• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Performance events:
3  *
4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
6  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
7  *
8  * Data type definitions, declarations, prototypes.
9  *
10  *    Started by: Thomas Gleixner and Ingo Molnar
11  *
12  * For licencing details see kernel-base/COPYING
13  */
14 #ifndef _UAPI_LINUX_PERF_EVENT_H
15 #define _UAPI_LINUX_PERF_EVENT_H
16 
17 #include <linux/types.h>
18 #include <linux/ioctl.h>
19 #include <asm/byteorder.h>
20 
21 /*
22  * User-space ABI bits:
23  */
24 
25 /*
26  * attr.type
27  */
28 enum perf_type_id {
29 	PERF_TYPE_HARDWARE			= 0,
30 	PERF_TYPE_SOFTWARE			= 1,
31 	PERF_TYPE_TRACEPOINT			= 2,
32 	PERF_TYPE_HW_CACHE			= 3,
33 	PERF_TYPE_RAW				= 4,
34 	PERF_TYPE_BREAKPOINT			= 5,
35 
36 	PERF_TYPE_MAX,				/* non-ABI */
37 };
38 
39 /*
40  * Generalized performance event event_id types, used by the
41  * attr.event_id parameter of the sys_perf_event_open()
42  * syscall:
43  */
44 enum perf_hw_id {
45 	/*
46 	 * Common hardware events, generalized by the kernel:
47 	 */
48 	PERF_COUNT_HW_CPU_CYCLES		= 0,
49 	PERF_COUNT_HW_INSTRUCTIONS		= 1,
50 	PERF_COUNT_HW_CACHE_REFERENCES		= 2,
51 	PERF_COUNT_HW_CACHE_MISSES		= 3,
52 	PERF_COUNT_HW_BRANCH_INSTRUCTIONS	= 4,
53 	PERF_COUNT_HW_BRANCH_MISSES		= 5,
54 	PERF_COUNT_HW_BUS_CYCLES		= 6,
55 	PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	= 7,
56 	PERF_COUNT_HW_STALLED_CYCLES_BACKEND	= 8,
57 	PERF_COUNT_HW_REF_CPU_CYCLES		= 9,
58 
59 	PERF_COUNT_HW_MAX,			/* non-ABI */
60 };
61 
62 /*
63  * Generalized hardware cache events:
64  *
65  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
66  *       { read, write, prefetch } x
67  *       { accesses, misses }
68  */
69 enum perf_hw_cache_id {
70 	PERF_COUNT_HW_CACHE_L1D			= 0,
71 	PERF_COUNT_HW_CACHE_L1I			= 1,
72 	PERF_COUNT_HW_CACHE_LL			= 2,
73 	PERF_COUNT_HW_CACHE_DTLB		= 3,
74 	PERF_COUNT_HW_CACHE_ITLB		= 4,
75 	PERF_COUNT_HW_CACHE_BPU			= 5,
76 	PERF_COUNT_HW_CACHE_NODE		= 6,
77 
78 	PERF_COUNT_HW_CACHE_MAX,		/* non-ABI */
79 };
80 
81 enum perf_hw_cache_op_id {
82 	PERF_COUNT_HW_CACHE_OP_READ		= 0,
83 	PERF_COUNT_HW_CACHE_OP_WRITE		= 1,
84 	PERF_COUNT_HW_CACHE_OP_PREFETCH		= 2,
85 
86 	PERF_COUNT_HW_CACHE_OP_MAX,		/* non-ABI */
87 };
88 
89 enum perf_hw_cache_op_result_id {
90 	PERF_COUNT_HW_CACHE_RESULT_ACCESS	= 0,
91 	PERF_COUNT_HW_CACHE_RESULT_MISS		= 1,
92 
93 	PERF_COUNT_HW_CACHE_RESULT_MAX,		/* non-ABI */
94 };
95 
96 /*
97  * Special "software" events provided by the kernel, even if the hardware
98  * does not support performance events. These events measure various
99  * physical and sw events of the kernel (and allow the profiling of them as
100  * well):
101  */
102 enum perf_sw_ids {
103 	PERF_COUNT_SW_CPU_CLOCK			= 0,
104 	PERF_COUNT_SW_TASK_CLOCK		= 1,
105 	PERF_COUNT_SW_PAGE_FAULTS		= 2,
106 	PERF_COUNT_SW_CONTEXT_SWITCHES		= 3,
107 	PERF_COUNT_SW_CPU_MIGRATIONS		= 4,
108 	PERF_COUNT_SW_PAGE_FAULTS_MIN		= 5,
109 	PERF_COUNT_SW_PAGE_FAULTS_MAJ		= 6,
110 	PERF_COUNT_SW_ALIGNMENT_FAULTS		= 7,
111 	PERF_COUNT_SW_EMULATION_FAULTS		= 8,
112 	PERF_COUNT_SW_DUMMY			= 9,
113 
114 	PERF_COUNT_SW_MAX,			/* non-ABI */
115 };
116 
117 /*
118  * Bits that can be set in attr.sample_type to request information
119  * in the overflow packets.
120  */
121 enum perf_event_sample_format {
122 	PERF_SAMPLE_IP				= 1U << 0,
123 	PERF_SAMPLE_TID				= 1U << 1,
124 	PERF_SAMPLE_TIME			= 1U << 2,
125 	PERF_SAMPLE_ADDR			= 1U << 3,
126 	PERF_SAMPLE_READ			= 1U << 4,
127 	PERF_SAMPLE_CALLCHAIN			= 1U << 5,
128 	PERF_SAMPLE_ID				= 1U << 6,
129 	PERF_SAMPLE_CPU				= 1U << 7,
130 	PERF_SAMPLE_PERIOD			= 1U << 8,
131 	PERF_SAMPLE_STREAM_ID			= 1U << 9,
132 	PERF_SAMPLE_RAW				= 1U << 10,
133 	PERF_SAMPLE_BRANCH_STACK		= 1U << 11,
134 	PERF_SAMPLE_REGS_USER			= 1U << 12,
135 	PERF_SAMPLE_STACK_USER			= 1U << 13,
136 	PERF_SAMPLE_WEIGHT			= 1U << 14,
137 	PERF_SAMPLE_DATA_SRC			= 1U << 15,
138 	PERF_SAMPLE_IDENTIFIER			= 1U << 16,
139 	PERF_SAMPLE_TRANSACTION			= 1U << 17,
140 
141 	PERF_SAMPLE_MAX = 1U << 18,		/* non-ABI */
142 };
143 
144 /*
145  * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
146  *
147  * If the user does not pass priv level information via branch_sample_type,
148  * the kernel uses the event's priv level. Branch and event priv levels do
149  * not have to match. Branch priv level is checked for permissions.
150  *
151  * The branch types can be combined, however BRANCH_ANY covers all types
152  * of branches and therefore it supersedes all the other types.
153  */
154 enum perf_branch_sample_type {
155 	PERF_SAMPLE_BRANCH_USER		= 1U << 0, /* user branches */
156 	PERF_SAMPLE_BRANCH_KERNEL	= 1U << 1, /* kernel branches */
157 	PERF_SAMPLE_BRANCH_HV		= 1U << 2, /* hypervisor branches */
158 
159 	PERF_SAMPLE_BRANCH_ANY		= 1U << 3, /* any branch types */
160 	PERF_SAMPLE_BRANCH_ANY_CALL	= 1U << 4, /* any call branch */
161 	PERF_SAMPLE_BRANCH_ANY_RETURN	= 1U << 5, /* any return branch */
162 	PERF_SAMPLE_BRANCH_IND_CALL	= 1U << 6, /* indirect calls */
163 	PERF_SAMPLE_BRANCH_ABORT_TX	= 1U << 7, /* transaction aborts */
164 	PERF_SAMPLE_BRANCH_IN_TX	= 1U << 8, /* in transaction */
165 	PERF_SAMPLE_BRANCH_NO_TX	= 1U << 9, /* not in transaction */
166 	PERF_SAMPLE_BRANCH_COND		= 1U << 10, /* conditional branches */
167 
168 	PERF_SAMPLE_BRANCH_MAX		= 1U << 11, /* non-ABI */
169 };
170 
171 #define PERF_SAMPLE_BRANCH_PLM_ALL \
172 	(PERF_SAMPLE_BRANCH_USER|\
173 	 PERF_SAMPLE_BRANCH_KERNEL|\
174 	 PERF_SAMPLE_BRANCH_HV)
175 
176 /*
177  * Values to determine ABI of the registers dump.
178  */
179 enum perf_sample_regs_abi {
180 	PERF_SAMPLE_REGS_ABI_NONE	= 0,
181 	PERF_SAMPLE_REGS_ABI_32		= 1,
182 	PERF_SAMPLE_REGS_ABI_64		= 2,
183 };
184 
185 /*
186  * Values for the memory transaction event qualifier, mostly for
187  * abort events. Multiple bits can be set.
188  */
189 enum {
190 	PERF_TXN_ELISION        = (1 << 0), /* From elision */
191 	PERF_TXN_TRANSACTION    = (1 << 1), /* From transaction */
192 	PERF_TXN_SYNC           = (1 << 2), /* Instruction is related */
193 	PERF_TXN_ASYNC          = (1 << 3), /* Instruction not related */
194 	PERF_TXN_RETRY          = (1 << 4), /* Retry possible */
195 	PERF_TXN_CONFLICT       = (1 << 5), /* Conflict abort */
196 	PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */
197 	PERF_TXN_CAPACITY_READ  = (1 << 7), /* Capacity read abort */
198 
199 	PERF_TXN_MAX	        = (1 << 8), /* non-ABI */
200 
201 	/* bits 32..63 are reserved for the abort code */
202 
203 	PERF_TXN_ABORT_MASK  = (0xffffffffULL << 32),
204 	PERF_TXN_ABORT_SHIFT = 32,
205 };
206 
207 /*
208  * The format of the data returned by read() on a perf event fd,
209  * as specified by attr.read_format:
210  *
211  * struct read_format {
212  *	{ u64		value;
213  *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
214  *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
215  *	  { u64		id;           } && PERF_FORMAT_ID
216  *	} && !PERF_FORMAT_GROUP
217  *
218  *	{ u64		nr;
219  *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
220  *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
221  *	  { u64		value;
222  *	    { u64	id;           } && PERF_FORMAT_ID
223  *	  }		cntr[nr];
224  *	} && PERF_FORMAT_GROUP
225  * };
226  */
227 enum perf_event_read_format {
228 	PERF_FORMAT_TOTAL_TIME_ENABLED		= 1U << 0,
229 	PERF_FORMAT_TOTAL_TIME_RUNNING		= 1U << 1,
230 	PERF_FORMAT_ID				= 1U << 2,
231 	PERF_FORMAT_GROUP			= 1U << 3,
232 
233 	PERF_FORMAT_MAX = 1U << 4,		/* non-ABI */
234 };
235 
236 #define PERF_ATTR_SIZE_VER0	64	/* sizeof first published struct */
237 #define PERF_ATTR_SIZE_VER1	72	/* add: config2 */
238 #define PERF_ATTR_SIZE_VER2	80	/* add: branch_sample_type */
239 #define PERF_ATTR_SIZE_VER3	96	/* add: sample_regs_user */
240 					/* add: sample_stack_user */
241 
242 /*
243  * Hardware event_id to monitor via a performance monitoring event:
244  */
245 struct perf_event_attr {
246 
247 	/*
248 	 * Major type: hardware/software/tracepoint/etc.
249 	 */
250 	__u32			type;
251 
252 	/*
253 	 * Size of the attr structure, for fwd/bwd compat.
254 	 */
255 	__u32			size;
256 
257 	/*
258 	 * Type specific configuration information.
259 	 */
260 	__u64			config;
261 
262 	union {
263 		__u64		sample_period;
264 		__u64		sample_freq;
265 	};
266 
267 	__u64			sample_type;
268 	__u64			read_format;
269 
270 	__u64			disabled       :  1, /* off by default        */
271 				inherit	       :  1, /* children inherit it   */
272 				pinned	       :  1, /* must always be on PMU */
273 				exclusive      :  1, /* only group on PMU     */
274 				exclude_user   :  1, /* don't count user      */
275 				exclude_kernel :  1, /* ditto kernel          */
276 				exclude_hv     :  1, /* ditto hypervisor      */
277 				exclude_idle   :  1, /* don't count when idle */
278 				mmap           :  1, /* include mmap data     */
279 				comm	       :  1, /* include comm data     */
280 				freq           :  1, /* use freq, not period  */
281 				inherit_stat   :  1, /* per task counts       */
282 				enable_on_exec :  1, /* next exec enables     */
283 				task           :  1, /* trace fork/exit       */
284 				watermark      :  1, /* wakeup_watermark      */
285 				/*
286 				 * precise_ip:
287 				 *
288 				 *  0 - SAMPLE_IP can have arbitrary skid
289 				 *  1 - SAMPLE_IP must have constant skid
290 				 *  2 - SAMPLE_IP requested to have 0 skid
291 				 *  3 - SAMPLE_IP must have 0 skid
292 				 *
293 				 *  See also PERF_RECORD_MISC_EXACT_IP
294 				 */
295 				precise_ip     :  2, /* skid constraint       */
296 				mmap_data      :  1, /* non-exec mmap data    */
297 				sample_id_all  :  1, /* sample_type all events */
298 
299 				exclude_host   :  1, /* don't count in host   */
300 				exclude_guest  :  1, /* don't count in guest  */
301 
302 				exclude_callchain_kernel : 1, /* exclude kernel callchains */
303 				exclude_callchain_user   : 1, /* exclude user callchains */
304 				mmap2          :  1, /* include mmap with inode data     */
305 				comm_exec      :  1, /* flag comm events that are due to an exec */
306 				__reserved_1   : 39;
307 
308 	union {
309 		__u32		wakeup_events;	  /* wakeup every n events */
310 		__u32		wakeup_watermark; /* bytes before wakeup   */
311 	};
312 
313 	__u32			bp_type;
314 	union {
315 		__u64		bp_addr;
316 		__u64		config1; /* extension of config */
317 	};
318 	union {
319 		__u64		bp_len;
320 		__u64		config2; /* extension of config1 */
321 	};
322 	__u64	branch_sample_type; /* enum perf_branch_sample_type */
323 
324 	/*
325 	 * Defines set of user regs to dump on samples.
326 	 * See asm/perf_regs.h for details.
327 	 */
328 	__u64	sample_regs_user;
329 
330 	/*
331 	 * Defines size of the user stack to dump on samples.
332 	 */
333 	__u32	sample_stack_user;
334 
335 	/* Align to u64. */
336 	__u32	__reserved_2;
337 };
338 
339 #define perf_flags(attr)	(*(&(attr)->read_format + 1))
340 
341 /*
342  * Ioctls that can be done on a perf event fd:
343  */
344 #define PERF_EVENT_IOC_ENABLE		_IO ('$', 0)
345 #define PERF_EVENT_IOC_DISABLE		_IO ('$', 1)
346 #define PERF_EVENT_IOC_REFRESH		_IO ('$', 2)
347 #define PERF_EVENT_IOC_RESET		_IO ('$', 3)
348 #define PERF_EVENT_IOC_PERIOD		_IOW('$', 4, __u64)
349 #define PERF_EVENT_IOC_SET_OUTPUT	_IO ('$', 5)
350 #define PERF_EVENT_IOC_SET_FILTER	_IOW('$', 6, char *)
351 #define PERF_EVENT_IOC_ID		_IOR('$', 7, __u64 *)
352 
353 enum perf_event_ioc_flags {
354 	PERF_IOC_FLAG_GROUP		= 1U << 0,
355 };
356 
357 /*
358  * Structure of the page that can be mapped via mmap
359  */
360 struct perf_event_mmap_page {
361 	__u32	version;		/* version number of this structure */
362 	__u32	compat_version;		/* lowest version this is compat with */
363 
364 	/*
365 	 * Bits needed to read the hw events in user-space.
366 	 *
367 	 *   u32 seq, time_mult, time_shift, index, width;
368 	 *   u64 count, enabled, running;
369 	 *   u64 cyc, time_offset;
370 	 *   s64 pmc = 0;
371 	 *
372 	 *   do {
373 	 *     seq = pc->lock;
374 	 *     barrier()
375 	 *
376 	 *     enabled = pc->time_enabled;
377 	 *     running = pc->time_running;
378 	 *
379 	 *     if (pc->cap_usr_time && enabled != running) {
380 	 *       cyc = rdtsc();
381 	 *       time_offset = pc->time_offset;
382 	 *       time_mult   = pc->time_mult;
383 	 *       time_shift  = pc->time_shift;
384 	 *     }
385 	 *
386 	 *     index = pc->index;
387 	 *     count = pc->offset;
388 	 *     if (pc->cap_user_rdpmc && index) {
389 	 *       width = pc->pmc_width;
390 	 *       pmc = rdpmc(index - 1);
391 	 *     }
392 	 *
393 	 *     barrier();
394 	 *   } while (pc->lock != seq);
395 	 *
396 	 * NOTE: for obvious reason this only works on self-monitoring
397 	 *       processes.
398 	 */
399 	__u32	lock;			/* seqlock for synchronization */
400 	__u32	index;			/* hardware event identifier */
401 	__s64	offset;			/* add to hardware event value */
402 	__u64	time_enabled;		/* time event active */
403 	__u64	time_running;		/* time event on cpu */
404 	union {
405 		__u64	capabilities;
406 		struct {
407 			__u64	cap_bit0		: 1, /* Always 0, deprecated, see commit 860f085b74e9 */
408 				cap_bit0_is_deprecated	: 1, /* Always 1, signals that bit 0 is zero */
409 
410 				cap_user_rdpmc		: 1, /* The RDPMC instruction can be used to read counts */
411 				cap_user_time		: 1, /* The time_* fields are used */
412 				cap_user_time_zero	: 1, /* The time_zero field is used */
413 				cap_____res		: 59;
414 		};
415 	};
416 
417 	/*
418 	 * If cap_user_rdpmc this field provides the bit-width of the value
419 	 * read using the rdpmc() or equivalent instruction. This can be used
420 	 * to sign extend the result like:
421 	 *
422 	 *   pmc <<= 64 - width;
423 	 *   pmc >>= 64 - width; // signed shift right
424 	 *   count += pmc;
425 	 */
426 	__u16	pmc_width;
427 
428 	/*
429 	 * If cap_usr_time the below fields can be used to compute the time
430 	 * delta since time_enabled (in ns) using rdtsc or similar.
431 	 *
432 	 *   u64 quot, rem;
433 	 *   u64 delta;
434 	 *
435 	 *   quot = (cyc >> time_shift);
436 	 *   rem = cyc & ((1 << time_shift) - 1);
437 	 *   delta = time_offset + quot * time_mult +
438 	 *              ((rem * time_mult) >> time_shift);
439 	 *
440 	 * Where time_offset,time_mult,time_shift and cyc are read in the
441 	 * seqcount loop described above. This delta can then be added to
442 	 * enabled and possible running (if index), improving the scaling:
443 	 *
444 	 *   enabled += delta;
445 	 *   if (index)
446 	 *     running += delta;
447 	 *
448 	 *   quot = count / running;
449 	 *   rem  = count % running;
450 	 *   count = quot * enabled + (rem * enabled) / running;
451 	 */
452 	__u16	time_shift;
453 	__u32	time_mult;
454 	__u64	time_offset;
455 	/*
456 	 * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated
457 	 * from sample timestamps.
458 	 *
459 	 *   time = timestamp - time_zero;
460 	 *   quot = time / time_mult;
461 	 *   rem  = time % time_mult;
462 	 *   cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
463 	 *
464 	 * And vice versa:
465 	 *
466 	 *   quot = cyc >> time_shift;
467 	 *   rem  = cyc & ((1 << time_shift) - 1);
468 	 *   timestamp = time_zero + quot * time_mult +
469 	 *               ((rem * time_mult) >> time_shift);
470 	 */
471 	__u64	time_zero;
472 	__u32	size;			/* Header size up to __reserved[] fields. */
473 
474 		/*
475 		 * Hole for extension of the self monitor capabilities
476 		 */
477 
478 	__u8	__reserved[118*8+4];	/* align to 1k. */
479 
480 	/*
481 	 * Control data for the mmap() data buffer.
482 	 *
483 	 * User-space reading the @data_head value should issue an smp_rmb(),
484 	 * after reading this value.
485 	 *
486 	 * When the mapping is PROT_WRITE the @data_tail value should be
487 	 * written by userspace to reflect the last read data, after issueing
488 	 * an smp_mb() to separate the data read from the ->data_tail store.
489 	 * In this case the kernel will not over-write unread data.
490 	 *
491 	 * See perf_output_put_handle() for the data ordering.
492 	 */
493 	__u64   data_head;		/* head in the data section */
494 	__u64	data_tail;		/* user-space written tail */
495 };
496 
497 #define PERF_RECORD_MISC_CPUMODE_MASK		(7 << 0)
498 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN	(0 << 0)
499 #define PERF_RECORD_MISC_KERNEL			(1 << 0)
500 #define PERF_RECORD_MISC_USER			(2 << 0)
501 #define PERF_RECORD_MISC_HYPERVISOR		(3 << 0)
502 #define PERF_RECORD_MISC_GUEST_KERNEL		(4 << 0)
503 #define PERF_RECORD_MISC_GUEST_USER		(5 << 0)
504 
505 /*
506  * PERF_RECORD_MISC_MMAP_DATA and PERF_RECORD_MISC_COMM_EXEC are used on
507  * different events so can reuse the same bit position.
508  */
509 #define PERF_RECORD_MISC_MMAP_DATA		(1 << 13)
510 #define PERF_RECORD_MISC_COMM_EXEC		(1 << 13)
511 /*
512  * Indicates that the content of PERF_SAMPLE_IP points to
513  * the actual instruction that triggered the event. See also
514  * perf_event_attr::precise_ip.
515  */
516 #define PERF_RECORD_MISC_EXACT_IP		(1 << 14)
517 /*
518  * Reserve the last bit to indicate some extended misc field
519  */
520 #define PERF_RECORD_MISC_EXT_RESERVED		(1 << 15)
521 
522 struct perf_event_header {
523 	__u32	type;
524 	__u16	misc;
525 	__u16	size;
526 };
527 
528 enum perf_event_type {
529 
530 	/*
531 	 * If perf_event_attr.sample_id_all is set then all event types will
532 	 * have the sample_type selected fields related to where/when
533 	 * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU,
534 	 * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed
535 	 * just after the perf_event_header and the fields already present for
536 	 * the existing fields, i.e. at the end of the payload. That way a newer
537 	 * perf.data file will be supported by older perf tools, with these new
538 	 * optional fields being ignored.
539 	 *
540 	 * struct sample_id {
541 	 * 	{ u32			pid, tid; } && PERF_SAMPLE_TID
542 	 * 	{ u64			time;     } && PERF_SAMPLE_TIME
543 	 * 	{ u64			id;       } && PERF_SAMPLE_ID
544 	 * 	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
545 	 * 	{ u32			cpu, res; } && PERF_SAMPLE_CPU
546 	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
547 	 * } && perf_event_attr::sample_id_all
548 	 *
549 	 * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.  The
550 	 * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed
551 	 * relative to header.size.
552 	 */
553 
554 	/*
555 	 * The MMAP events record the PROT_EXEC mappings so that we can
556 	 * correlate userspace IPs to code. They have the following structure:
557 	 *
558 	 * struct {
559 	 *	struct perf_event_header	header;
560 	 *
561 	 *	u32				pid, tid;
562 	 *	u64				addr;
563 	 *	u64				len;
564 	 *	u64				pgoff;
565 	 *	char				filename[];
566 	 * 	struct sample_id		sample_id;
567 	 * };
568 	 */
569 	PERF_RECORD_MMAP			= 1,
570 
571 	/*
572 	 * struct {
573 	 *	struct perf_event_header	header;
574 	 *	u64				id;
575 	 *	u64				lost;
576 	 * 	struct sample_id		sample_id;
577 	 * };
578 	 */
579 	PERF_RECORD_LOST			= 2,
580 
581 	/*
582 	 * struct {
583 	 *	struct perf_event_header	header;
584 	 *
585 	 *	u32				pid, tid;
586 	 *	char				comm[];
587 	 * 	struct sample_id		sample_id;
588 	 * };
589 	 */
590 	PERF_RECORD_COMM			= 3,
591 
592 	/*
593 	 * struct {
594 	 *	struct perf_event_header	header;
595 	 *	u32				pid, ppid;
596 	 *	u32				tid, ptid;
597 	 *	u64				time;
598 	 * 	struct sample_id		sample_id;
599 	 * };
600 	 */
601 	PERF_RECORD_EXIT			= 4,
602 
603 	/*
604 	 * struct {
605 	 *	struct perf_event_header	header;
606 	 *	u64				time;
607 	 *	u64				id;
608 	 *	u64				stream_id;
609 	 * 	struct sample_id		sample_id;
610 	 * };
611 	 */
612 	PERF_RECORD_THROTTLE			= 5,
613 	PERF_RECORD_UNTHROTTLE			= 6,
614 
615 	/*
616 	 * struct {
617 	 *	struct perf_event_header	header;
618 	 *	u32				pid, ppid;
619 	 *	u32				tid, ptid;
620 	 *	u64				time;
621 	 * 	struct sample_id		sample_id;
622 	 * };
623 	 */
624 	PERF_RECORD_FORK			= 7,
625 
626 	/*
627 	 * struct {
628 	 *	struct perf_event_header	header;
629 	 *	u32				pid, tid;
630 	 *
631 	 *	struct read_format		values;
632 	 * 	struct sample_id		sample_id;
633 	 * };
634 	 */
635 	PERF_RECORD_READ			= 8,
636 
637 	/*
638 	 * struct {
639 	 *	struct perf_event_header	header;
640 	 *
641 	 *	#
642 	 *	# Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.
643 	 *	# The advantage of PERF_SAMPLE_IDENTIFIER is that its position
644 	 *	# is fixed relative to header.
645 	 *	#
646 	 *
647 	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
648 	 *	{ u64			ip;	  } && PERF_SAMPLE_IP
649 	 *	{ u32			pid, tid; } && PERF_SAMPLE_TID
650 	 *	{ u64			time;     } && PERF_SAMPLE_TIME
651 	 *	{ u64			addr;     } && PERF_SAMPLE_ADDR
652 	 *	{ u64			id;	  } && PERF_SAMPLE_ID
653 	 *	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
654 	 *	{ u32			cpu, res; } && PERF_SAMPLE_CPU
655 	 *	{ u64			period;   } && PERF_SAMPLE_PERIOD
656 	 *
657 	 *	{ struct read_format	values;	  } && PERF_SAMPLE_READ
658 	 *
659 	 *	{ u64			nr,
660 	 *	  u64			ips[nr];  } && PERF_SAMPLE_CALLCHAIN
661 	 *
662 	 *	#
663 	 *	# The RAW record below is opaque data wrt the ABI
664 	 *	#
665 	 *	# That is, the ABI doesn't make any promises wrt to
666 	 *	# the stability of its content, it may vary depending
667 	 *	# on event, hardware, kernel version and phase of
668 	 *	# the moon.
669 	 *	#
670 	 *	# In other words, PERF_SAMPLE_RAW contents are not an ABI.
671 	 *	#
672 	 *
673 	 *	{ u32			size;
674 	 *	  char                  data[size];}&& PERF_SAMPLE_RAW
675 	 *
676 	 *	{ u64                   nr;
677 	 *        { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK
678 	 *
679 	 * 	{ u64			abi; # enum perf_sample_regs_abi
680 	 * 	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
681 	 *
682 	 * 	{ u64			size;
683 	 * 	  char			data[size];
684 	 * 	  u64			dyn_size; } && PERF_SAMPLE_STACK_USER
685 	 *
686 	 *	{ u64			weight;   } && PERF_SAMPLE_WEIGHT
687 	 *	{ u64			data_src; } && PERF_SAMPLE_DATA_SRC
688 	 *	{ u64			transaction; } && PERF_SAMPLE_TRANSACTION
689 	 * };
690 	 */
691 	PERF_RECORD_SAMPLE			= 9,
692 
693 	/*
694 	 * The MMAP2 records are an augmented version of MMAP, they add
695 	 * maj, min, ino numbers to be used to uniquely identify each mapping
696 	 *
697 	 * struct {
698 	 *	struct perf_event_header	header;
699 	 *
700 	 *	u32				pid, tid;
701 	 *	u64				addr;
702 	 *	u64				len;
703 	 *	u64				pgoff;
704 	 *	u32				maj;
705 	 *	u32				min;
706 	 *	u64				ino;
707 	 *	u64				ino_generation;
708 	 *	u32				prot, flags;
709 	 *	char				filename[];
710 	 * 	struct sample_id		sample_id;
711 	 * };
712 	 */
713 	PERF_RECORD_MMAP2			= 10,
714 
715 	PERF_RECORD_MAX,			/* non-ABI */
716 };
717 
718 #define PERF_MAX_STACK_DEPTH		127
719 
720 enum perf_callchain_context {
721 	PERF_CONTEXT_HV			= (__u64)-32,
722 	PERF_CONTEXT_KERNEL		= (__u64)-128,
723 	PERF_CONTEXT_USER		= (__u64)-512,
724 
725 	PERF_CONTEXT_GUEST		= (__u64)-2048,
726 	PERF_CONTEXT_GUEST_KERNEL	= (__u64)-2176,
727 	PERF_CONTEXT_GUEST_USER		= (__u64)-2560,
728 
729 	PERF_CONTEXT_MAX		= (__u64)-4095,
730 };
731 
732 #define PERF_FLAG_FD_NO_GROUP		(1UL << 0)
733 #define PERF_FLAG_FD_OUTPUT		(1UL << 1)
734 #define PERF_FLAG_PID_CGROUP		(1UL << 2) /* pid=cgroup id, per-cpu mode only */
735 #define PERF_FLAG_FD_CLOEXEC		(1UL << 3) /* O_CLOEXEC */
736 
737 union perf_mem_data_src {
738 	__u64 val;
739 	struct {
740 		__u64   mem_op:5,	/* type of opcode */
741 			mem_lvl:14,	/* memory hierarchy level */
742 			mem_snoop:5,	/* snoop mode */
743 			mem_lock:2,	/* lock instr */
744 			mem_dtlb:7,	/* tlb access */
745 			mem_rsvd:31;
746 	};
747 };
748 
749 /* type of opcode (load/store/prefetch,code) */
750 #define PERF_MEM_OP_NA		0x01 /* not available */
751 #define PERF_MEM_OP_LOAD	0x02 /* load instruction */
752 #define PERF_MEM_OP_STORE	0x04 /* store instruction */
753 #define PERF_MEM_OP_PFETCH	0x08 /* prefetch */
754 #define PERF_MEM_OP_EXEC	0x10 /* code (execution) */
755 #define PERF_MEM_OP_SHIFT	0
756 
757 /* memory hierarchy (memory level, hit or miss) */
758 #define PERF_MEM_LVL_NA		0x01  /* not available */
759 #define PERF_MEM_LVL_HIT	0x02  /* hit level */
760 #define PERF_MEM_LVL_MISS	0x04  /* miss level  */
761 #define PERF_MEM_LVL_L1		0x08  /* L1 */
762 #define PERF_MEM_LVL_LFB	0x10  /* Line Fill Buffer */
763 #define PERF_MEM_LVL_L2		0x20  /* L2 */
764 #define PERF_MEM_LVL_L3		0x40  /* L3 */
765 #define PERF_MEM_LVL_LOC_RAM	0x80  /* Local DRAM */
766 #define PERF_MEM_LVL_REM_RAM1	0x100 /* Remote DRAM (1 hop) */
767 #define PERF_MEM_LVL_REM_RAM2	0x200 /* Remote DRAM (2 hops) */
768 #define PERF_MEM_LVL_REM_CCE1	0x400 /* Remote Cache (1 hop) */
769 #define PERF_MEM_LVL_REM_CCE2	0x800 /* Remote Cache (2 hops) */
770 #define PERF_MEM_LVL_IO		0x1000 /* I/O memory */
771 #define PERF_MEM_LVL_UNC	0x2000 /* Uncached memory */
772 #define PERF_MEM_LVL_SHIFT	5
773 
774 /* snoop mode */
775 #define PERF_MEM_SNOOP_NA	0x01 /* not available */
776 #define PERF_MEM_SNOOP_NONE	0x02 /* no snoop */
777 #define PERF_MEM_SNOOP_HIT	0x04 /* snoop hit */
778 #define PERF_MEM_SNOOP_MISS	0x08 /* snoop miss */
779 #define PERF_MEM_SNOOP_HITM	0x10 /* snoop hit modified */
780 #define PERF_MEM_SNOOP_SHIFT	19
781 
782 /* locked instruction */
783 #define PERF_MEM_LOCK_NA	0x01 /* not available */
784 #define PERF_MEM_LOCK_LOCKED	0x02 /* locked transaction */
785 #define PERF_MEM_LOCK_SHIFT	24
786 
787 /* TLB access */
788 #define PERF_MEM_TLB_NA		0x01 /* not available */
789 #define PERF_MEM_TLB_HIT	0x02 /* hit level */
790 #define PERF_MEM_TLB_MISS	0x04 /* miss level */
791 #define PERF_MEM_TLB_L1		0x08 /* L1 */
792 #define PERF_MEM_TLB_L2		0x10 /* L2 */
793 #define PERF_MEM_TLB_WK		0x20 /* Hardware Walker*/
794 #define PERF_MEM_TLB_OS		0x40 /* OS fault handler */
795 #define PERF_MEM_TLB_SHIFT	26
796 
797 #define PERF_MEM_S(a, s) \
798 	(((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)
799 
800 /*
801  * single taken branch record layout:
802  *
803  *      from: source instruction (may not always be a branch insn)
804  *        to: branch target
805  *   mispred: branch target was mispredicted
806  * predicted: branch target was predicted
807  *
808  * support for mispred, predicted is optional. In case it
809  * is not supported mispred = predicted = 0.
810  *
811  *     in_tx: running in a hardware transaction
812  *     abort: aborting a hardware transaction
813  */
814 struct perf_branch_entry {
815 	__u64	from;
816 	__u64	to;
817 	__u64	mispred:1,  /* target mispredicted */
818 		predicted:1,/* target predicted */
819 		in_tx:1,    /* in transaction */
820 		abort:1,    /* transaction abort */
821 		reserved:60;
822 };
823 
824 #endif /* _UAPI_LINUX_PERF_EVENT_H */
825