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,
123 PERF_SAMPLE_STREAM_ID = 1U << 9,
124 PERF_SAMPLE_RAW = 1U << 10,
126 PERF_SAMPLE_MAX = 1U << 11, /* non-ABI */
130 * Bits that can be set in attr.read_format to request that
131 * reads on the counter should return the indicated quantities,
132 * in increasing order of bit value, after the counter value.
134 enum perf_counter_read_format {
135 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
136 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
137 PERF_FORMAT_ID = 1U << 2,
139 PERF_FORMAT_MAX = 1U << 3, /* non-ABI */
142 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
145 * Hardware event to monitor via a performance monitoring counter:
147 struct perf_counter_attr {
150 * Major type: hardware/software/tracepoint/etc.
155 * Size of the attr structure, for fwd/bwd compat.
160 * Type specific configuration information.
172 __u64 disabled : 1, /* off by default */
173 inherit : 1, /* children inherit it */
174 pinned : 1, /* must always be on PMU */
175 exclusive : 1, /* only group on PMU */
176 exclude_user : 1, /* don't count user */
177 exclude_kernel : 1, /* ditto kernel */
178 exclude_hv : 1, /* ditto hypervisor */
179 exclude_idle : 1, /* don't count when idle */
180 mmap : 1, /* include mmap data */
181 comm : 1, /* include comm data */
182 freq : 1, /* use freq, not period */
183 inherit_stat : 1, /* per task counts */
184 enable_on_exec : 1, /* next exec enables */
185 task : 1, /* trace fork/exit */
189 __u32 wakeup_events; /* wakeup every n events */
196 * Ioctls that can be done on a perf counter fd:
198 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
199 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
200 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
201 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
202 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
204 enum perf_counter_ioc_flags {
205 PERF_IOC_FLAG_GROUP = 1U << 0,
209 * Structure of the page that can be mapped via mmap
211 struct perf_counter_mmap_page {
212 __u32 version; /* version number of this structure */
213 __u32 compat_version; /* lowest version this is compat with */
216 * Bits needed to read the hw counters in user-space.
226 * count = pmc_read(pc->index - 1);
227 * count += pc->offset;
232 * } while (pc->lock != seq);
234 * NOTE: for obvious reason this only works on self-monitoring
237 __u32 lock; /* seqlock for synchronization */
238 __u32 index; /* hardware counter identifier */
239 __s64 offset; /* add to hardware counter value */
240 __u64 time_enabled; /* time counter active */
241 __u64 time_running; /* time counter on cpu */
244 * Hole for extension of the self monitor capabilities
247 __u64 __reserved[123]; /* align to 1k */
250 * Control data for the mmap() data buffer.
252 * User-space reading the @data_head value should issue an rmb(), on
253 * SMP capable platforms, after reading this value -- see
254 * perf_counter_wakeup().
256 * When the mapping is PROT_WRITE the @data_tail value should be
257 * written by userspace to reflect the last read data. In this case
258 * the kernel will not over-write unread data.
260 __u64 data_head; /* head in the data section */
261 __u64 data_tail; /* user-space written tail */
264 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
265 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
266 #define PERF_EVENT_MISC_KERNEL (1 << 0)
267 #define PERF_EVENT_MISC_USER (2 << 0)
268 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
270 struct perf_event_header {
276 enum perf_event_type {
279 * The MMAP events record the PROT_EXEC mappings so that we can
280 * correlate userspace IPs to code. They have the following structure:
283 * struct perf_event_header header;
296 * struct perf_event_header header;
305 * struct perf_event_header header;
315 * struct perf_event_header header;
324 * struct perf_event_header header;
330 PERF_EVENT_THROTTLE = 5,
331 PERF_EVENT_UNTHROTTLE = 6,
335 * struct perf_event_header header;
344 * struct perf_event_header header;
347 * { u64 time_enabled; } && PERF_FORMAT_ENABLED
348 * { u64 time_running; } && PERF_FORMAT_RUNNING
349 * { u64 parent_id; } && PERF_FORMAT_ID
356 * struct perf_event_header header;
358 * { u64 ip; } && PERF_SAMPLE_IP
359 * { u32 pid, tid; } && PERF_SAMPLE_TID
360 * { u64 time; } && PERF_SAMPLE_TIME
361 * { u64 addr; } && PERF_SAMPLE_ADDR
362 * { u64 id; } && PERF_SAMPLE_ID
363 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
364 * { u32 cpu, res; } && PERF_SAMPLE_CPU
365 * { u64 period; } && PERF_SAMPLE_PERIOD
368 * { u64 id, val; } cnt[nr]; } && PERF_SAMPLE_GROUP
371 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
373 * char data[size];}&& PERF_SAMPLE_RAW
376 PERF_EVENT_SAMPLE = 9,
378 PERF_EVENT_MAX, /* non-ABI */
381 enum perf_callchain_context {
382 PERF_CONTEXT_HV = (__u64)-32,
383 PERF_CONTEXT_KERNEL = (__u64)-128,
384 PERF_CONTEXT_USER = (__u64)-512,
386 PERF_CONTEXT_GUEST = (__u64)-2048,
387 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
388 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
390 PERF_CONTEXT_MAX = (__u64)-4095,
395 * Kernel-internal data types and definitions:
398 #ifdef CONFIG_PERF_COUNTERS
399 # include <asm/perf_counter.h>
402 #include <linux/list.h>
403 #include <linux/mutex.h>
404 #include <linux/rculist.h>
405 #include <linux/rcupdate.h>
406 #include <linux/spinlock.h>
407 #include <linux/hrtimer.h>
408 #include <linux/fs.h>
409 #include <linux/pid_namespace.h>
410 #include <asm/atomic.h>
412 #define PERF_MAX_STACK_DEPTH 255
414 struct perf_callchain_entry {
416 __u64 ip[PERF_MAX_STACK_DEPTH];
419 struct perf_raw_record {
427 * struct hw_perf_counter - performance counter hardware details:
429 struct hw_perf_counter {
430 #ifdef CONFIG_PERF_COUNTERS
432 struct { /* hardware */
434 unsigned long config_base;
435 unsigned long counter_base;
438 union { /* software */
440 struct hrtimer hrtimer;
443 atomic64_t prev_count;
446 atomic64_t period_left;
458 * struct pmu - generic performance monitoring unit
461 int (*enable) (struct perf_counter *counter);
462 void (*disable) (struct perf_counter *counter);
463 void (*read) (struct perf_counter *counter);
464 void (*unthrottle) (struct perf_counter *counter);
468 * enum perf_counter_active_state - the states of a counter
470 enum perf_counter_active_state {
471 PERF_COUNTER_STATE_ERROR = -2,
472 PERF_COUNTER_STATE_OFF = -1,
473 PERF_COUNTER_STATE_INACTIVE = 0,
474 PERF_COUNTER_STATE_ACTIVE = 1,
479 struct perf_mmap_data {
480 struct rcu_head rcu_head;
481 int nr_pages; /* nr of data pages */
482 int writable; /* are we writable */
483 int nr_locked; /* nr pages mlocked */
485 atomic_t poll; /* POLL_ for wakeups */
486 atomic_t events; /* event limit */
488 atomic_long_t head; /* write position */
489 atomic_long_t done_head; /* completed head */
491 atomic_t lock; /* concurrent writes */
492 atomic_t wakeup; /* needs a wakeup */
493 atomic_t lost; /* nr records lost */
495 struct perf_counter_mmap_page *user_page;
499 struct perf_pending_entry {
500 struct perf_pending_entry *next;
501 void (*func)(struct perf_pending_entry *);
505 * struct perf_counter - performance counter kernel representation:
507 struct perf_counter {
508 #ifdef CONFIG_PERF_COUNTERS
509 struct list_head list_entry;
510 struct list_head event_entry;
511 struct list_head sibling_list;
513 struct perf_counter *group_leader;
514 const struct pmu *pmu;
516 enum perf_counter_active_state state;
520 * These are the total time in nanoseconds that the counter
521 * has been enabled (i.e. eligible to run, and the task has
522 * been scheduled in, if this is a per-task counter)
523 * and running (scheduled onto the CPU), respectively.
525 * They are computed from tstamp_enabled, tstamp_running and
526 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
528 u64 total_time_enabled;
529 u64 total_time_running;
532 * These are timestamps used for computing total_time_enabled
533 * and total_time_running when the counter is in INACTIVE or
534 * ACTIVE state, measured in nanoseconds from an arbitrary point
536 * tstamp_enabled: the notional time when the counter was enabled
537 * tstamp_running: the notional time when the counter was scheduled on
538 * tstamp_stopped: in INACTIVE state, the notional time when the
539 * counter was scheduled off.
545 struct perf_counter_attr attr;
546 struct hw_perf_counter hw;
548 struct perf_counter_context *ctx;
552 * These accumulate total time (in nanoseconds) that children
553 * counters have been enabled and running, respectively.
555 atomic64_t child_total_time_enabled;
556 atomic64_t child_total_time_running;
559 * Protect attach/detach and child_list:
561 struct mutex child_mutex;
562 struct list_head child_list;
563 struct perf_counter *parent;
568 struct list_head owner_entry;
569 struct task_struct *owner;
572 struct mutex mmap_mutex;
574 struct perf_mmap_data *data;
577 wait_queue_head_t waitq;
578 struct fasync_struct *fasync;
580 /* delayed work for NMIs and such */
584 struct perf_pending_entry pending;
586 atomic_t event_limit;
588 void (*destroy)(struct perf_counter *);
589 struct rcu_head rcu_head;
591 struct pid_namespace *ns;
597 * struct perf_counter_context - counter context structure
599 * Used as a container for task counters and CPU counters as well:
601 struct perf_counter_context {
603 * Protect the states of the counters in the list,
604 * nr_active, and the list:
608 * Protect the list of counters. Locking either mutex or lock
609 * is sufficient to ensure the list doesn't change; to change
610 * the list you need to lock both the mutex and the spinlock.
614 struct list_head counter_list;
615 struct list_head event_list;
621 struct task_struct *task;
624 * Context clock, runs when context enabled.
630 * These fields let us detect when two contexts have both
631 * been cloned (inherited) from a common ancestor.
633 struct perf_counter_context *parent_ctx;
637 struct rcu_head rcu_head;
641 * struct perf_counter_cpu_context - per cpu counter context structure
643 struct perf_cpu_context {
644 struct perf_counter_context ctx;
645 struct perf_counter_context *task_ctx;
651 * Recursion avoidance:
653 * task, softirq, irq, nmi context
658 #ifdef CONFIG_PERF_COUNTERS
661 * Set by architecture code:
663 extern int perf_max_counters;
665 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
667 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
668 extern void perf_counter_task_sched_out(struct task_struct *task,
669 struct task_struct *next, int cpu);
670 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
671 extern int perf_counter_init_task(struct task_struct *child);
672 extern void perf_counter_exit_task(struct task_struct *child);
673 extern void perf_counter_free_task(struct task_struct *task);
674 extern void set_perf_counter_pending(void);
675 extern void perf_counter_do_pending(void);
676 extern void perf_counter_print_debug(void);
677 extern void __perf_disable(void);
678 extern bool __perf_enable(void);
679 extern void perf_disable(void);
680 extern void perf_enable(void);
681 extern int perf_counter_task_disable(void);
682 extern int perf_counter_task_enable(void);
683 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
684 struct perf_cpu_context *cpuctx,
685 struct perf_counter_context *ctx, int cpu);
686 extern void perf_counter_update_userpage(struct perf_counter *counter);
688 struct perf_sample_data {
689 struct pt_regs *regs;
692 struct perf_raw_record *raw;
695 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
696 struct perf_sample_data *data);
699 * Return 1 for a software counter, 0 for a hardware counter
701 static inline int is_software_counter(struct perf_counter *counter)
703 return (counter->attr.type != PERF_TYPE_RAW) &&
704 (counter->attr.type != PERF_TYPE_HARDWARE) &&
705 (counter->attr.type != PERF_TYPE_HW_CACHE);
708 extern atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
710 extern void __perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
713 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
715 if (atomic_read(&perf_swcounter_enabled[event]))
716 __perf_swcounter_event(event, nr, nmi, regs, addr);
719 extern void __perf_counter_mmap(struct vm_area_struct *vma);
721 static inline void perf_counter_mmap(struct vm_area_struct *vma)
723 if (vma->vm_flags & VM_EXEC)
724 __perf_counter_mmap(vma);
727 extern void perf_counter_comm(struct task_struct *tsk);
728 extern void perf_counter_fork(struct task_struct *tsk);
730 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
732 extern int sysctl_perf_counter_paranoid;
733 extern int sysctl_perf_counter_mlock;
734 extern int sysctl_perf_counter_sample_rate;
736 extern void perf_counter_init(void);
738 #ifndef perf_misc_flags
739 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
740 PERF_EVENT_MISC_KERNEL)
741 #define perf_instruction_pointer(regs) instruction_pointer(regs)
746 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
748 perf_counter_task_sched_out(struct task_struct *task,
749 struct task_struct *next, int cpu) { }
751 perf_counter_task_tick(struct task_struct *task, int cpu) { }
752 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
753 static inline void perf_counter_exit_task(struct task_struct *child) { }
754 static inline void perf_counter_free_task(struct task_struct *task) { }
755 static inline void perf_counter_do_pending(void) { }
756 static inline void perf_counter_print_debug(void) { }
757 static inline void perf_disable(void) { }
758 static inline void perf_enable(void) { }
759 static inline int perf_counter_task_disable(void) { return -EINVAL; }
760 static inline int perf_counter_task_enable(void) { return -EINVAL; }
763 perf_swcounter_event(u32 event, u64 nr, int nmi,
764 struct pt_regs *regs, u64 addr) { }
766 static inline void perf_counter_mmap(struct vm_area_struct *vma) { }
767 static inline void perf_counter_comm(struct task_struct *tsk) { }
768 static inline void perf_counter_fork(struct task_struct *tsk) { }
769 static inline void perf_counter_init(void) { }
772 #endif /* __KERNEL__ */
773 #endif /* _LINUX_PERF_COUNTER_H */