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 */
28enum 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 */
44enum 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 */
69enum 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
81enum 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
89enum 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 */
102enum 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 */
121enum 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	PERF_SAMPLE_REGS_INTR			= 1U << 18,
141
142	PERF_SAMPLE_MAX = 1U << 19,		/* non-ABI */
143};
144
145/*
146 * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
147 *
148 * If the user does not pass priv level information via branch_sample_type,
149 * the kernel uses the event's priv level. Branch and event priv levels do
150 * not have to match. Branch priv level is checked for permissions.
151 *
152 * The branch types can be combined, however BRANCH_ANY covers all types
153 * of branches and therefore it supersedes all the other types.
154 */
155enum perf_branch_sample_type_shift {
156	PERF_SAMPLE_BRANCH_USER_SHIFT		= 0, /* user branches */
157	PERF_SAMPLE_BRANCH_KERNEL_SHIFT		= 1, /* kernel branches */
158	PERF_SAMPLE_BRANCH_HV_SHIFT		= 2, /* hypervisor branches */
159
160	PERF_SAMPLE_BRANCH_ANY_SHIFT		= 3, /* any branch types */
161	PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT	= 4, /* any call branch */
162	PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT	= 5, /* any return branch */
163	PERF_SAMPLE_BRANCH_IND_CALL_SHIFT	= 6, /* indirect calls */
164	PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT	= 7, /* transaction aborts */
165	PERF_SAMPLE_BRANCH_IN_TX_SHIFT		= 8, /* in transaction */
166	PERF_SAMPLE_BRANCH_NO_TX_SHIFT		= 9, /* not in transaction */
167	PERF_SAMPLE_BRANCH_COND_SHIFT		= 10, /* conditional branches */
168
169	PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT	= 11, /* call/ret stack */
170
171	PERF_SAMPLE_BRANCH_MAX_SHIFT		/* non-ABI */
172};
173
174enum perf_branch_sample_type {
175	PERF_SAMPLE_BRANCH_USER		= 1U << PERF_SAMPLE_BRANCH_USER_SHIFT,
176	PERF_SAMPLE_BRANCH_KERNEL	= 1U << PERF_SAMPLE_BRANCH_KERNEL_SHIFT,
177	PERF_SAMPLE_BRANCH_HV		= 1U << PERF_SAMPLE_BRANCH_HV_SHIFT,
178
179	PERF_SAMPLE_BRANCH_ANY		= 1U << PERF_SAMPLE_BRANCH_ANY_SHIFT,
180	PERF_SAMPLE_BRANCH_ANY_CALL	= 1U << PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT,
181	PERF_SAMPLE_BRANCH_ANY_RETURN	= 1U << PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT,
182	PERF_SAMPLE_BRANCH_IND_CALL	= 1U << PERF_SAMPLE_BRANCH_IND_CALL_SHIFT,
183	PERF_SAMPLE_BRANCH_ABORT_TX	= 1U << PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT,
184	PERF_SAMPLE_BRANCH_IN_TX	= 1U << PERF_SAMPLE_BRANCH_IN_TX_SHIFT,
185	PERF_SAMPLE_BRANCH_NO_TX	= 1U << PERF_SAMPLE_BRANCH_NO_TX_SHIFT,
186	PERF_SAMPLE_BRANCH_COND		= 1U << PERF_SAMPLE_BRANCH_COND_SHIFT,
187
188	PERF_SAMPLE_BRANCH_CALL_STACK	= 1U << PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT,
189
190	PERF_SAMPLE_BRANCH_MAX		= 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT,
191};
192
193#define PERF_SAMPLE_BRANCH_PLM_ALL \
194	(PERF_SAMPLE_BRANCH_USER|\
195	 PERF_SAMPLE_BRANCH_KERNEL|\
196	 PERF_SAMPLE_BRANCH_HV)
197
198/*
199 * Values to determine ABI of the registers dump.
200 */
201enum perf_sample_regs_abi {
202	PERF_SAMPLE_REGS_ABI_NONE	= 0,
203	PERF_SAMPLE_REGS_ABI_32		= 1,
204	PERF_SAMPLE_REGS_ABI_64		= 2,
205};
206
207/*
208 * Values for the memory transaction event qualifier, mostly for
209 * abort events. Multiple bits can be set.
210 */
211enum {
212	PERF_TXN_ELISION        = (1 << 0), /* From elision */
213	PERF_TXN_TRANSACTION    = (1 << 1), /* From transaction */
214	PERF_TXN_SYNC           = (1 << 2), /* Instruction is related */
215	PERF_TXN_ASYNC          = (1 << 3), /* Instruction not related */
216	PERF_TXN_RETRY          = (1 << 4), /* Retry possible */
217	PERF_TXN_CONFLICT       = (1 << 5), /* Conflict abort */
218	PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */
219	PERF_TXN_CAPACITY_READ  = (1 << 7), /* Capacity read abort */
220
221	PERF_TXN_MAX	        = (1 << 8), /* non-ABI */
222
223	/* bits 32..63 are reserved for the abort code */
224
225	PERF_TXN_ABORT_MASK  = (0xffffffffULL << 32),
226	PERF_TXN_ABORT_SHIFT = 32,
227};
228
229/*
230 * The format of the data returned by read() on a perf event fd,
231 * as specified by attr.read_format:
232 *
233 * struct read_format {
234 *	{ u64		value;
235 *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
236 *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
237 *	  { u64		id;           } && PERF_FORMAT_ID
238 *	} && !PERF_FORMAT_GROUP
239 *
240 *	{ u64		nr;
241 *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
242 *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
243 *	  { u64		value;
244 *	    { u64	id;           } && PERF_FORMAT_ID
245 *	  }		cntr[nr];
246 *	} && PERF_FORMAT_GROUP
247 * };
248 */
249enum perf_event_read_format {
250	PERF_FORMAT_TOTAL_TIME_ENABLED		= 1U << 0,
251	PERF_FORMAT_TOTAL_TIME_RUNNING		= 1U << 1,
252	PERF_FORMAT_ID				= 1U << 2,
253	PERF_FORMAT_GROUP			= 1U << 3,
254
255	PERF_FORMAT_MAX = 1U << 4,		/* non-ABI */
256};
257
258#define PERF_ATTR_SIZE_VER0	64	/* sizeof first published struct */
259#define PERF_ATTR_SIZE_VER1	72	/* add: config2 */
260#define PERF_ATTR_SIZE_VER2	80	/* add: branch_sample_type */
261#define PERF_ATTR_SIZE_VER3	96	/* add: sample_regs_user */
262					/* add: sample_stack_user */
263#define PERF_ATTR_SIZE_VER4	104	/* add: sample_regs_intr */
264#define PERF_ATTR_SIZE_VER5	112	/* add: aux_watermark */
265
266/*
267 * Hardware event_id to monitor via a performance monitoring event:
268 */
269struct perf_event_attr {
270
271	/*
272	 * Major type: hardware/software/tracepoint/etc.
273	 */
274	__u32			type;
275
276	/*
277	 * Size of the attr structure, for fwd/bwd compat.
278	 */
279	__u32			size;
280
281	/*
282	 * Type specific configuration information.
283	 */
284	__u64			config;
285
286	union {
287		__u64		sample_period;
288		__u64		sample_freq;
289	};
290
291	__u64			sample_type;
292	__u64			read_format;
293
294	__u64			disabled       :  1, /* off by default        */
295				inherit	       :  1, /* children inherit it   */
296				pinned	       :  1, /* must always be on PMU */
297				exclusive      :  1, /* only group on PMU     */
298				exclude_user   :  1, /* don't count user      */
299				exclude_kernel :  1, /* ditto kernel          */
300				exclude_hv     :  1, /* ditto hypervisor      */
301				exclude_idle   :  1, /* don't count when idle */
302				mmap           :  1, /* include mmap data     */
303				comm	       :  1, /* include comm data     */
304				freq           :  1, /* use freq, not period  */
305				inherit_stat   :  1, /* per task counts       */
306				enable_on_exec :  1, /* next exec enables     */
307				task           :  1, /* trace fork/exit       */
308				watermark      :  1, /* wakeup_watermark      */
309				/*
310				 * precise_ip:
311				 *
312				 *  0 - SAMPLE_IP can have arbitrary skid
313				 *  1 - SAMPLE_IP must have constant skid
314				 *  2 - SAMPLE_IP requested to have 0 skid
315				 *  3 - SAMPLE_IP must have 0 skid
316				 *
317				 *  See also PERF_RECORD_MISC_EXACT_IP
318				 */
319				precise_ip     :  2, /* skid constraint       */
320				mmap_data      :  1, /* non-exec mmap data    */
321				sample_id_all  :  1, /* sample_type all events */
322
323				exclude_host   :  1, /* don't count in host   */
324				exclude_guest  :  1, /* don't count in guest  */
325
326				exclude_callchain_kernel : 1, /* exclude kernel callchains */
327				exclude_callchain_user   : 1, /* exclude user callchains */
328				mmap2          :  1, /* include mmap with inode data     */
329				comm_exec      :  1, /* flag comm events that are due to an exec */
330				use_clockid    :  1, /* use @clockid for time fields */
331				__reserved_1   : 38;
332
333	union {
334		__u32		wakeup_events;	  /* wakeup every n events */
335		__u32		wakeup_watermark; /* bytes before wakeup   */
336	};
337
338	__u32			bp_type;
339	union {
340		__u64		bp_addr;
341		__u64		config1; /* extension of config */
342	};
343	union {
344		__u64		bp_len;
345		__u64		config2; /* extension of config1 */
346	};
347	__u64	branch_sample_type; /* enum perf_branch_sample_type */
348
349	/*
350	 * Defines set of user regs to dump on samples.
351	 * See asm/perf_regs.h for details.
352	 */
353	__u64	sample_regs_user;
354
355	/*
356	 * Defines size of the user stack to dump on samples.
357	 */
358	__u32	sample_stack_user;
359
360	__s32	clockid;
361	/*
362	 * Defines set of regs to dump for each sample
363	 * state captured on:
364	 *  - precise = 0: PMU interrupt
365	 *  - precise > 0: sampled instruction
366	 *
367	 * See asm/perf_regs.h for details.
368	 */
369	__u64	sample_regs_intr;
370
371	/*
372	 * Wakeup watermark for AUX area
373	 */
374	__u32	aux_watermark;
375	__u32	__reserved_2;	/* align to __u64 */
376};
377
378#define perf_flags(attr)	(*(&(attr)->read_format + 1))
379
380/*
381 * Ioctls that can be done on a perf event fd:
382 */
383#define PERF_EVENT_IOC_ENABLE		_IO ('$', 0)
384#define PERF_EVENT_IOC_DISABLE		_IO ('$', 1)
385#define PERF_EVENT_IOC_REFRESH		_IO ('$', 2)
386#define PERF_EVENT_IOC_RESET		_IO ('$', 3)
387#define PERF_EVENT_IOC_PERIOD		_IOW('$', 4, __u64)
388#define PERF_EVENT_IOC_SET_OUTPUT	_IO ('$', 5)
389#define PERF_EVENT_IOC_SET_FILTER	_IOW('$', 6, char *)
390#define PERF_EVENT_IOC_ID		_IOR('$', 7, __u64 *)
391#define PERF_EVENT_IOC_SET_BPF		_IOW('$', 8, __u32)
392
393enum perf_event_ioc_flags {
394	PERF_IOC_FLAG_GROUP		= 1U << 0,
395};
396
397/*
398 * Structure of the page that can be mapped via mmap
399 */
400struct perf_event_mmap_page {
401	__u32	version;		/* version number of this structure */
402	__u32	compat_version;		/* lowest version this is compat with */
403
404	/*
405	 * Bits needed to read the hw events in user-space.
406	 *
407	 *   u32 seq, time_mult, time_shift, index, width;
408	 *   u64 count, enabled, running;
409	 *   u64 cyc, time_offset;
410	 *   s64 pmc = 0;
411	 *
412	 *   do {
413	 *     seq = pc->lock;
414	 *     barrier()
415	 *
416	 *     enabled = pc->time_enabled;
417	 *     running = pc->time_running;
418	 *
419	 *     if (pc->cap_usr_time && enabled != running) {
420	 *       cyc = rdtsc();
421	 *       time_offset = pc->time_offset;
422	 *       time_mult   = pc->time_mult;
423	 *       time_shift  = pc->time_shift;
424	 *     }
425	 *
426	 *     index = pc->index;
427	 *     count = pc->offset;
428	 *     if (pc->cap_user_rdpmc && index) {
429	 *       width = pc->pmc_width;
430	 *       pmc = rdpmc(index - 1);
431	 *     }
432	 *
433	 *     barrier();
434	 *   } while (pc->lock != seq);
435	 *
436	 * NOTE: for obvious reason this only works on self-monitoring
437	 *       processes.
438	 */
439	__u32	lock;			/* seqlock for synchronization */
440	__u32	index;			/* hardware event identifier */
441	__s64	offset;			/* add to hardware event value */
442	__u64	time_enabled;		/* time event active */
443	__u64	time_running;		/* time event on cpu */
444	union {
445		__u64	capabilities;
446		struct {
447			__u64	cap_bit0		: 1, /* Always 0, deprecated, see commit 860f085b74e9 */
448				cap_bit0_is_deprecated	: 1, /* Always 1, signals that bit 0 is zero */
449
450				cap_user_rdpmc		: 1, /* The RDPMC instruction can be used to read counts */
451				cap_user_time		: 1, /* The time_* fields are used */
452				cap_user_time_zero	: 1, /* The time_zero field is used */
453				cap_____res		: 59;
454		};
455	};
456
457	/*
458	 * If cap_user_rdpmc this field provides the bit-width of the value
459	 * read using the rdpmc() or equivalent instruction. This can be used
460	 * to sign extend the result like:
461	 *
462	 *   pmc <<= 64 - width;
463	 *   pmc >>= 64 - width; // signed shift right
464	 *   count += pmc;
465	 */
466	__u16	pmc_width;
467
468	/*
469	 * If cap_usr_time the below fields can be used to compute the time
470	 * delta since time_enabled (in ns) using rdtsc or similar.
471	 *
472	 *   u64 quot, rem;
473	 *   u64 delta;
474	 *
475	 *   quot = (cyc >> time_shift);
476	 *   rem = cyc & ((1 << time_shift) - 1);
477	 *   delta = time_offset + quot * time_mult +
478	 *              ((rem * time_mult) >> time_shift);
479	 *
480	 * Where time_offset,time_mult,time_shift and cyc are read in the
481	 * seqcount loop described above. This delta can then be added to
482	 * enabled and possible running (if index), improving the scaling:
483	 *
484	 *   enabled += delta;
485	 *   if (index)
486	 *     running += delta;
487	 *
488	 *   quot = count / running;
489	 *   rem  = count % running;
490	 *   count = quot * enabled + (rem * enabled) / running;
491	 */
492	__u16	time_shift;
493	__u32	time_mult;
494	__u64	time_offset;
495	/*
496	 * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated
497	 * from sample timestamps.
498	 *
499	 *   time = timestamp - time_zero;
500	 *   quot = time / time_mult;
501	 *   rem  = time % time_mult;
502	 *   cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
503	 *
504	 * And vice versa:
505	 *
506	 *   quot = cyc >> time_shift;
507	 *   rem  = cyc & ((1 << time_shift) - 1);
508	 *   timestamp = time_zero + quot * time_mult +
509	 *               ((rem * time_mult) >> time_shift);
510	 */
511	__u64	time_zero;
512	__u32	size;			/* Header size up to __reserved[] fields. */
513
514		/*
515		 * Hole for extension of the self monitor capabilities
516		 */
517
518	__u8	__reserved[118*8+4];	/* align to 1k. */
519
520	/*
521	 * Control data for the mmap() data buffer.
522	 *
523	 * User-space reading the @data_head value should issue an smp_rmb(),
524	 * after reading this value.
525	 *
526	 * When the mapping is PROT_WRITE the @data_tail value should be
527	 * written by userspace to reflect the last read data, after issueing
528	 * an smp_mb() to separate the data read from the ->data_tail store.
529	 * In this case the kernel will not over-write unread data.
530	 *
531	 * See perf_output_put_handle() for the data ordering.
532	 *
533	 * data_{offset,size} indicate the location and size of the perf record
534	 * buffer within the mmapped area.
535	 */
536	__u64   data_head;		/* head in the data section */
537	__u64	data_tail;		/* user-space written tail */
538	__u64	data_offset;		/* where the buffer starts */
539	__u64	data_size;		/* data buffer size */
540
541	/*
542	 * AUX area is defined by aux_{offset,size} fields that should be set
543	 * by the userspace, so that
544	 *
545	 *   aux_offset >= data_offset + data_size
546	 *
547	 * prior to mmap()ing it. Size of the mmap()ed area should be aux_size.
548	 *
549	 * Ring buffer pointers aux_{head,tail} have the same semantics as
550	 * data_{head,tail} and same ordering rules apply.
551	 */
552	__u64	aux_head;
553	__u64	aux_tail;
554	__u64	aux_offset;
555	__u64	aux_size;
556};
557
558#define PERF_RECORD_MISC_CPUMODE_MASK		(7 << 0)
559#define PERF_RECORD_MISC_CPUMODE_UNKNOWN	(0 << 0)
560#define PERF_RECORD_MISC_KERNEL			(1 << 0)
561#define PERF_RECORD_MISC_USER			(2 << 0)
562#define PERF_RECORD_MISC_HYPERVISOR		(3 << 0)
563#define PERF_RECORD_MISC_GUEST_KERNEL		(4 << 0)
564#define PERF_RECORD_MISC_GUEST_USER		(5 << 0)
565
566/*
567 * PERF_RECORD_MISC_MMAP_DATA and PERF_RECORD_MISC_COMM_EXEC are used on
568 * different events so can reuse the same bit position.
569 */
570#define PERF_RECORD_MISC_MMAP_DATA		(1 << 13)
571#define PERF_RECORD_MISC_COMM_EXEC		(1 << 13)
572/*
573 * Indicates that the content of PERF_SAMPLE_IP points to
574 * the actual instruction that triggered the event. See also
575 * perf_event_attr::precise_ip.
576 */
577#define PERF_RECORD_MISC_EXACT_IP		(1 << 14)
578/*
579 * Reserve the last bit to indicate some extended misc field
580 */
581#define PERF_RECORD_MISC_EXT_RESERVED		(1 << 15)
582
583struct perf_event_header {
584	__u32	type;
585	__u16	misc;
586	__u16	size;
587};
588
589enum perf_event_type {
590
591	/*
592	 * If perf_event_attr.sample_id_all is set then all event types will
593	 * have the sample_type selected fields related to where/when
594	 * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU,
595	 * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed
596	 * just after the perf_event_header and the fields already present for
597	 * the existing fields, i.e. at the end of the payload. That way a newer
598	 * perf.data file will be supported by older perf tools, with these new
599	 * optional fields being ignored.
600	 *
601	 * struct sample_id {
602	 * 	{ u32			pid, tid; } && PERF_SAMPLE_TID
603	 * 	{ u64			time;     } && PERF_SAMPLE_TIME
604	 * 	{ u64			id;       } && PERF_SAMPLE_ID
605	 * 	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
606	 * 	{ u32			cpu, res; } && PERF_SAMPLE_CPU
607	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
608	 * } && perf_event_attr::sample_id_all
609	 *
610	 * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.  The
611	 * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed
612	 * relative to header.size.
613	 */
614
615	/*
616	 * The MMAP events record the PROT_EXEC mappings so that we can
617	 * correlate userspace IPs to code. They have the following structure:
618	 *
619	 * struct {
620	 *	struct perf_event_header	header;
621	 *
622	 *	u32				pid, tid;
623	 *	u64				addr;
624	 *	u64				len;
625	 *	u64				pgoff;
626	 *	char				filename[];
627	 * 	struct sample_id		sample_id;
628	 * };
629	 */
630	PERF_RECORD_MMAP			= 1,
631
632	/*
633	 * struct {
634	 *	struct perf_event_header	header;
635	 *	u64				id;
636	 *	u64				lost;
637	 * 	struct sample_id		sample_id;
638	 * };
639	 */
640	PERF_RECORD_LOST			= 2,
641
642	/*
643	 * struct {
644	 *	struct perf_event_header	header;
645	 *
646	 *	u32				pid, tid;
647	 *	char				comm[];
648	 * 	struct sample_id		sample_id;
649	 * };
650	 */
651	PERF_RECORD_COMM			= 3,
652
653	/*
654	 * struct {
655	 *	struct perf_event_header	header;
656	 *	u32				pid, ppid;
657	 *	u32				tid, ptid;
658	 *	u64				time;
659	 * 	struct sample_id		sample_id;
660	 * };
661	 */
662	PERF_RECORD_EXIT			= 4,
663
664	/*
665	 * struct {
666	 *	struct perf_event_header	header;
667	 *	u64				time;
668	 *	u64				id;
669	 *	u64				stream_id;
670	 * 	struct sample_id		sample_id;
671	 * };
672	 */
673	PERF_RECORD_THROTTLE			= 5,
674	PERF_RECORD_UNTHROTTLE			= 6,
675
676	/*
677	 * struct {
678	 *	struct perf_event_header	header;
679	 *	u32				pid, ppid;
680	 *	u32				tid, ptid;
681	 *	u64				time;
682	 * 	struct sample_id		sample_id;
683	 * };
684	 */
685	PERF_RECORD_FORK			= 7,
686
687	/*
688	 * struct {
689	 *	struct perf_event_header	header;
690	 *	u32				pid, tid;
691	 *
692	 *	struct read_format		values;
693	 * 	struct sample_id		sample_id;
694	 * };
695	 */
696	PERF_RECORD_READ			= 8,
697
698	/*
699	 * struct {
700	 *	struct perf_event_header	header;
701	 *
702	 *	#
703	 *	# Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.
704	 *	# The advantage of PERF_SAMPLE_IDENTIFIER is that its position
705	 *	# is fixed relative to header.
706	 *	#
707	 *
708	 *	{ u64			id;	  } && PERF_SAMPLE_IDENTIFIER
709	 *	{ u64			ip;	  } && PERF_SAMPLE_IP
710	 *	{ u32			pid, tid; } && PERF_SAMPLE_TID
711	 *	{ u64			time;     } && PERF_SAMPLE_TIME
712	 *	{ u64			addr;     } && PERF_SAMPLE_ADDR
713	 *	{ u64			id;	  } && PERF_SAMPLE_ID
714	 *	{ u64			stream_id;} && PERF_SAMPLE_STREAM_ID
715	 *	{ u32			cpu, res; } && PERF_SAMPLE_CPU
716	 *	{ u64			period;   } && PERF_SAMPLE_PERIOD
717	 *
718	 *	{ struct read_format	values;	  } && PERF_SAMPLE_READ
719	 *
720	 *	{ u64			nr,
721	 *	  u64			ips[nr];  } && PERF_SAMPLE_CALLCHAIN
722	 *
723	 *	#
724	 *	# The RAW record below is opaque data wrt the ABI
725	 *	#
726	 *	# That is, the ABI doesn't make any promises wrt to
727	 *	# the stability of its content, it may vary depending
728	 *	# on event, hardware, kernel version and phase of
729	 *	# the moon.
730	 *	#
731	 *	# In other words, PERF_SAMPLE_RAW contents are not an ABI.
732	 *	#
733	 *
734	 *	{ u32			size;
735	 *	  char                  data[size];}&& PERF_SAMPLE_RAW
736	 *
737	 *	{ u64                   nr;
738	 *        { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK
739	 *
740	 * 	{ u64			abi; # enum perf_sample_regs_abi
741	 * 	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
742	 *
743	 * 	{ u64			size;
744	 * 	  char			data[size];
745	 * 	  u64			dyn_size; } && PERF_SAMPLE_STACK_USER
746	 *
747	 *	{ u64			weight;   } && PERF_SAMPLE_WEIGHT
748	 *	{ u64			data_src; } && PERF_SAMPLE_DATA_SRC
749	 *	{ u64			transaction; } && PERF_SAMPLE_TRANSACTION
750	 *	{ u64			abi; # enum perf_sample_regs_abi
751	 *	  u64			regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
752	 * };
753	 */
754	PERF_RECORD_SAMPLE			= 9,
755
756	/*
757	 * The MMAP2 records are an augmented version of MMAP, they add
758	 * maj, min, ino numbers to be used to uniquely identify each mapping
759	 *
760	 * struct {
761	 *	struct perf_event_header	header;
762	 *
763	 *	u32				pid, tid;
764	 *	u64				addr;
765	 *	u64				len;
766	 *	u64				pgoff;
767	 *	u32				maj;
768	 *	u32				min;
769	 *	u64				ino;
770	 *	u64				ino_generation;
771	 *	u32				prot, flags;
772	 *	char				filename[];
773	 * 	struct sample_id		sample_id;
774	 * };
775	 */
776	PERF_RECORD_MMAP2			= 10,
777
778	/*
779	 * Records that new data landed in the AUX buffer part.
780	 *
781	 * struct {
782	 * 	struct perf_event_header	header;
783	 *
784	 * 	u64				aux_offset;
785	 * 	u64				aux_size;
786	 *	u64				flags;
787	 * 	struct sample_id		sample_id;
788	 * };
789	 */
790	PERF_RECORD_AUX				= 11,
791
792	/*
793	 * Indicates that instruction trace has started
794	 *
795	 * struct {
796	 *	struct perf_event_header	header;
797	 *	u32				pid;
798	 *	u32				tid;
799	 * };
800	 */
801	PERF_RECORD_ITRACE_START		= 12,
802
803	PERF_RECORD_MAX,			/* non-ABI */
804};
805
806#define PERF_MAX_STACK_DEPTH		127
807
808enum perf_callchain_context {
809	PERF_CONTEXT_HV			= (__u64)-32,
810	PERF_CONTEXT_KERNEL		= (__u64)-128,
811	PERF_CONTEXT_USER		= (__u64)-512,
812
813	PERF_CONTEXT_GUEST		= (__u64)-2048,
814	PERF_CONTEXT_GUEST_KERNEL	= (__u64)-2176,
815	PERF_CONTEXT_GUEST_USER		= (__u64)-2560,
816
817	PERF_CONTEXT_MAX		= (__u64)-4095,
818};
819
820/**
821 * PERF_RECORD_AUX::flags bits
822 */
823#define PERF_AUX_FLAG_TRUNCATED		0x01	/* record was truncated to fit */
824#define PERF_AUX_FLAG_OVERWRITE		0x02	/* snapshot from overwrite mode */
825
826#define PERF_FLAG_FD_NO_GROUP		(1UL << 0)
827#define PERF_FLAG_FD_OUTPUT		(1UL << 1)
828#define PERF_FLAG_PID_CGROUP		(1UL << 2) /* pid=cgroup id, per-cpu mode only */
829#define PERF_FLAG_FD_CLOEXEC		(1UL << 3) /* O_CLOEXEC */
830
831union perf_mem_data_src {
832	__u64 val;
833	struct {
834		__u64   mem_op:5,	/* type of opcode */
835			mem_lvl:14,	/* memory hierarchy level */
836			mem_snoop:5,	/* snoop mode */
837			mem_lock:2,	/* lock instr */
838			mem_dtlb:7,	/* tlb access */
839			mem_rsvd:31;
840	};
841};
842
843/* type of opcode (load/store/prefetch,code) */
844#define PERF_MEM_OP_NA		0x01 /* not available */
845#define PERF_MEM_OP_LOAD	0x02 /* load instruction */
846#define PERF_MEM_OP_STORE	0x04 /* store instruction */
847#define PERF_MEM_OP_PFETCH	0x08 /* prefetch */
848#define PERF_MEM_OP_EXEC	0x10 /* code (execution) */
849#define PERF_MEM_OP_SHIFT	0
850
851/* memory hierarchy (memory level, hit or miss) */
852#define PERF_MEM_LVL_NA		0x01  /* not available */
853#define PERF_MEM_LVL_HIT	0x02  /* hit level */
854#define PERF_MEM_LVL_MISS	0x04  /* miss level  */
855#define PERF_MEM_LVL_L1		0x08  /* L1 */
856#define PERF_MEM_LVL_LFB	0x10  /* Line Fill Buffer */
857#define PERF_MEM_LVL_L2		0x20  /* L2 */
858#define PERF_MEM_LVL_L3		0x40  /* L3 */
859#define PERF_MEM_LVL_LOC_RAM	0x80  /* Local DRAM */
860#define PERF_MEM_LVL_REM_RAM1	0x100 /* Remote DRAM (1 hop) */
861#define PERF_MEM_LVL_REM_RAM2	0x200 /* Remote DRAM (2 hops) */
862#define PERF_MEM_LVL_REM_CCE1	0x400 /* Remote Cache (1 hop) */
863#define PERF_MEM_LVL_REM_CCE2	0x800 /* Remote Cache (2 hops) */
864#define PERF_MEM_LVL_IO		0x1000 /* I/O memory */
865#define PERF_MEM_LVL_UNC	0x2000 /* Uncached memory */
866#define PERF_MEM_LVL_SHIFT	5
867
868/* snoop mode */
869#define PERF_MEM_SNOOP_NA	0x01 /* not available */
870#define PERF_MEM_SNOOP_NONE	0x02 /* no snoop */
871#define PERF_MEM_SNOOP_HIT	0x04 /* snoop hit */
872#define PERF_MEM_SNOOP_MISS	0x08 /* snoop miss */
873#define PERF_MEM_SNOOP_HITM	0x10 /* snoop hit modified */
874#define PERF_MEM_SNOOP_SHIFT	19
875
876/* locked instruction */
877#define PERF_MEM_LOCK_NA	0x01 /* not available */
878#define PERF_MEM_LOCK_LOCKED	0x02 /* locked transaction */
879#define PERF_MEM_LOCK_SHIFT	24
880
881/* TLB access */
882#define PERF_MEM_TLB_NA		0x01 /* not available */
883#define PERF_MEM_TLB_HIT	0x02 /* hit level */
884#define PERF_MEM_TLB_MISS	0x04 /* miss level */
885#define PERF_MEM_TLB_L1		0x08 /* L1 */
886#define PERF_MEM_TLB_L2		0x10 /* L2 */
887#define PERF_MEM_TLB_WK		0x20 /* Hardware Walker*/
888#define PERF_MEM_TLB_OS		0x40 /* OS fault handler */
889#define PERF_MEM_TLB_SHIFT	26
890
891#define PERF_MEM_S(a, s) \
892	(((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)
893
894/*
895 * single taken branch record layout:
896 *
897 *      from: source instruction (may not always be a branch insn)
898 *        to: branch target
899 *   mispred: branch target was mispredicted
900 * predicted: branch target was predicted
901 *
902 * support for mispred, predicted is optional. In case it
903 * is not supported mispred = predicted = 0.
904 *
905 *     in_tx: running in a hardware transaction
906 *     abort: aborting a hardware transaction
907 */
908struct perf_branch_entry {
909	__u64	from;
910	__u64	to;
911	__u64	mispred:1,  /* target mispredicted */
912		predicted:1,/* target predicted */
913		in_tx:1,    /* in transaction */
914		abort:1,    /* transaction abort */
915		reserved:60;
916};
917
918#endif /* _UAPI_LINUX_PERF_EVENT_H */
919