850 	 *	# That is, the ABI doesn't make any promises wrt to
851 	 *	# the stability of its content, it may vary depending
852 	 *	# on event, hardware, kernel version and phase of
853 	 *	# the moon.
854 	 *	#
855 	 *	# In other words, PERF_SAMPLE_RAW contents are not an ABI.
856 	 *	#
857 	 *
858 	 *	{ u32			size;
859 	 *	  char                  data[size];}&& PERF_SAMPLE_RAW
860 	 *
861 	 *	{ u64                   nr;
862 	 *	  { u64	hw_idx; } && PERF_SAMPLE_BRANCH_HW_INDEX
863 	 *        { u64 from, to, flags } lbr[nr];
864 	 *      } && PERF_SAMPLE_BRANCH_STACK
865 	 *
866 	 * 	{ u64			abi; # enum perf_sample_regs_abi
867 	 * 	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
868 	 *
869 	 * 	{ u64			size;
870 	 * 	  char			data[size];
871 	 * 	  u64			dyn_size; } && PERF_SAMPLE_STACK_USER
872 	 *
873 	 *	{ u64			weight;   } && PERF_SAMPLE_WEIGHT
874 	 *	{ u64			data_src; } && PERF_SAMPLE_DATA_SRC
875 	 *	{ u64			transaction; } && PERF_SAMPLE_TRANSACTION
876 	 *	{ u64			abi; # enum perf_sample_regs_abi
877 	 *	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
878 	 *	{ u64			phys_addr;} && PERF_SAMPLE_PHYS_ADDR
879 	 *	{ u64			size;
880 	 *	  char			data[size]; } && PERF_SAMPLE_AUX
881 	 * };
882 	 */
883 	PERF_RECORD_SAMPLE			= 9,
884 
885 	/*
886 	 * The MMAP2 records are an augmented version of MMAP, they add
887 	 * maj, min, ino numbers to be used to uniquely identify each mapping
888 	 *
889 	 * struct {
890 	 *	struct perf_event_header	header;
891 	 *
892 	 *	u32				pid, tid;
893 	 *	u64				addr;
894 	 *	u64				len;
895 	 *	u64				pgoff;
896 	 *	u32				maj;
897 	 *	u32				min;
898 	 *	u64				ino;
899 	 *	u64				ino_generation;
900 	 *	u32				prot, flags;
901 	 *	char				filename[];
902 	 * 	struct sample_id		sample_id;
903 	 * };
904 	 */
905 	PERF_RECORD_MMAP2			= 10,
906 
907 	/*
908 	 * Records that new data landed in the AUX buffer part.
909 	 *
910 	 * struct {
911 	 * 	struct perf_event_header	header;
912 	 *
913 	 * 	u64				aux_offset;
914 	 * 	u64				aux_size;
915 	 *	u64				flags;
916 	 * 	struct sample_id		sample_id;
917 	 * };
918 	 */
919 	PERF_RECORD_AUX				= 11,
920 
921 	/*
922 	 * Indicates that instruction trace has started
923 	 *
924 	 * struct {
925 	 *	struct perf_event_header	header;
926 	 *	u32				pid;
927 	 *	u32				tid;
928 	 *	struct sample_id		sample_id;
929 	 * };
930 	 */
931 	PERF_RECORD_ITRACE_START		= 12,
932 
933 	/*
934 	 * Records the dropped/lost sample number.
935 	 *
936 	 * struct {
937 	 *	struct perf_event_header	header;
938 	 *
939 	 *	u64				lost;
940 	 *	struct sample_id		sample_id;
941 	 * };
942 	 */
943 	PERF_RECORD_LOST_SAMPLES		= 13,
944 
945 	/*
946 	 * Records a context switch in or out (flagged by
947 	 * PERF_RECORD_MISC_SWITCH_OUT). See also
948 	 * PERF_RECORD_SWITCH_CPU_WIDE.
949 	 *
950 	 * struct {
951 	 *	struct perf_event_header	header;
952 	 *	struct sample_id		sample_id;
953 	 * };
954 	 */
955 	PERF_RECORD_SWITCH			= 14,
956 
957 	/*
958 	 * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and
959 	 * next_prev_tid that are the next (switching out) or previous
960 	 * (switching in) pid/tid.
961 	 *
962 	 * struct {
963 	 *	struct perf_event_header	header;
964 	 *	u32				next_prev_pid;
965 	 *	u32				next_prev_tid;
966 	 *	struct sample_id		sample_id;
967 	 * };
968 	 */
969 	PERF_RECORD_SWITCH_CPU_WIDE		= 15,
970 
971 	/*
972 	 * struct {
973 	 *	struct perf_event_header	header;
974 	 *	u32				pid;
975 	 *	u32				tid;
976 	 *	u64				nr_namespaces;
977 	 *	{ u64				dev, inode; } [nr_namespaces];
978 	 *	struct sample_id		sample_id;
979 	 * };
980 	 */
981 	PERF_RECORD_NAMESPACES			= 16,
982 
983 	/*
984 	 * Record ksymbol register/unregister events:
985 	 *
986 	 * struct {
987 	 *	struct perf_event_header	header;
988 	 *	u64				addr;
989 	 *	u32				len;
990 	 *	u16				ksym_type;
991 	 *	u16				flags;
992 	 *	char				name[];
993 	 *	struct sample_id		sample_id;
994 	 * };
995 	 */
996 	PERF_RECORD_KSYMBOL			= 17,
997 
998 	/*
999 	 * Record bpf events:
1000 	 *  enum perf_bpf_event_type {
1001 	 *	PERF_BPF_EVENT_UNKNOWN		= 0,
1002 	 *	PERF_BPF_EVENT_PROG_LOAD	= 1,
1003 	 *	PERF_BPF_EVENT_PROG_UNLOAD	= 2,
1004 	 *  };
1005 	 *
1006 	 * struct {
1007 	 *	struct perf_event_header	header;
1008 	 *	u16				type;
1009 	 *	u16				flags;
1010 	 *	u32				id;
1011 	 *	u8				tag[BPF_TAG_SIZE];
1012 	 *	struct sample_id		sample_id;
1013 	 * };
1014 	 */
1015 	PERF_RECORD_BPF_EVENT			= 18,
1016 
1017 	/*
1018 	 * struct {
1019 	 *	struct perf_event_header	header;
1020 	 *	u64				id;
1021 	 *	char				path[];
1022 	 *	struct sample_id		sample_id;
1023 	 * };
1024 	 */
1025 	PERF_RECORD_CGROUP			= 19,
1026 
1027 	PERF_RECORD_MAX,			/* non-ABI */
1028 };
1029 
1030 enum perf_record_ksymbol_type {
1031 	PERF_RECORD_KSYMBOL_TYPE_UNKNOWN	= 0,
1032 	PERF_RECORD_KSYMBOL_TYPE_BPF		= 1,
1033 	PERF_RECORD_KSYMBOL_TYPE_MAX		/* non-ABI */
1034 };
1035 
1036 #define PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER	(1 << 0)
1037 
1038 enum perf_bpf_event_type {
1039 	PERF_BPF_EVENT_UNKNOWN		= 0,
1040 	PERF_BPF_EVENT_PROG_LOAD	= 1,
1041 	PERF_BPF_EVENT_PROG_UNLOAD	= 2,
1042 	PERF_BPF_EVENT_MAX,		/* non-ABI */
1043 };
1044 
1045 #define PERF_MAX_STACK_DEPTH		127
1046 #define PERF_MAX_CONTEXTS_PER_STACK	  8
1047 
1048 enum perf_callchain_context {
1049 	PERF_CONTEXT_HV			= (__u64)-32,
1050 	PERF_CONTEXT_KERNEL		= (__u64)-128,
1051 	PERF_CONTEXT_USER		= (__u64)-512,
1052 
1053 	PERF_CONTEXT_GUEST		= (__u64)-2048,
1054 	PERF_CONTEXT_GUEST_KERNEL	= (__u64)-2176,
1055 	PERF_CONTEXT_GUEST_USER		= (__u64)-2560,
1056 
1057 	PERF_CONTEXT_MAX		= (__u64)-4095,
1058 };
1059 
1060 /**
1061  * PERF_RECORD_AUX::flags bits
1062  */
1063 #define PERF_AUX_FLAG_TRUNCATED		0x01	/* record was truncated to fit */
1064 #define PERF_AUX_FLAG_OVERWRITE		0x02	/* snapshot from overwrite mode */
1065 #define PERF_AUX_FLAG_PARTIAL		0x04	/* record contains gaps */
1066 #define PERF_AUX_FLAG_COLLISION		0x08	/* sample collided with another */
1067 
1068 #define PERF_FLAG_FD_NO_GROUP		(1UL << 0)
1069 #define PERF_FLAG_FD_OUTPUT		(1UL << 1)
1070 #define PERF_FLAG_PID_CGROUP		(1UL << 2) /* pid=cgroup id, per-cpu mode only */
1071 #define PERF_FLAG_FD_CLOEXEC		(1UL << 3) /* O_CLOEXEC */
1072 
1073 #if defined(__LITTLE_ENDIAN_BITFIELD)
1074 union perf_mem_data_src {
1075 	__u64 val;
1076 	struct {
1077 		__u64   mem_op:5,	/* type of opcode */
1078 			mem_lvl:14,	/* memory hierarchy level */
1079 			mem_snoop:5,	/* snoop mode */
1080 			mem_lock:2,	/* lock instr */
1081 			mem_dtlb:7,	/* tlb access */
1082 			mem_lvl_num:4,	/* memory hierarchy level number */
1083 			mem_remote:1,   /* remote */
1084 			mem_snoopx:2,	/* snoop mode, ext */
1085 			mem_rsvd:24;
1086 	};
1087 };
1088 #elif defined(__BIG_ENDIAN_BITFIELD)
1089 union perf_mem_data_src {
1090 	__u64 val;
1091 	struct {
1092 		__u64	mem_rsvd:24,
1093 			mem_snoopx:2,	/* snoop mode, ext */
1094 			mem_remote:1,   /* remote */
1095 			mem_lvl_num:4,	/* memory hierarchy level number */
1096 			mem_dtlb:7,	/* tlb access */
1097 			mem_lock:2,	/* lock instr */
1098 			mem_snoop:5,	/* snoop mode */
1099 			mem_lvl:14,	/* memory hierarchy level */
1100 			mem_op:5;	/* type of opcode */
1101 	};
1102 };
1103 #else
1104 #error "Unknown endianness"
1105 #endif
1106 
1107 /* type of opcode (load/store/prefetch,code) */
1108 #define PERF_MEM_OP_NA		0x01 /* not available */
1109 #define PERF_MEM_OP_LOAD	0x02 /* load instruction */
1110 #define PERF_MEM_OP_STORE	0x04 /* store instruction */
1111 #define PERF_MEM_OP_PFETCH	0x08 /* prefetch */
1112 #define PERF_MEM_OP_EXEC	0x10 /* code (execution) */
1113 #define PERF_MEM_OP_SHIFT	0
1114 
1115 /* memory hierarchy (memory level, hit or miss) */
1116 #define PERF_MEM_LVL_NA		0x01  /* not available */
1117 #define PERF_MEM_LVL_HIT	0x02  /* hit level */
1118 #define PERF_MEM_LVL_MISS	0x04  /* miss level  */
1119 #define PERF_MEM_LVL_L1		0x08  /* L1 */
1120 #define PERF_MEM_LVL_LFB	0x10  /* Line Fill Buffer */
1121 #define PERF_MEM_LVL_L2		0x20  /* L2 */
1122 #define PERF_MEM_LVL_L3		0x40  /* L3 */
1123 #define PERF_MEM_LVL_LOC_RAM	0x80  /* Local DRAM */
1124 #define PERF_MEM_LVL_REM_RAM1	0x100 /* Remote DRAM (1 hop) */
1125 #define PERF_MEM_LVL_REM_RAM2	0x200 /* Remote DRAM (2 hops) */
1126 #define PERF_MEM_LVL_REM_CCE1	0x400 /* Remote Cache (1 hop) */
1127 #define PERF_MEM_LVL_REM_CCE2	0x800 /* Remote Cache (2 hops) */
1128 #define PERF_MEM_LVL_IO		0x1000 /* I/O memory */
1129 #define PERF_MEM_LVL_UNC	0x2000 /* Uncached memory */
1130 #define PERF_MEM_LVL_SHIFT	5
1131 
1132 #define PERF_MEM_REMOTE_REMOTE	0x01  /* Remote */
1133 #define PERF_MEM_REMOTE_SHIFT	37
1134 
1135 #define PERF_MEM_LVLNUM_L1	0x01 /* L1 */
1136 #define PERF_MEM_LVLNUM_L2	0x02 /* L2 */
1137 #define PERF_MEM_LVLNUM_L3	0x03 /* L3 */
1138 #define PERF_MEM_LVLNUM_L4	0x04 /* L4 */
1139 /* 5-0xa available */
1140 #define PERF_MEM_LVLNUM_ANY_CACHE 0x0b /* Any cache */
1141 #define PERF_MEM_LVLNUM_LFB	0x0c /* LFB */
1142 #define PERF_MEM_LVLNUM_RAM	0x0d /* RAM */
1143 #define PERF_MEM_LVLNUM_PMEM	0x0e /* PMEM */
1144 #define PERF_MEM_LVLNUM_NA	0x0f /* N/A */
1145 
1146 #define PERF_MEM_LVLNUM_SHIFT	33
1147 
1148 /* snoop mode */
1149 #define PERF_MEM_SNOOP_NA	0x01 /* not available */
1150 #define PERF_MEM_SNOOP_NONE	0x02 /* no snoop */
1151 #define PERF_MEM_SNOOP_HIT	0x04 /* snoop hit */
1152 #define PERF_MEM_SNOOP_MISS	0x08 /* snoop miss */
1153 #define PERF_MEM_SNOOP_HITM	0x10 /* snoop hit modified */
1154 #define PERF_MEM_SNOOP_SHIFT	19
1155 
1156 #define PERF_MEM_SNOOPX_FWD	0x01 /* forward */
1157 /* 1 free */
1158 #define PERF_MEM_SNOOPX_SHIFT	37
1159 
1160 /* locked instruction */
1161 #define PERF_MEM_LOCK_NA	0x01 /* not available */
1162 #define PERF_MEM_LOCK_LOCKED	0x02 /* locked transaction */
1163 #define PERF_MEM_LOCK_SHIFT	24
1164 
1165 /* TLB access */
1166 #define PERF_MEM_TLB_NA		0x01 /* not available */
1167 #define PERF_MEM_TLB_HIT	0x02 /* hit level */
1168 #define PERF_MEM_TLB_MISS	0x04 /* miss level */
1169 #define PERF_MEM_TLB_L1		0x08 /* L1 */
1170 #define PERF_MEM_TLB_L2		0x10 /* L2 */
1171 #define PERF_MEM_TLB_WK		0x20 /* Hardware Walker*/
1172 #define PERF_MEM_TLB_OS		0x40 /* OS fault handler */
1173 #define PERF_MEM_TLB_SHIFT	26
1174 
1175 #define PERF_MEM_S(a, s) \
1176 	(((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)
1177 
1178 /*
1179  * single taken branch record layout:
1180  *
1181  *      from: source instruction (may not always be a branch insn)
1182  *        to: branch target
1183  *   mispred: branch target was mispredicted
1184  * predicted: branch target was predicted
1185  *
1186  * support for mispred, predicted is optional. In case it
1187  * is not supported mispred = predicted = 0.
1188  *
1189  *     in_tx: running in a hardware transaction
1190  *     abort: aborting a hardware transaction
1191  *    cycles: cycles from last branch (or 0 if not supported)
1192  *      type: branch type
1193  */
1194 struct perf_branch_entry {
1195 	__u64	from;
1196 	__u64	to;
1197 	__u64	mispred:1,  /* target mispredicted */
1198 		predicted:1,/* target predicted */
1199 		in_tx:1,    /* in transaction */
1200 		abort:1,    /* transaction abort */
1201 		cycles:16,  /* cycle count to last branch */
1202 		type:4,     /* branch type */
1203 		reserved:40;
1204 };
1205 
1206 #endif /* _UAPI_LINUX_PERF_EVENT_H */
1207