1 /*
2 * Performance events:
3 *
4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
7 *
8 * Data type definitions, declarations, prototypes.
9 *
10 * Started by: Thomas Gleixner and Ingo Molnar
11 *
12 * For licencing details see kernel-base/COPYING
13 */
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
16
17 #include <linux/types.h>
18 #include <linux/ioctl.h>
19 #include <asm/byteorder.h>
20
21 /*
22 * User-space ABI bits:
23 */
24
25 /*
26 * attr.type
27 */
28 enum perf_type_id {
29 PERF_TYPE_HARDWARE = 0,
30 PERF_TYPE_SOFTWARE = 1,
31 PERF_TYPE_TRACEPOINT = 2,
32 PERF_TYPE_HW_CACHE = 3,
33 PERF_TYPE_RAW = 4,
34 PERF_TYPE_BREAKPOINT = 5,
35
36 PERF_TYPE_MAX, /* non-ABI */
37 };
38
39 /*
40 * Generalized performance event event_id types, used by the
41 * attr.event_id parameter of the sys_perf_event_open()
42 * syscall:
43 */
44 enum perf_hw_id {
45 /*
46 * Common hardware events, generalized by the kernel:
47 */
48 PERF_COUNT_HW_CPU_CYCLES = 0,
49 PERF_COUNT_HW_INSTRUCTIONS = 1,
50 PERF_COUNT_HW_CACHE_REFERENCES = 2,
51 PERF_COUNT_HW_CACHE_MISSES = 3,
52 PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4,
53 PERF_COUNT_HW_BRANCH_MISSES = 5,
54 PERF_COUNT_HW_BUS_CYCLES = 6,
55 PERF_COUNT_HW_STALLED_CYCLES_FRONTEND = 7,
56 PERF_COUNT_HW_STALLED_CYCLES_BACKEND = 8,
57 PERF_COUNT_HW_REF_CPU_CYCLES = 9,
58
59 PERF_COUNT_HW_MAX, /* non-ABI */
60 };
61
62 /*
63 * Generalized hardware cache events:
64 *
65 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
66 * { read, write, prefetch } x
67 * { accesses, misses }
68 */
69 enum perf_hw_cache_id {
70 PERF_COUNT_HW_CACHE_L1D = 0,
71 PERF_COUNT_HW_CACHE_L1I = 1,
72 PERF_COUNT_HW_CACHE_LL = 2,
73 PERF_COUNT_HW_CACHE_DTLB = 3,
74 PERF_COUNT_HW_CACHE_ITLB = 4,
75 PERF_COUNT_HW_CACHE_BPU = 5,
76 PERF_COUNT_HW_CACHE_NODE = 6,
77
78 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */
79 };
80
81 enum perf_hw_cache_op_id {
82 PERF_COUNT_HW_CACHE_OP_READ = 0,
83 PERF_COUNT_HW_CACHE_OP_WRITE = 1,
84 PERF_COUNT_HW_CACHE_OP_PREFETCH = 2,
85
86 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */
87 };
88
89 enum perf_hw_cache_op_result_id {
90 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0,
91 PERF_COUNT_HW_CACHE_RESULT_MISS = 1,
92
93 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */
94 };
95
96 /*
97 * Special "software" events provided by the kernel, even if the hardware
98 * does not support performance events. These events measure various
99 * physical and sw events of the kernel (and allow the profiling of them as
100 * well):
101 */
102 enum perf_sw_ids {
103 PERF_COUNT_SW_CPU_CLOCK = 0,
104 PERF_COUNT_SW_TASK_CLOCK = 1,
105 PERF_COUNT_SW_PAGE_FAULTS = 2,
106 PERF_COUNT_SW_CONTEXT_SWITCHES = 3,
107 PERF_COUNT_SW_CPU_MIGRATIONS = 4,
108 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5,
109 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6,
110 PERF_COUNT_SW_ALIGNMENT_FAULTS = 7,
111 PERF_COUNT_SW_EMULATION_FAULTS = 8,
112
113 PERF_COUNT_SW_MAX, /* non-ABI */
114 };
115
116 /*
117 * Bits that can be set in attr.sample_type to request information
118 * in the overflow packets.
119 */
120 enum perf_event_sample_format {
121 PERF_SAMPLE_IP = 1U << 0,
122 PERF_SAMPLE_TID = 1U << 1,
123 PERF_SAMPLE_TIME = 1U << 2,
124 PERF_SAMPLE_ADDR = 1U << 3,
125 PERF_SAMPLE_READ = 1U << 4,
126 PERF_SAMPLE_CALLCHAIN = 1U << 5,
127 PERF_SAMPLE_ID = 1U << 6,
128 PERF_SAMPLE_CPU = 1U << 7,
129 PERF_SAMPLE_PERIOD = 1U << 8,
130 PERF_SAMPLE_STREAM_ID = 1U << 9,
131 PERF_SAMPLE_RAW = 1U << 10,
132 PERF_SAMPLE_BRANCH_STACK = 1U << 11,
133
134 PERF_SAMPLE_MAX = 1U << 12, /* non-ABI */
135 };
136
137 /*
138 * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
139 *
140 * If the user does not pass priv level information via branch_sample_type,
141 * the kernel uses the event's priv level. Branch and event priv levels do
142 * not have to match. Branch priv level is checked for permissions.
143 *
144 * The branch types can be combined, however BRANCH_ANY covers all types
145 * of branches and therefore it supersedes all the other types.
146 */
147 enum perf_branch_sample_type {
148 PERF_SAMPLE_BRANCH_USER = 1U << 0, /* user branches */
149 PERF_SAMPLE_BRANCH_KERNEL = 1U << 1, /* kernel branches */
150 PERF_SAMPLE_BRANCH_HV = 1U << 2, /* hypervisor branches */
151
152 PERF_SAMPLE_BRANCH_ANY = 1U << 3, /* any branch types */
153 PERF_SAMPLE_BRANCH_ANY_CALL = 1U << 4, /* any call branch */
154 PERF_SAMPLE_BRANCH_ANY_RETURN = 1U << 5, /* any return branch */
155 PERF_SAMPLE_BRANCH_IND_CALL = 1U << 6, /* indirect calls */
156
157 PERF_SAMPLE_BRANCH_MAX = 1U << 7, /* non-ABI */
158 };
159
160 #define PERF_SAMPLE_BRANCH_PLM_ALL \
161 (PERF_SAMPLE_BRANCH_USER|\
162 PERF_SAMPLE_BRANCH_KERNEL|\
163 PERF_SAMPLE_BRANCH_HV)
164
165 /*
166 * The format of the data returned by read() on a perf event fd,
167 * as specified by attr.read_format:
168 *
169 * struct read_format {
170 * { u64 value;
171 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
172 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
173 * { u64 id; } && PERF_FORMAT_ID
174 * } && !PERF_FORMAT_GROUP
175 *
176 * { u64 nr;
177 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
178 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
179 * { u64 value;
180 * { u64 id; } && PERF_FORMAT_ID
181 * } cntr[nr];
182 * } && PERF_FORMAT_GROUP
183 * };
184 */
185 enum perf_event_read_format {
186 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0,
187 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1,
188 PERF_FORMAT_ID = 1U << 2,
189 PERF_FORMAT_GROUP = 1U << 3,
190
191 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */
192 };
193
194 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
195 #define PERF_ATTR_SIZE_VER1 72 /* add: config2 */
196 #define PERF_ATTR_SIZE_VER2 80 /* add: branch_sample_type */
197
198 /*
199 * Hardware event_id to monitor via a performance monitoring event:
200 */
201 struct perf_event_attr {
202
203 /*
204 * Major type: hardware/software/tracepoint/etc.
205 */
206 __u32 type;
207
208 /*
209 * Size of the attr structure, for fwd/bwd compat.
210 */
211 __u32 size;
212
213 /*
214 * Type specific configuration information.
215 */
216 __u64 config;
217
218 union {
219 __u64 sample_period;
220 __u64 sample_freq;
221 };
222
223 __u64 sample_type;
224 __u64 read_format;
225
226 __u64 disabled : 1, /* off by default */
227 inherit : 1, /* children inherit it */
228 pinned : 1, /* must always be on PMU */
229 exclusive : 1, /* only group on PMU */
230 exclude_user : 1, /* don't count user */
231 exclude_kernel : 1, /* ditto kernel */
232 exclude_hv : 1, /* ditto hypervisor */
233 exclude_idle : 1, /* don't count when idle */
234 mmap : 1, /* include mmap data */
235 comm : 1, /* include comm data */
236 freq : 1, /* use freq, not period */
237 inherit_stat : 1, /* per task counts */
238 enable_on_exec : 1, /* next exec enables */
239 task : 1, /* trace fork/exit */
240 watermark : 1, /* wakeup_watermark */
241 /*
242 * precise_ip:
243 *
244 * 0 - SAMPLE_IP can have arbitrary skid
245 * 1 - SAMPLE_IP must have constant skid
246 * 2 - SAMPLE_IP requested to have 0 skid
247 * 3 - SAMPLE_IP must have 0 skid
248 *
249 * See also PERF_RECORD_MISC_EXACT_IP
250 */
251 precise_ip : 2, /* skid constraint */
252 mmap_data : 1, /* non-exec mmap data */
253 sample_id_all : 1, /* sample_type all events */
254
255 exclude_host : 1, /* don't count in host */
256 exclude_guest : 1, /* don't count in guest */
257
258 __reserved_1 : 43;
259
260 union {
261 __u32 wakeup_events; /* wakeup every n events */
262 __u32 wakeup_watermark; /* bytes before wakeup */
263 };
264
265 __u32 bp_type;
266 union {
267 __u64 bp_addr;
268 __u64 config1; /* extension of config */
269 };
270 union {
271 __u64 bp_len;
272 __u64 config2; /* extension of config1 */
273 };
274 __u64 branch_sample_type; /* enum branch_sample_type */
275 };
276
277 /*
278 * Ioctls that can be done on a perf event fd:
279 */
280 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0)
281 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1)
282 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2)
283 #define PERF_EVENT_IOC_RESET _IO ('$', 3)
284 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64)
285 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5)
286 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)
287
288 enum perf_event_ioc_flags {
289 PERF_IOC_FLAG_GROUP = 1U << 0,
290 };
291
292 /*
293 * Structure of the page that can be mapped via mmap
294 */
295 struct perf_event_mmap_page {
296 __u32 version; /* version number of this structure */
297 __u32 compat_version; /* lowest version this is compat with */
298
299 /*
300 * Bits needed to read the hw events in user-space.
301 *
302 * u32 seq, time_mult, time_shift, idx, width;
303 * u64 count, enabled, running;
304 * u64 cyc, time_offset;
305 * s64 pmc = 0;
306 *
307 * do {
308 * seq = pc->lock;
309 * barrier()
310 *
311 * enabled = pc->time_enabled;
312 * running = pc->time_running;
313 *
314 * if (pc->cap_usr_time && enabled != running) {
315 * cyc = rdtsc();
316 * time_offset = pc->time_offset;
317 * time_mult = pc->time_mult;
318 * time_shift = pc->time_shift;
319 * }
320 *
321 * idx = pc->index;
322 * count = pc->offset;
323 * if (pc->cap_usr_rdpmc && idx) {
324 * width = pc->pmc_width;
325 * pmc = rdpmc(idx - 1);
326 * }
327 *
328 * barrier();
329 * } while (pc->lock != seq);
330 *
331 * NOTE: for obvious reason this only works on self-monitoring
332 * processes.
333 */
334 __u32 lock; /* seqlock for synchronization */
335 __u32 index; /* hardware event identifier */
336 __s64 offset; /* add to hardware event value */
337 __u64 time_enabled; /* time event active */
338 __u64 time_running; /* time event on cpu */
339 union {
340 __u64 capabilities;
341 __u64 cap_usr_time : 1,
342 cap_usr_rdpmc : 1,
343 cap_____res : 62;
344 };
345
346 /*
347 * If cap_usr_rdpmc this field provides the bit-width of the value
348 * read using the rdpmc() or equivalent instruction. This can be used
349 * to sign extend the result like:
350 *
351 * pmc <<= 64 - width;
352 * pmc >>= 64 - width; // signed shift right
353 * count += pmc;
354 */
355 __u16 pmc_width;
356
357 /*
358 * If cap_usr_time the below fields can be used to compute the time
359 * delta since time_enabled (in ns) using rdtsc or similar.
360 *
361 * u64 quot, rem;
362 * u64 delta;
363 *
364 * quot = (cyc >> time_shift);
365 * rem = cyc & ((1 << time_shift) - 1);
366 * delta = time_offset + quot * time_mult +
367 * ((rem * time_mult) >> time_shift);
368 *
369 * Where time_offset,time_mult,time_shift and cyc are read in the
370 * seqcount loop described above. This delta can then be added to
371 * enabled and possible running (if idx), improving the scaling:
372 *
373 * enabled += delta;
374 * if (idx)
375 * running += delta;
376 *
377 * quot = count / running;
378 * rem = count % running;
379 * count = quot * enabled + (rem * enabled) / running;
380 */
381 __u16 time_shift;
382 __u32 time_mult;
383 __u64 time_offset;
384
385 /*
386 * Hole for extension of the self monitor capabilities
387 */
388
389 __u64 __reserved[120]; /* align to 1k */
390
391 /*
392 * Control data for the mmap() data buffer.
393 *
394 * User-space reading the @data_head value should issue an smp_rmb(),
395 * after reading this value.
396 *
397 * When the mapping is PROT_WRITE the @data_tail value should be
398 * written by userspace to reflect the last read data, after issueing
399 * an smp_mb() to separate the data read from the ->data_tail store.
400 * In this case the kernel will not over-write unread data.
401 *
402 * See perf_output_put_handle() for the data ordering.
403 */
404 __u64 data_head; /* head in the data section */
405 __u64 data_tail; /* user-space written tail */
406 };
407
408 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0)
409 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0)
410 #define PERF_RECORD_MISC_KERNEL (1 << 0)
411 #define PERF_RECORD_MISC_USER (2 << 0)
412 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0)
413 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0)
414 #define PERF_RECORD_MISC_GUEST_USER (5 << 0)
415
416 /*
417 * Indicates that the content of PERF_SAMPLE_IP points to
418 * the actual instruction that triggered the event. See also
419 * perf_event_attr::precise_ip.
420 */
421 #define PERF_RECORD_MISC_EXACT_IP (1 << 14)
422 /*
423 * Reserve the last bit to indicate some extended misc field
424 */
425 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15)
426
427 struct perf_event_header {
428 __u32 type;
429 __u16 misc;
430 __u16 size;
431 };
432
433 enum perf_event_type {
434
435 /*
436 * If perf_event_attr.sample_id_all is set then all event types will
437 * have the sample_type selected fields related to where/when
438 * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
439 * described in PERF_RECORD_SAMPLE below, it will be stashed just after
440 * the perf_event_header and the fields already present for the existing
441 * fields, i.e. at the end of the payload. That way a newer perf.data
442 * file will be supported by older perf tools, with these new optional
443 * fields being ignored.
444 *
445 * The MMAP events record the PROT_EXEC mappings so that we can
446 * correlate userspace IPs to code. They have the following structure:
447 *
448 * struct {
449 * struct perf_event_header header;
450 *
451 * u32 pid, tid;
452 * u64 addr;
453 * u64 len;
454 * u64 pgoff;
455 * char filename[];
456 * };
457 */
458 PERF_RECORD_MMAP = 1,
459
460 /*
461 * struct {
462 * struct perf_event_header header;
463 * u64 id;
464 * u64 lost;
465 * };
466 */
467 PERF_RECORD_LOST = 2,
468
469 /*
470 * struct {
471 * struct perf_event_header header;
472 *
473 * u32 pid, tid;
474 * char comm[];
475 * };
476 */
477 PERF_RECORD_COMM = 3,
478
479 /*
480 * struct {
481 * struct perf_event_header header;
482 * u32 pid, ppid;
483 * u32 tid, ptid;
484 * u64 time;
485 * };
486 */
487 PERF_RECORD_EXIT = 4,
488
489 /*
490 * struct {
491 * struct perf_event_header header;
492 * u64 time;
493 * u64 id;
494 * u64 stream_id;
495 * };
496 */
497 PERF_RECORD_THROTTLE = 5,
498 PERF_RECORD_UNTHROTTLE = 6,
499
500 /*
501 * struct {
502 * struct perf_event_header header;
503 * u32 pid, ppid;
504 * u32 tid, ptid;
505 * u64 time;
506 * };
507 */
508 PERF_RECORD_FORK = 7,
509
510 /*
511 * struct {
512 * struct perf_event_header header;
513 * u32 pid, tid;
514 *
515 * struct read_format values;
516 * };
517 */
518 PERF_RECORD_READ = 8,
519
520 /*
521 * struct {
522 * struct perf_event_header header;
523 *
524 * { u64 ip; } && PERF_SAMPLE_IP
525 * { u32 pid, tid; } && PERF_SAMPLE_TID
526 * { u64 time; } && PERF_SAMPLE_TIME
527 * { u64 addr; } && PERF_SAMPLE_ADDR
528 * { u64 id; } && PERF_SAMPLE_ID
529 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID
530 * { u32 cpu, res; } && PERF_SAMPLE_CPU
531 * { u64 period; } && PERF_SAMPLE_PERIOD
532 *
533 * { struct read_format values; } && PERF_SAMPLE_READ
534 *
535 * { u64 nr,
536 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN
537 *
538 * #
539 * # The RAW record below is opaque data wrt the ABI
540 * #
541 * # That is, the ABI doesn't make any promises wrt to
542 * # the stability of its content, it may vary depending
543 * # on event, hardware, kernel version and phase of
544 * # the moon.
545 * #
546 * # In other words, PERF_SAMPLE_RAW contents are not an ABI.
547 * #
548 *
549 * { u32 size;
550 * char data[size];}&& PERF_SAMPLE_RAW
551 *
552 * { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK
553 * };
554 */
555 PERF_RECORD_SAMPLE = 9,
556
557 PERF_RECORD_MAX, /* non-ABI */
558 };
559
560 enum perf_callchain_context {
561 PERF_CONTEXT_HV = (__u64)-32,
562 PERF_CONTEXT_KERNEL = (__u64)-128,
563 PERF_CONTEXT_USER = (__u64)-512,
564
565 PERF_CONTEXT_GUEST = (__u64)-2048,
566 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176,
567 PERF_CONTEXT_GUEST_USER = (__u64)-2560,
568
569 PERF_CONTEXT_MAX = (__u64)-4095,
570 };
571
572 #define PERF_FLAG_FD_NO_GROUP (1U << 0)
573 #define PERF_FLAG_FD_OUTPUT (1U << 1)
574 #define PERF_FLAG_PID_CGROUP (1U << 2) /* pid=cgroup id, per-cpu mode only */
575
576 #ifdef __KERNEL__
577 /*
578 * Kernel-internal data types and definitions:
579 */
580
581 #ifdef CONFIG_PERF_EVENTS
582 # include <linux/cgroup.h>
583 # include <asm/perf_event.h>
584 # include <asm/local64.h>
585 #endif
586
587 struct perf_guest_info_callbacks {
588 int (*is_in_guest)(void);
589 int (*is_user_mode)(void);
590 unsigned long (*get_guest_ip)(void);
591 };
592
593 #ifdef CONFIG_HAVE_HW_BREAKPOINT
594 #include <asm/hw_breakpoint.h>
595 #endif
596
597 #include <linux/list.h>
598 #include <linux/mutex.h>
599 #include <linux/rculist.h>
600 #include <linux/rcupdate.h>
601 #include <linux/spinlock.h>
602 #include <linux/hrtimer.h>
603 #include <linux/fs.h>
604 #include <linux/pid_namespace.h>
605 #include <linux/workqueue.h>
606 #include <linux/ftrace.h>
607 #include <linux/cpu.h>
608 #include <linux/irq_work.h>
609 #include <linux/static_key.h>
610 #include <linux/atomic.h>
611 #include <linux/sysfs.h>
612 #include <asm/local.h>
613
614 #define PERF_MAX_STACK_DEPTH 255
615
616 struct perf_callchain_entry {
617 __u64 nr;
618 __u64 ip[PERF_MAX_STACK_DEPTH];
619 };
620
621 struct perf_raw_record {
622 u32 size;
623 void *data;
624 };
625
626 /*
627 * single taken branch record layout:
628 *
629 * from: source instruction (may not always be a branch insn)
630 * to: branch target
631 * mispred: branch target was mispredicted
632 * predicted: branch target was predicted
633 *
634 * support for mispred, predicted is optional. In case it
635 * is not supported mispred = predicted = 0.
636 */
637 struct perf_branch_entry {
638 __u64 from;
639 __u64 to;
640 __u64 mispred:1, /* target mispredicted */
641 predicted:1,/* target predicted */
642 reserved:62;
643 };
644
645 /*
646 * branch stack layout:
647 * nr: number of taken branches stored in entries[]
648 *
649 * Note that nr can vary from sample to sample
650 * branches (to, from) are stored from most recent
651 * to least recent, i.e., entries[0] contains the most
652 * recent branch.
653 */
654 struct perf_branch_stack {
655 __u64 nr;
656 struct perf_branch_entry entries[0];
657 };
658
659 struct task_struct;
660
661 /*
662 * extra PMU register associated with an event
663 */
664 struct hw_perf_event_extra {
665 u64 config; /* register value */
666 unsigned int reg; /* register address or index */
667 int alloc; /* extra register already allocated */
668 int idx; /* index in shared_regs->regs[] */
669 };
670
671 /**
672 * struct hw_perf_event - performance event hardware details:
673 */
674 struct hw_perf_event {
675 #ifdef CONFIG_PERF_EVENTS
676 union {
677 struct { /* hardware */
678 u64 config;
679 u64 last_tag;
680 unsigned long config_base;
681 unsigned long event_base;
682 int idx;
683 int last_cpu;
684
685 struct hw_perf_event_extra extra_reg;
686 struct hw_perf_event_extra branch_reg;
687 };
688 struct { /* software */
689 struct hrtimer hrtimer;
690 };
691 #ifdef CONFIG_HAVE_HW_BREAKPOINT
692 struct { /* breakpoint */
693 struct arch_hw_breakpoint info;
694 struct list_head bp_list;
695 /*
696 * Crufty hack to avoid the chicken and egg
697 * problem hw_breakpoint has with context
698 * creation and event initalization.
699 */
700 struct task_struct *bp_target;
701 };
702 #endif
703 };
704 int state;
705 local64_t prev_count;
706 u64 sample_period;
707 u64 last_period;
708 local64_t period_left;
709 u64 interrupts_seq;
710 u64 interrupts;
711
712 u64 freq_time_stamp;
713 u64 freq_count_stamp;
714 #endif
715 };
716
717 /*
718 * hw_perf_event::state flags
719 */
720 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
721 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
722 #define PERF_HES_ARCH 0x04
723
724 struct perf_event;
725
726 /*
727 * Common implementation detail of pmu::{start,commit,cancel}_txn
728 */
729 #define PERF_EVENT_TXN 0x1
730
731 /**
732 * struct pmu - generic performance monitoring unit
733 */
734 struct pmu {
735 struct list_head entry;
736
737 struct device *dev;
738 const struct attribute_group **attr_groups;
739 char *name;
740 int type;
741
742 int * __percpu pmu_disable_count;
743 struct perf_cpu_context * __percpu pmu_cpu_context;
744 int task_ctx_nr;
745
746 /*
747 * Fully disable/enable this PMU, can be used to protect from the PMI
748 * as well as for lazy/batch writing of the MSRs.
749 */
750 void (*pmu_enable) (struct pmu *pmu); /* optional */
751 void (*pmu_disable) (struct pmu *pmu); /* optional */
752
753 /*
754 * Try and initialize the event for this PMU.
755 * Should return -ENOENT when the @event doesn't match this PMU.
756 */
757 int (*event_init) (struct perf_event *event);
758
759 #define PERF_EF_START 0x01 /* start the counter when adding */
760 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
761 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
762
763 /*
764 * Adds/Removes a counter to/from the PMU, can be done inside
765 * a transaction, see the ->*_txn() methods.
766 */
767 int (*add) (struct perf_event *event, int flags);
768 void (*del) (struct perf_event *event, int flags);
769
770 /*
771 * Starts/Stops a counter present on the PMU. The PMI handler
772 * should stop the counter when perf_event_overflow() returns
773 * !0. ->start() will be used to continue.
774 */
775 void (*start) (struct perf_event *event, int flags);
776 void (*stop) (struct perf_event *event, int flags);
777
778 /*
779 * Updates the counter value of the event.
780 */
781 void (*read) (struct perf_event *event);
782
783 /*
784 * Group events scheduling is treated as a transaction, add
785 * group events as a whole and perform one schedulability test.
786 * If the test fails, roll back the whole group
787 *
788 * Start the transaction, after this ->add() doesn't need to
789 * do schedulability tests.
790 */
791 void (*start_txn) (struct pmu *pmu); /* optional */
792 /*
793 * If ->start_txn() disabled the ->add() schedulability test
794 * then ->commit_txn() is required to perform one. On success
795 * the transaction is closed. On error the transaction is kept
796 * open until ->cancel_txn() is called.
797 */
798 int (*commit_txn) (struct pmu *pmu); /* optional */
799 /*
800 * Will cancel the transaction, assumes ->del() is called
801 * for each successful ->add() during the transaction.
802 */
803 void (*cancel_txn) (struct pmu *pmu); /* optional */
804
805 /*
806 * Will return the value for perf_event_mmap_page::index for this event,
807 * if no implementation is provided it will default to: event->hw.idx + 1.
808 */
809 int (*event_idx) (struct perf_event *event); /*optional */
810
811 /*
812 * flush branch stack on context-switches (needed in cpu-wide mode)
813 */
814 void (*flush_branch_stack) (void);
815 };
816
817 /**
818 * enum perf_event_active_state - the states of a event
819 */
820 enum perf_event_active_state {
821 PERF_EVENT_STATE_ERROR = -2,
822 PERF_EVENT_STATE_OFF = -1,
823 PERF_EVENT_STATE_INACTIVE = 0,
824 PERF_EVENT_STATE_ACTIVE = 1,
825 };
826
827 struct file;
828 struct perf_sample_data;
829
830 typedef void (*perf_overflow_handler_t)(struct perf_event *,
831 struct perf_sample_data *,
832 struct pt_regs *regs);
833
834 enum perf_group_flag {
835 PERF_GROUP_SOFTWARE = 0x1,
836 };
837
838 #define SWEVENT_HLIST_BITS 8
839 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
840
841 struct swevent_hlist {
842 struct hlist_head heads[SWEVENT_HLIST_SIZE];
843 struct rcu_head rcu_head;
844 };
845
846 #define PERF_ATTACH_CONTEXT 0x01
847 #define PERF_ATTACH_GROUP 0x02
848 #define PERF_ATTACH_TASK 0x04
849
850 #ifdef CONFIG_CGROUP_PERF
851 /*
852 * perf_cgroup_info keeps track of time_enabled for a cgroup.
853 * This is a per-cpu dynamically allocated data structure.
854 */
855 struct perf_cgroup_info {
856 u64 time;
857 u64 timestamp;
858 };
859
860 struct perf_cgroup {
861 struct cgroup_subsys_state css;
862 struct perf_cgroup_info *info; /* timing info, one per cpu */
863 };
864 #endif
865
866 struct ring_buffer;
867
868 /**
869 * struct perf_event - performance event kernel representation:
870 */
871 struct perf_event {
872 #ifdef CONFIG_PERF_EVENTS
873 struct list_head group_entry;
874 struct list_head event_entry;
875 struct list_head sibling_list;
876 struct hlist_node hlist_entry;
877 int nr_siblings;
878 int group_flags;
879 struct perf_event *group_leader;
880 struct pmu *pmu;
881
882 enum perf_event_active_state state;
883 unsigned int attach_state;
884 local64_t count;
885 atomic64_t child_count;
886
887 /*
888 * These are the total time in nanoseconds that the event
889 * has been enabled (i.e. eligible to run, and the task has
890 * been scheduled in, if this is a per-task event)
891 * and running (scheduled onto the CPU), respectively.
892 *
893 * They are computed from tstamp_enabled, tstamp_running and
894 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
895 */
896 u64 total_time_enabled;
897 u64 total_time_running;
898
899 /*
900 * These are timestamps used for computing total_time_enabled
901 * and total_time_running when the event is in INACTIVE or
902 * ACTIVE state, measured in nanoseconds from an arbitrary point
903 * in time.
904 * tstamp_enabled: the notional time when the event was enabled
905 * tstamp_running: the notional time when the event was scheduled on
906 * tstamp_stopped: in INACTIVE state, the notional time when the
907 * event was scheduled off.
908 */
909 u64 tstamp_enabled;
910 u64 tstamp_running;
911 u64 tstamp_stopped;
912
913 /*
914 * timestamp shadows the actual context timing but it can
915 * be safely used in NMI interrupt context. It reflects the
916 * context time as it was when the event was last scheduled in.
917 *
918 * ctx_time already accounts for ctx->timestamp. Therefore to
919 * compute ctx_time for a sample, simply add perf_clock().
920 */
921 u64 shadow_ctx_time;
922
923 struct perf_event_attr attr;
924 u16 header_size;
925 u16 id_header_size;
926 u16 read_size;
927 struct hw_perf_event hw;
928
929 struct perf_event_context *ctx;
930 atomic_long_t refcount;
931
932 /*
933 * These accumulate total time (in nanoseconds) that children
934 * events have been enabled and running, respectively.
935 */
936 atomic64_t child_total_time_enabled;
937 atomic64_t child_total_time_running;
938
939 /*
940 * Protect attach/detach and child_list:
941 */
942 struct mutex child_mutex;
943 struct list_head child_list;
944 struct perf_event *parent;
945
946 int oncpu;
947 int cpu;
948
949 struct list_head owner_entry;
950 struct task_struct *owner;
951
952 /* mmap bits */
953 struct mutex mmap_mutex;
954 atomic_t mmap_count;
955
956 struct ring_buffer *rb;
957 struct list_head rb_entry;
958
959 /* poll related */
960 wait_queue_head_t waitq;
961 struct fasync_struct *fasync;
962
963 /* delayed work for NMIs and such */
964 int pending_wakeup;
965 int pending_kill;
966 int pending_disable;
967 struct irq_work pending;
968
969 atomic_t event_limit;
970
971 void (*destroy)(struct perf_event *);
972 struct rcu_head rcu_head;
973
974 struct pid_namespace *ns;
975 u64 id;
976
977 perf_overflow_handler_t overflow_handler;
978 void *overflow_handler_context;
979
980 #ifdef CONFIG_EVENT_TRACING
981 struct ftrace_event_call *tp_event;
982 struct event_filter *filter;
983 #ifdef CONFIG_FUNCTION_TRACER
984 struct ftrace_ops ftrace_ops;
985 #endif
986 #endif
987
988 #ifdef CONFIG_CGROUP_PERF
989 struct perf_cgroup *cgrp; /* cgroup event is attach to */
990 int cgrp_defer_enabled;
991 #endif
992
993 #endif /* CONFIG_PERF_EVENTS */
994 };
995
996 enum perf_event_context_type {
997 task_context,
998 cpu_context,
999 };
1000
1001 /**
1002 * struct perf_event_context - event context structure
1003 *
1004 * Used as a container for task events and CPU events as well:
1005 */
1006 struct perf_event_context {
1007 struct pmu *pmu;
1008 enum perf_event_context_type type;
1009 /*
1010 * Protect the states of the events in the list,
1011 * nr_active, and the list:
1012 */
1013 raw_spinlock_t lock;
1014 /*
1015 * Protect the list of events. Locking either mutex or lock
1016 * is sufficient to ensure the list doesn't change; to change
1017 * the list you need to lock both the mutex and the spinlock.
1018 */
1019 struct mutex mutex;
1020
1021 struct list_head pinned_groups;
1022 struct list_head flexible_groups;
1023 struct list_head event_list;
1024 int nr_events;
1025 int nr_active;
1026 int is_active;
1027 int nr_stat;
1028 int nr_freq;
1029 int rotate_disable;
1030 atomic_t refcount;
1031 struct task_struct *task;
1032
1033 /*
1034 * Context clock, runs when context enabled.
1035 */
1036 u64 time;
1037 u64 timestamp;
1038
1039 /*
1040 * These fields let us detect when two contexts have both
1041 * been cloned (inherited) from a common ancestor.
1042 */
1043 struct perf_event_context *parent_ctx;
1044 u64 parent_gen;
1045 u64 generation;
1046 int pin_count;
1047 int nr_cgroups; /* cgroup evts */
1048 int nr_branch_stack; /* branch_stack evt */
1049 struct rcu_head rcu_head;
1050 };
1051
1052 /*
1053 * Number of contexts where an event can trigger:
1054 * task, softirq, hardirq, nmi.
1055 */
1056 #define PERF_NR_CONTEXTS 4
1057
1058 /**
1059 * struct perf_event_cpu_context - per cpu event context structure
1060 */
1061 struct perf_cpu_context {
1062 struct perf_event_context ctx;
1063 struct perf_event_context *task_ctx;
1064 int active_oncpu;
1065 int exclusive;
1066 struct list_head rotation_list;
1067 int jiffies_interval;
1068 struct pmu *unique_pmu;
1069 struct perf_cgroup *cgrp;
1070 };
1071
1072 struct perf_output_handle {
1073 struct perf_event *event;
1074 struct ring_buffer *rb;
1075 unsigned long wakeup;
1076 unsigned long size;
1077 void *addr;
1078 int page;
1079 };
1080
1081 #ifdef CONFIG_PERF_EVENTS
1082
1083 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
1084 extern void perf_pmu_unregister(struct pmu *pmu);
1085
1086 extern int perf_num_counters(void);
1087 extern const char *perf_pmu_name(void);
1088 extern void __perf_event_task_sched_in(struct task_struct *prev,
1089 struct task_struct *task);
1090 extern void __perf_event_task_sched_out(struct task_struct *prev,
1091 struct task_struct *next);
1092 extern int perf_event_init_task(struct task_struct *child);
1093 extern void perf_event_exit_task(struct task_struct *child);
1094 extern void perf_event_free_task(struct task_struct *task);
1095 extern void perf_event_delayed_put(struct task_struct *task);
1096 extern void perf_event_print_debug(void);
1097 extern void perf_pmu_disable(struct pmu *pmu);
1098 extern void perf_pmu_enable(struct pmu *pmu);
1099 extern int perf_event_task_disable(void);
1100 extern int perf_event_task_enable(void);
1101 extern int perf_event_refresh(struct perf_event *event, int refresh);
1102 extern void perf_event_update_userpage(struct perf_event *event);
1103 extern int perf_event_release_kernel(struct perf_event *event);
1104 extern struct perf_event *
1105 perf_event_create_kernel_counter(struct perf_event_attr *attr,
1106 int cpu,
1107 struct task_struct *task,
1108 perf_overflow_handler_t callback,
1109 void *context);
1110 extern u64 perf_event_read_value(struct perf_event *event,
1111 u64 *enabled, u64 *running);
1112
1113
1114 struct perf_sample_data {
1115 u64 type;
1116
1117 u64 ip;
1118 struct {
1119 u32 pid;
1120 u32 tid;
1121 } tid_entry;
1122 u64 time;
1123 u64 addr;
1124 u64 id;
1125 u64 stream_id;
1126 struct {
1127 u32 cpu;
1128 u32 reserved;
1129 } cpu_entry;
1130 u64 period;
1131 struct perf_callchain_entry *callchain;
1132 struct perf_raw_record *raw;
1133 struct perf_branch_stack *br_stack;
1134 };
1135
perf_sample_data_init(struct perf_sample_data * data,u64 addr)1136 static inline void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
1137 {
1138 data->addr = addr;
1139 data->raw = NULL;
1140 data->br_stack = NULL;
1141 }
1142
1143 extern void perf_output_sample(struct perf_output_handle *handle,
1144 struct perf_event_header *header,
1145 struct perf_sample_data *data,
1146 struct perf_event *event);
1147 extern void perf_prepare_sample(struct perf_event_header *header,
1148 struct perf_sample_data *data,
1149 struct perf_event *event,
1150 struct pt_regs *regs);
1151
1152 extern int perf_event_overflow(struct perf_event *event,
1153 struct perf_sample_data *data,
1154 struct pt_regs *regs);
1155
is_sampling_event(struct perf_event * event)1156 static inline bool is_sampling_event(struct perf_event *event)
1157 {
1158 return event->attr.sample_period != 0;
1159 }
1160
1161 /*
1162 * Return 1 for a software event, 0 for a hardware event
1163 */
is_software_event(struct perf_event * event)1164 static inline int is_software_event(struct perf_event *event)
1165 {
1166 return event->pmu->task_ctx_nr == perf_sw_context;
1167 }
1168
1169 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1170
1171 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1172
1173 #ifndef perf_arch_fetch_caller_regs
perf_arch_fetch_caller_regs(struct pt_regs * regs,unsigned long ip)1174 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1175 #endif
1176
1177 /*
1178 * Take a snapshot of the regs. Skip ip and frame pointer to
1179 * the nth caller. We only need a few of the regs:
1180 * - ip for PERF_SAMPLE_IP
1181 * - cs for user_mode() tests
1182 * - bp for callchains
1183 * - eflags, for future purposes, just in case
1184 */
perf_fetch_caller_regs(struct pt_regs * regs)1185 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1186 {
1187 memset(regs, 0, sizeof(*regs));
1188
1189 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1190 }
1191
1192 static __always_inline void
perf_sw_event(u32 event_id,u64 nr,struct pt_regs * regs,u64 addr)1193 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1194 {
1195 struct pt_regs hot_regs;
1196
1197 if (static_key_false(&perf_swevent_enabled[event_id])) {
1198 if (!regs) {
1199 perf_fetch_caller_regs(&hot_regs);
1200 regs = &hot_regs;
1201 }
1202 __perf_sw_event(event_id, nr, regs, addr);
1203 }
1204 }
1205
1206 extern struct static_key_deferred perf_sched_events;
1207
perf_event_task_sched_in(struct task_struct * prev,struct task_struct * task)1208 static inline void perf_event_task_sched_in(struct task_struct *prev,
1209 struct task_struct *task)
1210 {
1211 if (static_key_false(&perf_sched_events.key))
1212 __perf_event_task_sched_in(prev, task);
1213 }
1214
perf_event_task_sched_out(struct task_struct * prev,struct task_struct * next)1215 static inline void perf_event_task_sched_out(struct task_struct *prev,
1216 struct task_struct *next)
1217 {
1218 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
1219
1220 if (static_key_false(&perf_sched_events.key))
1221 __perf_event_task_sched_out(prev, next);
1222 }
1223
1224 extern void perf_event_mmap(struct vm_area_struct *vma);
1225 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1226 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1227 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1228
1229 extern void perf_event_comm(struct task_struct *tsk);
1230 extern void perf_event_fork(struct task_struct *tsk);
1231
1232 /* Callchains */
1233 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1234
1235 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
1236 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
1237
perf_callchain_store(struct perf_callchain_entry * entry,u64 ip)1238 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1239 {
1240 if (entry->nr < PERF_MAX_STACK_DEPTH)
1241 entry->ip[entry->nr++] = ip;
1242 }
1243
1244 extern int sysctl_perf_event_paranoid;
1245 extern int sysctl_perf_event_mlock;
1246 extern int sysctl_perf_event_sample_rate;
1247
1248 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1249 void __user *buffer, size_t *lenp,
1250 loff_t *ppos);
1251
perf_paranoid_tracepoint_raw(void)1252 static inline bool perf_paranoid_tracepoint_raw(void)
1253 {
1254 return sysctl_perf_event_paranoid > -1;
1255 }
1256
perf_paranoid_cpu(void)1257 static inline bool perf_paranoid_cpu(void)
1258 {
1259 return sysctl_perf_event_paranoid > 0;
1260 }
1261
perf_paranoid_kernel(void)1262 static inline bool perf_paranoid_kernel(void)
1263 {
1264 return sysctl_perf_event_paranoid > 1;
1265 }
1266
1267 extern void perf_event_init(void);
1268 extern void perf_tp_event(u64 addr, u64 count, void *record,
1269 int entry_size, struct pt_regs *regs,
1270 struct hlist_head *head, int rctx);
1271 extern void perf_bp_event(struct perf_event *event, void *data);
1272
1273 #ifndef perf_misc_flags
1274 # define perf_misc_flags(regs) \
1275 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1276 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1277 #endif
1278
has_branch_stack(struct perf_event * event)1279 static inline bool has_branch_stack(struct perf_event *event)
1280 {
1281 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1282 }
1283
1284 extern int perf_output_begin(struct perf_output_handle *handle,
1285 struct perf_event *event, unsigned int size);
1286 extern void perf_output_end(struct perf_output_handle *handle);
1287 extern void perf_output_copy(struct perf_output_handle *handle,
1288 const void *buf, unsigned int len);
1289 extern int perf_swevent_get_recursion_context(void);
1290 extern void perf_swevent_put_recursion_context(int rctx);
1291 extern void perf_event_enable(struct perf_event *event);
1292 extern void perf_event_disable(struct perf_event *event);
1293 extern void perf_event_task_tick(void);
1294 #else
1295 static inline void
perf_event_task_sched_in(struct task_struct * prev,struct task_struct * task)1296 perf_event_task_sched_in(struct task_struct *prev,
1297 struct task_struct *task) { }
1298 static inline void
perf_event_task_sched_out(struct task_struct * prev,struct task_struct * next)1299 perf_event_task_sched_out(struct task_struct *prev,
1300 struct task_struct *next) { }
perf_event_init_task(struct task_struct * child)1301 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
perf_event_exit_task(struct task_struct * child)1302 static inline void perf_event_exit_task(struct task_struct *child) { }
perf_event_free_task(struct task_struct * task)1303 static inline void perf_event_free_task(struct task_struct *task) { }
perf_event_delayed_put(struct task_struct * task)1304 static inline void perf_event_delayed_put(struct task_struct *task) { }
perf_event_print_debug(void)1305 static inline void perf_event_print_debug(void) { }
perf_event_task_disable(void)1306 static inline int perf_event_task_disable(void) { return -EINVAL; }
perf_event_task_enable(void)1307 static inline int perf_event_task_enable(void) { return -EINVAL; }
perf_event_refresh(struct perf_event * event,int refresh)1308 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1309 {
1310 return -EINVAL;
1311 }
1312
1313 static inline void
perf_sw_event(u32 event_id,u64 nr,struct pt_regs * regs,u64 addr)1314 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1315 static inline void
perf_bp_event(struct perf_event * event,void * data)1316 perf_bp_event(struct perf_event *event, void *data) { }
1317
perf_register_guest_info_callbacks(struct perf_guest_info_callbacks * callbacks)1318 static inline int perf_register_guest_info_callbacks
1319 (struct perf_guest_info_callbacks *callbacks) { return 0; }
perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks * callbacks)1320 static inline int perf_unregister_guest_info_callbacks
1321 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1322
perf_event_mmap(struct vm_area_struct * vma)1323 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
perf_event_comm(struct task_struct * tsk)1324 static inline void perf_event_comm(struct task_struct *tsk) { }
perf_event_fork(struct task_struct * tsk)1325 static inline void perf_event_fork(struct task_struct *tsk) { }
perf_event_init(void)1326 static inline void perf_event_init(void) { }
perf_swevent_get_recursion_context(void)1327 static inline int perf_swevent_get_recursion_context(void) { return -1; }
perf_swevent_put_recursion_context(int rctx)1328 static inline void perf_swevent_put_recursion_context(int rctx) { }
perf_event_enable(struct perf_event * event)1329 static inline void perf_event_enable(struct perf_event *event) { }
perf_event_disable(struct perf_event * event)1330 static inline void perf_event_disable(struct perf_event *event) { }
perf_event_task_tick(void)1331 static inline void perf_event_task_tick(void) { }
1332 #endif
1333
1334 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1335 extern void perf_restore_debug_store(void);
1336 #else
perf_restore_debug_store(void)1337 static inline void perf_restore_debug_store(void) { }
1338 #endif
1339
1340 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1341
1342 /*
1343 * This has to have a higher priority than migration_notifier in sched.c.
1344 */
1345 #define perf_cpu_notifier(fn) \
1346 do { \
1347 static struct notifier_block fn##_nb __cpuinitdata = \
1348 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1349 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
1350 (void *)(unsigned long)smp_processor_id()); \
1351 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
1352 (void *)(unsigned long)smp_processor_id()); \
1353 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
1354 (void *)(unsigned long)smp_processor_id()); \
1355 register_cpu_notifier(&fn##_nb); \
1356 } while (0)
1357
1358
1359 #define PMU_FORMAT_ATTR(_name, _format) \
1360 static ssize_t \
1361 _name##_show(struct device *dev, \
1362 struct device_attribute *attr, \
1363 char *page) \
1364 { \
1365 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1366 return sprintf(page, _format "\n"); \
1367 } \
1368 \
1369 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1370
1371 #endif /* __KERNEL__ */
1372 #endif /* _LINUX_PERF_EVENT_H */
1373