Root/
Source at commit b05a5adf03613de371c77c3235f7d970d7cd0c71 created 13 years 1 month ago. By Lars-Peter Clausen, NAND: Optimize reading the eec data for the JZ4740 (evil hack) | |
---|---|
1 | /* |
2 | * Performance events core code: |
3 | * |
4 | * Copyright (C) 2008 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 <pzijlstr@redhat.com> |
7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
8 | * |
9 | * For licensing details see kernel-base/COPYING |
10 | */ |
11 | |
12 | #include <linux/fs.h> |
13 | #include <linux/mm.h> |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> |
16 | #include <linux/idr.h> |
17 | #include <linux/file.h> |
18 | #include <linux/poll.h> |
19 | #include <linux/slab.h> |
20 | #include <linux/hash.h> |
21 | #include <linux/sysfs.h> |
22 | #include <linux/dcache.h> |
23 | #include <linux/percpu.h> |
24 | #include <linux/ptrace.h> |
25 | #include <linux/reboot.h> |
26 | #include <linux/vmstat.h> |
27 | #include <linux/device.h> |
28 | #include <linux/vmalloc.h> |
29 | #include <linux/hardirq.h> |
30 | #include <linux/rculist.h> |
31 | #include <linux/uaccess.h> |
32 | #include <linux/syscalls.h> |
33 | #include <linux/anon_inodes.h> |
34 | #include <linux/kernel_stat.h> |
35 | #include <linux/perf_event.h> |
36 | #include <linux/ftrace_event.h> |
37 | #include <linux/hw_breakpoint.h> |
38 | |
39 | #include <asm/irq_regs.h> |
40 | |
41 | struct remote_function_call { |
42 | struct task_struct *p; |
43 | int (*func)(void *info); |
44 | void *info; |
45 | int ret; |
46 | }; |
47 | |
48 | static void remote_function(void *data) |
49 | { |
50 | struct remote_function_call *tfc = data; |
51 | struct task_struct *p = tfc->p; |
52 | |
53 | if (p) { |
54 | tfc->ret = -EAGAIN; |
55 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) |
56 | return; |
57 | } |
58 | |
59 | tfc->ret = tfc->func(tfc->info); |
60 | } |
61 | |
62 | /** |
63 | * task_function_call - call a function on the cpu on which a task runs |
64 | * @p: the task to evaluate |
65 | * @func: the function to be called |
66 | * @info: the function call argument |
67 | * |
68 | * Calls the function @func when the task is currently running. This might |
69 | * be on the current CPU, which just calls the function directly |
70 | * |
71 | * returns: @func return value, or |
72 | * -ESRCH - when the process isn't running |
73 | * -EAGAIN - when the process moved away |
74 | */ |
75 | static int |
76 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) |
77 | { |
78 | struct remote_function_call data = { |
79 | .p = p, |
80 | .func = func, |
81 | .info = info, |
82 | .ret = -ESRCH, /* No such (running) process */ |
83 | }; |
84 | |
85 | if (task_curr(p)) |
86 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); |
87 | |
88 | return data.ret; |
89 | } |
90 | |
91 | /** |
92 | * cpu_function_call - call a function on the cpu |
93 | * @func: the function to be called |
94 | * @info: the function call argument |
95 | * |
96 | * Calls the function @func on the remote cpu. |
97 | * |
98 | * returns: @func return value or -ENXIO when the cpu is offline |
99 | */ |
100 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) |
101 | { |
102 | struct remote_function_call data = { |
103 | .p = NULL, |
104 | .func = func, |
105 | .info = info, |
106 | .ret = -ENXIO, /* No such CPU */ |
107 | }; |
108 | |
109 | smp_call_function_single(cpu, remote_function, &data, 1); |
110 | |
111 | return data.ret; |
112 | } |
113 | |
114 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
115 | PERF_FLAG_FD_OUTPUT |\ |
116 | PERF_FLAG_PID_CGROUP) |
117 | |
118 | enum event_type_t { |
119 | EVENT_FLEXIBLE = 0x1, |
120 | EVENT_PINNED = 0x2, |
121 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, |
122 | }; |
123 | |
124 | /* |
125 | * perf_sched_events : >0 events exist |
126 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu |
127 | */ |
128 | atomic_t perf_sched_events __read_mostly; |
129 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
130 | |
131 | static atomic_t nr_mmap_events __read_mostly; |
132 | static atomic_t nr_comm_events __read_mostly; |
133 | static atomic_t nr_task_events __read_mostly; |
134 | |
135 | static LIST_HEAD(pmus); |
136 | static DEFINE_MUTEX(pmus_lock); |
137 | static struct srcu_struct pmus_srcu; |
138 | |
139 | /* |
140 | * perf event paranoia level: |
141 | * -1 - not paranoid at all |
142 | * 0 - disallow raw tracepoint access for unpriv |
143 | * 1 - disallow cpu events for unpriv |
144 | * 2 - disallow kernel profiling for unpriv |
145 | */ |
146 | int sysctl_perf_event_paranoid __read_mostly = 1; |
147 | |
148 | /* Minimum for 512 kiB + 1 user control page */ |
149 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ |
150 | |
151 | /* |
152 | * max perf event sample rate |
153 | */ |
154 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
155 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; |
156 | static int max_samples_per_tick __read_mostly = |
157 | DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); |
158 | |
159 | int perf_proc_update_handler(struct ctl_table *table, int write, |
160 | void __user *buffer, size_t *lenp, |
161 | loff_t *ppos) |
162 | { |
163 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); |
164 | |
165 | if (ret || !write) |
166 | return ret; |
167 | |
168 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); |
169 | |
170 | return 0; |
171 | } |
172 | |
173 | static atomic64_t perf_event_id; |
174 | |
175 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
176 | enum event_type_t event_type); |
177 | |
178 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
179 | enum event_type_t event_type, |
180 | struct task_struct *task); |
181 | |
182 | static void update_context_time(struct perf_event_context *ctx); |
183 | static u64 perf_event_time(struct perf_event *event); |
184 | |
185 | void __weak perf_event_print_debug(void) { } |
186 | |
187 | extern __weak const char *perf_pmu_name(void) |
188 | { |
189 | return "pmu"; |
190 | } |
191 | |
192 | static inline u64 perf_clock(void) |
193 | { |
194 | return local_clock(); |
195 | } |
196 | |
197 | static inline struct perf_cpu_context * |
198 | __get_cpu_context(struct perf_event_context *ctx) |
199 | { |
200 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); |
201 | } |
202 | |
203 | #ifdef CONFIG_CGROUP_PERF |
204 | |
205 | /* |
206 | * Must ensure cgroup is pinned (css_get) before calling |
207 | * this function. In other words, we cannot call this function |
208 | * if there is no cgroup event for the current CPU context. |
209 | */ |
210 | static inline struct perf_cgroup * |
211 | perf_cgroup_from_task(struct task_struct *task) |
212 | { |
213 | return container_of(task_subsys_state(task, perf_subsys_id), |
214 | struct perf_cgroup, css); |
215 | } |
216 | |
217 | static inline bool |
218 | perf_cgroup_match(struct perf_event *event) |
219 | { |
220 | struct perf_event_context *ctx = event->ctx; |
221 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
222 | |
223 | return !event->cgrp || event->cgrp == cpuctx->cgrp; |
224 | } |
225 | |
226 | static inline void perf_get_cgroup(struct perf_event *event) |
227 | { |
228 | css_get(&event->cgrp->css); |
229 | } |
230 | |
231 | static inline void perf_put_cgroup(struct perf_event *event) |
232 | { |
233 | css_put(&event->cgrp->css); |
234 | } |
235 | |
236 | static inline void perf_detach_cgroup(struct perf_event *event) |
237 | { |
238 | perf_put_cgroup(event); |
239 | event->cgrp = NULL; |
240 | } |
241 | |
242 | static inline int is_cgroup_event(struct perf_event *event) |
243 | { |
244 | return event->cgrp != NULL; |
245 | } |
246 | |
247 | static inline u64 perf_cgroup_event_time(struct perf_event *event) |
248 | { |
249 | struct perf_cgroup_info *t; |
250 | |
251 | t = per_cpu_ptr(event->cgrp->info, event->cpu); |
252 | return t->time; |
253 | } |
254 | |
255 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) |
256 | { |
257 | struct perf_cgroup_info *info; |
258 | u64 now; |
259 | |
260 | now = perf_clock(); |
261 | |
262 | info = this_cpu_ptr(cgrp->info); |
263 | |
264 | info->time += now - info->timestamp; |
265 | info->timestamp = now; |
266 | } |
267 | |
268 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) |
269 | { |
270 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; |
271 | if (cgrp_out) |
272 | __update_cgrp_time(cgrp_out); |
273 | } |
274 | |
275 | static inline void update_cgrp_time_from_event(struct perf_event *event) |
276 | { |
277 | struct perf_cgroup *cgrp; |
278 | |
279 | /* |
280 | * ensure we access cgroup data only when needed and |
281 | * when we know the cgroup is pinned (css_get) |
282 | */ |
283 | if (!is_cgroup_event(event)) |
284 | return; |
285 | |
286 | cgrp = perf_cgroup_from_task(current); |
287 | /* |
288 | * Do not update time when cgroup is not active |
289 | */ |
290 | if (cgrp == event->cgrp) |
291 | __update_cgrp_time(event->cgrp); |
292 | } |
293 | |
294 | static inline void |
295 | perf_cgroup_set_timestamp(struct task_struct *task, |
296 | struct perf_event_context *ctx) |
297 | { |
298 | struct perf_cgroup *cgrp; |
299 | struct perf_cgroup_info *info; |
300 | |
301 | /* |
302 | * ctx->lock held by caller |
303 | * ensure we do not access cgroup data |
304 | * unless we have the cgroup pinned (css_get) |
305 | */ |
306 | if (!task || !ctx->nr_cgroups) |
307 | return; |
308 | |
309 | cgrp = perf_cgroup_from_task(task); |
310 | info = this_cpu_ptr(cgrp->info); |
311 | info->timestamp = ctx->timestamp; |
312 | } |
313 | |
314 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ |
315 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ |
316 | |
317 | /* |
318 | * reschedule events based on the cgroup constraint of task. |
319 | * |
320 | * mode SWOUT : schedule out everything |
321 | * mode SWIN : schedule in based on cgroup for next |
322 | */ |
323 | void perf_cgroup_switch(struct task_struct *task, int mode) |
324 | { |
325 | struct perf_cpu_context *cpuctx; |
326 | struct pmu *pmu; |
327 | unsigned long flags; |
328 | |
329 | /* |
330 | * disable interrupts to avoid geting nr_cgroup |
331 | * changes via __perf_event_disable(). Also |
332 | * avoids preemption. |
333 | */ |
334 | local_irq_save(flags); |
335 | |
336 | /* |
337 | * we reschedule only in the presence of cgroup |
338 | * constrained events. |
339 | */ |
340 | rcu_read_lock(); |
341 | |
342 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
343 | |
344 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
345 | |
346 | perf_pmu_disable(cpuctx->ctx.pmu); |
347 | |
348 | /* |
349 | * perf_cgroup_events says at least one |
350 | * context on this CPU has cgroup events. |
351 | * |
352 | * ctx->nr_cgroups reports the number of cgroup |
353 | * events for a context. |
354 | */ |
355 | if (cpuctx->ctx.nr_cgroups > 0) { |
356 | |
357 | if (mode & PERF_CGROUP_SWOUT) { |
358 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
359 | /* |
360 | * must not be done before ctxswout due |
361 | * to event_filter_match() in event_sched_out() |
362 | */ |
363 | cpuctx->cgrp = NULL; |
364 | } |
365 | |
366 | if (mode & PERF_CGROUP_SWIN) { |
367 | WARN_ON_ONCE(cpuctx->cgrp); |
368 | /* set cgrp before ctxsw in to |
369 | * allow event_filter_match() to not |
370 | * have to pass task around |
371 | */ |
372 | cpuctx->cgrp = perf_cgroup_from_task(task); |
373 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); |
374 | } |
375 | } |
376 | |
377 | perf_pmu_enable(cpuctx->ctx.pmu); |
378 | } |
379 | |
380 | rcu_read_unlock(); |
381 | |
382 | local_irq_restore(flags); |
383 | } |
384 | |
385 | static inline void perf_cgroup_sched_out(struct task_struct *task) |
386 | { |
387 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); |
388 | } |
389 | |
390 | static inline void perf_cgroup_sched_in(struct task_struct *task) |
391 | { |
392 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); |
393 | } |
394 | |
395 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, |
396 | struct perf_event_attr *attr, |
397 | struct perf_event *group_leader) |
398 | { |
399 | struct perf_cgroup *cgrp; |
400 | struct cgroup_subsys_state *css; |
401 | struct file *file; |
402 | int ret = 0, fput_needed; |
403 | |
404 | file = fget_light(fd, &fput_needed); |
405 | if (!file) |
406 | return -EBADF; |
407 | |
408 | css = cgroup_css_from_dir(file, perf_subsys_id); |
409 | if (IS_ERR(css)) { |
410 | ret = PTR_ERR(css); |
411 | goto out; |
412 | } |
413 | |
414 | cgrp = container_of(css, struct perf_cgroup, css); |
415 | event->cgrp = cgrp; |
416 | |
417 | /* must be done before we fput() the file */ |
418 | perf_get_cgroup(event); |
419 | |
420 | /* |
421 | * all events in a group must monitor |
422 | * the same cgroup because a task belongs |
423 | * to only one perf cgroup at a time |
424 | */ |
425 | if (group_leader && group_leader->cgrp != cgrp) { |
426 | perf_detach_cgroup(event); |
427 | ret = -EINVAL; |
428 | } |
429 | out: |
430 | fput_light(file, fput_needed); |
431 | return ret; |
432 | } |
433 | |
434 | static inline void |
435 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) |
436 | { |
437 | struct perf_cgroup_info *t; |
438 | t = per_cpu_ptr(event->cgrp->info, event->cpu); |
439 | event->shadow_ctx_time = now - t->timestamp; |
440 | } |
441 | |
442 | static inline void |
443 | perf_cgroup_defer_enabled(struct perf_event *event) |
444 | { |
445 | /* |
446 | * when the current task's perf cgroup does not match |
447 | * the event's, we need to remember to call the |
448 | * perf_mark_enable() function the first time a task with |
449 | * a matching perf cgroup is scheduled in. |
450 | */ |
451 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) |
452 | event->cgrp_defer_enabled = 1; |
453 | } |
454 | |
455 | static inline void |
456 | perf_cgroup_mark_enabled(struct perf_event *event, |
457 | struct perf_event_context *ctx) |
458 | { |
459 | struct perf_event *sub; |
460 | u64 tstamp = perf_event_time(event); |
461 | |
462 | if (!event->cgrp_defer_enabled) |
463 | return; |
464 | |
465 | event->cgrp_defer_enabled = 0; |
466 | |
467 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
468 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
469 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { |
470 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; |
471 | sub->cgrp_defer_enabled = 0; |
472 | } |
473 | } |
474 | } |
475 | #else /* !CONFIG_CGROUP_PERF */ |
476 | |
477 | static inline bool |
478 | perf_cgroup_match(struct perf_event *event) |
479 | { |
480 | return true; |
481 | } |
482 | |
483 | static inline void perf_detach_cgroup(struct perf_event *event) |
484 | {} |
485 | |
486 | static inline int is_cgroup_event(struct perf_event *event) |
487 | { |
488 | return 0; |
489 | } |
490 | |
491 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) |
492 | { |
493 | return 0; |
494 | } |
495 | |
496 | static inline void update_cgrp_time_from_event(struct perf_event *event) |
497 | { |
498 | } |
499 | |
500 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) |
501 | { |
502 | } |
503 | |
504 | static inline void perf_cgroup_sched_out(struct task_struct *task) |
505 | { |
506 | } |
507 | |
508 | static inline void perf_cgroup_sched_in(struct task_struct *task) |
509 | { |
510 | } |
511 | |
512 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, |
513 | struct perf_event_attr *attr, |
514 | struct perf_event *group_leader) |
515 | { |
516 | return -EINVAL; |
517 | } |
518 | |
519 | static inline void |
520 | perf_cgroup_set_timestamp(struct task_struct *task, |
521 | struct perf_event_context *ctx) |
522 | { |
523 | } |
524 | |
525 | void |
526 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) |
527 | { |
528 | } |
529 | |
530 | static inline void |
531 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) |
532 | { |
533 | } |
534 | |
535 | static inline u64 perf_cgroup_event_time(struct perf_event *event) |
536 | { |
537 | return 0; |
538 | } |
539 | |
540 | static inline void |
541 | perf_cgroup_defer_enabled(struct perf_event *event) |
542 | { |
543 | } |
544 | |
545 | static inline void |
546 | perf_cgroup_mark_enabled(struct perf_event *event, |
547 | struct perf_event_context *ctx) |
548 | { |
549 | } |
550 | #endif |
551 | |
552 | void perf_pmu_disable(struct pmu *pmu) |
553 | { |
554 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
555 | if (!(*count)++) |
556 | pmu->pmu_disable(pmu); |
557 | } |
558 | |
559 | void perf_pmu_enable(struct pmu *pmu) |
560 | { |
561 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
562 | if (!--(*count)) |
563 | pmu->pmu_enable(pmu); |
564 | } |
565 | |
566 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
567 | |
568 | /* |
569 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
570 | * because they're strictly cpu affine and rotate_start is called with IRQs |
571 | * disabled, while rotate_context is called from IRQ context. |
572 | */ |
573 | static void perf_pmu_rotate_start(struct pmu *pmu) |
574 | { |
575 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
576 | struct list_head *head = &__get_cpu_var(rotation_list); |
577 | |
578 | WARN_ON(!irqs_disabled()); |
579 | |
580 | if (list_empty(&cpuctx->rotation_list)) |
581 | list_add(&cpuctx->rotation_list, head); |
582 | } |
583 | |
584 | static void get_ctx(struct perf_event_context *ctx) |
585 | { |
586 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
587 | } |
588 | |
589 | static void free_ctx(struct rcu_head *head) |
590 | { |
591 | struct perf_event_context *ctx; |
592 | |
593 | ctx = container_of(head, struct perf_event_context, rcu_head); |
594 | kfree(ctx); |
595 | } |
596 | |
597 | static void put_ctx(struct perf_event_context *ctx) |
598 | { |
599 | if (atomic_dec_and_test(&ctx->refcount)) { |
600 | if (ctx->parent_ctx) |
601 | put_ctx(ctx->parent_ctx); |
602 | if (ctx->task) |
603 | put_task_struct(ctx->task); |
604 | call_rcu(&ctx->rcu_head, free_ctx); |
605 | } |
606 | } |
607 | |
608 | static void unclone_ctx(struct perf_event_context *ctx) |
609 | { |
610 | if (ctx->parent_ctx) { |
611 | put_ctx(ctx->parent_ctx); |
612 | ctx->parent_ctx = NULL; |
613 | } |
614 | } |
615 | |
616 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
617 | { |
618 | /* |
619 | * only top level events have the pid namespace they were created in |
620 | */ |
621 | if (event->parent) |
622 | event = event->parent; |
623 | |
624 | return task_tgid_nr_ns(p, event->ns); |
625 | } |
626 | |
627 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) |
628 | { |
629 | /* |
630 | * only top level events have the pid namespace they were created in |
631 | */ |
632 | if (event->parent) |
633 | event = event->parent; |
634 | |
635 | return task_pid_nr_ns(p, event->ns); |
636 | } |
637 | |
638 | /* |
639 | * If we inherit events we want to return the parent event id |
640 | * to userspace. |
641 | */ |
642 | static u64 primary_event_id(struct perf_event *event) |
643 | { |
644 | u64 id = event->id; |
645 | |
646 | if (event->parent) |
647 | id = event->parent->id; |
648 | |
649 | return id; |
650 | } |
651 | |
652 | /* |
653 | * Get the perf_event_context for a task and lock it. |
654 | * This has to cope with with the fact that until it is locked, |
655 | * the context could get moved to another task. |
656 | */ |
657 | static struct perf_event_context * |
658 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
659 | { |
660 | struct perf_event_context *ctx; |
661 | |
662 | rcu_read_lock(); |
663 | retry: |
664 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
665 | if (ctx) { |
666 | /* |
667 | * If this context is a clone of another, it might |
668 | * get swapped for another underneath us by |
669 | * perf_event_task_sched_out, though the |
670 | * rcu_read_lock() protects us from any context |
671 | * getting freed. Lock the context and check if it |
672 | * got swapped before we could get the lock, and retry |
673 | * if so. If we locked the right context, then it |
674 | * can't get swapped on us any more. |
675 | */ |
676 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
677 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
678 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
679 | goto retry; |
680 | } |
681 | |
682 | if (!atomic_inc_not_zero(&ctx->refcount)) { |
683 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
684 | ctx = NULL; |
685 | } |
686 | } |
687 | rcu_read_unlock(); |
688 | return ctx; |
689 | } |
690 | |
691 | /* |
692 | * Get the context for a task and increment its pin_count so it |
693 | * can't get swapped to another task. This also increments its |
694 | * reference count so that the context can't get freed. |
695 | */ |
696 | static struct perf_event_context * |
697 | perf_pin_task_context(struct task_struct *task, int ctxn) |
698 | { |
699 | struct perf_event_context *ctx; |
700 | unsigned long flags; |
701 | |
702 | ctx = perf_lock_task_context(task, ctxn, &flags); |
703 | if (ctx) { |
704 | ++ctx->pin_count; |
705 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
706 | } |
707 | return ctx; |
708 | } |
709 | |
710 | static void perf_unpin_context(struct perf_event_context *ctx) |
711 | { |
712 | unsigned long flags; |
713 | |
714 | raw_spin_lock_irqsave(&ctx->lock, flags); |
715 | --ctx->pin_count; |
716 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
717 | } |
718 | |
719 | /* |
720 | * Update the record of the current time in a context. |
721 | */ |
722 | static void update_context_time(struct perf_event_context *ctx) |
723 | { |
724 | u64 now = perf_clock(); |
725 | |
726 | ctx->time += now - ctx->timestamp; |
727 | ctx->timestamp = now; |
728 | } |
729 | |
730 | static u64 perf_event_time(struct perf_event *event) |
731 | { |
732 | struct perf_event_context *ctx = event->ctx; |
733 | |
734 | if (is_cgroup_event(event)) |
735 | return perf_cgroup_event_time(event); |
736 | |
737 | return ctx ? ctx->time : 0; |
738 | } |
739 | |
740 | /* |
741 | * Update the total_time_enabled and total_time_running fields for a event. |
742 | */ |
743 | static void update_event_times(struct perf_event *event) |
744 | { |
745 | struct perf_event_context *ctx = event->ctx; |
746 | u64 run_end; |
747 | |
748 | if (event->state < PERF_EVENT_STATE_INACTIVE || |
749 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) |
750 | return; |
751 | /* |
752 | * in cgroup mode, time_enabled represents |
753 | * the time the event was enabled AND active |
754 | * tasks were in the monitored cgroup. This is |
755 | * independent of the activity of the context as |
756 | * there may be a mix of cgroup and non-cgroup events. |
757 | * |
758 | * That is why we treat cgroup events differently |
759 | * here. |
760 | */ |
761 | if (is_cgroup_event(event)) |
762 | run_end = perf_event_time(event); |
763 | else if (ctx->is_active) |
764 | run_end = ctx->time; |
765 | else |
766 | run_end = event->tstamp_stopped; |
767 | |
768 | event->total_time_enabled = run_end - event->tstamp_enabled; |
769 | |
770 | if (event->state == PERF_EVENT_STATE_INACTIVE) |
771 | run_end = event->tstamp_stopped; |
772 | else |
773 | run_end = perf_event_time(event); |
774 | |
775 | event->total_time_running = run_end - event->tstamp_running; |
776 | |
777 | } |
778 | |
779 | /* |
780 | * Update total_time_enabled and total_time_running for all events in a group. |
781 | */ |
782 | static void update_group_times(struct perf_event *leader) |
783 | { |
784 | struct perf_event *event; |
785 | |
786 | update_event_times(leader); |
787 | list_for_each_entry(event, &leader->sibling_list, group_entry) |
788 | update_event_times(event); |
789 | } |
790 | |
791 | static struct list_head * |
792 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) |
793 | { |
794 | if (event->attr.pinned) |
795 | return &ctx->pinned_groups; |
796 | else |
797 | return &ctx->flexible_groups; |
798 | } |
799 | |
800 | /* |
801 | * Add a event from the lists for its context. |
802 | * Must be called with ctx->mutex and ctx->lock held. |
803 | */ |
804 | static void |
805 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
806 | { |
807 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
808 | event->attach_state |= PERF_ATTACH_CONTEXT; |
809 | |
810 | /* |
811 | * If we're a stand alone event or group leader, we go to the context |
812 | * list, group events are kept attached to the group so that |
813 | * perf_group_detach can, at all times, locate all siblings. |
814 | */ |
815 | if (event->group_leader == event) { |
816 | struct list_head *list; |
817 | |
818 | if (is_software_event(event)) |
819 | event->group_flags |= PERF_GROUP_SOFTWARE; |
820 | |
821 | list = ctx_group_list(event, ctx); |
822 | list_add_tail(&event->group_entry, list); |
823 | } |
824 | |
825 | if (is_cgroup_event(event)) |
826 | ctx->nr_cgroups++; |
827 | |
828 | list_add_rcu(&event->event_entry, &ctx->event_list); |
829 | if (!ctx->nr_events) |
830 | perf_pmu_rotate_start(ctx->pmu); |
831 | ctx->nr_events++; |
832 | if (event->attr.inherit_stat) |
833 | ctx->nr_stat++; |
834 | } |
835 | |
836 | /* |
837 | * Called at perf_event creation and when events are attached/detached from a |
838 | * group. |
839 | */ |
840 | static void perf_event__read_size(struct perf_event *event) |
841 | { |
842 | int entry = sizeof(u64); /* value */ |
843 | int size = 0; |
844 | int nr = 1; |
845 | |
846 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
847 | size += sizeof(u64); |
848 | |
849 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
850 | size += sizeof(u64); |
851 | |
852 | if (event->attr.read_format & PERF_FORMAT_ID) |
853 | entry += sizeof(u64); |
854 | |
855 | if (event->attr.read_format & PERF_FORMAT_GROUP) { |
856 | nr += event->group_leader->nr_siblings; |
857 | size += sizeof(u64); |
858 | } |
859 | |
860 | size += entry * nr; |
861 | event->read_size = size; |
862 | } |
863 | |
864 | static void perf_event__header_size(struct perf_event *event) |
865 | { |
866 | struct perf_sample_data *data; |
867 | u64 sample_type = event->attr.sample_type; |
868 | u16 size = 0; |
869 | |
870 | perf_event__read_size(event); |
871 | |
872 | if (sample_type & PERF_SAMPLE_IP) |
873 | size += sizeof(data->ip); |
874 | |
875 | if (sample_type & PERF_SAMPLE_ADDR) |
876 | size += sizeof(data->addr); |
877 | |
878 | if (sample_type & PERF_SAMPLE_PERIOD) |
879 | size += sizeof(data->period); |
880 | |
881 | if (sample_type & PERF_SAMPLE_READ) |
882 | size += event->read_size; |
883 | |
884 | event->header_size = size; |
885 | } |
886 | |
887 | static void perf_event__id_header_size(struct perf_event *event) |
888 | { |
889 | struct perf_sample_data *data; |
890 | u64 sample_type = event->attr.sample_type; |
891 | u16 size = 0; |
892 | |
893 | if (sample_type & PERF_SAMPLE_TID) |
894 | size += sizeof(data->tid_entry); |
895 | |
896 | if (sample_type & PERF_SAMPLE_TIME) |
897 | size += sizeof(data->time); |
898 | |
899 | if (sample_type & PERF_SAMPLE_ID) |
900 | size += sizeof(data->id); |
901 | |
902 | if (sample_type & PERF_SAMPLE_STREAM_ID) |
903 | size += sizeof(data->stream_id); |
904 | |
905 | if (sample_type & PERF_SAMPLE_CPU) |
906 | size += sizeof(data->cpu_entry); |
907 | |
908 | event->id_header_size = size; |
909 | } |
910 | |
911 | static void perf_group_attach(struct perf_event *event) |
912 | { |
913 | struct perf_event *group_leader = event->group_leader, *pos; |
914 | |
915 | /* |
916 | * We can have double attach due to group movement in perf_event_open. |
917 | */ |
918 | if (event->attach_state & PERF_ATTACH_GROUP) |
919 | return; |
920 | |
921 | event->attach_state |= PERF_ATTACH_GROUP; |
922 | |
923 | if (group_leader == event) |
924 | return; |
925 | |
926 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && |
927 | !is_software_event(event)) |
928 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; |
929 | |
930 | list_add_tail(&event->group_entry, &group_leader->sibling_list); |
931 | group_leader->nr_siblings++; |
932 | |
933 | perf_event__header_size(group_leader); |
934 | |
935 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) |
936 | perf_event__header_size(pos); |
937 | } |
938 | |
939 | /* |
940 | * Remove a event from the lists for its context. |
941 | * Must be called with ctx->mutex and ctx->lock held. |
942 | */ |
943 | static void |
944 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
945 | { |
946 | struct perf_cpu_context *cpuctx; |
947 | /* |
948 | * We can have double detach due to exit/hot-unplug + close. |
949 | */ |
950 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) |
951 | return; |
952 | |
953 | event->attach_state &= ~PERF_ATTACH_CONTEXT; |
954 | |
955 | if (is_cgroup_event(event)) { |
956 | ctx->nr_cgroups--; |
957 | cpuctx = __get_cpu_context(ctx); |
958 | /* |
959 | * if there are no more cgroup events |
960 | * then cler cgrp to avoid stale pointer |
961 | * in update_cgrp_time_from_cpuctx() |
962 | */ |
963 | if (!ctx->nr_cgroups) |
964 | cpuctx->cgrp = NULL; |
965 | } |
966 | |
967 | ctx->nr_events--; |
968 | if (event->attr.inherit_stat) |
969 | ctx->nr_stat--; |
970 | |
971 | list_del_rcu(&event->event_entry); |
972 | |
973 | if (event->group_leader == event) |
974 | list_del_init(&event->group_entry); |
975 | |
976 | update_group_times(event); |
977 | |
978 | /* |
979 | * If event was in error state, then keep it |
980 | * that way, otherwise bogus counts will be |
981 | * returned on read(). The only way to get out |
982 | * of error state is by explicit re-enabling |
983 | * of the event |
984 | */ |
985 | if (event->state > PERF_EVENT_STATE_OFF) |
986 | event->state = PERF_EVENT_STATE_OFF; |
987 | } |
988 | |
989 | static void perf_group_detach(struct perf_event *event) |
990 | { |
991 | struct perf_event *sibling, *tmp; |
992 | struct list_head *list = NULL; |
993 | |
994 | /* |
995 | * We can have double detach due to exit/hot-unplug + close. |
996 | */ |
997 | if (!(event->attach_state & PERF_ATTACH_GROUP)) |
998 | return; |
999 | |
1000 | event->attach_state &= ~PERF_ATTACH_GROUP; |
1001 | |
1002 | /* |
1003 | * If this is a sibling, remove it from its group. |
1004 | */ |
1005 | if (event->group_leader != event) { |
1006 | list_del_init(&event->group_entry); |
1007 | event->group_leader->nr_siblings--; |
1008 | goto out; |
1009 | } |
1010 | |
1011 | if (!list_empty(&event->group_entry)) |
1012 | list = &event->group_entry; |
1013 | |
1014 | /* |
1015 | * If this was a group event with sibling events then |
1016 | * upgrade the siblings to singleton events by adding them |
1017 | * to whatever list we are on. |
1018 | */ |
1019 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
1020 | if (list) |
1021 | list_move_tail(&sibling->group_entry, list); |
1022 | sibling->group_leader = sibling; |
1023 | |
1024 | /* Inherit group flags from the previous leader */ |
1025 | sibling->group_flags = event->group_flags; |
1026 | } |
1027 | |
1028 | out: |
1029 | perf_event__header_size(event->group_leader); |
1030 | |
1031 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) |
1032 | perf_event__header_size(tmp); |
1033 | } |
1034 | |
1035 | static inline int |
1036 | event_filter_match(struct perf_event *event) |
1037 | { |
1038 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1039 | && perf_cgroup_match(event); |
1040 | } |
1041 | |
1042 | static void |
1043 | event_sched_out(struct perf_event *event, |
1044 | struct perf_cpu_context *cpuctx, |
1045 | struct perf_event_context *ctx) |
1046 | { |
1047 | u64 tstamp = perf_event_time(event); |
1048 | u64 delta; |
1049 | /* |
1050 | * An event which could not be activated because of |
1051 | * filter mismatch still needs to have its timings |
1052 | * maintained, otherwise bogus information is return |
1053 | * via read() for time_enabled, time_running: |
1054 | */ |
1055 | if (event->state == PERF_EVENT_STATE_INACTIVE |
1056 | && !event_filter_match(event)) { |
1057 | delta = tstamp - event->tstamp_stopped; |
1058 | event->tstamp_running += delta; |
1059 | event->tstamp_stopped = tstamp; |
1060 | } |
1061 | |
1062 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
1063 | return; |
1064 | |
1065 | event->state = PERF_EVENT_STATE_INACTIVE; |
1066 | if (event->pending_disable) { |
1067 | event->pending_disable = 0; |
1068 | event->state = PERF_EVENT_STATE_OFF; |
1069 | } |
1070 | event->tstamp_stopped = tstamp; |
1071 | event->pmu->del(event, 0); |
1072 | event->oncpu = -1; |
1073 | |
1074 | if (!is_software_event(event)) |
1075 | cpuctx->active_oncpu--; |
1076 | ctx->nr_active--; |
1077 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
1078 | cpuctx->exclusive = 0; |
1079 | } |
1080 | |
1081 | static void |
1082 | group_sched_out(struct perf_event *group_event, |
1083 | struct perf_cpu_context *cpuctx, |
1084 | struct perf_event_context *ctx) |
1085 | { |
1086 | struct perf_event *event; |
1087 | int state = group_event->state; |
1088 | |
1089 | event_sched_out(group_event, cpuctx, ctx); |
1090 | |
1091 | /* |
1092 | * Schedule out siblings (if any): |
1093 | */ |
1094 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1095 | event_sched_out(event, cpuctx, ctx); |
1096 | |
1097 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
1098 | cpuctx->exclusive = 0; |
1099 | } |
1100 | |
1101 | /* |
1102 | * Cross CPU call to remove a performance event |
1103 | * |
1104 | * We disable the event on the hardware level first. After that we |
1105 | * remove it from the context list. |
1106 | */ |
1107 | static int __perf_remove_from_context(void *info) |
1108 | { |
1109 | struct perf_event *event = info; |
1110 | struct perf_event_context *ctx = event->ctx; |
1111 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
1112 | |
1113 | raw_spin_lock(&ctx->lock); |
1114 | event_sched_out(event, cpuctx, ctx); |
1115 | list_del_event(event, ctx); |
1116 | raw_spin_unlock(&ctx->lock); |
1117 | |
1118 | return 0; |
1119 | } |
1120 | |
1121 | |
1122 | /* |
1123 | * Remove the event from a task's (or a CPU's) list of events. |
1124 | * |
1125 | * CPU events are removed with a smp call. For task events we only |
1126 | * call when the task is on a CPU. |
1127 | * |
1128 | * If event->ctx is a cloned context, callers must make sure that |
1129 | * every task struct that event->ctx->task could possibly point to |
1130 | * remains valid. This is OK when called from perf_release since |
1131 | * that only calls us on the top-level context, which can't be a clone. |
1132 | * When called from perf_event_exit_task, it's OK because the |
1133 | * context has been detached from its task. |
1134 | */ |
1135 | static void perf_remove_from_context(struct perf_event *event) |
1136 | { |
1137 | struct perf_event_context *ctx = event->ctx; |
1138 | struct task_struct *task = ctx->task; |
1139 | |
1140 | lockdep_assert_held(&ctx->mutex); |
1141 | |
1142 | if (!task) { |
1143 | /* |
1144 | * Per cpu events are removed via an smp call and |
1145 | * the removal is always successful. |
1146 | */ |
1147 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
1148 | return; |
1149 | } |
1150 | |
1151 | retry: |
1152 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1153 | return; |
1154 | |
1155 | raw_spin_lock_irq(&ctx->lock); |
1156 | /* |
1157 | * If we failed to find a running task, but find the context active now |
1158 | * that we've acquired the ctx->lock, retry. |
1159 | */ |
1160 | if (ctx->is_active) { |
1161 | raw_spin_unlock_irq(&ctx->lock); |
1162 | goto retry; |
1163 | } |
1164 | |
1165 | /* |
1166 | * Since the task isn't running, its safe to remove the event, us |
1167 | * holding the ctx->lock ensures the task won't get scheduled in. |
1168 | */ |
1169 | list_del_event(event, ctx); |
1170 | raw_spin_unlock_irq(&ctx->lock); |
1171 | } |
1172 | |
1173 | /* |
1174 | * Cross CPU call to disable a performance event |
1175 | */ |
1176 | static int __perf_event_disable(void *info) |
1177 | { |
1178 | struct perf_event *event = info; |
1179 | struct perf_event_context *ctx = event->ctx; |
1180 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
1181 | |
1182 | /* |
1183 | * If this is a per-task event, need to check whether this |
1184 | * event's task is the current task on this cpu. |
1185 | * |
1186 | * Can trigger due to concurrent perf_event_context_sched_out() |
1187 | * flipping contexts around. |
1188 | */ |
1189 | if (ctx->task && cpuctx->task_ctx != ctx) |
1190 | return -EINVAL; |
1191 | |
1192 | raw_spin_lock(&ctx->lock); |
1193 | |
1194 | /* |
1195 | * If the event is on, turn it off. |
1196 | * If it is in error state, leave it in error state. |
1197 | */ |
1198 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
1199 | update_context_time(ctx); |
1200 | update_cgrp_time_from_event(event); |
1201 | update_group_times(event); |
1202 | if (event == event->group_leader) |
1203 | group_sched_out(event, cpuctx, ctx); |
1204 | else |
1205 | event_sched_out(event, cpuctx, ctx); |
1206 | event->state = PERF_EVENT_STATE_OFF; |
1207 | } |
1208 | |
1209 | raw_spin_unlock(&ctx->lock); |
1210 | |
1211 | return 0; |
1212 | } |
1213 | |
1214 | /* |
1215 | * Disable a event. |
1216 | * |
1217 | * If event->ctx is a cloned context, callers must make sure that |
1218 | * every task struct that event->ctx->task could possibly point to |
1219 | * remains valid. This condition is satisifed when called through |
1220 | * perf_event_for_each_child or perf_event_for_each because they |
1221 | * hold the top-level event's child_mutex, so any descendant that |
1222 | * goes to exit will block in sync_child_event. |
1223 | * When called from perf_pending_event it's OK because event->ctx |
1224 | * is the current context on this CPU and preemption is disabled, |
1225 | * hence we can't get into perf_event_task_sched_out for this context. |
1226 | */ |
1227 | void perf_event_disable(struct perf_event *event) |
1228 | { |
1229 | struct perf_event_context *ctx = event->ctx; |
1230 | struct task_struct *task = ctx->task; |
1231 | |
1232 | if (!task) { |
1233 | /* |
1234 | * Disable the event on the cpu that it's on |
1235 | */ |
1236 | cpu_function_call(event->cpu, __perf_event_disable, event); |
1237 | return; |
1238 | } |
1239 | |
1240 | retry: |
1241 | if (!task_function_call(task, __perf_event_disable, event)) |
1242 | return; |
1243 | |
1244 | raw_spin_lock_irq(&ctx->lock); |
1245 | /* |
1246 | * If the event is still active, we need to retry the cross-call. |
1247 | */ |
1248 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
1249 | raw_spin_unlock_irq(&ctx->lock); |
1250 | /* |
1251 | * Reload the task pointer, it might have been changed by |
1252 | * a concurrent perf_event_context_sched_out(). |
1253 | */ |
1254 | task = ctx->task; |
1255 | goto retry; |
1256 | } |
1257 | |
1258 | /* |
1259 | * Since we have the lock this context can't be scheduled |
1260 | * in, so we can change the state safely. |
1261 | */ |
1262 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1263 | update_group_times(event); |
1264 | event->state = PERF_EVENT_STATE_OFF; |
1265 | } |
1266 | raw_spin_unlock_irq(&ctx->lock); |
1267 | } |
1268 | |
1269 | static void perf_set_shadow_time(struct perf_event *event, |
1270 | struct perf_event_context *ctx, |
1271 | u64 tstamp) |
1272 | { |
1273 | /* |
1274 | * use the correct time source for the time snapshot |
1275 | * |
1276 | * We could get by without this by leveraging the |
1277 | * fact that to get to this function, the caller |
1278 | * has most likely already called update_context_time() |
1279 | * and update_cgrp_time_xx() and thus both timestamp |
1280 | * are identical (or very close). Given that tstamp is, |
1281 | * already adjusted for cgroup, we could say that: |
1282 | * tstamp - ctx->timestamp |
1283 | * is equivalent to |
1284 | * tstamp - cgrp->timestamp. |
1285 | * |
1286 | * Then, in perf_output_read(), the calculation would |
1287 | * work with no changes because: |
1288 | * - event is guaranteed scheduled in |
1289 | * - no scheduled out in between |
1290 | * - thus the timestamp would be the same |
1291 | * |
1292 | * But this is a bit hairy. |
1293 | * |
1294 | * So instead, we have an explicit cgroup call to remain |
1295 | * within the time time source all along. We believe it |
1296 | * is cleaner and simpler to understand. |
1297 | */ |
1298 | if (is_cgroup_event(event)) |
1299 | perf_cgroup_set_shadow_time(event, tstamp); |
1300 | else |
1301 | event->shadow_ctx_time = tstamp - ctx->timestamp; |
1302 | } |
1303 | |
1304 | #define MAX_INTERRUPTS (~0ULL) |
1305 | |
1306 | static void perf_log_throttle(struct perf_event *event, int enable); |
1307 | |
1308 | static int |
1309 | event_sched_in(struct perf_event *event, |
1310 | struct perf_cpu_context *cpuctx, |
1311 | struct perf_event_context *ctx) |
1312 | { |
1313 | u64 tstamp = perf_event_time(event); |
1314 | |
1315 | if (event->state <= PERF_EVENT_STATE_OFF) |
1316 | return 0; |
1317 | |
1318 | event->state = PERF_EVENT_STATE_ACTIVE; |
1319 | event->oncpu = smp_processor_id(); |
1320 | |
1321 | /* |
1322 | * Unthrottle events, since we scheduled we might have missed several |
1323 | * ticks already, also for a heavily scheduling task there is little |
1324 | * guarantee it'll get a tick in a timely manner. |
1325 | */ |
1326 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { |
1327 | perf_log_throttle(event, 1); |
1328 | event->hw.interrupts = 0; |
1329 | } |
1330 | |
1331 | /* |
1332 | * The new state must be visible before we turn it on in the hardware: |
1333 | */ |
1334 | smp_wmb(); |
1335 | |
1336 | if (event->pmu->add(event, PERF_EF_START)) { |
1337 | event->state = PERF_EVENT_STATE_INACTIVE; |
1338 | event->oncpu = -1; |
1339 | return -EAGAIN; |
1340 | } |
1341 | |
1342 | event->tstamp_running += tstamp - event->tstamp_stopped; |
1343 | |
1344 | perf_set_shadow_time(event, ctx, tstamp); |
1345 | |
1346 | if (!is_software_event(event)) |
1347 | cpuctx->active_oncpu++; |
1348 | ctx->nr_active++; |
1349 | |
1350 | if (event->attr.exclusive) |
1351 | cpuctx->exclusive = 1; |
1352 | |
1353 | return 0; |
1354 | } |
1355 | |
1356 | static int |
1357 | group_sched_in(struct perf_event *group_event, |
1358 | struct perf_cpu_context *cpuctx, |
1359 | struct perf_event_context *ctx) |
1360 | { |
1361 | struct perf_event *event, *partial_group = NULL; |
1362 | struct pmu *pmu = group_event->pmu; |
1363 | u64 now = ctx->time; |
1364 | bool simulate = false; |
1365 | |
1366 | if (group_event->state == PERF_EVENT_STATE_OFF) |
1367 | return 0; |
1368 | |
1369 | pmu->start_txn(pmu); |
1370 | |
1371 | if (event_sched_in(group_event, cpuctx, ctx)) { |
1372 | pmu->cancel_txn(pmu); |
1373 | return -EAGAIN; |
1374 | } |
1375 | |
1376 | /* |
1377 | * Schedule in siblings as one group (if any): |
1378 | */ |
1379 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1380 | if (event_sched_in(event, cpuctx, ctx)) { |
1381 | partial_group = event; |
1382 | goto group_error; |
1383 | } |
1384 | } |
1385 | |
1386 | if (!pmu->commit_txn(pmu)) |
1387 | return 0; |
1388 | |
1389 | group_error: |
1390 | /* |
1391 | * Groups can be scheduled in as one unit only, so undo any |
1392 | * partial group before returning: |
1393 | * The events up to the failed event are scheduled out normally, |
1394 | * tstamp_stopped will be updated. |
1395 | * |
1396 | * The failed events and the remaining siblings need to have |
1397 | * their timings updated as if they had gone thru event_sched_in() |
1398 | * and event_sched_out(). This is required to get consistent timings |
1399 | * across the group. This also takes care of the case where the group |
1400 | * could never be scheduled by ensuring tstamp_stopped is set to mark |
1401 | * the time the event was actually stopped, such that time delta |
1402 | * calculation in update_event_times() is correct. |
1403 | */ |
1404 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1405 | if (event == partial_group) |
1406 | simulate = true; |
1407 | |
1408 | if (simulate) { |
1409 | event->tstamp_running += now - event->tstamp_stopped; |
1410 | event->tstamp_stopped = now; |
1411 | } else { |
1412 | event_sched_out(event, cpuctx, ctx); |
1413 | } |
1414 | } |
1415 | event_sched_out(group_event, cpuctx, ctx); |
1416 | |
1417 | pmu->cancel_txn(pmu); |
1418 | |
1419 | return -EAGAIN; |
1420 | } |
1421 | |
1422 | /* |
1423 | * Work out whether we can put this event group on the CPU now. |
1424 | */ |
1425 | static int group_can_go_on(struct perf_event *event, |
1426 | struct perf_cpu_context *cpuctx, |
1427 | int can_add_hw) |
1428 | { |
1429 | /* |
1430 | * Groups consisting entirely of software events can always go on. |
1431 | */ |
1432 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
1433 | return 1; |
1434 | /* |
1435 | * If an exclusive group is already on, no other hardware |
1436 | * events can go on. |
1437 | */ |
1438 | if (cpuctx->exclusive) |
1439 | return 0; |
1440 | /* |
1441 | * If this group is exclusive and there are already |
1442 | * events on the CPU, it can't go on. |
1443 | */ |
1444 | if (event->attr.exclusive && cpuctx->active_oncpu) |
1445 | return 0; |
1446 | /* |
1447 | * Otherwise, try to add it if all previous groups were able |
1448 | * to go on. |
1449 | */ |
1450 | return can_add_hw; |
1451 | } |
1452 | |
1453 | static void add_event_to_ctx(struct perf_event *event, |
1454 | struct perf_event_context *ctx) |
1455 | { |
1456 | u64 tstamp = perf_event_time(event); |
1457 | |
1458 | list_add_event(event, ctx); |
1459 | perf_group_attach(event); |
1460 | event->tstamp_enabled = tstamp; |
1461 | event->tstamp_running = tstamp; |
1462 | event->tstamp_stopped = tstamp; |
1463 | } |
1464 | |
1465 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
1466 | struct task_struct *tsk); |
1467 | |
1468 | /* |
1469 | * Cross CPU call to install and enable a performance event |
1470 | * |
1471 | * Must be called with ctx->mutex held |
1472 | */ |
1473 | static int __perf_install_in_context(void *info) |
1474 | { |
1475 | struct perf_event *event = info; |
1476 | struct perf_event_context *ctx = event->ctx; |
1477 | struct perf_event *leader = event->group_leader; |
1478 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
1479 | int err; |
1480 | |
1481 | /* |
1482 | * In case we're installing a new context to an already running task, |
1483 | * could also happen before perf_event_task_sched_in() on architectures |
1484 | * which do context switches with IRQs enabled. |
1485 | */ |
1486 | if (ctx->task && !cpuctx->task_ctx) |
1487 | perf_event_context_sched_in(ctx, ctx->task); |
1488 | |
1489 | raw_spin_lock(&ctx->lock); |
1490 | ctx->is_active = 1; |
1491 | update_context_time(ctx); |
1492 | /* |
1493 | * update cgrp time only if current cgrp |
1494 | * matches event->cgrp. Must be done before |
1495 | * calling add_event_to_ctx() |
1496 | */ |
1497 | update_cgrp_time_from_event(event); |
1498 | |
1499 | add_event_to_ctx(event, ctx); |
1500 | |
1501 | if (!event_filter_match(event)) |
1502 | goto unlock; |
1503 | |
1504 | /* |
1505 | * Don't put the event on if it is disabled or if |
1506 | * it is in a group and the group isn't on. |
1507 | */ |
1508 | if (event->state != PERF_EVENT_STATE_INACTIVE || |
1509 | (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)) |
1510 | goto unlock; |
1511 | |
1512 | /* |
1513 | * An exclusive event can't go on if there are already active |
1514 | * hardware events, and no hardware event can go on if there |
1515 | * is already an exclusive event on. |
1516 | */ |
1517 | if (!group_can_go_on(event, cpuctx, 1)) |
1518 | err = -EEXIST; |
1519 | else |
1520 | err = event_sched_in(event, cpuctx, ctx); |
1521 | |
1522 | if (err) { |
1523 | /* |
1524 | * This event couldn't go on. If it is in a group |
1525 | * then we have to pull the whole group off. |
1526 | * If the event group is pinned then put it in error state. |
1527 | */ |
1528 | if (leader != event) |
1529 | group_sched_out(leader, cpuctx, ctx); |
1530 | if (leader->attr.pinned) { |
1531 | update_group_times(leader); |
1532 | leader->state = PERF_EVENT_STATE_ERROR; |
1533 | } |
1534 | } |
1535 | |
1536 | unlock: |
1537 | raw_spin_unlock(&ctx->lock); |
1538 | |
1539 | return 0; |
1540 | } |
1541 | |
1542 | /* |
1543 | * Attach a performance event to a context |
1544 | * |
1545 | * First we add the event to the list with the hardware enable bit |
1546 | * in event->hw_config cleared. |
1547 | * |
1548 | * If the event is attached to a task which is on a CPU we use a smp |
1549 | * call to enable it in the task context. The task might have been |
1550 | * scheduled away, but we check this in the smp call again. |
1551 | */ |
1552 | static void |
1553 | perf_install_in_context(struct perf_event_context *ctx, |
1554 | struct perf_event *event, |
1555 | int cpu) |
1556 | { |
1557 | struct task_struct *task = ctx->task; |
1558 | |
1559 | lockdep_assert_held(&ctx->mutex); |
1560 | |
1561 | event->ctx = ctx; |
1562 | |
1563 | if (!task) { |
1564 | /* |
1565 | * Per cpu events are installed via an smp call and |
1566 | * the install is always successful. |
1567 | */ |
1568 | cpu_function_call(cpu, __perf_install_in_context, event); |
1569 | return; |
1570 | } |
1571 | |
1572 | retry: |
1573 | if (!task_function_call(task, __perf_install_in_context, event)) |
1574 | return; |
1575 | |
1576 | raw_spin_lock_irq(&ctx->lock); |
1577 | /* |
1578 | * If we failed to find a running task, but find the context active now |
1579 | * that we've acquired the ctx->lock, retry. |
1580 | */ |
1581 | if (ctx->is_active) { |
1582 | raw_spin_unlock_irq(&ctx->lock); |
1583 | goto retry; |
1584 | } |
1585 | |
1586 | /* |
1587 | * Since the task isn't running, its safe to add the event, us holding |
1588 | * the ctx->lock ensures the task won't get scheduled in. |
1589 | */ |
1590 | add_event_to_ctx(event, ctx); |
1591 | raw_spin_unlock_irq(&ctx->lock); |
1592 | } |
1593 | |
1594 | /* |
1595 | * Put a event into inactive state and update time fields. |
1596 | * Enabling the leader of a group effectively enables all |
1597 | * the group members that aren't explicitly disabled, so we |
1598 | * have to update their ->tstamp_enabled also. |
1599 | * Note: this works for group members as well as group leaders |
1600 | * since the non-leader members' sibling_lists will be empty. |
1601 | */ |
1602 | static void __perf_event_mark_enabled(struct perf_event *event, |
1603 | struct perf_event_context *ctx) |
1604 | { |
1605 | struct perf_event *sub; |
1606 | u64 tstamp = perf_event_time(event); |
1607 | |
1608 | event->state = PERF_EVENT_STATE_INACTIVE; |
1609 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
1610 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
1611 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1612 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; |
1613 | } |
1614 | } |
1615 | |
1616 | /* |
1617 | * Cross CPU call to enable a performance event |
1618 | */ |
1619 | static int __perf_event_enable(void *info) |
1620 | { |
1621 | struct perf_event *event = info; |
1622 | struct perf_event_context *ctx = event->ctx; |
1623 | struct perf_event *leader = event->group_leader; |
1624 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
1625 | int err; |
1626 | |
1627 | if (WARN_ON_ONCE(!ctx->is_active)) |
1628 | return -EINVAL; |
1629 | |
1630 | raw_spin_lock(&ctx->lock); |
1631 | update_context_time(ctx); |
1632 | |
1633 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
1634 | goto unlock; |
1635 | |
1636 | /* |
1637 | * set current task's cgroup time reference point |
1638 | */ |
1639 | perf_cgroup_set_timestamp(current, ctx); |
1640 | |
1641 | __perf_event_mark_enabled(event, ctx); |
1642 | |
1643 | if (!event_filter_match(event)) { |
1644 | if (is_cgroup_event(event)) |
1645 | perf_cgroup_defer_enabled(event); |
1646 | goto unlock; |
1647 | } |
1648 | |
1649 | /* |
1650 | * If the event is in a group and isn't the group leader, |
1651 | * then don't put it on unless the group is on. |
1652 | */ |
1653 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
1654 | goto unlock; |
1655 | |
1656 | if (!group_can_go_on(event, cpuctx, 1)) { |
1657 | err = -EEXIST; |
1658 | } else { |
1659 | if (event == leader) |
1660 | err = group_sched_in(event, cpuctx, ctx); |
1661 | else |
1662 | err = event_sched_in(event, cpuctx, ctx); |
1663 | } |
1664 | |
1665 | if (err) { |
1666 | /* |
1667 | * If this event can't go on and it's part of a |
1668 | * group, then the whole group has to come off. |
1669 | */ |
1670 | if (leader != event) |
1671 | group_sched_out(leader, cpuctx, ctx); |
1672 | if (leader->attr.pinned) { |
1673 | update_group_times(leader); |
1674 | leader->state = PERF_EVENT_STATE_ERROR; |
1675 | } |
1676 | } |
1677 | |
1678 | unlock: |
1679 | raw_spin_unlock(&ctx->lock); |
1680 | |
1681 | return 0; |
1682 | } |
1683 | |
1684 | /* |
1685 | * Enable a event. |
1686 | * |
1687 | * If event->ctx is a cloned context, callers must make sure that |
1688 | * every task struct that event->ctx->task could possibly point to |
1689 | * remains valid. This condition is satisfied when called through |
1690 | * perf_event_for_each_child or perf_event_for_each as described |
1691 | * for perf_event_disable. |
1692 | */ |
1693 | void perf_event_enable(struct perf_event *event) |
1694 | { |
1695 | struct perf_event_context *ctx = event->ctx; |
1696 | struct task_struct *task = ctx->task; |
1697 | |
1698 | if (!task) { |
1699 | /* |
1700 | * Enable the event on the cpu that it's on |
1701 | */ |
1702 | cpu_function_call(event->cpu, __perf_event_enable, event); |
1703 | return; |
1704 | } |
1705 | |
1706 | raw_spin_lock_irq(&ctx->lock); |
1707 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
1708 | goto out; |
1709 | |
1710 | /* |
1711 | * If the event is in error state, clear that first. |
1712 | * That way, if we see the event in error state below, we |
1713 | * know that it has gone back into error state, as distinct |
1714 | * from the task having been scheduled away before the |
1715 | * cross-call arrived. |
1716 | */ |
1717 | if (event->state == PERF_EVENT_STATE_ERROR) |
1718 | event->state = PERF_EVENT_STATE_OFF; |
1719 | |
1720 | retry: |
1721 | if (!ctx->is_active) { |
1722 | __perf_event_mark_enabled(event, ctx); |
1723 | goto out; |
1724 | } |
1725 | |
1726 | raw_spin_unlock_irq(&ctx->lock); |
1727 | |
1728 | if (!task_function_call(task, __perf_event_enable, event)) |
1729 | return; |
1730 | |
1731 | raw_spin_lock_irq(&ctx->lock); |
1732 | |
1733 | /* |
1734 | * If the context is active and the event is still off, |
1735 | * we need to retry the cross-call. |
1736 | */ |
1737 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
1738 | /* |
1739 | * task could have been flipped by a concurrent |
1740 | * perf_event_context_sched_out() |
1741 | */ |
1742 | task = ctx->task; |
1743 | goto retry; |
1744 | } |
1745 | |
1746 | out: |
1747 | raw_spin_unlock_irq(&ctx->lock); |
1748 | } |
1749 | |
1750 | static int perf_event_refresh(struct perf_event *event, int refresh) |
1751 | { |
1752 | /* |
1753 | * not supported on inherited events |
1754 | */ |
1755 | if (event->attr.inherit || !is_sampling_event(event)) |
1756 | return -EINVAL; |
1757 | |
1758 | atomic_add(refresh, &event->event_limit); |
1759 | perf_event_enable(event); |
1760 | |
1761 | return 0; |
1762 | } |
1763 | |
1764 | static void ctx_sched_out(struct perf_event_context *ctx, |
1765 | struct perf_cpu_context *cpuctx, |
1766 | enum event_type_t event_type) |
1767 | { |
1768 | struct perf_event *event; |
1769 | |
1770 | raw_spin_lock(&ctx->lock); |
1771 | perf_pmu_disable(ctx->pmu); |
1772 | ctx->is_active = 0; |
1773 | if (likely(!ctx->nr_events)) |
1774 | goto out; |
1775 | update_context_time(ctx); |
1776 | update_cgrp_time_from_cpuctx(cpuctx); |
1777 | |
1778 | if (!ctx->nr_active) |
1779 | goto out; |
1780 | |
1781 | if (event_type & EVENT_PINNED) { |
1782 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
1783 | group_sched_out(event, cpuctx, ctx); |
1784 | } |
1785 | |
1786 | if (event_type & EVENT_FLEXIBLE) { |
1787 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
1788 | group_sched_out(event, cpuctx, ctx); |
1789 | } |
1790 | out: |
1791 | perf_pmu_enable(ctx->pmu); |
1792 | raw_spin_unlock(&ctx->lock); |
1793 | } |
1794 | |
1795 | /* |
1796 | * Test whether two contexts are equivalent, i.e. whether they |
1797 | * have both been cloned from the same version of the same context |
1798 | * and they both have the same number of enabled events. |
1799 | * If the number of enabled events is the same, then the set |
1800 | * of enabled events should be the same, because these are both |
1801 | * inherited contexts, therefore we can't access individual events |
1802 | * in them directly with an fd; we can only enable/disable all |
1803 | * events via prctl, or enable/disable all events in a family |
1804 | * via ioctl, which will have the same effect on both contexts. |
1805 | */ |
1806 | static int context_equiv(struct perf_event_context *ctx1, |
1807 | struct perf_event_context *ctx2) |
1808 | { |
1809 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx |
1810 | && ctx1->parent_gen == ctx2->parent_gen |
1811 | && !ctx1->pin_count && !ctx2->pin_count; |
1812 | } |
1813 | |
1814 | static void __perf_event_sync_stat(struct perf_event *event, |
1815 | struct perf_event *next_event) |
1816 | { |
1817 | u64 value; |
1818 | |
1819 | if (!event->attr.inherit_stat) |
1820 | return; |
1821 | |
1822 | /* |
1823 | * Update the event value, we cannot use perf_event_read() |
1824 | * because we're in the middle of a context switch and have IRQs |
1825 | * disabled, which upsets smp_call_function_single(), however |
1826 | * we know the event must be on the current CPU, therefore we |
1827 | * don't need to use it. |
1828 | */ |
1829 | switch (event->state) { |
1830 | case PERF_EVENT_STATE_ACTIVE: |
1831 | event->pmu->read(event); |
1832 | /* fall-through */ |
1833 | |
1834 | case PERF_EVENT_STATE_INACTIVE: |
1835 | update_event_times(event); |
1836 | break; |
1837 | |
1838 | default: |
1839 | break; |
1840 | } |
1841 | |
1842 | /* |
1843 | * In order to keep per-task stats reliable we need to flip the event |
1844 | * values when we flip the contexts. |
1845 | */ |
1846 | value = local64_read(&next_event->count); |
1847 | value = local64_xchg(&event->count, value); |
1848 | local64_set(&next_event->count, value); |
1849 | |
1850 | swap(event->total_time_enabled, next_event->total_time_enabled); |
1851 | swap(event->total_time_running, next_event->total_time_running); |
1852 | |
1853 | /* |
1854 | * Since we swizzled the values, update the user visible data too. |
1855 | */ |
1856 | perf_event_update_userpage(event); |
1857 | perf_event_update_userpage(next_event); |
1858 | } |
1859 | |
1860 | #define list_next_entry(pos, member) \ |
1861 | list_entry(pos->member.next, typeof(*pos), member) |
1862 | |
1863 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
1864 | struct perf_event_context *next_ctx) |
1865 | { |
1866 | struct perf_event *event, *next_event; |
1867 | |
1868 | if (!ctx->nr_stat) |
1869 | return; |
1870 | |
1871 | update_context_time(ctx); |
1872 | |
1873 | event = list_first_entry(&ctx->event_list, |
1874 | struct perf_event, event_entry); |
1875 | |
1876 | next_event = list_first_entry(&next_ctx->event_list, |
1877 | struct perf_event, event_entry); |
1878 | |
1879 | while (&event->event_entry != &ctx->event_list && |
1880 | &next_event->event_entry != &next_ctx->event_list) { |
1881 | |
1882 | __perf_event_sync_stat(event, next_event); |
1883 | |
1884 | event = list_next_entry(event, event_entry); |
1885 | next_event = list_next_entry(next_event, event_entry); |
1886 | } |
1887 | } |
1888 | |
1889 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
1890 | struct task_struct *next) |
1891 | { |
1892 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
1893 | struct perf_event_context *next_ctx; |
1894 | struct perf_event_context *parent; |
1895 | struct perf_cpu_context *cpuctx; |
1896 | int do_switch = 1; |
1897 | |
1898 | if (likely(!ctx)) |
1899 | return; |
1900 | |
1901 | cpuctx = __get_cpu_context(ctx); |
1902 | if (!cpuctx->task_ctx) |
1903 | return; |
1904 | |
1905 | rcu_read_lock(); |
1906 | parent = rcu_dereference(ctx->parent_ctx); |
1907 | next_ctx = next->perf_event_ctxp[ctxn]; |
1908 | if (parent && next_ctx && |
1909 | rcu_dereference(next_ctx->parent_ctx) == parent) { |
1910 | /* |
1911 | * Looks like the two contexts are clones, so we might be |
1912 | * able to optimize the context switch. We lock both |
1913 | * contexts and check that they are clones under the |
1914 | * lock (including re-checking that neither has been |
1915 | * uncloned in the meantime). It doesn't matter which |
1916 | * order we take the locks because no other cpu could |
1917 | * be trying to lock both of these tasks. |
1918 | */ |
1919 | raw_spin_lock(&ctx->lock); |
1920 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); |
1921 | if (context_equiv(ctx, next_ctx)) { |
1922 | /* |
1923 | * XXX do we need a memory barrier of sorts |
1924 | * wrt to rcu_dereference() of perf_event_ctxp |
1925 | */ |
1926 | task->perf_event_ctxp[ctxn] = next_ctx; |
1927 | next->perf_event_ctxp[ctxn] = ctx; |
1928 | ctx->task = next; |
1929 | next_ctx->task = task; |
1930 | do_switch = 0; |
1931 | |
1932 | perf_event_sync_stat(ctx, next_ctx); |
1933 | } |
1934 | raw_spin_unlock(&next_ctx->lock); |
1935 | raw_spin_unlock(&ctx->lock); |
1936 | } |
1937 | rcu_read_unlock(); |
1938 | |
1939 | if (do_switch) { |
1940 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
1941 | cpuctx->task_ctx = NULL; |
1942 | } |
1943 | } |
1944 | |
1945 | #define for_each_task_context_nr(ctxn) \ |
1946 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) |
1947 | |
1948 | /* |
1949 | * Called from scheduler to remove the events of the current task, |
1950 | * with interrupts disabled. |
1951 | * |
1952 | * We stop each event and update the event value in event->count. |
1953 | * |
1954 | * This does not protect us against NMI, but disable() |
1955 | * sets the disabled bit in the control field of event _before_ |
1956 | * accessing the event control register. If a NMI hits, then it will |
1957 | * not restart the event. |
1958 | */ |
1959 | void __perf_event_task_sched_out(struct task_struct *task, |
1960 | struct task_struct *next) |
1961 | { |
1962 | int ctxn; |
1963 | |
1964 | for_each_task_context_nr(ctxn) |
1965 | perf_event_context_sched_out(task, ctxn, next); |
1966 | |
1967 | /* |
1968 | * if cgroup events exist on this CPU, then we need |
1969 | * to check if we have to switch out PMU state. |
1970 | * cgroup event are system-wide mode only |
1971 | */ |
1972 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) |
1973 | perf_cgroup_sched_out(task); |
1974 | } |
1975 | |
1976 | static void task_ctx_sched_out(struct perf_event_context *ctx, |
1977 | enum event_type_t event_type) |
1978 | { |
1979 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
1980 | |
1981 | if (!cpuctx->task_ctx) |
1982 | return; |
1983 | |
1984 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) |
1985 | return; |
1986 | |
1987 | ctx_sched_out(ctx, cpuctx, event_type); |
1988 | cpuctx->task_ctx = NULL; |
1989 | } |
1990 | |
1991 | /* |
1992 | * Called with IRQs disabled |
1993 | */ |
1994 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
1995 | enum event_type_t event_type) |
1996 | { |
1997 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); |
1998 | } |
1999 | |
2000 | static void |
2001 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
2002 | struct perf_cpu_context *cpuctx) |
2003 | { |
2004 | struct perf_event *event; |
2005 | |
2006 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2007 | if (event->state <= PERF_EVENT_STATE_OFF) |
2008 | continue; |
2009 | if (!event_filter_match(event)) |
2010 | continue; |
2011 | |
2012 | /* may need to reset tstamp_enabled */ |
2013 | if (is_cgroup_event(event)) |
2014 | perf_cgroup_mark_enabled(event, ctx); |
2015 | |
2016 | if (group_can_go_on(event, cpuctx, 1)) |
2017 | group_sched_in(event, cpuctx, ctx); |
2018 | |
2019 | /* |
2020 | * If this pinned group hasn't been scheduled, |
2021 | * put it in error state. |
2022 | */ |
2023 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2024 | update_group_times(event); |
2025 | event->state = PERF_EVENT_STATE_ERROR; |
2026 | } |
2027 | } |
2028 | } |
2029 | |
2030 | static void |
2031 | ctx_flexible_sched_in(struct perf_event_context *ctx, |
2032 | struct perf_cpu_context *cpuctx) |
2033 | { |
2034 | struct perf_event *event; |
2035 | int can_add_hw = 1; |
2036 | |
2037 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2038 | /* Ignore events in OFF or ERROR state */ |
2039 | if (event->state <= PERF_EVENT_STATE_OFF) |
2040 | continue; |
2041 | /* |
2042 | * Listen to the 'cpu' scheduling filter constraint |
2043 | * of events: |
2044 | */ |
2045 | if (!event_filter_match(event)) |
2046 | continue; |
2047 | |
2048 | /* may need to reset tstamp_enabled */ |
2049 | if (is_cgroup_event(event)) |
2050 | perf_cgroup_mark_enabled(event, ctx); |
2051 | |
2052 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
2053 | if (group_sched_in(event, cpuctx, ctx)) |
2054 | can_add_hw = 0; |
2055 | } |
2056 | } |
2057 | } |
2058 | |
2059 | static void |
2060 | ctx_sched_in(struct perf_event_context *ctx, |
2061 | struct perf_cpu_context *cpuctx, |
2062 | enum event_type_t event_type, |
2063 | struct task_struct *task) |
2064 | { |
2065 | u64 now; |
2066 | |
2067 | raw_spin_lock(&ctx->lock); |
2068 | ctx->is_active = 1; |
2069 | if (likely(!ctx->nr_events)) |
2070 | goto out; |
2071 | |
2072 | now = perf_clock(); |
2073 | ctx->timestamp = now; |
2074 | perf_cgroup_set_timestamp(task, ctx); |
2075 | /* |
2076 | * First go through the list and put on any pinned groups |
2077 | * in order to give them the best chance of going on. |
2078 | */ |
2079 | if (event_type & EVENT_PINNED) |
2080 | ctx_pinned_sched_in(ctx, cpuctx); |
2081 | |
2082 | /* Then walk through the lower prio flexible groups */ |
2083 | if (event_type & EVENT_FLEXIBLE) |
2084 | ctx_flexible_sched_in(ctx, cpuctx); |
2085 | |
2086 | out: |
2087 | raw_spin_unlock(&ctx->lock); |
2088 | } |
2089 | |
2090 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
2091 | enum event_type_t event_type, |
2092 | struct task_struct *task) |
2093 | { |
2094 | struct perf_event_context *ctx = &cpuctx->ctx; |
2095 | |
2096 | ctx_sched_in(ctx, cpuctx, event_type, task); |
2097 | } |
2098 | |
2099 | static void task_ctx_sched_in(struct perf_event_context *ctx, |
2100 | enum event_type_t event_type) |
2101 | { |
2102 | struct perf_cpu_context *cpuctx; |
2103 | |
2104 | cpuctx = __get_cpu_context(ctx); |
2105 | if (cpuctx->task_ctx == ctx) |
2106 | return; |
2107 | |
2108 | ctx_sched_in(ctx, cpuctx, event_type, NULL); |
2109 | cpuctx->task_ctx = ctx; |
2110 | } |
2111 | |
2112 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2113 | struct task_struct *task) |
2114 | { |
2115 | struct perf_cpu_context *cpuctx; |
2116 | |
2117 | cpuctx = __get_cpu_context(ctx); |
2118 | if (cpuctx->task_ctx == ctx) |
2119 | return; |
2120 | |
2121 | perf_pmu_disable(ctx->pmu); |
2122 | /* |
2123 | * We want to keep the following priority order: |
2124 | * cpu pinned (that don't need to move), task pinned, |
2125 | * cpu flexible, task flexible. |
2126 | */ |
2127 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2128 | |
2129 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); |
2130 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); |
2131 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); |
2132 | |
2133 | cpuctx->task_ctx = ctx; |
2134 | |
2135 | /* |
2136 | * Since these rotations are per-cpu, we need to ensure the |
2137 | * cpu-context we got scheduled on is actually rotating. |
2138 | */ |
2139 | perf_pmu_rotate_start(ctx->pmu); |
2140 | perf_pmu_enable(ctx->pmu); |
2141 | } |
2142 | |
2143 | /* |
2144 | * Called from scheduler to add the events of the current task |
2145 | * with interrupts disabled. |
2146 | * |
2147 | * We restore the event value and then enable it. |
2148 | * |
2149 | * This does not protect us against NMI, but enable() |
2150 | * sets the enabled bit in the control field of event _before_ |
2151 | * accessing the event control register. If a NMI hits, then it will |
2152 | * keep the event running. |
2153 | */ |
2154 | void __perf_event_task_sched_in(struct task_struct *task) |
2155 | { |
2156 | struct perf_event_context *ctx; |
2157 | int ctxn; |
2158 | |
2159 | for_each_task_context_nr(ctxn) { |
2160 | ctx = task->perf_event_ctxp[ctxn]; |
2161 | if (likely(!ctx)) |
2162 | continue; |
2163 | |
2164 | perf_event_context_sched_in(ctx, task); |
2165 | } |
2166 | /* |
2167 | * if cgroup events exist on this CPU, then we need |
2168 | * to check if we have to switch in PMU state. |
2169 | * cgroup event are system-wide mode only |
2170 | */ |
2171 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) |
2172 | perf_cgroup_sched_in(task); |
2173 | } |
2174 | |
2175 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2176 | { |
2177 | u64 frequency = event->attr.sample_freq; |
2178 | u64 sec = NSEC_PER_SEC; |
2179 | u64 divisor, dividend; |
2180 | |
2181 | int count_fls, nsec_fls, frequency_fls, sec_fls; |
2182 | |
2183 | count_fls = fls64(count); |
2184 | nsec_fls = fls64(nsec); |
2185 | frequency_fls = fls64(frequency); |
2186 | sec_fls = 30; |
2187 | |
2188 | /* |
2189 | * We got @count in @nsec, with a target of sample_freq HZ |
2190 | * the target period becomes: |
2191 | * |
2192 | * @count * 10^9 |
2193 | * period = ------------------- |
2194 | * @nsec * sample_freq |
2195 | * |
2196 | */ |
2197 | |
2198 | /* |
2199 | * Reduce accuracy by one bit such that @a and @b converge |
2200 | * to a similar magnitude. |
2201 | */ |
2202 | #define REDUCE_FLS(a, b) \ |
2203 | do { \ |
2204 | if (a##_fls > b##_fls) { \ |
2205 | a >>= 1; \ |
2206 | a##_fls--; \ |
2207 | } else { \ |
2208 | b >>= 1; \ |
2209 | b##_fls--; \ |
2210 | } \ |
2211 | } while (0) |
2212 | |
2213 | /* |
2214 | * Reduce accuracy until either term fits in a u64, then proceed with |
2215 | * the other, so that finally we can do a u64/u64 division. |
2216 | */ |
2217 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { |
2218 | REDUCE_FLS(nsec, frequency); |
2219 | REDUCE_FLS(sec, count); |
2220 | } |
2221 | |
2222 | if (count_fls + sec_fls > 64) { |
2223 | divisor = nsec * frequency; |
2224 | |
2225 | while (count_fls + sec_fls > 64) { |
2226 | REDUCE_FLS(count, sec); |
2227 | divisor >>= 1; |
2228 | } |
2229 | |
2230 | dividend = count * sec; |
2231 | } else { |
2232 | dividend = count * sec; |
2233 | |
2234 | while (nsec_fls + frequency_fls > 64) { |
2235 | REDUCE_FLS(nsec, frequency); |
2236 | dividend >>= 1; |
2237 | } |
2238 | |
2239 | divisor = nsec * frequency; |
2240 | } |
2241 | |
2242 | if (!divisor) |
2243 | return dividend; |
2244 | |
2245 | return div64_u64(dividend, divisor); |
2246 | } |
2247 | |
2248 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count) |
2249 | { |
2250 | struct hw_perf_event *hwc = &event->hw; |
2251 | s64 period, sample_period; |
2252 | s64 delta; |
2253 | |
2254 | period = perf_calculate_period(event, nsec, count); |
2255 | |
2256 | delta = (s64)(period - hwc->sample_period); |
2257 | delta = (delta + 7) / 8; /* low pass filter */ |
2258 | |
2259 | sample_period = hwc->sample_period + delta; |
2260 | |
2261 | if (!sample_period) |
2262 | sample_period = 1; |
2263 | |
2264 | hwc->sample_period = sample_period; |
2265 | |
2266 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
2267 | event->pmu->stop(event, PERF_EF_UPDATE); |
2268 | local64_set(&hwc->period_left, 0); |
2269 | event->pmu->start(event, PERF_EF_RELOAD); |
2270 | } |
2271 | } |
2272 | |
2273 | static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period) |
2274 | { |
2275 | struct perf_event *event; |
2276 | struct hw_perf_event *hwc; |
2277 | u64 interrupts, now; |
2278 | s64 delta; |
2279 | |
2280 | raw_spin_lock(&ctx->lock); |
2281 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
2282 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
2283 | continue; |
2284 | |
2285 | if (!event_filter_match(event)) |
2286 | continue; |
2287 | |
2288 | hwc = &event->hw; |
2289 | |
2290 | interrupts = hwc->interrupts; |
2291 | hwc->interrupts = 0; |
2292 | |
2293 | /* |
2294 | * unthrottle events on the tick |
2295 | */ |
2296 | if (interrupts == MAX_INTERRUPTS) { |
2297 | perf_log_throttle(event, 1); |
2298 | event->pmu->start(event, 0); |
2299 | } |
2300 | |
2301 | if (!event->attr.freq || !event->attr.sample_freq) |
2302 | continue; |
2303 | |
2304 | event->pmu->read(event); |
2305 | now = local64_read(&event->count); |
2306 | delta = now - hwc->freq_count_stamp; |
2307 | hwc->freq_count_stamp = now; |
2308 | |
2309 | if (delta > 0) |
2310 | perf_adjust_period(event, period, delta); |
2311 | } |
2312 | raw_spin_unlock(&ctx->lock); |
2313 | } |
2314 | |
2315 | /* |
2316 | * Round-robin a context's events: |
2317 | */ |
2318 | static void rotate_ctx(struct perf_event_context *ctx) |
2319 | { |
2320 | raw_spin_lock(&ctx->lock); |
2321 | |
2322 | /* |
2323 | * Rotate the first entry last of non-pinned groups. Rotation might be |
2324 | * disabled by the inheritance code. |
2325 | */ |
2326 | if (!ctx->rotate_disable) |
2327 | list_rotate_left(&ctx->flexible_groups); |
2328 | |
2329 | raw_spin_unlock(&ctx->lock); |
2330 | } |
2331 | |
2332 | /* |
2333 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2334 | * because they're strictly cpu affine and rotate_start is called with IRQs |
2335 | * disabled, while rotate_context is called from IRQ context. |
2336 | */ |
2337 | static void perf_rotate_context(struct perf_cpu_context *cpuctx) |
2338 | { |
2339 | u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC; |
2340 | struct perf_event_context *ctx = NULL; |
2341 | int rotate = 0, remove = 1; |
2342 | |
2343 | if (cpuctx->ctx.nr_events) { |
2344 | remove = 0; |
2345 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2346 | rotate = 1; |
2347 | } |
2348 | |
2349 | ctx = cpuctx->task_ctx; |
2350 | if (ctx && ctx->nr_events) { |
2351 | remove = 0; |
2352 | if (ctx->nr_events != ctx->nr_active) |
2353 | rotate = 1; |
2354 | } |
2355 | |
2356 | perf_pmu_disable(cpuctx->ctx.pmu); |
2357 | perf_ctx_adjust_freq(&cpuctx->ctx, interval); |
2358 | if (ctx) |
2359 | perf_ctx_adjust_freq(ctx, interval); |
2360 | |
2361 | if (!rotate) |
2362 | goto done; |
2363 | |
2364 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2365 | if (ctx) |
2366 | task_ctx_sched_out(ctx, EVENT_FLEXIBLE); |
2367 | |
2368 | rotate_ctx(&cpuctx->ctx); |
2369 | if (ctx) |
2370 | rotate_ctx(ctx); |
2371 | |
2372 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, current); |
2373 | if (ctx) |
2374 | task_ctx_sched_in(ctx, EVENT_FLEXIBLE); |
2375 | |
2376 | done: |
2377 | if (remove) |
2378 | list_del_init(&cpuctx->rotation_list); |
2379 | |
2380 | perf_pmu_enable(cpuctx->ctx.pmu); |
2381 | } |
2382 | |
2383 | void perf_event_task_tick(void) |
2384 | { |
2385 | struct list_head *head = &__get_cpu_var(rotation_list); |
2386 | struct perf_cpu_context *cpuctx, *tmp; |
2387 | |
2388 | WARN_ON(!irqs_disabled()); |
2389 | |
2390 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
2391 | if (cpuctx->jiffies_interval == 1 || |
2392 | !(jiffies % cpuctx->jiffies_interval)) |
2393 | perf_rotate_context(cpuctx); |
2394 | } |
2395 | } |
2396 | |
2397 | static int event_enable_on_exec(struct perf_event *event, |
2398 | struct perf_event_context *ctx) |
2399 | { |
2400 | if (!event->attr.enable_on_exec) |
2401 | return 0; |
2402 | |
2403 | event->attr.enable_on_exec = 0; |
2404 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
2405 | return 0; |
2406 | |
2407 | __perf_event_mark_enabled(event, ctx); |
2408 | |
2409 | return 1; |
2410 | } |
2411 | |
2412 | /* |
2413 | * Enable all of a task's events that have been marked enable-on-exec. |
2414 | * This expects task == current. |
2415 | */ |
2416 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
2417 | { |
2418 | struct perf_event *event; |
2419 | unsigned long flags; |
2420 | int enabled = 0; |
2421 | int ret; |
2422 | |
2423 | local_irq_save(flags); |
2424 | if (!ctx || !ctx->nr_events) |
2425 | goto out; |
2426 | |
2427 | /* |
2428 | * We must ctxsw out cgroup events to avoid conflict |
2429 | * when invoking perf_task_event_sched_in() later on |
2430 | * in this function. Otherwise we end up trying to |
2431 | * ctxswin cgroup events which are already scheduled |
2432 | * in. |
2433 | */ |
2434 | perf_cgroup_sched_out(current); |
2435 | task_ctx_sched_out(ctx, EVENT_ALL); |
2436 | |
2437 | raw_spin_lock(&ctx->lock); |
2438 | |
2439 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2440 | ret = event_enable_on_exec(event, ctx); |
2441 | if (ret) |
2442 | enabled = 1; |
2443 | } |
2444 | |
2445 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2446 | ret = event_enable_on_exec(event, ctx); |
2447 | if (ret) |
2448 | enabled = 1; |
2449 | } |
2450 | |
2451 | /* |
2452 | * Unclone this context if we enabled any event. |
2453 | */ |
2454 | if (enabled) |
2455 | unclone_ctx(ctx); |
2456 | |
2457 | raw_spin_unlock(&ctx->lock); |
2458 | |
2459 | /* |
2460 | * Also calls ctxswin for cgroup events, if any: |
2461 | */ |
2462 | perf_event_context_sched_in(ctx, ctx->task); |
2463 | out: |
2464 | local_irq_restore(flags); |
2465 | } |
2466 | |
2467 | /* |
2468 | * Cross CPU call to read the hardware event |
2469 | */ |
2470 | static void __perf_event_read(void *info) |
2471 | { |
2472 | struct perf_event *event = info; |
2473 | struct perf_event_context *ctx = event->ctx; |
2474 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2475 | |
2476 | /* |
2477 | * If this is a task context, we need to check whether it is |
2478 | * the current task context of this cpu. If not it has been |
2479 | * scheduled out before the smp call arrived. In that case |
2480 | * event->count would have been updated to a recent sample |
2481 | * when the event was scheduled out. |
2482 | */ |
2483 | if (ctx->task && cpuctx->task_ctx != ctx) |
2484 | return; |
2485 | |
2486 | raw_spin_lock(&ctx->lock); |
2487 | if (ctx->is_active) { |
2488 | update_context_time(ctx); |
2489 | update_cgrp_time_from_event(event); |
2490 | } |
2491 | update_event_times(event); |
2492 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2493 | event->pmu->read(event); |
2494 | raw_spin_unlock(&ctx->lock); |
2495 | } |
2496 | |
2497 | static inline u64 perf_event_count(struct perf_event *event) |
2498 | { |
2499 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
2500 | } |
2501 | |
2502 | static u64 perf_event_read(struct perf_event *event) |
2503 | { |
2504 | /* |
2505 | * If event is enabled and currently active on a CPU, update the |
2506 | * value in the event structure: |
2507 | */ |
2508 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2509 | smp_call_function_single(event->oncpu, |
2510 | __perf_event_read, event, 1); |
2511 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2512 | struct perf_event_context *ctx = event->ctx; |
2513 | unsigned long flags; |
2514 | |
2515 | raw_spin_lock_irqsave(&ctx->lock, flags); |
2516 | /* |
2517 | * may read while context is not active |
2518 | * (e.g., thread is blocked), in that case |
2519 | * we cannot update context time |
2520 | */ |
2521 | if (ctx->is_active) { |
2522 | update_context_time(ctx); |
2523 | update_cgrp_time_from_event(event); |
2524 | } |
2525 | update_event_times(event); |
2526 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
2527 | } |
2528 | |
2529 | return perf_event_count(event); |
2530 | } |
2531 | |
2532 | /* |
2533 | * Callchain support |
2534 | */ |
2535 | |
2536 | struct callchain_cpus_entries { |
2537 | struct rcu_head rcu_head; |
2538 | struct perf_callchain_entry *cpu_entries[0]; |
2539 | }; |
2540 | |
2541 | static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]); |
2542 | static atomic_t nr_callchain_events; |
2543 | static DEFINE_MUTEX(callchain_mutex); |
2544 | struct callchain_cpus_entries *callchain_cpus_entries; |
2545 | |
2546 | |
2547 | __weak void perf_callchain_kernel(struct perf_callchain_entry *entry, |
2548 | struct pt_regs *regs) |
2549 | { |
2550 | } |
2551 | |
2552 | __weak void perf_callchain_user(struct perf_callchain_entry *entry, |
2553 | struct pt_regs *regs) |
2554 | { |
2555 | } |
2556 | |
2557 | static void release_callchain_buffers_rcu(struct rcu_head *head) |
2558 | { |
2559 | struct callchain_cpus_entries *entries; |
2560 | int cpu; |
2561 | |
2562 | entries = container_of(head, struct callchain_cpus_entries, rcu_head); |
2563 | |
2564 | for_each_possible_cpu(cpu) |
2565 | kfree(entries->cpu_entries[cpu]); |
2566 | |
2567 | kfree(entries); |
2568 | } |
2569 | |
2570 | static void release_callchain_buffers(void) |
2571 | { |
2572 | struct callchain_cpus_entries *entries; |
2573 | |
2574 | entries = callchain_cpus_entries; |
2575 | rcu_assign_pointer(callchain_cpus_entries, NULL); |
2576 | call_rcu(&entries->rcu_head, release_callchain_buffers_rcu); |
2577 | } |
2578 | |
2579 | static int alloc_callchain_buffers(void) |
2580 | { |
2581 | int cpu; |
2582 | int size; |
2583 | struct callchain_cpus_entries *entries; |
2584 | |
2585 | /* |
2586 | * We can't use the percpu allocation API for data that can be |
2587 | * accessed from NMI. Use a temporary manual per cpu allocation |
2588 | * until that gets sorted out. |
2589 | */ |
2590 | size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]); |
2591 | |
2592 | entries = kzalloc(size, GFP_KERNEL); |
2593 | if (!entries) |
2594 | return -ENOMEM; |
2595 | |
2596 | size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS; |
2597 | |
2598 | for_each_possible_cpu(cpu) { |
2599 | entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL, |
2600 | cpu_to_node(cpu)); |
2601 | if (!entries->cpu_entries[cpu]) |
2602 | goto fail; |
2603 | } |
2604 | |
2605 | rcu_assign_pointer(callchain_cpus_entries, entries); |
2606 | |
2607 | return 0; |
2608 | |
2609 | fail: |
2610 | for_each_possible_cpu(cpu) |
2611 | kfree(entries->cpu_entries[cpu]); |
2612 | kfree(entries); |
2613 | |
2614 | return -ENOMEM; |
2615 | } |
2616 | |
2617 | static int get_callchain_buffers(void) |
2618 | { |
2619 | int err = 0; |
2620 | int count; |
2621 | |
2622 | mutex_lock(&callchain_mutex); |
2623 | |
2624 | count = atomic_inc_return(&nr_callchain_events); |
2625 | if (WARN_ON_ONCE(count < 1)) { |
2626 | err = -EINVAL; |
2627 | goto exit; |
2628 | } |
2629 | |
2630 | if (count > 1) { |
2631 | /* If the allocation failed, give up */ |
2632 | if (!callchain_cpus_entries) |
2633 | err = -ENOMEM; |
2634 | goto exit; |
2635 | } |
2636 | |
2637 | err = alloc_callchain_buffers(); |
2638 | if (err) |
2639 | release_callchain_buffers(); |
2640 | exit: |
2641 | mutex_unlock(&callchain_mutex); |
2642 | |
2643 | return err; |
2644 | } |
2645 | |
2646 | static void put_callchain_buffers(void) |
2647 | { |
2648 | if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) { |
2649 | release_callchain_buffers(); |
2650 | mutex_unlock(&callchain_mutex); |
2651 | } |
2652 | } |
2653 | |
2654 | static int get_recursion_context(int *recursion) |
2655 | { |
2656 | int rctx; |
2657 | |
2658 | if (in_nmi()) |
2659 | rctx = 3; |
2660 | else if (in_irq()) |
2661 | rctx = 2; |
2662 | else if (in_softirq()) |
2663 | rctx = 1; |
2664 | else |
2665 | rctx = 0; |
2666 | |
2667 | if (recursion[rctx]) |
2668 | return -1; |
2669 | |
2670 | recursion[rctx]++; |
2671 | barrier(); |
2672 | |
2673 | return rctx; |
2674 | } |
2675 | |
2676 | static inline void put_recursion_context(int *recursion, int rctx) |
2677 | { |
2678 | barrier(); |
2679 | recursion[rctx]--; |
2680 | } |
2681 | |
2682 | static struct perf_callchain_entry *get_callchain_entry(int *rctx) |
2683 | { |
2684 | int cpu; |
2685 | struct callchain_cpus_entries *entries; |
2686 | |
2687 | *rctx = get_recursion_context(__get_cpu_var(callchain_recursion)); |
2688 | if (*rctx == -1) |
2689 | return NULL; |
2690 | |
2691 | entries = rcu_dereference(callchain_cpus_entries); |
2692 | if (!entries) |
2693 | return NULL; |
2694 | |
2695 | cpu = smp_processor_id(); |
2696 | |
2697 | return &entries->cpu_entries[cpu][*rctx]; |
2698 | } |
2699 | |
2700 | static void |
2701 | put_callchain_entry(int rctx) |
2702 | { |
2703 | put_recursion_context(__get_cpu_var(callchain_recursion), rctx); |
2704 | } |
2705 | |
2706 | static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs) |
2707 | { |
2708 | int rctx; |
2709 | struct perf_callchain_entry *entry; |
2710 | |
2711 | |
2712 | entry = get_callchain_entry(&rctx); |
2713 | if (rctx == -1) |
2714 | return NULL; |
2715 | |
2716 | if (!entry) |
2717 | goto exit_put; |
2718 | |
2719 | entry->nr = 0; |
2720 | |
2721 | if (!user_mode(regs)) { |
2722 | perf_callchain_store(entry, PERF_CONTEXT_KERNEL); |
2723 | perf_callchain_kernel(entry, regs); |
2724 | if (current->mm) |
2725 | regs = task_pt_regs(current); |
2726 | else |
2727 | regs = NULL; |
2728 | } |
2729 | |
2730 | if (regs) { |
2731 | perf_callchain_store(entry, PERF_CONTEXT_USER); |
2732 | perf_callchain_user(entry, regs); |
2733 | } |
2734 | |
2735 | exit_put: |
2736 | put_callchain_entry(rctx); |
2737 | |
2738 | return entry; |
2739 | } |
2740 | |
2741 | /* |
2742 | * Initialize the perf_event context in a task_struct: |
2743 | */ |
2744 | static void __perf_event_init_context(struct perf_event_context *ctx) |
2745 | { |
2746 | raw_spin_lock_init(&ctx->lock); |
2747 | mutex_init(&ctx->mutex); |
2748 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2749 | INIT_LIST_HEAD(&ctx->flexible_groups); |
2750 | INIT_LIST_HEAD(&ctx->event_list); |
2751 | atomic_set(&ctx->refcount, 1); |
2752 | } |
2753 | |
2754 | static struct perf_event_context * |
2755 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) |
2756 | { |
2757 | struct perf_event_context *ctx; |
2758 | |
2759 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); |
2760 | if (!ctx) |
2761 | return NULL; |
2762 | |
2763 | __perf_event_init_context(ctx); |
2764 | if (task) { |
2765 | ctx->task = task; |
2766 | get_task_struct(task); |
2767 | } |
2768 | ctx->pmu = pmu; |
2769 | |
2770 | return ctx; |
2771 | } |
2772 | |
2773 | static struct task_struct * |
2774 | find_lively_task_by_vpid(pid_t vpid) |
2775 | { |
2776 | struct task_struct *task; |
2777 | int err; |
2778 | |
2779 | rcu_read_lock(); |
2780 | if (!vpid) |
2781 | task = current; |
2782 | else |
2783 | task = find_task_by_vpid(vpid); |
2784 | if (task) |
2785 | get_task_struct(task); |
2786 | rcu_read_unlock(); |
2787 | |
2788 | if (!task) |
2789 | return ERR_PTR(-ESRCH); |
2790 | |
2791 | /* Reuse ptrace permission checks for now. */ |
2792 | err = -EACCES; |
2793 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) |
2794 | goto errout; |
2795 | |
2796 | return task; |
2797 | errout: |
2798 | put_task_struct(task); |
2799 | return ERR_PTR(err); |
2800 | |
2801 | } |
2802 | |
2803 | /* |
2804 | * Returns a matching context with refcount and pincount. |
2805 | */ |
2806 | static struct perf_event_context * |
2807 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
2808 | { |
2809 | struct perf_event_context *ctx; |
2810 | struct perf_cpu_context *cpuctx; |
2811 | unsigned long flags; |
2812 | int ctxn, err; |
2813 | |
2814 | if (!task) { |
2815 | /* Must be root to operate on a CPU event: */ |
2816 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
2817 | return ERR_PTR(-EACCES); |
2818 | |
2819 | /* |
2820 | * We could be clever and allow to attach a event to an |
2821 | * offline CPU and activate it when the CPU comes up, but |
2822 | * that's for later. |
2823 | */ |
2824 | if (!cpu_online(cpu)) |
2825 | return ERR_PTR(-ENODEV); |
2826 | |
2827 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
2828 | ctx = &cpuctx->ctx; |
2829 | get_ctx(ctx); |
2830 | ++ctx->pin_count; |
2831 | |
2832 | return ctx; |
2833 | } |
2834 | |
2835 | err = -EINVAL; |
2836 | ctxn = pmu->task_ctx_nr; |
2837 | if (ctxn < 0) |
2838 | goto errout; |
2839 | |
2840 | retry: |
2841 | ctx = perf_lock_task_context(task, ctxn, &flags); |
2842 | if (ctx) { |
2843 | unclone_ctx(ctx); |
2844 | ++ctx->pin_count; |
2845 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
2846 | } |
2847 | |
2848 | if (!ctx) { |
2849 | ctx = alloc_perf_context(pmu, task); |
2850 | err = -ENOMEM; |
2851 | if (!ctx) |
2852 | goto errout; |
2853 | |
2854 | get_ctx(ctx); |
2855 | |
2856 | err = 0; |
2857 | mutex_lock(&task->perf_event_mutex); |
2858 | /* |
2859 | * If it has already passed perf_event_exit_task(). |
2860 | * we must see PF_EXITING, it takes this mutex too. |
2861 | */ |
2862 | if (task->flags & PF_EXITING) |
2863 | err = -ESRCH; |
2864 | else if (task->perf_event_ctxp[ctxn]) |
2865 | err = -EAGAIN; |
2866 | else { |
2867 | ++ctx->pin_count; |
2868 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
2869 | } |
2870 | mutex_unlock(&task->perf_event_mutex); |
2871 | |
2872 | if (unlikely(err)) { |
2873 | put_task_struct(task); |
2874 | kfree(ctx); |
2875 | |
2876 | if (err == -EAGAIN) |
2877 | goto retry; |
2878 | goto errout; |
2879 | } |
2880 | } |
2881 | |
2882 | return ctx; |
2883 | |
2884 | errout: |
2885 | return ERR_PTR(err); |
2886 | } |
2887 | |
2888 | static void perf_event_free_filter(struct perf_event *event); |
2889 | |
2890 | static void free_event_rcu(struct rcu_head *head) |
2891 | { |
2892 | struct perf_event *event; |
2893 | |
2894 | event = container_of(head, struct perf_event, rcu_head); |
2895 | if (event->ns) |
2896 | put_pid_ns(event->ns); |
2897 | perf_event_free_filter(event); |
2898 | kfree(event); |
2899 | } |
2900 | |
2901 | static void perf_buffer_put(struct perf_buffer *buffer); |
2902 | |
2903 | static void free_event(struct perf_event *event) |
2904 | { |
2905 | irq_work_sync(&event->pending); |
2906 | |
2907 | if (!event->parent) { |
2908 | if (event->attach_state & PERF_ATTACH_TASK) |
2909 | jump_label_dec(&perf_sched_events); |
2910 | if (event->attr.mmap || event->attr.mmap_data) |
2911 | atomic_dec(&nr_mmap_events); |
2912 | if (event->attr.comm) |
2913 | atomic_dec(&nr_comm_events); |
2914 | if (event->attr.task) |
2915 | atomic_dec(&nr_task_events); |
2916 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
2917 | put_callchain_buffers(); |
2918 | if (is_cgroup_event(event)) { |
2919 | atomic_dec(&per_cpu(perf_cgroup_events, event->cpu)); |
2920 | jump_label_dec(&perf_sched_events); |
2921 | } |
2922 | } |
2923 | |
2924 | if (event->buffer) { |
2925 | perf_buffer_put(event->buffer); |
2926 | event->buffer = NULL; |
2927 | } |
2928 | |
2929 | if (is_cgroup_event(event)) |
2930 | perf_detach_cgroup(event); |
2931 | |
2932 | if (event->destroy) |
2933 | event->destroy(event); |
2934 | |
2935 | if (event->ctx) |
2936 | put_ctx(event->ctx); |
2937 | |
2938 | call_rcu(&event->rcu_head, free_event_rcu); |
2939 | } |
2940 | |
2941 | int perf_event_release_kernel(struct perf_event *event) |
2942 | { |
2943 | struct perf_event_context *ctx = event->ctx; |
2944 | |
2945 | /* |
2946 | * Remove from the PMU, can't get re-enabled since we got |
2947 | * here because the last ref went. |
2948 | */ |
2949 | perf_event_disable(event); |
2950 | |
2951 | WARN_ON_ONCE(ctx->parent_ctx); |
2952 | /* |
2953 | * There are two ways this annotation is useful: |
2954 | * |
2955 | * 1) there is a lock recursion from perf_event_exit_task |
2956 | * see the comment there. |
2957 | * |
2958 | * 2) there is a lock-inversion with mmap_sem through |
2959 | * perf_event_read_group(), which takes faults while |
2960 | * holding ctx->mutex, however this is called after |
2961 | * the last filedesc died, so there is no possibility |
2962 | * to trigger the AB-BA case. |
2963 | */ |
2964 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); |
2965 | raw_spin_lock_irq(&ctx->lock); |
2966 | perf_group_detach(event); |
2967 | list_del_event(event, ctx); |
2968 | raw_spin_unlock_irq(&ctx->lock); |
2969 | mutex_unlock(&ctx->mutex); |
2970 | |
2971 | free_event(event); |
2972 | |
2973 | return 0; |
2974 | } |
2975 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
2976 | |
2977 | /* |
2978 | * Called when the last reference to the file is gone. |
2979 | */ |
2980 | static int perf_release(struct inode *inode, struct file *file) |
2981 | { |
2982 | struct perf_event *event = file->private_data; |
2983 | struct task_struct *owner; |
2984 | |
2985 | file->private_data = NULL; |
2986 | |
2987 | rcu_read_lock(); |
2988 | owner = ACCESS_ONCE(event->owner); |
2989 | /* |
2990 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe |
2991 | * !owner it means the list deletion is complete and we can indeed |
2992 | * free this event, otherwise we need to serialize on |
2993 | * owner->perf_event_mutex. |
2994 | */ |
2995 | smp_read_barrier_depends(); |
2996 | if (owner) { |
2997 | /* |
2998 | * Since delayed_put_task_struct() also drops the last |
2999 | * task reference we can safely take a new reference |
3000 | * while holding the rcu_read_lock(). |
3001 | */ |
3002 | get_task_struct(owner); |
3003 | } |
3004 | rcu_read_unlock(); |
3005 | |
3006 | if (owner) { |
3007 | mutex_lock(&owner->perf_event_mutex); |
3008 | /* |
3009 | * We have to re-check the event->owner field, if it is cleared |
3010 | * we raced with perf_event_exit_task(), acquiring the mutex |
3011 | * ensured they're done, and we can proceed with freeing the |
3012 | * event. |
3013 | */ |
3014 | if (event->owner) |
3015 | list_del_init(&event->owner_entry); |
3016 | mutex_unlock(&owner->perf_event_mutex); |
3017 | put_task_struct(owner); |
3018 | } |
3019 | |
3020 | return perf_event_release_kernel(event); |
3021 | } |
3022 | |
3023 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
3024 | { |
3025 | struct perf_event *child; |
3026 | u64 total = 0; |
3027 | |
3028 | *enabled = 0; |
3029 | *running = 0; |
3030 | |
3031 | mutex_lock(&event->child_mutex); |
3032 | total += perf_event_read(event); |
3033 | *enabled += event->total_time_enabled + |
3034 | atomic64_read(&event->child_total_time_enabled); |
3035 | *running += event->total_time_running + |
3036 | atomic64_read(&event->child_total_time_running); |
3037 | |
3038 | list_for_each_entry(child, &event->child_list, child_list) { |
3039 | total += perf_event_read(child); |
3040 | *enabled += child->total_time_enabled; |
3041 | *running += child->total_time_running; |
3042 | } |
3043 | mutex_unlock(&event->child_mutex); |
3044 | |
3045 | return total; |
3046 | } |
3047 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
3048 | |
3049 | static int perf_event_read_group(struct perf_event *event, |
3050 | u64 read_format, char __user *buf) |
3051 | { |
3052 | struct perf_event *leader = event->group_leader, *sub; |
3053 | int n = 0, size = 0, ret = -EFAULT; |
3054 | struct perf_event_context *ctx = leader->ctx; |
3055 | u64 values[5]; |
3056 | u64 count, enabled, running; |
3057 | |
3058 | mutex_lock(&ctx->mutex); |
3059 | count = perf_event_read_value(leader, &enabled, &running); |
3060 | |
3061 | values[n++] = 1 + leader->nr_siblings; |
3062 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3063 | values[n++] = enabled; |
3064 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
3065 | values[n++] = running; |
3066 | values[n++] = count; |
3067 | if (read_format & PERF_FORMAT_ID) |
3068 | values[n++] = primary_event_id(leader); |
3069 | |
3070 | size = n * sizeof(u64); |
3071 | |
3072 | if (copy_to_user(buf, values, size)) |
3073 | goto unlock; |
3074 | |
3075 | ret = size; |
3076 | |
3077 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3078 | n = 0; |
3079 | |
3080 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
3081 | if (read_format & PERF_FORMAT_ID) |
3082 | values[n++] = primary_event_id(sub); |
3083 | |
3084 | size = n * sizeof(u64); |
3085 | |
3086 | if (copy_to_user(buf + ret, values, size)) { |
3087 | ret = -EFAULT; |
3088 | goto unlock; |
3089 | } |
3090 | |
3091 | ret += size; |
3092 | } |
3093 | unlock: |
3094 | mutex_unlock(&ctx->mutex); |
3095 | |
3096 | return ret; |
3097 | } |
3098 | |
3099 | static int perf_event_read_one(struct perf_event *event, |
3100 | u64 read_format, char __user *buf) |
3101 | { |
3102 | u64 enabled, running; |
3103 | u64 values[4]; |
3104 | int n = 0; |
3105 | |
3106 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3107 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3108 | values[n++] = enabled; |
3109 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
3110 | values[n++] = running; |
3111 | if (read_format & PERF_FORMAT_ID) |
3112 | values[n++] = primary_event_id(event); |
3113 | |
3114 | if (copy_to_user(buf, values, n * sizeof(u64))) |
3115 | return -EFAULT; |
3116 | |
3117 | return n * sizeof(u64); |
3118 | } |
3119 | |
3120 | /* |
3121 | * Read the performance event - simple non blocking version for now |
3122 | */ |
3123 | static ssize_t |
3124 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
3125 | { |
3126 | u64 read_format = event->attr.read_format; |
3127 | int ret; |
3128 | |
3129 | /* |
3130 | * Return end-of-file for a read on a event that is in |
3131 | * error state (i.e. because it was pinned but it couldn't be |
3132 | * scheduled on to the CPU at some point). |
3133 | */ |
3134 | if (event->state == PERF_EVENT_STATE_ERROR) |
3135 | return 0; |
3136 | |
3137 | if (count < event->read_size) |
3138 | return -ENOSPC; |
3139 | |
3140 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3141 | if (read_format & PERF_FORMAT_GROUP) |
3142 | ret = perf_event_read_group(event, read_format, buf); |
3143 | else |
3144 | ret = perf_event_read_one(event, read_format, buf); |
3145 | |
3146 | return ret; |
3147 | } |
3148 | |
3149 | static ssize_t |
3150 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) |
3151 | { |
3152 | struct perf_event *event = file->private_data; |
3153 | |
3154 | return perf_read_hw(event, buf, count); |
3155 | } |
3156 | |
3157 | static unsigned int perf_poll(struct file *file, poll_table *wait) |
3158 | { |
3159 | struct perf_event *event = file->private_data; |
3160 | struct perf_buffer *buffer; |
3161 | unsigned int events = POLL_HUP; |
3162 | |
3163 | rcu_read_lock(); |
3164 | buffer = rcu_dereference(event->buffer); |
3165 | if (buffer) |
3166 | events = atomic_xchg(&buffer->poll, 0); |
3167 | rcu_read_unlock(); |
3168 | |
3169 | poll_wait(file, &event->waitq, wait); |
3170 | |
3171 | return events; |
3172 | } |
3173 | |
3174 | static void perf_event_reset(struct perf_event *event) |
3175 | { |
3176 | (void)perf_event_read(event); |
3177 | local64_set(&event->count, 0); |
3178 | perf_event_update_userpage(event); |
3179 | } |
3180 | |
3181 | /* |
3182 | * Holding the top-level event's child_mutex means that any |
3183 | * descendant process that has inherited this event will block |
3184 | * in sync_child_event if it goes to exit, thus satisfying the |
3185 | * task existence requirements of perf_event_enable/disable. |
3186 | */ |
3187 | static void perf_event_for_each_child(struct perf_event *event, |
3188 | void (*func)(struct perf_event *)) |
3189 | { |
3190 | struct perf_event *child; |
3191 | |
3192 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3193 | mutex_lock(&event->child_mutex); |
3194 | func(event); |
3195 | list_for_each_entry(child, &event->child_list, child_list) |
3196 | func(child); |
3197 | mutex_unlock(&event->child_mutex); |
3198 | } |
3199 | |
3200 | static void perf_event_for_each(struct perf_event *event, |
3201 | void (*func)(struct perf_event *)) |
3202 | { |
3203 | struct perf_event_context *ctx = event->ctx; |
3204 | struct perf_event *sibling; |
3205 | |
3206 | WARN_ON_ONCE(ctx->parent_ctx); |
3207 | mutex_lock(&ctx->mutex); |
3208 | event = event->group_leader; |
3209 | |
3210 | perf_event_for_each_child(event, func); |
3211 | func(event); |
3212 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
3213 | perf_event_for_each_child(event, func); |
3214 | mutex_unlock(&ctx->mutex); |
3215 | } |
3216 | |
3217 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
3218 | { |
3219 | struct perf_event_context *ctx = event->ctx; |
3220 | int ret = 0; |
3221 | u64 value; |
3222 | |
3223 | if (!is_sampling_event(event)) |
3224 | return -EINVAL; |
3225 | |
3226 | if (copy_from_user(&value, arg, sizeof(value))) |
3227 | return -EFAULT; |
3228 | |
3229 | if (!value) |
3230 | return -EINVAL; |
3231 | |
3232 | raw_spin_lock_irq(&ctx->lock); |
3233 | if (event->attr.freq) { |
3234 | if (value > sysctl_perf_event_sample_rate) { |
3235 | ret = -EINVAL; |
3236 | goto unlock; |
3237 | } |
3238 | |
3239 | event->attr.sample_freq = value; |
3240 | } else { |
3241 | event->attr.sample_period = value; |
3242 | event->hw.sample_period = value; |
3243 | } |
3244 | unlock: |
3245 | raw_spin_unlock_irq(&ctx->lock); |
3246 | |
3247 | return ret; |
3248 | } |
3249 | |
3250 | static const struct file_operations perf_fops; |
3251 | |
3252 | static struct perf_event *perf_fget_light(int fd, int *fput_needed) |
3253 | { |
3254 | struct file *file; |
3255 | |
3256 | file = fget_light(fd, fput_needed); |
3257 | if (!file) |
3258 | return ERR_PTR(-EBADF); |
3259 | |
3260 | if (file->f_op != &perf_fops) { |
3261 | fput_light(file, *fput_needed); |
3262 | *fput_needed = 0; |
3263 | return ERR_PTR(-EBADF); |
3264 | } |
3265 | |
3266 | return file->private_data; |
3267 | } |
3268 | |
3269 | static int perf_event_set_output(struct perf_event *event, |
3270 | struct perf_event *output_event); |
3271 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
3272 | |
3273 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3274 | { |
3275 | struct perf_event *event = file->private_data; |
3276 | void (*func)(struct perf_event *); |
3277 | u32 flags = arg; |
3278 | |
3279 | switch (cmd) { |
3280 | case PERF_EVENT_IOC_ENABLE: |
3281 | func = perf_event_enable; |
3282 | break; |
3283 | case PERF_EVENT_IOC_DISABLE: |
3284 | func = perf_event_disable; |
3285 | break; |
3286 | case PERF_EVENT_IOC_RESET: |
3287 | func = perf_event_reset; |
3288 | break; |
3289 | |
3290 | case PERF_EVENT_IOC_REFRESH: |
3291 | return perf_event_refresh(event, arg); |
3292 | |
3293 | case PERF_EVENT_IOC_PERIOD: |
3294 | return perf_event_period(event, (u64 __user *)arg); |
3295 | |
3296 | case PERF_EVENT_IOC_SET_OUTPUT: |
3297 | { |
3298 | struct perf_event *output_event = NULL; |
3299 | int fput_needed = 0; |
3300 | int ret; |
3301 | |
3302 | if (arg != -1) { |
3303 | output_event = perf_fget_light(arg, &fput_needed); |
3304 | if (IS_ERR(output_event)) |
3305 | return PTR_ERR(output_event); |
3306 | } |
3307 | |
3308 | ret = perf_event_set_output(event, output_event); |
3309 | if (output_event) |
3310 | fput_light(output_event->filp, fput_needed); |
3311 | |
3312 | return ret; |
3313 | } |
3314 | |
3315 | case PERF_EVENT_IOC_SET_FILTER: |
3316 | return perf_event_set_filter(event, (void __user *)arg); |
3317 | |
3318 | default: |
3319 | return -ENOTTY; |
3320 | } |
3321 | |
3322 | if (flags & PERF_IOC_FLAG_GROUP) |
3323 | perf_event_for_each(event, func); |
3324 | else |
3325 | perf_event_for_each_child(event, func); |
3326 | |
3327 | return 0; |
3328 | } |
3329 | |
3330 | int perf_event_task_enable(void) |
3331 | { |
3332 | struct perf_event *event; |
3333 | |
3334 | mutex_lock(¤t->perf_event_mutex); |
3335 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) |
3336 | perf_event_for_each_child(event, perf_event_enable); |
3337 | mutex_unlock(¤t->perf_event_mutex); |
3338 | |
3339 | return 0; |
3340 | } |
3341 | |
3342 | int perf_event_task_disable(void) |
3343 | { |
3344 | struct perf_event *event; |
3345 | |
3346 | mutex_lock(¤t->perf_event_mutex); |
3347 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) |
3348 | perf_event_for_each_child(event, perf_event_disable); |
3349 | mutex_unlock(¤t->perf_event_mutex); |
3350 | |
3351 | return 0; |
3352 | } |
3353 | |
3354 | #ifndef PERF_EVENT_INDEX_OFFSET |
3355 | # define PERF_EVENT_INDEX_OFFSET 0 |
3356 | #endif |
3357 | |
3358 | static int perf_event_index(struct perf_event *event) |
3359 | { |
3360 | if (event->hw.state & PERF_HES_STOPPED) |
3361 | return 0; |
3362 | |
3363 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
3364 | return 0; |
3365 | |
3366 | return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET; |
3367 | } |
3368 | |
3369 | /* |
3370 | * Callers need to ensure there can be no nesting of this function, otherwise |
3371 | * the seqlock logic goes bad. We can not serialize this because the arch |
3372 | * code calls this from NMI context. |
3373 | */ |
3374 | void perf_event_update_userpage(struct perf_event *event) |
3375 | { |
3376 | struct perf_event_mmap_page *userpg; |
3377 | struct perf_buffer *buffer; |
3378 | |
3379 | rcu_read_lock(); |
3380 | buffer = rcu_dereference(event->buffer); |
3381 | if (!buffer) |
3382 | goto unlock; |
3383 | |
3384 | userpg = buffer->user_page; |
3385 | |
3386 | /* |
3387 | * Disable preemption so as to not let the corresponding user-space |
3388 | * spin too long if we get preempted. |
3389 | */ |
3390 | preempt_disable(); |
3391 | ++userpg->lock; |
3392 | barrier(); |
3393 | userpg->index = perf_event_index(event); |
3394 | userpg->offset = perf_event_count(event); |
3395 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
3396 | userpg->offset -= local64_read(&event->hw.prev_count); |
3397 | |
3398 | userpg->time_enabled = event->total_time_enabled + |
3399 | atomic64_read(&event->child_total_time_enabled); |
3400 | |
3401 | userpg->time_running = event->total_time_running + |
3402 | atomic64_read(&event->child_total_time_running); |
3403 | |
3404 | barrier(); |
3405 | ++userpg->lock; |
3406 | preempt_enable(); |
3407 | unlock: |
3408 | rcu_read_unlock(); |
3409 | } |
3410 | |
3411 | static unsigned long perf_data_size(struct perf_buffer *buffer); |
3412 | |
3413 | static void |
3414 | perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags) |
3415 | { |
3416 | long max_size = perf_data_size(buffer); |
3417 | |
3418 | if (watermark) |
3419 | buffer->watermark = min(max_size, watermark); |
3420 | |
3421 | if (!buffer->watermark) |
3422 | buffer->watermark = max_size / 2; |
3423 | |
3424 | if (flags & PERF_BUFFER_WRITABLE) |
3425 | buffer->writable = 1; |
3426 | |
3427 | atomic_set(&buffer->refcount, 1); |
3428 | } |
3429 | |
3430 | #ifndef CONFIG_PERF_USE_VMALLOC |
3431 | |
3432 | /* |
3433 | * Back perf_mmap() with regular GFP_KERNEL-0 pages. |
3434 | */ |
3435 | |
3436 | static struct page * |
3437 | perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff) |
3438 | { |
3439 | if (pgoff > buffer->nr_pages) |
3440 | return NULL; |
3441 | |
3442 | if (pgoff == 0) |
3443 | return virt_to_page(buffer->user_page); |
3444 | |
3445 | return virt_to_page(buffer->data_pages[pgoff - 1]); |
3446 | } |
3447 | |
3448 | static void *perf_mmap_alloc_page(int cpu) |
3449 | { |
3450 | struct page *page; |
3451 | int node; |
3452 | |
3453 | node = (cpu == -1) ? cpu : cpu_to_node(cpu); |
3454 | page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0); |
3455 | if (!page) |
3456 | return NULL; |
3457 | |
3458 | return page_address(page); |
3459 | } |
3460 | |
3461 | static struct perf_buffer * |
3462 | perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags) |
3463 | { |
3464 | struct perf_buffer *buffer; |
3465 | unsigned long size; |
3466 | int i; |
3467 | |
3468 | size = sizeof(struct perf_buffer); |
3469 | size += nr_pages * sizeof(void *); |
3470 | |
3471 | buffer = kzalloc(size, GFP_KERNEL); |
3472 | if (!buffer) |
3473 | goto fail; |
3474 | |
3475 | buffer->user_page = perf_mmap_alloc_page(cpu); |
3476 | if (!buffer->user_page) |
3477 | goto fail_user_page; |
3478 | |
3479 | for (i = 0; i < nr_pages; i++) { |
3480 | buffer->data_pages[i] = perf_mmap_alloc_page(cpu); |
3481 | if (!buffer->data_pages[i]) |
3482 | goto fail_data_pages; |
3483 | } |
3484 | |
3485 | buffer->nr_pages = nr_pages; |
3486 | |
3487 | perf_buffer_init(buffer, watermark, flags); |
3488 | |
3489 | return buffer; |
3490 | |
3491 | fail_data_pages: |
3492 | for (i--; i >= 0; i--) |
3493 | free_page((unsigned long)buffer->data_pages[i]); |
3494 | |
3495 | free_page((unsigned long)buffer->user_page); |
3496 | |
3497 | fail_user_page: |
3498 | kfree(buffer); |
3499 | |
3500 | fail: |
3501 | return NULL; |
3502 | } |
3503 | |
3504 | static void perf_mmap_free_page(unsigned long addr) |
3505 | { |
3506 | struct page *page = virt_to_page((void *)addr); |
3507 | |
3508 | page->mapping = NULL; |
3509 | __free_page(page); |
3510 | } |
3511 | |
3512 | static void perf_buffer_free(struct perf_buffer *buffer) |
3513 | { |
3514 | int i; |
3515 | |
3516 | perf_mmap_free_page((unsigned long)buffer->user_page); |
3517 | for (i = 0; i < buffer->nr_pages; i++) |
3518 | perf_mmap_free_page((unsigned long)buffer->data_pages[i]); |
3519 | kfree(buffer); |
3520 | } |
3521 | |
3522 | static inline int page_order(struct perf_buffer *buffer) |
3523 | { |
3524 | return 0; |
3525 | } |
3526 | |
3527 | #else |
3528 | |
3529 | /* |
3530 | * Back perf_mmap() with vmalloc memory. |
3531 | * |
3532 | * Required for architectures that have d-cache aliasing issues. |
3533 | */ |
3534 | |
3535 | static inline int page_order(struct perf_buffer *buffer) |
3536 | { |
3537 | return buffer->page_order; |
3538 | } |
3539 | |
3540 | static struct page * |
3541 | perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff) |
3542 | { |
3543 | if (pgoff > (1UL << page_order(buffer))) |
3544 | return NULL; |
3545 | |
3546 | return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE); |
3547 | } |
3548 | |
3549 | static void perf_mmap_unmark_page(void *addr) |
3550 | { |
3551 | struct page *page = vmalloc_to_page(addr); |
3552 | |
3553 | page->mapping = NULL; |
3554 | } |
3555 | |
3556 | static void perf_buffer_free_work(struct work_struct *work) |
3557 | { |
3558 | struct perf_buffer *buffer; |
3559 | void *base; |
3560 | int i, nr; |
3561 | |
3562 | buffer = container_of(work, struct perf_buffer, work); |
3563 | nr = 1 << page_order(buffer); |
3564 | |
3565 | base = buffer->user_page; |
3566 | for (i = 0; i < nr + 1; i++) |
3567 | perf_mmap_unmark_page(base + (i * PAGE_SIZE)); |
3568 | |
3569 | vfree(base); |
3570 | kfree(buffer); |
3571 | } |
3572 | |
3573 | static void perf_buffer_free(struct perf_buffer *buffer) |
3574 | { |
3575 | schedule_work(&buffer->work); |
3576 | } |
3577 | |
3578 | static struct perf_buffer * |
3579 | perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags) |
3580 | { |
3581 | struct perf_buffer *buffer; |
3582 | unsigned long size; |
3583 | void *all_buf; |
3584 | |
3585 | size = sizeof(struct perf_buffer); |
3586 | size += sizeof(void *); |
3587 | |
3588 | buffer = kzalloc(size, GFP_KERNEL); |
3589 | if (!buffer) |
3590 | goto fail; |
3591 | |
3592 | INIT_WORK(&buffer->work, perf_buffer_free_work); |
3593 | |
3594 | all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE); |
3595 | if (!all_buf) |
3596 | goto fail_all_buf; |
3597 | |
3598 | buffer->user_page = all_buf; |
3599 | buffer->data_pages[0] = all_buf + PAGE_SIZE; |
3600 | buffer->page_order = ilog2(nr_pages); |
3601 | buffer->nr_pages = 1; |
3602 | |
3603 | perf_buffer_init(buffer, watermark, flags); |
3604 | |
3605 | return buffer; |
3606 | |
3607 | fail_all_buf: |
3608 | kfree(buffer); |
3609 | |
3610 | fail: |
3611 | return NULL; |
3612 | } |
3613 | |
3614 | #endif |
3615 | |
3616 | static unsigned long perf_data_size(struct perf_buffer *buffer) |
3617 | { |
3618 | return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer)); |
3619 | } |
3620 | |
3621 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3622 | { |
3623 | struct perf_event *event = vma->vm_file->private_data; |
3624 | struct perf_buffer *buffer; |
3625 | int ret = VM_FAULT_SIGBUS; |
3626 | |
3627 | if (vmf->flags & FAULT_FLAG_MKWRITE) { |
3628 | if (vmf->pgoff == 0) |
3629 | ret = 0; |
3630 | return ret; |
3631 | } |
3632 | |
3633 | rcu_read_lock(); |
3634 | buffer = rcu_dereference(event->buffer); |
3635 | if (!buffer) |
3636 | goto unlock; |
3637 | |
3638 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) |
3639 | goto unlock; |
3640 | |
3641 | vmf->page = perf_mmap_to_page(buffer, vmf->pgoff); |
3642 | if (!vmf->page) |
3643 | goto unlock; |
3644 | |
3645 | get_page(vmf->page); |
3646 | vmf->page->mapping = vma->vm_file->f_mapping; |
3647 | vmf->page->index = vmf->pgoff; |
3648 | |
3649 | ret = 0; |
3650 | unlock: |
3651 | rcu_read_unlock(); |
3652 | |
3653 | return ret; |
3654 | } |
3655 | |
3656 | static void perf_buffer_free_rcu(struct rcu_head *rcu_head) |
3657 | { |
3658 | struct perf_buffer *buffer; |
3659 | |
3660 | buffer = container_of(rcu_head, struct perf_buffer, rcu_head); |
3661 | perf_buffer_free(buffer); |
3662 | } |
3663 | |
3664 | static struct perf_buffer *perf_buffer_get(struct perf_event *event) |
3665 | { |
3666 | struct perf_buffer *buffer; |
3667 | |
3668 | rcu_read_lock(); |
3669 | buffer = rcu_dereference(event->buffer); |
3670 | if (buffer) { |
3671 | if (!atomic_inc_not_zero(&buffer->refcount)) |
3672 | buffer = NULL; |
3673 | } |
3674 | rcu_read_unlock(); |
3675 | |
3676 | return buffer; |
3677 | } |
3678 | |
3679 | static void perf_buffer_put(struct perf_buffer *buffer) |
3680 | { |
3681 | if (!atomic_dec_and_test(&buffer->refcount)) |
3682 | return; |
3683 | |
3684 | call_rcu(&buffer->rcu_head, perf_buffer_free_rcu); |
3685 | } |
3686 | |
3687 | static void perf_mmap_open(struct vm_area_struct *vma) |
3688 | { |
3689 | struct perf_event *event = vma->vm_file->private_data; |
3690 | |
3691 | atomic_inc(&event->mmap_count); |
3692 | } |
3693 | |
3694 | static void perf_mmap_close(struct vm_area_struct *vma) |
3695 | { |
3696 | struct perf_event *event = vma->vm_file->private_data; |
3697 | |
3698 | if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) { |
3699 | unsigned long size = perf_data_size(event->buffer); |
3700 | struct user_struct *user = event->mmap_user; |
3701 | struct perf_buffer *buffer = event->buffer; |
3702 | |
3703 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm); |
3704 | vma->vm_mm->locked_vm -= event->mmap_locked; |
3705 | rcu_assign_pointer(event->buffer, NULL); |
3706 | mutex_unlock(&event->mmap_mutex); |
3707 | |
3708 | perf_buffer_put(buffer); |
3709 | free_uid(user); |
3710 | } |
3711 | } |
3712 | |
3713 | static const struct vm_operations_struct perf_mmap_vmops = { |
3714 | .open = perf_mmap_open, |
3715 | .close = perf_mmap_close, |
3716 | .fault = perf_mmap_fault, |
3717 | .page_mkwrite = perf_mmap_fault, |
3718 | }; |
3719 | |
3720 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) |
3721 | { |
3722 | struct perf_event *event = file->private_data; |
3723 | unsigned long user_locked, user_lock_limit; |
3724 | struct user_struct *user = current_user(); |
3725 | unsigned long locked, lock_limit; |
3726 | struct perf_buffer *buffer; |
3727 | unsigned long vma_size; |
3728 | unsigned long nr_pages; |
3729 | long user_extra, extra; |
3730 | int ret = 0, flags = 0; |
3731 | |
3732 | /* |
3733 | * Don't allow mmap() of inherited per-task counters. This would |
3734 | * create a performance issue due to all children writing to the |
3735 | * same buffer. |
3736 | */ |
3737 | if (event->cpu == -1 && event->attr.inherit) |
3738 | return -EINVAL; |
3739 | |
3740 | if (!(vma->vm_flags & VM_SHARED)) |
3741 | return -EINVAL; |
3742 | |
3743 | vma_size = vma->vm_end - vma->vm_start; |
3744 | nr_pages = (vma_size / PAGE_SIZE) - 1; |
3745 | |
3746 | /* |
3747 | * If we have buffer pages ensure they're a power-of-two number, so we |
3748 | * can do bitmasks instead of modulo. |
3749 | */ |
3750 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) |
3751 | return -EINVAL; |
3752 | |
3753 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
3754 | return -EINVAL; |
3755 | |
3756 | if (vma->vm_pgoff != 0) |
3757 | return -EINVAL; |
3758 | |
3759 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3760 | mutex_lock(&event->mmap_mutex); |
3761 | if (event->buffer) { |
3762 | if (event->buffer->nr_pages == nr_pages) |
3763 | atomic_inc(&event->buffer->refcount); |
3764 | else |
3765 | ret = -EINVAL; |
3766 | goto unlock; |
3767 | } |
3768 | |
3769 | user_extra = nr_pages + 1; |
3770 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
3771 | |
3772 | /* |
3773 | * Increase the limit linearly with more CPUs: |
3774 | */ |
3775 | user_lock_limit *= num_online_cpus(); |
3776 | |
3777 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
3778 | |
3779 | extra = 0; |
3780 | if (user_locked > user_lock_limit) |
3781 | extra = user_locked - user_lock_limit; |
3782 | |
3783 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
3784 | lock_limit >>= PAGE_SHIFT; |
3785 | locked = vma->vm_mm->locked_vm + extra; |
3786 | |
3787 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
3788 | !capable(CAP_IPC_LOCK)) { |
3789 | ret = -EPERM; |
3790 | goto unlock; |
3791 | } |
3792 | |
3793 | WARN_ON(event->buffer); |
3794 | |
3795 | if (vma->vm_flags & VM_WRITE) |
3796 | flags |= PERF_BUFFER_WRITABLE; |
3797 | |
3798 | buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark, |
3799 | event->cpu, flags); |
3800 | if (!buffer) { |
3801 | ret = -ENOMEM; |
3802 | goto unlock; |
3803 | } |
3804 | rcu_assign_pointer(event->buffer, buffer); |
3805 | |
3806 | atomic_long_add(user_extra, &user->locked_vm); |
3807 | event->mmap_locked = extra; |
3808 | event->mmap_user = get_current_user(); |
3809 | vma->vm_mm->locked_vm += event->mmap_locked; |
3810 | |
3811 | unlock: |
3812 | if (!ret) |
3813 | atomic_inc(&event->mmap_count); |
3814 | mutex_unlock(&event->mmap_mutex); |
3815 | |
3816 | vma->vm_flags |= VM_RESERVED; |
3817 | vma->vm_ops = &perf_mmap_vmops; |
3818 | |
3819 | return ret; |
3820 | } |
3821 | |
3822 | static int perf_fasync(int fd, struct file *filp, int on) |
3823 | { |
3824 | struct inode *inode = filp->f_path.dentry->d_inode; |
3825 | struct perf_event *event = filp->private_data; |
3826 | int retval; |
3827 | |
3828 | mutex_lock(&inode->i_mutex); |
3829 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3830 | mutex_unlock(&inode->i_mutex); |
3831 | |
3832 | if (retval < 0) |
3833 | return retval; |
3834 | |
3835 | return 0; |
3836 | } |
3837 | |
3838 | static const struct file_operations perf_fops = { |
3839 | .llseek = no_llseek, |
3840 | .release = perf_release, |
3841 | .read = perf_read, |
3842 | .poll = perf_poll, |
3843 | .unlocked_ioctl = perf_ioctl, |
3844 | .compat_ioctl = perf_ioctl, |
3845 | .mmap = perf_mmap, |
3846 | .fasync = perf_fasync, |
3847 | }; |
3848 | |
3849 | /* |
3850 | * Perf event wakeup |
3851 | * |
3852 | * If there's data, ensure we set the poll() state and publish everything |
3853 | * to user-space before waking everybody up. |
3854 | */ |
3855 | |
3856 | void perf_event_wakeup(struct perf_event *event) |
3857 | { |
3858 | wake_up_all(&event->waitq); |
3859 | |
3860 | if (event->pending_kill) { |
3861 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); |
3862 | event->pending_kill = 0; |
3863 | } |
3864 | } |
3865 | |
3866 | static void perf_pending_event(struct irq_work *entry) |
3867 | { |
3868 | struct perf_event *event = container_of(entry, |
3869 | struct perf_event, pending); |
3870 | |
3871 | if (event->pending_disable) { |
3872 | event->pending_disable = 0; |
3873 | __perf_event_disable(event); |
3874 | } |
3875 | |
3876 | if (event->pending_wakeup) { |
3877 | event->pending_wakeup = 0; |
3878 | perf_event_wakeup(event); |
3879 | } |
3880 | } |
3881 | |
3882 | /* |
3883 | * We assume there is only KVM supporting the callbacks. |
3884 | * Later on, we might change it to a list if there is |
3885 | * another virtualization implementation supporting the callbacks. |
3886 | */ |
3887 | struct perf_guest_info_callbacks *perf_guest_cbs; |
3888 | |
3889 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) |
3890 | { |
3891 | perf_guest_cbs = cbs; |
3892 | return 0; |
3893 | } |
3894 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); |
3895 | |
3896 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) |
3897 | { |
3898 | perf_guest_cbs = NULL; |
3899 | return 0; |
3900 | } |
3901 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); |
3902 | |
3903 | /* |
3904 | * Output |
3905 | */ |
3906 | static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail, |
3907 | unsigned long offset, unsigned long head) |
3908 | { |
3909 | unsigned long mask; |
3910 | |
3911 | if (!buffer->writable) |
3912 | return true; |
3913 | |
3914 | mask = perf_data_size(buffer) - 1; |
3915 | |
3916 | offset = (offset - tail) & mask; |
3917 | head = (head - tail) & mask; |
3918 | |
3919 | if ((int)(head - offset) < 0) |
3920 | return false; |
3921 | |
3922 | return true; |
3923 | } |
3924 | |
3925 | static void perf_output_wakeup(struct perf_output_handle *handle) |
3926 | { |
3927 | atomic_set(&handle->buffer->poll, POLL_IN); |
3928 | |
3929 | if (handle->nmi) { |
3930 | handle->event->pending_wakeup = 1; |
3931 | irq_work_queue(&handle->event->pending); |
3932 | } else |
3933 | perf_event_wakeup(handle->event); |
3934 | } |
3935 | |
3936 | /* |
3937 | * We need to ensure a later event_id doesn't publish a head when a former |
3938 | * event isn't done writing. However since we need to deal with NMIs we |
3939 | * cannot fully serialize things. |
3940 | * |
3941 | * We only publish the head (and generate a wakeup) when the outer-most |
3942 | * event completes. |
3943 | */ |
3944 | static void perf_output_get_handle(struct perf_output_handle *handle) |
3945 | { |
3946 | struct perf_buffer *buffer = handle->buffer; |
3947 | |
3948 | preempt_disable(); |
3949 | local_inc(&buffer->nest); |
3950 | handle->wakeup = local_read(&buffer->wakeup); |
3951 | } |
3952 | |
3953 | static void perf_output_put_handle(struct perf_output_handle *handle) |
3954 | { |
3955 | struct perf_buffer *buffer = handle->buffer; |
3956 | unsigned long head; |
3957 | |
3958 | again: |
3959 | head = local_read(&buffer->head); |
3960 | |
3961 | /* |
3962 | * IRQ/NMI can happen here, which means we can miss a head update. |
3963 | */ |
3964 | |
3965 | if (!local_dec_and_test(&buffer->nest)) |
3966 | goto out; |
3967 | |
3968 | /* |
3969 | * Publish the known good head. Rely on the full barrier implied |
3970 | * by atomic_dec_and_test() order the buffer->head read and this |
3971 | * write. |
3972 | */ |
3973 | buffer->user_page->data_head = head; |
3974 | |
3975 | /* |
3976 | * Now check if we missed an update, rely on the (compiler) |
3977 | * barrier in atomic_dec_and_test() to re-read buffer->head. |
3978 | */ |
3979 | if (unlikely(head != local_read(&buffer->head))) { |
3980 | local_inc(&buffer->nest); |
3981 | goto again; |
3982 | } |
3983 | |
3984 | if (handle->wakeup != local_read(&buffer->wakeup)) |
3985 | perf_output_wakeup(handle); |
3986 | |
3987 | out: |
3988 | preempt_enable(); |
3989 | } |
3990 | |
3991 | __always_inline void perf_output_copy(struct perf_output_handle *handle, |
3992 | const void *buf, unsigned int len) |
3993 | { |
3994 | do { |
3995 | unsigned long size = min_t(unsigned long, handle->size, len); |
3996 | |
3997 | memcpy(handle->addr, buf, size); |
3998 | |
3999 | len -= size; |
4000 | handle->addr += size; |
4001 | buf += size; |
4002 | handle->size -= size; |
4003 | if (!handle->size) { |
4004 | struct perf_buffer *buffer = handle->buffer; |
4005 | |
4006 | handle->page++; |
4007 | handle->page &= buffer->nr_pages - 1; |
4008 | handle->addr = buffer->data_pages[handle->page]; |
4009 | handle->size = PAGE_SIZE << page_order(buffer); |
4010 | } |
4011 | } while (len); |
4012 | } |
4013 | |
4014 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4015 | struct perf_sample_data *data, |
4016 | struct perf_event *event) |
4017 | { |
4018 | u64 sample_type = event->attr.sample_type; |
4019 | |
4020 | data->type = sample_type; |
4021 | header->size += event->id_header_size; |
4022 | |
4023 | if (sample_type & PERF_SAMPLE_TID) { |
4024 | /* namespace issues */ |
4025 | data->tid_entry.pid = perf_event_pid(event, current); |
4026 | data->tid_entry.tid = perf_event_tid(event, current); |
4027 | } |
4028 | |
4029 | if (sample_type & PERF_SAMPLE_TIME) |
4030 | data->time = perf_clock(); |
4031 | |
4032 | if (sample_type & PERF_SAMPLE_ID) |
4033 | data->id = primary_event_id(event); |
4034 | |
4035 | if (sample_type & PERF_SAMPLE_STREAM_ID) |
4036 | data->stream_id = event->id; |
4037 | |
4038 | if (sample_type & PERF_SAMPLE_CPU) { |
4039 | data->cpu_entry.cpu = raw_smp_processor_id(); |
4040 | data->cpu_entry.reserved = 0; |
4041 | } |
4042 | } |
4043 | |
4044 | static void perf_event_header__init_id(struct perf_event_header *header, |
4045 | struct perf_sample_data *data, |
4046 | struct perf_event *event) |
4047 | { |
4048 | if (event->attr.sample_id_all) |
4049 | __perf_event_header__init_id(header, data, event); |
4050 | } |
4051 | |
4052 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, |
4053 | struct perf_sample_data *data) |
4054 | { |
4055 | u64 sample_type = data->type; |
4056 | |
4057 | if (sample_type & PERF_SAMPLE_TID) |
4058 | perf_output_put(handle, data->tid_entry); |
4059 | |
4060 | if (sample_type & PERF_SAMPLE_TIME) |
4061 | perf_output_put(handle, data->time); |
4062 | |
4063 | if (sample_type & PERF_SAMPLE_ID) |
4064 | perf_output_put(handle, data->id); |
4065 | |
4066 | if (sample_type & PERF_SAMPLE_STREAM_ID) |
4067 | perf_output_put(handle, data->stream_id); |
4068 | |
4069 | if (sample_type & PERF_SAMPLE_CPU) |
4070 | perf_output_put(handle, data->cpu_entry); |
4071 | } |
4072 | |
4073 | static void perf_event__output_id_sample(struct perf_event *event, |
4074 | struct perf_output_handle *handle, |
4075 | struct perf_sample_data *sample) |
4076 | { |
4077 | if (event->attr.sample_id_all) |
4078 | __perf_event__output_id_sample(handle, sample); |
4079 | } |
4080 | |
4081 | int perf_output_begin(struct perf_output_handle *handle, |
4082 | struct perf_event *event, unsigned int size, |
4083 | int nmi, int sample) |
4084 | { |
4085 | struct perf_buffer *buffer; |
4086 | unsigned long tail, offset, head; |
4087 | int have_lost; |
4088 | struct perf_sample_data sample_data; |
4089 | struct { |
4090 | struct perf_event_header header; |
4091 | u64 id; |
4092 | u64 lost; |
4093 | } lost_event; |
4094 | |
4095 | rcu_read_lock(); |
4096 | /* |
4097 | * For inherited events we send all the output towards the parent. |
4098 | */ |
4099 | if (event->parent) |
4100 | event = event->parent; |
4101 | |
4102 | buffer = rcu_dereference(event->buffer); |
4103 | if (!buffer) |
4104 | goto out; |
4105 | |
4106 | handle->buffer = buffer; |
4107 | handle->event = event; |
4108 | handle->nmi = nmi; |
4109 | handle->sample = sample; |
4110 | |
4111 | if (!buffer->nr_pages) |
4112 | goto out; |
4113 | |
4114 | have_lost = local_read(&buffer->lost); |
4115 | if (have_lost) { |
4116 | lost_event.header.size = sizeof(lost_event); |
4117 | perf_event_header__init_id(&lost_event.header, &sample_data, |
4118 | event); |
4119 | size += lost_event.header.size; |
4120 | } |
4121 | |
4122 | perf_output_get_handle(handle); |
4123 | |
4124 | do { |
4125 | /* |
4126 | * Userspace could choose to issue a mb() before updating the |
4127 | * tail pointer. So that all reads will be completed before the |
4128 | * write is issued. |
4129 | */ |
4130 | tail = ACCESS_ONCE(buffer->user_page->data_tail); |
4131 | smp_rmb(); |
4132 | offset = head = local_read(&buffer->head); |
4133 | head += size; |
4134 | if (unlikely(!perf_output_space(buffer, tail, offset, head))) |
4135 | goto fail; |
4136 | } while (local_cmpxchg(&buffer->head, offset, head) != offset); |
4137 | |
4138 | if (head - local_read(&buffer->wakeup) > buffer->watermark) |
4139 | local_add(buffer->watermark, &buffer->wakeup); |
4140 | |
4141 | handle->page = offset >> (PAGE_SHIFT + page_order(buffer)); |
4142 | handle->page &= buffer->nr_pages - 1; |
4143 | handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1); |
4144 | handle->addr = buffer->data_pages[handle->page]; |
4145 | handle->addr += handle->size; |
4146 | handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size; |
4147 | |
4148 | if (have_lost) { |
4149 | lost_event.header.type = PERF_RECORD_LOST; |
4150 | lost_event.header.misc = 0; |
4151 | lost_event.id = event->id; |
4152 | lost_event.lost = local_xchg(&buffer->lost, 0); |
4153 | |
4154 | perf_output_put(handle, lost_event); |
4155 | perf_event__output_id_sample(event, handle, &sample_data); |
4156 | } |
4157 | |
4158 | return 0; |
4159 | |
4160 | fail: |
4161 | local_inc(&buffer->lost); |
4162 | perf_output_put_handle(handle); |
4163 | out: |
4164 | rcu_read_unlock(); |
4165 | |
4166 | return -ENOSPC; |
4167 | } |
4168 | |
4169 | void perf_output_end(struct perf_output_handle *handle) |
4170 | { |
4171 | struct perf_event *event = handle->event; |
4172 | struct perf_buffer *buffer = handle->buffer; |
4173 | |
4174 | int wakeup_events = event->attr.wakeup_events; |
4175 | |
4176 | if (handle->sample && wakeup_events) { |
4177 | int events = local_inc_return(&buffer->events); |
4178 | if (events >= wakeup_events) { |
4179 | local_sub(wakeup_events, &buffer->events); |
4180 | local_inc(&buffer->wakeup); |
4181 | } |
4182 | } |
4183 | |
4184 | perf_output_put_handle(handle); |
4185 | rcu_read_unlock(); |
4186 | } |
4187 | |
4188 | static void perf_output_read_one(struct perf_output_handle *handle, |
4189 | struct perf_event *event, |
4190 | u64 enabled, u64 running) |
4191 | { |
4192 | u64 read_format = event->attr.read_format; |
4193 | u64 values[4]; |
4194 | int n = 0; |
4195 | |
4196 | values[n++] = perf_event_count(event); |
4197 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
4198 | values[n++] = enabled + |
4199 | atomic64_read(&event->child_total_time_enabled); |
4200 | } |
4201 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { |
4202 | values[n++] = running + |
4203 | atomic64_read(&event->child_total_time_running); |
4204 | } |
4205 | if (read_format & PERF_FORMAT_ID) |
4206 | values[n++] = primary_event_id(event); |
4207 | |
4208 | perf_output_copy(handle, values, n * sizeof(u64)); |
4209 | } |
4210 | |
4211 | /* |
4212 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
4213 | */ |
4214 | static void perf_output_read_group(struct perf_output_handle *handle, |
4215 | struct perf_event *event, |
4216 | u64 enabled, u64 running) |
4217 | { |
4218 | struct perf_event *leader = event->group_leader, *sub; |
4219 | u64 read_format = event->attr.read_format; |
4220 | u64 values[5]; |
4221 | int n = 0; |
4222 | |
4223 | values[n++] = 1 + leader->nr_siblings; |
4224 | |
4225 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
4226 | values[n++] = enabled; |
4227 | |
4228 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
4229 | values[n++] = running; |
4230 | |
4231 | if (leader != event) |
4232 | leader->pmu->read(leader); |
4233 | |
4234 | values[n++] = perf_event_count(leader); |
4235 | if (read_format & PERF_FORMAT_ID) |
4236 | values[n++] = primary_event_id(leader); |
4237 | |
4238 | perf_output_copy(handle, values, n * sizeof(u64)); |
4239 | |
4240 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
4241 | n = 0; |
4242 | |
4243 | if (sub != event) |
4244 | sub->pmu->read(sub); |
4245 | |
4246 | values[n++] = perf_event_count(sub); |
4247 | if (read_format & PERF_FORMAT_ID) |
4248 | values[n++] = primary_event_id(sub); |
4249 | |
4250 | perf_output_copy(handle, values, n * sizeof(u64)); |
4251 | } |
4252 | } |
4253 | |
4254 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4255 | PERF_FORMAT_TOTAL_TIME_RUNNING) |
4256 | |
4257 | static void perf_output_read(struct perf_output_handle *handle, |
4258 | struct perf_event *event) |
4259 | { |
4260 | u64 enabled = 0, running = 0, now, ctx_time; |
4261 | u64 read_format = event->attr.read_format; |
4262 | |
4263 | /* |
4264 | * compute total_time_enabled, total_time_running |
4265 | * based on snapshot values taken when the event |
4266 | * was last scheduled in. |
4267 | * |
4268 | * we cannot simply called update_context_time() |
4269 | * because of locking issue as we are called in |
4270 | * NMI context |
4271 | */ |
4272 | if (read_format & PERF_FORMAT_TOTAL_TIMES) { |
4273 | now = perf_clock(); |
4274 | ctx_time = event->shadow_ctx_time + now; |
4275 | enabled = ctx_time - event->tstamp_enabled; |
4276 | running = ctx_time - event->tstamp_running; |
4277 | } |
4278 | |
4279 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
4280 | perf_output_read_group(handle, event, enabled, running); |
4281 | else |
4282 | perf_output_read_one(handle, event, enabled, running); |
4283 | } |
4284 | |
4285 | void perf_output_sample(struct perf_output_handle *handle, |
4286 | struct perf_event_header *header, |
4287 | struct perf_sample_data *data, |
4288 | struct perf_event *event) |
4289 | { |
4290 | u64 sample_type = data->type; |
4291 | |
4292 | perf_output_put(handle, *header); |
4293 | |
4294 | if (sample_type & PERF_SAMPLE_IP) |
4295 | perf_output_put(handle, data->ip); |
4296 | |
4297 | if (sample_type & PERF_SAMPLE_TID) |
4298 | perf_output_put(handle, data->tid_entry); |
4299 | |
4300 | if (sample_type & PERF_SAMPLE_TIME) |
4301 | perf_output_put(handle, data->time); |
4302 | |
4303 | if (sample_type & PERF_SAMPLE_ADDR) |
4304 | perf_output_put(handle, data->addr); |
4305 | |
4306 | if (sample_type & PERF_SAMPLE_ID) |
4307 | perf_output_put(handle, data->id); |
4308 | |
4309 | if (sample_type & PERF_SAMPLE_STREAM_ID) |
4310 | perf_output_put(handle, data->stream_id); |
4311 | |
4312 | if (sample_type & PERF_SAMPLE_CPU) |
4313 | perf_output_put(handle, data->cpu_entry); |
4314 | |
4315 | if (sample_type & PERF_SAMPLE_PERIOD) |
4316 | perf_output_put(handle, data->period); |
4317 | |
4318 | if (sample_type & PERF_SAMPLE_READ) |
4319 | perf_output_read(handle, event); |
4320 | |
4321 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
4322 | if (data->callchain) { |
4323 | int size = 1; |
4324 | |
4325 | if (data->callchain) |
4326 | size += data->callchain->nr; |
4327 | |
4328 | size *= sizeof(u64); |
4329 | |
4330 | perf_output_copy(handle, data->callchain, size); |
4331 | } else { |
4332 | u64 nr = 0; |
4333 | perf_output_put(handle, nr); |
4334 | } |
4335 | } |
4336 | |
4337 | if (sample_type & PERF_SAMPLE_RAW) { |
4338 | if (data->raw) { |
4339 | perf_output_put(handle, data->raw->size); |
4340 | perf_output_copy(handle, data->raw->data, |
4341 | data->raw->size); |
4342 | } else { |
4343 | struct { |
4344 | u32 size; |
4345 | u32 data; |
4346 | } raw = { |
4347 | .size = sizeof(u32), |
4348 | .data = 0, |
4349 | }; |
4350 | perf_output_put(handle, raw); |
4351 | } |
4352 | } |
4353 | } |
4354 | |
4355 | void perf_prepare_sample(struct perf_event_header *header, |
4356 | struct perf_sample_data *data, |
4357 | struct perf_event *event, |
4358 | struct pt_regs *regs) |
4359 | { |
4360 | u64 sample_type = event->attr.sample_type; |
4361 | |
4362 | header->type = PERF_RECORD_SAMPLE; |
4363 | header->size = sizeof(*header) + event->header_size; |
4364 | |
4365 | header->misc = 0; |
4366 | header->misc |= perf_misc_flags(regs); |
4367 | |
4368 | __perf_event_header__init_id(header, data, event); |
4369 | |
4370 | if (sample_type & PERF_SAMPLE_IP) |
4371 | data->ip = perf_instruction_pointer(regs); |
4372 | |
4373 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
4374 | int size = 1; |
4375 | |
4376 | data->callchain = perf_callchain(regs); |
4377 | |
4378 | if (data->callchain) |
4379 | size += data->callchain->nr; |
4380 | |
4381 | header->size += size * sizeof(u64); |
4382 | } |
4383 | |
4384 | if (sample_type & PERF_SAMPLE_RAW) { |
4385 | int size = sizeof(u32); |
4386 | |
4387 | if (data->raw) |
4388 | size += data->raw->size; |
4389 | else |
4390 | size += sizeof(u32); |
4391 | |
4392 | WARN_ON_ONCE(size & (sizeof(u64)-1)); |
4393 | header->size += size; |
4394 | } |
4395 | } |
4396 | |
4397 | static void perf_event_output(struct perf_event *event, int nmi, |
4398 | struct perf_sample_data *data, |
4399 | struct pt_regs *regs) |
4400 | { |
4401 | struct perf_output_handle handle; |
4402 | struct perf_event_header header; |
4403 | |
4404 | /* protect the callchain buffers */ |
4405 | rcu_read_lock(); |
4406 | |
4407 | perf_prepare_sample(&header, data, event, regs); |
4408 | |
4409 | if (perf_output_begin(&handle, event, header.size, nmi, 1)) |
4410 | goto exit; |
4411 | |
4412 | perf_output_sample(&handle, &header, data, event); |
4413 | |
4414 | perf_output_end(&handle); |
4415 | |
4416 | exit: |
4417 | rcu_read_unlock(); |
4418 | } |
4419 | |
4420 | /* |
4421 | * read event_id |
4422 | */ |
4423 | |
4424 | struct perf_read_event { |
4425 | struct perf_event_header header; |
4426 | |
4427 | u32 pid; |
4428 | u32 tid; |
4429 | }; |
4430 | |
4431 | static void |
4432 | perf_event_read_event(struct perf_event *event, |
4433 | struct task_struct *task) |
4434 | { |
4435 | struct perf_output_handle handle; |
4436 | struct perf_sample_data sample; |
4437 | struct perf_read_event read_event = { |
4438 | .header = { |
4439 | .type = PERF_RECORD_READ, |
4440 | .misc = 0, |
4441 | .size = sizeof(read_event) + event->read_size, |
4442 | }, |
4443 | .pid = perf_event_pid(event, task), |
4444 | .tid = perf_event_tid(event, task), |
4445 | }; |
4446 | int ret; |
4447 | |
4448 | perf_event_header__init_id(&read_event.header, &sample, event); |
4449 | ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0); |
4450 | if (ret) |
4451 | return; |
4452 | |
4453 | perf_output_put(&handle, read_event); |
4454 | perf_output_read(&handle, event); |
4455 | perf_event__output_id_sample(event, &handle, &sample); |
4456 | |
4457 | perf_output_end(&handle); |
4458 | } |
4459 | |
4460 | /* |
4461 | * task tracking -- fork/exit |
4462 | * |
4463 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
4464 | */ |
4465 | |
4466 | struct perf_task_event { |
4467 | struct task_struct *task; |
4468 | struct perf_event_context *task_ctx; |
4469 | |
4470 | struct { |
4471 | struct perf_event_header header; |
4472 | |
4473 | u32 pid; |
4474 | u32 ppid; |
4475 | u32 tid; |
4476 | u32 ptid; |
4477 | u64 time; |
4478 | } event_id; |
4479 | }; |
4480 | |
4481 | static void perf_event_task_output(struct perf_event *event, |
4482 | struct perf_task_event *task_event) |
4483 | { |
4484 | struct perf_output_handle handle; |
4485 | struct perf_sample_data sample; |
4486 | struct task_struct *task = task_event->task; |
4487 | int ret, size = task_event->event_id.header.size; |
4488 | |
4489 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
4490 | |
4491 | ret = perf_output_begin(&handle, event, |
4492 | task_event->event_id.header.size, 0, 0); |
4493 | if (ret) |
4494 | goto out; |
4495 | |
4496 | task_event->event_id.pid = perf_event_pid(event, task); |
4497 | task_event->event_id.ppid = perf_event_pid(event, current); |
4498 | |
4499 | task_event->event_id.tid = perf_event_tid(event, task); |
4500 | task_event->event_id.ptid = perf_event_tid(event, current); |
4501 | |
4502 | perf_output_put(&handle, task_event->event_id); |
4503 | |
4504 | perf_event__output_id_sample(event, &handle, &sample); |
4505 | |
4506 | perf_output_end(&handle); |
4507 | out: |
4508 | task_event->event_id.header.size = size; |
4509 | } |
4510 | |
4511 | static int perf_event_task_match(struct perf_event *event) |
4512 | { |
4513 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
4514 | return 0; |
4515 | |
4516 | if (!event_filter_match(event)) |
4517 | return 0; |
4518 | |
4519 | if (event->attr.comm || event->attr.mmap || |
4520 | event->attr.mmap_data || event->attr.task) |
4521 | return 1; |
4522 | |
4523 | return 0; |
4524 | } |
4525 | |
4526 | static void perf_event_task_ctx(struct perf_event_context *ctx, |
4527 | struct perf_task_event *task_event) |
4528 | { |
4529 | struct perf_event *event; |
4530 | |
4531 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4532 | if (perf_event_task_match(event)) |
4533 | perf_event_task_output(event, task_event); |
4534 | } |
4535 | } |
4536 | |
4537 | static void perf_event_task_event(struct perf_task_event *task_event) |
4538 | { |
4539 | struct perf_cpu_context *cpuctx; |
4540 | struct perf_event_context *ctx; |
4541 | struct pmu *pmu; |
4542 | int ctxn; |
4543 | |
4544 | rcu_read_lock(); |
4545 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
4546 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
4547 | if (cpuctx->active_pmu != pmu) |
4548 | goto next; |
4549 | perf_event_task_ctx(&cpuctx->ctx, task_event); |
4550 | |
4551 | ctx = task_event->task_ctx; |
4552 | if (!ctx) { |
4553 | ctxn = pmu->task_ctx_nr; |
4554 | if (ctxn < 0) |
4555 | goto next; |
4556 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); |
4557 | } |
4558 | if (ctx) |
4559 | perf_event_task_ctx(ctx, task_event); |
4560 | next: |
4561 | put_cpu_ptr(pmu->pmu_cpu_context); |
4562 | } |
4563 | rcu_read_unlock(); |
4564 | } |
4565 | |
4566 | static void perf_event_task(struct task_struct *task, |
4567 | struct perf_event_context *task_ctx, |
4568 | int new) |
4569 | { |
4570 | struct perf_task_event task_event; |
4571 | |
4572 | if (!atomic_read(&nr_comm_events) && |
4573 | !atomic_read(&nr_mmap_events) && |
4574 | !atomic_read(&nr_task_events)) |
4575 | return; |
4576 | |
4577 | task_event = (struct perf_task_event){ |
4578 | .task = task, |
4579 | .task_ctx = task_ctx, |
4580 | .event_id = { |
4581 | .header = { |
4582 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
4583 | .misc = 0, |
4584 | .size = sizeof(task_event.event_id), |
4585 | }, |
4586 | /* .pid */ |
4587 | /* .ppid */ |
4588 | /* .tid */ |
4589 | /* .ptid */ |
4590 | .time = perf_clock(), |
4591 | }, |
4592 | }; |
4593 | |
4594 | perf_event_task_event(&task_event); |
4595 | } |
4596 | |
4597 | void perf_event_fork(struct task_struct *task) |
4598 | { |
4599 | perf_event_task(task, NULL, 1); |
4600 | } |
4601 | |
4602 | /* |
4603 | * comm tracking |
4604 | */ |
4605 | |
4606 | struct perf_comm_event { |
4607 | struct task_struct *task; |
4608 | char *comm; |
4609 | int comm_size; |
4610 | |
4611 | struct { |
4612 | struct perf_event_header header; |
4613 | |
4614 | u32 pid; |
4615 | u32 tid; |
4616 | } event_id; |
4617 | }; |
4618 | |
4619 | static void perf_event_comm_output(struct perf_event *event, |
4620 | struct perf_comm_event *comm_event) |
4621 | { |
4622 | struct perf_output_handle handle; |
4623 | struct perf_sample_data sample; |
4624 | int size = comm_event->event_id.header.size; |
4625 | int ret; |
4626 | |
4627 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
4628 | ret = perf_output_begin(&handle, event, |
4629 | comm_event->event_id.header.size, 0, 0); |
4630 | |
4631 | if (ret) |
4632 | goto out; |
4633 | |
4634 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4635 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); |
4636 | |
4637 | perf_output_put(&handle, comm_event->event_id); |
4638 | perf_output_copy(&handle, comm_event->comm, |
4639 | comm_event->comm_size); |
4640 | |
4641 | perf_event__output_id_sample(event, &handle, &sample); |
4642 | |
4643 | perf_output_end(&handle); |
4644 | out: |
4645 | comm_event->event_id.header.size = size; |
4646 | } |
4647 | |
4648 | static int perf_event_comm_match(struct perf_event *event) |
4649 | { |
4650 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
4651 | return 0; |
4652 | |
4653 | if (!event_filter_match(event)) |
4654 | return 0; |
4655 | |
4656 | if (event->attr.comm) |
4657 | return 1; |
4658 | |
4659 | return 0; |
4660 | } |
4661 | |
4662 | static void perf_event_comm_ctx(struct perf_event_context *ctx, |
4663 | struct perf_comm_event *comm_event) |
4664 | { |
4665 | struct perf_event *event; |
4666 | |
4667 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4668 | if (perf_event_comm_match(event)) |
4669 | perf_event_comm_output(event, comm_event); |
4670 | } |
4671 | } |
4672 | |
4673 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
4674 | { |
4675 | struct perf_cpu_context *cpuctx; |
4676 | struct perf_event_context *ctx; |
4677 | char comm[TASK_COMM_LEN]; |
4678 | unsigned int size; |
4679 | struct pmu *pmu; |
4680 | int ctxn; |
4681 | |
4682 | memset(comm, 0, sizeof(comm)); |
4683 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
4684 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
4685 | |
4686 | comm_event->comm = comm; |
4687 | comm_event->comm_size = size; |
4688 | |
4689 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
4690 | rcu_read_lock(); |
4691 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
4692 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
4693 | if (cpuctx->active_pmu != pmu) |
4694 | goto next; |
4695 | perf_event_comm_ctx(&cpuctx->ctx, comm_event); |
4696 | |
4697 | ctxn = pmu->task_ctx_nr; |
4698 | if (ctxn < 0) |
4699 | goto next; |
4700 | |
4701 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); |
4702 | if (ctx) |
4703 | perf_event_comm_ctx(ctx, comm_event); |
4704 | next: |
4705 | put_cpu_ptr(pmu->pmu_cpu_context); |
4706 | } |
4707 | rcu_read_unlock(); |
4708 | } |
4709 | |
4710 | void perf_event_comm(struct task_struct *task) |
4711 | { |
4712 | struct perf_comm_event comm_event; |
4713 | struct perf_event_context *ctx; |
4714 | int ctxn; |
4715 | |
4716 | for_each_task_context_nr(ctxn) { |
4717 | ctx = task->perf_event_ctxp[ctxn]; |
4718 | if (!ctx) |
4719 | continue; |
4720 | |
4721 | perf_event_enable_on_exec(ctx); |
4722 | } |
4723 | |
4724 | if (!atomic_read(&nr_comm_events)) |
4725 | return; |
4726 | |
4727 | comm_event = (struct perf_comm_event){ |
4728 | .task = task, |
4729 | /* .comm */ |
4730 | /* .comm_size */ |
4731 | .event_id = { |
4732 | .header = { |
4733 | .type = PERF_RECORD_COMM, |
4734 | .misc = 0, |
4735 | /* .size */ |
4736 | }, |
4737 | /* .pid */ |
4738 | /* .tid */ |
4739 | }, |
4740 | }; |
4741 | |
4742 | perf_event_comm_event(&comm_event); |
4743 | } |
4744 | |
4745 | /* |
4746 | * mmap tracking |
4747 | */ |
4748 | |
4749 | struct perf_mmap_event { |
4750 | struct vm_area_struct *vma; |
4751 | |
4752 | const char *file_name; |
4753 | int file_size; |
4754 | |
4755 | struct { |
4756 | struct perf_event_header header; |
4757 | |
4758 | u32 pid; |
4759 | u32 tid; |
4760 | u64 start; |
4761 | u64 len; |
4762 | u64 pgoff; |
4763 | } event_id; |
4764 | }; |
4765 | |
4766 | static void perf_event_mmap_output(struct perf_event *event, |
4767 | struct perf_mmap_event *mmap_event) |
4768 | { |
4769 | struct perf_output_handle handle; |
4770 | struct perf_sample_data sample; |
4771 | int size = mmap_event->event_id.header.size; |
4772 | int ret; |
4773 | |
4774 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
4775 | ret = perf_output_begin(&handle, event, |
4776 | mmap_event->event_id.header.size, 0, 0); |
4777 | if (ret) |
4778 | goto out; |
4779 | |
4780 | mmap_event->event_id.pid = perf_event_pid(event, current); |
4781 | mmap_event->event_id.tid = perf_event_tid(event, current); |
4782 | |
4783 | perf_output_put(&handle, mmap_event->event_id); |
4784 | perf_output_copy(&handle, mmap_event->file_name, |
4785 | mmap_event->file_size); |
4786 | |
4787 | perf_event__output_id_sample(event, &handle, &sample); |
4788 | |
4789 | perf_output_end(&handle); |
4790 | out: |
4791 | mmap_event->event_id.header.size = size; |
4792 | } |
4793 | |
4794 | static int perf_event_mmap_match(struct perf_event *event, |
4795 | struct perf_mmap_event *mmap_event, |
4796 | int executable) |
4797 | { |
4798 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
4799 | return 0; |
4800 | |
4801 | if (!event_filter_match(event)) |
4802 | return 0; |
4803 | |
4804 | if ((!executable && event->attr.mmap_data) || |
4805 | (executable && event->attr.mmap)) |
4806 | return 1; |
4807 | |
4808 | return 0; |
4809 | } |
4810 | |
4811 | static void perf_event_mmap_ctx(struct perf_event_context *ctx, |
4812 | struct perf_mmap_event *mmap_event, |
4813 | int executable) |
4814 | { |
4815 | struct perf_event *event; |
4816 | |
4817 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4818 | if (perf_event_mmap_match(event, mmap_event, executable)) |
4819 | perf_event_mmap_output(event, mmap_event); |
4820 | } |
4821 | } |
4822 | |
4823 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
4824 | { |
4825 | struct perf_cpu_context *cpuctx; |
4826 | struct perf_event_context *ctx; |
4827 | struct vm_area_struct *vma = mmap_event->vma; |
4828 | struct file *file = vma->vm_file; |
4829 | unsigned int size; |
4830 | char tmp[16]; |
4831 | char *buf = NULL; |
4832 | const char *name; |
4833 | struct pmu *pmu; |
4834 | int ctxn; |
4835 | |
4836 | memset(tmp, 0, sizeof(tmp)); |
4837 | |
4838 | if (file) { |
4839 | /* |
4840 | * d_path works from the end of the buffer backwards, so we |
4841 | * need to add enough zero bytes after the string to handle |
4842 | * the 64bit alignment we do later. |
4843 | */ |
4844 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); |
4845 | if (!buf) { |
4846 | name = strncpy(tmp, "//enomem", sizeof(tmp)); |
4847 | goto got_name; |
4848 | } |
4849 | name = d_path(&file->f_path, buf, PATH_MAX); |
4850 | if (IS_ERR(name)) { |
4851 | name = strncpy(tmp, "//toolong", sizeof(tmp)); |
4852 | goto got_name; |
4853 | } |
4854 | } else { |
4855 | if (arch_vma_name(mmap_event->vma)) { |
4856 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), |
4857 | sizeof(tmp)); |
4858 | goto got_name; |
4859 | } |
4860 | |
4861 | if (!vma->vm_mm) { |
4862 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); |
4863 | goto got_name; |
4864 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
4865 | vma->vm_end >= vma->vm_mm->brk) { |
4866 | name = strncpy(tmp, "[heap]", sizeof(tmp)); |
4867 | goto got_name; |
4868 | } else if (vma->vm_start <= vma->vm_mm->start_stack && |
4869 | vma->vm_end >= vma->vm_mm->start_stack) { |
4870 | name = strncpy(tmp, "[stack]", sizeof(tmp)); |
4871 | goto got_name; |
4872 | } |
4873 | |
4874 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
4875 | goto got_name; |
4876 | } |
4877 | |
4878 | got_name: |
4879 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
4880 | |
4881 | mmap_event->file_name = name; |
4882 | mmap_event->file_size = size; |
4883 | |
4884 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
4885 | |
4886 | rcu_read_lock(); |
4887 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
4888 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
4889 | if (cpuctx->active_pmu != pmu) |
4890 | goto next; |
4891 | perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, |
4892 | vma->vm_flags & VM_EXEC); |
4893 | |
4894 | ctxn = pmu->task_ctx_nr; |
4895 | if (ctxn < 0) |
4896 | goto next; |
4897 | |
4898 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); |
4899 | if (ctx) { |
4900 | perf_event_mmap_ctx(ctx, mmap_event, |
4901 | vma->vm_flags & VM_EXEC); |
4902 | } |
4903 | next: |
4904 | put_cpu_ptr(pmu->pmu_cpu_context); |
4905 | } |
4906 | rcu_read_unlock(); |
4907 | |
4908 | kfree(buf); |
4909 | } |
4910 | |
4911 | void perf_event_mmap(struct vm_area_struct *vma) |
4912 | { |
4913 | struct perf_mmap_event mmap_event; |
4914 | |
4915 | if (!atomic_read(&nr_mmap_events)) |
4916 | return; |
4917 | |
4918 | mmap_event = (struct perf_mmap_event){ |
4919 | .vma = vma, |
4920 | /* .file_name */ |
4921 | /* .file_size */ |
4922 | .event_id = { |
4923 | .header = { |
4924 | .type = PERF_RECORD_MMAP, |
4925 | .misc = PERF_RECORD_MISC_USER, |
4926 | /* .size */ |
4927 | }, |
4928 | /* .pid */ |
4929 | /* .tid */ |
4930 | .start = vma->vm_start, |
4931 | .len = vma->vm_end - vma->vm_start, |
4932 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
4933 | }, |
4934 | }; |
4935 | |
4936 | perf_event_mmap_event(&mmap_event); |
4937 | } |
4938 | |
4939 | /* |
4940 | * IRQ throttle logging |
4941 | */ |
4942 | |
4943 | static void perf_log_throttle(struct perf_event *event, int enable) |
4944 | { |
4945 | struct perf_output_handle handle; |
4946 | struct perf_sample_data sample; |
4947 | int ret; |
4948 | |
4949 | struct { |
4950 | struct perf_event_header header; |
4951 | u64 time; |
4952 | u64 id; |
4953 | u64 stream_id; |
4954 | } throttle_event = { |
4955 | .header = { |
4956 | .type = PERF_RECORD_THROTTLE, |
4957 | .misc = 0, |
4958 | .size = sizeof(throttle_event), |
4959 | }, |
4960 | .time = perf_clock(), |
4961 | .id = primary_event_id(event), |
4962 | .stream_id = event->id, |
4963 | }; |
4964 | |
4965 | if (enable) |
4966 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
4967 | |
4968 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
4969 | |
4970 | ret = perf_output_begin(&handle, event, |
4971 | throttle_event.header.size, 1, 0); |
4972 | if (ret) |
4973 | return; |
4974 | |
4975 | perf_output_put(&handle, throttle_event); |
4976 | perf_event__output_id_sample(event, &handle, &sample); |
4977 | perf_output_end(&handle); |
4978 | } |
4979 | |
4980 | /* |
4981 | * Generic event overflow handling, sampling. |
4982 | */ |
4983 | |
4984 | static int __perf_event_overflow(struct perf_event *event, int nmi, |
4985 | int throttle, struct perf_sample_data *data, |
4986 | struct pt_regs *regs) |
4987 | { |
4988 | int events = atomic_read(&event->event_limit); |
4989 | struct hw_perf_event *hwc = &event->hw; |
4990 | int ret = 0; |
4991 | |
4992 | /* |
4993 | * Non-sampling counters might still use the PMI to fold short |
4994 | * hardware counters, ignore those. |
4995 | */ |
4996 | if (unlikely(!is_sampling_event(event))) |
4997 | return 0; |
4998 | |
4999 | if (unlikely(hwc->interrupts >= max_samples_per_tick)) { |
5000 | if (throttle) { |
5001 | hwc->interrupts = MAX_INTERRUPTS; |
5002 | perf_log_throttle(event, 0); |
5003 | ret = 1; |
5004 | } |
5005 | } else |
5006 | hwc->interrupts++; |
5007 | |
5008 | if (event->attr.freq) { |
5009 | u64 now = perf_clock(); |
5010 | s64 delta = now - hwc->freq_time_stamp; |
5011 | |
5012 | hwc->freq_time_stamp = now; |
5013 | |
5014 | if (delta > 0 && delta < 2*TICK_NSEC) |
5015 | perf_adjust_period(event, delta, hwc->last_period); |
5016 | } |
5017 | |
5018 | /* |
5019 | * XXX event_limit might not quite work as expected on inherited |
5020 | * events |
5021 | */ |
5022 | |
5023 | event->pending_kill = POLL_IN; |
5024 | if (events && atomic_dec_and_test(&event->event_limit)) { |
5025 | ret = 1; |
5026 | event->pending_kill = POLL_HUP; |
5027 | if (nmi) { |
5028 | event->pending_disable = 1; |
5029 | irq_work_queue(&event->pending); |
5030 | } else |
5031 | perf_event_disable(event); |
5032 | } |
5033 | |
5034 | if (event->overflow_handler) |
5035 | event->overflow_handler(event, nmi, data, regs); |
5036 | else |
5037 | perf_event_output(event, nmi, data, regs); |
5038 | |
5039 | return ret; |
5040 | } |
5041 | |
5042 | int perf_event_overflow(struct perf_event *event, int nmi, |
5043 | struct perf_sample_data *data, |
5044 | struct pt_regs *regs) |
5045 | { |
5046 | return __perf_event_overflow(event, nmi, 1, data, regs); |
5047 | } |
5048 | |
5049 | /* |
5050 | * Generic software event infrastructure |
5051 | */ |
5052 | |
5053 | struct swevent_htable { |
5054 | struct swevent_hlist *swevent_hlist; |
5055 | struct mutex hlist_mutex; |
5056 | int hlist_refcount; |
5057 | |
5058 | /* Recursion avoidance in each contexts */ |
5059 | int recursion[PERF_NR_CONTEXTS]; |
5060 | }; |
5061 | |
5062 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); |
5063 | |
5064 | /* |
5065 | * We directly increment event->count and keep a second value in |
5066 | * event->hw.period_left to count intervals. This period event |
5067 | * is kept in the range [-sample_period, 0] so that we can use the |
5068 | * sign as trigger. |
5069 | */ |
5070 | |
5071 | static u64 perf_swevent_set_period(struct perf_event *event) |
5072 | { |
5073 | struct hw_perf_event *hwc = &event->hw; |
5074 | u64 period = hwc->last_period; |
5075 | u64 nr, offset; |
5076 | s64 old, val; |
5077 | |
5078 | hwc->last_period = hwc->sample_period; |
5079 | |
5080 | again: |
5081 | old = val = local64_read(&hwc->period_left); |
5082 | if (val < 0) |
5083 | return 0; |
5084 | |
5085 | nr = div64_u64(period + val, period); |
5086 | offset = nr * period; |
5087 | val -= offset; |
5088 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
5089 | goto again; |
5090 | |
5091 | return nr; |
5092 | } |
5093 | |
5094 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
5095 | int nmi, struct perf_sample_data *data, |
5096 | struct pt_regs *regs) |
5097 | { |
5098 | struct hw_perf_event *hwc = &event->hw; |
5099 | int throttle = 0; |
5100 | |
5101 | data->period = event->hw.last_period; |
5102 | if (!overflow) |
5103 | overflow = perf_swevent_set_period(event); |
5104 | |
5105 | if (hwc->interrupts == MAX_INTERRUPTS) |
5106 | return; |
5107 | |
5108 | for (; overflow; overflow--) { |
5109 | if (__perf_event_overflow(event, nmi, throttle, |
5110 | data, regs)) { |
5111 | /* |
5112 | * We inhibit the overflow from happening when |
5113 | * hwc->interrupts == MAX_INTERRUPTS. |
5114 | */ |
5115 | break; |
5116 | } |
5117 | throttle = 1; |
5118 | } |
5119 | } |
5120 | |
5121 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
5122 | int nmi, struct perf_sample_data *data, |
5123 | struct pt_regs *regs) |
5124 | { |
5125 | struct hw_perf_event *hwc = &event->hw; |
5126 | |
5127 | local64_add(nr, &event->count); |
5128 | |
5129 | if (!regs) |
5130 | return; |
5131 | |
5132 | if (!is_sampling_event(event)) |
5133 | return; |
5134 | |
5135 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
5136 | return perf_swevent_overflow(event, 1, nmi, data, regs); |
5137 | |
5138 | if (local64_add_negative(nr, &hwc->period_left)) |
5139 | return; |
5140 | |
5141 | perf_swevent_overflow(event, 0, nmi, data, regs); |
5142 | } |
5143 | |
5144 | static int perf_exclude_event(struct perf_event *event, |
5145 | struct pt_regs *regs) |
5146 | { |
5147 | if (event->hw.state & PERF_HES_STOPPED) |
5148 | return 1; |
5149 | |
5150 | if (regs) { |
5151 | if (event->attr.exclude_user && user_mode(regs)) |
5152 | return 1; |
5153 | |
5154 | if (event->attr.exclude_kernel && !user_mode(regs)) |
5155 | return 1; |
5156 | } |
5157 | |
5158 | return 0; |
5159 | } |
5160 | |
5161 | static int perf_swevent_match(struct perf_event *event, |
5162 | enum perf_type_id type, |
5163 | u32 event_id, |
5164 | struct perf_sample_data *data, |
5165 | struct pt_regs *regs) |
5166 | { |
5167 | if (event->attr.type != type) |
5168 | return 0; |
5169 | |
5170 | if (event->attr.config != event_id) |
5171 | return 0; |
5172 | |
5173 | if (perf_exclude_event(event, regs)) |
5174 | return 0; |
5175 | |
5176 | return 1; |
5177 | } |
5178 | |
5179 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5180 | { |
5181 | u64 val = event_id | (type << 32); |
5182 | |
5183 | return hash_64(val, SWEVENT_HLIST_BITS); |
5184 | } |
5185 | |
5186 | static inline struct hlist_head * |
5187 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) |
5188 | { |
5189 | u64 hash = swevent_hash(type, event_id); |
5190 | |
5191 | return &hlist->heads[hash]; |
5192 | } |
5193 | |
5194 | /* For the read side: events when they trigger */ |
5195 | static inline struct hlist_head * |
5196 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
5197 | { |
5198 | struct swevent_hlist *hlist; |
5199 | |
5200 | hlist = rcu_dereference(swhash->swevent_hlist); |
5201 | if (!hlist) |
5202 | return NULL; |
5203 | |
5204 | return __find_swevent_head(hlist, type, event_id); |
5205 | } |
5206 | |
5207 | /* For the event head insertion and removal in the hlist */ |
5208 | static inline struct hlist_head * |
5209 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
5210 | { |
5211 | struct swevent_hlist *hlist; |
5212 | u32 event_id = event->attr.config; |
5213 | u64 type = event->attr.type; |
5214 | |
5215 | /* |
5216 | * Event scheduling is always serialized against hlist allocation |
5217 | * and release. Which makes the protected version suitable here. |
5218 | * The context lock guarantees that. |
5219 | */ |
5220 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
5221 | lockdep_is_held(&event->ctx->lock)); |
5222 | if (!hlist) |
5223 | return NULL; |
5224 | |
5225 | return __find_swevent_head(hlist, type, event_id); |
5226 | } |
5227 | |
5228 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, |
5229 | u64 nr, int nmi, |
5230 | struct perf_sample_data *data, |
5231 | struct pt_regs *regs) |
5232 | { |
5233 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
5234 | struct perf_event *event; |
5235 | struct hlist_node *node; |
5236 | struct hlist_head *head; |
5237 | |
5238 | rcu_read_lock(); |
5239 | head = find_swevent_head_rcu(swhash, type, event_id); |
5240 | if (!head) |
5241 | goto end; |
5242 | |
5243 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { |
5244 | if (perf_swevent_match(event, type, event_id, data, regs)) |
5245 | perf_swevent_event(event, nr, nmi, data, regs); |
5246 | } |
5247 | end: |
5248 | rcu_read_unlock(); |
5249 | } |
5250 | |
5251 | int perf_swevent_get_recursion_context(void) |
5252 | { |
5253 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
5254 | |
5255 | return get_recursion_context(swhash->recursion); |
5256 | } |
5257 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
5258 | |
5259 | inline void perf_swevent_put_recursion_context(int rctx) |
5260 | { |
5261 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
5262 | |
5263 | put_recursion_context(swhash->recursion, rctx); |
5264 | } |
5265 | |
5266 | void __perf_sw_event(u32 event_id, u64 nr, int nmi, |
5267 | struct pt_regs *regs, u64 addr) |
5268 | { |
5269 | struct perf_sample_data data; |
5270 | int rctx; |
5271 | |
5272 | preempt_disable_notrace(); |
5273 | rctx = perf_swevent_get_recursion_context(); |
5274 | if (rctx < 0) |
5275 | return; |
5276 | |
5277 | perf_sample_data_init(&data, addr); |
5278 | |
5279 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs); |
5280 | |
5281 | perf_swevent_put_recursion_context(rctx); |
5282 | preempt_enable_notrace(); |
5283 | } |
5284 | |
5285 | static void perf_swevent_read(struct perf_event *event) |
5286 | { |
5287 | } |
5288 | |
5289 | static int perf_swevent_add(struct perf_event *event, int flags) |
5290 | { |
5291 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
5292 | struct hw_perf_event *hwc = &event->hw; |
5293 | struct hlist_head *head; |
5294 | |
5295 | if (is_sampling_event(event)) { |
5296 | hwc->last_period = hwc->sample_period; |
5297 | perf_swevent_set_period(event); |
5298 | } |
5299 | |
5300 | hwc->state = !(flags & PERF_EF_START); |
5301 | |
5302 | head = find_swevent_head(swhash, event); |
5303 | if (WARN_ON_ONCE(!head)) |
5304 | return -EINVAL; |
5305 | |
5306 | hlist_add_head_rcu(&event->hlist_entry, head); |
5307 | |
5308 | return 0; |
5309 | } |
5310 | |
5311 | static void perf_swevent_del(struct perf_event *event, int flags) |
5312 | { |
5313 | hlist_del_rcu(&event->hlist_entry); |
5314 | } |
5315 | |
5316 | static void perf_swevent_start(struct perf_event *event, int flags) |
5317 | { |
5318 | event->hw.state = 0; |
5319 | } |
5320 | |
5321 | static void perf_swevent_stop(struct perf_event *event, int flags) |
5322 | { |
5323 | event->hw.state = PERF_HES_STOPPED; |
5324 | } |
5325 | |
5326 | /* Deref the hlist from the update side */ |
5327 | static inline struct swevent_hlist * |
5328 | swevent_hlist_deref(struct swevent_htable *swhash) |
5329 | { |
5330 | return rcu_dereference_protected(swhash->swevent_hlist, |
5331 | lockdep_is_held(&swhash->hlist_mutex)); |
5332 | } |
5333 | |
5334 | static void swevent_hlist_release_rcu(struct rcu_head *rcu_head) |
5335 | { |
5336 | struct swevent_hlist *hlist; |
5337 | |
5338 | hlist = container_of(rcu_head, struct swevent_hlist, rcu_head); |
5339 | kfree(hlist); |
5340 | } |
5341 | |
5342 | static void swevent_hlist_release(struct swevent_htable *swhash) |
5343 | { |
5344 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
5345 | |
5346 | if (!hlist) |
5347 | return; |
5348 | |
5349 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
5350 | call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu); |
5351 | } |
5352 | |
5353 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) |
5354 | { |
5355 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
5356 | |
5357 | mutex_lock(&swhash->hlist_mutex); |
5358 | |
5359 | if (!--swhash->hlist_refcount) |
5360 | swevent_hlist_release(swhash); |
5361 | |
5362 | mutex_unlock(&swhash->hlist_mutex); |
5363 | } |
5364 | |
5365 | static void swevent_hlist_put(struct perf_event *event) |
5366 | { |
5367 | int cpu; |
5368 | |
5369 | if (event->cpu != -1) { |
5370 | swevent_hlist_put_cpu(event, event->cpu); |
5371 | return; |
5372 | } |
5373 | |
5374 | for_each_possible_cpu(cpu) |
5375 | swevent_hlist_put_cpu(event, cpu); |
5376 | } |
5377 | |
5378 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) |
5379 | { |
5380 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
5381 | int err = 0; |
5382 | |
5383 | mutex_lock(&swhash->hlist_mutex); |
5384 | |
5385 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
5386 | struct swevent_hlist *hlist; |
5387 | |
5388 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); |
5389 | if (!hlist) { |
5390 | err = -ENOMEM; |
5391 | goto exit; |
5392 | } |
5393 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
5394 | } |
5395 | swhash->hlist_refcount++; |
5396 | exit: |
5397 | mutex_unlock(&swhash->hlist_mutex); |
5398 | |
5399 | return err; |
5400 | } |
5401 | |
5402 | static int swevent_hlist_get(struct perf_event *event) |
5403 | { |
5404 | int err; |
5405 | int cpu, failed_cpu; |
5406 | |
5407 | if (event->cpu != -1) |
5408 | return swevent_hlist_get_cpu(event, event->cpu); |
5409 | |
5410 | get_online_cpus(); |
5411 | for_each_possible_cpu(cpu) { |
5412 | err = swevent_hlist_get_cpu(event, cpu); |
5413 | if (err) { |
5414 | failed_cpu = cpu; |
5415 | goto fail; |
5416 | } |
5417 | } |
5418 | put_online_cpus(); |
5419 | |
5420 | return 0; |
5421 | fail: |
5422 | for_each_possible_cpu(cpu) { |
5423 | if (cpu == failed_cpu) |
5424 | break; |
5425 | swevent_hlist_put_cpu(event, cpu); |
5426 | } |
5427 | |
5428 | put_online_cpus(); |
5429 | return err; |
5430 | } |
5431 | |
5432 | atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
5433 | |
5434 | static void sw_perf_event_destroy(struct perf_event *event) |
5435 | { |
5436 | u64 event_id = event->attr.config; |
5437 | |
5438 | WARN_ON(event->parent); |
5439 | |
5440 | jump_label_dec(&perf_swevent_enabled[event_id]); |
5441 | swevent_hlist_put(event); |
5442 | } |
5443 | |
5444 | static int perf_swevent_init(struct perf_event *event) |
5445 | { |
5446 | int event_id = event->attr.config; |
5447 | |
5448 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5449 | return -ENOENT; |
5450 | |
5451 | switch (event_id) { |
5452 | case PERF_COUNT_SW_CPU_CLOCK: |
5453 | case PERF_COUNT_SW_TASK_CLOCK: |
5454 | return -ENOENT; |
5455 | |
5456 | default: |
5457 | break; |
5458 | } |
5459 | |
5460 | if (event_id >= PERF_COUNT_SW_MAX) |
5461 | return -ENOENT; |
5462 | |
5463 | if (!event->parent) { |
5464 | int err; |
5465 | |
5466 | err = swevent_hlist_get(event); |
5467 | if (err) |
5468 | return err; |
5469 | |
5470 | jump_label_inc(&perf_swevent_enabled[event_id]); |
5471 | event->destroy = sw_perf_event_destroy; |
5472 | } |
5473 | |
5474 | return 0; |
5475 | } |
5476 | |
5477 | static struct pmu perf_swevent = { |
5478 | .task_ctx_nr = perf_sw_context, |
5479 | |
5480 | .event_init = perf_swevent_init, |
5481 | .add = perf_swevent_add, |
5482 | .del = perf_swevent_del, |
5483 | .start = perf_swevent_start, |
5484 | .stop = perf_swevent_stop, |
5485 | .read = perf_swevent_read, |
5486 | }; |
5487 | |
5488 | #ifdef CONFIG_EVENT_TRACING |
5489 | |
5490 | static int perf_tp_filter_match(struct perf_event *event, |
5491 | struct perf_sample_data *data) |
5492 | { |
5493 | void *record = data->raw->data; |
5494 | |
5495 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) |
5496 | return 1; |
5497 | return 0; |
5498 | } |
5499 | |
5500 | static int perf_tp_event_match(struct perf_event *event, |
5501 | struct perf_sample_data *data, |
5502 | struct pt_regs *regs) |
5503 | { |
5504 | if (event->hw.state & PERF_HES_STOPPED) |
5505 | return 0; |
5506 | /* |
5507 | * All tracepoints are from kernel-space. |
5508 | */ |
5509 | if (event->attr.exclude_kernel) |
5510 | return 0; |
5511 | |
5512 | if (!perf_tp_filter_match(event, data)) |
5513 | return 0; |
5514 | |
5515 | return 1; |
5516 | } |
5517 | |
5518 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, |
5519 | struct pt_regs *regs, struct hlist_head *head, int rctx) |
5520 | { |
5521 | struct perf_sample_data data; |
5522 | struct perf_event *event; |
5523 | struct hlist_node *node; |
5524 | |
5525 | struct perf_raw_record raw = { |
5526 | .size = entry_size, |
5527 | .data = record, |
5528 | }; |
5529 | |
5530 | perf_sample_data_init(&data, addr); |
5531 | data.raw = &raw; |
5532 | |
5533 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { |
5534 | if (perf_tp_event_match(event, &data, regs)) |
5535 | perf_swevent_event(event, count, 1, &data, regs); |
5536 | } |
5537 | |
5538 | perf_swevent_put_recursion_context(rctx); |
5539 | } |
5540 | EXPORT_SYMBOL_GPL(perf_tp_event); |
5541 | |
5542 | static void tp_perf_event_destroy(struct perf_event *event) |
5543 | { |
5544 | perf_trace_destroy(event); |
5545 | } |
5546 | |
5547 | static int perf_tp_event_init(struct perf_event *event) |
5548 | { |
5549 | int err; |
5550 | |
5551 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5552 | return -ENOENT; |
5553 | |
5554 | err = perf_trace_init(event); |
5555 | if (err) |
5556 | return err; |
5557 | |
5558 | event->destroy = tp_perf_event_destroy; |
5559 | |
5560 | return 0; |
5561 | } |
5562 | |
5563 | static struct pmu perf_tracepoint = { |
5564 | .task_ctx_nr = perf_sw_context, |
5565 | |
5566 | .event_init = perf_tp_event_init, |
5567 | .add = perf_trace_add, |
5568 | .del = perf_trace_del, |
5569 | .start = perf_swevent_start, |
5570 | .stop = perf_swevent_stop, |
5571 | .read = perf_swevent_read, |
5572 | }; |
5573 | |
5574 | static inline void perf_tp_register(void) |
5575 | { |
5576 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
5577 | } |
5578 | |
5579 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) |
5580 | { |
5581 | char *filter_str; |
5582 | int ret; |
5583 | |
5584 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5585 | return -EINVAL; |
5586 | |
5587 | filter_str = strndup_user(arg, PAGE_SIZE); |
5588 | if (IS_ERR(filter_str)) |
5589 | return PTR_ERR(filter_str); |
5590 | |
5591 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); |
5592 | |
5593 | kfree(filter_str); |
5594 | return ret; |
5595 | } |
5596 | |
5597 | static void perf_event_free_filter(struct perf_event *event) |
5598 | { |
5599 | ftrace_profile_free_filter(event); |
5600 | } |
5601 | |
5602 | #else |
5603 | |
5604 | static inline void perf_tp_register(void) |
5605 | { |
5606 | } |
5607 | |
5608 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) |
5609 | { |
5610 | return -ENOENT; |
5611 | } |
5612 | |
5613 | static void perf_event_free_filter(struct perf_event *event) |
5614 | { |
5615 | } |
5616 | |
5617 | #endif /* CONFIG_EVENT_TRACING */ |
5618 | |
5619 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
5620 | void perf_bp_event(struct perf_event *bp, void *data) |
5621 | { |
5622 | struct perf_sample_data sample; |
5623 | struct pt_regs *regs = data; |
5624 | |
5625 | perf_sample_data_init(&sample, bp->attr.bp_addr); |
5626 | |
5627 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
5628 | perf_swevent_event(bp, 1, 1, &sample, regs); |
5629 | } |
5630 | #endif |
5631 | |
5632 | /* |
5633 | * hrtimer based swevent callback |
5634 | */ |
5635 | |
5636 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
5637 | { |
5638 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5639 | struct perf_sample_data data; |
5640 | struct pt_regs *regs; |
5641 | struct perf_event *event; |
5642 | u64 period; |
5643 | |
5644 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
5645 | |
5646 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
5647 | return HRTIMER_NORESTART; |
5648 | |
5649 | event->pmu->read(event); |
5650 | |
5651 | perf_sample_data_init(&data, 0); |
5652 | data.period = event->hw.last_period; |
5653 | regs = get_irq_regs(); |
5654 | |
5655 | if (regs && !perf_exclude_event(event, regs)) { |
5656 | if (!(event->attr.exclude_idle && current->pid == 0)) |
5657 | if (perf_event_overflow(event, 0, &data, regs)) |
5658 | ret = HRTIMER_NORESTART; |
5659 | } |
5660 | |
5661 | period = max_t(u64, 10000, event->hw.sample_period); |
5662 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); |
5663 | |
5664 | return ret; |
5665 | } |
5666 | |
5667 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5668 | { |
5669 | struct hw_perf_event *hwc = &event->hw; |
5670 | s64 period; |
5671 | |
5672 | if (!is_sampling_event(event)) |
5673 | return; |
5674 | |
5675 | period = local64_read(&hwc->period_left); |
5676 | if (period) { |
5677 | if (period < 0) |
5678 | period = 10000; |
5679 | |
5680 | local64_set(&hwc->period_left, 0); |
5681 | } else { |
5682 | period = max_t(u64, 10000, hwc->sample_period); |
5683 | } |
5684 | __hrtimer_start_range_ns(&hwc->hrtimer, |
5685 | ns_to_ktime(period), 0, |
5686 | HRTIMER_MODE_REL_PINNED, 0); |
5687 | } |
5688 | |
5689 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) |
5690 | { |
5691 | struct hw_perf_event *hwc = &event->hw; |
5692 | |
5693 | if (is_sampling_event(event)) { |
5694 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
5695 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
5696 | |
5697 | hrtimer_cancel(&hwc->hrtimer); |
5698 | } |
5699 | } |
5700 | |
5701 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5702 | { |
5703 | struct hw_perf_event *hwc = &event->hw; |
5704 | |
5705 | if (!is_sampling_event(event)) |
5706 | return; |
5707 | |
5708 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
5709 | hwc->hrtimer.function = perf_swevent_hrtimer; |
5710 | |
5711 | /* |
5712 | * Since hrtimers have a fixed rate, we can do a static freq->period |
5713 | * mapping and avoid the whole period adjust feedback stuff. |
5714 | */ |
5715 | if (event->attr.freq) { |
5716 | long freq = event->attr.sample_freq; |
5717 | |
5718 | event->attr.sample_period = NSEC_PER_SEC / freq; |
5719 | hwc->sample_period = event->attr.sample_period; |
5720 | local64_set(&hwc->period_left, hwc->sample_period); |
5721 | event->attr.freq = 0; |
5722 | } |
5723 | } |
5724 | |
5725 | /* |
5726 | * Software event: cpu wall time clock |
5727 | */ |
5728 | |
5729 | static void cpu_clock_event_update(struct perf_event *event) |
5730 | { |
5731 | s64 prev; |
5732 | u64 now; |
5733 | |
5734 | now = local_clock(); |
5735 | prev = local64_xchg(&event->hw.prev_count, now); |
5736 | local64_add(now - prev, &event->count); |
5737 | } |
5738 | |
5739 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
5740 | { |
5741 | local64_set(&event->hw.prev_count, local_clock()); |
5742 | perf_swevent_start_hrtimer(event); |
5743 | } |
5744 | |
5745 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
5746 | { |
5747 | perf_swevent_cancel_hrtimer(event); |
5748 | cpu_clock_event_update(event); |
5749 | } |
5750 | |
5751 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
5752 | { |
5753 | if (flags & PERF_EF_START) |
5754 | cpu_clock_event_start(event, flags); |
5755 | |
5756 | return 0; |
5757 | } |
5758 | |
5759 | static void cpu_clock_event_del(struct perf_event *event, int flags) |
5760 | { |
5761 | cpu_clock_event_stop(event, flags); |
5762 | } |
5763 | |
5764 | static void cpu_clock_event_read(struct perf_event *event) |
5765 | { |
5766 | cpu_clock_event_update(event); |
5767 | } |
5768 | |
5769 | static int cpu_clock_event_init(struct perf_event *event) |
5770 | { |
5771 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5772 | return -ENOENT; |
5773 | |
5774 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) |
5775 | return -ENOENT; |
5776 | |
5777 | perf_swevent_init_hrtimer(event); |
5778 | |
5779 | return 0; |
5780 | } |
5781 | |
5782 | static struct pmu perf_cpu_clock = { |
5783 | .task_ctx_nr = perf_sw_context, |
5784 | |
5785 | .event_init = cpu_clock_event_init, |
5786 | .add = cpu_clock_event_add, |
5787 | .del = cpu_clock_event_del, |
5788 | .start = cpu_clock_event_start, |
5789 | .stop = cpu_clock_event_stop, |
5790 | .read = cpu_clock_event_read, |
5791 | }; |
5792 | |
5793 | /* |
5794 | * Software event: task time clock |
5795 | */ |
5796 | |
5797 | static void task_clock_event_update(struct perf_event *event, u64 now) |
5798 | { |
5799 | u64 prev; |
5800 | s64 delta; |
5801 | |
5802 | prev = local64_xchg(&event->hw.prev_count, now); |
5803 | delta = now - prev; |
5804 | local64_add(delta, &event->count); |
5805 | } |
5806 | |
5807 | static void task_clock_event_start(struct perf_event *event, int flags) |
5808 | { |
5809 | local64_set(&event->hw.prev_count, event->ctx->time); |
5810 | perf_swevent_start_hrtimer(event); |
5811 | } |
5812 | |
5813 | static void task_clock_event_stop(struct perf_event *event, int flags) |
5814 | { |
5815 | perf_swevent_cancel_hrtimer(event); |
5816 | task_clock_event_update(event, event->ctx->time); |
5817 | } |
5818 | |
5819 | static int task_clock_event_add(struct perf_event *event, int flags) |
5820 | { |
5821 | if (flags & PERF_EF_START) |
5822 | task_clock_event_start(event, flags); |
5823 | |
5824 | return 0; |
5825 | } |
5826 | |
5827 | static void task_clock_event_del(struct perf_event *event, int flags) |
5828 | { |
5829 | task_clock_event_stop(event, PERF_EF_UPDATE); |
5830 | } |
5831 | |
5832 | static void task_clock_event_read(struct perf_event *event) |
5833 | { |
5834 | u64 now = perf_clock(); |
5835 | u64 delta = now - event->ctx->timestamp; |
5836 | u64 time = event->ctx->time + delta; |
5837 | |
5838 | task_clock_event_update(event, time); |
5839 | } |
5840 | |
5841 | static int task_clock_event_init(struct perf_event *event) |
5842 | { |
5843 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5844 | return -ENOENT; |
5845 | |
5846 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) |
5847 | return -ENOENT; |
5848 | |
5849 | perf_swevent_init_hrtimer(event); |
5850 | |
5851 | return 0; |
5852 | } |
5853 | |
5854 | static struct pmu perf_task_clock = { |
5855 | .task_ctx_nr = perf_sw_context, |
5856 | |
5857 | .event_init = task_clock_event_init, |
5858 | .add = task_clock_event_add, |
5859 | .del = task_clock_event_del, |
5860 | .start = task_clock_event_start, |
5861 | .stop = task_clock_event_stop, |
5862 | .read = task_clock_event_read, |
5863 | }; |
5864 | |
5865 | static void perf_pmu_nop_void(struct pmu *pmu) |
5866 | { |
5867 | } |
5868 | |
5869 | static int perf_pmu_nop_int(struct pmu *pmu) |
5870 | { |
5871 | return 0; |
5872 | } |
5873 | |
5874 | static void perf_pmu_start_txn(struct pmu *pmu) |
5875 | { |
5876 | perf_pmu_disable(pmu); |
5877 | } |
5878 | |
5879 | static int perf_pmu_commit_txn(struct pmu *pmu) |
5880 | { |
5881 | perf_pmu_enable(pmu); |
5882 | return 0; |
5883 | } |
5884 | |
5885 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
5886 | { |
5887 | perf_pmu_enable(pmu); |
5888 | } |
5889 | |
5890 | /* |
5891 | * Ensures all contexts with the same task_ctx_nr have the same |
5892 | * pmu_cpu_context too. |
5893 | */ |
5894 | static void *find_pmu_context(int ctxn) |
5895 | { |
5896 | struct pmu *pmu; |
5897 | |
5898 | if (ctxn < 0) |
5899 | return NULL; |
5900 | |
5901 | list_for_each_entry(pmu, &pmus, entry) { |
5902 | if (pmu->task_ctx_nr == ctxn) |
5903 | return pmu->pmu_cpu_context; |
5904 | } |
5905 | |
5906 | return NULL; |
5907 | } |
5908 | |
5909 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
5910 | { |
5911 | int cpu; |
5912 | |
5913 | for_each_possible_cpu(cpu) { |
5914 | struct perf_cpu_context *cpuctx; |
5915 | |
5916 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
5917 | |
5918 | if (cpuctx->active_pmu == old_pmu) |
5919 | cpuctx->active_pmu = pmu; |
5920 | } |
5921 | } |
5922 | |
5923 | static void free_pmu_context(struct pmu *pmu) |
5924 | { |
5925 | struct pmu *i; |
5926 | |
5927 | mutex_lock(&pmus_lock); |
5928 | /* |
5929 | * Like a real lame refcount. |
5930 | */ |
5931 | list_for_each_entry(i, &pmus, entry) { |
5932 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { |
5933 | update_pmu_context(i, pmu); |
5934 | goto out; |
5935 | } |
5936 | } |
5937 | |
5938 | free_percpu(pmu->pmu_cpu_context); |
5939 | out: |
5940 | mutex_unlock(&pmus_lock); |
5941 | } |
5942 | static struct idr pmu_idr; |
5943 | |
5944 | static ssize_t |
5945 | type_show(struct device *dev, struct device_attribute *attr, char *page) |
5946 | { |
5947 | struct pmu *pmu = dev_get_drvdata(dev); |
5948 | |
5949 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); |
5950 | } |
5951 | |
5952 | static struct device_attribute pmu_dev_attrs[] = { |
5953 | __ATTR_RO(type), |
5954 | __ATTR_NULL, |
5955 | }; |
5956 | |
5957 | static int pmu_bus_running; |
5958 | static struct bus_type pmu_bus = { |
5959 | .name = "event_source", |
5960 | .dev_attrs = pmu_dev_attrs, |
5961 | }; |
5962 | |
5963 | static void pmu_dev_release(struct device *dev) |
5964 | { |
5965 | kfree(dev); |
5966 | } |
5967 | |
5968 | static int pmu_dev_alloc(struct pmu *pmu) |
5969 | { |
5970 | int ret = -ENOMEM; |
5971 | |
5972 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); |
5973 | if (!pmu->dev) |
5974 | goto out; |
5975 | |
5976 | device_initialize(pmu->dev); |
5977 | ret = dev_set_name(pmu->dev, "%s", pmu->name); |
5978 | if (ret) |
5979 | goto free_dev; |
5980 | |
5981 | dev_set_drvdata(pmu->dev, pmu); |
5982 | pmu->dev->bus = &pmu_bus; |
5983 | pmu->dev->release = pmu_dev_release; |
5984 | ret = device_add(pmu->dev); |
5985 | if (ret) |
5986 | goto free_dev; |
5987 | |
5988 | out: |
5989 | return ret; |
5990 | |
5991 | free_dev: |
5992 | put_device(pmu->dev); |
5993 | goto out; |
5994 | } |
5995 | |
5996 | static struct lock_class_key cpuctx_mutex; |
5997 | |
5998 | int perf_pmu_register(struct pmu *pmu, char *name, int type) |
5999 | { |
6000 | int cpu, ret; |
6001 | |
6002 | mutex_lock(&pmus_lock); |
6003 | ret = -ENOMEM; |
6004 | pmu->pmu_disable_count = alloc_percpu(int); |
6005 | if (!pmu->pmu_disable_count) |
6006 | goto unlock; |
6007 | |
6008 | pmu->type = -1; |
6009 | if (!name) |
6010 | goto skip_type; |
6011 | pmu->name = name; |
6012 | |
6013 | if (type < 0) { |
6014 | int err = idr_pre_get(&pmu_idr, GFP_KERNEL); |
6015 | if (!err) |
6016 | goto free_pdc; |
6017 | |
6018 | err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type); |
6019 | if (err) { |
6020 | ret = err; |
6021 | goto free_pdc; |
6022 | } |
6023 | } |
6024 | pmu->type = type; |
6025 | |
6026 | if (pmu_bus_running) { |
6027 | ret = pmu_dev_alloc(pmu); |
6028 | if (ret) |
6029 | goto free_idr; |
6030 | } |
6031 | |
6032 | skip_type: |
6033 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6034 | if (pmu->pmu_cpu_context) |
6035 | goto got_cpu_context; |
6036 | |
6037 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6038 | if (!pmu->pmu_cpu_context) |
6039 | goto free_dev; |
6040 | |
6041 | for_each_possible_cpu(cpu) { |
6042 | struct perf_cpu_context *cpuctx; |
6043 | |
6044 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
6045 | __perf_event_init_context(&cpuctx->ctx); |
6046 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
6047 | cpuctx->ctx.type = cpu_context; |
6048 | cpuctx->ctx.pmu = pmu; |
6049 | cpuctx->jiffies_interval = 1; |
6050 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
6051 | cpuctx->active_pmu = pmu; |
6052 | } |
6053 | |
6054 | got_cpu_context: |
6055 | if (!pmu->start_txn) { |
6056 | if (pmu->pmu_enable) { |
6057 | /* |
6058 | * If we have pmu_enable/pmu_disable calls, install |
6059 | * transaction stubs that use that to try and batch |
6060 | * hardware accesses. |
6061 | */ |
6062 | pmu->start_txn = perf_pmu_start_txn; |
6063 | pmu->commit_txn = perf_pmu_commit_txn; |
6064 | pmu->cancel_txn = perf_pmu_cancel_txn; |
6065 | } else { |
6066 | pmu->start_txn = perf_pmu_nop_void; |
6067 | pmu->commit_txn = perf_pmu_nop_int; |
6068 | pmu->cancel_txn = perf_pmu_nop_void; |
6069 | } |
6070 | } |
6071 | |
6072 | if (!pmu->pmu_enable) { |
6073 | pmu->pmu_enable = perf_pmu_nop_void; |
6074 | pmu->pmu_disable = perf_pmu_nop_void; |
6075 | } |
6076 | |
6077 | list_add_rcu(&pmu->entry, &pmus); |
6078 | ret = 0; |
6079 | unlock: |
6080 | mutex_unlock(&pmus_lock); |
6081 | |
6082 | return ret; |
6083 | |
6084 | free_dev: |
6085 | device_del(pmu->dev); |
6086 | put_device(pmu->dev); |
6087 | |
6088 | free_idr: |
6089 | if (pmu->type >= PERF_TYPE_MAX) |
6090 | idr_remove(&pmu_idr, pmu->type); |
6091 | |
6092 | free_pdc: |
6093 | free_percpu(pmu->pmu_disable_count); |
6094 | goto unlock; |
6095 | } |
6096 | |
6097 | void perf_pmu_unregister(struct pmu *pmu) |
6098 | { |
6099 | mutex_lock(&pmus_lock); |
6100 | list_del_rcu(&pmu->entry); |
6101 | mutex_unlock(&pmus_lock); |
6102 | |
6103 | /* |
6104 | * We dereference the pmu list under both SRCU and regular RCU, so |
6105 | * synchronize against both of those. |
6106 | */ |
6107 | synchronize_srcu(&pmus_srcu); |
6108 | synchronize_rcu(); |
6109 | |
6110 | free_percpu(pmu->pmu_disable_count); |
6111 | if (pmu->type >= PERF_TYPE_MAX) |
6112 | idr_remove(&pmu_idr, pmu->type); |
6113 | device_del(pmu->dev); |
6114 | put_device(pmu->dev); |
6115 | free_pmu_context(pmu); |
6116 | } |
6117 | |
6118 | struct pmu *perf_init_event(struct perf_event *event) |
6119 | { |
6120 | struct pmu *pmu = NULL; |
6121 | int idx; |
6122 | int ret; |
6123 | |
6124 | idx = srcu_read_lock(&pmus_srcu); |
6125 | |
6126 | rcu_read_lock(); |
6127 | pmu = idr_find(&pmu_idr, event->attr.type); |
6128 | rcu_read_unlock(); |
6129 | if (pmu) { |
6130 | ret = pmu->event_init(event); |
6131 | if (ret) |
6132 | pmu = ERR_PTR(ret); |
6133 | goto unlock; |
6134 | } |
6135 | |
6136 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
6137 | ret = pmu->event_init(event); |
6138 | if (!ret) |
6139 | goto unlock; |
6140 | |
6141 | if (ret != -ENOENT) { |
6142 | pmu = ERR_PTR(ret); |
6143 | goto unlock; |
6144 | } |
6145 | } |
6146 | pmu = ERR_PTR(-ENOENT); |
6147 | unlock: |
6148 | srcu_read_unlock(&pmus_srcu, idx); |
6149 | |
6150 | return pmu; |
6151 | } |
6152 | |
6153 | /* |
6154 | * Allocate and initialize a event structure |
6155 | */ |
6156 | static struct perf_event * |
6157 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
6158 | struct task_struct *task, |
6159 | struct perf_event *group_leader, |
6160 | struct perf_event *parent_event, |
6161 | perf_overflow_handler_t overflow_handler) |
6162 | { |
6163 | struct pmu *pmu; |
6164 | struct perf_event *event; |
6165 | struct hw_perf_event *hwc; |
6166 | long err; |
6167 | |
6168 | if ((unsigned)cpu >= nr_cpu_ids) { |
6169 | if (!task || cpu != -1) |
6170 | return ERR_PTR(-EINVAL); |
6171 | } |
6172 | |
6173 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
6174 | if (!event) |
6175 | return ERR_PTR(-ENOMEM); |
6176 | |
6177 | /* |
6178 | * Single events are their own group leaders, with an |
6179 | * empty sibling list: |
6180 | */ |
6181 | if (!group_leader) |
6182 | group_leader = event; |
6183 | |
6184 | mutex_init(&event->child_mutex); |
6185 | INIT_LIST_HEAD(&event->child_list); |
6186 | |
6187 | INIT_LIST_HEAD(&event->group_entry); |
6188 | INIT_LIST_HEAD(&event->event_entry); |
6189 | INIT_LIST_HEAD(&event->sibling_list); |
6190 | init_waitqueue_head(&event->waitq); |
6191 | init_irq_work(&event->pending, perf_pending_event); |
6192 | |
6193 | mutex_init(&event->mmap_mutex); |
6194 | |
6195 | event->cpu = cpu; |
6196 | event->attr = *attr; |
6197 | event->group_leader = group_leader; |
6198 | event->pmu = NULL; |
6199 | event->oncpu = -1; |
6200 | |
6201 | event->parent = parent_event; |
6202 | |
6203 | event->ns = get_pid_ns(current->nsproxy->pid_ns); |
6204 | event->id = atomic64_inc_return(&perf_event_id); |
6205 | |
6206 | event->state = PERF_EVENT_STATE_INACTIVE; |
6207 | |
6208 | if (task) { |
6209 | event->attach_state = PERF_ATTACH_TASK; |
6210 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6211 | /* |
6212 | * hw_breakpoint is a bit difficult here.. |
6213 | */ |
6214 | if (attr->type == PERF_TYPE_BREAKPOINT) |
6215 | event->hw.bp_target = task; |
6216 | #endif |
6217 | } |
6218 | |
6219 | if (!overflow_handler && parent_event) |
6220 | overflow_handler = parent_event->overflow_handler; |
6221 | |
6222 | event->overflow_handler = overflow_handler; |
6223 | |
6224 | if (attr->disabled) |
6225 | event->state = PERF_EVENT_STATE_OFF; |
6226 | |
6227 | pmu = NULL; |
6228 | |
6229 | hwc = &event->hw; |
6230 | hwc->sample_period = attr->sample_period; |
6231 | if (attr->freq && attr->sample_freq) |
6232 | hwc->sample_period = 1; |
6233 | hwc->last_period = hwc->sample_period; |
6234 | |
6235 | local64_set(&hwc->period_left, hwc->sample_period); |
6236 | |
6237 | /* |
6238 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
6239 | */ |
6240 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
6241 | goto done; |
6242 | |
6243 | pmu = perf_init_event(event); |
6244 | |
6245 | done: |
6246 | err = 0; |
6247 | if (!pmu) |
6248 | err = -EINVAL; |
6249 | else if (IS_ERR(pmu)) |
6250 | err = PTR_ERR(pmu); |
6251 | |
6252 | if (err) { |
6253 | if (event->ns) |
6254 | put_pid_ns(event->ns); |
6255 | kfree(event); |
6256 | return ERR_PTR(err); |
6257 | } |
6258 | |
6259 | event->pmu = pmu; |
6260 | |
6261 | if (!event->parent) { |
6262 | if (event->attach_state & PERF_ATTACH_TASK) |
6263 | jump_label_inc(&perf_sched_events); |
6264 | if (event->attr.mmap || event->attr.mmap_data) |
6265 | atomic_inc(&nr_mmap_events); |
6266 | if (event->attr.comm) |
6267 | atomic_inc(&nr_comm_events); |
6268 | if (event->attr.task) |
6269 | atomic_inc(&nr_task_events); |
6270 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6271 | err = get_callchain_buffers(); |
6272 | if (err) { |
6273 | free_event(event); |
6274 | return ERR_PTR(err); |
6275 | } |
6276 | } |
6277 | } |
6278 | |
6279 | return event; |
6280 | } |
6281 | |
6282 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6283 | struct perf_event_attr *attr) |
6284 | { |
6285 | u32 size; |
6286 | int ret; |
6287 | |
6288 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) |
6289 | return -EFAULT; |
6290 | |
6291 | /* |
6292 | * zero the full structure, so that a short copy will be nice. |
6293 | */ |
6294 | memset(attr, 0, sizeof(*attr)); |
6295 | |
6296 | ret = get_user(size, &uattr->size); |
6297 | if (ret) |
6298 | return ret; |
6299 | |
6300 | if (size > PAGE_SIZE) /* silly large */ |
6301 | goto err_size; |
6302 | |
6303 | if (!size) /* abi compat */ |
6304 | size = PERF_ATTR_SIZE_VER0; |
6305 | |
6306 | if (size < PERF_ATTR_SIZE_VER0) |
6307 | goto err_size; |
6308 | |
6309 | /* |
6310 | * If we're handed a bigger struct than we know of, |
6311 | * ensure all the unknown bits are 0 - i.e. new |
6312 | * user-space does not rely on any kernel feature |
6313 | * extensions we dont know about yet. |
6314 | */ |
6315 | if (size > sizeof(*attr)) { |
6316 | unsigned char __user *addr; |
6317 | unsigned char __user *end; |
6318 | unsigned char val; |
6319 | |
6320 | addr = (void __user *)uattr + sizeof(*attr); |
6321 | end = (void __user *)uattr + size; |
6322 | |
6323 | for (; addr < end; addr++) { |
6324 | ret = get_user(val, addr); |
6325 | if (ret) |
6326 | return ret; |
6327 | if (val) |
6328 | goto err_size; |
6329 | } |
6330 | size = sizeof(*attr); |
6331 | } |
6332 | |
6333 | ret = copy_from_user(attr, uattr, size); |
6334 | if (ret) |
6335 | return -EFAULT; |
6336 | |
6337 | /* |
6338 | * If the type exists, the corresponding creation will verify |
6339 | * the attr->config. |
6340 | */ |
6341 | if (attr->type >= PERF_TYPE_MAX) |
6342 | return -EINVAL; |
6343 | |
6344 | if (attr->__reserved_1) |
6345 | return -EINVAL; |
6346 | |
6347 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) |
6348 | return -EINVAL; |
6349 | |
6350 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) |
6351 | return -EINVAL; |
6352 | |
6353 | out: |
6354 | return ret; |
6355 | |
6356 | err_size: |
6357 | put_user(sizeof(*attr), &uattr->size); |
6358 | ret = -E2BIG; |
6359 | goto out; |
6360 | } |
6361 | |
6362 | static int |
6363 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) |
6364 | { |
6365 | struct perf_buffer *buffer = NULL, *old_buffer = NULL; |
6366 | int ret = -EINVAL; |
6367 | |
6368 | if (!output_event) |
6369 | goto set; |
6370 | |
6371 | /* don't allow circular references */ |
6372 | if (event == output_event) |
6373 | goto out; |
6374 | |
6375 | /* |
6376 | * Don't allow cross-cpu buffers |
6377 | */ |
6378 | if (output_event->cpu != event->cpu) |
6379 | goto out; |
6380 | |
6381 | /* |
6382 | * If its not a per-cpu buffer, it must be the same task. |
6383 | */ |
6384 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) |
6385 | goto out; |
6386 | |
6387 | set: |
6388 | mutex_lock(&event->mmap_mutex); |
6389 | /* Can't redirect output if we've got an active mmap() */ |
6390 | if (atomic_read(&event->mmap_count)) |
6391 | goto unlock; |
6392 | |
6393 | if (output_event) { |
6394 | /* get the buffer we want to redirect to */ |
6395 | buffer = perf_buffer_get(output_event); |
6396 | if (!buffer) |
6397 | goto unlock; |
6398 | } |
6399 | |
6400 | old_buffer = event->buffer; |
6401 | rcu_assign_pointer(event->buffer, buffer); |
6402 | ret = 0; |
6403 | unlock: |
6404 | mutex_unlock(&event->mmap_mutex); |
6405 | |
6406 | if (old_buffer) |
6407 | perf_buffer_put(old_buffer); |
6408 | out: |
6409 | return ret; |
6410 | } |
6411 | |
6412 | /** |
6413 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
6414 | * |
6415 | * @attr_uptr: event_id type attributes for monitoring/sampling |
6416 | * @pid: target pid |
6417 | * @cpu: target cpu |
6418 | * @group_fd: group leader event fd |
6419 | */ |
6420 | SYSCALL_DEFINE5(perf_event_open, |
6421 | struct perf_event_attr __user *, attr_uptr, |
6422 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
6423 | { |
6424 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6425 | struct perf_event *event, *sibling; |
6426 | struct perf_event_attr attr; |
6427 | struct perf_event_context *ctx; |
6428 | struct file *event_file = NULL; |
6429 | struct file *group_file = NULL; |
6430 | struct task_struct *task = NULL; |
6431 | struct pmu *pmu; |
6432 | int event_fd; |
6433 | int move_group = 0; |
6434 | int fput_needed = 0; |
6435 | int err; |
6436 | |
6437 | /* for future expandability... */ |
6438 | if (flags & ~PERF_FLAG_ALL) |
6439 | return -EINVAL; |
6440 | |
6441 | err = perf_copy_attr(attr_uptr, &attr); |
6442 | if (err) |
6443 | return err; |
6444 | |
6445 | if (!attr.exclude_kernel) { |
6446 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
6447 | return -EACCES; |
6448 | } |
6449 | |
6450 | if (attr.freq) { |
6451 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
6452 | return -EINVAL; |
6453 | } |
6454 | |
6455 | /* |
6456 | * In cgroup mode, the pid argument is used to pass the fd |
6457 | * opened to the cgroup directory in cgroupfs. The cpu argument |
6458 | * designates the cpu on which to monitor threads from that |
6459 | * cgroup. |
6460 | */ |
6461 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) |
6462 | return -EINVAL; |
6463 | |
6464 | event_fd = get_unused_fd_flags(O_RDWR); |
6465 | if (event_fd < 0) |
6466 | return event_fd; |
6467 | |
6468 | if (group_fd != -1) { |
6469 | group_leader = perf_fget_light(group_fd, &fput_needed); |
6470 | if (IS_ERR(group_leader)) { |
6471 | err = PTR_ERR(group_leader); |
6472 | goto err_fd; |
6473 | } |
6474 | group_file = group_leader->filp; |
6475 | if (flags & PERF_FLAG_FD_OUTPUT) |
6476 | output_event = group_leader; |
6477 | if (flags & PERF_FLAG_FD_NO_GROUP) |
6478 | group_leader = NULL; |
6479 | } |
6480 | |
6481 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
6482 | task = find_lively_task_by_vpid(pid); |
6483 | if (IS_ERR(task)) { |
6484 | err = PTR_ERR(task); |
6485 | goto err_group_fd; |
6486 | } |
6487 | } |
6488 | |
6489 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL); |
6490 | if (IS_ERR(event)) { |
6491 | err = PTR_ERR(event); |
6492 | goto err_task; |
6493 | } |
6494 | |
6495 | if (flags & PERF_FLAG_PID_CGROUP) { |
6496 | err = perf_cgroup_connect(pid, event, &attr, group_leader); |
6497 | if (err) |
6498 | goto err_alloc; |
6499 | /* |
6500 | * one more event: |
6501 | * - that has cgroup constraint on event->cpu |
6502 | * - that may need work on context switch |
6503 | */ |
6504 | atomic_inc(&per_cpu(perf_cgroup_events, event->cpu)); |
6505 | jump_label_inc(&perf_sched_events); |
6506 | } |
6507 | |
6508 | /* |
6509 | * Special case software events and allow them to be part of |
6510 | * any hardware group. |
6511 | */ |
6512 | pmu = event->pmu; |
6513 | |
6514 | if (group_leader && |
6515 | (is_software_event(event) != is_software_event(group_leader))) { |
6516 | if (is_software_event(event)) { |
6517 | /* |
6518 | * If event and group_leader are not both a software |
6519 | * event, and event is, then group leader is not. |
6520 | * |
6521 | * Allow the addition of software events to !software |
6522 | * groups, this is safe because software events never |
6523 | * fail to schedule. |
6524 | */ |
6525 | pmu = group_leader->pmu; |
6526 | } else if (is_software_event(group_leader) && |
6527 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { |
6528 | /* |
6529 | * In case the group is a pure software group, and we |
6530 | * try to add a hardware event, move the whole group to |
6531 | * the hardware context. |
6532 | */ |
6533 | move_group = 1; |
6534 | } |
6535 | } |
6536 | |
6537 | /* |
6538 | * Get the target context (task or percpu): |
6539 | */ |
6540 | ctx = find_get_context(pmu, task, cpu); |
6541 | if (IS_ERR(ctx)) { |
6542 | err = PTR_ERR(ctx); |
6543 | goto err_alloc; |
6544 | } |
6545 | |
6546 | if (task) { |
6547 | put_task_struct(task); |
6548 | task = NULL; |
6549 | } |
6550 | |
6551 | /* |
6552 | * Look up the group leader (we will attach this event to it): |
6553 | */ |
6554 | if (group_leader) { |
6555 | err = -EINVAL; |
6556 | |
6557 | /* |
6558 | * Do not allow a recursive hierarchy (this new sibling |
6559 | * becoming part of another group-sibling): |
6560 | */ |
6561 | if (group_leader->group_leader != group_leader) |
6562 | goto err_context; |
6563 | /* |
6564 | * Do not allow to attach to a group in a different |
6565 | * task or CPU context: |
6566 | */ |
6567 | if (move_group) { |
6568 | if (group_leader->ctx->type != ctx->type) |
6569 | goto err_context; |
6570 | } else { |
6571 | if (group_leader->ctx != ctx) |
6572 | goto err_context; |
6573 | } |
6574 | |
6575 | /* |
6576 | * Only a group leader can be exclusive or pinned |
6577 | */ |
6578 | if (attr.exclusive || attr.pinned) |
6579 | goto err_context; |
6580 | } |
6581 | |
6582 | if (output_event) { |
6583 | err = perf_event_set_output(event, output_event); |
6584 | if (err) |
6585 | goto err_context; |
6586 | } |
6587 | |
6588 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
6589 | if (IS_ERR(event_file)) { |
6590 | err = PTR_ERR(event_file); |
6591 | goto err_context; |
6592 | } |
6593 | |
6594 | if (move_group) { |
6595 | struct perf_event_context *gctx = group_leader->ctx; |
6596 | |
6597 | mutex_lock(&gctx->mutex); |
6598 | perf_remove_from_context(group_leader); |
6599 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6600 | group_entry) { |
6601 | perf_remove_from_context(sibling); |
6602 | put_ctx(gctx); |
6603 | } |
6604 | mutex_unlock(&gctx->mutex); |
6605 | put_ctx(gctx); |
6606 | } |
6607 | |
6608 | event->filp = event_file; |
6609 | WARN_ON_ONCE(ctx->parent_ctx); |
6610 | mutex_lock(&ctx->mutex); |
6611 | |
6612 | if (move_group) { |
6613 | perf_install_in_context(ctx, group_leader, cpu); |
6614 | get_ctx(ctx); |
6615 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6616 | group_entry) { |
6617 | perf_install_in_context(ctx, sibling, cpu); |
6618 | get_ctx(ctx); |
6619 | } |
6620 | } |
6621 | |
6622 | perf_install_in_context(ctx, event, cpu); |
6623 | ++ctx->generation; |
6624 | perf_unpin_context(ctx); |
6625 | mutex_unlock(&ctx->mutex); |
6626 | |
6627 | event->owner = current; |
6628 | |
6629 | mutex_lock(¤t->perf_event_mutex); |
6630 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); |
6631 | mutex_unlock(¤t->perf_event_mutex); |
6632 | |
6633 | /* |
6634 | * Precalculate sample_data sizes |
6635 | */ |
6636 | perf_event__header_size(event); |
6637 | perf_event__id_header_size(event); |
6638 | |
6639 | /* |
6640 | * Drop the reference on the group_event after placing the |
6641 | * new event on the sibling_list. This ensures destruction |
6642 | * of the group leader will find the pointer to itself in |
6643 | * perf_group_detach(). |
6644 | */ |
6645 | fput_light(group_file, fput_needed); |
6646 | fd_install(event_fd, event_file); |
6647 | return event_fd; |
6648 | |
6649 | err_context: |
6650 | perf_unpin_context(ctx); |
6651 | put_ctx(ctx); |
6652 | err_alloc: |
6653 | free_event(event); |
6654 | err_task: |
6655 | if (task) |
6656 | put_task_struct(task); |
6657 | err_group_fd: |
6658 | fput_light(group_file, fput_needed); |
6659 | err_fd: |
6660 | put_unused_fd(event_fd); |
6661 | return err; |
6662 | } |
6663 | |
6664 | /** |
6665 | * perf_event_create_kernel_counter |
6666 | * |
6667 | * @attr: attributes of the counter to create |
6668 | * @cpu: cpu in which the counter is bound |
6669 | * @task: task to profile (NULL for percpu) |
6670 | */ |
6671 | struct perf_event * |
6672 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, |
6673 | struct task_struct *task, |
6674 | perf_overflow_handler_t overflow_handler) |
6675 | { |
6676 | struct perf_event_context *ctx; |
6677 | struct perf_event *event; |
6678 | int err; |
6679 | |
6680 | /* |
6681 | * Get the target context (task or percpu): |
6682 | */ |
6683 | |
6684 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler); |
6685 | if (IS_ERR(event)) { |
6686 | err = PTR_ERR(event); |
6687 | goto err; |
6688 | } |
6689 | |
6690 | ctx = find_get_context(event->pmu, task, cpu); |
6691 | if (IS_ERR(ctx)) { |
6692 | err = PTR_ERR(ctx); |
6693 | goto err_free; |
6694 | } |
6695 | |
6696 | event->filp = NULL; |
6697 | WARN_ON_ONCE(ctx->parent_ctx); |
6698 | mutex_lock(&ctx->mutex); |
6699 | perf_install_in_context(ctx, event, cpu); |
6700 | ++ctx->generation; |
6701 | perf_unpin_context(ctx); |
6702 | mutex_unlock(&ctx->mutex); |
6703 | |
6704 | return event; |
6705 | |
6706 | err_free: |
6707 | free_event(event); |
6708 | err: |
6709 | return ERR_PTR(err); |
6710 | } |
6711 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
6712 | |
6713 | static void sync_child_event(struct perf_event *child_event, |
6714 | struct task_struct *child) |
6715 | { |
6716 | struct perf_event *parent_event = child_event->parent; |
6717 | u64 child_val; |
6718 | |
6719 | if (child_event->attr.inherit_stat) |
6720 | perf_event_read_event(child_event, child); |
6721 | |
6722 | child_val = perf_event_count(child_event); |
6723 | |
6724 | /* |
6725 | * Add back the child's count to the parent's count: |
6726 | */ |
6727 | atomic64_add(child_val, &parent_event->child_count); |
6728 | atomic64_add(child_event->total_time_enabled, |
6729 | &parent_event->child_total_time_enabled); |
6730 | atomic64_add(child_event->total_time_running, |
6731 | &parent_event->child_total_time_running); |
6732 | |
6733 | /* |
6734 | * Remove this event from the parent's list |
6735 | */ |
6736 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
6737 | mutex_lock(&parent_event->child_mutex); |
6738 | list_del_init(&child_event->child_list); |
6739 | mutex_unlock(&parent_event->child_mutex); |
6740 | |
6741 | /* |
6742 | * Release the parent event, if this was the last |
6743 | * reference to it. |
6744 | */ |
6745 | fput(parent_event->filp); |
6746 | } |
6747 | |
6748 | static void |
6749 | __perf_event_exit_task(struct perf_event *child_event, |
6750 | struct perf_event_context *child_ctx, |
6751 | struct task_struct *child) |
6752 | { |
6753 | if (child_event->parent) { |
6754 | raw_spin_lock_irq(&child_ctx->lock); |
6755 | perf_group_detach(child_event); |
6756 | raw_spin_unlock_irq(&child_ctx->lock); |
6757 | } |
6758 | |
6759 | perf_remove_from_context(child_event); |
6760 | |
6761 | /* |
6762 | * It can happen that the parent exits first, and has events |
6763 | * that are still around due to the child reference. These |
6764 | * events need to be zapped. |
6765 | */ |
6766 | if (child_event->parent) { |
6767 | sync_child_event(child_event, child); |
6768 | free_event(child_event); |
6769 | } |
6770 | } |
6771 | |
6772 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
6773 | { |
6774 | struct perf_event *child_event, *tmp; |
6775 | struct perf_event_context *child_ctx; |
6776 | unsigned long flags; |
6777 | |
6778 | if (likely(!child->perf_event_ctxp[ctxn])) { |
6779 | perf_event_task(child, NULL, 0); |
6780 | return; |
6781 | } |
6782 | |
6783 | local_irq_save(flags); |
6784 | /* |
6785 | * We can't reschedule here because interrupts are disabled, |
6786 | * and either child is current or it is a task that can't be |
6787 | * scheduled, so we are now safe from rescheduling changing |
6788 | * our context. |
6789 | */ |
6790 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
6791 | task_ctx_sched_out(child_ctx, EVENT_ALL); |
6792 | |
6793 | /* |
6794 | * Take the context lock here so that if find_get_context is |
6795 | * reading child->perf_event_ctxp, we wait until it has |
6796 | * incremented the context's refcount before we do put_ctx below. |
6797 | */ |
6798 | raw_spin_lock(&child_ctx->lock); |
6799 | child->perf_event_ctxp[ctxn] = NULL; |
6800 | /* |
6801 | * If this context is a clone; unclone it so it can't get |
6802 | * swapped to another process while we're removing all |
6803 | * the events from it. |
6804 | */ |
6805 | unclone_ctx(child_ctx); |
6806 | update_context_time(child_ctx); |
6807 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
6808 | |
6809 | /* |
6810 | * Report the task dead after unscheduling the events so that we |
6811 | * won't get any samples after PERF_RECORD_EXIT. We can however still |
6812 | * get a few PERF_RECORD_READ events. |
6813 | */ |
6814 | perf_event_task(child, child_ctx, 0); |
6815 | |
6816 | /* |
6817 | * We can recurse on the same lock type through: |
6818 | * |
6819 | * __perf_event_exit_task() |
6820 | * sync_child_event() |
6821 | * fput(parent_event->filp) |
6822 | * perf_release() |
6823 | * mutex_lock(&ctx->mutex) |
6824 | * |
6825 | * But since its the parent context it won't be the same instance. |
6826 | */ |
6827 | mutex_lock(&child_ctx->mutex); |
6828 | |
6829 | again: |
6830 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
6831 | group_entry) |
6832 | __perf_event_exit_task(child_event, child_ctx, child); |
6833 | |
6834 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, |
6835 | group_entry) |
6836 | __perf_event_exit_task(child_event, child_ctx, child); |
6837 | |
6838 | /* |
6839 | * If the last event was a group event, it will have appended all |
6840 | * its siblings to the list, but we obtained 'tmp' before that which |
6841 | * will still point to the list head terminating the iteration. |
6842 | */ |
6843 | if (!list_empty(&child_ctx->pinned_groups) || |
6844 | !list_empty(&child_ctx->flexible_groups)) |
6845 | goto again; |
6846 | |
6847 | mutex_unlock(&child_ctx->mutex); |
6848 | |
6849 | put_ctx(child_ctx); |
6850 | } |
6851 | |
6852 | /* |
6853 | * When a child task exits, feed back event values to parent events. |
6854 | */ |
6855 | void perf_event_exit_task(struct task_struct *child) |
6856 | { |
6857 | struct perf_event *event, *tmp; |
6858 | int ctxn; |
6859 | |
6860 | mutex_lock(&child->perf_event_mutex); |
6861 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, |
6862 | owner_entry) { |
6863 | list_del_init(&event->owner_entry); |
6864 | |
6865 | /* |
6866 | * Ensure the list deletion is visible before we clear |
6867 | * the owner, closes a race against perf_release() where |
6868 | * we need to serialize on the owner->perf_event_mutex. |
6869 | */ |
6870 | smp_wmb(); |
6871 | event->owner = NULL; |
6872 | } |
6873 | mutex_unlock(&child->perf_event_mutex); |
6874 | |
6875 | for_each_task_context_nr(ctxn) |
6876 | perf_event_exit_task_context(child, ctxn); |
6877 | } |
6878 | |
6879 | static void perf_free_event(struct perf_event *event, |
6880 | struct perf_event_context *ctx) |
6881 | { |
6882 | struct perf_event *parent = event->parent; |
6883 | |
6884 | if (WARN_ON_ONCE(!parent)) |
6885 | return; |
6886 | |
6887 | mutex_lock(&parent->child_mutex); |
6888 | list_del_init(&event->child_list); |
6889 | mutex_unlock(&parent->child_mutex); |
6890 | |
6891 | fput(parent->filp); |
6892 | |
6893 | perf_group_detach(event); |
6894 | list_del_event(event, ctx); |
6895 | free_event(event); |
6896 | } |
6897 | |
6898 | /* |
6899 | * free an unexposed, unused context as created by inheritance by |
6900 | * perf_event_init_task below, used by fork() in case of fail. |
6901 | */ |
6902 | void perf_event_free_task(struct task_struct *task) |
6903 | { |
6904 | struct perf_event_context *ctx; |
6905 | struct perf_event *event, *tmp; |
6906 | int ctxn; |
6907 | |
6908 | for_each_task_context_nr(ctxn) { |
6909 | ctx = task->perf_event_ctxp[ctxn]; |
6910 | if (!ctx) |
6911 | continue; |
6912 | |
6913 | mutex_lock(&ctx->mutex); |
6914 | again: |
6915 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
6916 | group_entry) |
6917 | perf_free_event(event, ctx); |
6918 | |
6919 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
6920 | group_entry) |
6921 | perf_free_event(event, ctx); |
6922 | |
6923 | if (!list_empty(&ctx->pinned_groups) || |
6924 | !list_empty(&ctx->flexible_groups)) |
6925 | goto again; |
6926 | |
6927 | mutex_unlock(&ctx->mutex); |
6928 | |
6929 | put_ctx(ctx); |
6930 | } |
6931 | } |
6932 | |
6933 | void perf_event_delayed_put(struct task_struct *task) |
6934 | { |
6935 | int ctxn; |
6936 | |
6937 | for_each_task_context_nr(ctxn) |
6938 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); |
6939 | } |
6940 | |
6941 | /* |
6942 | * inherit a event from parent task to child task: |
6943 | */ |
6944 | static struct perf_event * |
6945 | inherit_event(struct perf_event *parent_event, |
6946 | struct task_struct *parent, |
6947 | struct perf_event_context *parent_ctx, |
6948 | struct task_struct *child, |
6949 | struct perf_event *group_leader, |
6950 | struct perf_event_context *child_ctx) |
6951 | { |
6952 | struct perf_event *child_event; |
6953 | unsigned long flags; |
6954 | |
6955 | /* |
6956 | * Instead of creating recursive hierarchies of events, |
6957 | * we link inherited events back to the original parent, |
6958 | * which has a filp for sure, which we use as the reference |
6959 | * count: |
6960 | */ |
6961 | if (parent_event->parent) |
6962 | parent_event = parent_event->parent; |
6963 | |
6964 | child_event = perf_event_alloc(&parent_event->attr, |
6965 | parent_event->cpu, |
6966 | child, |
6967 | group_leader, parent_event, |
6968 | NULL); |
6969 | if (IS_ERR(child_event)) |
6970 | return child_event; |
6971 | get_ctx(child_ctx); |
6972 | |
6973 | /* |
6974 | * Make the child state follow the state of the parent event, |
6975 | * not its attr.disabled bit. We hold the parent's mutex, |
6976 | * so we won't race with perf_event_{en, dis}able_family. |
6977 | */ |
6978 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) |
6979 | child_event->state = PERF_EVENT_STATE_INACTIVE; |
6980 | else |
6981 | child_event->state = PERF_EVENT_STATE_OFF; |
6982 | |
6983 | if (parent_event->attr.freq) { |
6984 | u64 sample_period = parent_event->hw.sample_period; |
6985 | struct hw_perf_event *hwc = &child_event->hw; |
6986 | |
6987 | hwc->sample_period = sample_period; |
6988 | hwc->last_period = sample_period; |
6989 | |
6990 | local64_set(&hwc->period_left, sample_period); |
6991 | } |
6992 | |
6993 | child_event->ctx = child_ctx; |
6994 | child_event->overflow_handler = parent_event->overflow_handler; |
6995 | |
6996 | /* |
6997 | * Precalculate sample_data sizes |
6998 | */ |
6999 | perf_event__header_size(child_event); |
7000 | perf_event__id_header_size(child_event); |
7001 | |
7002 | /* |
7003 | * Link it up in the child's context: |
7004 | */ |
7005 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
7006 | add_event_to_ctx(child_event, child_ctx); |
7007 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
7008 | |
7009 | /* |
7010 | * Get a reference to the parent filp - we will fput it |
7011 | * when the child event exits. This is safe to do because |
7012 | * we are in the parent and we know that the filp still |
7013 | * exists and has a nonzero count: |
7014 | */ |
7015 | atomic_long_inc(&parent_event->filp->f_count); |
7016 | |
7017 | /* |
7018 | * Link this into the parent event's child list |
7019 | */ |
7020 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7021 | mutex_lock(&parent_event->child_mutex); |
7022 | list_add_tail(&child_event->child_list, &parent_event->child_list); |
7023 | mutex_unlock(&parent_event->child_mutex); |
7024 | |
7025 | return child_event; |
7026 | } |
7027 | |
7028 | static int inherit_group(struct perf_event *parent_event, |
7029 | struct task_struct *parent, |
7030 | struct perf_event_context *parent_ctx, |
7031 | struct task_struct *child, |
7032 | struct perf_event_context *child_ctx) |
7033 | { |
7034 | struct perf_event *leader; |
7035 | struct perf_event *sub; |
7036 | struct perf_event *child_ctr; |
7037 | |
7038 | leader = inherit_event(parent_event, parent, parent_ctx, |
7039 | child, NULL, child_ctx); |
7040 | if (IS_ERR(leader)) |
7041 | return PTR_ERR(leader); |
7042 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { |
7043 | child_ctr = inherit_event(sub, parent, parent_ctx, |
7044 | child, leader, child_ctx); |
7045 | if (IS_ERR(child_ctr)) |
7046 | return PTR_ERR(child_ctr); |
7047 | } |
7048 | return 0; |
7049 | } |
7050 | |
7051 | static int |
7052 | inherit_task_group(struct perf_event *event, struct task_struct *parent, |
7053 | struct perf_event_context *parent_ctx, |
7054 | struct task_struct *child, int ctxn, |
7055 | int *inherited_all) |
7056 | { |
7057 | int ret; |
7058 | struct perf_event_context *child_ctx; |
7059 | |
7060 | if (!event->attr.inherit) { |
7061 | *inherited_all = 0; |
7062 | return 0; |
7063 | } |
7064 | |
7065 | child_ctx = child->perf_event_ctxp[ctxn]; |
7066 | if (!child_ctx) { |
7067 | /* |
7068 | * This is executed from the parent task context, so |
7069 | * inherit events that have been marked for cloning. |
7070 | * First allocate and initialize a context for the |
7071 | * child. |
7072 | */ |
7073 | |
7074 | child_ctx = alloc_perf_context(event->pmu, child); |
7075 | if (!child_ctx) |
7076 | return -ENOMEM; |
7077 | |
7078 | child->perf_event_ctxp[ctxn] = child_ctx; |
7079 | } |
7080 | |
7081 | ret = inherit_group(event, parent, parent_ctx, |
7082 | child, child_ctx); |
7083 | |
7084 | if (ret) |
7085 | *inherited_all = 0; |
7086 | |
7087 | return ret; |
7088 | } |
7089 | |
7090 | /* |
7091 | * Initialize the perf_event context in task_struct |
7092 | */ |
7093 | int perf_event_init_context(struct task_struct *child, int ctxn) |
7094 | { |
7095 | struct perf_event_context *child_ctx, *parent_ctx; |
7096 | struct perf_event_context *cloned_ctx; |
7097 | struct perf_event *event; |
7098 | struct task_struct *parent = current; |
7099 | int inherited_all = 1; |
7100 | unsigned long flags; |
7101 | int ret = 0; |
7102 | |
7103 | if (likely(!parent->perf_event_ctxp[ctxn])) |
7104 | return 0; |
7105 | |
7106 | /* |
7107 | * If the parent's context is a clone, pin it so it won't get |
7108 | * swapped under us. |
7109 | */ |
7110 | parent_ctx = perf_pin_task_context(parent, ctxn); |
7111 | |
7112 | /* |
7113 | * No need to check if parent_ctx != NULL here; since we saw |
7114 | * it non-NULL earlier, the only reason for it to become NULL |
7115 | * is if we exit, and since we're currently in the middle of |
7116 | * a fork we can't be exiting at the same time. |
7117 | */ |
7118 | |
7119 | /* |
7120 | * Lock the parent list. No need to lock the child - not PID |
7121 | * hashed yet and not running, so nobody can access it. |
7122 | */ |
7123 | mutex_lock(&parent_ctx->mutex); |
7124 | |
7125 | /* |
7126 | * We dont have to disable NMIs - we are only looking at |
7127 | * the list, not manipulating it: |
7128 | */ |
7129 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
7130 | ret = inherit_task_group(event, parent, parent_ctx, |
7131 | child, ctxn, &inherited_all); |
7132 | if (ret) |
7133 | break; |
7134 | } |
7135 | |
7136 | /* |
7137 | * We can't hold ctx->lock when iterating the ->flexible_group list due |
7138 | * to allocations, but we need to prevent rotation because |
7139 | * rotate_ctx() will change the list from interrupt context. |
7140 | */ |
7141 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7142 | parent_ctx->rotate_disable = 1; |
7143 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
7144 | |
7145 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
7146 | ret = inherit_task_group(event, parent, parent_ctx, |
7147 | child, ctxn, &inherited_all); |
7148 | if (ret) |
7149 | break; |
7150 | } |
7151 | |
7152 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7153 | parent_ctx->rotate_disable = 0; |
7154 | |
7155 | child_ctx = child->perf_event_ctxp[ctxn]; |
7156 | |
7157 | if (child_ctx && inherited_all) { |
7158 | /* |
7159 | * Mark the child context as a clone of the parent |
7160 | * context, or of whatever the parent is a clone of. |
7161 | * |
7162 | * Note that if the parent is a clone, the holding of |
7163 | * parent_ctx->lock avoids it from being uncloned. |
7164 | */ |
7165 | cloned_ctx = parent_ctx->parent_ctx; |
7166 | if (cloned_ctx) { |
7167 | child_ctx->parent_ctx = cloned_ctx; |
7168 | child_ctx->parent_gen = parent_ctx->parent_gen; |
7169 | } else { |
7170 | child_ctx->parent_ctx = parent_ctx; |
7171 | child_ctx->parent_gen = parent_ctx->generation; |
7172 | } |
7173 | get_ctx(child_ctx->parent_ctx); |
7174 | } |
7175 | |
7176 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
7177 | mutex_unlock(&parent_ctx->mutex); |
7178 | |
7179 | perf_unpin_context(parent_ctx); |
7180 | put_ctx(parent_ctx); |
7181 | |
7182 | return ret; |
7183 | } |
7184 | |
7185 | /* |
7186 | * Initialize the perf_event context in task_struct |
7187 | */ |
7188 | int perf_event_init_task(struct task_struct *child) |
7189 | { |
7190 | int ctxn, ret; |
7191 | |
7192 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7193 | mutex_init(&child->perf_event_mutex); |
7194 | INIT_LIST_HEAD(&child->perf_event_list); |
7195 | |
7196 | for_each_task_context_nr(ctxn) { |
7197 | ret = perf_event_init_context(child, ctxn); |
7198 | if (ret) |
7199 | return ret; |
7200 | } |
7201 | |
7202 | return 0; |
7203 | } |
7204 | |
7205 | static void __init perf_event_init_all_cpus(void) |
7206 | { |
7207 | struct swevent_htable *swhash; |
7208 | int cpu; |
7209 | |
7210 | for_each_possible_cpu(cpu) { |
7211 | swhash = &per_cpu(swevent_htable, cpu); |
7212 | mutex_init(&swhash->hlist_mutex); |
7213 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
7214 | } |
7215 | } |
7216 | |
7217 | static void __cpuinit perf_event_init_cpu(int cpu) |
7218 | { |
7219 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
7220 | |
7221 | mutex_lock(&swhash->hlist_mutex); |
7222 | if (swhash->hlist_refcount > 0) { |
7223 | struct swevent_hlist *hlist; |
7224 | |
7225 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7226 | WARN_ON(!hlist); |
7227 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
7228 | } |
7229 | mutex_unlock(&swhash->hlist_mutex); |
7230 | } |
7231 | |
7232 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
7233 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
7234 | { |
7235 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7236 | |
7237 | WARN_ON(!irqs_disabled()); |
7238 | |
7239 | list_del_init(&cpuctx->rotation_list); |
7240 | } |
7241 | |
7242 | static void __perf_event_exit_context(void *__info) |
7243 | { |
7244 | struct perf_event_context *ctx = __info; |
7245 | struct perf_event *event, *tmp; |
7246 | |
7247 | perf_pmu_rotate_stop(ctx->pmu); |
7248 | |
7249 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
7250 | __perf_remove_from_context(event); |
7251 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
7252 | __perf_remove_from_context(event); |
7253 | } |
7254 | |
7255 | static void perf_event_exit_cpu_context(int cpu) |
7256 | { |
7257 | struct perf_event_context *ctx; |
7258 | struct pmu *pmu; |
7259 | int idx; |
7260 | |
7261 | idx = srcu_read_lock(&pmus_srcu); |
7262 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7263 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
7264 | |
7265 | mutex_lock(&ctx->mutex); |
7266 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); |
7267 | mutex_unlock(&ctx->mutex); |
7268 | } |
7269 | srcu_read_unlock(&pmus_srcu, idx); |
7270 | } |
7271 | |
7272 | static void perf_event_exit_cpu(int cpu) |
7273 | { |
7274 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
7275 | |
7276 | mutex_lock(&swhash->hlist_mutex); |
7277 | swevent_hlist_release(swhash); |
7278 | mutex_unlock(&swhash->hlist_mutex); |
7279 | |
7280 | perf_event_exit_cpu_context(cpu); |
7281 | } |
7282 | #else |
7283 | static inline void perf_event_exit_cpu(int cpu) { } |
7284 | #endif |
7285 | |
7286 | static int |
7287 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) |
7288 | { |
7289 | int cpu; |
7290 | |
7291 | for_each_online_cpu(cpu) |
7292 | perf_event_exit_cpu(cpu); |
7293 | |
7294 | return NOTIFY_OK; |
7295 | } |
7296 | |
7297 | /* |
7298 | * Run the perf reboot notifier at the very last possible moment so that |
7299 | * the generic watchdog code runs as long as possible. |
7300 | */ |
7301 | static struct notifier_block perf_reboot_notifier = { |
7302 | .notifier_call = perf_reboot, |
7303 | .priority = INT_MIN, |
7304 | }; |
7305 | |
7306 | static int __cpuinit |
7307 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
7308 | { |
7309 | unsigned int cpu = (long)hcpu; |
7310 | |
7311 | switch (action & ~CPU_TASKS_FROZEN) { |
7312 | |
7313 | case CPU_UP_PREPARE: |
7314 | case CPU_DOWN_FAILED: |
7315 | perf_event_init_cpu(cpu); |
7316 | break; |
7317 | |
7318 | case CPU_UP_CANCELED: |
7319 | case CPU_DOWN_PREPARE: |
7320 | perf_event_exit_cpu(cpu); |
7321 | break; |
7322 | |
7323 | default: |
7324 | break; |
7325 | } |
7326 | |
7327 | return NOTIFY_OK; |
7328 | } |
7329 | |
7330 | void __init perf_event_init(void) |
7331 | { |
7332 | int ret; |
7333 | |
7334 | idr_init(&pmu_idr); |
7335 | |
7336 | perf_event_init_all_cpus(); |
7337 | init_srcu_struct(&pmus_srcu); |
7338 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7339 | perf_pmu_register(&perf_cpu_clock, NULL, -1); |
7340 | perf_pmu_register(&perf_task_clock, NULL, -1); |
7341 | perf_tp_register(); |
7342 | perf_cpu_notifier(perf_cpu_notify); |
7343 | register_reboot_notifier(&perf_reboot_notifier); |
7344 | |
7345 | ret = init_hw_breakpoint(); |
7346 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); |
7347 | } |
7348 | |
7349 | static int __init perf_event_sysfs_init(void) |
7350 | { |
7351 | struct pmu *pmu; |
7352 | int ret; |
7353 | |
7354 | mutex_lock(&pmus_lock); |
7355 | |
7356 | ret = bus_register(&pmu_bus); |
7357 | if (ret) |
7358 | goto unlock; |
7359 | |
7360 | list_for_each_entry(pmu, &pmus, entry) { |
7361 | if (!pmu->name || pmu->type < 0) |
7362 | continue; |
7363 | |
7364 | ret = pmu_dev_alloc(pmu); |
7365 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); |
7366 | } |
7367 | pmu_bus_running = 1; |
7368 | ret = 0; |
7369 | |
7370 | unlock: |
7371 | mutex_unlock(&pmus_lock); |
7372 | |
7373 | return ret; |
7374 | } |
7375 | device_initcall(perf_event_sysfs_init); |
7376 | |
7377 | #ifdef CONFIG_CGROUP_PERF |
7378 | static struct cgroup_subsys_state *perf_cgroup_create( |
7379 | struct cgroup_subsys *ss, struct cgroup *cont) |
7380 | { |
7381 | struct perf_cgroup *jc; |
7382 | |
7383 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
7384 | if (!jc) |
7385 | return ERR_PTR(-ENOMEM); |
7386 | |
7387 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7388 | if (!jc->info) { |
7389 | kfree(jc); |
7390 | return ERR_PTR(-ENOMEM); |
7391 | } |
7392 | |
7393 | return &jc->css; |
7394 | } |
7395 | |
7396 | static void perf_cgroup_destroy(struct cgroup_subsys *ss, |
7397 | struct cgroup *cont) |
7398 | { |
7399 | struct perf_cgroup *jc; |
7400 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), |
7401 | struct perf_cgroup, css); |
7402 | free_percpu(jc->info); |
7403 | kfree(jc); |
7404 | } |
7405 | |
7406 | static int __perf_cgroup_move(void *info) |
7407 | { |
7408 | struct task_struct *task = info; |
7409 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); |
7410 | return 0; |
7411 | } |
7412 | |
7413 | static void perf_cgroup_move(struct task_struct *task) |
7414 | { |
7415 | task_function_call(task, __perf_cgroup_move, task); |
7416 | } |
7417 | |
7418 | static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
7419 | struct cgroup *old_cgrp, struct task_struct *task, |
7420 | bool threadgroup) |
7421 | { |
7422 | perf_cgroup_move(task); |
7423 | if (threadgroup) { |
7424 | struct task_struct *c; |
7425 | rcu_read_lock(); |
7426 | list_for_each_entry_rcu(c, &task->thread_group, thread_group) { |
7427 | perf_cgroup_move(c); |
7428 | } |
7429 | rcu_read_unlock(); |
7430 | } |
7431 | } |
7432 | |
7433 | static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp, |
7434 | struct cgroup *old_cgrp, struct task_struct *task) |
7435 | { |
7436 | /* |
7437 | * cgroup_exit() is called in the copy_process() failure path. |
7438 | * Ignore this case since the task hasn't ran yet, this avoids |
7439 | * trying to poke a half freed task state from generic code. |
7440 | */ |
7441 | if (!(task->flags & PF_EXITING)) |
7442 | return; |
7443 | |
7444 | perf_cgroup_move(task); |
7445 | } |
7446 | |
7447 | struct cgroup_subsys perf_subsys = { |
7448 | .name = "perf_event", |
7449 | .subsys_id = perf_subsys_id, |
7450 | .create = perf_cgroup_create, |
7451 | .destroy = perf_cgroup_destroy, |
7452 | .exit = perf_cgroup_exit, |
7453 | .attach = perf_cgroup_attach, |
7454 | }; |
7455 | #endif /* CONFIG_CGROUP_PERF */ |
7456 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9