perf_counter: add an mmap method to allow userspace to read hardware counters
[safe/jmp/linux-2.6] / include / linux / perf_counter.h
1 /*
2  *  Performance counters:
3  *
4  *   Copyright(C) 2008, Thomas Gleixner <tglx@linutronix.de>
5  *   Copyright(C) 2008, Red Hat, Inc., Ingo Molnar
6  *
7  *  Data type definitions, declarations, prototypes.
8  *
9  *  Started by: Thomas Gleixner and Ingo Molnar
10  *
11  *  For licencing details see kernel-base/COPYING
12  */
13 #ifndef _LINUX_PERF_COUNTER_H
14 #define _LINUX_PERF_COUNTER_H
15
16 #include <linux/types.h>
17 #include <linux/ioctl.h>
18 #include <asm/byteorder.h>
19
20 /*
21  * User-space ABI bits:
22  */
23
24 /*
25  * hw_event.type
26  */
27 enum perf_event_types {
28         PERF_TYPE_HARDWARE              = 0,
29         PERF_TYPE_SOFTWARE              = 1,
30         PERF_TYPE_TRACEPOINT            = 2,
31
32         /*
33          * available TYPE space, raw is the max value.
34          */
35
36         PERF_TYPE_RAW                   = 128,
37 };
38
39 /*
40  * Generalized performance counter event types, used by the hw_event.event_id
41  * parameter of the sys_perf_counter_open() syscall:
42  */
43 enum hw_event_ids {
44         /*
45          * Common hardware events, generalized by the kernel:
46          */
47         PERF_COUNT_CPU_CYCLES           = 0,
48         PERF_COUNT_INSTRUCTIONS         = 1,
49         PERF_COUNT_CACHE_REFERENCES     = 2,
50         PERF_COUNT_CACHE_MISSES         = 3,
51         PERF_COUNT_BRANCH_INSTRUCTIONS  = 4,
52         PERF_COUNT_BRANCH_MISSES        = 5,
53         PERF_COUNT_BUS_CYCLES           = 6,
54
55         PERF_HW_EVENTS_MAX              = 7,
56 };
57
58 /*
59  * Special "software" counters provided by the kernel, even if the hardware
60  * does not support performance counters. These counters measure various
61  * physical and sw events of the kernel (and allow the profiling of them as
62  * well):
63  */
64 enum sw_event_ids {
65         PERF_COUNT_CPU_CLOCK            = 0,
66         PERF_COUNT_TASK_CLOCK           = 1,
67         PERF_COUNT_PAGE_FAULTS          = 2,
68         PERF_COUNT_CONTEXT_SWITCHES     = 3,
69         PERF_COUNT_CPU_MIGRATIONS       = 4,
70         PERF_COUNT_PAGE_FAULTS_MIN      = 5,
71         PERF_COUNT_PAGE_FAULTS_MAJ      = 6,
72
73         PERF_SW_EVENTS_MAX              = 7,
74 };
75
76 /*
77  * IRQ-notification data record type:
78  */
79 enum perf_counter_record_type {
80         PERF_RECORD_SIMPLE              = 0,
81         PERF_RECORD_IRQ                 = 1,
82         PERF_RECORD_GROUP               = 2,
83 };
84
85 #define __PERF_COUNTER_MASK(name)                       \
86         (((1ULL << PERF_COUNTER_##name##_BITS) - 1) <<  \
87          PERF_COUNTER_##name##_SHIFT)
88
89 #define PERF_COUNTER_RAW_BITS           1
90 #define PERF_COUNTER_RAW_SHIFT          63
91 #define PERF_COUNTER_RAW_MASK           __PERF_COUNTER_MASK(RAW)
92
93 #define PERF_COUNTER_CONFIG_BITS        63
94 #define PERF_COUNTER_CONFIG_SHIFT       0
95 #define PERF_COUNTER_CONFIG_MASK        __PERF_COUNTER_MASK(CONFIG)
96
97 #define PERF_COUNTER_TYPE_BITS          7
98 #define PERF_COUNTER_TYPE_SHIFT         56
99 #define PERF_COUNTER_TYPE_MASK          __PERF_COUNTER_MASK(TYPE)
100
101 #define PERF_COUNTER_EVENT_BITS         56
102 #define PERF_COUNTER_EVENT_SHIFT        0
103 #define PERF_COUNTER_EVENT_MASK         __PERF_COUNTER_MASK(EVENT)
104
105 /*
106  * Hardware event to monitor via a performance monitoring counter:
107  */
108 struct perf_counter_hw_event {
109         /*
110          * The MSB of the config word signifies if the rest contains cpu
111          * specific (raw) counter configuration data, if unset, the next
112          * 7 bits are an event type and the rest of the bits are the event
113          * identifier.
114          */
115         __u64                   config;
116
117         __u64                   irq_period;
118         __u64                   record_type;
119         __u64                   read_format;
120
121         __u64                   disabled       :  1, /* off by default        */
122                                 nmi            :  1, /* NMI sampling          */
123                                 inherit        :  1, /* children inherit it   */
124                                 pinned         :  1, /* must always be on PMU */
125                                 exclusive      :  1, /* only group on PMU     */
126                                 exclude_user   :  1, /* don't count user      */
127                                 exclude_kernel :  1, /* ditto kernel          */
128                                 exclude_hv     :  1, /* ditto hypervisor      */
129                                 exclude_idle   :  1, /* don't count when idle */
130
131                                 __reserved_1   : 55;
132
133         __u32                   extra_config_len;
134         __u32                   __reserved_4;
135
136         __u64                   __reserved_2;
137         __u64                   __reserved_3;
138 };
139
140 /*
141  * Ioctls that can be done on a perf counter fd:
142  */
143 #define PERF_COUNTER_IOC_ENABLE         _IO('$', 0)
144 #define PERF_COUNTER_IOC_DISABLE        _IO('$', 1)
145
146 /*
147  * Structure of the page that can be mapped via mmap
148  */
149 struct perf_counter_mmap_page {
150         __u32   version;                /* version number of this structure */
151         __u32   compat_version;         /* lowest version this is compat with */
152         __u32   lock;                   /* seqlock for synchronization */
153         __u32   index;                  /* hardware counter identifier */
154         __s64   offset;                 /* add to hardware counter value */
155 };
156
157 #ifdef __KERNEL__
158 /*
159  * Kernel-internal data types and definitions:
160  */
161
162 #ifdef CONFIG_PERF_COUNTERS
163 # include <asm/perf_counter.h>
164 #endif
165
166 #include <linux/list.h>
167 #include <linux/mutex.h>
168 #include <linux/rculist.h>
169 #include <linux/rcupdate.h>
170 #include <linux/spinlock.h>
171 #include <linux/hrtimer.h>
172 #include <asm/atomic.h>
173
174 struct task_struct;
175
176 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
177 {
178         return hw_event->config & PERF_COUNTER_RAW_MASK;
179 }
180
181 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
182 {
183         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
184 }
185
186 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
187 {
188         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
189                 PERF_COUNTER_TYPE_SHIFT;
190 }
191
192 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
193 {
194         return hw_event->config & PERF_COUNTER_EVENT_MASK;
195 }
196
197 /**
198  * struct hw_perf_counter - performance counter hardware details:
199  */
200 struct hw_perf_counter {
201 #ifdef CONFIG_PERF_COUNTERS
202         union {
203                 struct { /* hardware */
204                         u64                             config;
205                         unsigned long                   config_base;
206                         unsigned long                   counter_base;
207                         int                             nmi;
208                         unsigned int                    idx;
209                 };
210                 union { /* software */
211                         atomic64_t                      count;
212                         struct hrtimer                  hrtimer;
213                 };
214         };
215         atomic64_t                      prev_count;
216         u64                             irq_period;
217         atomic64_t                      period_left;
218 #endif
219 };
220
221 /*
222  * Hardcoded buffer length limit for now, for IRQ-fed events:
223  */
224 #define PERF_DATA_BUFLEN                2048
225
226 /**
227  * struct perf_data - performance counter IRQ data sampling ...
228  */
229 struct perf_data {
230         int                             len;
231         int                             rd_idx;
232         int                             overrun;
233         u8                              data[PERF_DATA_BUFLEN];
234 };
235
236 struct perf_counter;
237
238 /**
239  * struct hw_perf_counter_ops - performance counter hw ops
240  */
241 struct hw_perf_counter_ops {
242         int (*enable)                   (struct perf_counter *counter);
243         void (*disable)                 (struct perf_counter *counter);
244         void (*read)                    (struct perf_counter *counter);
245 };
246
247 /**
248  * enum perf_counter_active_state - the states of a counter
249  */
250 enum perf_counter_active_state {
251         PERF_COUNTER_STATE_ERROR        = -2,
252         PERF_COUNTER_STATE_OFF          = -1,
253         PERF_COUNTER_STATE_INACTIVE     =  0,
254         PERF_COUNTER_STATE_ACTIVE       =  1,
255 };
256
257 struct file;
258
259 /**
260  * struct perf_counter - performance counter kernel representation:
261  */
262 struct perf_counter {
263 #ifdef CONFIG_PERF_COUNTERS
264         struct list_head                list_entry;
265         struct list_head                event_entry;
266         struct list_head                sibling_list;
267         struct perf_counter             *group_leader;
268         const struct hw_perf_counter_ops *hw_ops;
269
270         enum perf_counter_active_state  state;
271         enum perf_counter_active_state  prev_state;
272         atomic64_t                      count;
273
274         struct perf_counter_hw_event    hw_event;
275         struct hw_perf_counter          hw;
276
277         struct perf_counter_context     *ctx;
278         struct task_struct              *task;
279         struct file                     *filp;
280
281         struct perf_counter             *parent;
282         struct list_head                child_list;
283
284         /*
285          * Protect attach/detach and child_list:
286          */
287         struct mutex                    mutex;
288
289         int                             oncpu;
290         int                             cpu;
291
292         /* pointer to page shared with userspace via mmap */
293         unsigned long                   user_page;
294
295         /* read() / irq related data */
296         wait_queue_head_t               waitq;
297         /* optional: for NMIs */
298         int                             wakeup_pending;
299         struct perf_data                *irqdata;
300         struct perf_data                *usrdata;
301         struct perf_data                data[2];
302
303         void (*destroy)(struct perf_counter *);
304         struct rcu_head                 rcu_head;
305 #endif
306 };
307
308 /**
309  * struct perf_counter_context - counter context structure
310  *
311  * Used as a container for task counters and CPU counters as well:
312  */
313 struct perf_counter_context {
314 #ifdef CONFIG_PERF_COUNTERS
315         /*
316          * Protect the states of the counters in the list,
317          * nr_active, and the list:
318          */
319         spinlock_t              lock;
320         /*
321          * Protect the list of counters.  Locking either mutex or lock
322          * is sufficient to ensure the list doesn't change; to change
323          * the list you need to lock both the mutex and the spinlock.
324          */
325         struct mutex            mutex;
326
327         struct list_head        counter_list;
328         struct list_head        event_list;
329         int                     nr_counters;
330         int                     nr_active;
331         int                     is_active;
332         struct task_struct      *task;
333 #endif
334 };
335
336 /**
337  * struct perf_counter_cpu_context - per cpu counter context structure
338  */
339 struct perf_cpu_context {
340         struct perf_counter_context     ctx;
341         struct perf_counter_context     *task_ctx;
342         int                             active_oncpu;
343         int                             max_pertask;
344         int                             exclusive;
345
346         /*
347          * Recursion avoidance:
348          *
349          * task, softirq, irq, nmi context
350          */
351         int                     recursion[4];
352 };
353
354 /*
355  * Set by architecture code:
356  */
357 extern int perf_max_counters;
358
359 #ifdef CONFIG_PERF_COUNTERS
360 extern const struct hw_perf_counter_ops *
361 hw_perf_counter_init(struct perf_counter *counter);
362
363 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
364 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
365 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
366 extern void perf_counter_init_task(struct task_struct *child);
367 extern void perf_counter_exit_task(struct task_struct *child);
368 extern void perf_counter_notify(struct pt_regs *regs);
369 extern void perf_counter_print_debug(void);
370 extern void perf_counter_unthrottle(void);
371 extern u64 hw_perf_save_disable(void);
372 extern void hw_perf_restore(u64 ctrl);
373 extern int perf_counter_task_disable(void);
374 extern int perf_counter_task_enable(void);
375 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
376                struct perf_cpu_context *cpuctx,
377                struct perf_counter_context *ctx, int cpu);
378 extern void perf_counter_update_userpage(struct perf_counter *counter);
379
380 extern void perf_counter_output(struct perf_counter *counter,
381                                 int nmi, struct pt_regs *regs);
382 /*
383  * Return 1 for a software counter, 0 for a hardware counter
384  */
385 static inline int is_software_counter(struct perf_counter *counter)
386 {
387         return !perf_event_raw(&counter->hw_event) &&
388                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
389 }
390
391 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *);
392
393 #else
394 static inline void
395 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
396 static inline void
397 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
398 static inline void
399 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
400 static inline void perf_counter_init_task(struct task_struct *child)    { }
401 static inline void perf_counter_exit_task(struct task_struct *child)    { }
402 static inline void perf_counter_notify(struct pt_regs *regs)            { }
403 static inline void perf_counter_print_debug(void)                       { }
404 static inline void perf_counter_unthrottle(void)                        { }
405 static inline void hw_perf_restore(u64 ctrl)                            { }
406 static inline u64 hw_perf_save_disable(void)                  { return 0; }
407 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
408 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
409
410 static inline void perf_swcounter_event(u32 event, u64 nr,
411                                         int nmi, struct pt_regs *regs)  { }
412 #endif
413
414 #endif /* __KERNEL__ */
415 #endif /* _LINUX_PERF_COUNTER_H */