1 /* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
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
9 * Data type definitions, declarations, prototypes.
11 * Started by: Thomas Gleixner and Ingo Molnar
13 * For licencing details see kernel-base/COPYING
15 #ifndef _UAPI_LINUX_PERF_EVENT_H
16 #define _UAPI_LINUX_PERF_EVENT_H
18 #include <linux/types.h>
19 #include <linux/ioctl.h>
20 #include <asm/byteorder.h>
23 * User-space ABI bits:
30 PERF_TYPE_HARDWARE = 0,
31 PERF_TYPE_SOFTWARE = 1,
32 PERF_TYPE_TRACEPOINT = 2,
33 PERF_TYPE_HW_CACHE = 3,
35 PERF_TYPE_BREAKPOINT = 5,
37 PERF_TYPE_MAX, /* non-ABI */
41 * Generalized performance event event_id types, used by the
42 * attr.event_id parameter of the sys_perf_event_open()
47 * Common hardware events, generalized by the kernel:
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,
60 PERF_COUNT_HW_MAX, /* non-ABI */
64 * Generalized hardware cache events:
66 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
67 * { read, write, prefetch } x
68 * { accesses, misses }
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,
79 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */
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,
87 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */
90 enum perf_hw_cache_op_result_id {
91 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0,
92 PERF_COUNT_HW_CACHE_RESULT_MISS = 1,
94 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */
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
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,
116 PERF_COUNT_SW_MAX, /* non-ABI */
120 * Bits that can be set in attr.sample_type to request information
121 * in the overflow packets.
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,
147 PERF_SAMPLE_MAX = 1U << 22, /* non-ABI */
149 __PERF_SAMPLE_CALLCHAIN_EARLY = 1ULL << 63, /* non-ABI; internal use */
153 * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
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.
159 * The branch types can be combined, however BRANCH_ANY covers all types
160 * of branches and therefore it supersedes all the other types.
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 */
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 */
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 */
180 PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT = 14, /* no flags */
181 PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT = 15, /* no cycles */
183 PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT = 16, /* save branch type */
185 PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT = 17, /* save low level index of raw branch records */
187 PERF_SAMPLE_BRANCH_MAX_SHIFT /* non-ABI */
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,
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,
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,
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,
211 PERF_SAMPLE_BRANCH_TYPE_SAVE =
212 1U << PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT,
214 PERF_SAMPLE_BRANCH_HW_INDEX = 1U << PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT,
216 PERF_SAMPLE_BRANCH_MAX = 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT,
220 * Common flow change classification
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 */
237 #define PERF_SAMPLE_BRANCH_PLM_ALL \
238 (PERF_SAMPLE_BRANCH_USER|\
239 PERF_SAMPLE_BRANCH_KERNEL|\
240 PERF_SAMPLE_BRANCH_HV)
243 * Values to determine ABI of the registers dump.
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,
252 * Values for the memory transaction event qualifier, mostly for
253 * abort events. Multiple bits can be set.
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 */
265 PERF_TXN_MAX = (1 << 8), /* non-ABI */
267 /* bits 32..63 are reserved for the abort code */
269 PERF_TXN_ABORT_MASK = (0xffffffffULL << 32),
270 PERF_TXN_ABORT_SHIFT = 32,
274 * The format of the data returned by read() on a perf event fd,
275 * as specified by attr.read_format:
277 * struct read_format {
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
285 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
286 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
288 * { u64 id; } && PERF_FORMAT_ID
290 * } && PERF_FORMAT_GROUP
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,
299 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */
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 */
312 * Hardware event_id to monitor via a performance monitoring event:
314 * @sample_max_stack: Max number of frame pointers in a callchain,
315 * should be < /proc/sys/kernel/perf_event_max_stack
317 struct perf_event_attr {
320 * Major type: hardware/software/tracepoint/etc.
325 * Size of the attr structure, for fwd/bwd compat.
330 * Type specific configuration information.
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 */
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
365 * See also PERF_RECORD_MISC_EXACT_IP
367 precise_ip : 2, /* skid constraint */
368 mmap_data : 1, /* non-exec mmap data */
369 sample_id_all : 1, /* sample_type all events */
371 exclude_host : 1, /* don't count in host */
372 exclude_guest : 1, /* don't count in guest */
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 text_poke : 1, /* include text poke events */
390 __u32 wakeup_events; /* wakeup every n events */
391 __u32 wakeup_watermark; /* bytes before wakeup */
397 __u64 kprobe_func; /* for perf_kprobe */
398 __u64 uprobe_path; /* for perf_uprobe */
399 __u64 config1; /* extension of config */
403 __u64 kprobe_addr; /* when kprobe_func == NULL */
404 __u64 probe_offset; /* for perf_[k,u]probe */
405 __u64 config2; /* extension of config1 */
407 __u64 branch_sample_type; /* enum perf_branch_sample_type */
410 * Defines set of user regs to dump on samples.
411 * See asm/perf_regs.h for details.
413 __u64 sample_regs_user;
416 * Defines size of the user stack to dump on samples.
418 __u32 sample_stack_user;
422 * Defines set of regs to dump for each sample
424 * - precise = 0: PMU interrupt
425 * - precise > 0: sampled instruction
427 * See asm/perf_regs.h for details.
429 __u64 sample_regs_intr;
432 * Wakeup watermark for AUX area
435 __u16 sample_max_stack;
437 __u32 aux_sample_size;
442 * Structure used by below PERF_EVENT_IOC_QUERY_BPF command
443 * to query bpf programs attached to the same perf tracepoint
444 * as the given perf event.
446 struct perf_event_query_bpf {
448 * The below ids array length
452 * Set by the kernel to indicate the number of
457 * User provided buffer to store program ids
463 * Ioctls that can be done on a perf event fd:
465 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
466 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
467 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
468 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
469 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64)
470 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
471 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
472 #define PERF_EVENT_IOC_ID _IOR('$', 7, __u64 *)
473 #define PERF_EVENT_IOC_SET_BPF _IOW('$', 8, __u32)
474 #define PERF_EVENT_IOC_PAUSE_OUTPUT _IOW('$', 9, __u32)
475 #define PERF_EVENT_IOC_QUERY_BPF _IOWR('$', 10, struct perf_event_query_bpf *)
476 #define PERF_EVENT_IOC_MODIFY_ATTRIBUTES _IOW('$', 11, struct perf_event_attr *)
478 enum perf_event_ioc_flags {
479 PERF_IOC_FLAG_GROUP = 1U << 0,
483 * Structure of the page that can be mapped via mmap
485 struct perf_event_mmap_page {
486 __u32 version; /* version number of this structure */
487 __u32 compat_version; /* lowest version this is compat with */
490 * Bits needed to read the hw events in user-space.
492 * u32 seq, time_mult, time_shift, index, width;
493 * u64 count, enabled, running;
494 * u64 cyc, time_offset;
501 * enabled = pc->time_enabled;
502 * running = pc->time_running;
504 * if (pc->cap_usr_time && enabled != running) {
506 * time_offset = pc->time_offset;
507 * time_mult = pc->time_mult;
508 * time_shift = pc->time_shift;
512 * count = pc->offset;
513 * if (pc->cap_user_rdpmc && index) {
514 * width = pc->pmc_width;
515 * pmc = rdpmc(index - 1);
519 * } while (pc->lock != seq);
521 * NOTE: for obvious reason this only works on self-monitoring
524 __u32 lock; /* seqlock for synchronization */
525 __u32 index; /* hardware event identifier */
526 __s64 offset; /* add to hardware event value */
527 __u64 time_enabled; /* time event active */
528 __u64 time_running; /* time event on cpu */
532 __u64 cap_bit0 : 1, /* Always 0, deprecated, see commit 860f085b74e9 */
533 cap_bit0_is_deprecated : 1, /* Always 1, signals that bit 0 is zero */
535 cap_user_rdpmc : 1, /* The RDPMC instruction can be used to read counts */
536 cap_user_time : 1, /* The time_{shift,mult,offset} fields are used */
537 cap_user_time_zero : 1, /* The time_zero field is used */
538 cap_user_time_short : 1, /* the time_{cycle,mask} fields are used */
544 * If cap_user_rdpmc this field provides the bit-width of the value
545 * read using the rdpmc() or equivalent instruction. This can be used
546 * to sign extend the result like:
548 * pmc <<= 64 - width;
549 * pmc >>= 64 - width; // signed shift right
555 * If cap_usr_time the below fields can be used to compute the time
556 * delta since time_enabled (in ns) using rdtsc or similar.
561 * quot = (cyc >> time_shift);
562 * rem = cyc & (((u64)1 << time_shift) - 1);
563 * delta = time_offset + quot * time_mult +
564 * ((rem * time_mult) >> time_shift);
566 * Where time_offset,time_mult,time_shift and cyc are read in the
567 * seqcount loop described above. This delta can then be added to
568 * enabled and possible running (if index), improving the scaling:
574 * quot = count / running;
575 * rem = count % running;
576 * count = quot * enabled + (rem * enabled) / running;
582 * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated
583 * from sample timestamps.
585 * time = timestamp - time_zero;
586 * quot = time / time_mult;
587 * rem = time % time_mult;
588 * cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
592 * quot = cyc >> time_shift;
593 * rem = cyc & (((u64)1 << time_shift) - 1);
594 * timestamp = time_zero + quot * time_mult +
595 * ((rem * time_mult) >> time_shift);
599 __u32 size; /* Header size up to __reserved[] fields. */
603 * If cap_usr_time_short, the hardware clock is less than 64bit wide
604 * and we must compute the 'cyc' value, as used by cap_usr_time, as:
606 * cyc = time_cycles + ((cyc - time_cycles) & time_mask)
608 * NOTE: this form is explicitly chosen such that cap_usr_time_short
609 * is a correction on top of cap_usr_time, and code that doesn't
610 * know about cap_usr_time_short still works under the assumption
611 * the counter doesn't wrap.
617 * Hole for extension of the self monitor capabilities
620 __u8 __reserved[116*8]; /* align to 1k. */
623 * Control data for the mmap() data buffer.
625 * User-space reading the @data_head value should issue an smp_rmb(),
626 * after reading this value.
628 * When the mapping is PROT_WRITE the @data_tail value should be
629 * written by userspace to reflect the last read data, after issueing
630 * an smp_mb() to separate the data read from the ->data_tail store.
631 * In this case the kernel will not over-write unread data.
633 * See perf_output_put_handle() for the data ordering.
635 * data_{offset,size} indicate the location and size of the perf record
636 * buffer within the mmapped area.
638 __u64 data_head; /* head in the data section */
639 __u64 data_tail; /* user-space written tail */
640 __u64 data_offset; /* where the buffer starts */
641 __u64 data_size; /* data buffer size */
644 * AUX area is defined by aux_{offset,size} fields that should be set
645 * by the userspace, so that
647 * aux_offset >= data_offset + data_size
649 * prior to mmap()ing it. Size of the mmap()ed area should be aux_size.
651 * Ring buffer pointers aux_{head,tail} have the same semantics as
652 * data_{head,tail} and same ordering rules apply.
660 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0)
661 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
662 #define PERF_RECORD_MISC_KERNEL (1 << 0)
663 #define PERF_RECORD_MISC_USER (2 << 0)
664 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
665 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0)
666 #define PERF_RECORD_MISC_GUEST_USER (5 << 0)
669 * Indicates that /proc/PID/maps parsing are truncated by time out.
671 #define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT (1 << 12)
673 * Following PERF_RECORD_MISC_* are used on different
674 * events, so can reuse the same bit position:
676 * PERF_RECORD_MISC_MMAP_DATA - PERF_RECORD_MMAP* events
677 * PERF_RECORD_MISC_COMM_EXEC - PERF_RECORD_COMM event
678 * PERF_RECORD_MISC_FORK_EXEC - PERF_RECORD_FORK event (perf internal)
679 * PERF_RECORD_MISC_SWITCH_OUT - PERF_RECORD_SWITCH* events
681 #define PERF_RECORD_MISC_MMAP_DATA (1 << 13)
682 #define PERF_RECORD_MISC_COMM_EXEC (1 << 13)
683 #define PERF_RECORD_MISC_FORK_EXEC (1 << 13)
684 #define PERF_RECORD_MISC_SWITCH_OUT (1 << 13)
686 * These PERF_RECORD_MISC_* flags below are safely reused
687 * for the following events:
689 * PERF_RECORD_MISC_EXACT_IP - PERF_RECORD_SAMPLE of precise events
690 * PERF_RECORD_MISC_SWITCH_OUT_PREEMPT - PERF_RECORD_SWITCH* events
693 * PERF_RECORD_MISC_EXACT_IP:
694 * Indicates that the content of PERF_SAMPLE_IP points to
695 * the actual instruction that triggered the event. See also
696 * perf_event_attr::precise_ip.
698 * PERF_RECORD_MISC_SWITCH_OUT_PREEMPT:
699 * Indicates that thread was preempted in TASK_RUNNING state.
701 #define PERF_RECORD_MISC_EXACT_IP (1 << 14)
702 #define PERF_RECORD_MISC_SWITCH_OUT_PREEMPT (1 << 14)
704 * Reserve the last bit to indicate some extended misc field
706 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15)
708 struct perf_event_header {
714 struct perf_ns_link_info {
728 NR_NAMESPACES, /* number of available namespaces */
731 enum perf_event_type {
734 * If perf_event_attr.sample_id_all is set then all event types will
735 * have the sample_type selected fields related to where/when
736 * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU,
737 * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed
738 * just after the perf_event_header and the fields already present for
739 * the existing fields, i.e. at the end of the payload. That way a newer
740 * perf.data file will be supported by older perf tools, with these new
741 * optional fields being ignored.
744 * { u32 pid, tid; } && PERF_SAMPLE_TID
745 * { u64 time; } && PERF_SAMPLE_TIME
746 * { u64 id; } && PERF_SAMPLE_ID
747 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
748 * { u32 cpu, res; } && PERF_SAMPLE_CPU
749 * { u64 id; } && PERF_SAMPLE_IDENTIFIER
750 * } && perf_event_attr::sample_id_all
752 * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID. The
753 * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed
754 * relative to header.size.
758 * The MMAP events record the PROT_EXEC mappings so that we can
759 * correlate userspace IPs to code. They have the following structure:
762 * struct perf_event_header header;
769 * struct sample_id sample_id;
772 PERF_RECORD_MMAP = 1,
776 * struct perf_event_header header;
779 * struct sample_id sample_id;
782 PERF_RECORD_LOST = 2,
786 * struct perf_event_header header;
790 * struct sample_id sample_id;
793 PERF_RECORD_COMM = 3,
797 * struct perf_event_header header;
801 * struct sample_id sample_id;
804 PERF_RECORD_EXIT = 4,
808 * struct perf_event_header header;
812 * struct sample_id sample_id;
815 PERF_RECORD_THROTTLE = 5,
816 PERF_RECORD_UNTHROTTLE = 6,
820 * struct perf_event_header header;
824 * struct sample_id sample_id;
827 PERF_RECORD_FORK = 7,
831 * struct perf_event_header header;
834 * struct read_format values;
835 * struct sample_id sample_id;
838 PERF_RECORD_READ = 8,
842 * struct perf_event_header header;
845 * # Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.
846 * # The advantage of PERF_SAMPLE_IDENTIFIER is that its position
847 * # is fixed relative to header.
850 * { u64 id; } && PERF_SAMPLE_IDENTIFIER
851 * { u64 ip; } && PERF_SAMPLE_IP
852 * { u32 pid, tid; } && PERF_SAMPLE_TID
853 * { u64 time; } && PERF_SAMPLE_TIME
854 * { u64 addr; } && PERF_SAMPLE_ADDR
855 * { u64 id; } && PERF_SAMPLE_ID
856 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
857 * { u32 cpu, res; } && PERF_SAMPLE_CPU
858 * { u64 period; } && PERF_SAMPLE_PERIOD
860 * { struct read_format values; } && PERF_SAMPLE_READ
863 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
866 * # The RAW record below is opaque data wrt the ABI
868 * # That is, the ABI doesn't make any promises wrt to
869 * # the stability of its content, it may vary depending
870 * # on event, hardware, kernel version and phase of
873 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
877 * char data[size];}&& PERF_SAMPLE_RAW
880 * { u64 hw_idx; } && PERF_SAMPLE_BRANCH_HW_INDEX
881 * { u64 from, to, flags } lbr[nr];
882 * } && PERF_SAMPLE_BRANCH_STACK
884 * { u64 abi; # enum perf_sample_regs_abi
885 * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
889 * u64 dyn_size; } && PERF_SAMPLE_STACK_USER
891 * { u64 weight; } && PERF_SAMPLE_WEIGHT
892 * { u64 data_src; } && PERF_SAMPLE_DATA_SRC
893 * { u64 transaction; } && PERF_SAMPLE_TRANSACTION
894 * { u64 abi; # enum perf_sample_regs_abi
895 * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
896 * { u64 phys_addr;} && PERF_SAMPLE_PHYS_ADDR
898 * char data[size]; } && PERF_SAMPLE_AUX
901 PERF_RECORD_SAMPLE = 9,
904 * The MMAP2 records are an augmented version of MMAP, they add
905 * maj, min, ino numbers to be used to uniquely identify each mapping
908 * struct perf_event_header header;
917 * u64 ino_generation;
920 * struct sample_id sample_id;
923 PERF_RECORD_MMAP2 = 10,
926 * Records that new data landed in the AUX buffer part.
929 * struct perf_event_header header;
934 * struct sample_id sample_id;
937 PERF_RECORD_AUX = 11,
940 * Indicates that instruction trace has started
943 * struct perf_event_header header;
946 * struct sample_id sample_id;
949 PERF_RECORD_ITRACE_START = 12,
952 * Records the dropped/lost sample number.
955 * struct perf_event_header header;
958 * struct sample_id sample_id;
961 PERF_RECORD_LOST_SAMPLES = 13,
964 * Records a context switch in or out (flagged by
965 * PERF_RECORD_MISC_SWITCH_OUT). See also
966 * PERF_RECORD_SWITCH_CPU_WIDE.
969 * struct perf_event_header header;
970 * struct sample_id sample_id;
973 PERF_RECORD_SWITCH = 14,
976 * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and
977 * next_prev_tid that are the next (switching out) or previous
978 * (switching in) pid/tid.
981 * struct perf_event_header header;
984 * struct sample_id sample_id;
987 PERF_RECORD_SWITCH_CPU_WIDE = 15,
991 * struct perf_event_header header;
995 * { u64 dev, inode; } [nr_namespaces];
996 * struct sample_id sample_id;
999 PERF_RECORD_NAMESPACES = 16,
1002 * Record ksymbol register/unregister events:
1005 * struct perf_event_header header;
1011 * struct sample_id sample_id;
1014 PERF_RECORD_KSYMBOL = 17,
1017 * Record bpf events:
1018 * enum perf_bpf_event_type {
1019 * PERF_BPF_EVENT_UNKNOWN = 0,
1020 * PERF_BPF_EVENT_PROG_LOAD = 1,
1021 * PERF_BPF_EVENT_PROG_UNLOAD = 2,
1025 * struct perf_event_header header;
1029 * u8 tag[BPF_TAG_SIZE];
1030 * struct sample_id sample_id;
1033 PERF_RECORD_BPF_EVENT = 18,
1037 * struct perf_event_header header;
1040 * struct sample_id sample_id;
1043 PERF_RECORD_CGROUP = 19,
1046 * Records changes to kernel text i.e. self-modified code. 'old_len' is
1047 * the number of old bytes, 'new_len' is the number of new bytes. Either
1048 * 'old_len' or 'new_len' may be zero to indicate, for example, the
1049 * addition or removal of a trampoline. 'bytes' contains the old bytes
1050 * followed immediately by the new bytes.
1053 * struct perf_event_header header;
1058 * struct sample_id sample_id;
1061 PERF_RECORD_TEXT_POKE = 20,
1063 PERF_RECORD_MAX, /* non-ABI */
1066 enum perf_record_ksymbol_type {
1067 PERF_RECORD_KSYMBOL_TYPE_UNKNOWN = 0,
1068 PERF_RECORD_KSYMBOL_TYPE_BPF = 1,
1070 * Out of line code such as kprobe-replaced instructions or optimized
1071 * kprobes or ftrace trampolines.
1073 PERF_RECORD_KSYMBOL_TYPE_OOL = 2,
1074 PERF_RECORD_KSYMBOL_TYPE_MAX /* non-ABI */
1077 #define PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER (1 << 0)
1079 enum perf_bpf_event_type {
1080 PERF_BPF_EVENT_UNKNOWN = 0,
1081 PERF_BPF_EVENT_PROG_LOAD = 1,
1082 PERF_BPF_EVENT_PROG_UNLOAD = 2,
1083 PERF_BPF_EVENT_MAX, /* non-ABI */
1086 #define PERF_MAX_STACK_DEPTH 127
1087 #define PERF_MAX_CONTEXTS_PER_STACK 8
1089 enum perf_callchain_context {
1090 PERF_CONTEXT_HV = (__u64)-32,
1091 PERF_CONTEXT_KERNEL = (__u64)-128,
1092 PERF_CONTEXT_USER = (__u64)-512,
1094 PERF_CONTEXT_GUEST = (__u64)-2048,
1095 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
1096 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
1098 PERF_CONTEXT_MAX = (__u64)-4095,
1102 * PERF_RECORD_AUX::flags bits
1104 #define PERF_AUX_FLAG_TRUNCATED 0x01 /* record was truncated to fit */
1105 #define PERF_AUX_FLAG_OVERWRITE 0x02 /* snapshot from overwrite mode */
1106 #define PERF_AUX_FLAG_PARTIAL 0x04 /* record contains gaps */
1107 #define PERF_AUX_FLAG_COLLISION 0x08 /* sample collided with another */
1109 #define PERF_FLAG_FD_NO_GROUP (1UL << 0)
1110 #define PERF_FLAG_FD_OUTPUT (1UL << 1)
1111 #define PERF_FLAG_PID_CGROUP (1UL << 2) /* pid=cgroup id, per-cpu mode only */
1112 #define PERF_FLAG_FD_CLOEXEC (1UL << 3) /* O_CLOEXEC */
1114 #if defined(__LITTLE_ENDIAN_BITFIELD)
1115 union perf_mem_data_src {
1118 __u64 mem_op:5, /* type of opcode */
1119 mem_lvl:14, /* memory hierarchy level */
1120 mem_snoop:5, /* snoop mode */
1121 mem_lock:2, /* lock instr */
1122 mem_dtlb:7, /* tlb access */
1123 mem_lvl_num:4, /* memory hierarchy level number */
1124 mem_remote:1, /* remote */
1125 mem_snoopx:2, /* snoop mode, ext */
1129 #elif defined(__BIG_ENDIAN_BITFIELD)
1130 union perf_mem_data_src {
1134 mem_snoopx:2, /* snoop mode, ext */
1135 mem_remote:1, /* remote */
1136 mem_lvl_num:4, /* memory hierarchy level number */
1137 mem_dtlb:7, /* tlb access */
1138 mem_lock:2, /* lock instr */
1139 mem_snoop:5, /* snoop mode */
1140 mem_lvl:14, /* memory hierarchy level */
1141 mem_op:5; /* type of opcode */
1145 #error "Unknown endianness"
1148 /* type of opcode (load/store/prefetch,code) */
1149 #define PERF_MEM_OP_NA 0x01 /* not available */
1150 #define PERF_MEM_OP_LOAD 0x02 /* load instruction */
1151 #define PERF_MEM_OP_STORE 0x04 /* store instruction */
1152 #define PERF_MEM_OP_PFETCH 0x08 /* prefetch */
1153 #define PERF_MEM_OP_EXEC 0x10 /* code (execution) */
1154 #define PERF_MEM_OP_SHIFT 0
1156 /* memory hierarchy (memory level, hit or miss) */
1157 #define PERF_MEM_LVL_NA 0x01 /* not available */
1158 #define PERF_MEM_LVL_HIT 0x02 /* hit level */
1159 #define PERF_MEM_LVL_MISS 0x04 /* miss level */
1160 #define PERF_MEM_LVL_L1 0x08 /* L1 */
1161 #define PERF_MEM_LVL_LFB 0x10 /* Line Fill Buffer */
1162 #define PERF_MEM_LVL_L2 0x20 /* L2 */
1163 #define PERF_MEM_LVL_L3 0x40 /* L3 */
1164 #define PERF_MEM_LVL_LOC_RAM 0x80 /* Local DRAM */
1165 #define PERF_MEM_LVL_REM_RAM1 0x100 /* Remote DRAM (1 hop) */
1166 #define PERF_MEM_LVL_REM_RAM2 0x200 /* Remote DRAM (2 hops) */
1167 #define PERF_MEM_LVL_REM_CCE1 0x400 /* Remote Cache (1 hop) */
1168 #define PERF_MEM_LVL_REM_CCE2 0x800 /* Remote Cache (2 hops) */
1169 #define PERF_MEM_LVL_IO 0x1000 /* I/O memory */
1170 #define PERF_MEM_LVL_UNC 0x2000 /* Uncached memory */
1171 #define PERF_MEM_LVL_SHIFT 5
1173 #define PERF_MEM_REMOTE_REMOTE 0x01 /* Remote */
1174 #define PERF_MEM_REMOTE_SHIFT 37
1176 #define PERF_MEM_LVLNUM_L1 0x01 /* L1 */
1177 #define PERF_MEM_LVLNUM_L2 0x02 /* L2 */
1178 #define PERF_MEM_LVLNUM_L3 0x03 /* L3 */
1179 #define PERF_MEM_LVLNUM_L4 0x04 /* L4 */
1180 /* 5-0xa available */
1181 #define PERF_MEM_LVLNUM_ANY_CACHE 0x0b /* Any cache */
1182 #define PERF_MEM_LVLNUM_LFB 0x0c /* LFB */
1183 #define PERF_MEM_LVLNUM_RAM 0x0d /* RAM */
1184 #define PERF_MEM_LVLNUM_PMEM 0x0e /* PMEM */
1185 #define PERF_MEM_LVLNUM_NA 0x0f /* N/A */
1187 #define PERF_MEM_LVLNUM_SHIFT 33
1190 #define PERF_MEM_SNOOP_NA 0x01 /* not available */
1191 #define PERF_MEM_SNOOP_NONE 0x02 /* no snoop */
1192 #define PERF_MEM_SNOOP_HIT 0x04 /* snoop hit */
1193 #define PERF_MEM_SNOOP_MISS 0x08 /* snoop miss */
1194 #define PERF_MEM_SNOOP_HITM 0x10 /* snoop hit modified */
1195 #define PERF_MEM_SNOOP_SHIFT 19
1197 #define PERF_MEM_SNOOPX_FWD 0x01 /* forward */
1199 #define PERF_MEM_SNOOPX_SHIFT 38
1201 /* locked instruction */
1202 #define PERF_MEM_LOCK_NA 0x01 /* not available */
1203 #define PERF_MEM_LOCK_LOCKED 0x02 /* locked transaction */
1204 #define PERF_MEM_LOCK_SHIFT 24
1207 #define PERF_MEM_TLB_NA 0x01 /* not available */
1208 #define PERF_MEM_TLB_HIT 0x02 /* hit level */
1209 #define PERF_MEM_TLB_MISS 0x04 /* miss level */
1210 #define PERF_MEM_TLB_L1 0x08 /* L1 */
1211 #define PERF_MEM_TLB_L2 0x10 /* L2 */
1212 #define PERF_MEM_TLB_WK 0x20 /* Hardware Walker*/
1213 #define PERF_MEM_TLB_OS 0x40 /* OS fault handler */
1214 #define PERF_MEM_TLB_SHIFT 26
1216 #define PERF_MEM_S(a, s) \
1217 (((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)
1220 * single taken branch record layout:
1222 * from: source instruction (may not always be a branch insn)
1224 * mispred: branch target was mispredicted
1225 * predicted: branch target was predicted
1227 * support for mispred, predicted is optional. In case it
1228 * is not supported mispred = predicted = 0.
1230 * in_tx: running in a hardware transaction
1231 * abort: aborting a hardware transaction
1232 * cycles: cycles from last branch (or 0 if not supported)
1235 struct perf_branch_entry {
1238 __u64 mispred:1, /* target mispredicted */
1239 predicted:1,/* target predicted */
1240 in_tx:1, /* in transaction */
1241 abort:1, /* transaction abort */
1242 cycles:16, /* cycle count to last branch */
1243 type:4, /* branch type */
1247 #endif /* _UAPI_LINUX_PERF_EVENT_H */