2 * Performance counters:
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_COUNTER_H
15 #define _LINUX_PERF_COUNTER_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 counter event types, used by the
40 * attr.event_id parameter of the sys_perf_counter_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 counters:
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" counters provided by the kernel, even if the hardware
93 * does not support performance counters. These counters 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,
106 PERF_COUNT_SW_MAX, /* non-ABI */
110 * Bits that can be set in attr.sample_type to request information
111 * in the overflow packets.
113 enum perf_counter_sample_format {
114 PERF_SAMPLE_IP = 1U << 0,
115 PERF_SAMPLE_TID = 1U << 1,
116 PERF_SAMPLE_TIME = 1U << 2,
117 PERF_SAMPLE_ADDR = 1U << 3,
118 PERF_SAMPLE_GROUP = 1U << 4,
119 PERF_SAMPLE_CALLCHAIN = 1U << 5,
120 PERF_SAMPLE_ID = 1U << 6,
121 PERF_SAMPLE_CPU = 1U << 7,
122 PERF_SAMPLE_PERIOD = 1U << 8,
124 PERF_SAMPLE_MAX = 1U << 9, /* non-ABI */
128 * Bits that can be set in attr.read_format to request that
129 * reads on the counter should return the indicated quantities,
130 * in increasing order of bit value, after the counter value.
132 enum perf_counter_read_format {
133 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
134 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
135 PERF_FORMAT_ID = 1U << 2,
137 PERF_FORMAT_MAX = 1U << 3, /* non-ABI */
140 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
143 * Hardware event to monitor via a performance monitoring counter:
145 struct perf_counter_attr {
148 * Major type: hardware/software/tracepoint/etc.
153 * Size of the attr structure, for fwd/bwd compat.
158 * Type specific configuration information.
170 __u64 disabled : 1, /* off by default */
171 inherit : 1, /* children inherit it */
172 pinned : 1, /* must always be on PMU */
173 exclusive : 1, /* only group on PMU */
174 exclude_user : 1, /* don't count user */
175 exclude_kernel : 1, /* ditto kernel */
176 exclude_hv : 1, /* ditto hypervisor */
177 exclude_idle : 1, /* don't count when idle */
178 mmap : 1, /* include mmap data */
179 comm : 1, /* include comm data */
180 freq : 1, /* use freq, not period */
184 __u32 wakeup_events; /* wakeup every n events */
191 * Ioctls that can be done on a perf counter fd:
193 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
194 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
195 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
196 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
197 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
199 enum perf_counter_ioc_flags {
200 PERF_IOC_FLAG_GROUP = 1U << 0,
204 * Structure of the page that can be mapped via mmap
206 struct perf_counter_mmap_page {
207 __u32 version; /* version number of this structure */
208 __u32 compat_version; /* lowest version this is compat with */
211 * Bits needed to read the hw counters in user-space.
221 * count = pmc_read(pc->index - 1);
222 * count += pc->offset;
227 * } while (pc->lock != seq);
229 * NOTE: for obvious reason this only works on self-monitoring
232 __u32 lock; /* seqlock for synchronization */
233 __u32 index; /* hardware counter identifier */
234 __s64 offset; /* add to hardware counter value */
235 __u64 time_enabled; /* time counter active */
236 __u64 time_running; /* time counter on cpu */
239 * Hole for extension of the self monitor capabilities
242 __u64 __reserved[123]; /* align to 1k */
245 * Control data for the mmap() data buffer.
247 * User-space reading the @data_head value should issue an rmb(), on
248 * SMP capable platforms, after reading this value -- see
249 * perf_counter_wakeup().
251 * When the mapping is PROT_WRITE the @data_tail value should be
252 * written by userspace to reflect the last read data. In this case
253 * the kernel will not over-write unread data.
255 __u64 data_head; /* head in the data section */
256 __u64 data_tail; /* user-space written tail */
259 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
260 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
261 #define PERF_EVENT_MISC_KERNEL (1 << 0)
262 #define PERF_EVENT_MISC_USER (2 << 0)
263 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
264 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
266 struct perf_event_header {
272 enum perf_event_type {
275 * The MMAP events record the PROT_EXEC mappings so that we can
276 * correlate userspace IPs to code. They have the following structure:
279 * struct perf_event_header header;
292 * struct perf_event_header header;
301 * struct perf_event_header header;
311 * struct perf_event_header header;
317 PERF_EVENT_PERIOD = 4,
321 * struct perf_event_header header;
326 PERF_EVENT_THROTTLE = 5,
327 PERF_EVENT_UNTHROTTLE = 6,
331 * struct perf_event_header header;
339 * struct perf_event_header header;
342 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
343 * { u64 time_running; } && PERF_FORMAT_RUNNING
344 * { u64 parent_id; } && PERF_FORMAT_ID
350 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
351 * will be PERF_SAMPLE_*
354 * struct perf_event_header header;
356 * { u64 ip; } && PERF_SAMPLE_IP
357 * { u32 pid, tid; } && PERF_SAMPLE_TID
358 * { u64 time; } && PERF_SAMPLE_TIME
359 * { u64 addr; } && PERF_SAMPLE_ADDR
360 * { u64 config; } && PERF_SAMPLE_CONFIG
361 * { u32 cpu, res; } && PERF_SAMPLE_CPU
364 * { u64 id, val; } cnt[nr]; } && PERF_SAMPLE_GROUP
367 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
372 enum perf_callchain_context {
373 PERF_CONTEXT_HV = (__u64)-32,
374 PERF_CONTEXT_KERNEL = (__u64)-128,
375 PERF_CONTEXT_USER = (__u64)-512,
377 PERF_CONTEXT_GUEST = (__u64)-2048,
378 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
379 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
381 PERF_CONTEXT_MAX = (__u64)-4095,
386 * Kernel-internal data types and definitions:
389 #ifdef CONFIG_PERF_COUNTERS
390 # include <asm/perf_counter.h>
393 #include <linux/list.h>
394 #include <linux/mutex.h>
395 #include <linux/rculist.h>
396 #include <linux/rcupdate.h>
397 #include <linux/spinlock.h>
398 #include <linux/hrtimer.h>
399 #include <linux/fs.h>
400 #include <linux/pid_namespace.h>
401 #include <asm/atomic.h>
403 #define PERF_MAX_STACK_DEPTH 255
405 struct perf_callchain_entry {
407 __u64 ip[PERF_MAX_STACK_DEPTH];
413 * struct hw_perf_counter - performance counter hardware details:
415 struct hw_perf_counter {
416 #ifdef CONFIG_PERF_COUNTERS
418 struct { /* hardware */
420 unsigned long config_base;
421 unsigned long counter_base;
424 union { /* software */
426 struct hrtimer hrtimer;
429 atomic64_t prev_count;
432 atomic64_t period_left;
444 * struct pmu - generic performance monitoring unit
447 int (*enable) (struct perf_counter *counter);
448 void (*disable) (struct perf_counter *counter);
449 void (*read) (struct perf_counter *counter);
450 void (*unthrottle) (struct perf_counter *counter);
454 * enum perf_counter_active_state - the states of a counter
456 enum perf_counter_active_state {
457 PERF_COUNTER_STATE_ERROR = -2,
458 PERF_COUNTER_STATE_OFF = -1,
459 PERF_COUNTER_STATE_INACTIVE = 0,
460 PERF_COUNTER_STATE_ACTIVE = 1,
465 struct perf_mmap_data {
466 struct rcu_head rcu_head;
467 int nr_pages; /* nr of data pages */
468 int writable; /* are we writable */
469 int nr_locked; /* nr pages mlocked */
471 atomic_t poll; /* POLL_ for wakeups */
472 atomic_t events; /* event limit */
474 atomic_long_t head; /* write position */
475 atomic_long_t done_head; /* completed head */
477 atomic_t lock; /* concurrent writes */
478 atomic_t wakeup; /* needs a wakeup */
479 atomic_t lost; /* nr records lost */
481 struct perf_counter_mmap_page *user_page;
485 struct perf_pending_entry {
486 struct perf_pending_entry *next;
487 void (*func)(struct perf_pending_entry *);
491 * struct perf_counter - performance counter kernel representation:
493 struct perf_counter {
494 #ifdef CONFIG_PERF_COUNTERS
495 struct list_head list_entry;
496 struct list_head event_entry;
497 struct list_head sibling_list;
499 struct perf_counter *group_leader;
500 const struct pmu *pmu;
502 enum perf_counter_active_state state;
506 * These are the total time in nanoseconds that the counter
507 * has been enabled (i.e. eligible to run, and the task has
508 * been scheduled in, if this is a per-task counter)
509 * and running (scheduled onto the CPU), respectively.
511 * They are computed from tstamp_enabled, tstamp_running and
512 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
514 u64 total_time_enabled;
515 u64 total_time_running;
518 * These are timestamps used for computing total_time_enabled
519 * and total_time_running when the counter is in INACTIVE or
520 * ACTIVE state, measured in nanoseconds from an arbitrary point
522 * tstamp_enabled: the notional time when the counter was enabled
523 * tstamp_running: the notional time when the counter was scheduled on
524 * tstamp_stopped: in INACTIVE state, the notional time when the
525 * counter was scheduled off.
531 struct perf_counter_attr attr;
532 struct hw_perf_counter hw;
534 struct perf_counter_context *ctx;
538 * These accumulate total time (in nanoseconds) that children
539 * counters have been enabled and running, respectively.
541 atomic64_t child_total_time_enabled;
542 atomic64_t child_total_time_running;
545 * Protect attach/detach and child_list:
547 struct mutex child_mutex;
548 struct list_head child_list;
549 struct perf_counter *parent;
554 struct list_head owner_entry;
555 struct task_struct *owner;
558 struct mutex mmap_mutex;
560 struct perf_mmap_data *data;
563 wait_queue_head_t waitq;
564 struct fasync_struct *fasync;
566 /* delayed work for NMIs and such */
570 struct perf_pending_entry pending;
572 atomic_t event_limit;
574 void (*destroy)(struct perf_counter *);
575 struct rcu_head rcu_head;
577 struct pid_namespace *ns;
583 * struct perf_counter_context - counter context structure
585 * Used as a container for task counters and CPU counters as well:
587 struct perf_counter_context {
589 * Protect the states of the counters in the list,
590 * nr_active, and the list:
594 * Protect the list of counters. Locking either mutex or lock
595 * is sufficient to ensure the list doesn't change; to change
596 * the list you need to lock both the mutex and the spinlock.
600 struct list_head counter_list;
601 struct list_head event_list;
606 struct task_struct *task;
609 * Context clock, runs when context enabled.
615 * These fields let us detect when two contexts have both
616 * been cloned (inherited) from a common ancestor.
618 struct perf_counter_context *parent_ctx;
622 struct rcu_head rcu_head;
626 * struct perf_counter_cpu_context - per cpu counter context structure
628 struct perf_cpu_context {
629 struct perf_counter_context ctx;
630 struct perf_counter_context *task_ctx;
636 * Recursion avoidance:
638 * task, softirq, irq, nmi context
643 #ifdef CONFIG_PERF_COUNTERS
646 * Set by architecture code:
648 extern int perf_max_counters;
650 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
652 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
653 extern void perf_counter_task_sched_out(struct task_struct *task,
654 struct task_struct *next, int cpu);
655 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
656 extern int perf_counter_init_task(struct task_struct *child);
657 extern void perf_counter_exit_task(struct task_struct *child);
658 extern void perf_counter_free_task(struct task_struct *task);
659 extern void set_perf_counter_pending(void);
660 extern void perf_counter_do_pending(void);
661 extern void perf_counter_print_debug(void);
662 extern void __perf_disable(void);
663 extern bool __perf_enable(void);
664 extern void perf_disable(void);
665 extern void perf_enable(void);
666 extern int perf_counter_task_disable(void);
667 extern int perf_counter_task_enable(void);
668 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
669 struct perf_cpu_context *cpuctx,
670 struct perf_counter_context *ctx, int cpu);
671 extern void perf_counter_update_userpage(struct perf_counter *counter);
673 struct perf_sample_data {
674 struct pt_regs *regs;
679 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
680 struct perf_sample_data *data);
683 * Return 1 for a software counter, 0 for a hardware counter
685 static inline int is_software_counter(struct perf_counter *counter)
687 return (counter->attr.type != PERF_TYPE_RAW) &&
688 (counter->attr.type != PERF_TYPE_HARDWARE) &&
689 (counter->attr.type != PERF_TYPE_HW_CACHE);
692 extern atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
694 extern void __perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
697 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
699 if (atomic_read(&perf_swcounter_enabled[event]))
700 __perf_swcounter_event(event, nr, nmi, regs, addr);
703 extern void __perf_counter_mmap(struct vm_area_struct *vma);
705 static inline void perf_counter_mmap(struct vm_area_struct *vma)
707 if (vma->vm_flags & VM_EXEC)
708 __perf_counter_mmap(vma);
711 extern void perf_counter_comm(struct task_struct *tsk);
712 extern void perf_counter_fork(struct task_struct *tsk);
714 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
716 extern int sysctl_perf_counter_paranoid;
717 extern int sysctl_perf_counter_mlock;
718 extern int sysctl_perf_counter_sample_rate;
720 extern void perf_counter_init(void);
722 #ifndef perf_misc_flags
723 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
724 PERF_EVENT_MISC_KERNEL)
725 #define perf_instruction_pointer(regs) instruction_pointer(regs)
730 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
732 perf_counter_task_sched_out(struct task_struct *task,
733 struct task_struct *next, int cpu) { }
735 perf_counter_task_tick(struct task_struct *task, int cpu) { }
736 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
737 static inline void perf_counter_exit_task(struct task_struct *child) { }
738 static inline void perf_counter_free_task(struct task_struct *task) { }
739 static inline void perf_counter_do_pending(void) { }
740 static inline void perf_counter_print_debug(void) { }
741 static inline void perf_disable(void) { }
742 static inline void perf_enable(void) { }
743 static inline int perf_counter_task_disable(void) { return -EINVAL; }
744 static inline int perf_counter_task_enable(void) { return -EINVAL; }
747 perf_swcounter_event(u32 event, u64 nr, int nmi,
748 struct pt_regs *regs, u64 addr) { }
750 static inline void perf_counter_mmap(struct vm_area_struct *vma) { }
751 static inline void perf_counter_comm(struct task_struct *tsk) { }
752 static inline void perf_counter_fork(struct task_struct *tsk) { }
753 static inline void perf_counter_init(void) { }
756 #endif /* __KERNEL__ */
757 #endif /* _LINUX_PERF_COUNTER_H */