2 * Performance counters:
4 * Copyright(C) 2008, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008, Red Hat, Inc., Ingo Molnar
7 * Data type definitions, declarations, prototypes.
9 * Started by: Thomas Gleixner and Ingo Molnar
11 * For licencing details see kernel-base/COPYING
13 #ifndef _LINUX_PERF_COUNTER_H
14 #define _LINUX_PERF_COUNTER_H
16 #include <linux/types.h>
17 #include <linux/ioctl.h>
18 #include <asm/byteorder.h>
21 * User-space ABI bits:
27 enum perf_event_types {
28 PERF_TYPE_HARDWARE = 0,
29 PERF_TYPE_SOFTWARE = 1,
30 PERF_TYPE_TRACEPOINT = 2,
31 PERF_TYPE_HW_CACHE = 3,
34 * available TYPE space, raw is the max value.
41 * Generalized performance counter event types, used by the attr.event_id
42 * parameter of the sys_perf_counter_open() syscall:
46 * Common hardware events, generalized by the kernel:
48 PERF_COUNT_CPU_CYCLES = 0,
49 PERF_COUNT_INSTRUCTIONS = 1,
50 PERF_COUNT_CACHE_REFERENCES = 2,
51 PERF_COUNT_CACHE_MISSES = 3,
52 PERF_COUNT_BRANCH_INSTRUCTIONS = 4,
53 PERF_COUNT_BRANCH_MISSES = 5,
54 PERF_COUNT_BUS_CYCLES = 6,
56 PERF_HW_EVENTS_MAX = 7,
60 * Generalized hardware cache counters:
62 * { L1-D, L1-I, L2, LLC, ITLB, DTLB, BPU } x
63 * { read, write, prefetch } x
64 * { accesses, misses }
67 PERF_COUNT_HW_CACHE_L1D,
68 PERF_COUNT_HW_CACHE_L1I,
69 PERF_COUNT_HW_CACHE_L2,
70 PERF_COUNT_HW_CACHE_DTLB,
71 PERF_COUNT_HW_CACHE_ITLB,
72 PERF_COUNT_HW_CACHE_BPU,
74 PERF_COUNT_HW_CACHE_MAX,
78 PERF_COUNT_HW_CACHE_OP_READ,
79 PERF_COUNT_HW_CACHE_OP_WRITE,
80 PERF_COUNT_HW_CACHE_OP_PREFETCH,
82 PERF_COUNT_HW_CACHE_OP_MAX,
85 enum hw_cache_op_result_id {
86 PERF_COUNT_HW_CACHE_RESULT_ACCESS,
87 PERF_COUNT_HW_CACHE_RESULT_MISS,
89 PERF_COUNT_HW_CACHE_RESULT_MAX,
93 * Special "software" counters provided by the kernel, even if the hardware
94 * does not support performance counters. These counters measure various
95 * physical and sw events of the kernel (and allow the profiling of them as
99 PERF_COUNT_CPU_CLOCK = 0,
100 PERF_COUNT_TASK_CLOCK = 1,
101 PERF_COUNT_PAGE_FAULTS = 2,
102 PERF_COUNT_CONTEXT_SWITCHES = 3,
103 PERF_COUNT_CPU_MIGRATIONS = 4,
104 PERF_COUNT_PAGE_FAULTS_MIN = 5,
105 PERF_COUNT_PAGE_FAULTS_MAJ = 6,
107 PERF_SW_EVENTS_MAX = 7,
111 * Bits that can be set in attr.sample_type to request information
112 * in the overflow packets.
114 enum perf_counter_sample_format {
115 PERF_SAMPLE_IP = 1U << 0,
116 PERF_SAMPLE_TID = 1U << 1,
117 PERF_SAMPLE_TIME = 1U << 2,
118 PERF_SAMPLE_ADDR = 1U << 3,
119 PERF_SAMPLE_GROUP = 1U << 4,
120 PERF_SAMPLE_CALLCHAIN = 1U << 5,
121 PERF_SAMPLE_ID = 1U << 6,
122 PERF_SAMPLE_CPU = 1U << 7,
123 PERF_SAMPLE_PERIOD = 1U << 8,
127 * Bits that can be set in attr.read_format to request that
128 * reads on the counter should return the indicated quantities,
129 * in increasing order of bit value, after the counter value.
131 enum perf_counter_read_format {
132 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
133 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
134 PERF_FORMAT_ID = 1U << 2,
138 * Hardware event to monitor via a performance monitoring counter:
140 struct perf_counter_attr {
142 * Major type: hardware/software/tracepoint/etc.
148 * Type specific configuration information.
160 __u64 disabled : 1, /* off by default */
161 inherit : 1, /* children inherit it */
162 pinned : 1, /* must always be on PMU */
163 exclusive : 1, /* only group on PMU */
164 exclude_user : 1, /* don't count user */
165 exclude_kernel : 1, /* ditto kernel */
166 exclude_hv : 1, /* ditto hypervisor */
167 exclude_idle : 1, /* don't count when idle */
168 mmap : 1, /* include mmap data */
169 comm : 1, /* include comm data */
170 freq : 1, /* use freq, not period */
174 __u32 wakeup_events; /* wakeup every n events */
181 * Ioctls that can be done on a perf counter fd:
183 #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0)
184 #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1)
185 #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2)
186 #define PERF_COUNTER_IOC_RESET _IO ('$', 3)
187 #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64)
189 enum perf_counter_ioc_flags {
190 PERF_IOC_FLAG_GROUP = 1U << 0,
194 * Structure of the page that can be mapped via mmap
196 struct perf_counter_mmap_page {
197 __u32 version; /* version number of this structure */
198 __u32 compat_version; /* lowest version this is compat with */
201 * Bits needed to read the hw counters in user-space.
211 * count = pmc_read(pc->index - 1);
212 * count += pc->offset;
217 * } while (pc->lock != seq);
219 * NOTE: for obvious reason this only works on self-monitoring
222 __u32 lock; /* seqlock for synchronization */
223 __u32 index; /* hardware counter identifier */
224 __s64 offset; /* add to hardware counter value */
227 * Control data for the mmap() data buffer.
229 * User-space reading this value should issue an rmb(), on SMP capable
230 * platforms, after reading this value -- see perf_counter_wakeup().
232 __u64 data_head; /* head in the data section */
235 #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0)
236 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
237 #define PERF_EVENT_MISC_KERNEL (1 << 0)
238 #define PERF_EVENT_MISC_USER (2 << 0)
239 #define PERF_EVENT_MISC_HYPERVISOR (3 << 0)
240 #define PERF_EVENT_MISC_OVERFLOW (1 << 2)
242 struct perf_event_header {
248 enum perf_event_type {
251 * The MMAP events record the PROT_EXEC mappings so that we can
252 * correlate userspace IPs to code. They have the following structure:
255 * struct perf_event_header header;
268 * struct perf_event_header header;
278 * struct perf_event_header header;
284 PERF_EVENT_PERIOD = 4,
288 * struct perf_event_header header;
292 PERF_EVENT_THROTTLE = 5,
293 PERF_EVENT_UNTHROTTLE = 6,
297 * struct perf_event_header header;
304 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
305 * will be PERF_RECORD_*
308 * struct perf_event_header header;
310 * { u64 ip; } && PERF_RECORD_IP
311 * { u32 pid, tid; } && PERF_RECORD_TID
312 * { u64 time; } && PERF_RECORD_TIME
313 * { u64 addr; } && PERF_RECORD_ADDR
314 * { u64 config; } && PERF_RECORD_CONFIG
315 * { u32 cpu, res; } && PERF_RECORD_CPU
318 * { u64 id, val; } cnt[nr]; } && PERF_RECORD_GROUP
324 * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN
331 * Kernel-internal data types and definitions:
334 #ifdef CONFIG_PERF_COUNTERS
335 # include <asm/perf_counter.h>
338 #include <linux/list.h>
339 #include <linux/mutex.h>
340 #include <linux/rculist.h>
341 #include <linux/rcupdate.h>
342 #include <linux/spinlock.h>
343 #include <linux/hrtimer.h>
344 #include <linux/fs.h>
345 #include <linux/pid_namespace.h>
346 #include <asm/atomic.h>
351 * struct hw_perf_counter - performance counter hardware details:
353 struct hw_perf_counter {
354 #ifdef CONFIG_PERF_COUNTERS
356 struct { /* hardware */
358 unsigned long config_base;
359 unsigned long counter_base;
362 union { /* software */
364 struct hrtimer hrtimer;
367 atomic64_t prev_count;
369 atomic64_t period_left;
380 * struct pmu - generic performance monitoring unit
383 int (*enable) (struct perf_counter *counter);
384 void (*disable) (struct perf_counter *counter);
385 void (*read) (struct perf_counter *counter);
386 void (*unthrottle) (struct perf_counter *counter);
390 * enum perf_counter_active_state - the states of a counter
392 enum perf_counter_active_state {
393 PERF_COUNTER_STATE_ERROR = -2,
394 PERF_COUNTER_STATE_OFF = -1,
395 PERF_COUNTER_STATE_INACTIVE = 0,
396 PERF_COUNTER_STATE_ACTIVE = 1,
401 struct perf_mmap_data {
402 struct rcu_head rcu_head;
403 int nr_pages; /* nr of data pages */
404 int nr_locked; /* nr pages mlocked */
406 atomic_t poll; /* POLL_ for wakeups */
407 atomic_t events; /* event limit */
409 atomic_long_t head; /* write position */
410 atomic_long_t done_head; /* completed head */
412 atomic_t lock; /* concurrent writes */
414 atomic_t wakeup; /* needs a wakeup */
416 struct perf_counter_mmap_page *user_page;
420 struct perf_pending_entry {
421 struct perf_pending_entry *next;
422 void (*func)(struct perf_pending_entry *);
426 * struct perf_counter - performance counter kernel representation:
428 struct perf_counter {
429 #ifdef CONFIG_PERF_COUNTERS
430 struct list_head list_entry;
431 struct list_head event_entry;
432 struct list_head sibling_list;
434 struct perf_counter *group_leader;
435 const struct pmu *pmu;
437 enum perf_counter_active_state state;
441 * These are the total time in nanoseconds that the counter
442 * has been enabled (i.e. eligible to run, and the task has
443 * been scheduled in, if this is a per-task counter)
444 * and running (scheduled onto the CPU), respectively.
446 * They are computed from tstamp_enabled, tstamp_running and
447 * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
449 u64 total_time_enabled;
450 u64 total_time_running;
453 * These are timestamps used for computing total_time_enabled
454 * and total_time_running when the counter is in INACTIVE or
455 * ACTIVE state, measured in nanoseconds from an arbitrary point
457 * tstamp_enabled: the notional time when the counter was enabled
458 * tstamp_running: the notional time when the counter was scheduled on
459 * tstamp_stopped: in INACTIVE state, the notional time when the
460 * counter was scheduled off.
466 struct perf_counter_attr attr;
467 struct hw_perf_counter hw;
469 struct perf_counter_context *ctx;
473 * These accumulate total time (in nanoseconds) that children
474 * counters have been enabled and running, respectively.
476 atomic64_t child_total_time_enabled;
477 atomic64_t child_total_time_running;
480 * Protect attach/detach and child_list:
482 struct mutex child_mutex;
483 struct list_head child_list;
484 struct perf_counter *parent;
489 struct list_head owner_entry;
490 struct task_struct *owner;
493 struct mutex mmap_mutex;
495 struct perf_mmap_data *data;
498 wait_queue_head_t waitq;
499 struct fasync_struct *fasync;
501 /* delayed work for NMIs and such */
505 struct perf_pending_entry pending;
507 atomic_t event_limit;
509 void (*destroy)(struct perf_counter *);
510 struct rcu_head rcu_head;
512 struct pid_namespace *ns;
518 * struct perf_counter_context - counter context structure
520 * Used as a container for task counters and CPU counters as well:
522 struct perf_counter_context {
524 * Protect the states of the counters in the list,
525 * nr_active, and the list:
529 * Protect the list of counters. Locking either mutex or lock
530 * is sufficient to ensure the list doesn't change; to change
531 * the list you need to lock both the mutex and the spinlock.
535 struct list_head counter_list;
536 struct list_head event_list;
541 struct task_struct *task;
544 * Context clock, runs when context enabled.
550 * These fields let us detect when two contexts have both
551 * been cloned (inherited) from a common ancestor.
553 struct perf_counter_context *parent_ctx;
557 struct rcu_head rcu_head;
561 * struct perf_counter_cpu_context - per cpu counter context structure
563 struct perf_cpu_context {
564 struct perf_counter_context ctx;
565 struct perf_counter_context *task_ctx;
571 * Recursion avoidance:
573 * task, softirq, irq, nmi context
578 #ifdef CONFIG_PERF_COUNTERS
581 * Set by architecture code:
583 extern int perf_max_counters;
585 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
587 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
588 extern void perf_counter_task_sched_out(struct task_struct *task,
589 struct task_struct *next, int cpu);
590 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
591 extern int perf_counter_init_task(struct task_struct *child);
592 extern void perf_counter_exit_task(struct task_struct *child);
593 extern void perf_counter_free_task(struct task_struct *task);
594 extern void perf_counter_do_pending(void);
595 extern void perf_counter_print_debug(void);
596 extern void __perf_disable(void);
597 extern bool __perf_enable(void);
598 extern void perf_disable(void);
599 extern void perf_enable(void);
600 extern int perf_counter_task_disable(void);
601 extern int perf_counter_task_enable(void);
602 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
603 struct perf_cpu_context *cpuctx,
604 struct perf_counter_context *ctx, int cpu);
605 extern void perf_counter_update_userpage(struct perf_counter *counter);
607 extern int perf_counter_overflow(struct perf_counter *counter,
608 int nmi, struct pt_regs *regs, u64 addr);
610 * Return 1 for a software counter, 0 for a hardware counter
612 static inline int is_software_counter(struct perf_counter *counter)
614 return (counter->attr.type != PERF_TYPE_RAW) &&
615 (counter->attr.type != PERF_TYPE_HARDWARE);
618 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
620 extern void __perf_counter_mmap(struct vm_area_struct *vma);
622 static inline void perf_counter_mmap(struct vm_area_struct *vma)
624 if (vma->vm_flags & VM_EXEC)
625 __perf_counter_mmap(vma);
628 extern void perf_counter_comm(struct task_struct *tsk);
629 extern void perf_counter_fork(struct task_struct *tsk);
631 extern void perf_counter_task_migration(struct task_struct *task, int cpu);
633 #define MAX_STACK_DEPTH 255
635 struct perf_callchain_entry {
636 u16 nr, hv, kernel, user;
637 u64 ip[MAX_STACK_DEPTH];
640 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
642 extern int sysctl_perf_counter_priv;
643 extern int sysctl_perf_counter_mlock;
644 extern int sysctl_perf_counter_limit;
646 extern void perf_counter_init(void);
648 #ifndef perf_misc_flags
649 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_EVENT_MISC_USER : \
650 PERF_EVENT_MISC_KERNEL)
651 #define perf_instruction_pointer(regs) instruction_pointer(regs)
656 perf_counter_task_sched_in(struct task_struct *task, int cpu) { }
658 perf_counter_task_sched_out(struct task_struct *task,
659 struct task_struct *next, int cpu) { }
661 perf_counter_task_tick(struct task_struct *task, int cpu) { }
662 static inline int perf_counter_init_task(struct task_struct *child) { return 0; }
663 static inline void perf_counter_exit_task(struct task_struct *child) { }
664 static inline void perf_counter_free_task(struct task_struct *task) { }
665 static inline void perf_counter_do_pending(void) { }
666 static inline void perf_counter_print_debug(void) { }
667 static inline void perf_disable(void) { }
668 static inline void perf_enable(void) { }
669 static inline int perf_counter_task_disable(void) { return -EINVAL; }
670 static inline int perf_counter_task_enable(void) { return -EINVAL; }
673 perf_swcounter_event(u32 event, u64 nr, int nmi,
674 struct pt_regs *regs, u64 addr) { }
676 static inline void perf_counter_mmap(struct vm_area_struct *vma) { }
677 static inline void perf_counter_comm(struct task_struct *tsk) { }
678 static inline void perf_counter_fork(struct task_struct *tsk) { }
679 static inline void perf_counter_init(void) { }
680 static inline void perf_counter_task_migration(struct task_struct *task,
684 #endif /* __KERNEL__ */
685 #endif /* _LINUX_PERF_COUNTER_H */