4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009, Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009, 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,
35 PERF_TYPE_MAX, /* non-ABI */
39 * Generalized performance event event_id types, used by the
40 * attr.event_id parameter of the sys_perf_event_open()
45 * Common hardware events, generalized by the kernel:
47 PERF_COUNT_HW_CPU_CYCLES = 0,
48 PERF_COUNT_HW_INSTRUCTIONS = 1,
49 PERF_COUNT_HW_CACHE_REFERENCES = 2,
50 PERF_COUNT_HW_CACHE_MISSES = 3,
51 PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4,
52 PERF_COUNT_HW_BRANCH_MISSES = 5,
53 PERF_COUNT_HW_BUS_CYCLES = 6,
55 PERF_COUNT_HW_MAX, /* non-ABI */
59 * Generalized hardware cache events:
61 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
62 * { read, write, prefetch } x
63 * { accesses, misses }
65 enum perf_hw_cache_id {
66 PERF_COUNT_HW_CACHE_L1D = 0,
67 PERF_COUNT_HW_CACHE_L1I = 1,
68 PERF_COUNT_HW_CACHE_LL = 2,
69 PERF_COUNT_HW_CACHE_DTLB = 3,
70 PERF_COUNT_HW_CACHE_ITLB = 4,
71 PERF_COUNT_HW_CACHE_BPU = 5,
73 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */
76 enum perf_hw_cache_op_id {
77 PERF_COUNT_HW_CACHE_OP_READ = 0,
78 PERF_COUNT_HW_CACHE_OP_WRITE = 1,
79 PERF_COUNT_HW_CACHE_OP_PREFETCH = 2,
81 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */
84 enum perf_hw_cache_op_result_id {
85 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0,
86 PERF_COUNT_HW_CACHE_RESULT_MISS = 1,
88 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */
92 * Special "software" events provided by the kernel, even if the hardware
93 * does not support performance events. These events measure various
94 * physical and sw events of the kernel (and allow the profiling of them as
98 PERF_COUNT_SW_CPU_CLOCK = 0,
99 PERF_COUNT_SW_TASK_CLOCK = 1,
100 PERF_COUNT_SW_PAGE_FAULTS = 2,
101 PERF_COUNT_SW_CONTEXT_SWITCHES = 3,
102 PERF_COUNT_SW_CPU_MIGRATIONS = 4,
103 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5,
104 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6,
105 PERF_COUNT_SW_ALIGNMENT_FAULTS = 7,
106 PERF_COUNT_SW_EMULATION_FAULTS = 8,
108 PERF_COUNT_SW_MAX, /* non-ABI */
112 * Bits that can be set in attr.sample_type to request information
113 * in the overflow packets.
115 enum perf_event_sample_format {
116 PERF_SAMPLE_IP = 1U << 0,
117 PERF_SAMPLE_TID = 1U << 1,
118 PERF_SAMPLE_TIME = 1U << 2,
119 PERF_SAMPLE_ADDR = 1U << 3,
120 PERF_SAMPLE_READ = 1U << 4,
121 PERF_SAMPLE_CALLCHAIN = 1U << 5,
122 PERF_SAMPLE_ID = 1U << 6,
123 PERF_SAMPLE_CPU = 1U << 7,
124 PERF_SAMPLE_PERIOD = 1U << 8,
125 PERF_SAMPLE_STREAM_ID = 1U << 9,
126 PERF_SAMPLE_RAW = 1U << 10,
128 PERF_SAMPLE_MAX = 1U << 11, /* non-ABI */
132 * The format of the data returned by read() on a perf event fd,
133 * as specified by attr.read_format:
135 * struct read_format {
137 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
138 * { u64 time_running; } && PERF_FORMAT_RUNNING
139 * { u64 id; } && PERF_FORMAT_ID
140 * } && !PERF_FORMAT_GROUP
143 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
144 * { u64 time_running; } && PERF_FORMAT_RUNNING
146 * { u64 id; } && PERF_FORMAT_ID
148 * } && PERF_FORMAT_GROUP
151 enum perf_event_read_format {
152 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
153 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
154 PERF_FORMAT_ID = 1U << 2,
155 PERF_FORMAT_GROUP = 1U << 3,
157 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */
160 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
163 * Hardware event_id to monitor via a performance monitoring event:
165 struct perf_event_attr {
168 * Major type: hardware/software/tracepoint/etc.
173 * Size of the attr structure, for fwd/bwd compat.
178 * Type specific configuration information.
190 __u64 disabled : 1, /* off by default */
191 inherit : 1, /* children inherit it */
192 pinned : 1, /* must always be on PMU */
193 exclusive : 1, /* only group on PMU */
194 exclude_user : 1, /* don't count user */
195 exclude_kernel : 1, /* ditto kernel */
196 exclude_hv : 1, /* ditto hypervisor */
197 exclude_idle : 1, /* don't count when idle */
198 mmap : 1, /* include mmap data */
199 comm : 1, /* include comm data */
200 freq : 1, /* use freq, not period */
201 inherit_stat : 1, /* per task counts */
202 enable_on_exec : 1, /* next exec enables */
203 task : 1, /* trace fork/exit */
204 watermark : 1, /* wakeup_watermark */
209 __u32 wakeup_events; /* wakeup every n events */
210 __u32 wakeup_watermark; /* bytes before wakeup */
218 * Ioctls that can be done on a perf event fd:
220 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
221 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
222 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
223 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
224 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64)
225 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
226 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
228 enum perf_event_ioc_flags {
229 PERF_IOC_FLAG_GROUP = 1U << 0,
233 * Structure of the page that can be mapped via mmap
235 struct perf_event_mmap_page {
236 __u32 version; /* version number of this structure */
237 __u32 compat_version; /* lowest version this is compat with */
240 * Bits needed to read the hw events in user-space.
250 * count = pmc_read(pc->index - 1);
251 * count += pc->offset;
256 * } while (pc->lock != seq);
258 * NOTE: for obvious reason this only works on self-monitoring
261 __u32 lock; /* seqlock for synchronization */
262 __u32 index; /* hardware event identifier */
263 __s64 offset; /* add to hardware event value */
264 __u64 time_enabled; /* time event active */
265 __u64 time_running; /* time event on cpu */
268 * Hole for extension of the self monitor capabilities
271 __u64 __reserved[123]; /* align to 1k */
274 * Control data for the mmap() data buffer.
276 * User-space reading the @data_head value should issue an rmb(), on
277 * SMP capable platforms, after reading this value -- see
278 * perf_event_wakeup().
280 * When the mapping is PROT_WRITE the @data_tail value should be
281 * written by userspace to reflect the last read data. In this case
282 * the kernel will not over-write unread data.
284 __u64 data_head; /* head in the data section */
285 __u64 data_tail; /* user-space written tail */
288 #define PERF_RECORD_MISC_CPUMODE_MASK (3 << 0)
289 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
290 #define PERF_RECORD_MISC_KERNEL (1 << 0)
291 #define PERF_RECORD_MISC_USER (2 << 0)
292 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
294 struct perf_event_header {
300 enum perf_event_type {
303 * The MMAP events record the PROT_EXEC mappings so that we can
304 * correlate userspace IPs to code. They have the following structure:
307 * struct perf_event_header header;
316 PERF_RECORD_MMAP = 1,
320 * struct perf_event_header header;
325 PERF_RECORD_LOST = 2,
329 * struct perf_event_header header;
335 PERF_RECORD_COMM = 3,
339 * struct perf_event_header header;
345 PERF_RECORD_EXIT = 4,
349 * struct perf_event_header header;
355 PERF_RECORD_THROTTLE = 5,
356 PERF_RECORD_UNTHROTTLE = 6,
360 * struct perf_event_header header;
366 PERF_RECORD_FORK = 7,
370 * struct perf_event_header header;
373 * struct read_format values;
376 PERF_RECORD_READ = 8,
380 * struct perf_event_header header;
382 * { u64 ip; } && PERF_SAMPLE_IP
383 * { u32 pid, tid; } && PERF_SAMPLE_TID
384 * { u64 time; } && PERF_SAMPLE_TIME
385 * { u64 addr; } && PERF_SAMPLE_ADDR
386 * { u64 id; } && PERF_SAMPLE_ID
387 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
388 * { u32 cpu, res; } && PERF_SAMPLE_CPU
389 * { u64 period; } && PERF_SAMPLE_PERIOD
391 * { struct read_format values; } && PERF_SAMPLE_READ
394 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
397 * # The RAW record below is opaque data wrt the ABI
399 * # That is, the ABI doesn't make any promises wrt to
400 * # the stability of its content, it may vary depending
401 * # on event, hardware, kernel version and phase of
404 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
408 * char data[size];}&& PERF_SAMPLE_RAW
411 PERF_RECORD_SAMPLE = 9,
413 PERF_RECORD_MAX, /* non-ABI */
416 enum perf_callchain_context {
417 PERF_CONTEXT_HV = (__u64)-32,
418 PERF_CONTEXT_KERNEL = (__u64)-128,
419 PERF_CONTEXT_USER = (__u64)-512,
421 PERF_CONTEXT_GUEST = (__u64)-2048,
422 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
423 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
425 PERF_CONTEXT_MAX = (__u64)-4095,
428 #define PERF_FLAG_FD_NO_GROUP (1U << 0)
429 #define PERF_FLAG_FD_OUTPUT (1U << 1)
433 * Kernel-internal data types and definitions:
436 #ifdef CONFIG_PERF_EVENTS
437 # include <asm/perf_event.h>
440 #include <linux/list.h>
441 #include <linux/mutex.h>
442 #include <linux/rculist.h>
443 #include <linux/rcupdate.h>
444 #include <linux/spinlock.h>
445 #include <linux/hrtimer.h>
446 #include <linux/fs.h>
447 #include <linux/pid_namespace.h>
448 #include <linux/workqueue.h>
449 #include <asm/atomic.h>
451 #define PERF_MAX_STACK_DEPTH 255
453 struct perf_callchain_entry {
455 __u64 ip[PERF_MAX_STACK_DEPTH];
458 struct perf_raw_record {
466 * struct hw_perf_event - performance event hardware details:
468 struct hw_perf_event {
469 #ifdef CONFIG_PERF_EVENTS
471 struct { /* hardware */
473 unsigned long config_base;
474 unsigned long event_base;
477 struct { /* software */
479 struct hrtimer hrtimer;
482 atomic64_t prev_count;
485 atomic64_t period_left;
497 * struct pmu - generic performance monitoring unit
500 int (*enable) (struct perf_event *event);
501 void (*disable) (struct perf_event *event);
502 void (*read) (struct perf_event *event);
503 void (*unthrottle) (struct perf_event *event);
507 * enum perf_event_active_state - the states of a event
509 enum perf_event_active_state {
510 PERF_EVENT_STATE_ERROR = -2,
511 PERF_EVENT_STATE_OFF = -1,
512 PERF_EVENT_STATE_INACTIVE = 0,
513 PERF_EVENT_STATE_ACTIVE = 1,
518 struct perf_mmap_data {
519 struct rcu_head rcu_head;
520 #ifdef CONFIG_PERF_USE_VMALLOC
521 struct work_struct work;
524 int nr_pages; /* nr of data pages */
525 int writable; /* are we writable */
526 int nr_locked; /* nr pages mlocked */
528 atomic_t poll; /* POLL_ for wakeups */
529 atomic_t events; /* event_id limit */
531 atomic_long_t head; /* write position */
532 atomic_long_t done_head; /* completed head */
534 atomic_t lock; /* concurrent writes */
535 atomic_t wakeup; /* needs a wakeup */
536 atomic_t lost; /* nr records lost */
538 long watermark; /* wakeup watermark */
540 struct perf_event_mmap_page *user_page;
544 struct perf_pending_entry {
545 struct perf_pending_entry *next;
546 void (*func)(struct perf_pending_entry *);
550 * struct perf_event - performance event kernel representation:
553 #ifdef CONFIG_PERF_EVENTS
554 struct list_head group_entry;
555 struct list_head event_entry;
556 struct list_head sibling_list;
558 struct perf_event *group_leader;
559 struct perf_event *output;
560 const struct pmu *pmu;
562 enum perf_event_active_state state;
566 * These are the total time in nanoseconds that the event
567 * has been enabled (i.e. eligible to run, and the task has
568 * been scheduled in, if this is a per-task event)
569 * and running (scheduled onto the CPU), respectively.
571 * They are computed from tstamp_enabled, tstamp_running and
572 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
574 u64 total_time_enabled;
575 u64 total_time_running;
578 * These are timestamps used for computing total_time_enabled
579 * and total_time_running when the event is in INACTIVE or
580 * ACTIVE state, measured in nanoseconds from an arbitrary point
582 * tstamp_enabled: the notional time when the event was enabled
583 * tstamp_running: the notional time when the event was scheduled on
584 * tstamp_stopped: in INACTIVE state, the notional time when the
585 * event was scheduled off.
591 struct perf_event_attr attr;
592 struct hw_perf_event hw;
594 struct perf_event_context *ctx;
598 * These accumulate total time (in nanoseconds) that children
599 * events have been enabled and running, respectively.
601 atomic64_t child_total_time_enabled;
602 atomic64_t child_total_time_running;
605 * Protect attach/detach and child_list:
607 struct mutex child_mutex;
608 struct list_head child_list;
609 struct perf_event *parent;
614 struct list_head owner_entry;
615 struct task_struct *owner;
618 struct mutex mmap_mutex;
620 struct perf_mmap_data *data;
623 wait_queue_head_t waitq;
624 struct fasync_struct *fasync;
626 /* delayed work for NMIs and such */
630 struct perf_pending_entry pending;
632 atomic_t event_limit;
634 void (*destroy)(struct perf_event *);
635 struct rcu_head rcu_head;
637 struct pid_namespace *ns;
640 #ifdef CONFIG_EVENT_PROFILE
641 struct event_filter *filter;
644 #endif /* CONFIG_PERF_EVENTS */
648 * struct perf_event_context - event context structure
650 * Used as a container for task events and CPU events as well:
652 struct perf_event_context {
654 * Protect the states of the events in the list,
655 * nr_active, and the list:
659 * Protect the list of events. Locking either mutex or lock
660 * is sufficient to ensure the list doesn't change; to change
661 * the list you need to lock both the mutex and the spinlock.
665 struct list_head group_list;
666 struct list_head event_list;
672 struct task_struct *task;
675 * Context clock, runs when context enabled.
681 * These fields let us detect when two contexts have both
682 * been cloned (inherited) from a common ancestor.
684 struct perf_event_context *parent_ctx;
688 struct rcu_head rcu_head;
692 * struct perf_event_cpu_context - per cpu event context structure
694 struct perf_cpu_context {
695 struct perf_event_context ctx;
696 struct perf_event_context *task_ctx;
702 * Recursion avoidance:
704 * task, softirq, irq, nmi context
709 struct perf_output_handle {
710 struct perf_event *event;
711 struct perf_mmap_data *data;
713 unsigned long offset;
719 #ifdef CONFIG_PERF_EVENTS
722 * Set by architecture code:
724 extern int perf_max_events;
726 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
728 extern void perf_event_task_sched_in(struct task_struct *task, int cpu);
729 extern void perf_event_task_sched_out(struct task_struct *task,
730 struct task_struct *next, int cpu);
731 extern void perf_event_task_tick(struct task_struct *task, int cpu);
732 extern int perf_event_init_task(struct task_struct *child);
733 extern void perf_event_exit_task(struct task_struct *child);
734 extern void perf_event_free_task(struct task_struct *task);
735 extern void set_perf_event_pending(void);
736 extern void perf_event_do_pending(void);
737 extern void perf_event_print_debug(void);
738 extern void __perf_disable(void);
739 extern bool __perf_enable(void);
740 extern void perf_disable(void);
741 extern void perf_enable(void);
742 extern int perf_event_task_disable(void);
743 extern int perf_event_task_enable(void);
744 extern int hw_perf_group_sched_in(struct perf_event *group_leader,
745 struct perf_cpu_context *cpuctx,
746 struct perf_event_context *ctx, int cpu);
747 extern void perf_event_update_userpage(struct perf_event *event);
749 struct perf_sample_data {
766 struct perf_callchain_entry *callchain;
767 struct perf_raw_record *raw;
770 extern void perf_output_sample(struct perf_output_handle *handle,
771 struct perf_event_header *header,
772 struct perf_sample_data *data,
773 struct perf_event *event);
774 extern void perf_prepare_sample(struct perf_event_header *header,
775 struct perf_sample_data *data,
776 struct perf_event *event,
777 struct pt_regs *regs);
779 extern int perf_event_overflow(struct perf_event *event, int nmi,
780 struct perf_sample_data *data,
781 struct pt_regs *regs);
784 * Return 1 for a software event, 0 for a hardware event
786 static inline int is_software_event(struct perf_event *event)
788 return (event->attr.type != PERF_TYPE_RAW) &&
789 (event->attr.type != PERF_TYPE_HARDWARE) &&
790 (event->attr.type != PERF_TYPE_HW_CACHE);
793 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
795 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
798 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
800 if (atomic_read(&perf_swevent_enabled[event_id]))
801 __perf_sw_event(event_id, nr, nmi, regs, addr);
804 extern void __perf_event_mmap(struct vm_area_struct *vma);
806 static inline void perf_event_mmap(struct vm_area_struct *vma)
808 if (vma->vm_flags & VM_EXEC)
809 __perf_event_mmap(vma);
812 extern void perf_event_comm(struct task_struct *tsk);
813 extern void perf_event_fork(struct task_struct *tsk);
815 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
817 extern int sysctl_perf_event_paranoid;
818 extern int sysctl_perf_event_mlock;
819 extern int sysctl_perf_event_sample_rate;
821 extern void perf_event_init(void);
822 extern void perf_tp_event(int event_id, u64 addr, u64 count,
823 void *record, int entry_size);
825 #ifndef perf_misc_flags
826 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_RECORD_MISC_USER : \
827 PERF_RECORD_MISC_KERNEL)
828 #define perf_instruction_pointer(regs) instruction_pointer(regs)
831 extern int perf_output_begin(struct perf_output_handle *handle,
832 struct perf_event *event, unsigned int size,
833 int nmi, int sample);
834 extern void perf_output_end(struct perf_output_handle *handle);
835 extern void perf_output_copy(struct perf_output_handle *handle,
836 const void *buf, unsigned int len);
839 perf_event_task_sched_in(struct task_struct *task, int cpu) { }
841 perf_event_task_sched_out(struct task_struct *task,
842 struct task_struct *next, int cpu) { }
844 perf_event_task_tick(struct task_struct *task, int cpu) { }
845 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
846 static inline void perf_event_exit_task(struct task_struct *child) { }
847 static inline void perf_event_free_task(struct task_struct *task) { }
848 static inline void perf_event_do_pending(void) { }
849 static inline void perf_event_print_debug(void) { }
850 static inline void perf_disable(void) { }
851 static inline void perf_enable(void) { }
852 static inline int perf_event_task_disable(void) { return -EINVAL; }
853 static inline int perf_event_task_enable(void) { return -EINVAL; }
856 perf_sw_event(u32 event_id, u64 nr, int nmi,
857 struct pt_regs *regs, u64 addr) { }
859 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
860 static inline void perf_event_comm(struct task_struct *tsk) { }
861 static inline void perf_event_fork(struct task_struct *tsk) { }
862 static inline void perf_event_init(void) { }
866 #define perf_output_put(handle, x) \
867 perf_output_copy((handle), &(x), sizeof(x))
869 #endif /* __KERNEL__ */
870 #endif /* _LINUX_PERF_EVENT_H */