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
8 * Data type definitions, declarations, prototypes.
10 * Started by: Thomas Gleixner and Ingo Molnar
12 * For licencing details see kernel-base/COPYING
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
17 #include <linux/types.h>
18 #include <linux/ioctl.h>
19 #include <asm/byteorder.h>
22 * User-space ABI bits:
29 PERF_TYPE_HARDWARE = 0,
30 PERF_TYPE_SOFTWARE = 1,
31 PERF_TYPE_TRACEPOINT = 2,
32 PERF_TYPE_HW_CACHE = 3,
34 PERF_TYPE_BREAKPOINT = 5,
36 PERF_TYPE_MAX, /* non-ABI */
40 * Generalized performance event event_id types, used by the
41 * attr.event_id parameter of the sys_perf_event_open()
46 * Common hardware events, generalized by the kernel:
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,
59 PERF_COUNT_HW_MAX, /* non-ABI */
63 * Generalized hardware cache events:
65 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
66 * { read, write, prefetch } x
67 * { accesses, misses }
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,
78 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */
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,
86 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */
89 enum perf_hw_cache_op_result_id {
90 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0,
91 PERF_COUNT_HW_CACHE_RESULT_MISS = 1,
93 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */
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
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,
113 PERF_COUNT_SW_MAX, /* non-ABI */
117 * Bits that can be set in attr.sample_type to request information
118 * in the overflow packets.
120 enum perf_event_sample_format {
121 PERF_SAMPLE_IP = 1U << 0,
122 PERF_SAMPLE_TID = 1U << 1,
123 PERF_SAMPLE_TIME = 1U << 2,
124 PERF_SAMPLE_ADDR = 1U << 3,
125 PERF_SAMPLE_READ = 1U << 4,
126 PERF_SAMPLE_CALLCHAIN = 1U << 5,
127 PERF_SAMPLE_ID = 1U << 6,
128 PERF_SAMPLE_CPU = 1U << 7,
129 PERF_SAMPLE_PERIOD = 1U << 8,
130 PERF_SAMPLE_STREAM_ID = 1U << 9,
131 PERF_SAMPLE_RAW = 1U << 10,
132 PERF_SAMPLE_BRANCH_STACK = 1U << 11,
134 PERF_SAMPLE_MAX = 1U << 12, /* non-ABI */
138 * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
140 * If the user does not pass priv level information via branch_sample_type,
141 * the kernel uses the event's priv level. Branch and event priv levels do
142 * not have to match. Branch priv level is checked for permissions.
144 * The branch types can be combined, however BRANCH_ANY covers all types
145 * of branches and therefore it supersedes all the other types.
147 enum perf_branch_sample_type {
148 PERF_SAMPLE_BRANCH_USER = 1U << 0, /* user branches */
149 PERF_SAMPLE_BRANCH_KERNEL = 1U << 1, /* kernel branches */
150 PERF_SAMPLE_BRANCH_HV = 1U << 2, /* hypervisor branches */
152 PERF_SAMPLE_BRANCH_ANY = 1U << 3, /* any branch types */
153 PERF_SAMPLE_BRANCH_ANY_CALL = 1U << 4, /* any call branch */
154 PERF_SAMPLE_BRANCH_ANY_RETURN = 1U << 5, /* any return branch */
155 PERF_SAMPLE_BRANCH_IND_CALL = 1U << 6, /* indirect calls */
157 PERF_SAMPLE_BRANCH_MAX = 1U << 7, /* non-ABI */
160 #define PERF_SAMPLE_BRANCH_PLM_ALL \
161 (PERF_SAMPLE_BRANCH_USER|\
162 PERF_SAMPLE_BRANCH_KERNEL|\
163 PERF_SAMPLE_BRANCH_HV)
166 * The format of the data returned by read() on a perf event fd,
167 * as specified by attr.read_format:
169 * struct read_format {
171 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
172 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
173 * { u64 id; } && PERF_FORMAT_ID
174 * } && !PERF_FORMAT_GROUP
177 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
178 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
180 * { u64 id; } && PERF_FORMAT_ID
182 * } && PERF_FORMAT_GROUP
185 enum perf_event_read_format {
186 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
187 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
188 PERF_FORMAT_ID = 1U << 2,
189 PERF_FORMAT_GROUP = 1U << 3,
191 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */
194 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
195 #define PERF_ATTR_SIZE_VER1 72 /* add: config2 */
196 #define PERF_ATTR_SIZE_VER2 80 /* add: branch_sample_type */
199 * Hardware event_id to monitor via a performance monitoring event:
201 struct perf_event_attr {
204 * Major type: hardware/software/tracepoint/etc.
209 * Size of the attr structure, for fwd/bwd compat.
214 * Type specific configuration information.
226 __u64 disabled : 1, /* off by default */
227 inherit : 1, /* children inherit it */
228 pinned : 1, /* must always be on PMU */
229 exclusive : 1, /* only group on PMU */
230 exclude_user : 1, /* don't count user */
231 exclude_kernel : 1, /* ditto kernel */
232 exclude_hv : 1, /* ditto hypervisor */
233 exclude_idle : 1, /* don't count when idle */
234 mmap : 1, /* include mmap data */
235 comm : 1, /* include comm data */
236 freq : 1, /* use freq, not period */
237 inherit_stat : 1, /* per task counts */
238 enable_on_exec : 1, /* next exec enables */
239 task : 1, /* trace fork/exit */
240 watermark : 1, /* wakeup_watermark */
244 * 0 - SAMPLE_IP can have arbitrary skid
245 * 1 - SAMPLE_IP must have constant skid
246 * 2 - SAMPLE_IP requested to have 0 skid
247 * 3 - SAMPLE_IP must have 0 skid
249 * See also PERF_RECORD_MISC_EXACT_IP
251 precise_ip : 2, /* skid constraint */
252 mmap_data : 1, /* non-exec mmap data */
253 sample_id_all : 1, /* sample_type all events */
255 exclude_host : 1, /* don't count in host */
256 exclude_guest : 1, /* don't count in guest */
261 __u32 wakeup_events; /* wakeup every n events */
262 __u32 wakeup_watermark; /* bytes before wakeup */
268 __u64 config1; /* extension of config */
272 __u64 config2; /* extension of config1 */
274 __u64 branch_sample_type; /* enum branch_sample_type */
278 * Ioctls that can be done on a perf event fd:
280 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
281 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
282 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
283 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
284 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64)
285 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
286 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
288 enum perf_event_ioc_flags {
289 PERF_IOC_FLAG_GROUP = 1U << 0,
293 * Structure of the page that can be mapped via mmap
295 struct perf_event_mmap_page {
296 __u32 version; /* version number of this structure */
297 __u32 compat_version; /* lowest version this is compat with */
300 * Bits needed to read the hw events in user-space.
302 * u32 seq, time_mult, time_shift, idx, width;
303 * u64 count, enabled, running;
304 * u64 cyc, time_offset;
311 * enabled = pc->time_enabled;
312 * running = pc->time_running;
314 * if (pc->cap_usr_time && enabled != running) {
316 * time_offset = pc->time_offset;
317 * time_mult = pc->time_mult;
318 * time_shift = pc->time_shift;
322 * count = pc->offset;
323 * if (pc->cap_usr_rdpmc && idx) {
324 * width = pc->pmc_width;
325 * pmc = rdpmc(idx - 1);
329 * } while (pc->lock != seq);
331 * NOTE: for obvious reason this only works on self-monitoring
334 __u32 lock; /* seqlock for synchronization */
335 __u32 index; /* hardware event identifier */
336 __s64 offset; /* add to hardware event value */
337 __u64 time_enabled; /* time event active */
338 __u64 time_running; /* time event on cpu */
341 __u64 cap_usr_time : 1,
347 * If cap_usr_rdpmc this field provides the bit-width of the value
348 * read using the rdpmc() or equivalent instruction. This can be used
349 * to sign extend the result like:
351 * pmc <<= 64 - width;
352 * pmc >>= 64 - width; // signed shift right
358 * If cap_usr_time the below fields can be used to compute the time
359 * delta since time_enabled (in ns) using rdtsc or similar.
364 * quot = (cyc >> time_shift);
365 * rem = cyc & ((1 << time_shift) - 1);
366 * delta = time_offset + quot * time_mult +
367 * ((rem * time_mult) >> time_shift);
369 * Where time_offset,time_mult,time_shift and cyc are read in the
370 * seqcount loop described above. This delta can then be added to
371 * enabled and possible running (if idx), improving the scaling:
377 * quot = count / running;
378 * rem = count % running;
379 * count = quot * enabled + (rem * enabled) / running;
386 * Hole for extension of the self monitor capabilities
389 __u64 __reserved[120]; /* align to 1k */
392 * Control data for the mmap() data buffer.
394 * User-space reading the @data_head value should issue an rmb(), on
395 * SMP capable platforms, after reading this value -- see
396 * perf_event_wakeup().
398 * When the mapping is PROT_WRITE the @data_tail value should be
399 * written by userspace to reflect the last read data. In this case
400 * the kernel will not over-write unread data.
402 __u64 data_head; /* head in the data section */
403 __u64 data_tail; /* user-space written tail */
406 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0)
407 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
408 #define PERF_RECORD_MISC_KERNEL (1 << 0)
409 #define PERF_RECORD_MISC_USER (2 << 0)
410 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
411 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0)
412 #define PERF_RECORD_MISC_GUEST_USER (5 << 0)
415 * Indicates that the content of PERF_SAMPLE_IP points to
416 * the actual instruction that triggered the event. See also
417 * perf_event_attr::precise_ip.
419 #define PERF_RECORD_MISC_EXACT_IP (1 << 14)
421 * Reserve the last bit to indicate some extended misc field
423 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15)
425 struct perf_event_header {
431 enum perf_event_type {
434 * If perf_event_attr.sample_id_all is set then all event types will
435 * have the sample_type selected fields related to where/when
436 * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
437 * described in PERF_RECORD_SAMPLE below, it will be stashed just after
438 * the perf_event_header and the fields already present for the existing
439 * fields, i.e. at the end of the payload. That way a newer perf.data
440 * file will be supported by older perf tools, with these new optional
441 * fields being ignored.
443 * The MMAP events record the PROT_EXEC mappings so that we can
444 * correlate userspace IPs to code. They have the following structure:
447 * struct perf_event_header header;
456 PERF_RECORD_MMAP = 1,
460 * struct perf_event_header header;
465 PERF_RECORD_LOST = 2,
469 * struct perf_event_header header;
475 PERF_RECORD_COMM = 3,
479 * struct perf_event_header header;
485 PERF_RECORD_EXIT = 4,
489 * struct perf_event_header header;
495 PERF_RECORD_THROTTLE = 5,
496 PERF_RECORD_UNTHROTTLE = 6,
500 * struct perf_event_header header;
506 PERF_RECORD_FORK = 7,
510 * struct perf_event_header header;
513 * struct read_format values;
516 PERF_RECORD_READ = 8,
520 * struct perf_event_header header;
522 * { u64 ip; } && PERF_SAMPLE_IP
523 * { u32 pid, tid; } && PERF_SAMPLE_TID
524 * { u64 time; } && PERF_SAMPLE_TIME
525 * { u64 addr; } && PERF_SAMPLE_ADDR
526 * { u64 id; } && PERF_SAMPLE_ID
527 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
528 * { u32 cpu, res; } && PERF_SAMPLE_CPU
529 * { u64 period; } && PERF_SAMPLE_PERIOD
531 * { struct read_format values; } && PERF_SAMPLE_READ
534 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
537 * # The RAW record below is opaque data wrt the ABI
539 * # That is, the ABI doesn't make any promises wrt to
540 * # the stability of its content, it may vary depending
541 * # on event, hardware, kernel version and phase of
544 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
548 * char data[size];}&& PERF_SAMPLE_RAW
550 * { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK
553 PERF_RECORD_SAMPLE = 9,
555 PERF_RECORD_MAX, /* non-ABI */
558 enum perf_callchain_context {
559 PERF_CONTEXT_HV = (__u64)-32,
560 PERF_CONTEXT_KERNEL = (__u64)-128,
561 PERF_CONTEXT_USER = (__u64)-512,
563 PERF_CONTEXT_GUEST = (__u64)-2048,
564 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
565 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
567 PERF_CONTEXT_MAX = (__u64)-4095,
570 #define PERF_FLAG_FD_NO_GROUP (1U << 0)
571 #define PERF_FLAG_FD_OUTPUT (1U << 1)
572 #define PERF_FLAG_PID_CGROUP (1U << 2) /* pid=cgroup id, per-cpu mode only */
576 * Kernel-internal data types and definitions:
579 #ifdef CONFIG_PERF_EVENTS
580 # include <linux/cgroup.h>
581 # include <asm/perf_event.h>
582 # include <asm/local64.h>
585 struct perf_guest_info_callbacks {
586 int (*is_in_guest)(void);
587 int (*is_user_mode)(void);
588 unsigned long (*get_guest_ip)(void);
591 #ifdef CONFIG_HAVE_HW_BREAKPOINT
592 #include <asm/hw_breakpoint.h>
595 #include <linux/list.h>
596 #include <linux/mutex.h>
597 #include <linux/rculist.h>
598 #include <linux/rcupdate.h>
599 #include <linux/spinlock.h>
600 #include <linux/hrtimer.h>
601 #include <linux/fs.h>
602 #include <linux/pid_namespace.h>
603 #include <linux/workqueue.h>
604 #include <linux/ftrace.h>
605 #include <linux/cpu.h>
606 #include <linux/irq_work.h>
607 #include <linux/static_key.h>
608 #include <linux/atomic.h>
609 #include <linux/sysfs.h>
610 #include <asm/local.h>
612 #define PERF_MAX_STACK_DEPTH 255
614 struct perf_callchain_entry {
616 __u64 ip[PERF_MAX_STACK_DEPTH];
619 struct perf_raw_record {
625 * single taken branch record layout:
627 * from: source instruction (may not always be a branch insn)
629 * mispred: branch target was mispredicted
630 * predicted: branch target was predicted
632 * support for mispred, predicted is optional. In case it
633 * is not supported mispred = predicted = 0.
635 struct perf_branch_entry {
638 __u64 mispred:1, /* target mispredicted */
639 predicted:1,/* target predicted */
644 * branch stack layout:
645 * nr: number of taken branches stored in entries[]
647 * Note that nr can vary from sample to sample
648 * branches (to, from) are stored from most recent
649 * to least recent, i.e., entries[0] contains the most
652 struct perf_branch_stack {
654 struct perf_branch_entry entries[0];
660 * extra PMU register associated with an event
662 struct hw_perf_event_extra {
663 u64 config; /* register value */
664 unsigned int reg; /* register address or index */
665 int alloc; /* extra register already allocated */
666 int idx; /* index in shared_regs->regs[] */
670 * struct hw_perf_event - performance event hardware details:
672 struct hw_perf_event {
673 #ifdef CONFIG_PERF_EVENTS
675 struct { /* hardware */
678 unsigned long config_base;
679 unsigned long event_base;
683 struct hw_perf_event_extra extra_reg;
684 struct hw_perf_event_extra branch_reg;
686 struct { /* software */
687 struct hrtimer hrtimer;
689 #ifdef CONFIG_HAVE_HW_BREAKPOINT
690 struct { /* breakpoint */
691 struct arch_hw_breakpoint info;
692 struct list_head bp_list;
694 * Crufty hack to avoid the chicken and egg
695 * problem hw_breakpoint has with context
696 * creation and event initalization.
698 struct task_struct *bp_target;
703 local64_t prev_count;
706 local64_t period_left;
711 u64 freq_count_stamp;
716 * hw_perf_event::state flags
718 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
719 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
720 #define PERF_HES_ARCH 0x04
725 * Common implementation detail of pmu::{start,commit,cancel}_txn
727 #define PERF_EVENT_TXN 0x1
730 * struct pmu - generic performance monitoring unit
733 struct list_head entry;
736 const struct attribute_group **attr_groups;
740 int * __percpu pmu_disable_count;
741 struct perf_cpu_context * __percpu pmu_cpu_context;
745 * Fully disable/enable this PMU, can be used to protect from the PMI
746 * as well as for lazy/batch writing of the MSRs.
748 void (*pmu_enable) (struct pmu *pmu); /* optional */
749 void (*pmu_disable) (struct pmu *pmu); /* optional */
752 * Try and initialize the event for this PMU.
753 * Should return -ENOENT when the @event doesn't match this PMU.
755 int (*event_init) (struct perf_event *event);
757 #define PERF_EF_START 0x01 /* start the counter when adding */
758 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
759 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
762 * Adds/Removes a counter to/from the PMU, can be done inside
763 * a transaction, see the ->*_txn() methods.
765 int (*add) (struct perf_event *event, int flags);
766 void (*del) (struct perf_event *event, int flags);
769 * Starts/Stops a counter present on the PMU. The PMI handler
770 * should stop the counter when perf_event_overflow() returns
771 * !0. ->start() will be used to continue.
773 void (*start) (struct perf_event *event, int flags);
774 void (*stop) (struct perf_event *event, int flags);
777 * Updates the counter value of the event.
779 void (*read) (struct perf_event *event);
782 * Group events scheduling is treated as a transaction, add
783 * group events as a whole and perform one schedulability test.
784 * If the test fails, roll back the whole group
786 * Start the transaction, after this ->add() doesn't need to
787 * do schedulability tests.
789 void (*start_txn) (struct pmu *pmu); /* optional */
791 * If ->start_txn() disabled the ->add() schedulability test
792 * then ->commit_txn() is required to perform one. On success
793 * the transaction is closed. On error the transaction is kept
794 * open until ->cancel_txn() is called.
796 int (*commit_txn) (struct pmu *pmu); /* optional */
798 * Will cancel the transaction, assumes ->del() is called
799 * for each successful ->add() during the transaction.
801 void (*cancel_txn) (struct pmu *pmu); /* optional */
804 * Will return the value for perf_event_mmap_page::index for this event,
805 * if no implementation is provided it will default to: event->hw.idx + 1.
807 int (*event_idx) (struct perf_event *event); /*optional */
810 * flush branch stack on context-switches (needed in cpu-wide mode)
812 void (*flush_branch_stack) (void);
816 * enum perf_event_active_state - the states of a event
818 enum perf_event_active_state {
819 PERF_EVENT_STATE_ERROR = -2,
820 PERF_EVENT_STATE_OFF = -1,
821 PERF_EVENT_STATE_INACTIVE = 0,
822 PERF_EVENT_STATE_ACTIVE = 1,
826 struct perf_sample_data;
828 typedef void (*perf_overflow_handler_t)(struct perf_event *,
829 struct perf_sample_data *,
830 struct pt_regs *regs);
832 enum perf_group_flag {
833 PERF_GROUP_SOFTWARE = 0x1,
836 #define SWEVENT_HLIST_BITS 8
837 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
839 struct swevent_hlist {
840 struct hlist_head heads[SWEVENT_HLIST_SIZE];
841 struct rcu_head rcu_head;
844 #define PERF_ATTACH_CONTEXT 0x01
845 #define PERF_ATTACH_GROUP 0x02
846 #define PERF_ATTACH_TASK 0x04
848 #ifdef CONFIG_CGROUP_PERF
850 * perf_cgroup_info keeps track of time_enabled for a cgroup.
851 * This is a per-cpu dynamically allocated data structure.
853 struct perf_cgroup_info {
859 struct cgroup_subsys_state css;
860 struct perf_cgroup_info *info; /* timing info, one per cpu */
867 * struct perf_event - performance event kernel representation:
870 #ifdef CONFIG_PERF_EVENTS
871 struct list_head group_entry;
872 struct list_head event_entry;
873 struct list_head sibling_list;
874 struct hlist_node hlist_entry;
877 struct perf_event *group_leader;
880 enum perf_event_active_state state;
881 unsigned int attach_state;
883 atomic64_t child_count;
886 * These are the total time in nanoseconds that the event
887 * has been enabled (i.e. eligible to run, and the task has
888 * been scheduled in, if this is a per-task event)
889 * and running (scheduled onto the CPU), respectively.
891 * They are computed from tstamp_enabled, tstamp_running and
892 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
894 u64 total_time_enabled;
895 u64 total_time_running;
898 * These are timestamps used for computing total_time_enabled
899 * and total_time_running when the event is in INACTIVE or
900 * ACTIVE state, measured in nanoseconds from an arbitrary point
902 * tstamp_enabled: the notional time when the event was enabled
903 * tstamp_running: the notional time when the event was scheduled on
904 * tstamp_stopped: in INACTIVE state, the notional time when the
905 * event was scheduled off.
912 * timestamp shadows the actual context timing but it can
913 * be safely used in NMI interrupt context. It reflects the
914 * context time as it was when the event was last scheduled in.
916 * ctx_time already accounts for ctx->timestamp. Therefore to
917 * compute ctx_time for a sample, simply add perf_clock().
921 struct perf_event_attr attr;
925 struct hw_perf_event hw;
927 struct perf_event_context *ctx;
931 * These accumulate total time (in nanoseconds) that children
932 * events have been enabled and running, respectively.
934 atomic64_t child_total_time_enabled;
935 atomic64_t child_total_time_running;
938 * Protect attach/detach and child_list:
940 struct mutex child_mutex;
941 struct list_head child_list;
942 struct perf_event *parent;
947 struct list_head owner_entry;
948 struct task_struct *owner;
951 struct mutex mmap_mutex;
954 struct user_struct *mmap_user;
955 struct ring_buffer *rb;
956 struct list_head rb_entry;
959 wait_queue_head_t waitq;
960 struct fasync_struct *fasync;
962 /* delayed work for NMIs and such */
966 struct irq_work pending;
968 atomic_t event_limit;
970 void (*destroy)(struct perf_event *);
971 struct rcu_head rcu_head;
973 struct pid_namespace *ns;
976 perf_overflow_handler_t overflow_handler;
977 void *overflow_handler_context;
979 #ifdef CONFIG_EVENT_TRACING
980 struct ftrace_event_call *tp_event;
981 struct event_filter *filter;
982 #ifdef CONFIG_FUNCTION_TRACER
983 struct ftrace_ops ftrace_ops;
987 #ifdef CONFIG_CGROUP_PERF
988 struct perf_cgroup *cgrp; /* cgroup event is attach to */
989 int cgrp_defer_enabled;
992 #endif /* CONFIG_PERF_EVENTS */
995 enum perf_event_context_type {
1001 * struct perf_event_context - event context structure
1003 * Used as a container for task events and CPU events as well:
1005 struct perf_event_context {
1007 enum perf_event_context_type type;
1009 * Protect the states of the events in the list,
1010 * nr_active, and the list:
1012 raw_spinlock_t lock;
1014 * Protect the list of events. Locking either mutex or lock
1015 * is sufficient to ensure the list doesn't change; to change
1016 * the list you need to lock both the mutex and the spinlock.
1020 struct list_head pinned_groups;
1021 struct list_head flexible_groups;
1022 struct list_head event_list;
1030 struct task_struct *task;
1033 * Context clock, runs when context enabled.
1039 * These fields let us detect when two contexts have both
1040 * been cloned (inherited) from a common ancestor.
1042 struct perf_event_context *parent_ctx;
1046 int nr_cgroups; /* cgroup evts */
1047 int nr_branch_stack; /* branch_stack evt */
1048 struct rcu_head rcu_head;
1052 * Number of contexts where an event can trigger:
1053 * task, softirq, hardirq, nmi.
1055 #define PERF_NR_CONTEXTS 4
1058 * struct perf_event_cpu_context - per cpu event context structure
1060 struct perf_cpu_context {
1061 struct perf_event_context ctx;
1062 struct perf_event_context *task_ctx;
1065 struct list_head rotation_list;
1066 int jiffies_interval;
1067 struct pmu *active_pmu;
1068 struct perf_cgroup *cgrp;
1071 struct perf_output_handle {
1072 struct perf_event *event;
1073 struct ring_buffer *rb;
1074 unsigned long wakeup;
1080 #ifdef CONFIG_PERF_EVENTS
1082 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
1083 extern void perf_pmu_unregister(struct pmu *pmu);
1085 extern int perf_num_counters(void);
1086 extern const char *perf_pmu_name(void);
1087 extern void __perf_event_task_sched_in(struct task_struct *prev,
1088 struct task_struct *task);
1089 extern void __perf_event_task_sched_out(struct task_struct *prev,
1090 struct task_struct *next);
1091 extern int perf_event_init_task(struct task_struct *child);
1092 extern void perf_event_exit_task(struct task_struct *child);
1093 extern void perf_event_free_task(struct task_struct *task);
1094 extern void perf_event_delayed_put(struct task_struct *task);
1095 extern void perf_event_print_debug(void);
1096 extern void perf_pmu_disable(struct pmu *pmu);
1097 extern void perf_pmu_enable(struct pmu *pmu);
1098 extern int perf_event_task_disable(void);
1099 extern int perf_event_task_enable(void);
1100 extern int perf_event_refresh(struct perf_event *event, int refresh);
1101 extern void perf_event_update_userpage(struct perf_event *event);
1102 extern int perf_event_release_kernel(struct perf_event *event);
1103 extern struct perf_event *
1104 perf_event_create_kernel_counter(struct perf_event_attr *attr,
1106 struct task_struct *task,
1107 perf_overflow_handler_t callback,
1109 extern u64 perf_event_read_value(struct perf_event *event,
1110 u64 *enabled, u64 *running);
1113 struct perf_sample_data {
1130 struct perf_callchain_entry *callchain;
1131 struct perf_raw_record *raw;
1132 struct perf_branch_stack *br_stack;
1135 static inline void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
1139 data->br_stack = NULL;
1142 extern void perf_output_sample(struct perf_output_handle *handle,
1143 struct perf_event_header *header,
1144 struct perf_sample_data *data,
1145 struct perf_event *event);
1146 extern void perf_prepare_sample(struct perf_event_header *header,
1147 struct perf_sample_data *data,
1148 struct perf_event *event,
1149 struct pt_regs *regs);
1151 extern int perf_event_overflow(struct perf_event *event,
1152 struct perf_sample_data *data,
1153 struct pt_regs *regs);
1155 static inline bool is_sampling_event(struct perf_event *event)
1157 return event->attr.sample_period != 0;
1161 * Return 1 for a software event, 0 for a hardware event
1163 static inline int is_software_event(struct perf_event *event)
1165 return event->pmu->task_ctx_nr == perf_sw_context;
1168 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1170 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1172 #ifndef perf_arch_fetch_caller_regs
1173 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1177 * Take a snapshot of the regs. Skip ip and frame pointer to
1178 * the nth caller. We only need a few of the regs:
1179 * - ip for PERF_SAMPLE_IP
1180 * - cs for user_mode() tests
1181 * - bp for callchains
1182 * - eflags, for future purposes, just in case
1184 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1186 memset(regs, 0, sizeof(*regs));
1188 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1191 static __always_inline void
1192 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1194 struct pt_regs hot_regs;
1196 if (static_key_false(&perf_swevent_enabled[event_id])) {
1198 perf_fetch_caller_regs(&hot_regs);
1201 __perf_sw_event(event_id, nr, regs, addr);
1205 extern struct static_key_deferred perf_sched_events;
1207 static inline void perf_event_task_sched_in(struct task_struct *prev,
1208 struct task_struct *task)
1210 if (static_key_false(&perf_sched_events.key))
1211 __perf_event_task_sched_in(prev, task);
1214 static inline void perf_event_task_sched_out(struct task_struct *prev,
1215 struct task_struct *next)
1217 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
1219 if (static_key_false(&perf_sched_events.key))
1220 __perf_event_task_sched_out(prev, next);
1223 extern void perf_event_mmap(struct vm_area_struct *vma);
1224 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1225 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1226 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1228 extern void perf_event_comm(struct task_struct *tsk);
1229 extern void perf_event_fork(struct task_struct *tsk);
1232 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1234 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
1235 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
1237 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1239 if (entry->nr < PERF_MAX_STACK_DEPTH)
1240 entry->ip[entry->nr++] = ip;
1243 extern int sysctl_perf_event_paranoid;
1244 extern int sysctl_perf_event_mlock;
1245 extern int sysctl_perf_event_sample_rate;
1247 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1248 void __user *buffer, size_t *lenp,
1251 static inline bool perf_paranoid_tracepoint_raw(void)
1253 return sysctl_perf_event_paranoid > -1;
1256 static inline bool perf_paranoid_cpu(void)
1258 return sysctl_perf_event_paranoid > 0;
1261 static inline bool perf_paranoid_kernel(void)
1263 return sysctl_perf_event_paranoid > 1;
1266 extern void perf_event_init(void);
1267 extern void perf_tp_event(u64 addr, u64 count, void *record,
1268 int entry_size, struct pt_regs *regs,
1269 struct hlist_head *head, int rctx);
1270 extern void perf_bp_event(struct perf_event *event, void *data);
1272 #ifndef perf_misc_flags
1273 # define perf_misc_flags(regs) \
1274 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1275 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1278 static inline bool has_branch_stack(struct perf_event *event)
1280 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1283 extern int perf_output_begin(struct perf_output_handle *handle,
1284 struct perf_event *event, unsigned int size);
1285 extern void perf_output_end(struct perf_output_handle *handle);
1286 extern void perf_output_copy(struct perf_output_handle *handle,
1287 const void *buf, unsigned int len);
1288 extern int perf_swevent_get_recursion_context(void);
1289 extern void perf_swevent_put_recursion_context(int rctx);
1290 extern void perf_event_enable(struct perf_event *event);
1291 extern void perf_event_disable(struct perf_event *event);
1292 extern void perf_event_task_tick(void);
1295 perf_event_task_sched_in(struct task_struct *prev,
1296 struct task_struct *task) { }
1298 perf_event_task_sched_out(struct task_struct *prev,
1299 struct task_struct *next) { }
1300 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1301 static inline void perf_event_exit_task(struct task_struct *child) { }
1302 static inline void perf_event_free_task(struct task_struct *task) { }
1303 static inline void perf_event_delayed_put(struct task_struct *task) { }
1304 static inline void perf_event_print_debug(void) { }
1305 static inline int perf_event_task_disable(void) { return -EINVAL; }
1306 static inline int perf_event_task_enable(void) { return -EINVAL; }
1307 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1313 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1315 perf_bp_event(struct perf_event *event, void *data) { }
1317 static inline int perf_register_guest_info_callbacks
1318 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1319 static inline int perf_unregister_guest_info_callbacks
1320 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1322 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1323 static inline void perf_event_comm(struct task_struct *tsk) { }
1324 static inline void perf_event_fork(struct task_struct *tsk) { }
1325 static inline void perf_event_init(void) { }
1326 static inline int perf_swevent_get_recursion_context(void) { return -1; }
1327 static inline void perf_swevent_put_recursion_context(int rctx) { }
1328 static inline void perf_event_enable(struct perf_event *event) { }
1329 static inline void perf_event_disable(struct perf_event *event) { }
1330 static inline void perf_event_task_tick(void) { }
1333 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1336 * This has to have a higher priority than migration_notifier in sched.c.
1338 #define perf_cpu_notifier(fn) \
1340 static struct notifier_block fn##_nb __cpuinitdata = \
1341 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1342 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1343 (void *)(unsigned long)smp_processor_id()); \
1344 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1345 (void *)(unsigned long)smp_processor_id()); \
1346 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1347 (void *)(unsigned long)smp_processor_id()); \
1348 register_cpu_notifier(&fn##_nb); \
1352 #define PMU_FORMAT_ATTR(_name, _format) \
1354 _name##_show(struct device *dev, \
1355 struct device_attribute *attr, \
1358 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1359 return sprintf(page, _format "\n"); \
1362 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1364 #endif /* __KERNEL__ */
1365 #endif /* _LINUX_PERF_EVENT_H */