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 */
237 * Control data for the mmap() data buffer.
239 * User-space reading this value should issue an rmb(), on SMP capable
240 * platforms, after reading this value -- see perf_counter_wakeup().
242 __u64 data_head; /* head in the data section */
245 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
246 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
247 #define PERF_EVENT_MISC_KERNEL (1 << 0)
248 #define PERF_EVENT_MISC_USER (2 << 0)
249 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
250 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
252 struct perf_event_header {
258 enum perf_event_type {
261 * The MMAP events record the PROT_EXEC mappings so that we can
262 * correlate userspace IPs to code. They have the following structure:
265 * struct perf_event_header header;
278 * struct perf_event_header header;
288 * struct perf_event_header header;
294 PERF_EVENT_PERIOD = 4,
298 * struct perf_event_header header;
303 PERF_EVENT_THROTTLE = 5,
304 PERF_EVENT_UNTHROTTLE = 6,
308 * struct perf_event_header header;
315 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
316 * will be PERF_RECORD_*
319 * struct perf_event_header header;
321 * { u64 ip; } && PERF_RECORD_IP
322 * { u32 pid, tid; } && PERF_RECORD_TID
323 * { u64 time; } && PERF_RECORD_TIME
324 * { u64 addr; } && PERF_RECORD_ADDR
325 * { u64 config; } && PERF_RECORD_CONFIG
326 * { u32 cpu, res; } && PERF_RECORD_CPU
329 * { u64 id, val; } cnt[nr]; } && PERF_RECORD_GROUP
335 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
342 * Kernel-internal data types and definitions:
345 #ifdef CONFIG_PERF_COUNTERS
346 # include <asm/perf_counter.h>
349 #include <linux/list.h>
350 #include <linux/mutex.h>
351 #include <linux/rculist.h>
352 #include <linux/rcupdate.h>
353 #include <linux/spinlock.h>
354 #include <linux/hrtimer.h>
355 #include <linux/fs.h>
356 #include <linux/pid_namespace.h>
357 #include <asm/atomic.h>
362 * struct hw_perf_counter - performance counter hardware details:
364 struct hw_perf_counter {
365 #ifdef CONFIG_PERF_COUNTERS
367 struct { /* hardware */
369 unsigned long config_base;
370 unsigned long counter_base;
373 union { /* software */
375 struct hrtimer hrtimer;
378 atomic64_t prev_count;
381 atomic64_t period_left;
393 * struct pmu - generic performance monitoring unit
396 int (*enable) (struct perf_counter *counter);
397 void (*disable) (struct perf_counter *counter);
398 void (*read) (struct perf_counter *counter);
399 void (*unthrottle) (struct perf_counter *counter);
403 * enum perf_counter_active_state - the states of a counter
405 enum perf_counter_active_state {
406 PERF_COUNTER_STATE_ERROR = -2,
407 PERF_COUNTER_STATE_OFF = -1,
408 PERF_COUNTER_STATE_INACTIVE = 0,
409 PERF_COUNTER_STATE_ACTIVE = 1,
414 struct perf_mmap_data {
415 struct rcu_head rcu_head;
416 int nr_pages; /* nr of data pages */
417 int nr_locked; /* nr pages mlocked */
419 atomic_t poll; /* POLL_ for wakeups */
420 atomic_t events; /* event limit */
422 atomic_long_t head; /* write position */
423 atomic_long_t done_head; /* completed head */
425 atomic_t lock; /* concurrent writes */
427 atomic_t wakeup; /* needs a wakeup */
429 struct perf_counter_mmap_page *user_page;
433 struct perf_pending_entry {
434 struct perf_pending_entry *next;
435 void (*func)(struct perf_pending_entry *);
439 * struct perf_counter - performance counter kernel representation:
441 struct perf_counter {
442 #ifdef CONFIG_PERF_COUNTERS
443 struct list_head list_entry;
444 struct list_head event_entry;
445 struct list_head sibling_list;
447 struct perf_counter *group_leader;
448 const struct pmu *pmu;
450 enum perf_counter_active_state state;
454 * These are the total time in nanoseconds that the counter
455 * has been enabled (i.e. eligible to run, and the task has
456 * been scheduled in, if this is a per-task counter)
457 * and running (scheduled onto the CPU), respectively.
459 * They are computed from tstamp_enabled, tstamp_running and
460 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
462 u64 total_time_enabled;
463 u64 total_time_running;
466 * These are timestamps used for computing total_time_enabled
467 * and total_time_running when the counter is in INACTIVE or
468 * ACTIVE state, measured in nanoseconds from an arbitrary point
470 * tstamp_enabled: the notional time when the counter was enabled
471 * tstamp_running: the notional time when the counter was scheduled on
472 * tstamp_stopped: in INACTIVE state, the notional time when the
473 * counter was scheduled off.
479 struct perf_counter_attr attr;
480 struct hw_perf_counter hw;
482 struct perf_counter_context *ctx;
486 * These accumulate total time (in nanoseconds) that children
487 * counters have been enabled and running, respectively.
489 atomic64_t child_total_time_enabled;
490 atomic64_t child_total_time_running;
493 * Protect attach/detach and child_list:
495 struct mutex child_mutex;
496 struct list_head child_list;
497 struct perf_counter *parent;
502 struct list_head owner_entry;
503 struct task_struct *owner;
506 struct mutex mmap_mutex;
508 struct perf_mmap_data *data;
511 wait_queue_head_t waitq;
512 struct fasync_struct *fasync;
514 /* delayed work for NMIs and such */
518 struct perf_pending_entry pending;
520 atomic_t event_limit;
522 void (*destroy)(struct perf_counter *);
523 struct rcu_head rcu_head;
525 struct pid_namespace *ns;
531 * struct perf_counter_context - counter context structure
533 * Used as a container for task counters and CPU counters as well:
535 struct perf_counter_context {
537 * Protect the states of the counters in the list,
538 * nr_active, and the list:
542 * Protect the list of counters. Locking either mutex or lock
543 * is sufficient to ensure the list doesn't change; to change
544 * the list you need to lock both the mutex and the spinlock.
548 struct list_head counter_list;
549 struct list_head event_list;
554 struct task_struct *task;
557 * Context clock, runs when context enabled.
563 * These fields let us detect when two contexts have both
564 * been cloned (inherited) from a common ancestor.
566 struct perf_counter_context *parent_ctx;
570 struct rcu_head rcu_head;
574 * struct perf_counter_cpu_context - per cpu counter context structure
576 struct perf_cpu_context {
577 struct perf_counter_context ctx;
578 struct perf_counter_context *task_ctx;
584 * Recursion avoidance:
586 * task, softirq, irq, nmi context
591 #ifdef CONFIG_PERF_COUNTERS
594 * Set by architecture code:
596 extern int perf_max_counters;
598 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
600 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
601 extern void perf_counter_task_sched_out(struct task_struct *task,
602 struct task_struct *next, int cpu);
603 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
604 extern int perf_counter_init_task(struct task_struct *child);
605 extern void perf_counter_exit_task(struct task_struct *child);
606 extern void perf_counter_free_task(struct task_struct *task);
607 extern void perf_counter_do_pending(void);
608 extern void perf_counter_print_debug(void);
609 extern void __perf_disable(void);
610 extern bool __perf_enable(void);
611 extern void perf_disable(void);
612 extern void perf_enable(void);
613 extern int perf_counter_task_disable(void);
614 extern int perf_counter_task_enable(void);
615 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
616 struct perf_cpu_context *cpuctx,
617 struct perf_counter_context *ctx, int cpu);
618 extern void perf_counter_update_userpage(struct perf_counter *counter);
620 struct perf_sample_data {
621 struct pt_regs *regs;
626 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
627 struct perf_sample_data *data);
630 * Return 1 for a software counter, 0 for a hardware counter
632 static inline int is_software_counter(struct perf_counter *counter)
634 return (counter->attr.type != PERF_TYPE_RAW) &&
635 (counter->attr.type != PERF_TYPE_HARDWARE) &&
636 (counter->attr.type != PERF_TYPE_HW_CACHE);
639 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
641 extern void __perf_counter_mmap(struct vm_area_struct *vma);
643 static inline void perf_counter_mmap(struct vm_area_struct *vma)
645 if (vma->vm_flags & VM_EXEC)
646 __perf_counter_mmap(vma);
649 extern void perf_counter_comm(struct task_struct *tsk);
650 extern void perf_counter_fork(struct task_struct *tsk);
652 extern void perf_counter_task_migration(struct task_struct *task, int cpu);
654 #define MAX_STACK_DEPTH 255
656 struct perf_callchain_entry {
661 u64 ip[MAX_STACK_DEPTH];
664 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
666 extern int sysctl_perf_counter_paranoid;
667 extern int sysctl_perf_counter_mlock;
668 extern int sysctl_perf_counter_sample_rate;
670 extern void perf_counter_init(void);
672 #ifndef perf_misc_flags
673 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
674 PERF_EVENT_MISC_KERNEL)
675 #define perf_instruction_pointer(regs) instruction_pointer(regs)
680 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
682 perf_counter_task_sched_out(struct task_struct *task,
683 struct task_struct *next, int cpu) { }
685 perf_counter_task_tick(struct task_struct *task, int cpu) { }
686 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
687 static inline void perf_counter_exit_task(struct task_struct *child) { }
688 static inline void perf_counter_free_task(struct task_struct *task) { }
689 static inline void perf_counter_do_pending(void) { }
690 static inline void perf_counter_print_debug(void) { }
691 static inline void perf_disable(void) { }
692 static inline void perf_enable(void) { }
693 static inline int perf_counter_task_disable(void) { return -EINVAL; }
694 static inline int perf_counter_task_enable(void) { return -EINVAL; }
697 perf_swcounter_event(u32 event, u64 nr, int nmi,
698 struct pt_regs *regs, u64 addr) { }
700 static inline void perf_counter_mmap(struct vm_area_struct *vma) { }
701 static inline void perf_counter_comm(struct task_struct *tsk) { }
702 static inline void perf_counter_fork(struct task_struct *tsk) { }
703 static inline void perf_counter_init(void) { }
704 static inline void perf_counter_task_migration(struct task_struct *task,
708 #endif /* __KERNEL__ */
709 #endif /* _LINUX_PERF_COUNTER_H */