]>
Commit | Line | Data |
---|---|---|
0793a61d | 1 | /* |
57c0c15b | 2 | * Performance events core code: |
0793a61d | 3 | * |
98144511 | 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
e7e7ee2e IM |
5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
7b732a75 | 8 | * |
57c0c15b | 9 | * For licensing details see kernel-base/COPYING |
0793a61d TG |
10 | */ |
11 | ||
12 | #include <linux/fs.h> | |
b9cacc7b | 13 | #include <linux/mm.h> |
0793a61d TG |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> | |
2e80a82a | 16 | #include <linux/idr.h> |
04289bb9 | 17 | #include <linux/file.h> |
0793a61d | 18 | #include <linux/poll.h> |
5a0e3ad6 | 19 | #include <linux/slab.h> |
76e1d904 | 20 | #include <linux/hash.h> |
0793a61d | 21 | #include <linux/sysfs.h> |
22a4f650 | 22 | #include <linux/dcache.h> |
0793a61d | 23 | #include <linux/percpu.h> |
22a4f650 | 24 | #include <linux/ptrace.h> |
c277443c | 25 | #include <linux/reboot.h> |
b9cacc7b | 26 | #include <linux/vmstat.h> |
abe43400 | 27 | #include <linux/device.h> |
6e5fdeed | 28 | #include <linux/export.h> |
906010b2 | 29 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
30 | #include <linux/hardirq.h> |
31 | #include <linux/rculist.h> | |
0793a61d TG |
32 | #include <linux/uaccess.h> |
33 | #include <linux/syscalls.h> | |
34 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 35 | #include <linux/kernel_stat.h> |
cdd6c482 | 36 | #include <linux/perf_event.h> |
6fb2915d | 37 | #include <linux/ftrace_event.h> |
3c502e7a | 38 | #include <linux/hw_breakpoint.h> |
0793a61d | 39 | |
76369139 FW |
40 | #include "internal.h" |
41 | ||
4e193bd4 TB |
42 | #include <asm/irq_regs.h> |
43 | ||
fe4b04fa | 44 | struct remote_function_call { |
e7e7ee2e IM |
45 | struct task_struct *p; |
46 | int (*func)(void *info); | |
47 | void *info; | |
48 | int ret; | |
fe4b04fa PZ |
49 | }; |
50 | ||
51 | static void remote_function(void *data) | |
52 | { | |
53 | struct remote_function_call *tfc = data; | |
54 | struct task_struct *p = tfc->p; | |
55 | ||
56 | if (p) { | |
57 | tfc->ret = -EAGAIN; | |
58 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
59 | return; | |
60 | } | |
61 | ||
62 | tfc->ret = tfc->func(tfc->info); | |
63 | } | |
64 | ||
65 | /** | |
66 | * task_function_call - call a function on the cpu on which a task runs | |
67 | * @p: the task to evaluate | |
68 | * @func: the function to be called | |
69 | * @info: the function call argument | |
70 | * | |
71 | * Calls the function @func when the task is currently running. This might | |
72 | * be on the current CPU, which just calls the function directly | |
73 | * | |
74 | * returns: @func return value, or | |
75 | * -ESRCH - when the process isn't running | |
76 | * -EAGAIN - when the process moved away | |
77 | */ | |
78 | static int | |
79 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
80 | { | |
81 | struct remote_function_call data = { | |
e7e7ee2e IM |
82 | .p = p, |
83 | .func = func, | |
84 | .info = info, | |
85 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
86 | }; |
87 | ||
88 | if (task_curr(p)) | |
89 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
90 | ||
91 | return data.ret; | |
92 | } | |
93 | ||
94 | /** | |
95 | * cpu_function_call - call a function on the cpu | |
96 | * @func: the function to be called | |
97 | * @info: the function call argument | |
98 | * | |
99 | * Calls the function @func on the remote cpu. | |
100 | * | |
101 | * returns: @func return value or -ENXIO when the cpu is offline | |
102 | */ | |
103 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
104 | { | |
105 | struct remote_function_call data = { | |
e7e7ee2e IM |
106 | .p = NULL, |
107 | .func = func, | |
108 | .info = info, | |
109 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
110 | }; |
111 | ||
112 | smp_call_function_single(cpu, remote_function, &data, 1); | |
113 | ||
114 | return data.ret; | |
115 | } | |
116 | ||
e5d1367f SE |
117 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
118 | PERF_FLAG_FD_OUTPUT |\ | |
119 | PERF_FLAG_PID_CGROUP) | |
120 | ||
bce38cd5 SE |
121 | /* |
122 | * branch priv levels that need permission checks | |
123 | */ | |
124 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
125 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
126 | PERF_SAMPLE_BRANCH_HV) | |
127 | ||
0b3fcf17 SE |
128 | enum event_type_t { |
129 | EVENT_FLEXIBLE = 0x1, | |
130 | EVENT_PINNED = 0x2, | |
131 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
132 | }; | |
133 | ||
e5d1367f SE |
134 | /* |
135 | * perf_sched_events : >0 events exist | |
136 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
137 | */ | |
c5905afb | 138 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 139 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 140 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 141 | |
cdd6c482 IM |
142 | static atomic_t nr_mmap_events __read_mostly; |
143 | static atomic_t nr_comm_events __read_mostly; | |
144 | static atomic_t nr_task_events __read_mostly; | |
9ee318a7 | 145 | |
108b02cf PZ |
146 | static LIST_HEAD(pmus); |
147 | static DEFINE_MUTEX(pmus_lock); | |
148 | static struct srcu_struct pmus_srcu; | |
149 | ||
0764771d | 150 | /* |
cdd6c482 | 151 | * perf event paranoia level: |
0fbdea19 IM |
152 | * -1 - not paranoid at all |
153 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 154 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 155 | * 2 - disallow kernel profiling for unpriv |
0764771d | 156 | */ |
cdd6c482 | 157 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 158 | |
20443384 FW |
159 | /* Minimum for 512 kiB + 1 user control page */ |
160 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
161 | |
162 | /* | |
cdd6c482 | 163 | * max perf event sample rate |
df58ab24 | 164 | */ |
163ec435 PZ |
165 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
166 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
167 | static int max_samples_per_tick __read_mostly = | |
168 | DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
169 | ||
170 | int perf_proc_update_handler(struct ctl_table *table, int write, | |
171 | void __user *buffer, size_t *lenp, | |
172 | loff_t *ppos) | |
173 | { | |
174 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
175 | ||
176 | if (ret || !write) | |
177 | return ret; | |
178 | ||
179 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
180 | ||
181 | return 0; | |
182 | } | |
1ccd1549 | 183 | |
cdd6c482 | 184 | static atomic64_t perf_event_id; |
a96bbc16 | 185 | |
0b3fcf17 SE |
186 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
187 | enum event_type_t event_type); | |
188 | ||
189 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
190 | enum event_type_t event_type, |
191 | struct task_struct *task); | |
192 | ||
193 | static void update_context_time(struct perf_event_context *ctx); | |
194 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 195 | |
10c6db11 PZ |
196 | static void ring_buffer_attach(struct perf_event *event, |
197 | struct ring_buffer *rb); | |
198 | ||
cdd6c482 | 199 | void __weak perf_event_print_debug(void) { } |
0793a61d | 200 | |
84c79910 | 201 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 202 | { |
84c79910 | 203 | return "pmu"; |
0793a61d TG |
204 | } |
205 | ||
0b3fcf17 SE |
206 | static inline u64 perf_clock(void) |
207 | { | |
208 | return local_clock(); | |
209 | } | |
210 | ||
e5d1367f SE |
211 | static inline struct perf_cpu_context * |
212 | __get_cpu_context(struct perf_event_context *ctx) | |
213 | { | |
214 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
215 | } | |
216 | ||
facc4307 PZ |
217 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
218 | struct perf_event_context *ctx) | |
219 | { | |
220 | raw_spin_lock(&cpuctx->ctx.lock); | |
221 | if (ctx) | |
222 | raw_spin_lock(&ctx->lock); | |
223 | } | |
224 | ||
225 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
226 | struct perf_event_context *ctx) | |
227 | { | |
228 | if (ctx) | |
229 | raw_spin_unlock(&ctx->lock); | |
230 | raw_spin_unlock(&cpuctx->ctx.lock); | |
231 | } | |
232 | ||
e5d1367f SE |
233 | #ifdef CONFIG_CGROUP_PERF |
234 | ||
3f7cce3c SE |
235 | /* |
236 | * Must ensure cgroup is pinned (css_get) before calling | |
237 | * this function. In other words, we cannot call this function | |
238 | * if there is no cgroup event for the current CPU context. | |
239 | */ | |
e5d1367f SE |
240 | static inline struct perf_cgroup * |
241 | perf_cgroup_from_task(struct task_struct *task) | |
242 | { | |
243 | return container_of(task_subsys_state(task, perf_subsys_id), | |
244 | struct perf_cgroup, css); | |
245 | } | |
246 | ||
247 | static inline bool | |
248 | perf_cgroup_match(struct perf_event *event) | |
249 | { | |
250 | struct perf_event_context *ctx = event->ctx; | |
251 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
252 | ||
253 | return !event->cgrp || event->cgrp == cpuctx->cgrp; | |
254 | } | |
255 | ||
256 | static inline void perf_get_cgroup(struct perf_event *event) | |
257 | { | |
258 | css_get(&event->cgrp->css); | |
259 | } | |
260 | ||
261 | static inline void perf_put_cgroup(struct perf_event *event) | |
262 | { | |
263 | css_put(&event->cgrp->css); | |
264 | } | |
265 | ||
266 | static inline void perf_detach_cgroup(struct perf_event *event) | |
267 | { | |
268 | perf_put_cgroup(event); | |
269 | event->cgrp = NULL; | |
270 | } | |
271 | ||
272 | static inline int is_cgroup_event(struct perf_event *event) | |
273 | { | |
274 | return event->cgrp != NULL; | |
275 | } | |
276 | ||
277 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
278 | { | |
279 | struct perf_cgroup_info *t; | |
280 | ||
281 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
282 | return t->time; | |
283 | } | |
284 | ||
285 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
286 | { | |
287 | struct perf_cgroup_info *info; | |
288 | u64 now; | |
289 | ||
290 | now = perf_clock(); | |
291 | ||
292 | info = this_cpu_ptr(cgrp->info); | |
293 | ||
294 | info->time += now - info->timestamp; | |
295 | info->timestamp = now; | |
296 | } | |
297 | ||
298 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
299 | { | |
300 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
301 | if (cgrp_out) | |
302 | __update_cgrp_time(cgrp_out); | |
303 | } | |
304 | ||
305 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
306 | { | |
3f7cce3c SE |
307 | struct perf_cgroup *cgrp; |
308 | ||
e5d1367f | 309 | /* |
3f7cce3c SE |
310 | * ensure we access cgroup data only when needed and |
311 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 312 | */ |
3f7cce3c | 313 | if (!is_cgroup_event(event)) |
e5d1367f SE |
314 | return; |
315 | ||
3f7cce3c SE |
316 | cgrp = perf_cgroup_from_task(current); |
317 | /* | |
318 | * Do not update time when cgroup is not active | |
319 | */ | |
320 | if (cgrp == event->cgrp) | |
321 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
322 | } |
323 | ||
324 | static inline void | |
3f7cce3c SE |
325 | perf_cgroup_set_timestamp(struct task_struct *task, |
326 | struct perf_event_context *ctx) | |
e5d1367f SE |
327 | { |
328 | struct perf_cgroup *cgrp; | |
329 | struct perf_cgroup_info *info; | |
330 | ||
3f7cce3c SE |
331 | /* |
332 | * ctx->lock held by caller | |
333 | * ensure we do not access cgroup data | |
334 | * unless we have the cgroup pinned (css_get) | |
335 | */ | |
336 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
337 | return; |
338 | ||
339 | cgrp = perf_cgroup_from_task(task); | |
340 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 341 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
342 | } |
343 | ||
344 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
345 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
346 | ||
347 | /* | |
348 | * reschedule events based on the cgroup constraint of task. | |
349 | * | |
350 | * mode SWOUT : schedule out everything | |
351 | * mode SWIN : schedule in based on cgroup for next | |
352 | */ | |
353 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
354 | { | |
355 | struct perf_cpu_context *cpuctx; | |
356 | struct pmu *pmu; | |
357 | unsigned long flags; | |
358 | ||
359 | /* | |
360 | * disable interrupts to avoid geting nr_cgroup | |
361 | * changes via __perf_event_disable(). Also | |
362 | * avoids preemption. | |
363 | */ | |
364 | local_irq_save(flags); | |
365 | ||
366 | /* | |
367 | * we reschedule only in the presence of cgroup | |
368 | * constrained events. | |
369 | */ | |
370 | rcu_read_lock(); | |
371 | ||
372 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f SE |
373 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
374 | ||
e5d1367f SE |
375 | /* |
376 | * perf_cgroup_events says at least one | |
377 | * context on this CPU has cgroup events. | |
378 | * | |
379 | * ctx->nr_cgroups reports the number of cgroup | |
380 | * events for a context. | |
381 | */ | |
382 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
383 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
384 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
385 | |
386 | if (mode & PERF_CGROUP_SWOUT) { | |
387 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
388 | /* | |
389 | * must not be done before ctxswout due | |
390 | * to event_filter_match() in event_sched_out() | |
391 | */ | |
392 | cpuctx->cgrp = NULL; | |
393 | } | |
394 | ||
395 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 396 | WARN_ON_ONCE(cpuctx->cgrp); |
e5d1367f SE |
397 | /* set cgrp before ctxsw in to |
398 | * allow event_filter_match() to not | |
399 | * have to pass task around | |
400 | */ | |
401 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
402 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
403 | } | |
facc4307 PZ |
404 | perf_pmu_enable(cpuctx->ctx.pmu); |
405 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 406 | } |
e5d1367f SE |
407 | } |
408 | ||
409 | rcu_read_unlock(); | |
410 | ||
411 | local_irq_restore(flags); | |
412 | } | |
413 | ||
a8d757ef SE |
414 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
415 | struct task_struct *next) | |
e5d1367f | 416 | { |
a8d757ef SE |
417 | struct perf_cgroup *cgrp1; |
418 | struct perf_cgroup *cgrp2 = NULL; | |
419 | ||
420 | /* | |
421 | * we come here when we know perf_cgroup_events > 0 | |
422 | */ | |
423 | cgrp1 = perf_cgroup_from_task(task); | |
424 | ||
425 | /* | |
426 | * next is NULL when called from perf_event_enable_on_exec() | |
427 | * that will systematically cause a cgroup_switch() | |
428 | */ | |
429 | if (next) | |
430 | cgrp2 = perf_cgroup_from_task(next); | |
431 | ||
432 | /* | |
433 | * only schedule out current cgroup events if we know | |
434 | * that we are switching to a different cgroup. Otherwise, | |
435 | * do no touch the cgroup events. | |
436 | */ | |
437 | if (cgrp1 != cgrp2) | |
438 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
439 | } |
440 | ||
a8d757ef SE |
441 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
442 | struct task_struct *task) | |
e5d1367f | 443 | { |
a8d757ef SE |
444 | struct perf_cgroup *cgrp1; |
445 | struct perf_cgroup *cgrp2 = NULL; | |
446 | ||
447 | /* | |
448 | * we come here when we know perf_cgroup_events > 0 | |
449 | */ | |
450 | cgrp1 = perf_cgroup_from_task(task); | |
451 | ||
452 | /* prev can never be NULL */ | |
453 | cgrp2 = perf_cgroup_from_task(prev); | |
454 | ||
455 | /* | |
456 | * only need to schedule in cgroup events if we are changing | |
457 | * cgroup during ctxsw. Cgroup events were not scheduled | |
458 | * out of ctxsw out if that was not the case. | |
459 | */ | |
460 | if (cgrp1 != cgrp2) | |
461 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
462 | } |
463 | ||
464 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
465 | struct perf_event_attr *attr, | |
466 | struct perf_event *group_leader) | |
467 | { | |
468 | struct perf_cgroup *cgrp; | |
469 | struct cgroup_subsys_state *css; | |
470 | struct file *file; | |
471 | int ret = 0, fput_needed; | |
472 | ||
473 | file = fget_light(fd, &fput_needed); | |
474 | if (!file) | |
475 | return -EBADF; | |
476 | ||
477 | css = cgroup_css_from_dir(file, perf_subsys_id); | |
3db272c0 LZ |
478 | if (IS_ERR(css)) { |
479 | ret = PTR_ERR(css); | |
480 | goto out; | |
481 | } | |
e5d1367f SE |
482 | |
483 | cgrp = container_of(css, struct perf_cgroup, css); | |
484 | event->cgrp = cgrp; | |
485 | ||
f75e18cb LZ |
486 | /* must be done before we fput() the file */ |
487 | perf_get_cgroup(event); | |
488 | ||
e5d1367f SE |
489 | /* |
490 | * all events in a group must monitor | |
491 | * the same cgroup because a task belongs | |
492 | * to only one perf cgroup at a time | |
493 | */ | |
494 | if (group_leader && group_leader->cgrp != cgrp) { | |
495 | perf_detach_cgroup(event); | |
496 | ret = -EINVAL; | |
e5d1367f | 497 | } |
3db272c0 | 498 | out: |
e5d1367f SE |
499 | fput_light(file, fput_needed); |
500 | return ret; | |
501 | } | |
502 | ||
503 | static inline void | |
504 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
505 | { | |
506 | struct perf_cgroup_info *t; | |
507 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
508 | event->shadow_ctx_time = now - t->timestamp; | |
509 | } | |
510 | ||
511 | static inline void | |
512 | perf_cgroup_defer_enabled(struct perf_event *event) | |
513 | { | |
514 | /* | |
515 | * when the current task's perf cgroup does not match | |
516 | * the event's, we need to remember to call the | |
517 | * perf_mark_enable() function the first time a task with | |
518 | * a matching perf cgroup is scheduled in. | |
519 | */ | |
520 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
521 | event->cgrp_defer_enabled = 1; | |
522 | } | |
523 | ||
524 | static inline void | |
525 | perf_cgroup_mark_enabled(struct perf_event *event, | |
526 | struct perf_event_context *ctx) | |
527 | { | |
528 | struct perf_event *sub; | |
529 | u64 tstamp = perf_event_time(event); | |
530 | ||
531 | if (!event->cgrp_defer_enabled) | |
532 | return; | |
533 | ||
534 | event->cgrp_defer_enabled = 0; | |
535 | ||
536 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
537 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
538 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
539 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
540 | sub->cgrp_defer_enabled = 0; | |
541 | } | |
542 | } | |
543 | } | |
544 | #else /* !CONFIG_CGROUP_PERF */ | |
545 | ||
546 | static inline bool | |
547 | perf_cgroup_match(struct perf_event *event) | |
548 | { | |
549 | return true; | |
550 | } | |
551 | ||
552 | static inline void perf_detach_cgroup(struct perf_event *event) | |
553 | {} | |
554 | ||
555 | static inline int is_cgroup_event(struct perf_event *event) | |
556 | { | |
557 | return 0; | |
558 | } | |
559 | ||
560 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
561 | { | |
562 | return 0; | |
563 | } | |
564 | ||
565 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
566 | { | |
567 | } | |
568 | ||
569 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
570 | { | |
571 | } | |
572 | ||
a8d757ef SE |
573 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
574 | struct task_struct *next) | |
e5d1367f SE |
575 | { |
576 | } | |
577 | ||
a8d757ef SE |
578 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
579 | struct task_struct *task) | |
e5d1367f SE |
580 | { |
581 | } | |
582 | ||
583 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
584 | struct perf_event_attr *attr, | |
585 | struct perf_event *group_leader) | |
586 | { | |
587 | return -EINVAL; | |
588 | } | |
589 | ||
590 | static inline void | |
3f7cce3c SE |
591 | perf_cgroup_set_timestamp(struct task_struct *task, |
592 | struct perf_event_context *ctx) | |
e5d1367f SE |
593 | { |
594 | } | |
595 | ||
596 | void | |
597 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
598 | { | |
599 | } | |
600 | ||
601 | static inline void | |
602 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
603 | { | |
604 | } | |
605 | ||
606 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
607 | { | |
608 | return 0; | |
609 | } | |
610 | ||
611 | static inline void | |
612 | perf_cgroup_defer_enabled(struct perf_event *event) | |
613 | { | |
614 | } | |
615 | ||
616 | static inline void | |
617 | perf_cgroup_mark_enabled(struct perf_event *event, | |
618 | struct perf_event_context *ctx) | |
619 | { | |
620 | } | |
621 | #endif | |
622 | ||
33696fc0 | 623 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 624 | { |
33696fc0 PZ |
625 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
626 | if (!(*count)++) | |
627 | pmu->pmu_disable(pmu); | |
9e35ad38 | 628 | } |
9e35ad38 | 629 | |
33696fc0 | 630 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 631 | { |
33696fc0 PZ |
632 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
633 | if (!--(*count)) | |
634 | pmu->pmu_enable(pmu); | |
9e35ad38 | 635 | } |
9e35ad38 | 636 | |
e9d2b064 PZ |
637 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
638 | ||
639 | /* | |
640 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
641 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
642 | * disabled, while rotate_context is called from IRQ context. | |
643 | */ | |
108b02cf | 644 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 645 | { |
108b02cf | 646 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 647 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 648 | |
e9d2b064 | 649 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 650 | |
e9d2b064 PZ |
651 | if (list_empty(&cpuctx->rotation_list)) |
652 | list_add(&cpuctx->rotation_list, head); | |
9e35ad38 | 653 | } |
9e35ad38 | 654 | |
cdd6c482 | 655 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 656 | { |
e5289d4a | 657 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
658 | } |
659 | ||
cdd6c482 | 660 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 661 | { |
564c2b21 PM |
662 | if (atomic_dec_and_test(&ctx->refcount)) { |
663 | if (ctx->parent_ctx) | |
664 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
665 | if (ctx->task) |
666 | put_task_struct(ctx->task); | |
cb796ff3 | 667 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 668 | } |
a63eaf34 PM |
669 | } |
670 | ||
cdd6c482 | 671 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
672 | { |
673 | if (ctx->parent_ctx) { | |
674 | put_ctx(ctx->parent_ctx); | |
675 | ctx->parent_ctx = NULL; | |
676 | } | |
677 | } | |
678 | ||
6844c09d ACM |
679 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
680 | { | |
681 | /* | |
682 | * only top level events have the pid namespace they were created in | |
683 | */ | |
684 | if (event->parent) | |
685 | event = event->parent; | |
686 | ||
687 | return task_tgid_nr_ns(p, event->ns); | |
688 | } | |
689 | ||
690 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
691 | { | |
692 | /* | |
693 | * only top level events have the pid namespace they were created in | |
694 | */ | |
695 | if (event->parent) | |
696 | event = event->parent; | |
697 | ||
698 | return task_pid_nr_ns(p, event->ns); | |
699 | } | |
700 | ||
7f453c24 | 701 | /* |
cdd6c482 | 702 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
703 | * to userspace. |
704 | */ | |
cdd6c482 | 705 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 706 | { |
cdd6c482 | 707 | u64 id = event->id; |
7f453c24 | 708 | |
cdd6c482 IM |
709 | if (event->parent) |
710 | id = event->parent->id; | |
7f453c24 PZ |
711 | |
712 | return id; | |
713 | } | |
714 | ||
25346b93 | 715 | /* |
cdd6c482 | 716 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
717 | * This has to cope with with the fact that until it is locked, |
718 | * the context could get moved to another task. | |
719 | */ | |
cdd6c482 | 720 | static struct perf_event_context * |
8dc85d54 | 721 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 722 | { |
cdd6c482 | 723 | struct perf_event_context *ctx; |
25346b93 PM |
724 | |
725 | rcu_read_lock(); | |
9ed6060d | 726 | retry: |
8dc85d54 | 727 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
728 | if (ctx) { |
729 | /* | |
730 | * If this context is a clone of another, it might | |
731 | * get swapped for another underneath us by | |
cdd6c482 | 732 | * perf_event_task_sched_out, though the |
25346b93 PM |
733 | * rcu_read_lock() protects us from any context |
734 | * getting freed. Lock the context and check if it | |
735 | * got swapped before we could get the lock, and retry | |
736 | * if so. If we locked the right context, then it | |
737 | * can't get swapped on us any more. | |
738 | */ | |
e625cce1 | 739 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 740 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 741 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
25346b93 PM |
742 | goto retry; |
743 | } | |
b49a9e7e PZ |
744 | |
745 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 746 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
747 | ctx = NULL; |
748 | } | |
25346b93 PM |
749 | } |
750 | rcu_read_unlock(); | |
751 | return ctx; | |
752 | } | |
753 | ||
754 | /* | |
755 | * Get the context for a task and increment its pin_count so it | |
756 | * can't get swapped to another task. This also increments its | |
757 | * reference count so that the context can't get freed. | |
758 | */ | |
8dc85d54 PZ |
759 | static struct perf_event_context * |
760 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 761 | { |
cdd6c482 | 762 | struct perf_event_context *ctx; |
25346b93 PM |
763 | unsigned long flags; |
764 | ||
8dc85d54 | 765 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
766 | if (ctx) { |
767 | ++ctx->pin_count; | |
e625cce1 | 768 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
769 | } |
770 | return ctx; | |
771 | } | |
772 | ||
cdd6c482 | 773 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
774 | { |
775 | unsigned long flags; | |
776 | ||
e625cce1 | 777 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 778 | --ctx->pin_count; |
e625cce1 | 779 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
780 | } |
781 | ||
f67218c3 PZ |
782 | /* |
783 | * Update the record of the current time in a context. | |
784 | */ | |
785 | static void update_context_time(struct perf_event_context *ctx) | |
786 | { | |
787 | u64 now = perf_clock(); | |
788 | ||
789 | ctx->time += now - ctx->timestamp; | |
790 | ctx->timestamp = now; | |
791 | } | |
792 | ||
4158755d SE |
793 | static u64 perf_event_time(struct perf_event *event) |
794 | { | |
795 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
796 | |
797 | if (is_cgroup_event(event)) | |
798 | return perf_cgroup_event_time(event); | |
799 | ||
4158755d SE |
800 | return ctx ? ctx->time : 0; |
801 | } | |
802 | ||
f67218c3 PZ |
803 | /* |
804 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 805 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
806 | */ |
807 | static void update_event_times(struct perf_event *event) | |
808 | { | |
809 | struct perf_event_context *ctx = event->ctx; | |
810 | u64 run_end; | |
811 | ||
812 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
813 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
814 | return; | |
e5d1367f SE |
815 | /* |
816 | * in cgroup mode, time_enabled represents | |
817 | * the time the event was enabled AND active | |
818 | * tasks were in the monitored cgroup. This is | |
819 | * independent of the activity of the context as | |
820 | * there may be a mix of cgroup and non-cgroup events. | |
821 | * | |
822 | * That is why we treat cgroup events differently | |
823 | * here. | |
824 | */ | |
825 | if (is_cgroup_event(event)) | |
46cd6a7f | 826 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
827 | else if (ctx->is_active) |
828 | run_end = ctx->time; | |
acd1d7c1 PZ |
829 | else |
830 | run_end = event->tstamp_stopped; | |
831 | ||
832 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
833 | |
834 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
835 | run_end = event->tstamp_stopped; | |
836 | else | |
4158755d | 837 | run_end = perf_event_time(event); |
f67218c3 PZ |
838 | |
839 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 840 | |
f67218c3 PZ |
841 | } |
842 | ||
96c21a46 PZ |
843 | /* |
844 | * Update total_time_enabled and total_time_running for all events in a group. | |
845 | */ | |
846 | static void update_group_times(struct perf_event *leader) | |
847 | { | |
848 | struct perf_event *event; | |
849 | ||
850 | update_event_times(leader); | |
851 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
852 | update_event_times(event); | |
853 | } | |
854 | ||
889ff015 FW |
855 | static struct list_head * |
856 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
857 | { | |
858 | if (event->attr.pinned) | |
859 | return &ctx->pinned_groups; | |
860 | else | |
861 | return &ctx->flexible_groups; | |
862 | } | |
863 | ||
fccc714b | 864 | /* |
cdd6c482 | 865 | * Add a event from the lists for its context. |
fccc714b PZ |
866 | * Must be called with ctx->mutex and ctx->lock held. |
867 | */ | |
04289bb9 | 868 | static void |
cdd6c482 | 869 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 870 | { |
8a49542c PZ |
871 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
872 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
873 | |
874 | /* | |
8a49542c PZ |
875 | * If we're a stand alone event or group leader, we go to the context |
876 | * list, group events are kept attached to the group so that | |
877 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 878 | */ |
8a49542c | 879 | if (event->group_leader == event) { |
889ff015 FW |
880 | struct list_head *list; |
881 | ||
d6f962b5 FW |
882 | if (is_software_event(event)) |
883 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
884 | ||
889ff015 FW |
885 | list = ctx_group_list(event, ctx); |
886 | list_add_tail(&event->group_entry, list); | |
5c148194 | 887 | } |
592903cd | 888 | |
08309379 | 889 | if (is_cgroup_event(event)) |
e5d1367f | 890 | ctx->nr_cgroups++; |
e5d1367f | 891 | |
d010b332 SE |
892 | if (has_branch_stack(event)) |
893 | ctx->nr_branch_stack++; | |
894 | ||
cdd6c482 | 895 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 896 | if (!ctx->nr_events) |
108b02cf | 897 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
898 | ctx->nr_events++; |
899 | if (event->attr.inherit_stat) | |
bfbd3381 | 900 | ctx->nr_stat++; |
04289bb9 IM |
901 | } |
902 | ||
c320c7b7 ACM |
903 | /* |
904 | * Called at perf_event creation and when events are attached/detached from a | |
905 | * group. | |
906 | */ | |
907 | static void perf_event__read_size(struct perf_event *event) | |
908 | { | |
909 | int entry = sizeof(u64); /* value */ | |
910 | int size = 0; | |
911 | int nr = 1; | |
912 | ||
913 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
914 | size += sizeof(u64); | |
915 | ||
916 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
917 | size += sizeof(u64); | |
918 | ||
919 | if (event->attr.read_format & PERF_FORMAT_ID) | |
920 | entry += sizeof(u64); | |
921 | ||
922 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
923 | nr += event->group_leader->nr_siblings; | |
924 | size += sizeof(u64); | |
925 | } | |
926 | ||
927 | size += entry * nr; | |
928 | event->read_size = size; | |
929 | } | |
930 | ||
931 | static void perf_event__header_size(struct perf_event *event) | |
932 | { | |
933 | struct perf_sample_data *data; | |
934 | u64 sample_type = event->attr.sample_type; | |
935 | u16 size = 0; | |
936 | ||
937 | perf_event__read_size(event); | |
938 | ||
939 | if (sample_type & PERF_SAMPLE_IP) | |
940 | size += sizeof(data->ip); | |
941 | ||
6844c09d ACM |
942 | if (sample_type & PERF_SAMPLE_ADDR) |
943 | size += sizeof(data->addr); | |
944 | ||
945 | if (sample_type & PERF_SAMPLE_PERIOD) | |
946 | size += sizeof(data->period); | |
947 | ||
948 | if (sample_type & PERF_SAMPLE_READ) | |
949 | size += event->read_size; | |
950 | ||
951 | event->header_size = size; | |
952 | } | |
953 | ||
954 | static void perf_event__id_header_size(struct perf_event *event) | |
955 | { | |
956 | struct perf_sample_data *data; | |
957 | u64 sample_type = event->attr.sample_type; | |
958 | u16 size = 0; | |
959 | ||
c320c7b7 ACM |
960 | if (sample_type & PERF_SAMPLE_TID) |
961 | size += sizeof(data->tid_entry); | |
962 | ||
963 | if (sample_type & PERF_SAMPLE_TIME) | |
964 | size += sizeof(data->time); | |
965 | ||
c320c7b7 ACM |
966 | if (sample_type & PERF_SAMPLE_ID) |
967 | size += sizeof(data->id); | |
968 | ||
969 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
970 | size += sizeof(data->stream_id); | |
971 | ||
972 | if (sample_type & PERF_SAMPLE_CPU) | |
973 | size += sizeof(data->cpu_entry); | |
974 | ||
6844c09d | 975 | event->id_header_size = size; |
c320c7b7 ACM |
976 | } |
977 | ||
8a49542c PZ |
978 | static void perf_group_attach(struct perf_event *event) |
979 | { | |
c320c7b7 | 980 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 981 | |
74c3337c PZ |
982 | /* |
983 | * We can have double attach due to group movement in perf_event_open. | |
984 | */ | |
985 | if (event->attach_state & PERF_ATTACH_GROUP) | |
986 | return; | |
987 | ||
8a49542c PZ |
988 | event->attach_state |= PERF_ATTACH_GROUP; |
989 | ||
990 | if (group_leader == event) | |
991 | return; | |
992 | ||
993 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
994 | !is_software_event(event)) | |
995 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
996 | ||
997 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
998 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
999 | |
1000 | perf_event__header_size(group_leader); | |
1001 | ||
1002 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1003 | perf_event__header_size(pos); | |
8a49542c PZ |
1004 | } |
1005 | ||
a63eaf34 | 1006 | /* |
cdd6c482 | 1007 | * Remove a event from the lists for its context. |
fccc714b | 1008 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1009 | */ |
04289bb9 | 1010 | static void |
cdd6c482 | 1011 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1012 | { |
68cacd29 | 1013 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1014 | /* |
1015 | * We can have double detach due to exit/hot-unplug + close. | |
1016 | */ | |
1017 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1018 | return; |
8a49542c PZ |
1019 | |
1020 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1021 | ||
68cacd29 | 1022 | if (is_cgroup_event(event)) { |
e5d1367f | 1023 | ctx->nr_cgroups--; |
68cacd29 SE |
1024 | cpuctx = __get_cpu_context(ctx); |
1025 | /* | |
1026 | * if there are no more cgroup events | |
1027 | * then cler cgrp to avoid stale pointer | |
1028 | * in update_cgrp_time_from_cpuctx() | |
1029 | */ | |
1030 | if (!ctx->nr_cgroups) | |
1031 | cpuctx->cgrp = NULL; | |
1032 | } | |
e5d1367f | 1033 | |
d010b332 SE |
1034 | if (has_branch_stack(event)) |
1035 | ctx->nr_branch_stack--; | |
1036 | ||
cdd6c482 IM |
1037 | ctx->nr_events--; |
1038 | if (event->attr.inherit_stat) | |
bfbd3381 | 1039 | ctx->nr_stat--; |
8bc20959 | 1040 | |
cdd6c482 | 1041 | list_del_rcu(&event->event_entry); |
04289bb9 | 1042 | |
8a49542c PZ |
1043 | if (event->group_leader == event) |
1044 | list_del_init(&event->group_entry); | |
5c148194 | 1045 | |
96c21a46 | 1046 | update_group_times(event); |
b2e74a26 SE |
1047 | |
1048 | /* | |
1049 | * If event was in error state, then keep it | |
1050 | * that way, otherwise bogus counts will be | |
1051 | * returned on read(). The only way to get out | |
1052 | * of error state is by explicit re-enabling | |
1053 | * of the event | |
1054 | */ | |
1055 | if (event->state > PERF_EVENT_STATE_OFF) | |
1056 | event->state = PERF_EVENT_STATE_OFF; | |
050735b0 PZ |
1057 | } |
1058 | ||
8a49542c | 1059 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1060 | { |
1061 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1062 | struct list_head *list = NULL; |
1063 | ||
1064 | /* | |
1065 | * We can have double detach due to exit/hot-unplug + close. | |
1066 | */ | |
1067 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1068 | return; | |
1069 | ||
1070 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1071 | ||
1072 | /* | |
1073 | * If this is a sibling, remove it from its group. | |
1074 | */ | |
1075 | if (event->group_leader != event) { | |
1076 | list_del_init(&event->group_entry); | |
1077 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1078 | goto out; |
8a49542c PZ |
1079 | } |
1080 | ||
1081 | if (!list_empty(&event->group_entry)) | |
1082 | list = &event->group_entry; | |
2e2af50b | 1083 | |
04289bb9 | 1084 | /* |
cdd6c482 IM |
1085 | * If this was a group event with sibling events then |
1086 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1087 | * to whatever list we are on. |
04289bb9 | 1088 | */ |
cdd6c482 | 1089 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1090 | if (list) |
1091 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1092 | sibling->group_leader = sibling; |
d6f962b5 FW |
1093 | |
1094 | /* Inherit group flags from the previous leader */ | |
1095 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1096 | } |
c320c7b7 ACM |
1097 | |
1098 | out: | |
1099 | perf_event__header_size(event->group_leader); | |
1100 | ||
1101 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1102 | perf_event__header_size(tmp); | |
04289bb9 IM |
1103 | } |
1104 | ||
fa66f07a SE |
1105 | static inline int |
1106 | event_filter_match(struct perf_event *event) | |
1107 | { | |
e5d1367f SE |
1108 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1109 | && perf_cgroup_match(event); | |
fa66f07a SE |
1110 | } |
1111 | ||
9ffcfa6f SE |
1112 | static void |
1113 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1114 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1115 | struct perf_event_context *ctx) |
3b6f9e5c | 1116 | { |
4158755d | 1117 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1118 | u64 delta; |
1119 | /* | |
1120 | * An event which could not be activated because of | |
1121 | * filter mismatch still needs to have its timings | |
1122 | * maintained, otherwise bogus information is return | |
1123 | * via read() for time_enabled, time_running: | |
1124 | */ | |
1125 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1126 | && !event_filter_match(event)) { | |
e5d1367f | 1127 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1128 | event->tstamp_running += delta; |
4158755d | 1129 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1130 | } |
1131 | ||
cdd6c482 | 1132 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1133 | return; |
3b6f9e5c | 1134 | |
cdd6c482 IM |
1135 | event->state = PERF_EVENT_STATE_INACTIVE; |
1136 | if (event->pending_disable) { | |
1137 | event->pending_disable = 0; | |
1138 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1139 | } |
4158755d | 1140 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1141 | event->pmu->del(event, 0); |
cdd6c482 | 1142 | event->oncpu = -1; |
3b6f9e5c | 1143 | |
cdd6c482 | 1144 | if (!is_software_event(event)) |
3b6f9e5c PM |
1145 | cpuctx->active_oncpu--; |
1146 | ctx->nr_active--; | |
0f5a2601 PZ |
1147 | if (event->attr.freq && event->attr.sample_freq) |
1148 | ctx->nr_freq--; | |
cdd6c482 | 1149 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c PM |
1150 | cpuctx->exclusive = 0; |
1151 | } | |
1152 | ||
d859e29f | 1153 | static void |
cdd6c482 | 1154 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1155 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1156 | struct perf_event_context *ctx) |
d859e29f | 1157 | { |
cdd6c482 | 1158 | struct perf_event *event; |
fa66f07a | 1159 | int state = group_event->state; |
d859e29f | 1160 | |
cdd6c482 | 1161 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1162 | |
1163 | /* | |
1164 | * Schedule out siblings (if any): | |
1165 | */ | |
cdd6c482 IM |
1166 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1167 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1168 | |
fa66f07a | 1169 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1170 | cpuctx->exclusive = 0; |
1171 | } | |
1172 | ||
0793a61d | 1173 | /* |
cdd6c482 | 1174 | * Cross CPU call to remove a performance event |
0793a61d | 1175 | * |
cdd6c482 | 1176 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1177 | * remove it from the context list. |
1178 | */ | |
fe4b04fa | 1179 | static int __perf_remove_from_context(void *info) |
0793a61d | 1180 | { |
cdd6c482 IM |
1181 | struct perf_event *event = info; |
1182 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1183 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1184 | |
e625cce1 | 1185 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1186 | event_sched_out(event, cpuctx, ctx); |
cdd6c482 | 1187 | list_del_event(event, ctx); |
64ce3126 PZ |
1188 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1189 | ctx->is_active = 0; | |
1190 | cpuctx->task_ctx = NULL; | |
1191 | } | |
e625cce1 | 1192 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1193 | |
1194 | return 0; | |
0793a61d TG |
1195 | } |
1196 | ||
1197 | ||
1198 | /* | |
cdd6c482 | 1199 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1200 | * |
cdd6c482 | 1201 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1202 | * call when the task is on a CPU. |
c93f7669 | 1203 | * |
cdd6c482 IM |
1204 | * If event->ctx is a cloned context, callers must make sure that |
1205 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1206 | * remains valid. This is OK when called from perf_release since |
1207 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1208 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1209 | * context has been detached from its task. |
0793a61d | 1210 | */ |
fe4b04fa | 1211 | static void perf_remove_from_context(struct perf_event *event) |
0793a61d | 1212 | { |
cdd6c482 | 1213 | struct perf_event_context *ctx = event->ctx; |
0793a61d TG |
1214 | struct task_struct *task = ctx->task; |
1215 | ||
fe4b04fa PZ |
1216 | lockdep_assert_held(&ctx->mutex); |
1217 | ||
0793a61d TG |
1218 | if (!task) { |
1219 | /* | |
cdd6c482 | 1220 | * Per cpu events are removed via an smp call and |
af901ca1 | 1221 | * the removal is always successful. |
0793a61d | 1222 | */ |
fe4b04fa | 1223 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
0793a61d TG |
1224 | return; |
1225 | } | |
1226 | ||
1227 | retry: | |
fe4b04fa PZ |
1228 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1229 | return; | |
0793a61d | 1230 | |
e625cce1 | 1231 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1232 | /* |
fe4b04fa PZ |
1233 | * If we failed to find a running task, but find the context active now |
1234 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1235 | */ |
fe4b04fa | 1236 | if (ctx->is_active) { |
e625cce1 | 1237 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1238 | goto retry; |
1239 | } | |
1240 | ||
1241 | /* | |
fe4b04fa PZ |
1242 | * Since the task isn't running, its safe to remove the event, us |
1243 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1244 | */ |
fe4b04fa | 1245 | list_del_event(event, ctx); |
e625cce1 | 1246 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1247 | } |
1248 | ||
d859e29f | 1249 | /* |
cdd6c482 | 1250 | * Cross CPU call to disable a performance event |
d859e29f | 1251 | */ |
fe4b04fa | 1252 | static int __perf_event_disable(void *info) |
d859e29f | 1253 | { |
cdd6c482 | 1254 | struct perf_event *event = info; |
cdd6c482 | 1255 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1256 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1257 | |
1258 | /* | |
cdd6c482 IM |
1259 | * If this is a per-task event, need to check whether this |
1260 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1261 | * |
1262 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1263 | * flipping contexts around. | |
d859e29f | 1264 | */ |
665c2142 | 1265 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1266 | return -EINVAL; |
d859e29f | 1267 | |
e625cce1 | 1268 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1269 | |
1270 | /* | |
cdd6c482 | 1271 | * If the event is on, turn it off. |
d859e29f PM |
1272 | * If it is in error state, leave it in error state. |
1273 | */ | |
cdd6c482 | 1274 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1275 | update_context_time(ctx); |
e5d1367f | 1276 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1277 | update_group_times(event); |
1278 | if (event == event->group_leader) | |
1279 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1280 | else |
cdd6c482 IM |
1281 | event_sched_out(event, cpuctx, ctx); |
1282 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1283 | } |
1284 | ||
e625cce1 | 1285 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1286 | |
1287 | return 0; | |
d859e29f PM |
1288 | } |
1289 | ||
1290 | /* | |
cdd6c482 | 1291 | * Disable a event. |
c93f7669 | 1292 | * |
cdd6c482 IM |
1293 | * If event->ctx is a cloned context, callers must make sure that |
1294 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1295 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1296 | * perf_event_for_each_child or perf_event_for_each because they |
1297 | * hold the top-level event's child_mutex, so any descendant that | |
1298 | * goes to exit will block in sync_child_event. | |
1299 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1300 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1301 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1302 | */ |
44234adc | 1303 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1304 | { |
cdd6c482 | 1305 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1306 | struct task_struct *task = ctx->task; |
1307 | ||
1308 | if (!task) { | |
1309 | /* | |
cdd6c482 | 1310 | * Disable the event on the cpu that it's on |
d859e29f | 1311 | */ |
fe4b04fa | 1312 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1313 | return; |
1314 | } | |
1315 | ||
9ed6060d | 1316 | retry: |
fe4b04fa PZ |
1317 | if (!task_function_call(task, __perf_event_disable, event)) |
1318 | return; | |
d859e29f | 1319 | |
e625cce1 | 1320 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1321 | /* |
cdd6c482 | 1322 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1323 | */ |
cdd6c482 | 1324 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1325 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1326 | /* |
1327 | * Reload the task pointer, it might have been changed by | |
1328 | * a concurrent perf_event_context_sched_out(). | |
1329 | */ | |
1330 | task = ctx->task; | |
d859e29f PM |
1331 | goto retry; |
1332 | } | |
1333 | ||
1334 | /* | |
1335 | * Since we have the lock this context can't be scheduled | |
1336 | * in, so we can change the state safely. | |
1337 | */ | |
cdd6c482 IM |
1338 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1339 | update_group_times(event); | |
1340 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1341 | } |
e625cce1 | 1342 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1343 | } |
dcfce4a0 | 1344 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1345 | |
e5d1367f SE |
1346 | static void perf_set_shadow_time(struct perf_event *event, |
1347 | struct perf_event_context *ctx, | |
1348 | u64 tstamp) | |
1349 | { | |
1350 | /* | |
1351 | * use the correct time source for the time snapshot | |
1352 | * | |
1353 | * We could get by without this by leveraging the | |
1354 | * fact that to get to this function, the caller | |
1355 | * has most likely already called update_context_time() | |
1356 | * and update_cgrp_time_xx() and thus both timestamp | |
1357 | * are identical (or very close). Given that tstamp is, | |
1358 | * already adjusted for cgroup, we could say that: | |
1359 | * tstamp - ctx->timestamp | |
1360 | * is equivalent to | |
1361 | * tstamp - cgrp->timestamp. | |
1362 | * | |
1363 | * Then, in perf_output_read(), the calculation would | |
1364 | * work with no changes because: | |
1365 | * - event is guaranteed scheduled in | |
1366 | * - no scheduled out in between | |
1367 | * - thus the timestamp would be the same | |
1368 | * | |
1369 | * But this is a bit hairy. | |
1370 | * | |
1371 | * So instead, we have an explicit cgroup call to remain | |
1372 | * within the time time source all along. We believe it | |
1373 | * is cleaner and simpler to understand. | |
1374 | */ | |
1375 | if (is_cgroup_event(event)) | |
1376 | perf_cgroup_set_shadow_time(event, tstamp); | |
1377 | else | |
1378 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1379 | } | |
1380 | ||
4fe757dd PZ |
1381 | #define MAX_INTERRUPTS (~0ULL) |
1382 | ||
1383 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1384 | ||
235c7fc7 | 1385 | static int |
9ffcfa6f | 1386 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1387 | struct perf_cpu_context *cpuctx, |
6e37738a | 1388 | struct perf_event_context *ctx) |
235c7fc7 | 1389 | { |
4158755d SE |
1390 | u64 tstamp = perf_event_time(event); |
1391 | ||
cdd6c482 | 1392 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1393 | return 0; |
1394 | ||
cdd6c482 | 1395 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1396 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1397 | |
1398 | /* | |
1399 | * Unthrottle events, since we scheduled we might have missed several | |
1400 | * ticks already, also for a heavily scheduling task there is little | |
1401 | * guarantee it'll get a tick in a timely manner. | |
1402 | */ | |
1403 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1404 | perf_log_throttle(event, 1); | |
1405 | event->hw.interrupts = 0; | |
1406 | } | |
1407 | ||
235c7fc7 IM |
1408 | /* |
1409 | * The new state must be visible before we turn it on in the hardware: | |
1410 | */ | |
1411 | smp_wmb(); | |
1412 | ||
a4eaf7f1 | 1413 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1414 | event->state = PERF_EVENT_STATE_INACTIVE; |
1415 | event->oncpu = -1; | |
235c7fc7 IM |
1416 | return -EAGAIN; |
1417 | } | |
1418 | ||
4158755d | 1419 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1420 | |
e5d1367f | 1421 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1422 | |
cdd6c482 | 1423 | if (!is_software_event(event)) |
3b6f9e5c | 1424 | cpuctx->active_oncpu++; |
235c7fc7 | 1425 | ctx->nr_active++; |
0f5a2601 PZ |
1426 | if (event->attr.freq && event->attr.sample_freq) |
1427 | ctx->nr_freq++; | |
235c7fc7 | 1428 | |
cdd6c482 | 1429 | if (event->attr.exclusive) |
3b6f9e5c PM |
1430 | cpuctx->exclusive = 1; |
1431 | ||
235c7fc7 IM |
1432 | return 0; |
1433 | } | |
1434 | ||
6751b71e | 1435 | static int |
cdd6c482 | 1436 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1437 | struct perf_cpu_context *cpuctx, |
6e37738a | 1438 | struct perf_event_context *ctx) |
6751b71e | 1439 | { |
6bde9b6c | 1440 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1441 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1442 | u64 now = ctx->time; |
1443 | bool simulate = false; | |
6751b71e | 1444 | |
cdd6c482 | 1445 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1446 | return 0; |
1447 | ||
ad5133b7 | 1448 | pmu->start_txn(pmu); |
6bde9b6c | 1449 | |
9ffcfa6f | 1450 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1451 | pmu->cancel_txn(pmu); |
6751b71e | 1452 | return -EAGAIN; |
90151c35 | 1453 | } |
6751b71e PM |
1454 | |
1455 | /* | |
1456 | * Schedule in siblings as one group (if any): | |
1457 | */ | |
cdd6c482 | 1458 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1459 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1460 | partial_group = event; |
6751b71e PM |
1461 | goto group_error; |
1462 | } | |
1463 | } | |
1464 | ||
9ffcfa6f | 1465 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1466 | return 0; |
9ffcfa6f | 1467 | |
6751b71e PM |
1468 | group_error: |
1469 | /* | |
1470 | * Groups can be scheduled in as one unit only, so undo any | |
1471 | * partial group before returning: | |
d7842da4 SE |
1472 | * The events up to the failed event are scheduled out normally, |
1473 | * tstamp_stopped will be updated. | |
1474 | * | |
1475 | * The failed events and the remaining siblings need to have | |
1476 | * their timings updated as if they had gone thru event_sched_in() | |
1477 | * and event_sched_out(). This is required to get consistent timings | |
1478 | * across the group. This also takes care of the case where the group | |
1479 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1480 | * the time the event was actually stopped, such that time delta | |
1481 | * calculation in update_event_times() is correct. | |
6751b71e | 1482 | */ |
cdd6c482 IM |
1483 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1484 | if (event == partial_group) | |
d7842da4 SE |
1485 | simulate = true; |
1486 | ||
1487 | if (simulate) { | |
1488 | event->tstamp_running += now - event->tstamp_stopped; | |
1489 | event->tstamp_stopped = now; | |
1490 | } else { | |
1491 | event_sched_out(event, cpuctx, ctx); | |
1492 | } | |
6751b71e | 1493 | } |
9ffcfa6f | 1494 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1495 | |
ad5133b7 | 1496 | pmu->cancel_txn(pmu); |
90151c35 | 1497 | |
6751b71e PM |
1498 | return -EAGAIN; |
1499 | } | |
1500 | ||
3b6f9e5c | 1501 | /* |
cdd6c482 | 1502 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1503 | */ |
cdd6c482 | 1504 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1505 | struct perf_cpu_context *cpuctx, |
1506 | int can_add_hw) | |
1507 | { | |
1508 | /* | |
cdd6c482 | 1509 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1510 | */ |
d6f962b5 | 1511 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1512 | return 1; |
1513 | /* | |
1514 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1515 | * events can go on. |
3b6f9e5c PM |
1516 | */ |
1517 | if (cpuctx->exclusive) | |
1518 | return 0; | |
1519 | /* | |
1520 | * If this group is exclusive and there are already | |
cdd6c482 | 1521 | * events on the CPU, it can't go on. |
3b6f9e5c | 1522 | */ |
cdd6c482 | 1523 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1524 | return 0; |
1525 | /* | |
1526 | * Otherwise, try to add it if all previous groups were able | |
1527 | * to go on. | |
1528 | */ | |
1529 | return can_add_hw; | |
1530 | } | |
1531 | ||
cdd6c482 IM |
1532 | static void add_event_to_ctx(struct perf_event *event, |
1533 | struct perf_event_context *ctx) | |
53cfbf59 | 1534 | { |
4158755d SE |
1535 | u64 tstamp = perf_event_time(event); |
1536 | ||
cdd6c482 | 1537 | list_add_event(event, ctx); |
8a49542c | 1538 | perf_group_attach(event); |
4158755d SE |
1539 | event->tstamp_enabled = tstamp; |
1540 | event->tstamp_running = tstamp; | |
1541 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1542 | } |
1543 | ||
2c29ef0f PZ |
1544 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1545 | static void | |
1546 | ctx_sched_in(struct perf_event_context *ctx, | |
1547 | struct perf_cpu_context *cpuctx, | |
1548 | enum event_type_t event_type, | |
1549 | struct task_struct *task); | |
fe4b04fa | 1550 | |
dce5855b PZ |
1551 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1552 | struct perf_event_context *ctx, | |
1553 | struct task_struct *task) | |
1554 | { | |
1555 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1556 | if (ctx) | |
1557 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1558 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1559 | if (ctx) | |
1560 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1561 | } | |
1562 | ||
0793a61d | 1563 | /* |
cdd6c482 | 1564 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1565 | * |
1566 | * Must be called with ctx->mutex held | |
0793a61d | 1567 | */ |
fe4b04fa | 1568 | static int __perf_install_in_context(void *info) |
0793a61d | 1569 | { |
cdd6c482 IM |
1570 | struct perf_event *event = info; |
1571 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1572 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1573 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1574 | struct task_struct *task = current; | |
1575 | ||
b58f6b0d | 1576 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1577 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1578 | |
1579 | /* | |
2c29ef0f | 1580 | * If there was an active task_ctx schedule it out. |
0793a61d | 1581 | */ |
b58f6b0d | 1582 | if (task_ctx) |
2c29ef0f | 1583 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1584 | |
1585 | /* | |
1586 | * If the context we're installing events in is not the | |
1587 | * active task_ctx, flip them. | |
1588 | */ | |
1589 | if (ctx->task && task_ctx != ctx) { | |
1590 | if (task_ctx) | |
1591 | raw_spin_unlock(&task_ctx->lock); | |
1592 | raw_spin_lock(&ctx->lock); | |
1593 | task_ctx = ctx; | |
1594 | } | |
1595 | ||
1596 | if (task_ctx) { | |
1597 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1598 | task = task_ctx->task; |
1599 | } | |
b58f6b0d | 1600 | |
2c29ef0f | 1601 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1602 | |
4af4998b | 1603 | update_context_time(ctx); |
e5d1367f SE |
1604 | /* |
1605 | * update cgrp time only if current cgrp | |
1606 | * matches event->cgrp. Must be done before | |
1607 | * calling add_event_to_ctx() | |
1608 | */ | |
1609 | update_cgrp_time_from_event(event); | |
0793a61d | 1610 | |
cdd6c482 | 1611 | add_event_to_ctx(event, ctx); |
0793a61d | 1612 | |
d859e29f | 1613 | /* |
2c29ef0f | 1614 | * Schedule everything back in |
d859e29f | 1615 | */ |
dce5855b | 1616 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1617 | |
1618 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1619 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1620 | |
1621 | return 0; | |
0793a61d TG |
1622 | } |
1623 | ||
1624 | /* | |
cdd6c482 | 1625 | * Attach a performance event to a context |
0793a61d | 1626 | * |
cdd6c482 IM |
1627 | * First we add the event to the list with the hardware enable bit |
1628 | * in event->hw_config cleared. | |
0793a61d | 1629 | * |
cdd6c482 | 1630 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1631 | * call to enable it in the task context. The task might have been |
1632 | * scheduled away, but we check this in the smp call again. | |
1633 | */ | |
1634 | static void | |
cdd6c482 IM |
1635 | perf_install_in_context(struct perf_event_context *ctx, |
1636 | struct perf_event *event, | |
0793a61d TG |
1637 | int cpu) |
1638 | { | |
1639 | struct task_struct *task = ctx->task; | |
1640 | ||
fe4b04fa PZ |
1641 | lockdep_assert_held(&ctx->mutex); |
1642 | ||
c3f00c70 PZ |
1643 | event->ctx = ctx; |
1644 | ||
0793a61d TG |
1645 | if (!task) { |
1646 | /* | |
cdd6c482 | 1647 | * Per cpu events are installed via an smp call and |
af901ca1 | 1648 | * the install is always successful. |
0793a61d | 1649 | */ |
fe4b04fa | 1650 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1651 | return; |
1652 | } | |
1653 | ||
0793a61d | 1654 | retry: |
fe4b04fa PZ |
1655 | if (!task_function_call(task, __perf_install_in_context, event)) |
1656 | return; | |
0793a61d | 1657 | |
e625cce1 | 1658 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1659 | /* |
fe4b04fa PZ |
1660 | * If we failed to find a running task, but find the context active now |
1661 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1662 | */ |
fe4b04fa | 1663 | if (ctx->is_active) { |
e625cce1 | 1664 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1665 | goto retry; |
1666 | } | |
1667 | ||
1668 | /* | |
fe4b04fa PZ |
1669 | * Since the task isn't running, its safe to add the event, us holding |
1670 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1671 | */ |
fe4b04fa | 1672 | add_event_to_ctx(event, ctx); |
e625cce1 | 1673 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1674 | } |
1675 | ||
fa289bec | 1676 | /* |
cdd6c482 | 1677 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1678 | * Enabling the leader of a group effectively enables all |
1679 | * the group members that aren't explicitly disabled, so we | |
1680 | * have to update their ->tstamp_enabled also. | |
1681 | * Note: this works for group members as well as group leaders | |
1682 | * since the non-leader members' sibling_lists will be empty. | |
1683 | */ | |
1d9b482e | 1684 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1685 | { |
cdd6c482 | 1686 | struct perf_event *sub; |
4158755d | 1687 | u64 tstamp = perf_event_time(event); |
fa289bec | 1688 | |
cdd6c482 | 1689 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1690 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1691 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1692 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1693 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1694 | } |
fa289bec PM |
1695 | } |
1696 | ||
d859e29f | 1697 | /* |
cdd6c482 | 1698 | * Cross CPU call to enable a performance event |
d859e29f | 1699 | */ |
fe4b04fa | 1700 | static int __perf_event_enable(void *info) |
04289bb9 | 1701 | { |
cdd6c482 | 1702 | struct perf_event *event = info; |
cdd6c482 IM |
1703 | struct perf_event_context *ctx = event->ctx; |
1704 | struct perf_event *leader = event->group_leader; | |
108b02cf | 1705 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 1706 | int err; |
04289bb9 | 1707 | |
fe4b04fa PZ |
1708 | if (WARN_ON_ONCE(!ctx->is_active)) |
1709 | return -EINVAL; | |
3cbed429 | 1710 | |
e625cce1 | 1711 | raw_spin_lock(&ctx->lock); |
4af4998b | 1712 | update_context_time(ctx); |
d859e29f | 1713 | |
cdd6c482 | 1714 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 1715 | goto unlock; |
e5d1367f SE |
1716 | |
1717 | /* | |
1718 | * set current task's cgroup time reference point | |
1719 | */ | |
3f7cce3c | 1720 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 1721 | |
1d9b482e | 1722 | __perf_event_mark_enabled(event); |
04289bb9 | 1723 | |
e5d1367f SE |
1724 | if (!event_filter_match(event)) { |
1725 | if (is_cgroup_event(event)) | |
1726 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 1727 | goto unlock; |
e5d1367f | 1728 | } |
f4c4176f | 1729 | |
04289bb9 | 1730 | /* |
cdd6c482 | 1731 | * If the event is in a group and isn't the group leader, |
d859e29f | 1732 | * then don't put it on unless the group is on. |
04289bb9 | 1733 | */ |
cdd6c482 | 1734 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 1735 | goto unlock; |
3b6f9e5c | 1736 | |
cdd6c482 | 1737 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 1738 | err = -EEXIST; |
e758a33d | 1739 | } else { |
cdd6c482 | 1740 | if (event == leader) |
6e37738a | 1741 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 1742 | else |
6e37738a | 1743 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 1744 | } |
d859e29f PM |
1745 | |
1746 | if (err) { | |
1747 | /* | |
cdd6c482 | 1748 | * If this event can't go on and it's part of a |
d859e29f PM |
1749 | * group, then the whole group has to come off. |
1750 | */ | |
cdd6c482 | 1751 | if (leader != event) |
d859e29f | 1752 | group_sched_out(leader, cpuctx, ctx); |
0d48696f | 1753 | if (leader->attr.pinned) { |
53cfbf59 | 1754 | update_group_times(leader); |
cdd6c482 | 1755 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 1756 | } |
d859e29f PM |
1757 | } |
1758 | ||
9ed6060d | 1759 | unlock: |
e625cce1 | 1760 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1761 | |
1762 | return 0; | |
d859e29f PM |
1763 | } |
1764 | ||
1765 | /* | |
cdd6c482 | 1766 | * Enable a event. |
c93f7669 | 1767 | * |
cdd6c482 IM |
1768 | * If event->ctx is a cloned context, callers must make sure that |
1769 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1770 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
1771 | * perf_event_for_each_child or perf_event_for_each as described |
1772 | * for perf_event_disable. | |
d859e29f | 1773 | */ |
44234adc | 1774 | void perf_event_enable(struct perf_event *event) |
d859e29f | 1775 | { |
cdd6c482 | 1776 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1777 | struct task_struct *task = ctx->task; |
1778 | ||
1779 | if (!task) { | |
1780 | /* | |
cdd6c482 | 1781 | * Enable the event on the cpu that it's on |
d859e29f | 1782 | */ |
fe4b04fa | 1783 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
1784 | return; |
1785 | } | |
1786 | ||
e625cce1 | 1787 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 1788 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
1789 | goto out; |
1790 | ||
1791 | /* | |
cdd6c482 IM |
1792 | * If the event is in error state, clear that first. |
1793 | * That way, if we see the event in error state below, we | |
d859e29f PM |
1794 | * know that it has gone back into error state, as distinct |
1795 | * from the task having been scheduled away before the | |
1796 | * cross-call arrived. | |
1797 | */ | |
cdd6c482 IM |
1798 | if (event->state == PERF_EVENT_STATE_ERROR) |
1799 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 1800 | |
9ed6060d | 1801 | retry: |
fe4b04fa | 1802 | if (!ctx->is_active) { |
1d9b482e | 1803 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
1804 | goto out; |
1805 | } | |
1806 | ||
e625cce1 | 1807 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1808 | |
1809 | if (!task_function_call(task, __perf_event_enable, event)) | |
1810 | return; | |
d859e29f | 1811 | |
e625cce1 | 1812 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
1813 | |
1814 | /* | |
cdd6c482 | 1815 | * If the context is active and the event is still off, |
d859e29f PM |
1816 | * we need to retry the cross-call. |
1817 | */ | |
fe4b04fa PZ |
1818 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
1819 | /* | |
1820 | * task could have been flipped by a concurrent | |
1821 | * perf_event_context_sched_out() | |
1822 | */ | |
1823 | task = ctx->task; | |
d859e29f | 1824 | goto retry; |
fe4b04fa | 1825 | } |
fa289bec | 1826 | |
9ed6060d | 1827 | out: |
e625cce1 | 1828 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1829 | } |
dcfce4a0 | 1830 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 1831 | |
26ca5c11 | 1832 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 1833 | { |
2023b359 | 1834 | /* |
cdd6c482 | 1835 | * not supported on inherited events |
2023b359 | 1836 | */ |
2e939d1d | 1837 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
1838 | return -EINVAL; |
1839 | ||
cdd6c482 IM |
1840 | atomic_add(refresh, &event->event_limit); |
1841 | perf_event_enable(event); | |
2023b359 PZ |
1842 | |
1843 | return 0; | |
79f14641 | 1844 | } |
26ca5c11 | 1845 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 1846 | |
5b0311e1 FW |
1847 | static void ctx_sched_out(struct perf_event_context *ctx, |
1848 | struct perf_cpu_context *cpuctx, | |
1849 | enum event_type_t event_type) | |
235c7fc7 | 1850 | { |
cdd6c482 | 1851 | struct perf_event *event; |
db24d33e | 1852 | int is_active = ctx->is_active; |
235c7fc7 | 1853 | |
db24d33e | 1854 | ctx->is_active &= ~event_type; |
cdd6c482 | 1855 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
1856 | return; |
1857 | ||
4af4998b | 1858 | update_context_time(ctx); |
e5d1367f | 1859 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 1860 | if (!ctx->nr_active) |
facc4307 | 1861 | return; |
5b0311e1 | 1862 | |
075e0b00 | 1863 | perf_pmu_disable(ctx->pmu); |
db24d33e | 1864 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
1865 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
1866 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 1867 | } |
889ff015 | 1868 | |
db24d33e | 1869 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 1870 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 1871 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 1872 | } |
1b9a644f | 1873 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
1874 | } |
1875 | ||
564c2b21 PM |
1876 | /* |
1877 | * Test whether two contexts are equivalent, i.e. whether they | |
1878 | * have both been cloned from the same version of the same context | |
cdd6c482 IM |
1879 | * and they both have the same number of enabled events. |
1880 | * If the number of enabled events is the same, then the set | |
1881 | * of enabled events should be the same, because these are both | |
1882 | * inherited contexts, therefore we can't access individual events | |
564c2b21 | 1883 | * in them directly with an fd; we can only enable/disable all |
cdd6c482 | 1884 | * events via prctl, or enable/disable all events in a family |
564c2b21 PM |
1885 | * via ioctl, which will have the same effect on both contexts. |
1886 | */ | |
cdd6c482 IM |
1887 | static int context_equiv(struct perf_event_context *ctx1, |
1888 | struct perf_event_context *ctx2) | |
564c2b21 PM |
1889 | { |
1890 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx | |
ad3a37de | 1891 | && ctx1->parent_gen == ctx2->parent_gen |
25346b93 | 1892 | && !ctx1->pin_count && !ctx2->pin_count; |
564c2b21 PM |
1893 | } |
1894 | ||
cdd6c482 IM |
1895 | static void __perf_event_sync_stat(struct perf_event *event, |
1896 | struct perf_event *next_event) | |
bfbd3381 PZ |
1897 | { |
1898 | u64 value; | |
1899 | ||
cdd6c482 | 1900 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
1901 | return; |
1902 | ||
1903 | /* | |
cdd6c482 | 1904 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
1905 | * because we're in the middle of a context switch and have IRQs |
1906 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 1907 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
1908 | * don't need to use it. |
1909 | */ | |
cdd6c482 IM |
1910 | switch (event->state) { |
1911 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
1912 | event->pmu->read(event); |
1913 | /* fall-through */ | |
bfbd3381 | 1914 | |
cdd6c482 IM |
1915 | case PERF_EVENT_STATE_INACTIVE: |
1916 | update_event_times(event); | |
bfbd3381 PZ |
1917 | break; |
1918 | ||
1919 | default: | |
1920 | break; | |
1921 | } | |
1922 | ||
1923 | /* | |
cdd6c482 | 1924 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
1925 | * values when we flip the contexts. |
1926 | */ | |
e7850595 PZ |
1927 | value = local64_read(&next_event->count); |
1928 | value = local64_xchg(&event->count, value); | |
1929 | local64_set(&next_event->count, value); | |
bfbd3381 | 1930 | |
cdd6c482 IM |
1931 | swap(event->total_time_enabled, next_event->total_time_enabled); |
1932 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 1933 | |
bfbd3381 | 1934 | /* |
19d2e755 | 1935 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 1936 | */ |
cdd6c482 IM |
1937 | perf_event_update_userpage(event); |
1938 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
1939 | } |
1940 | ||
1941 | #define list_next_entry(pos, member) \ | |
1942 | list_entry(pos->member.next, typeof(*pos), member) | |
1943 | ||
cdd6c482 IM |
1944 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
1945 | struct perf_event_context *next_ctx) | |
bfbd3381 | 1946 | { |
cdd6c482 | 1947 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
1948 | |
1949 | if (!ctx->nr_stat) | |
1950 | return; | |
1951 | ||
02ffdbc8 PZ |
1952 | update_context_time(ctx); |
1953 | ||
cdd6c482 IM |
1954 | event = list_first_entry(&ctx->event_list, |
1955 | struct perf_event, event_entry); | |
bfbd3381 | 1956 | |
cdd6c482 IM |
1957 | next_event = list_first_entry(&next_ctx->event_list, |
1958 | struct perf_event, event_entry); | |
bfbd3381 | 1959 | |
cdd6c482 IM |
1960 | while (&event->event_entry != &ctx->event_list && |
1961 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 1962 | |
cdd6c482 | 1963 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 1964 | |
cdd6c482 IM |
1965 | event = list_next_entry(event, event_entry); |
1966 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
1967 | } |
1968 | } | |
1969 | ||
fe4b04fa PZ |
1970 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
1971 | struct task_struct *next) | |
0793a61d | 1972 | { |
8dc85d54 | 1973 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 IM |
1974 | struct perf_event_context *next_ctx; |
1975 | struct perf_event_context *parent; | |
108b02cf | 1976 | struct perf_cpu_context *cpuctx; |
c93f7669 | 1977 | int do_switch = 1; |
0793a61d | 1978 | |
108b02cf PZ |
1979 | if (likely(!ctx)) |
1980 | return; | |
10989fb2 | 1981 | |
108b02cf PZ |
1982 | cpuctx = __get_cpu_context(ctx); |
1983 | if (!cpuctx->task_ctx) | |
0793a61d TG |
1984 | return; |
1985 | ||
c93f7669 PM |
1986 | rcu_read_lock(); |
1987 | parent = rcu_dereference(ctx->parent_ctx); | |
8dc85d54 | 1988 | next_ctx = next->perf_event_ctxp[ctxn]; |
c93f7669 PM |
1989 | if (parent && next_ctx && |
1990 | rcu_dereference(next_ctx->parent_ctx) == parent) { | |
1991 | /* | |
1992 | * Looks like the two contexts are clones, so we might be | |
1993 | * able to optimize the context switch. We lock both | |
1994 | * contexts and check that they are clones under the | |
1995 | * lock (including re-checking that neither has been | |
1996 | * uncloned in the meantime). It doesn't matter which | |
1997 | * order we take the locks because no other cpu could | |
1998 | * be trying to lock both of these tasks. | |
1999 | */ | |
e625cce1 TG |
2000 | raw_spin_lock(&ctx->lock); |
2001 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2002 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2003 | /* |
2004 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2005 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2006 | */ |
8dc85d54 PZ |
2007 | task->perf_event_ctxp[ctxn] = next_ctx; |
2008 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2009 | ctx->task = next; |
2010 | next_ctx->task = task; | |
2011 | do_switch = 0; | |
bfbd3381 | 2012 | |
cdd6c482 | 2013 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2014 | } |
e625cce1 TG |
2015 | raw_spin_unlock(&next_ctx->lock); |
2016 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2017 | } |
c93f7669 | 2018 | rcu_read_unlock(); |
564c2b21 | 2019 | |
c93f7669 | 2020 | if (do_switch) { |
facc4307 | 2021 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2022 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2023 | cpuctx->task_ctx = NULL; |
facc4307 | 2024 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2025 | } |
0793a61d TG |
2026 | } |
2027 | ||
8dc85d54 PZ |
2028 | #define for_each_task_context_nr(ctxn) \ |
2029 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2030 | ||
2031 | /* | |
2032 | * Called from scheduler to remove the events of the current task, | |
2033 | * with interrupts disabled. | |
2034 | * | |
2035 | * We stop each event and update the event value in event->count. | |
2036 | * | |
2037 | * This does not protect us against NMI, but disable() | |
2038 | * sets the disabled bit in the control field of event _before_ | |
2039 | * accessing the event control register. If a NMI hits, then it will | |
2040 | * not restart the event. | |
2041 | */ | |
82cd6def PZ |
2042 | void __perf_event_task_sched_out(struct task_struct *task, |
2043 | struct task_struct *next) | |
8dc85d54 PZ |
2044 | { |
2045 | int ctxn; | |
2046 | ||
8dc85d54 PZ |
2047 | for_each_task_context_nr(ctxn) |
2048 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2049 | |
2050 | /* | |
2051 | * if cgroup events exist on this CPU, then we need | |
2052 | * to check if we have to switch out PMU state. | |
2053 | * cgroup event are system-wide mode only | |
2054 | */ | |
2055 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2056 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2057 | } |
2058 | ||
04dc2dbb | 2059 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2060 | { |
108b02cf | 2061 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2062 | |
a63eaf34 PM |
2063 | if (!cpuctx->task_ctx) |
2064 | return; | |
012b84da IM |
2065 | |
2066 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2067 | return; | |
2068 | ||
04dc2dbb | 2069 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2070 | cpuctx->task_ctx = NULL; |
2071 | } | |
2072 | ||
5b0311e1 FW |
2073 | /* |
2074 | * Called with IRQs disabled | |
2075 | */ | |
2076 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2077 | enum event_type_t event_type) | |
2078 | { | |
2079 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2080 | } |
2081 | ||
235c7fc7 | 2082 | static void |
5b0311e1 | 2083 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2084 | struct perf_cpu_context *cpuctx) |
0793a61d | 2085 | { |
cdd6c482 | 2086 | struct perf_event *event; |
0793a61d | 2087 | |
889ff015 FW |
2088 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2089 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2090 | continue; |
5632ab12 | 2091 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2092 | continue; |
2093 | ||
e5d1367f SE |
2094 | /* may need to reset tstamp_enabled */ |
2095 | if (is_cgroup_event(event)) | |
2096 | perf_cgroup_mark_enabled(event, ctx); | |
2097 | ||
8c9ed8e1 | 2098 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2099 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2100 | |
2101 | /* | |
2102 | * If this pinned group hasn't been scheduled, | |
2103 | * put it in error state. | |
2104 | */ | |
cdd6c482 IM |
2105 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2106 | update_group_times(event); | |
2107 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2108 | } |
3b6f9e5c | 2109 | } |
5b0311e1 FW |
2110 | } |
2111 | ||
2112 | static void | |
2113 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2114 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2115 | { |
2116 | struct perf_event *event; | |
2117 | int can_add_hw = 1; | |
3b6f9e5c | 2118 | |
889ff015 FW |
2119 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2120 | /* Ignore events in OFF or ERROR state */ | |
2121 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2122 | continue; |
04289bb9 IM |
2123 | /* |
2124 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2125 | * of events: |
04289bb9 | 2126 | */ |
5632ab12 | 2127 | if (!event_filter_match(event)) |
0793a61d TG |
2128 | continue; |
2129 | ||
e5d1367f SE |
2130 | /* may need to reset tstamp_enabled */ |
2131 | if (is_cgroup_event(event)) | |
2132 | perf_cgroup_mark_enabled(event, ctx); | |
2133 | ||
9ed6060d | 2134 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2135 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2136 | can_add_hw = 0; |
9ed6060d | 2137 | } |
0793a61d | 2138 | } |
5b0311e1 FW |
2139 | } |
2140 | ||
2141 | static void | |
2142 | ctx_sched_in(struct perf_event_context *ctx, | |
2143 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2144 | enum event_type_t event_type, |
2145 | struct task_struct *task) | |
5b0311e1 | 2146 | { |
e5d1367f | 2147 | u64 now; |
db24d33e | 2148 | int is_active = ctx->is_active; |
e5d1367f | 2149 | |
db24d33e | 2150 | ctx->is_active |= event_type; |
5b0311e1 | 2151 | if (likely(!ctx->nr_events)) |
facc4307 | 2152 | return; |
5b0311e1 | 2153 | |
e5d1367f SE |
2154 | now = perf_clock(); |
2155 | ctx->timestamp = now; | |
3f7cce3c | 2156 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2157 | /* |
2158 | * First go through the list and put on any pinned groups | |
2159 | * in order to give them the best chance of going on. | |
2160 | */ | |
db24d33e | 2161 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2162 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2163 | |
2164 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2165 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2166 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2167 | } |
2168 | ||
329c0e01 | 2169 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2170 | enum event_type_t event_type, |
2171 | struct task_struct *task) | |
329c0e01 FW |
2172 | { |
2173 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2174 | ||
e5d1367f | 2175 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2176 | } |
2177 | ||
e5d1367f SE |
2178 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2179 | struct task_struct *task) | |
235c7fc7 | 2180 | { |
108b02cf | 2181 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2182 | |
108b02cf | 2183 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2184 | if (cpuctx->task_ctx == ctx) |
2185 | return; | |
2186 | ||
facc4307 | 2187 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2188 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2189 | /* |
2190 | * We want to keep the following priority order: | |
2191 | * cpu pinned (that don't need to move), task pinned, | |
2192 | * cpu flexible, task flexible. | |
2193 | */ | |
2194 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2195 | ||
1d5f003f GN |
2196 | if (ctx->nr_events) |
2197 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2198 | |
86b47c25 GN |
2199 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2200 | ||
facc4307 PZ |
2201 | perf_pmu_enable(ctx->pmu); |
2202 | perf_ctx_unlock(cpuctx, ctx); | |
2203 | ||
b5ab4cd5 PZ |
2204 | /* |
2205 | * Since these rotations are per-cpu, we need to ensure the | |
2206 | * cpu-context we got scheduled on is actually rotating. | |
2207 | */ | |
108b02cf | 2208 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2209 | } |
2210 | ||
d010b332 SE |
2211 | /* |
2212 | * When sampling the branck stack in system-wide, it may be necessary | |
2213 | * to flush the stack on context switch. This happens when the branch | |
2214 | * stack does not tag its entries with the pid of the current task. | |
2215 | * Otherwise it becomes impossible to associate a branch entry with a | |
2216 | * task. This ambiguity is more likely to appear when the branch stack | |
2217 | * supports priv level filtering and the user sets it to monitor only | |
2218 | * at the user level (which could be a useful measurement in system-wide | |
2219 | * mode). In that case, the risk is high of having a branch stack with | |
2220 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2221 | * entries or stashing them somewhere in the PMU specific code layer. | |
2222 | * | |
2223 | * This function provides the context switch callback to the lower code | |
2224 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2225 | * with at least one active event using taken branch sampling. | |
2226 | */ | |
2227 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2228 | struct task_struct *task) | |
2229 | { | |
2230 | struct perf_cpu_context *cpuctx; | |
2231 | struct pmu *pmu; | |
2232 | unsigned long flags; | |
2233 | ||
2234 | /* no need to flush branch stack if not changing task */ | |
2235 | if (prev == task) | |
2236 | return; | |
2237 | ||
2238 | local_irq_save(flags); | |
2239 | ||
2240 | rcu_read_lock(); | |
2241 | ||
2242 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2243 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2244 | ||
2245 | /* | |
2246 | * check if the context has at least one | |
2247 | * event using PERF_SAMPLE_BRANCH_STACK | |
2248 | */ | |
2249 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2250 | && pmu->flush_branch_stack) { | |
2251 | ||
2252 | pmu = cpuctx->ctx.pmu; | |
2253 | ||
2254 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2255 | ||
2256 | perf_pmu_disable(pmu); | |
2257 | ||
2258 | pmu->flush_branch_stack(); | |
2259 | ||
2260 | perf_pmu_enable(pmu); | |
2261 | ||
2262 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2263 | } | |
2264 | } | |
2265 | ||
2266 | rcu_read_unlock(); | |
2267 | ||
2268 | local_irq_restore(flags); | |
2269 | } | |
2270 | ||
8dc85d54 PZ |
2271 | /* |
2272 | * Called from scheduler to add the events of the current task | |
2273 | * with interrupts disabled. | |
2274 | * | |
2275 | * We restore the event value and then enable it. | |
2276 | * | |
2277 | * This does not protect us against NMI, but enable() | |
2278 | * sets the enabled bit in the control field of event _before_ | |
2279 | * accessing the event control register. If a NMI hits, then it will | |
2280 | * keep the event running. | |
2281 | */ | |
a8d757ef SE |
2282 | void __perf_event_task_sched_in(struct task_struct *prev, |
2283 | struct task_struct *task) | |
8dc85d54 PZ |
2284 | { |
2285 | struct perf_event_context *ctx; | |
2286 | int ctxn; | |
2287 | ||
2288 | for_each_task_context_nr(ctxn) { | |
2289 | ctx = task->perf_event_ctxp[ctxn]; | |
2290 | if (likely(!ctx)) | |
2291 | continue; | |
2292 | ||
e5d1367f | 2293 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2294 | } |
e5d1367f SE |
2295 | /* |
2296 | * if cgroup events exist on this CPU, then we need | |
2297 | * to check if we have to switch in PMU state. | |
2298 | * cgroup event are system-wide mode only | |
2299 | */ | |
2300 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2301 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2302 | |
2303 | /* check for system-wide branch_stack events */ | |
2304 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2305 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2306 | } |
2307 | ||
abd50713 PZ |
2308 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2309 | { | |
2310 | u64 frequency = event->attr.sample_freq; | |
2311 | u64 sec = NSEC_PER_SEC; | |
2312 | u64 divisor, dividend; | |
2313 | ||
2314 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2315 | ||
2316 | count_fls = fls64(count); | |
2317 | nsec_fls = fls64(nsec); | |
2318 | frequency_fls = fls64(frequency); | |
2319 | sec_fls = 30; | |
2320 | ||
2321 | /* | |
2322 | * We got @count in @nsec, with a target of sample_freq HZ | |
2323 | * the target period becomes: | |
2324 | * | |
2325 | * @count * 10^9 | |
2326 | * period = ------------------- | |
2327 | * @nsec * sample_freq | |
2328 | * | |
2329 | */ | |
2330 | ||
2331 | /* | |
2332 | * Reduce accuracy by one bit such that @a and @b converge | |
2333 | * to a similar magnitude. | |
2334 | */ | |
fe4b04fa | 2335 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2336 | do { \ |
2337 | if (a##_fls > b##_fls) { \ | |
2338 | a >>= 1; \ | |
2339 | a##_fls--; \ | |
2340 | } else { \ | |
2341 | b >>= 1; \ | |
2342 | b##_fls--; \ | |
2343 | } \ | |
2344 | } while (0) | |
2345 | ||
2346 | /* | |
2347 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2348 | * the other, so that finally we can do a u64/u64 division. | |
2349 | */ | |
2350 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2351 | REDUCE_FLS(nsec, frequency); | |
2352 | REDUCE_FLS(sec, count); | |
2353 | } | |
2354 | ||
2355 | if (count_fls + sec_fls > 64) { | |
2356 | divisor = nsec * frequency; | |
2357 | ||
2358 | while (count_fls + sec_fls > 64) { | |
2359 | REDUCE_FLS(count, sec); | |
2360 | divisor >>= 1; | |
2361 | } | |
2362 | ||
2363 | dividend = count * sec; | |
2364 | } else { | |
2365 | dividend = count * sec; | |
2366 | ||
2367 | while (nsec_fls + frequency_fls > 64) { | |
2368 | REDUCE_FLS(nsec, frequency); | |
2369 | dividend >>= 1; | |
2370 | } | |
2371 | ||
2372 | divisor = nsec * frequency; | |
2373 | } | |
2374 | ||
f6ab91ad PZ |
2375 | if (!divisor) |
2376 | return dividend; | |
2377 | ||
abd50713 PZ |
2378 | return div64_u64(dividend, divisor); |
2379 | } | |
2380 | ||
e050e3f0 SE |
2381 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2382 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2383 | ||
f39d47ff | 2384 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2385 | { |
cdd6c482 | 2386 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2387 | s64 period, sample_period; |
bd2b5b12 PZ |
2388 | s64 delta; |
2389 | ||
abd50713 | 2390 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2391 | |
2392 | delta = (s64)(period - hwc->sample_period); | |
2393 | delta = (delta + 7) / 8; /* low pass filter */ | |
2394 | ||
2395 | sample_period = hwc->sample_period + delta; | |
2396 | ||
2397 | if (!sample_period) | |
2398 | sample_period = 1; | |
2399 | ||
bd2b5b12 | 2400 | hwc->sample_period = sample_period; |
abd50713 | 2401 | |
e7850595 | 2402 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2403 | if (disable) |
2404 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2405 | ||
e7850595 | 2406 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2407 | |
2408 | if (disable) | |
2409 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2410 | } |
bd2b5b12 PZ |
2411 | } |
2412 | ||
e050e3f0 SE |
2413 | /* |
2414 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2415 | * events. At the same time, make sure, having freq events does not change | |
2416 | * the rate of unthrottling as that would introduce bias. | |
2417 | */ | |
2418 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2419 | int needs_unthr) | |
60db5e09 | 2420 | { |
cdd6c482 IM |
2421 | struct perf_event *event; |
2422 | struct hw_perf_event *hwc; | |
e050e3f0 | 2423 | u64 now, period = TICK_NSEC; |
abd50713 | 2424 | s64 delta; |
60db5e09 | 2425 | |
e050e3f0 SE |
2426 | /* |
2427 | * only need to iterate over all events iff: | |
2428 | * - context have events in frequency mode (needs freq adjust) | |
2429 | * - there are events to unthrottle on this cpu | |
2430 | */ | |
2431 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2432 | return; |
2433 | ||
e050e3f0 | 2434 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2435 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2436 | |
03541f8b | 2437 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2438 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2439 | continue; |
2440 | ||
5632ab12 | 2441 | if (!event_filter_match(event)) |
5d27c23d PZ |
2442 | continue; |
2443 | ||
cdd6c482 | 2444 | hwc = &event->hw; |
6a24ed6c | 2445 | |
e050e3f0 SE |
2446 | if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) { |
2447 | hwc->interrupts = 0; | |
cdd6c482 | 2448 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2449 | event->pmu->start(event, 0); |
a78ac325 PZ |
2450 | } |
2451 | ||
cdd6c482 | 2452 | if (!event->attr.freq || !event->attr.sample_freq) |
60db5e09 PZ |
2453 | continue; |
2454 | ||
e050e3f0 SE |
2455 | /* |
2456 | * stop the event and update event->count | |
2457 | */ | |
2458 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2459 | ||
e7850595 | 2460 | now = local64_read(&event->count); |
abd50713 PZ |
2461 | delta = now - hwc->freq_count_stamp; |
2462 | hwc->freq_count_stamp = now; | |
60db5e09 | 2463 | |
e050e3f0 SE |
2464 | /* |
2465 | * restart the event | |
2466 | * reload only if value has changed | |
f39d47ff SE |
2467 | * we have stopped the event so tell that |
2468 | * to perf_adjust_period() to avoid stopping it | |
2469 | * twice. | |
e050e3f0 | 2470 | */ |
abd50713 | 2471 | if (delta > 0) |
f39d47ff | 2472 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2473 | |
2474 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
60db5e09 | 2475 | } |
e050e3f0 | 2476 | |
f39d47ff | 2477 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2478 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2479 | } |
2480 | ||
235c7fc7 | 2481 | /* |
cdd6c482 | 2482 | * Round-robin a context's events: |
235c7fc7 | 2483 | */ |
cdd6c482 | 2484 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2485 | { |
dddd3379 TG |
2486 | /* |
2487 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2488 | * disabled by the inheritance code. | |
2489 | */ | |
2490 | if (!ctx->rotate_disable) | |
2491 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2492 | } |
2493 | ||
b5ab4cd5 | 2494 | /* |
e9d2b064 PZ |
2495 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2496 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2497 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2498 | */ |
e9d2b064 | 2499 | static void perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2500 | { |
8dc85d54 | 2501 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2502 | int rotate = 0, remove = 1; |
7fc23a53 | 2503 | |
b5ab4cd5 | 2504 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2505 | remove = 0; |
b5ab4cd5 PZ |
2506 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2507 | rotate = 1; | |
2508 | } | |
235c7fc7 | 2509 | |
8dc85d54 | 2510 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2511 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2512 | remove = 0; |
b5ab4cd5 PZ |
2513 | if (ctx->nr_events != ctx->nr_active) |
2514 | rotate = 1; | |
2515 | } | |
9717e6cd | 2516 | |
e050e3f0 | 2517 | if (!rotate) |
0f5a2601 PZ |
2518 | goto done; |
2519 | ||
facc4307 | 2520 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2521 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2522 | |
e050e3f0 SE |
2523 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2524 | if (ctx) | |
2525 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2526 | |
e050e3f0 SE |
2527 | rotate_ctx(&cpuctx->ctx); |
2528 | if (ctx) | |
2529 | rotate_ctx(ctx); | |
235c7fc7 | 2530 | |
e050e3f0 | 2531 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2532 | |
0f5a2601 PZ |
2533 | perf_pmu_enable(cpuctx->ctx.pmu); |
2534 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2535 | done: |
e9d2b064 PZ |
2536 | if (remove) |
2537 | list_del_init(&cpuctx->rotation_list); | |
e9d2b064 PZ |
2538 | } |
2539 | ||
2540 | void perf_event_task_tick(void) | |
2541 | { | |
2542 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2543 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2544 | struct perf_event_context *ctx; |
2545 | int throttled; | |
b5ab4cd5 | 2546 | |
e9d2b064 PZ |
2547 | WARN_ON(!irqs_disabled()); |
2548 | ||
e050e3f0 SE |
2549 | __this_cpu_inc(perf_throttled_seq); |
2550 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2551 | ||
e9d2b064 | 2552 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2553 | ctx = &cpuctx->ctx; |
2554 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2555 | ||
2556 | ctx = cpuctx->task_ctx; | |
2557 | if (ctx) | |
2558 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2559 | ||
e9d2b064 PZ |
2560 | if (cpuctx->jiffies_interval == 1 || |
2561 | !(jiffies % cpuctx->jiffies_interval)) | |
2562 | perf_rotate_context(cpuctx); | |
2563 | } | |
0793a61d TG |
2564 | } |
2565 | ||
889ff015 FW |
2566 | static int event_enable_on_exec(struct perf_event *event, |
2567 | struct perf_event_context *ctx) | |
2568 | { | |
2569 | if (!event->attr.enable_on_exec) | |
2570 | return 0; | |
2571 | ||
2572 | event->attr.enable_on_exec = 0; | |
2573 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2574 | return 0; | |
2575 | ||
1d9b482e | 2576 | __perf_event_mark_enabled(event); |
889ff015 FW |
2577 | |
2578 | return 1; | |
2579 | } | |
2580 | ||
57e7986e | 2581 | /* |
cdd6c482 | 2582 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2583 | * This expects task == current. |
2584 | */ | |
8dc85d54 | 2585 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2586 | { |
cdd6c482 | 2587 | struct perf_event *event; |
57e7986e PM |
2588 | unsigned long flags; |
2589 | int enabled = 0; | |
889ff015 | 2590 | int ret; |
57e7986e PM |
2591 | |
2592 | local_irq_save(flags); | |
cdd6c482 | 2593 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2594 | goto out; |
2595 | ||
e566b76e SE |
2596 | /* |
2597 | * We must ctxsw out cgroup events to avoid conflict | |
2598 | * when invoking perf_task_event_sched_in() later on | |
2599 | * in this function. Otherwise we end up trying to | |
2600 | * ctxswin cgroup events which are already scheduled | |
2601 | * in. | |
2602 | */ | |
a8d757ef | 2603 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2604 | |
e625cce1 | 2605 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2606 | task_ctx_sched_out(ctx); |
57e7986e | 2607 | |
b79387ef | 2608 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2609 | ret = event_enable_on_exec(event, ctx); |
2610 | if (ret) | |
2611 | enabled = 1; | |
57e7986e PM |
2612 | } |
2613 | ||
2614 | /* | |
cdd6c482 | 2615 | * Unclone this context if we enabled any event. |
57e7986e | 2616 | */ |
71a851b4 PZ |
2617 | if (enabled) |
2618 | unclone_ctx(ctx); | |
57e7986e | 2619 | |
e625cce1 | 2620 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2621 | |
e566b76e SE |
2622 | /* |
2623 | * Also calls ctxswin for cgroup events, if any: | |
2624 | */ | |
e5d1367f | 2625 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2626 | out: |
57e7986e PM |
2627 | local_irq_restore(flags); |
2628 | } | |
2629 | ||
0793a61d | 2630 | /* |
cdd6c482 | 2631 | * Cross CPU call to read the hardware event |
0793a61d | 2632 | */ |
cdd6c482 | 2633 | static void __perf_event_read(void *info) |
0793a61d | 2634 | { |
cdd6c482 IM |
2635 | struct perf_event *event = info; |
2636 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2637 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2638 | |
e1ac3614 PM |
2639 | /* |
2640 | * If this is a task context, we need to check whether it is | |
2641 | * the current task context of this cpu. If not it has been | |
2642 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2643 | * event->count would have been updated to a recent sample |
2644 | * when the event was scheduled out. | |
e1ac3614 PM |
2645 | */ |
2646 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2647 | return; | |
2648 | ||
e625cce1 | 2649 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2650 | if (ctx->is_active) { |
542e72fc | 2651 | update_context_time(ctx); |
e5d1367f SE |
2652 | update_cgrp_time_from_event(event); |
2653 | } | |
cdd6c482 | 2654 | update_event_times(event); |
542e72fc PZ |
2655 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2656 | event->pmu->read(event); | |
e625cce1 | 2657 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2658 | } |
2659 | ||
b5e58793 PZ |
2660 | static inline u64 perf_event_count(struct perf_event *event) |
2661 | { | |
e7850595 | 2662 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
2663 | } |
2664 | ||
cdd6c482 | 2665 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
2666 | { |
2667 | /* | |
cdd6c482 IM |
2668 | * If event is enabled and currently active on a CPU, update the |
2669 | * value in the event structure: | |
0793a61d | 2670 | */ |
cdd6c482 IM |
2671 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2672 | smp_call_function_single(event->oncpu, | |
2673 | __perf_event_read, event, 1); | |
2674 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
2675 | struct perf_event_context *ctx = event->ctx; |
2676 | unsigned long flags; | |
2677 | ||
e625cce1 | 2678 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
2679 | /* |
2680 | * may read while context is not active | |
2681 | * (e.g., thread is blocked), in that case | |
2682 | * we cannot update context time | |
2683 | */ | |
e5d1367f | 2684 | if (ctx->is_active) { |
c530ccd9 | 2685 | update_context_time(ctx); |
e5d1367f SE |
2686 | update_cgrp_time_from_event(event); |
2687 | } | |
cdd6c482 | 2688 | update_event_times(event); |
e625cce1 | 2689 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
2690 | } |
2691 | ||
b5e58793 | 2692 | return perf_event_count(event); |
0793a61d TG |
2693 | } |
2694 | ||
a63eaf34 | 2695 | /* |
cdd6c482 | 2696 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 2697 | */ |
eb184479 | 2698 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 2699 | { |
e625cce1 | 2700 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 2701 | mutex_init(&ctx->mutex); |
889ff015 FW |
2702 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2703 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
2704 | INIT_LIST_HEAD(&ctx->event_list); |
2705 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
2706 | } |
2707 | ||
2708 | static struct perf_event_context * | |
2709 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
2710 | { | |
2711 | struct perf_event_context *ctx; | |
2712 | ||
2713 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
2714 | if (!ctx) | |
2715 | return NULL; | |
2716 | ||
2717 | __perf_event_init_context(ctx); | |
2718 | if (task) { | |
2719 | ctx->task = task; | |
2720 | get_task_struct(task); | |
0793a61d | 2721 | } |
eb184479 PZ |
2722 | ctx->pmu = pmu; |
2723 | ||
2724 | return ctx; | |
a63eaf34 PM |
2725 | } |
2726 | ||
2ebd4ffb MH |
2727 | static struct task_struct * |
2728 | find_lively_task_by_vpid(pid_t vpid) | |
2729 | { | |
2730 | struct task_struct *task; | |
2731 | int err; | |
0793a61d TG |
2732 | |
2733 | rcu_read_lock(); | |
2ebd4ffb | 2734 | if (!vpid) |
0793a61d TG |
2735 | task = current; |
2736 | else | |
2ebd4ffb | 2737 | task = find_task_by_vpid(vpid); |
0793a61d TG |
2738 | if (task) |
2739 | get_task_struct(task); | |
2740 | rcu_read_unlock(); | |
2741 | ||
2742 | if (!task) | |
2743 | return ERR_PTR(-ESRCH); | |
2744 | ||
0793a61d | 2745 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
2746 | err = -EACCES; |
2747 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
2748 | goto errout; | |
2749 | ||
2ebd4ffb MH |
2750 | return task; |
2751 | errout: | |
2752 | put_task_struct(task); | |
2753 | return ERR_PTR(err); | |
2754 | ||
2755 | } | |
2756 | ||
fe4b04fa PZ |
2757 | /* |
2758 | * Returns a matching context with refcount and pincount. | |
2759 | */ | |
108b02cf | 2760 | static struct perf_event_context * |
38a81da2 | 2761 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 2762 | { |
cdd6c482 | 2763 | struct perf_event_context *ctx; |
22a4f650 | 2764 | struct perf_cpu_context *cpuctx; |
25346b93 | 2765 | unsigned long flags; |
8dc85d54 | 2766 | int ctxn, err; |
0793a61d | 2767 | |
22a4ec72 | 2768 | if (!task) { |
cdd6c482 | 2769 | /* Must be root to operate on a CPU event: */ |
0764771d | 2770 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
2771 | return ERR_PTR(-EACCES); |
2772 | ||
0793a61d | 2773 | /* |
cdd6c482 | 2774 | * We could be clever and allow to attach a event to an |
0793a61d TG |
2775 | * offline CPU and activate it when the CPU comes up, but |
2776 | * that's for later. | |
2777 | */ | |
f6325e30 | 2778 | if (!cpu_online(cpu)) |
0793a61d TG |
2779 | return ERR_PTR(-ENODEV); |
2780 | ||
108b02cf | 2781 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 2782 | ctx = &cpuctx->ctx; |
c93f7669 | 2783 | get_ctx(ctx); |
fe4b04fa | 2784 | ++ctx->pin_count; |
0793a61d | 2785 | |
0793a61d TG |
2786 | return ctx; |
2787 | } | |
2788 | ||
8dc85d54 PZ |
2789 | err = -EINVAL; |
2790 | ctxn = pmu->task_ctx_nr; | |
2791 | if (ctxn < 0) | |
2792 | goto errout; | |
2793 | ||
9ed6060d | 2794 | retry: |
8dc85d54 | 2795 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 2796 | if (ctx) { |
71a851b4 | 2797 | unclone_ctx(ctx); |
fe4b04fa | 2798 | ++ctx->pin_count; |
e625cce1 | 2799 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 2800 | } else { |
eb184479 | 2801 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
2802 | err = -ENOMEM; |
2803 | if (!ctx) | |
2804 | goto errout; | |
eb184479 | 2805 | |
dbe08d82 ON |
2806 | err = 0; |
2807 | mutex_lock(&task->perf_event_mutex); | |
2808 | /* | |
2809 | * If it has already passed perf_event_exit_task(). | |
2810 | * we must see PF_EXITING, it takes this mutex too. | |
2811 | */ | |
2812 | if (task->flags & PF_EXITING) | |
2813 | err = -ESRCH; | |
2814 | else if (task->perf_event_ctxp[ctxn]) | |
2815 | err = -EAGAIN; | |
fe4b04fa | 2816 | else { |
9137fb28 | 2817 | get_ctx(ctx); |
fe4b04fa | 2818 | ++ctx->pin_count; |
dbe08d82 | 2819 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 2820 | } |
dbe08d82 ON |
2821 | mutex_unlock(&task->perf_event_mutex); |
2822 | ||
2823 | if (unlikely(err)) { | |
9137fb28 | 2824 | put_ctx(ctx); |
dbe08d82 ON |
2825 | |
2826 | if (err == -EAGAIN) | |
2827 | goto retry; | |
2828 | goto errout; | |
a63eaf34 PM |
2829 | } |
2830 | } | |
2831 | ||
0793a61d | 2832 | return ctx; |
c93f7669 | 2833 | |
9ed6060d | 2834 | errout: |
c93f7669 | 2835 | return ERR_PTR(err); |
0793a61d TG |
2836 | } |
2837 | ||
6fb2915d LZ |
2838 | static void perf_event_free_filter(struct perf_event *event); |
2839 | ||
cdd6c482 | 2840 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 2841 | { |
cdd6c482 | 2842 | struct perf_event *event; |
592903cd | 2843 | |
cdd6c482 IM |
2844 | event = container_of(head, struct perf_event, rcu_head); |
2845 | if (event->ns) | |
2846 | put_pid_ns(event->ns); | |
6fb2915d | 2847 | perf_event_free_filter(event); |
cdd6c482 | 2848 | kfree(event); |
592903cd PZ |
2849 | } |
2850 | ||
76369139 | 2851 | static void ring_buffer_put(struct ring_buffer *rb); |
925d519a | 2852 | |
cdd6c482 | 2853 | static void free_event(struct perf_event *event) |
f1600952 | 2854 | { |
e360adbe | 2855 | irq_work_sync(&event->pending); |
925d519a | 2856 | |
cdd6c482 | 2857 | if (!event->parent) { |
82cd6def | 2858 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 2859 | static_key_slow_dec_deferred(&perf_sched_events); |
3af9e859 | 2860 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
2861 | atomic_dec(&nr_mmap_events); |
2862 | if (event->attr.comm) | |
2863 | atomic_dec(&nr_comm_events); | |
2864 | if (event->attr.task) | |
2865 | atomic_dec(&nr_task_events); | |
927c7a9e FW |
2866 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
2867 | put_callchain_buffers(); | |
08309379 PZ |
2868 | if (is_cgroup_event(event)) { |
2869 | atomic_dec(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 2870 | static_key_slow_dec_deferred(&perf_sched_events); |
08309379 | 2871 | } |
d010b332 SE |
2872 | |
2873 | if (has_branch_stack(event)) { | |
2874 | static_key_slow_dec_deferred(&perf_sched_events); | |
2875 | /* is system-wide event */ | |
2876 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
2877 | atomic_dec(&per_cpu(perf_branch_stack_events, | |
2878 | event->cpu)); | |
2879 | } | |
f344011c | 2880 | } |
9ee318a7 | 2881 | |
76369139 FW |
2882 | if (event->rb) { |
2883 | ring_buffer_put(event->rb); | |
2884 | event->rb = NULL; | |
a4be7c27 PZ |
2885 | } |
2886 | ||
e5d1367f SE |
2887 | if (is_cgroup_event(event)) |
2888 | perf_detach_cgroup(event); | |
2889 | ||
cdd6c482 IM |
2890 | if (event->destroy) |
2891 | event->destroy(event); | |
e077df4f | 2892 | |
0c67b408 PZ |
2893 | if (event->ctx) |
2894 | put_ctx(event->ctx); | |
2895 | ||
cdd6c482 | 2896 | call_rcu(&event->rcu_head, free_event_rcu); |
f1600952 PZ |
2897 | } |
2898 | ||
a66a3052 | 2899 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 2900 | { |
cdd6c482 | 2901 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 2902 | |
ad3a37de | 2903 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
2904 | /* |
2905 | * There are two ways this annotation is useful: | |
2906 | * | |
2907 | * 1) there is a lock recursion from perf_event_exit_task | |
2908 | * see the comment there. | |
2909 | * | |
2910 | * 2) there is a lock-inversion with mmap_sem through | |
2911 | * perf_event_read_group(), which takes faults while | |
2912 | * holding ctx->mutex, however this is called after | |
2913 | * the last filedesc died, so there is no possibility | |
2914 | * to trigger the AB-BA case. | |
2915 | */ | |
2916 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
050735b0 | 2917 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 2918 | perf_group_detach(event); |
050735b0 | 2919 | raw_spin_unlock_irq(&ctx->lock); |
e03a9a55 | 2920 | perf_remove_from_context(event); |
d859e29f | 2921 | mutex_unlock(&ctx->mutex); |
0793a61d | 2922 | |
cdd6c482 | 2923 | free_event(event); |
0793a61d TG |
2924 | |
2925 | return 0; | |
2926 | } | |
a66a3052 | 2927 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 2928 | |
a66a3052 PZ |
2929 | /* |
2930 | * Called when the last reference to the file is gone. | |
2931 | */ | |
2932 | static int perf_release(struct inode *inode, struct file *file) | |
fb0459d7 | 2933 | { |
a66a3052 | 2934 | struct perf_event *event = file->private_data; |
8882135b | 2935 | struct task_struct *owner; |
fb0459d7 | 2936 | |
a66a3052 | 2937 | file->private_data = NULL; |
fb0459d7 | 2938 | |
8882135b PZ |
2939 | rcu_read_lock(); |
2940 | owner = ACCESS_ONCE(event->owner); | |
2941 | /* | |
2942 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
2943 | * !owner it means the list deletion is complete and we can indeed | |
2944 | * free this event, otherwise we need to serialize on | |
2945 | * owner->perf_event_mutex. | |
2946 | */ | |
2947 | smp_read_barrier_depends(); | |
2948 | if (owner) { | |
2949 | /* | |
2950 | * Since delayed_put_task_struct() also drops the last | |
2951 | * task reference we can safely take a new reference | |
2952 | * while holding the rcu_read_lock(). | |
2953 | */ | |
2954 | get_task_struct(owner); | |
2955 | } | |
2956 | rcu_read_unlock(); | |
2957 | ||
2958 | if (owner) { | |
2959 | mutex_lock(&owner->perf_event_mutex); | |
2960 | /* | |
2961 | * We have to re-check the event->owner field, if it is cleared | |
2962 | * we raced with perf_event_exit_task(), acquiring the mutex | |
2963 | * ensured they're done, and we can proceed with freeing the | |
2964 | * event. | |
2965 | */ | |
2966 | if (event->owner) | |
2967 | list_del_init(&event->owner_entry); | |
2968 | mutex_unlock(&owner->perf_event_mutex); | |
2969 | put_task_struct(owner); | |
2970 | } | |
2971 | ||
a66a3052 | 2972 | return perf_event_release_kernel(event); |
fb0459d7 | 2973 | } |
fb0459d7 | 2974 | |
59ed446f | 2975 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 2976 | { |
cdd6c482 | 2977 | struct perf_event *child; |
e53c0994 PZ |
2978 | u64 total = 0; |
2979 | ||
59ed446f PZ |
2980 | *enabled = 0; |
2981 | *running = 0; | |
2982 | ||
6f10581a | 2983 | mutex_lock(&event->child_mutex); |
cdd6c482 | 2984 | total += perf_event_read(event); |
59ed446f PZ |
2985 | *enabled += event->total_time_enabled + |
2986 | atomic64_read(&event->child_total_time_enabled); | |
2987 | *running += event->total_time_running + | |
2988 | atomic64_read(&event->child_total_time_running); | |
2989 | ||
2990 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 2991 | total += perf_event_read(child); |
59ed446f PZ |
2992 | *enabled += child->total_time_enabled; |
2993 | *running += child->total_time_running; | |
2994 | } | |
6f10581a | 2995 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
2996 | |
2997 | return total; | |
2998 | } | |
fb0459d7 | 2999 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3000 | |
cdd6c482 | 3001 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3002 | u64 read_format, char __user *buf) |
3003 | { | |
cdd6c482 | 3004 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3005 | int n = 0, size = 0, ret = -EFAULT; |
3006 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3007 | u64 values[5]; |
59ed446f | 3008 | u64 count, enabled, running; |
abf4868b | 3009 | |
6f10581a | 3010 | mutex_lock(&ctx->mutex); |
59ed446f | 3011 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3012 | |
3013 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3014 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3015 | values[n++] = enabled; | |
3016 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3017 | values[n++] = running; | |
abf4868b PZ |
3018 | values[n++] = count; |
3019 | if (read_format & PERF_FORMAT_ID) | |
3020 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3021 | |
3022 | size = n * sizeof(u64); | |
3023 | ||
3024 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3025 | goto unlock; |
3dab77fb | 3026 | |
6f10581a | 3027 | ret = size; |
3dab77fb | 3028 | |
65abc865 | 3029 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3030 | n = 0; |
3dab77fb | 3031 | |
59ed446f | 3032 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3033 | if (read_format & PERF_FORMAT_ID) |
3034 | values[n++] = primary_event_id(sub); | |
3035 | ||
3036 | size = n * sizeof(u64); | |
3037 | ||
184d3da8 | 3038 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3039 | ret = -EFAULT; |
3040 | goto unlock; | |
3041 | } | |
abf4868b PZ |
3042 | |
3043 | ret += size; | |
3dab77fb | 3044 | } |
6f10581a PZ |
3045 | unlock: |
3046 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3047 | |
abf4868b | 3048 | return ret; |
3dab77fb PZ |
3049 | } |
3050 | ||
cdd6c482 | 3051 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3052 | u64 read_format, char __user *buf) |
3053 | { | |
59ed446f | 3054 | u64 enabled, running; |
3dab77fb PZ |
3055 | u64 values[4]; |
3056 | int n = 0; | |
3057 | ||
59ed446f PZ |
3058 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3059 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3060 | values[n++] = enabled; | |
3061 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3062 | values[n++] = running; | |
3dab77fb | 3063 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3064 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3065 | |
3066 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3067 | return -EFAULT; | |
3068 | ||
3069 | return n * sizeof(u64); | |
3070 | } | |
3071 | ||
0793a61d | 3072 | /* |
cdd6c482 | 3073 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3074 | */ |
3075 | static ssize_t | |
cdd6c482 | 3076 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3077 | { |
cdd6c482 | 3078 | u64 read_format = event->attr.read_format; |
3dab77fb | 3079 | int ret; |
0793a61d | 3080 | |
3b6f9e5c | 3081 | /* |
cdd6c482 | 3082 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3083 | * error state (i.e. because it was pinned but it couldn't be |
3084 | * scheduled on to the CPU at some point). | |
3085 | */ | |
cdd6c482 | 3086 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3087 | return 0; |
3088 | ||
c320c7b7 | 3089 | if (count < event->read_size) |
3dab77fb PZ |
3090 | return -ENOSPC; |
3091 | ||
cdd6c482 | 3092 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3093 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3094 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3095 | else |
cdd6c482 | 3096 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3097 | |
3dab77fb | 3098 | return ret; |
0793a61d TG |
3099 | } |
3100 | ||
0793a61d TG |
3101 | static ssize_t |
3102 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3103 | { | |
cdd6c482 | 3104 | struct perf_event *event = file->private_data; |
0793a61d | 3105 | |
cdd6c482 | 3106 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3107 | } |
3108 | ||
3109 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3110 | { | |
cdd6c482 | 3111 | struct perf_event *event = file->private_data; |
76369139 | 3112 | struct ring_buffer *rb; |
c33a0bc4 | 3113 | unsigned int events = POLL_HUP; |
c7138f37 | 3114 | |
10c6db11 PZ |
3115 | /* |
3116 | * Race between perf_event_set_output() and perf_poll(): perf_poll() | |
3117 | * grabs the rb reference but perf_event_set_output() overrides it. | |
3118 | * Here is the timeline for two threads T1, T2: | |
3119 | * t0: T1, rb = rcu_dereference(event->rb) | |
3120 | * t1: T2, old_rb = event->rb | |
3121 | * t2: T2, event->rb = new rb | |
3122 | * t3: T2, ring_buffer_detach(old_rb) | |
3123 | * t4: T1, ring_buffer_attach(rb1) | |
3124 | * t5: T1, poll_wait(event->waitq) | |
3125 | * | |
3126 | * To avoid this problem, we grab mmap_mutex in perf_poll() | |
3127 | * thereby ensuring that the assignment of the new ring buffer | |
3128 | * and the detachment of the old buffer appear atomic to perf_poll() | |
3129 | */ | |
3130 | mutex_lock(&event->mmap_mutex); | |
3131 | ||
c7138f37 | 3132 | rcu_read_lock(); |
76369139 | 3133 | rb = rcu_dereference(event->rb); |
10c6db11 PZ |
3134 | if (rb) { |
3135 | ring_buffer_attach(event, rb); | |
76369139 | 3136 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3137 | } |
c7138f37 | 3138 | rcu_read_unlock(); |
0793a61d | 3139 | |
10c6db11 PZ |
3140 | mutex_unlock(&event->mmap_mutex); |
3141 | ||
cdd6c482 | 3142 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3143 | |
0793a61d TG |
3144 | return events; |
3145 | } | |
3146 | ||
cdd6c482 | 3147 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3148 | { |
cdd6c482 | 3149 | (void)perf_event_read(event); |
e7850595 | 3150 | local64_set(&event->count, 0); |
cdd6c482 | 3151 | perf_event_update_userpage(event); |
3df5edad PZ |
3152 | } |
3153 | ||
c93f7669 | 3154 | /* |
cdd6c482 IM |
3155 | * Holding the top-level event's child_mutex means that any |
3156 | * descendant process that has inherited this event will block | |
3157 | * in sync_child_event if it goes to exit, thus satisfying the | |
3158 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3159 | */ |
cdd6c482 IM |
3160 | static void perf_event_for_each_child(struct perf_event *event, |
3161 | void (*func)(struct perf_event *)) | |
3df5edad | 3162 | { |
cdd6c482 | 3163 | struct perf_event *child; |
3df5edad | 3164 | |
cdd6c482 IM |
3165 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3166 | mutex_lock(&event->child_mutex); | |
3167 | func(event); | |
3168 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3169 | func(child); |
cdd6c482 | 3170 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3171 | } |
3172 | ||
cdd6c482 IM |
3173 | static void perf_event_for_each(struct perf_event *event, |
3174 | void (*func)(struct perf_event *)) | |
3df5edad | 3175 | { |
cdd6c482 IM |
3176 | struct perf_event_context *ctx = event->ctx; |
3177 | struct perf_event *sibling; | |
3df5edad | 3178 | |
75f937f2 PZ |
3179 | WARN_ON_ONCE(ctx->parent_ctx); |
3180 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3181 | event = event->group_leader; |
75f937f2 | 3182 | |
cdd6c482 IM |
3183 | perf_event_for_each_child(event, func); |
3184 | func(event); | |
3185 | list_for_each_entry(sibling, &event->sibling_list, group_entry) | |
3186 | perf_event_for_each_child(event, func); | |
75f937f2 | 3187 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3188 | } |
3189 | ||
cdd6c482 | 3190 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3191 | { |
cdd6c482 | 3192 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3193 | int ret = 0; |
3194 | u64 value; | |
3195 | ||
6c7e550f | 3196 | if (!is_sampling_event(event)) |
08247e31 PZ |
3197 | return -EINVAL; |
3198 | ||
ad0cf347 | 3199 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3200 | return -EFAULT; |
3201 | ||
3202 | if (!value) | |
3203 | return -EINVAL; | |
3204 | ||
e625cce1 | 3205 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3206 | if (event->attr.freq) { |
3207 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3208 | ret = -EINVAL; |
3209 | goto unlock; | |
3210 | } | |
3211 | ||
cdd6c482 | 3212 | event->attr.sample_freq = value; |
08247e31 | 3213 | } else { |
cdd6c482 IM |
3214 | event->attr.sample_period = value; |
3215 | event->hw.sample_period = value; | |
08247e31 PZ |
3216 | } |
3217 | unlock: | |
e625cce1 | 3218 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3219 | |
3220 | return ret; | |
3221 | } | |
3222 | ||
ac9721f3 PZ |
3223 | static const struct file_operations perf_fops; |
3224 | ||
3225 | static struct perf_event *perf_fget_light(int fd, int *fput_needed) | |
3226 | { | |
3227 | struct file *file; | |
3228 | ||
3229 | file = fget_light(fd, fput_needed); | |
3230 | if (!file) | |
3231 | return ERR_PTR(-EBADF); | |
3232 | ||
3233 | if (file->f_op != &perf_fops) { | |
3234 | fput_light(file, *fput_needed); | |
3235 | *fput_needed = 0; | |
3236 | return ERR_PTR(-EBADF); | |
3237 | } | |
3238 | ||
3239 | return file->private_data; | |
3240 | } | |
3241 | ||
3242 | static int perf_event_set_output(struct perf_event *event, | |
3243 | struct perf_event *output_event); | |
6fb2915d | 3244 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3245 | |
d859e29f PM |
3246 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3247 | { | |
cdd6c482 IM |
3248 | struct perf_event *event = file->private_data; |
3249 | void (*func)(struct perf_event *); | |
3df5edad | 3250 | u32 flags = arg; |
d859e29f PM |
3251 | |
3252 | switch (cmd) { | |
cdd6c482 IM |
3253 | case PERF_EVENT_IOC_ENABLE: |
3254 | func = perf_event_enable; | |
d859e29f | 3255 | break; |
cdd6c482 IM |
3256 | case PERF_EVENT_IOC_DISABLE: |
3257 | func = perf_event_disable; | |
79f14641 | 3258 | break; |
cdd6c482 IM |
3259 | case PERF_EVENT_IOC_RESET: |
3260 | func = perf_event_reset; | |
6de6a7b9 | 3261 | break; |
3df5edad | 3262 | |
cdd6c482 IM |
3263 | case PERF_EVENT_IOC_REFRESH: |
3264 | return perf_event_refresh(event, arg); | |
08247e31 | 3265 | |
cdd6c482 IM |
3266 | case PERF_EVENT_IOC_PERIOD: |
3267 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3268 | |
cdd6c482 | 3269 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 PZ |
3270 | { |
3271 | struct perf_event *output_event = NULL; | |
3272 | int fput_needed = 0; | |
3273 | int ret; | |
3274 | ||
3275 | if (arg != -1) { | |
3276 | output_event = perf_fget_light(arg, &fput_needed); | |
3277 | if (IS_ERR(output_event)) | |
3278 | return PTR_ERR(output_event); | |
3279 | } | |
3280 | ||
3281 | ret = perf_event_set_output(event, output_event); | |
3282 | if (output_event) | |
3283 | fput_light(output_event->filp, fput_needed); | |
3284 | ||
3285 | return ret; | |
3286 | } | |
a4be7c27 | 3287 | |
6fb2915d LZ |
3288 | case PERF_EVENT_IOC_SET_FILTER: |
3289 | return perf_event_set_filter(event, (void __user *)arg); | |
3290 | ||
d859e29f | 3291 | default: |
3df5edad | 3292 | return -ENOTTY; |
d859e29f | 3293 | } |
3df5edad PZ |
3294 | |
3295 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3296 | perf_event_for_each(event, func); |
3df5edad | 3297 | else |
cdd6c482 | 3298 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3299 | |
3300 | return 0; | |
d859e29f PM |
3301 | } |
3302 | ||
cdd6c482 | 3303 | int perf_event_task_enable(void) |
771d7cde | 3304 | { |
cdd6c482 | 3305 | struct perf_event *event; |
771d7cde | 3306 | |
cdd6c482 IM |
3307 | mutex_lock(¤t->perf_event_mutex); |
3308 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3309 | perf_event_for_each_child(event, perf_event_enable); | |
3310 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3311 | |
3312 | return 0; | |
3313 | } | |
3314 | ||
cdd6c482 | 3315 | int perf_event_task_disable(void) |
771d7cde | 3316 | { |
cdd6c482 | 3317 | struct perf_event *event; |
771d7cde | 3318 | |
cdd6c482 IM |
3319 | mutex_lock(¤t->perf_event_mutex); |
3320 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3321 | perf_event_for_each_child(event, perf_event_disable); | |
3322 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3323 | |
3324 | return 0; | |
3325 | } | |
3326 | ||
cdd6c482 | 3327 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3328 | { |
a4eaf7f1 PZ |
3329 | if (event->hw.state & PERF_HES_STOPPED) |
3330 | return 0; | |
3331 | ||
cdd6c482 | 3332 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3333 | return 0; |
3334 | ||
35edc2a5 | 3335 | return event->pmu->event_idx(event); |
194002b2 PZ |
3336 | } |
3337 | ||
c4794295 | 3338 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3339 | u64 *now, |
7f310a5d EM |
3340 | u64 *enabled, |
3341 | u64 *running) | |
c4794295 | 3342 | { |
e3f3541c | 3343 | u64 ctx_time; |
c4794295 | 3344 | |
e3f3541c PZ |
3345 | *now = perf_clock(); |
3346 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3347 | *enabled = ctx_time - event->tstamp_enabled; |
3348 | *running = ctx_time - event->tstamp_running; | |
3349 | } | |
3350 | ||
c7206205 | 3351 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3352 | { |
3353 | } | |
3354 | ||
38ff667b PZ |
3355 | /* |
3356 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3357 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3358 | * code calls this from NMI context. | |
3359 | */ | |
cdd6c482 | 3360 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3361 | { |
cdd6c482 | 3362 | struct perf_event_mmap_page *userpg; |
76369139 | 3363 | struct ring_buffer *rb; |
e3f3541c | 3364 | u64 enabled, running, now; |
38ff667b PZ |
3365 | |
3366 | rcu_read_lock(); | |
0d641208 EM |
3367 | /* |
3368 | * compute total_time_enabled, total_time_running | |
3369 | * based on snapshot values taken when the event | |
3370 | * was last scheduled in. | |
3371 | * | |
3372 | * we cannot simply called update_context_time() | |
3373 | * because of locking issue as we can be called in | |
3374 | * NMI context | |
3375 | */ | |
e3f3541c | 3376 | calc_timer_values(event, &now, &enabled, &running); |
76369139 FW |
3377 | rb = rcu_dereference(event->rb); |
3378 | if (!rb) | |
38ff667b PZ |
3379 | goto unlock; |
3380 | ||
76369139 | 3381 | userpg = rb->user_page; |
37d81828 | 3382 | |
7b732a75 PZ |
3383 | /* |
3384 | * Disable preemption so as to not let the corresponding user-space | |
3385 | * spin too long if we get preempted. | |
3386 | */ | |
3387 | preempt_disable(); | |
37d81828 | 3388 | ++userpg->lock; |
92f22a38 | 3389 | barrier(); |
cdd6c482 | 3390 | userpg->index = perf_event_index(event); |
b5e58793 | 3391 | userpg->offset = perf_event_count(event); |
365a4038 | 3392 | if (userpg->index) |
e7850595 | 3393 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3394 | |
0d641208 | 3395 | userpg->time_enabled = enabled + |
cdd6c482 | 3396 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3397 | |
0d641208 | 3398 | userpg->time_running = running + |
cdd6c482 | 3399 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3400 | |
c7206205 | 3401 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3402 | |
92f22a38 | 3403 | barrier(); |
37d81828 | 3404 | ++userpg->lock; |
7b732a75 | 3405 | preempt_enable(); |
38ff667b | 3406 | unlock: |
7b732a75 | 3407 | rcu_read_unlock(); |
37d81828 PM |
3408 | } |
3409 | ||
906010b2 PZ |
3410 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3411 | { | |
3412 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3413 | struct ring_buffer *rb; |
906010b2 PZ |
3414 | int ret = VM_FAULT_SIGBUS; |
3415 | ||
3416 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3417 | if (vmf->pgoff == 0) | |
3418 | ret = 0; | |
3419 | return ret; | |
3420 | } | |
3421 | ||
3422 | rcu_read_lock(); | |
76369139 FW |
3423 | rb = rcu_dereference(event->rb); |
3424 | if (!rb) | |
906010b2 PZ |
3425 | goto unlock; |
3426 | ||
3427 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3428 | goto unlock; | |
3429 | ||
76369139 | 3430 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3431 | if (!vmf->page) |
3432 | goto unlock; | |
3433 | ||
3434 | get_page(vmf->page); | |
3435 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3436 | vmf->page->index = vmf->pgoff; | |
3437 | ||
3438 | ret = 0; | |
3439 | unlock: | |
3440 | rcu_read_unlock(); | |
3441 | ||
3442 | return ret; | |
3443 | } | |
3444 | ||
10c6db11 PZ |
3445 | static void ring_buffer_attach(struct perf_event *event, |
3446 | struct ring_buffer *rb) | |
3447 | { | |
3448 | unsigned long flags; | |
3449 | ||
3450 | if (!list_empty(&event->rb_entry)) | |
3451 | return; | |
3452 | ||
3453 | spin_lock_irqsave(&rb->event_lock, flags); | |
3454 | if (!list_empty(&event->rb_entry)) | |
3455 | goto unlock; | |
3456 | ||
3457 | list_add(&event->rb_entry, &rb->event_list); | |
3458 | unlock: | |
3459 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3460 | } | |
3461 | ||
3462 | static void ring_buffer_detach(struct perf_event *event, | |
3463 | struct ring_buffer *rb) | |
3464 | { | |
3465 | unsigned long flags; | |
3466 | ||
3467 | if (list_empty(&event->rb_entry)) | |
3468 | return; | |
3469 | ||
3470 | spin_lock_irqsave(&rb->event_lock, flags); | |
3471 | list_del_init(&event->rb_entry); | |
3472 | wake_up_all(&event->waitq); | |
3473 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3474 | } | |
3475 | ||
3476 | static void ring_buffer_wakeup(struct perf_event *event) | |
3477 | { | |
3478 | struct ring_buffer *rb; | |
3479 | ||
3480 | rcu_read_lock(); | |
3481 | rb = rcu_dereference(event->rb); | |
44b7f4b9 WD |
3482 | if (!rb) |
3483 | goto unlock; | |
3484 | ||
3485 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
10c6db11 | 3486 | wake_up_all(&event->waitq); |
44b7f4b9 WD |
3487 | |
3488 | unlock: | |
10c6db11 PZ |
3489 | rcu_read_unlock(); |
3490 | } | |
3491 | ||
76369139 | 3492 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3493 | { |
76369139 | 3494 | struct ring_buffer *rb; |
906010b2 | 3495 | |
76369139 FW |
3496 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3497 | rb_free(rb); | |
7b732a75 PZ |
3498 | } |
3499 | ||
76369139 | 3500 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3501 | { |
76369139 | 3502 | struct ring_buffer *rb; |
7b732a75 | 3503 | |
ac9721f3 | 3504 | rcu_read_lock(); |
76369139 FW |
3505 | rb = rcu_dereference(event->rb); |
3506 | if (rb) { | |
3507 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3508 | rb = NULL; | |
ac9721f3 PZ |
3509 | } |
3510 | rcu_read_unlock(); | |
3511 | ||
76369139 | 3512 | return rb; |
ac9721f3 PZ |
3513 | } |
3514 | ||
76369139 | 3515 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3516 | { |
10c6db11 PZ |
3517 | struct perf_event *event, *n; |
3518 | unsigned long flags; | |
3519 | ||
76369139 | 3520 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3521 | return; |
7b732a75 | 3522 | |
10c6db11 PZ |
3523 | spin_lock_irqsave(&rb->event_lock, flags); |
3524 | list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) { | |
3525 | list_del_init(&event->rb_entry); | |
3526 | wake_up_all(&event->waitq); | |
3527 | } | |
3528 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3529 | ||
76369139 | 3530 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3531 | } |
3532 | ||
3533 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3534 | { | |
cdd6c482 | 3535 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3536 | |
cdd6c482 | 3537 | atomic_inc(&event->mmap_count); |
7b732a75 PZ |
3538 | } |
3539 | ||
3540 | static void perf_mmap_close(struct vm_area_struct *vma) | |
3541 | { | |
cdd6c482 | 3542 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3543 | |
cdd6c482 | 3544 | if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) { |
76369139 | 3545 | unsigned long size = perf_data_size(event->rb); |
ac9721f3 | 3546 | struct user_struct *user = event->mmap_user; |
76369139 | 3547 | struct ring_buffer *rb = event->rb; |
789f90fc | 3548 | |
906010b2 | 3549 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm); |
bc3e53f6 | 3550 | vma->vm_mm->pinned_vm -= event->mmap_locked; |
76369139 | 3551 | rcu_assign_pointer(event->rb, NULL); |
10c6db11 | 3552 | ring_buffer_detach(event, rb); |
cdd6c482 | 3553 | mutex_unlock(&event->mmap_mutex); |
ac9721f3 | 3554 | |
76369139 | 3555 | ring_buffer_put(rb); |
ac9721f3 | 3556 | free_uid(user); |
7b732a75 | 3557 | } |
37d81828 PM |
3558 | } |
3559 | ||
f0f37e2f | 3560 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3561 | .open = perf_mmap_open, |
3562 | .close = perf_mmap_close, | |
3563 | .fault = perf_mmap_fault, | |
3564 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3565 | }; |
3566 | ||
3567 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3568 | { | |
cdd6c482 | 3569 | struct perf_event *event = file->private_data; |
22a4f650 | 3570 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3571 | struct user_struct *user = current_user(); |
22a4f650 | 3572 | unsigned long locked, lock_limit; |
76369139 | 3573 | struct ring_buffer *rb; |
7b732a75 PZ |
3574 | unsigned long vma_size; |
3575 | unsigned long nr_pages; | |
789f90fc | 3576 | long user_extra, extra; |
d57e34fd | 3577 | int ret = 0, flags = 0; |
37d81828 | 3578 | |
c7920614 PZ |
3579 | /* |
3580 | * Don't allow mmap() of inherited per-task counters. This would | |
3581 | * create a performance issue due to all children writing to the | |
76369139 | 3582 | * same rb. |
c7920614 PZ |
3583 | */ |
3584 | if (event->cpu == -1 && event->attr.inherit) | |
3585 | return -EINVAL; | |
3586 | ||
43a21ea8 | 3587 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3588 | return -EINVAL; |
7b732a75 PZ |
3589 | |
3590 | vma_size = vma->vm_end - vma->vm_start; | |
3591 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3592 | ||
7730d865 | 3593 | /* |
76369139 | 3594 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3595 | * can do bitmasks instead of modulo. |
3596 | */ | |
3597 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3598 | return -EINVAL; |
3599 | ||
7b732a75 | 3600 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3601 | return -EINVAL; |
3602 | ||
7b732a75 PZ |
3603 | if (vma->vm_pgoff != 0) |
3604 | return -EINVAL; | |
37d81828 | 3605 | |
cdd6c482 IM |
3606 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3607 | mutex_lock(&event->mmap_mutex); | |
76369139 FW |
3608 | if (event->rb) { |
3609 | if (event->rb->nr_pages == nr_pages) | |
3610 | atomic_inc(&event->rb->refcount); | |
ac9721f3 | 3611 | else |
ebb3c4c4 PZ |
3612 | ret = -EINVAL; |
3613 | goto unlock; | |
3614 | } | |
3615 | ||
789f90fc | 3616 | user_extra = nr_pages + 1; |
cdd6c482 | 3617 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
3618 | |
3619 | /* | |
3620 | * Increase the limit linearly with more CPUs: | |
3621 | */ | |
3622 | user_lock_limit *= num_online_cpus(); | |
3623 | ||
789f90fc | 3624 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 3625 | |
789f90fc PZ |
3626 | extra = 0; |
3627 | if (user_locked > user_lock_limit) | |
3628 | extra = user_locked - user_lock_limit; | |
7b732a75 | 3629 | |
78d7d407 | 3630 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 3631 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 3632 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 3633 | |
459ec28a IM |
3634 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
3635 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
3636 | ret = -EPERM; |
3637 | goto unlock; | |
3638 | } | |
7b732a75 | 3639 | |
76369139 | 3640 | WARN_ON(event->rb); |
906010b2 | 3641 | |
d57e34fd | 3642 | if (vma->vm_flags & VM_WRITE) |
76369139 | 3643 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 3644 | |
4ec8363d VW |
3645 | rb = rb_alloc(nr_pages, |
3646 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
3647 | event->cpu, flags); | |
3648 | ||
76369139 | 3649 | if (!rb) { |
ac9721f3 | 3650 | ret = -ENOMEM; |
ebb3c4c4 | 3651 | goto unlock; |
ac9721f3 | 3652 | } |
76369139 | 3653 | rcu_assign_pointer(event->rb, rb); |
43a21ea8 | 3654 | |
ac9721f3 PZ |
3655 | atomic_long_add(user_extra, &user->locked_vm); |
3656 | event->mmap_locked = extra; | |
3657 | event->mmap_user = get_current_user(); | |
bc3e53f6 | 3658 | vma->vm_mm->pinned_vm += event->mmap_locked; |
ac9721f3 | 3659 | |
9a0f05cb PZ |
3660 | perf_event_update_userpage(event); |
3661 | ||
ebb3c4c4 | 3662 | unlock: |
ac9721f3 PZ |
3663 | if (!ret) |
3664 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 3665 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 3666 | |
37d81828 PM |
3667 | vma->vm_flags |= VM_RESERVED; |
3668 | vma->vm_ops = &perf_mmap_vmops; | |
7b732a75 PZ |
3669 | |
3670 | return ret; | |
37d81828 PM |
3671 | } |
3672 | ||
3c446b3d PZ |
3673 | static int perf_fasync(int fd, struct file *filp, int on) |
3674 | { | |
3c446b3d | 3675 | struct inode *inode = filp->f_path.dentry->d_inode; |
cdd6c482 | 3676 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
3677 | int retval; |
3678 | ||
3679 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 3680 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
3681 | mutex_unlock(&inode->i_mutex); |
3682 | ||
3683 | if (retval < 0) | |
3684 | return retval; | |
3685 | ||
3686 | return 0; | |
3687 | } | |
3688 | ||
0793a61d | 3689 | static const struct file_operations perf_fops = { |
3326c1ce | 3690 | .llseek = no_llseek, |
0793a61d TG |
3691 | .release = perf_release, |
3692 | .read = perf_read, | |
3693 | .poll = perf_poll, | |
d859e29f PM |
3694 | .unlocked_ioctl = perf_ioctl, |
3695 | .compat_ioctl = perf_ioctl, | |
37d81828 | 3696 | .mmap = perf_mmap, |
3c446b3d | 3697 | .fasync = perf_fasync, |
0793a61d TG |
3698 | }; |
3699 | ||
925d519a | 3700 | /* |
cdd6c482 | 3701 | * Perf event wakeup |
925d519a PZ |
3702 | * |
3703 | * If there's data, ensure we set the poll() state and publish everything | |
3704 | * to user-space before waking everybody up. | |
3705 | */ | |
3706 | ||
cdd6c482 | 3707 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 3708 | { |
10c6db11 | 3709 | ring_buffer_wakeup(event); |
4c9e2542 | 3710 | |
cdd6c482 IM |
3711 | if (event->pending_kill) { |
3712 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
3713 | event->pending_kill = 0; | |
4c9e2542 | 3714 | } |
925d519a PZ |
3715 | } |
3716 | ||
e360adbe | 3717 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 3718 | { |
cdd6c482 IM |
3719 | struct perf_event *event = container_of(entry, |
3720 | struct perf_event, pending); | |
79f14641 | 3721 | |
cdd6c482 IM |
3722 | if (event->pending_disable) { |
3723 | event->pending_disable = 0; | |
3724 | __perf_event_disable(event); | |
79f14641 PZ |
3725 | } |
3726 | ||
cdd6c482 IM |
3727 | if (event->pending_wakeup) { |
3728 | event->pending_wakeup = 0; | |
3729 | perf_event_wakeup(event); | |
79f14641 PZ |
3730 | } |
3731 | } | |
3732 | ||
39447b38 ZY |
3733 | /* |
3734 | * We assume there is only KVM supporting the callbacks. | |
3735 | * Later on, we might change it to a list if there is | |
3736 | * another virtualization implementation supporting the callbacks. | |
3737 | */ | |
3738 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
3739 | ||
3740 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3741 | { | |
3742 | perf_guest_cbs = cbs; | |
3743 | return 0; | |
3744 | } | |
3745 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
3746 | ||
3747 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3748 | { | |
3749 | perf_guest_cbs = NULL; | |
3750 | return 0; | |
3751 | } | |
3752 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
3753 | ||
c980d109 ACM |
3754 | static void __perf_event_header__init_id(struct perf_event_header *header, |
3755 | struct perf_sample_data *data, | |
3756 | struct perf_event *event) | |
6844c09d ACM |
3757 | { |
3758 | u64 sample_type = event->attr.sample_type; | |
3759 | ||
3760 | data->type = sample_type; | |
3761 | header->size += event->id_header_size; | |
3762 | ||
3763 | if (sample_type & PERF_SAMPLE_TID) { | |
3764 | /* namespace issues */ | |
3765 | data->tid_entry.pid = perf_event_pid(event, current); | |
3766 | data->tid_entry.tid = perf_event_tid(event, current); | |
3767 | } | |
3768 | ||
3769 | if (sample_type & PERF_SAMPLE_TIME) | |
3770 | data->time = perf_clock(); | |
3771 | ||
3772 | if (sample_type & PERF_SAMPLE_ID) | |
3773 | data->id = primary_event_id(event); | |
3774 | ||
3775 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3776 | data->stream_id = event->id; | |
3777 | ||
3778 | if (sample_type & PERF_SAMPLE_CPU) { | |
3779 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
3780 | data->cpu_entry.reserved = 0; | |
3781 | } | |
3782 | } | |
3783 | ||
76369139 FW |
3784 | void perf_event_header__init_id(struct perf_event_header *header, |
3785 | struct perf_sample_data *data, | |
3786 | struct perf_event *event) | |
c980d109 ACM |
3787 | { |
3788 | if (event->attr.sample_id_all) | |
3789 | __perf_event_header__init_id(header, data, event); | |
3790 | } | |
3791 | ||
3792 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
3793 | struct perf_sample_data *data) | |
3794 | { | |
3795 | u64 sample_type = data->type; | |
3796 | ||
3797 | if (sample_type & PERF_SAMPLE_TID) | |
3798 | perf_output_put(handle, data->tid_entry); | |
3799 | ||
3800 | if (sample_type & PERF_SAMPLE_TIME) | |
3801 | perf_output_put(handle, data->time); | |
3802 | ||
3803 | if (sample_type & PERF_SAMPLE_ID) | |
3804 | perf_output_put(handle, data->id); | |
3805 | ||
3806 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3807 | perf_output_put(handle, data->stream_id); | |
3808 | ||
3809 | if (sample_type & PERF_SAMPLE_CPU) | |
3810 | perf_output_put(handle, data->cpu_entry); | |
3811 | } | |
3812 | ||
76369139 FW |
3813 | void perf_event__output_id_sample(struct perf_event *event, |
3814 | struct perf_output_handle *handle, | |
3815 | struct perf_sample_data *sample) | |
c980d109 ACM |
3816 | { |
3817 | if (event->attr.sample_id_all) | |
3818 | __perf_event__output_id_sample(handle, sample); | |
3819 | } | |
3820 | ||
3dab77fb | 3821 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
3822 | struct perf_event *event, |
3823 | u64 enabled, u64 running) | |
3dab77fb | 3824 | { |
cdd6c482 | 3825 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
3826 | u64 values[4]; |
3827 | int n = 0; | |
3828 | ||
b5e58793 | 3829 | values[n++] = perf_event_count(event); |
3dab77fb | 3830 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 3831 | values[n++] = enabled + |
cdd6c482 | 3832 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
3833 | } |
3834 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 3835 | values[n++] = running + |
cdd6c482 | 3836 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
3837 | } |
3838 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 3839 | values[n++] = primary_event_id(event); |
3dab77fb | 3840 | |
76369139 | 3841 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
3842 | } |
3843 | ||
3844 | /* | |
cdd6c482 | 3845 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
3846 | */ |
3847 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
3848 | struct perf_event *event, |
3849 | u64 enabled, u64 running) | |
3dab77fb | 3850 | { |
cdd6c482 IM |
3851 | struct perf_event *leader = event->group_leader, *sub; |
3852 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
3853 | u64 values[5]; |
3854 | int n = 0; | |
3855 | ||
3856 | values[n++] = 1 + leader->nr_siblings; | |
3857 | ||
3858 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 3859 | values[n++] = enabled; |
3dab77fb PZ |
3860 | |
3861 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 3862 | values[n++] = running; |
3dab77fb | 3863 | |
cdd6c482 | 3864 | if (leader != event) |
3dab77fb PZ |
3865 | leader->pmu->read(leader); |
3866 | ||
b5e58793 | 3867 | values[n++] = perf_event_count(leader); |
3dab77fb | 3868 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3869 | values[n++] = primary_event_id(leader); |
3dab77fb | 3870 | |
76369139 | 3871 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 3872 | |
65abc865 | 3873 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
3874 | n = 0; |
3875 | ||
cdd6c482 | 3876 | if (sub != event) |
3dab77fb PZ |
3877 | sub->pmu->read(sub); |
3878 | ||
b5e58793 | 3879 | values[n++] = perf_event_count(sub); |
3dab77fb | 3880 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3881 | values[n++] = primary_event_id(sub); |
3dab77fb | 3882 | |
76369139 | 3883 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
3884 | } |
3885 | } | |
3886 | ||
eed01528 SE |
3887 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
3888 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3889 | ||
3dab77fb | 3890 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 3891 | struct perf_event *event) |
3dab77fb | 3892 | { |
e3f3541c | 3893 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
3894 | u64 read_format = event->attr.read_format; |
3895 | ||
3896 | /* | |
3897 | * compute total_time_enabled, total_time_running | |
3898 | * based on snapshot values taken when the event | |
3899 | * was last scheduled in. | |
3900 | * | |
3901 | * we cannot simply called update_context_time() | |
3902 | * because of locking issue as we are called in | |
3903 | * NMI context | |
3904 | */ | |
c4794295 | 3905 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 3906 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 3907 | |
cdd6c482 | 3908 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 3909 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 3910 | else |
eed01528 | 3911 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
3912 | } |
3913 | ||
5622f295 MM |
3914 | void perf_output_sample(struct perf_output_handle *handle, |
3915 | struct perf_event_header *header, | |
3916 | struct perf_sample_data *data, | |
cdd6c482 | 3917 | struct perf_event *event) |
5622f295 MM |
3918 | { |
3919 | u64 sample_type = data->type; | |
3920 | ||
3921 | perf_output_put(handle, *header); | |
3922 | ||
3923 | if (sample_type & PERF_SAMPLE_IP) | |
3924 | perf_output_put(handle, data->ip); | |
3925 | ||
3926 | if (sample_type & PERF_SAMPLE_TID) | |
3927 | perf_output_put(handle, data->tid_entry); | |
3928 | ||
3929 | if (sample_type & PERF_SAMPLE_TIME) | |
3930 | perf_output_put(handle, data->time); | |
3931 | ||
3932 | if (sample_type & PERF_SAMPLE_ADDR) | |
3933 | perf_output_put(handle, data->addr); | |
3934 | ||
3935 | if (sample_type & PERF_SAMPLE_ID) | |
3936 | perf_output_put(handle, data->id); | |
3937 | ||
3938 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3939 | perf_output_put(handle, data->stream_id); | |
3940 | ||
3941 | if (sample_type & PERF_SAMPLE_CPU) | |
3942 | perf_output_put(handle, data->cpu_entry); | |
3943 | ||
3944 | if (sample_type & PERF_SAMPLE_PERIOD) | |
3945 | perf_output_put(handle, data->period); | |
3946 | ||
3947 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 3948 | perf_output_read(handle, event); |
5622f295 MM |
3949 | |
3950 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
3951 | if (data->callchain) { | |
3952 | int size = 1; | |
3953 | ||
3954 | if (data->callchain) | |
3955 | size += data->callchain->nr; | |
3956 | ||
3957 | size *= sizeof(u64); | |
3958 | ||
76369139 | 3959 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
3960 | } else { |
3961 | u64 nr = 0; | |
3962 | perf_output_put(handle, nr); | |
3963 | } | |
3964 | } | |
3965 | ||
3966 | if (sample_type & PERF_SAMPLE_RAW) { | |
3967 | if (data->raw) { | |
3968 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
3969 | __output_copy(handle, data->raw->data, |
3970 | data->raw->size); | |
5622f295 MM |
3971 | } else { |
3972 | struct { | |
3973 | u32 size; | |
3974 | u32 data; | |
3975 | } raw = { | |
3976 | .size = sizeof(u32), | |
3977 | .data = 0, | |
3978 | }; | |
3979 | perf_output_put(handle, raw); | |
3980 | } | |
3981 | } | |
a7ac67ea PZ |
3982 | |
3983 | if (!event->attr.watermark) { | |
3984 | int wakeup_events = event->attr.wakeup_events; | |
3985 | ||
3986 | if (wakeup_events) { | |
3987 | struct ring_buffer *rb = handle->rb; | |
3988 | int events = local_inc_return(&rb->events); | |
3989 | ||
3990 | if (events >= wakeup_events) { | |
3991 | local_sub(wakeup_events, &rb->events); | |
3992 | local_inc(&rb->wakeup); | |
3993 | } | |
3994 | } | |
3995 | } | |
bce38cd5 SE |
3996 | |
3997 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
3998 | if (data->br_stack) { | |
3999 | size_t size; | |
4000 | ||
4001 | size = data->br_stack->nr | |
4002 | * sizeof(struct perf_branch_entry); | |
4003 | ||
4004 | perf_output_put(handle, data->br_stack->nr); | |
4005 | perf_output_copy(handle, data->br_stack->entries, size); | |
4006 | } else { | |
4007 | /* | |
4008 | * we always store at least the value of nr | |
4009 | */ | |
4010 | u64 nr = 0; | |
4011 | perf_output_put(handle, nr); | |
4012 | } | |
4013 | } | |
5622f295 MM |
4014 | } |
4015 | ||
4016 | void perf_prepare_sample(struct perf_event_header *header, | |
4017 | struct perf_sample_data *data, | |
cdd6c482 | 4018 | struct perf_event *event, |
5622f295 | 4019 | struct pt_regs *regs) |
7b732a75 | 4020 | { |
cdd6c482 | 4021 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4022 | |
cdd6c482 | 4023 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4024 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4025 | |
4026 | header->misc = 0; | |
4027 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4028 | |
c980d109 | 4029 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4030 | |
c320c7b7 | 4031 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4032 | data->ip = perf_instruction_pointer(regs); |
4033 | ||
b23f3325 | 4034 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4035 | int size = 1; |
394ee076 | 4036 | |
5622f295 MM |
4037 | data->callchain = perf_callchain(regs); |
4038 | ||
4039 | if (data->callchain) | |
4040 | size += data->callchain->nr; | |
4041 | ||
4042 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4043 | } |
4044 | ||
3a43ce68 | 4045 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4046 | int size = sizeof(u32); |
4047 | ||
4048 | if (data->raw) | |
4049 | size += data->raw->size; | |
4050 | else | |
4051 | size += sizeof(u32); | |
4052 | ||
4053 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4054 | header->size += size; |
7f453c24 | 4055 | } |
bce38cd5 SE |
4056 | |
4057 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4058 | int size = sizeof(u64); /* nr */ | |
4059 | if (data->br_stack) { | |
4060 | size += data->br_stack->nr | |
4061 | * sizeof(struct perf_branch_entry); | |
4062 | } | |
4063 | header->size += size; | |
4064 | } | |
5622f295 | 4065 | } |
7f453c24 | 4066 | |
a8b0ca17 | 4067 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4068 | struct perf_sample_data *data, |
4069 | struct pt_regs *regs) | |
4070 | { | |
4071 | struct perf_output_handle handle; | |
4072 | struct perf_event_header header; | |
689802b2 | 4073 | |
927c7a9e FW |
4074 | /* protect the callchain buffers */ |
4075 | rcu_read_lock(); | |
4076 | ||
cdd6c482 | 4077 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4078 | |
a7ac67ea | 4079 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4080 | goto exit; |
0322cd6e | 4081 | |
cdd6c482 | 4082 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4083 | |
8a057d84 | 4084 | perf_output_end(&handle); |
927c7a9e FW |
4085 | |
4086 | exit: | |
4087 | rcu_read_unlock(); | |
0322cd6e PZ |
4088 | } |
4089 | ||
38b200d6 | 4090 | /* |
cdd6c482 | 4091 | * read event_id |
38b200d6 PZ |
4092 | */ |
4093 | ||
4094 | struct perf_read_event { | |
4095 | struct perf_event_header header; | |
4096 | ||
4097 | u32 pid; | |
4098 | u32 tid; | |
38b200d6 PZ |
4099 | }; |
4100 | ||
4101 | static void | |
cdd6c482 | 4102 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4103 | struct task_struct *task) |
4104 | { | |
4105 | struct perf_output_handle handle; | |
c980d109 | 4106 | struct perf_sample_data sample; |
dfc65094 | 4107 | struct perf_read_event read_event = { |
38b200d6 | 4108 | .header = { |
cdd6c482 | 4109 | .type = PERF_RECORD_READ, |
38b200d6 | 4110 | .misc = 0, |
c320c7b7 | 4111 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4112 | }, |
cdd6c482 IM |
4113 | .pid = perf_event_pid(event, task), |
4114 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4115 | }; |
3dab77fb | 4116 | int ret; |
38b200d6 | 4117 | |
c980d109 | 4118 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4119 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4120 | if (ret) |
4121 | return; | |
4122 | ||
dfc65094 | 4123 | perf_output_put(&handle, read_event); |
cdd6c482 | 4124 | perf_output_read(&handle, event); |
c980d109 | 4125 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4126 | |
38b200d6 PZ |
4127 | perf_output_end(&handle); |
4128 | } | |
4129 | ||
60313ebe | 4130 | /* |
9f498cc5 PZ |
4131 | * task tracking -- fork/exit |
4132 | * | |
3af9e859 | 4133 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
60313ebe PZ |
4134 | */ |
4135 | ||
9f498cc5 | 4136 | struct perf_task_event { |
3a80b4a3 | 4137 | struct task_struct *task; |
cdd6c482 | 4138 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4139 | |
4140 | struct { | |
4141 | struct perf_event_header header; | |
4142 | ||
4143 | u32 pid; | |
4144 | u32 ppid; | |
9f498cc5 PZ |
4145 | u32 tid; |
4146 | u32 ptid; | |
393b2ad8 | 4147 | u64 time; |
cdd6c482 | 4148 | } event_id; |
60313ebe PZ |
4149 | }; |
4150 | ||
cdd6c482 | 4151 | static void perf_event_task_output(struct perf_event *event, |
9f498cc5 | 4152 | struct perf_task_event *task_event) |
60313ebe PZ |
4153 | { |
4154 | struct perf_output_handle handle; | |
c980d109 | 4155 | struct perf_sample_data sample; |
9f498cc5 | 4156 | struct task_struct *task = task_event->task; |
c980d109 | 4157 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4158 | |
c980d109 | 4159 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4160 | |
c980d109 | 4161 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4162 | task_event->event_id.header.size); |
ef60777c | 4163 | if (ret) |
c980d109 | 4164 | goto out; |
60313ebe | 4165 | |
cdd6c482 IM |
4166 | task_event->event_id.pid = perf_event_pid(event, task); |
4167 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4168 | |
cdd6c482 IM |
4169 | task_event->event_id.tid = perf_event_tid(event, task); |
4170 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4171 | |
cdd6c482 | 4172 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4173 | |
c980d109 ACM |
4174 | perf_event__output_id_sample(event, &handle, &sample); |
4175 | ||
60313ebe | 4176 | perf_output_end(&handle); |
c980d109 ACM |
4177 | out: |
4178 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4179 | } |
4180 | ||
cdd6c482 | 4181 | static int perf_event_task_match(struct perf_event *event) |
60313ebe | 4182 | { |
6f93d0a7 | 4183 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4184 | return 0; |
4185 | ||
5632ab12 | 4186 | if (!event_filter_match(event)) |
5d27c23d PZ |
4187 | return 0; |
4188 | ||
3af9e859 EM |
4189 | if (event->attr.comm || event->attr.mmap || |
4190 | event->attr.mmap_data || event->attr.task) | |
60313ebe PZ |
4191 | return 1; |
4192 | ||
4193 | return 0; | |
4194 | } | |
4195 | ||
cdd6c482 | 4196 | static void perf_event_task_ctx(struct perf_event_context *ctx, |
9f498cc5 | 4197 | struct perf_task_event *task_event) |
60313ebe | 4198 | { |
cdd6c482 | 4199 | struct perf_event *event; |
60313ebe | 4200 | |
cdd6c482 IM |
4201 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4202 | if (perf_event_task_match(event)) | |
4203 | perf_event_task_output(event, task_event); | |
60313ebe | 4204 | } |
60313ebe PZ |
4205 | } |
4206 | ||
cdd6c482 | 4207 | static void perf_event_task_event(struct perf_task_event *task_event) |
60313ebe PZ |
4208 | { |
4209 | struct perf_cpu_context *cpuctx; | |
8dc85d54 | 4210 | struct perf_event_context *ctx; |
108b02cf | 4211 | struct pmu *pmu; |
8dc85d54 | 4212 | int ctxn; |
60313ebe | 4213 | |
d6ff86cf | 4214 | rcu_read_lock(); |
108b02cf | 4215 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4216 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4217 | if (cpuctx->active_pmu != pmu) |
4218 | goto next; | |
108b02cf | 4219 | perf_event_task_ctx(&cpuctx->ctx, task_event); |
8dc85d54 PZ |
4220 | |
4221 | ctx = task_event->task_ctx; | |
4222 | if (!ctx) { | |
4223 | ctxn = pmu->task_ctx_nr; | |
4224 | if (ctxn < 0) | |
41945f6c | 4225 | goto next; |
8dc85d54 PZ |
4226 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); |
4227 | } | |
4228 | if (ctx) | |
4229 | perf_event_task_ctx(ctx, task_event); | |
41945f6c PZ |
4230 | next: |
4231 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4232 | } |
60313ebe PZ |
4233 | rcu_read_unlock(); |
4234 | } | |
4235 | ||
cdd6c482 IM |
4236 | static void perf_event_task(struct task_struct *task, |
4237 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4238 | int new) |
60313ebe | 4239 | { |
9f498cc5 | 4240 | struct perf_task_event task_event; |
60313ebe | 4241 | |
cdd6c482 IM |
4242 | if (!atomic_read(&nr_comm_events) && |
4243 | !atomic_read(&nr_mmap_events) && | |
4244 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4245 | return; |
4246 | ||
9f498cc5 | 4247 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4248 | .task = task, |
4249 | .task_ctx = task_ctx, | |
cdd6c482 | 4250 | .event_id = { |
60313ebe | 4251 | .header = { |
cdd6c482 | 4252 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4253 | .misc = 0, |
cdd6c482 | 4254 | .size = sizeof(task_event.event_id), |
60313ebe | 4255 | }, |
573402db PZ |
4256 | /* .pid */ |
4257 | /* .ppid */ | |
9f498cc5 PZ |
4258 | /* .tid */ |
4259 | /* .ptid */ | |
6f93d0a7 | 4260 | .time = perf_clock(), |
60313ebe PZ |
4261 | }, |
4262 | }; | |
4263 | ||
cdd6c482 | 4264 | perf_event_task_event(&task_event); |
9f498cc5 PZ |
4265 | } |
4266 | ||
cdd6c482 | 4267 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4268 | { |
cdd6c482 | 4269 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4270 | } |
4271 | ||
8d1b2d93 PZ |
4272 | /* |
4273 | * comm tracking | |
4274 | */ | |
4275 | ||
4276 | struct perf_comm_event { | |
22a4f650 IM |
4277 | struct task_struct *task; |
4278 | char *comm; | |
8d1b2d93 PZ |
4279 | int comm_size; |
4280 | ||
4281 | struct { | |
4282 | struct perf_event_header header; | |
4283 | ||
4284 | u32 pid; | |
4285 | u32 tid; | |
cdd6c482 | 4286 | } event_id; |
8d1b2d93 PZ |
4287 | }; |
4288 | ||
cdd6c482 | 4289 | static void perf_event_comm_output(struct perf_event *event, |
8d1b2d93 PZ |
4290 | struct perf_comm_event *comm_event) |
4291 | { | |
4292 | struct perf_output_handle handle; | |
c980d109 | 4293 | struct perf_sample_data sample; |
cdd6c482 | 4294 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4295 | int ret; |
4296 | ||
4297 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); | |
4298 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4299 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4300 | |
4301 | if (ret) | |
c980d109 | 4302 | goto out; |
8d1b2d93 | 4303 | |
cdd6c482 IM |
4304 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4305 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4306 | |
cdd6c482 | 4307 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4308 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4309 | comm_event->comm_size); |
c980d109 ACM |
4310 | |
4311 | perf_event__output_id_sample(event, &handle, &sample); | |
4312 | ||
8d1b2d93 | 4313 | perf_output_end(&handle); |
c980d109 ACM |
4314 | out: |
4315 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4316 | } |
4317 | ||
cdd6c482 | 4318 | static int perf_event_comm_match(struct perf_event *event) |
8d1b2d93 | 4319 | { |
6f93d0a7 | 4320 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4321 | return 0; |
4322 | ||
5632ab12 | 4323 | if (!event_filter_match(event)) |
5d27c23d PZ |
4324 | return 0; |
4325 | ||
cdd6c482 | 4326 | if (event->attr.comm) |
8d1b2d93 PZ |
4327 | return 1; |
4328 | ||
4329 | return 0; | |
4330 | } | |
4331 | ||
cdd6c482 | 4332 | static void perf_event_comm_ctx(struct perf_event_context *ctx, |
8d1b2d93 PZ |
4333 | struct perf_comm_event *comm_event) |
4334 | { | |
cdd6c482 | 4335 | struct perf_event *event; |
8d1b2d93 | 4336 | |
cdd6c482 IM |
4337 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4338 | if (perf_event_comm_match(event)) | |
4339 | perf_event_comm_output(event, comm_event); | |
8d1b2d93 | 4340 | } |
8d1b2d93 PZ |
4341 | } |
4342 | ||
cdd6c482 | 4343 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 PZ |
4344 | { |
4345 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4346 | struct perf_event_context *ctx; |
413ee3b4 | 4347 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4348 | unsigned int size; |
108b02cf | 4349 | struct pmu *pmu; |
8dc85d54 | 4350 | int ctxn; |
8d1b2d93 | 4351 | |
413ee3b4 | 4352 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4353 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4354 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4355 | |
4356 | comm_event->comm = comm; | |
4357 | comm_event->comm_size = size; | |
4358 | ||
cdd6c482 | 4359 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
f6595f3a | 4360 | rcu_read_lock(); |
108b02cf | 4361 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4362 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4363 | if (cpuctx->active_pmu != pmu) |
4364 | goto next; | |
108b02cf | 4365 | perf_event_comm_ctx(&cpuctx->ctx, comm_event); |
8dc85d54 PZ |
4366 | |
4367 | ctxn = pmu->task_ctx_nr; | |
4368 | if (ctxn < 0) | |
41945f6c | 4369 | goto next; |
8dc85d54 PZ |
4370 | |
4371 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4372 | if (ctx) | |
4373 | perf_event_comm_ctx(ctx, comm_event); | |
41945f6c PZ |
4374 | next: |
4375 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4376 | } |
665c2142 | 4377 | rcu_read_unlock(); |
8d1b2d93 PZ |
4378 | } |
4379 | ||
cdd6c482 | 4380 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4381 | { |
9ee318a7 | 4382 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4383 | struct perf_event_context *ctx; |
4384 | int ctxn; | |
9ee318a7 | 4385 | |
8dc85d54 PZ |
4386 | for_each_task_context_nr(ctxn) { |
4387 | ctx = task->perf_event_ctxp[ctxn]; | |
4388 | if (!ctx) | |
4389 | continue; | |
9ee318a7 | 4390 | |
8dc85d54 PZ |
4391 | perf_event_enable_on_exec(ctx); |
4392 | } | |
9ee318a7 | 4393 | |
cdd6c482 | 4394 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4395 | return; |
a63eaf34 | 4396 | |
9ee318a7 | 4397 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4398 | .task = task, |
573402db PZ |
4399 | /* .comm */ |
4400 | /* .comm_size */ | |
cdd6c482 | 4401 | .event_id = { |
573402db | 4402 | .header = { |
cdd6c482 | 4403 | .type = PERF_RECORD_COMM, |
573402db PZ |
4404 | .misc = 0, |
4405 | /* .size */ | |
4406 | }, | |
4407 | /* .pid */ | |
4408 | /* .tid */ | |
8d1b2d93 PZ |
4409 | }, |
4410 | }; | |
4411 | ||
cdd6c482 | 4412 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
4413 | } |
4414 | ||
0a4a9391 PZ |
4415 | /* |
4416 | * mmap tracking | |
4417 | */ | |
4418 | ||
4419 | struct perf_mmap_event { | |
089dd79d PZ |
4420 | struct vm_area_struct *vma; |
4421 | ||
4422 | const char *file_name; | |
4423 | int file_size; | |
0a4a9391 PZ |
4424 | |
4425 | struct { | |
4426 | struct perf_event_header header; | |
4427 | ||
4428 | u32 pid; | |
4429 | u32 tid; | |
4430 | u64 start; | |
4431 | u64 len; | |
4432 | u64 pgoff; | |
cdd6c482 | 4433 | } event_id; |
0a4a9391 PZ |
4434 | }; |
4435 | ||
cdd6c482 | 4436 | static void perf_event_mmap_output(struct perf_event *event, |
0a4a9391 PZ |
4437 | struct perf_mmap_event *mmap_event) |
4438 | { | |
4439 | struct perf_output_handle handle; | |
c980d109 | 4440 | struct perf_sample_data sample; |
cdd6c482 | 4441 | int size = mmap_event->event_id.header.size; |
c980d109 | 4442 | int ret; |
0a4a9391 | 4443 | |
c980d109 ACM |
4444 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
4445 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4446 | mmap_event->event_id.header.size); |
0a4a9391 | 4447 | if (ret) |
c980d109 | 4448 | goto out; |
0a4a9391 | 4449 | |
cdd6c482 IM |
4450 | mmap_event->event_id.pid = perf_event_pid(event, current); |
4451 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 4452 | |
cdd6c482 | 4453 | perf_output_put(&handle, mmap_event->event_id); |
76369139 | 4454 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 4455 | mmap_event->file_size); |
c980d109 ACM |
4456 | |
4457 | perf_event__output_id_sample(event, &handle, &sample); | |
4458 | ||
78d613eb | 4459 | perf_output_end(&handle); |
c980d109 ACM |
4460 | out: |
4461 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
4462 | } |
4463 | ||
cdd6c482 | 4464 | static int perf_event_mmap_match(struct perf_event *event, |
3af9e859 EM |
4465 | struct perf_mmap_event *mmap_event, |
4466 | int executable) | |
0a4a9391 | 4467 | { |
6f93d0a7 | 4468 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4469 | return 0; |
4470 | ||
5632ab12 | 4471 | if (!event_filter_match(event)) |
5d27c23d PZ |
4472 | return 0; |
4473 | ||
3af9e859 EM |
4474 | if ((!executable && event->attr.mmap_data) || |
4475 | (executable && event->attr.mmap)) | |
0a4a9391 PZ |
4476 | return 1; |
4477 | ||
4478 | return 0; | |
4479 | } | |
4480 | ||
cdd6c482 | 4481 | static void perf_event_mmap_ctx(struct perf_event_context *ctx, |
3af9e859 EM |
4482 | struct perf_mmap_event *mmap_event, |
4483 | int executable) | |
0a4a9391 | 4484 | { |
cdd6c482 | 4485 | struct perf_event *event; |
0a4a9391 | 4486 | |
cdd6c482 | 4487 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
3af9e859 | 4488 | if (perf_event_mmap_match(event, mmap_event, executable)) |
cdd6c482 | 4489 | perf_event_mmap_output(event, mmap_event); |
0a4a9391 | 4490 | } |
0a4a9391 PZ |
4491 | } |
4492 | ||
cdd6c482 | 4493 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 PZ |
4494 | { |
4495 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4496 | struct perf_event_context *ctx; |
089dd79d PZ |
4497 | struct vm_area_struct *vma = mmap_event->vma; |
4498 | struct file *file = vma->vm_file; | |
0a4a9391 PZ |
4499 | unsigned int size; |
4500 | char tmp[16]; | |
4501 | char *buf = NULL; | |
089dd79d | 4502 | const char *name; |
108b02cf | 4503 | struct pmu *pmu; |
8dc85d54 | 4504 | int ctxn; |
0a4a9391 | 4505 | |
413ee3b4 AB |
4506 | memset(tmp, 0, sizeof(tmp)); |
4507 | ||
0a4a9391 | 4508 | if (file) { |
413ee3b4 | 4509 | /* |
76369139 | 4510 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
4511 | * need to add enough zero bytes after the string to handle |
4512 | * the 64bit alignment we do later. | |
4513 | */ | |
4514 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
4515 | if (!buf) { |
4516 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
4517 | goto got_name; | |
4518 | } | |
d3d21c41 | 4519 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
4520 | if (IS_ERR(name)) { |
4521 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
4522 | goto got_name; | |
4523 | } | |
4524 | } else { | |
413ee3b4 AB |
4525 | if (arch_vma_name(mmap_event->vma)) { |
4526 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
4527 | sizeof(tmp)); | |
089dd79d | 4528 | goto got_name; |
413ee3b4 | 4529 | } |
089dd79d PZ |
4530 | |
4531 | if (!vma->vm_mm) { | |
4532 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
4533 | goto got_name; | |
3af9e859 EM |
4534 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
4535 | vma->vm_end >= vma->vm_mm->brk) { | |
4536 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
4537 | goto got_name; | |
4538 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
4539 | vma->vm_end >= vma->vm_mm->start_stack) { | |
4540 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
4541 | goto got_name; | |
089dd79d PZ |
4542 | } |
4543 | ||
0a4a9391 PZ |
4544 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
4545 | goto got_name; | |
4546 | } | |
4547 | ||
4548 | got_name: | |
888fcee0 | 4549 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
4550 | |
4551 | mmap_event->file_name = name; | |
4552 | mmap_event->file_size = size; | |
4553 | ||
cdd6c482 | 4554 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 4555 | |
f6d9dd23 | 4556 | rcu_read_lock(); |
108b02cf | 4557 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4558 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4559 | if (cpuctx->active_pmu != pmu) |
4560 | goto next; | |
108b02cf PZ |
4561 | perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, |
4562 | vma->vm_flags & VM_EXEC); | |
8dc85d54 PZ |
4563 | |
4564 | ctxn = pmu->task_ctx_nr; | |
4565 | if (ctxn < 0) | |
41945f6c | 4566 | goto next; |
8dc85d54 PZ |
4567 | |
4568 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4569 | if (ctx) { | |
4570 | perf_event_mmap_ctx(ctx, mmap_event, | |
4571 | vma->vm_flags & VM_EXEC); | |
4572 | } | |
41945f6c PZ |
4573 | next: |
4574 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4575 | } |
665c2142 PZ |
4576 | rcu_read_unlock(); |
4577 | ||
0a4a9391 PZ |
4578 | kfree(buf); |
4579 | } | |
4580 | ||
3af9e859 | 4581 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 4582 | { |
9ee318a7 PZ |
4583 | struct perf_mmap_event mmap_event; |
4584 | ||
cdd6c482 | 4585 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
4586 | return; |
4587 | ||
4588 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 4589 | .vma = vma, |
573402db PZ |
4590 | /* .file_name */ |
4591 | /* .file_size */ | |
cdd6c482 | 4592 | .event_id = { |
573402db | 4593 | .header = { |
cdd6c482 | 4594 | .type = PERF_RECORD_MMAP, |
39447b38 | 4595 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
4596 | /* .size */ |
4597 | }, | |
4598 | /* .pid */ | |
4599 | /* .tid */ | |
089dd79d PZ |
4600 | .start = vma->vm_start, |
4601 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 4602 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 PZ |
4603 | }, |
4604 | }; | |
4605 | ||
cdd6c482 | 4606 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
4607 | } |
4608 | ||
a78ac325 PZ |
4609 | /* |
4610 | * IRQ throttle logging | |
4611 | */ | |
4612 | ||
cdd6c482 | 4613 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
4614 | { |
4615 | struct perf_output_handle handle; | |
c980d109 | 4616 | struct perf_sample_data sample; |
a78ac325 PZ |
4617 | int ret; |
4618 | ||
4619 | struct { | |
4620 | struct perf_event_header header; | |
4621 | u64 time; | |
cca3f454 | 4622 | u64 id; |
7f453c24 | 4623 | u64 stream_id; |
a78ac325 PZ |
4624 | } throttle_event = { |
4625 | .header = { | |
cdd6c482 | 4626 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
4627 | .misc = 0, |
4628 | .size = sizeof(throttle_event), | |
4629 | }, | |
def0a9b2 | 4630 | .time = perf_clock(), |
cdd6c482 IM |
4631 | .id = primary_event_id(event), |
4632 | .stream_id = event->id, | |
a78ac325 PZ |
4633 | }; |
4634 | ||
966ee4d6 | 4635 | if (enable) |
cdd6c482 | 4636 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 4637 | |
c980d109 ACM |
4638 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
4639 | ||
4640 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4641 | throttle_event.header.size); |
a78ac325 PZ |
4642 | if (ret) |
4643 | return; | |
4644 | ||
4645 | perf_output_put(&handle, throttle_event); | |
c980d109 | 4646 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
4647 | perf_output_end(&handle); |
4648 | } | |
4649 | ||
f6c7d5fe | 4650 | /* |
cdd6c482 | 4651 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
4652 | */ |
4653 | ||
a8b0ca17 | 4654 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4655 | int throttle, struct perf_sample_data *data, |
4656 | struct pt_regs *regs) | |
f6c7d5fe | 4657 | { |
cdd6c482 IM |
4658 | int events = atomic_read(&event->event_limit); |
4659 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 4660 | u64 seq; |
79f14641 PZ |
4661 | int ret = 0; |
4662 | ||
96398826 PZ |
4663 | /* |
4664 | * Non-sampling counters might still use the PMI to fold short | |
4665 | * hardware counters, ignore those. | |
4666 | */ | |
4667 | if (unlikely(!is_sampling_event(event))) | |
4668 | return 0; | |
4669 | ||
e050e3f0 SE |
4670 | seq = __this_cpu_read(perf_throttled_seq); |
4671 | if (seq != hwc->interrupts_seq) { | |
4672 | hwc->interrupts_seq = seq; | |
4673 | hwc->interrupts = 1; | |
4674 | } else { | |
4675 | hwc->interrupts++; | |
4676 | if (unlikely(throttle | |
4677 | && hwc->interrupts >= max_samples_per_tick)) { | |
4678 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
4679 | hwc->interrupts = MAX_INTERRUPTS; |
4680 | perf_log_throttle(event, 0); | |
a78ac325 PZ |
4681 | ret = 1; |
4682 | } | |
e050e3f0 | 4683 | } |
60db5e09 | 4684 | |
cdd6c482 | 4685 | if (event->attr.freq) { |
def0a9b2 | 4686 | u64 now = perf_clock(); |
abd50713 | 4687 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 4688 | |
abd50713 | 4689 | hwc->freq_time_stamp = now; |
bd2b5b12 | 4690 | |
abd50713 | 4691 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 4692 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
4693 | } |
4694 | ||
2023b359 PZ |
4695 | /* |
4696 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 4697 | * events |
2023b359 PZ |
4698 | */ |
4699 | ||
cdd6c482 IM |
4700 | event->pending_kill = POLL_IN; |
4701 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 4702 | ret = 1; |
cdd6c482 | 4703 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
4704 | event->pending_disable = 1; |
4705 | irq_work_queue(&event->pending); | |
79f14641 PZ |
4706 | } |
4707 | ||
453f19ee | 4708 | if (event->overflow_handler) |
a8b0ca17 | 4709 | event->overflow_handler(event, data, regs); |
453f19ee | 4710 | else |
a8b0ca17 | 4711 | perf_event_output(event, data, regs); |
453f19ee | 4712 | |
f506b3dc | 4713 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
4714 | event->pending_wakeup = 1; |
4715 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
4716 | } |
4717 | ||
79f14641 | 4718 | return ret; |
f6c7d5fe PZ |
4719 | } |
4720 | ||
a8b0ca17 | 4721 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4722 | struct perf_sample_data *data, |
4723 | struct pt_regs *regs) | |
850bc73f | 4724 | { |
a8b0ca17 | 4725 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
4726 | } |
4727 | ||
15dbf27c | 4728 | /* |
cdd6c482 | 4729 | * Generic software event infrastructure |
15dbf27c PZ |
4730 | */ |
4731 | ||
b28ab83c PZ |
4732 | struct swevent_htable { |
4733 | struct swevent_hlist *swevent_hlist; | |
4734 | struct mutex hlist_mutex; | |
4735 | int hlist_refcount; | |
4736 | ||
4737 | /* Recursion avoidance in each contexts */ | |
4738 | int recursion[PERF_NR_CONTEXTS]; | |
4739 | }; | |
4740 | ||
4741 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
4742 | ||
7b4b6658 | 4743 | /* |
cdd6c482 IM |
4744 | * We directly increment event->count and keep a second value in |
4745 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
4746 | * is kept in the range [-sample_period, 0] so that we can use the |
4747 | * sign as trigger. | |
4748 | */ | |
4749 | ||
cdd6c482 | 4750 | static u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 4751 | { |
cdd6c482 | 4752 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
4753 | u64 period = hwc->last_period; |
4754 | u64 nr, offset; | |
4755 | s64 old, val; | |
4756 | ||
4757 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
4758 | |
4759 | again: | |
e7850595 | 4760 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
4761 | if (val < 0) |
4762 | return 0; | |
15dbf27c | 4763 | |
7b4b6658 PZ |
4764 | nr = div64_u64(period + val, period); |
4765 | offset = nr * period; | |
4766 | val -= offset; | |
e7850595 | 4767 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 4768 | goto again; |
15dbf27c | 4769 | |
7b4b6658 | 4770 | return nr; |
15dbf27c PZ |
4771 | } |
4772 | ||
0cff784a | 4773 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 4774 | struct perf_sample_data *data, |
5622f295 | 4775 | struct pt_regs *regs) |
15dbf27c | 4776 | { |
cdd6c482 | 4777 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 4778 | int throttle = 0; |
15dbf27c | 4779 | |
0cff784a PZ |
4780 | if (!overflow) |
4781 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 4782 | |
7b4b6658 PZ |
4783 | if (hwc->interrupts == MAX_INTERRUPTS) |
4784 | return; | |
15dbf27c | 4785 | |
7b4b6658 | 4786 | for (; overflow; overflow--) { |
a8b0ca17 | 4787 | if (__perf_event_overflow(event, throttle, |
5622f295 | 4788 | data, regs)) { |
7b4b6658 PZ |
4789 | /* |
4790 | * We inhibit the overflow from happening when | |
4791 | * hwc->interrupts == MAX_INTERRUPTS. | |
4792 | */ | |
4793 | break; | |
4794 | } | |
cf450a73 | 4795 | throttle = 1; |
7b4b6658 | 4796 | } |
15dbf27c PZ |
4797 | } |
4798 | ||
a4eaf7f1 | 4799 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 4800 | struct perf_sample_data *data, |
5622f295 | 4801 | struct pt_regs *regs) |
7b4b6658 | 4802 | { |
cdd6c482 | 4803 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 4804 | |
e7850595 | 4805 | local64_add(nr, &event->count); |
d6d020e9 | 4806 | |
0cff784a PZ |
4807 | if (!regs) |
4808 | return; | |
4809 | ||
6c7e550f | 4810 | if (!is_sampling_event(event)) |
7b4b6658 | 4811 | return; |
d6d020e9 | 4812 | |
5d81e5cf AV |
4813 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
4814 | data->period = nr; | |
4815 | return perf_swevent_overflow(event, 1, data, regs); | |
4816 | } else | |
4817 | data->period = event->hw.last_period; | |
4818 | ||
0cff784a | 4819 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 4820 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 4821 | |
e7850595 | 4822 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 4823 | return; |
df1a132b | 4824 | |
a8b0ca17 | 4825 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
4826 | } |
4827 | ||
f5ffe02e FW |
4828 | static int perf_exclude_event(struct perf_event *event, |
4829 | struct pt_regs *regs) | |
4830 | { | |
a4eaf7f1 | 4831 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 4832 | return 1; |
a4eaf7f1 | 4833 | |
f5ffe02e FW |
4834 | if (regs) { |
4835 | if (event->attr.exclude_user && user_mode(regs)) | |
4836 | return 1; | |
4837 | ||
4838 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
4839 | return 1; | |
4840 | } | |
4841 | ||
4842 | return 0; | |
4843 | } | |
4844 | ||
cdd6c482 | 4845 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 4846 | enum perf_type_id type, |
6fb2915d LZ |
4847 | u32 event_id, |
4848 | struct perf_sample_data *data, | |
4849 | struct pt_regs *regs) | |
15dbf27c | 4850 | { |
cdd6c482 | 4851 | if (event->attr.type != type) |
a21ca2ca | 4852 | return 0; |
f5ffe02e | 4853 | |
cdd6c482 | 4854 | if (event->attr.config != event_id) |
15dbf27c PZ |
4855 | return 0; |
4856 | ||
f5ffe02e FW |
4857 | if (perf_exclude_event(event, regs)) |
4858 | return 0; | |
15dbf27c PZ |
4859 | |
4860 | return 1; | |
4861 | } | |
4862 | ||
76e1d904 FW |
4863 | static inline u64 swevent_hash(u64 type, u32 event_id) |
4864 | { | |
4865 | u64 val = event_id | (type << 32); | |
4866 | ||
4867 | return hash_64(val, SWEVENT_HLIST_BITS); | |
4868 | } | |
4869 | ||
49f135ed FW |
4870 | static inline struct hlist_head * |
4871 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 4872 | { |
49f135ed FW |
4873 | u64 hash = swevent_hash(type, event_id); |
4874 | ||
4875 | return &hlist->heads[hash]; | |
4876 | } | |
76e1d904 | 4877 | |
49f135ed FW |
4878 | /* For the read side: events when they trigger */ |
4879 | static inline struct hlist_head * | |
b28ab83c | 4880 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
4881 | { |
4882 | struct swevent_hlist *hlist; | |
76e1d904 | 4883 | |
b28ab83c | 4884 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
4885 | if (!hlist) |
4886 | return NULL; | |
4887 | ||
49f135ed FW |
4888 | return __find_swevent_head(hlist, type, event_id); |
4889 | } | |
4890 | ||
4891 | /* For the event head insertion and removal in the hlist */ | |
4892 | static inline struct hlist_head * | |
b28ab83c | 4893 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
4894 | { |
4895 | struct swevent_hlist *hlist; | |
4896 | u32 event_id = event->attr.config; | |
4897 | u64 type = event->attr.type; | |
4898 | ||
4899 | /* | |
4900 | * Event scheduling is always serialized against hlist allocation | |
4901 | * and release. Which makes the protected version suitable here. | |
4902 | * The context lock guarantees that. | |
4903 | */ | |
b28ab83c | 4904 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
4905 | lockdep_is_held(&event->ctx->lock)); |
4906 | if (!hlist) | |
4907 | return NULL; | |
4908 | ||
4909 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
4910 | } |
4911 | ||
4912 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 4913 | u64 nr, |
76e1d904 FW |
4914 | struct perf_sample_data *data, |
4915 | struct pt_regs *regs) | |
15dbf27c | 4916 | { |
b28ab83c | 4917 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 4918 | struct perf_event *event; |
76e1d904 FW |
4919 | struct hlist_node *node; |
4920 | struct hlist_head *head; | |
15dbf27c | 4921 | |
76e1d904 | 4922 | rcu_read_lock(); |
b28ab83c | 4923 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
4924 | if (!head) |
4925 | goto end; | |
4926 | ||
4927 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { | |
6fb2915d | 4928 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 4929 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 4930 | } |
76e1d904 FW |
4931 | end: |
4932 | rcu_read_unlock(); | |
15dbf27c PZ |
4933 | } |
4934 | ||
4ed7c92d | 4935 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 4936 | { |
b28ab83c | 4937 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 4938 | |
b28ab83c | 4939 | return get_recursion_context(swhash->recursion); |
96f6d444 | 4940 | } |
645e8cc0 | 4941 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 4942 | |
fa9f90be | 4943 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 4944 | { |
b28ab83c | 4945 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 4946 | |
b28ab83c | 4947 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 4948 | } |
15dbf27c | 4949 | |
a8b0ca17 | 4950 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 4951 | { |
a4234bfc | 4952 | struct perf_sample_data data; |
4ed7c92d PZ |
4953 | int rctx; |
4954 | ||
1c024eca | 4955 | preempt_disable_notrace(); |
4ed7c92d PZ |
4956 | rctx = perf_swevent_get_recursion_context(); |
4957 | if (rctx < 0) | |
4958 | return; | |
a4234bfc | 4959 | |
dc1d628a | 4960 | perf_sample_data_init(&data, addr); |
92bf309a | 4961 | |
a8b0ca17 | 4962 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
4963 | |
4964 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 4965 | preempt_enable_notrace(); |
b8e83514 PZ |
4966 | } |
4967 | ||
cdd6c482 | 4968 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 4969 | { |
15dbf27c PZ |
4970 | } |
4971 | ||
a4eaf7f1 | 4972 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 4973 | { |
b28ab83c | 4974 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 4975 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
4976 | struct hlist_head *head; |
4977 | ||
6c7e550f | 4978 | if (is_sampling_event(event)) { |
7b4b6658 | 4979 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 4980 | perf_swevent_set_period(event); |
7b4b6658 | 4981 | } |
76e1d904 | 4982 | |
a4eaf7f1 PZ |
4983 | hwc->state = !(flags & PERF_EF_START); |
4984 | ||
b28ab83c | 4985 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
4986 | if (WARN_ON_ONCE(!head)) |
4987 | return -EINVAL; | |
4988 | ||
4989 | hlist_add_head_rcu(&event->hlist_entry, head); | |
4990 | ||
15dbf27c PZ |
4991 | return 0; |
4992 | } | |
4993 | ||
a4eaf7f1 | 4994 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 4995 | { |
76e1d904 | 4996 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
4997 | } |
4998 | ||
a4eaf7f1 | 4999 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5000 | { |
a4eaf7f1 | 5001 | event->hw.state = 0; |
d6d020e9 | 5002 | } |
aa9c4c0f | 5003 | |
a4eaf7f1 | 5004 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5005 | { |
a4eaf7f1 | 5006 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5007 | } |
5008 | ||
49f135ed FW |
5009 | /* Deref the hlist from the update side */ |
5010 | static inline struct swevent_hlist * | |
b28ab83c | 5011 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5012 | { |
b28ab83c PZ |
5013 | return rcu_dereference_protected(swhash->swevent_hlist, |
5014 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5015 | } |
5016 | ||
b28ab83c | 5017 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5018 | { |
b28ab83c | 5019 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5020 | |
49f135ed | 5021 | if (!hlist) |
76e1d904 FW |
5022 | return; |
5023 | ||
b28ab83c | 5024 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5025 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5026 | } |
5027 | ||
5028 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5029 | { | |
b28ab83c | 5030 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5031 | |
b28ab83c | 5032 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5033 | |
b28ab83c PZ |
5034 | if (!--swhash->hlist_refcount) |
5035 | swevent_hlist_release(swhash); | |
76e1d904 | 5036 | |
b28ab83c | 5037 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5038 | } |
5039 | ||
5040 | static void swevent_hlist_put(struct perf_event *event) | |
5041 | { | |
5042 | int cpu; | |
5043 | ||
5044 | if (event->cpu != -1) { | |
5045 | swevent_hlist_put_cpu(event, event->cpu); | |
5046 | return; | |
5047 | } | |
5048 | ||
5049 | for_each_possible_cpu(cpu) | |
5050 | swevent_hlist_put_cpu(event, cpu); | |
5051 | } | |
5052 | ||
5053 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5054 | { | |
b28ab83c | 5055 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5056 | int err = 0; |
5057 | ||
b28ab83c | 5058 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5059 | |
b28ab83c | 5060 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5061 | struct swevent_hlist *hlist; |
5062 | ||
5063 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5064 | if (!hlist) { | |
5065 | err = -ENOMEM; | |
5066 | goto exit; | |
5067 | } | |
b28ab83c | 5068 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5069 | } |
b28ab83c | 5070 | swhash->hlist_refcount++; |
9ed6060d | 5071 | exit: |
b28ab83c | 5072 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5073 | |
5074 | return err; | |
5075 | } | |
5076 | ||
5077 | static int swevent_hlist_get(struct perf_event *event) | |
5078 | { | |
5079 | int err; | |
5080 | int cpu, failed_cpu; | |
5081 | ||
5082 | if (event->cpu != -1) | |
5083 | return swevent_hlist_get_cpu(event, event->cpu); | |
5084 | ||
5085 | get_online_cpus(); | |
5086 | for_each_possible_cpu(cpu) { | |
5087 | err = swevent_hlist_get_cpu(event, cpu); | |
5088 | if (err) { | |
5089 | failed_cpu = cpu; | |
5090 | goto fail; | |
5091 | } | |
5092 | } | |
5093 | put_online_cpus(); | |
5094 | ||
5095 | return 0; | |
9ed6060d | 5096 | fail: |
76e1d904 FW |
5097 | for_each_possible_cpu(cpu) { |
5098 | if (cpu == failed_cpu) | |
5099 | break; | |
5100 | swevent_hlist_put_cpu(event, cpu); | |
5101 | } | |
5102 | ||
5103 | put_online_cpus(); | |
5104 | return err; | |
5105 | } | |
5106 | ||
c5905afb | 5107 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5108 | |
b0a873eb PZ |
5109 | static void sw_perf_event_destroy(struct perf_event *event) |
5110 | { | |
5111 | u64 event_id = event->attr.config; | |
95476b64 | 5112 | |
b0a873eb PZ |
5113 | WARN_ON(event->parent); |
5114 | ||
c5905afb | 5115 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5116 | swevent_hlist_put(event); |
5117 | } | |
5118 | ||
5119 | static int perf_swevent_init(struct perf_event *event) | |
5120 | { | |
5121 | int event_id = event->attr.config; | |
5122 | ||
5123 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5124 | return -ENOENT; | |
5125 | ||
2481c5fa SE |
5126 | /* |
5127 | * no branch sampling for software events | |
5128 | */ | |
5129 | if (has_branch_stack(event)) | |
5130 | return -EOPNOTSUPP; | |
5131 | ||
b0a873eb PZ |
5132 | switch (event_id) { |
5133 | case PERF_COUNT_SW_CPU_CLOCK: | |
5134 | case PERF_COUNT_SW_TASK_CLOCK: | |
5135 | return -ENOENT; | |
5136 | ||
5137 | default: | |
5138 | break; | |
5139 | } | |
5140 | ||
ce677831 | 5141 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5142 | return -ENOENT; |
5143 | ||
5144 | if (!event->parent) { | |
5145 | int err; | |
5146 | ||
5147 | err = swevent_hlist_get(event); | |
5148 | if (err) | |
5149 | return err; | |
5150 | ||
c5905afb | 5151 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5152 | event->destroy = sw_perf_event_destroy; |
5153 | } | |
5154 | ||
5155 | return 0; | |
5156 | } | |
5157 | ||
35edc2a5 PZ |
5158 | static int perf_swevent_event_idx(struct perf_event *event) |
5159 | { | |
5160 | return 0; | |
5161 | } | |
5162 | ||
b0a873eb | 5163 | static struct pmu perf_swevent = { |
89a1e187 | 5164 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5165 | |
b0a873eb | 5166 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5167 | .add = perf_swevent_add, |
5168 | .del = perf_swevent_del, | |
5169 | .start = perf_swevent_start, | |
5170 | .stop = perf_swevent_stop, | |
1c024eca | 5171 | .read = perf_swevent_read, |
35edc2a5 PZ |
5172 | |
5173 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5174 | }; |
5175 | ||
b0a873eb PZ |
5176 | #ifdef CONFIG_EVENT_TRACING |
5177 | ||
1c024eca PZ |
5178 | static int perf_tp_filter_match(struct perf_event *event, |
5179 | struct perf_sample_data *data) | |
5180 | { | |
5181 | void *record = data->raw->data; | |
5182 | ||
5183 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5184 | return 1; | |
5185 | return 0; | |
5186 | } | |
5187 | ||
5188 | static int perf_tp_event_match(struct perf_event *event, | |
5189 | struct perf_sample_data *data, | |
5190 | struct pt_regs *regs) | |
5191 | { | |
a0f7d0f7 FW |
5192 | if (event->hw.state & PERF_HES_STOPPED) |
5193 | return 0; | |
580d607c PZ |
5194 | /* |
5195 | * All tracepoints are from kernel-space. | |
5196 | */ | |
5197 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5198 | return 0; |
5199 | ||
5200 | if (!perf_tp_filter_match(event, data)) | |
5201 | return 0; | |
5202 | ||
5203 | return 1; | |
5204 | } | |
5205 | ||
5206 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
ecc55f84 | 5207 | struct pt_regs *regs, struct hlist_head *head, int rctx) |
95476b64 FW |
5208 | { |
5209 | struct perf_sample_data data; | |
1c024eca PZ |
5210 | struct perf_event *event; |
5211 | struct hlist_node *node; | |
5212 | ||
95476b64 FW |
5213 | struct perf_raw_record raw = { |
5214 | .size = entry_size, | |
5215 | .data = record, | |
5216 | }; | |
5217 | ||
5218 | perf_sample_data_init(&data, addr); | |
5219 | data.raw = &raw; | |
5220 | ||
1c024eca PZ |
5221 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { |
5222 | if (perf_tp_event_match(event, &data, regs)) | |
a8b0ca17 | 5223 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5224 | } |
ecc55f84 PZ |
5225 | |
5226 | perf_swevent_put_recursion_context(rctx); | |
95476b64 FW |
5227 | } |
5228 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5229 | ||
cdd6c482 | 5230 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5231 | { |
1c024eca | 5232 | perf_trace_destroy(event); |
e077df4f PZ |
5233 | } |
5234 | ||
b0a873eb | 5235 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5236 | { |
76e1d904 FW |
5237 | int err; |
5238 | ||
b0a873eb PZ |
5239 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5240 | return -ENOENT; | |
5241 | ||
2481c5fa SE |
5242 | /* |
5243 | * no branch sampling for tracepoint events | |
5244 | */ | |
5245 | if (has_branch_stack(event)) | |
5246 | return -EOPNOTSUPP; | |
5247 | ||
1c024eca PZ |
5248 | err = perf_trace_init(event); |
5249 | if (err) | |
b0a873eb | 5250 | return err; |
e077df4f | 5251 | |
cdd6c482 | 5252 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5253 | |
b0a873eb PZ |
5254 | return 0; |
5255 | } | |
5256 | ||
5257 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5258 | .task_ctx_nr = perf_sw_context, |
5259 | ||
b0a873eb | 5260 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5261 | .add = perf_trace_add, |
5262 | .del = perf_trace_del, | |
5263 | .start = perf_swevent_start, | |
5264 | .stop = perf_swevent_stop, | |
b0a873eb | 5265 | .read = perf_swevent_read, |
35edc2a5 PZ |
5266 | |
5267 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5268 | }; |
5269 | ||
5270 | static inline void perf_tp_register(void) | |
5271 | { | |
2e80a82a | 5272 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5273 | } |
6fb2915d LZ |
5274 | |
5275 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5276 | { | |
5277 | char *filter_str; | |
5278 | int ret; | |
5279 | ||
5280 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5281 | return -EINVAL; | |
5282 | ||
5283 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5284 | if (IS_ERR(filter_str)) | |
5285 | return PTR_ERR(filter_str); | |
5286 | ||
5287 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5288 | ||
5289 | kfree(filter_str); | |
5290 | return ret; | |
5291 | } | |
5292 | ||
5293 | static void perf_event_free_filter(struct perf_event *event) | |
5294 | { | |
5295 | ftrace_profile_free_filter(event); | |
5296 | } | |
5297 | ||
e077df4f | 5298 | #else |
6fb2915d | 5299 | |
b0a873eb | 5300 | static inline void perf_tp_register(void) |
e077df4f | 5301 | { |
e077df4f | 5302 | } |
6fb2915d LZ |
5303 | |
5304 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5305 | { | |
5306 | return -ENOENT; | |
5307 | } | |
5308 | ||
5309 | static void perf_event_free_filter(struct perf_event *event) | |
5310 | { | |
5311 | } | |
5312 | ||
07b139c8 | 5313 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5314 | |
24f1e32c | 5315 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5316 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5317 | { |
f5ffe02e FW |
5318 | struct perf_sample_data sample; |
5319 | struct pt_regs *regs = data; | |
5320 | ||
dc1d628a | 5321 | perf_sample_data_init(&sample, bp->attr.bp_addr); |
f5ffe02e | 5322 | |
a4eaf7f1 | 5323 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5324 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5325 | } |
5326 | #endif | |
5327 | ||
b0a873eb PZ |
5328 | /* |
5329 | * hrtimer based swevent callback | |
5330 | */ | |
f29ac756 | 5331 | |
b0a873eb | 5332 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5333 | { |
b0a873eb PZ |
5334 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5335 | struct perf_sample_data data; | |
5336 | struct pt_regs *regs; | |
5337 | struct perf_event *event; | |
5338 | u64 period; | |
f29ac756 | 5339 | |
b0a873eb | 5340 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5341 | |
5342 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5343 | return HRTIMER_NORESTART; | |
5344 | ||
b0a873eb | 5345 | event->pmu->read(event); |
f344011c | 5346 | |
b0a873eb PZ |
5347 | perf_sample_data_init(&data, 0); |
5348 | data.period = event->hw.last_period; | |
5349 | regs = get_irq_regs(); | |
5350 | ||
5351 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5352 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
a8b0ca17 | 5353 | if (perf_event_overflow(event, &data, regs)) |
b0a873eb PZ |
5354 | ret = HRTIMER_NORESTART; |
5355 | } | |
24f1e32c | 5356 | |
b0a873eb PZ |
5357 | period = max_t(u64, 10000, event->hw.sample_period); |
5358 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5359 | |
b0a873eb | 5360 | return ret; |
f29ac756 PZ |
5361 | } |
5362 | ||
b0a873eb | 5363 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5364 | { |
b0a873eb | 5365 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5366 | s64 period; |
5367 | ||
5368 | if (!is_sampling_event(event)) | |
5369 | return; | |
f5ffe02e | 5370 | |
5d508e82 FBH |
5371 | period = local64_read(&hwc->period_left); |
5372 | if (period) { | |
5373 | if (period < 0) | |
5374 | period = 10000; | |
fa407f35 | 5375 | |
5d508e82 FBH |
5376 | local64_set(&hwc->period_left, 0); |
5377 | } else { | |
5378 | period = max_t(u64, 10000, hwc->sample_period); | |
5379 | } | |
5380 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 5381 | ns_to_ktime(period), 0, |
b5ab4cd5 | 5382 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 5383 | } |
b0a873eb PZ |
5384 | |
5385 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 5386 | { |
b0a873eb PZ |
5387 | struct hw_perf_event *hwc = &event->hw; |
5388 | ||
6c7e550f | 5389 | if (is_sampling_event(event)) { |
b0a873eb | 5390 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 5391 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
5392 | |
5393 | hrtimer_cancel(&hwc->hrtimer); | |
5394 | } | |
24f1e32c FW |
5395 | } |
5396 | ||
ba3dd36c PZ |
5397 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5398 | { | |
5399 | struct hw_perf_event *hwc = &event->hw; | |
5400 | ||
5401 | if (!is_sampling_event(event)) | |
5402 | return; | |
5403 | ||
5404 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5405 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
5406 | ||
5407 | /* | |
5408 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
5409 | * mapping and avoid the whole period adjust feedback stuff. | |
5410 | */ | |
5411 | if (event->attr.freq) { | |
5412 | long freq = event->attr.sample_freq; | |
5413 | ||
5414 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
5415 | hwc->sample_period = event->attr.sample_period; | |
5416 | local64_set(&hwc->period_left, hwc->sample_period); | |
5417 | event->attr.freq = 0; | |
5418 | } | |
5419 | } | |
5420 | ||
b0a873eb PZ |
5421 | /* |
5422 | * Software event: cpu wall time clock | |
5423 | */ | |
5424 | ||
5425 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 5426 | { |
b0a873eb PZ |
5427 | s64 prev; |
5428 | u64 now; | |
5429 | ||
a4eaf7f1 | 5430 | now = local_clock(); |
b0a873eb PZ |
5431 | prev = local64_xchg(&event->hw.prev_count, now); |
5432 | local64_add(now - prev, &event->count); | |
24f1e32c | 5433 | } |
24f1e32c | 5434 | |
a4eaf7f1 | 5435 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5436 | { |
a4eaf7f1 | 5437 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 5438 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5439 | } |
5440 | ||
a4eaf7f1 | 5441 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 5442 | { |
b0a873eb PZ |
5443 | perf_swevent_cancel_hrtimer(event); |
5444 | cpu_clock_event_update(event); | |
5445 | } | |
f29ac756 | 5446 | |
a4eaf7f1 PZ |
5447 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
5448 | { | |
5449 | if (flags & PERF_EF_START) | |
5450 | cpu_clock_event_start(event, flags); | |
5451 | ||
5452 | return 0; | |
5453 | } | |
5454 | ||
5455 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
5456 | { | |
5457 | cpu_clock_event_stop(event, flags); | |
5458 | } | |
5459 | ||
b0a873eb PZ |
5460 | static void cpu_clock_event_read(struct perf_event *event) |
5461 | { | |
5462 | cpu_clock_event_update(event); | |
5463 | } | |
f344011c | 5464 | |
b0a873eb PZ |
5465 | static int cpu_clock_event_init(struct perf_event *event) |
5466 | { | |
5467 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5468 | return -ENOENT; | |
5469 | ||
5470 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
5471 | return -ENOENT; | |
5472 | ||
2481c5fa SE |
5473 | /* |
5474 | * no branch sampling for software events | |
5475 | */ | |
5476 | if (has_branch_stack(event)) | |
5477 | return -EOPNOTSUPP; | |
5478 | ||
ba3dd36c PZ |
5479 | perf_swevent_init_hrtimer(event); |
5480 | ||
b0a873eb | 5481 | return 0; |
f29ac756 PZ |
5482 | } |
5483 | ||
b0a873eb | 5484 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
5485 | .task_ctx_nr = perf_sw_context, |
5486 | ||
b0a873eb | 5487 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
5488 | .add = cpu_clock_event_add, |
5489 | .del = cpu_clock_event_del, | |
5490 | .start = cpu_clock_event_start, | |
5491 | .stop = cpu_clock_event_stop, | |
b0a873eb | 5492 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
5493 | |
5494 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5495 | }; |
5496 | ||
5497 | /* | |
5498 | * Software event: task time clock | |
5499 | */ | |
5500 | ||
5501 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 5502 | { |
b0a873eb PZ |
5503 | u64 prev; |
5504 | s64 delta; | |
5c92d124 | 5505 | |
b0a873eb PZ |
5506 | prev = local64_xchg(&event->hw.prev_count, now); |
5507 | delta = now - prev; | |
5508 | local64_add(delta, &event->count); | |
5509 | } | |
5c92d124 | 5510 | |
a4eaf7f1 | 5511 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5512 | { |
a4eaf7f1 | 5513 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 5514 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5515 | } |
5516 | ||
a4eaf7f1 | 5517 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
5518 | { |
5519 | perf_swevent_cancel_hrtimer(event); | |
5520 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
5521 | } |
5522 | ||
5523 | static int task_clock_event_add(struct perf_event *event, int flags) | |
5524 | { | |
5525 | if (flags & PERF_EF_START) | |
5526 | task_clock_event_start(event, flags); | |
b0a873eb | 5527 | |
a4eaf7f1 PZ |
5528 | return 0; |
5529 | } | |
5530 | ||
5531 | static void task_clock_event_del(struct perf_event *event, int flags) | |
5532 | { | |
5533 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
5534 | } |
5535 | ||
5536 | static void task_clock_event_read(struct perf_event *event) | |
5537 | { | |
768a06e2 PZ |
5538 | u64 now = perf_clock(); |
5539 | u64 delta = now - event->ctx->timestamp; | |
5540 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
5541 | |
5542 | task_clock_event_update(event, time); | |
5543 | } | |
5544 | ||
5545 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 5546 | { |
b0a873eb PZ |
5547 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5548 | return -ENOENT; | |
5549 | ||
5550 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
5551 | return -ENOENT; | |
5552 | ||
2481c5fa SE |
5553 | /* |
5554 | * no branch sampling for software events | |
5555 | */ | |
5556 | if (has_branch_stack(event)) | |
5557 | return -EOPNOTSUPP; | |
5558 | ||
ba3dd36c PZ |
5559 | perf_swevent_init_hrtimer(event); |
5560 | ||
b0a873eb | 5561 | return 0; |
6fb2915d LZ |
5562 | } |
5563 | ||
b0a873eb | 5564 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
5565 | .task_ctx_nr = perf_sw_context, |
5566 | ||
b0a873eb | 5567 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
5568 | .add = task_clock_event_add, |
5569 | .del = task_clock_event_del, | |
5570 | .start = task_clock_event_start, | |
5571 | .stop = task_clock_event_stop, | |
b0a873eb | 5572 | .read = task_clock_event_read, |
35edc2a5 PZ |
5573 | |
5574 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 5575 | }; |
6fb2915d | 5576 | |
ad5133b7 | 5577 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 5578 | { |
e077df4f | 5579 | } |
6fb2915d | 5580 | |
ad5133b7 | 5581 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 5582 | { |
ad5133b7 | 5583 | return 0; |
6fb2915d LZ |
5584 | } |
5585 | ||
ad5133b7 | 5586 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 5587 | { |
ad5133b7 | 5588 | perf_pmu_disable(pmu); |
6fb2915d LZ |
5589 | } |
5590 | ||
ad5133b7 PZ |
5591 | static int perf_pmu_commit_txn(struct pmu *pmu) |
5592 | { | |
5593 | perf_pmu_enable(pmu); | |
5594 | return 0; | |
5595 | } | |
e077df4f | 5596 | |
ad5133b7 | 5597 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 5598 | { |
ad5133b7 | 5599 | perf_pmu_enable(pmu); |
24f1e32c FW |
5600 | } |
5601 | ||
35edc2a5 PZ |
5602 | static int perf_event_idx_default(struct perf_event *event) |
5603 | { | |
5604 | return event->hw.idx + 1; | |
5605 | } | |
5606 | ||
8dc85d54 PZ |
5607 | /* |
5608 | * Ensures all contexts with the same task_ctx_nr have the same | |
5609 | * pmu_cpu_context too. | |
5610 | */ | |
5611 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 5612 | { |
8dc85d54 | 5613 | struct pmu *pmu; |
b326e956 | 5614 | |
8dc85d54 PZ |
5615 | if (ctxn < 0) |
5616 | return NULL; | |
24f1e32c | 5617 | |
8dc85d54 PZ |
5618 | list_for_each_entry(pmu, &pmus, entry) { |
5619 | if (pmu->task_ctx_nr == ctxn) | |
5620 | return pmu->pmu_cpu_context; | |
5621 | } | |
24f1e32c | 5622 | |
8dc85d54 | 5623 | return NULL; |
24f1e32c FW |
5624 | } |
5625 | ||
51676957 | 5626 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 5627 | { |
51676957 PZ |
5628 | int cpu; |
5629 | ||
5630 | for_each_possible_cpu(cpu) { | |
5631 | struct perf_cpu_context *cpuctx; | |
5632 | ||
5633 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
5634 | ||
5635 | if (cpuctx->active_pmu == old_pmu) | |
5636 | cpuctx->active_pmu = pmu; | |
5637 | } | |
5638 | } | |
5639 | ||
5640 | static void free_pmu_context(struct pmu *pmu) | |
5641 | { | |
5642 | struct pmu *i; | |
f5ffe02e | 5643 | |
8dc85d54 | 5644 | mutex_lock(&pmus_lock); |
0475f9ea | 5645 | /* |
8dc85d54 | 5646 | * Like a real lame refcount. |
0475f9ea | 5647 | */ |
51676957 PZ |
5648 | list_for_each_entry(i, &pmus, entry) { |
5649 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
5650 | update_pmu_context(i, pmu); | |
8dc85d54 | 5651 | goto out; |
51676957 | 5652 | } |
8dc85d54 | 5653 | } |
d6d020e9 | 5654 | |
51676957 | 5655 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
5656 | out: |
5657 | mutex_unlock(&pmus_lock); | |
24f1e32c | 5658 | } |
2e80a82a | 5659 | static struct idr pmu_idr; |
d6d020e9 | 5660 | |
abe43400 PZ |
5661 | static ssize_t |
5662 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
5663 | { | |
5664 | struct pmu *pmu = dev_get_drvdata(dev); | |
5665 | ||
5666 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
5667 | } | |
5668 | ||
5669 | static struct device_attribute pmu_dev_attrs[] = { | |
5670 | __ATTR_RO(type), | |
5671 | __ATTR_NULL, | |
5672 | }; | |
5673 | ||
5674 | static int pmu_bus_running; | |
5675 | static struct bus_type pmu_bus = { | |
5676 | .name = "event_source", | |
5677 | .dev_attrs = pmu_dev_attrs, | |
5678 | }; | |
5679 | ||
5680 | static void pmu_dev_release(struct device *dev) | |
5681 | { | |
5682 | kfree(dev); | |
5683 | } | |
5684 | ||
5685 | static int pmu_dev_alloc(struct pmu *pmu) | |
5686 | { | |
5687 | int ret = -ENOMEM; | |
5688 | ||
5689 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
5690 | if (!pmu->dev) | |
5691 | goto out; | |
5692 | ||
0c9d42ed | 5693 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
5694 | device_initialize(pmu->dev); |
5695 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
5696 | if (ret) | |
5697 | goto free_dev; | |
5698 | ||
5699 | dev_set_drvdata(pmu->dev, pmu); | |
5700 | pmu->dev->bus = &pmu_bus; | |
5701 | pmu->dev->release = pmu_dev_release; | |
5702 | ret = device_add(pmu->dev); | |
5703 | if (ret) | |
5704 | goto free_dev; | |
5705 | ||
5706 | out: | |
5707 | return ret; | |
5708 | ||
5709 | free_dev: | |
5710 | put_device(pmu->dev); | |
5711 | goto out; | |
5712 | } | |
5713 | ||
547e9fd7 | 5714 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 5715 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 5716 | |
2e80a82a | 5717 | int perf_pmu_register(struct pmu *pmu, char *name, int type) |
24f1e32c | 5718 | { |
108b02cf | 5719 | int cpu, ret; |
24f1e32c | 5720 | |
b0a873eb | 5721 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
5722 | ret = -ENOMEM; |
5723 | pmu->pmu_disable_count = alloc_percpu(int); | |
5724 | if (!pmu->pmu_disable_count) | |
5725 | goto unlock; | |
f29ac756 | 5726 | |
2e80a82a PZ |
5727 | pmu->type = -1; |
5728 | if (!name) | |
5729 | goto skip_type; | |
5730 | pmu->name = name; | |
5731 | ||
5732 | if (type < 0) { | |
5733 | int err = idr_pre_get(&pmu_idr, GFP_KERNEL); | |
5734 | if (!err) | |
5735 | goto free_pdc; | |
5736 | ||
5737 | err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type); | |
5738 | if (err) { | |
5739 | ret = err; | |
5740 | goto free_pdc; | |
5741 | } | |
5742 | } | |
5743 | pmu->type = type; | |
5744 | ||
abe43400 PZ |
5745 | if (pmu_bus_running) { |
5746 | ret = pmu_dev_alloc(pmu); | |
5747 | if (ret) | |
5748 | goto free_idr; | |
5749 | } | |
5750 | ||
2e80a82a | 5751 | skip_type: |
8dc85d54 PZ |
5752 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
5753 | if (pmu->pmu_cpu_context) | |
5754 | goto got_cpu_context; | |
f29ac756 | 5755 | |
108b02cf PZ |
5756 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
5757 | if (!pmu->pmu_cpu_context) | |
abe43400 | 5758 | goto free_dev; |
f344011c | 5759 | |
108b02cf PZ |
5760 | for_each_possible_cpu(cpu) { |
5761 | struct perf_cpu_context *cpuctx; | |
5762 | ||
5763 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 5764 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 5765 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 5766 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 5767 | cpuctx->ctx.type = cpu_context; |
108b02cf | 5768 | cpuctx->ctx.pmu = pmu; |
e9d2b064 PZ |
5769 | cpuctx->jiffies_interval = 1; |
5770 | INIT_LIST_HEAD(&cpuctx->rotation_list); | |
51676957 | 5771 | cpuctx->active_pmu = pmu; |
108b02cf | 5772 | } |
76e1d904 | 5773 | |
8dc85d54 | 5774 | got_cpu_context: |
ad5133b7 PZ |
5775 | if (!pmu->start_txn) { |
5776 | if (pmu->pmu_enable) { | |
5777 | /* | |
5778 | * If we have pmu_enable/pmu_disable calls, install | |
5779 | * transaction stubs that use that to try and batch | |
5780 | * hardware accesses. | |
5781 | */ | |
5782 | pmu->start_txn = perf_pmu_start_txn; | |
5783 | pmu->commit_txn = perf_pmu_commit_txn; | |
5784 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
5785 | } else { | |
5786 | pmu->start_txn = perf_pmu_nop_void; | |
5787 | pmu->commit_txn = perf_pmu_nop_int; | |
5788 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 5789 | } |
5c92d124 | 5790 | } |
15dbf27c | 5791 | |
ad5133b7 PZ |
5792 | if (!pmu->pmu_enable) { |
5793 | pmu->pmu_enable = perf_pmu_nop_void; | |
5794 | pmu->pmu_disable = perf_pmu_nop_void; | |
5795 | } | |
5796 | ||
35edc2a5 PZ |
5797 | if (!pmu->event_idx) |
5798 | pmu->event_idx = perf_event_idx_default; | |
5799 | ||
b0a873eb | 5800 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
5801 | ret = 0; |
5802 | unlock: | |
b0a873eb PZ |
5803 | mutex_unlock(&pmus_lock); |
5804 | ||
33696fc0 | 5805 | return ret; |
108b02cf | 5806 | |
abe43400 PZ |
5807 | free_dev: |
5808 | device_del(pmu->dev); | |
5809 | put_device(pmu->dev); | |
5810 | ||
2e80a82a PZ |
5811 | free_idr: |
5812 | if (pmu->type >= PERF_TYPE_MAX) | |
5813 | idr_remove(&pmu_idr, pmu->type); | |
5814 | ||
108b02cf PZ |
5815 | free_pdc: |
5816 | free_percpu(pmu->pmu_disable_count); | |
5817 | goto unlock; | |
f29ac756 PZ |
5818 | } |
5819 | ||
b0a873eb | 5820 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 5821 | { |
b0a873eb PZ |
5822 | mutex_lock(&pmus_lock); |
5823 | list_del_rcu(&pmu->entry); | |
5824 | mutex_unlock(&pmus_lock); | |
5c92d124 | 5825 | |
0475f9ea | 5826 | /* |
cde8e884 PZ |
5827 | * We dereference the pmu list under both SRCU and regular RCU, so |
5828 | * synchronize against both of those. | |
0475f9ea | 5829 | */ |
b0a873eb | 5830 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 5831 | synchronize_rcu(); |
d6d020e9 | 5832 | |
33696fc0 | 5833 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
5834 | if (pmu->type >= PERF_TYPE_MAX) |
5835 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
5836 | device_del(pmu->dev); |
5837 | put_device(pmu->dev); | |
51676957 | 5838 | free_pmu_context(pmu); |
b0a873eb | 5839 | } |
d6d020e9 | 5840 | |
b0a873eb PZ |
5841 | struct pmu *perf_init_event(struct perf_event *event) |
5842 | { | |
5843 | struct pmu *pmu = NULL; | |
5844 | int idx; | |
940c5b29 | 5845 | int ret; |
b0a873eb PZ |
5846 | |
5847 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
5848 | |
5849 | rcu_read_lock(); | |
5850 | pmu = idr_find(&pmu_idr, event->attr.type); | |
5851 | rcu_read_unlock(); | |
940c5b29 | 5852 | if (pmu) { |
7e5b2a01 | 5853 | event->pmu = pmu; |
940c5b29 LM |
5854 | ret = pmu->event_init(event); |
5855 | if (ret) | |
5856 | pmu = ERR_PTR(ret); | |
2e80a82a | 5857 | goto unlock; |
940c5b29 | 5858 | } |
2e80a82a | 5859 | |
b0a873eb | 5860 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 5861 | event->pmu = pmu; |
940c5b29 | 5862 | ret = pmu->event_init(event); |
b0a873eb | 5863 | if (!ret) |
e5f4d339 | 5864 | goto unlock; |
76e1d904 | 5865 | |
b0a873eb PZ |
5866 | if (ret != -ENOENT) { |
5867 | pmu = ERR_PTR(ret); | |
e5f4d339 | 5868 | goto unlock; |
f344011c | 5869 | } |
5c92d124 | 5870 | } |
e5f4d339 PZ |
5871 | pmu = ERR_PTR(-ENOENT); |
5872 | unlock: | |
b0a873eb | 5873 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 5874 | |
4aeb0b42 | 5875 | return pmu; |
5c92d124 IM |
5876 | } |
5877 | ||
0793a61d | 5878 | /* |
cdd6c482 | 5879 | * Allocate and initialize a event structure |
0793a61d | 5880 | */ |
cdd6c482 | 5881 | static struct perf_event * |
c3f00c70 | 5882 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
5883 | struct task_struct *task, |
5884 | struct perf_event *group_leader, | |
5885 | struct perf_event *parent_event, | |
4dc0da86 AK |
5886 | perf_overflow_handler_t overflow_handler, |
5887 | void *context) | |
0793a61d | 5888 | { |
51b0fe39 | 5889 | struct pmu *pmu; |
cdd6c482 IM |
5890 | struct perf_event *event; |
5891 | struct hw_perf_event *hwc; | |
d5d2bc0d | 5892 | long err; |
0793a61d | 5893 | |
66832eb4 ON |
5894 | if ((unsigned)cpu >= nr_cpu_ids) { |
5895 | if (!task || cpu != -1) | |
5896 | return ERR_PTR(-EINVAL); | |
5897 | } | |
5898 | ||
c3f00c70 | 5899 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 5900 | if (!event) |
d5d2bc0d | 5901 | return ERR_PTR(-ENOMEM); |
0793a61d | 5902 | |
04289bb9 | 5903 | /* |
cdd6c482 | 5904 | * Single events are their own group leaders, with an |
04289bb9 IM |
5905 | * empty sibling list: |
5906 | */ | |
5907 | if (!group_leader) | |
cdd6c482 | 5908 | group_leader = event; |
04289bb9 | 5909 | |
cdd6c482 IM |
5910 | mutex_init(&event->child_mutex); |
5911 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 5912 | |
cdd6c482 IM |
5913 | INIT_LIST_HEAD(&event->group_entry); |
5914 | INIT_LIST_HEAD(&event->event_entry); | |
5915 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
5916 | INIT_LIST_HEAD(&event->rb_entry); |
5917 | ||
cdd6c482 | 5918 | init_waitqueue_head(&event->waitq); |
e360adbe | 5919 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 5920 | |
cdd6c482 | 5921 | mutex_init(&event->mmap_mutex); |
7b732a75 | 5922 | |
cdd6c482 IM |
5923 | event->cpu = cpu; |
5924 | event->attr = *attr; | |
5925 | event->group_leader = group_leader; | |
5926 | event->pmu = NULL; | |
cdd6c482 | 5927 | event->oncpu = -1; |
a96bbc16 | 5928 | |
cdd6c482 | 5929 | event->parent = parent_event; |
b84fbc9f | 5930 | |
cdd6c482 IM |
5931 | event->ns = get_pid_ns(current->nsproxy->pid_ns); |
5932 | event->id = atomic64_inc_return(&perf_event_id); | |
a96bbc16 | 5933 | |
cdd6c482 | 5934 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 5935 | |
d580ff86 PZ |
5936 | if (task) { |
5937 | event->attach_state = PERF_ATTACH_TASK; | |
5938 | #ifdef CONFIG_HAVE_HW_BREAKPOINT | |
5939 | /* | |
5940 | * hw_breakpoint is a bit difficult here.. | |
5941 | */ | |
5942 | if (attr->type == PERF_TYPE_BREAKPOINT) | |
5943 | event->hw.bp_target = task; | |
5944 | #endif | |
5945 | } | |
5946 | ||
4dc0da86 | 5947 | if (!overflow_handler && parent_event) { |
b326e956 | 5948 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
5949 | context = parent_event->overflow_handler_context; |
5950 | } | |
66832eb4 | 5951 | |
b326e956 | 5952 | event->overflow_handler = overflow_handler; |
4dc0da86 | 5953 | event->overflow_handler_context = context; |
97eaf530 | 5954 | |
0d48696f | 5955 | if (attr->disabled) |
cdd6c482 | 5956 | event->state = PERF_EVENT_STATE_OFF; |
a86ed508 | 5957 | |
4aeb0b42 | 5958 | pmu = NULL; |
b8e83514 | 5959 | |
cdd6c482 | 5960 | hwc = &event->hw; |
bd2b5b12 | 5961 | hwc->sample_period = attr->sample_period; |
0d48696f | 5962 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 5963 | hwc->sample_period = 1; |
eced1dfc | 5964 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 5965 | |
e7850595 | 5966 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 5967 | |
2023b359 | 5968 | /* |
cdd6c482 | 5969 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 5970 | */ |
3dab77fb | 5971 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
2023b359 PZ |
5972 | goto done; |
5973 | ||
b0a873eb | 5974 | pmu = perf_init_event(event); |
974802ea | 5975 | |
d5d2bc0d PM |
5976 | done: |
5977 | err = 0; | |
4aeb0b42 | 5978 | if (!pmu) |
d5d2bc0d | 5979 | err = -EINVAL; |
4aeb0b42 RR |
5980 | else if (IS_ERR(pmu)) |
5981 | err = PTR_ERR(pmu); | |
5c92d124 | 5982 | |
d5d2bc0d | 5983 | if (err) { |
cdd6c482 IM |
5984 | if (event->ns) |
5985 | put_pid_ns(event->ns); | |
5986 | kfree(event); | |
d5d2bc0d | 5987 | return ERR_PTR(err); |
621a01ea | 5988 | } |
d5d2bc0d | 5989 | |
cdd6c482 | 5990 | if (!event->parent) { |
82cd6def | 5991 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 5992 | static_key_slow_inc(&perf_sched_events.key); |
3af9e859 | 5993 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
5994 | atomic_inc(&nr_mmap_events); |
5995 | if (event->attr.comm) | |
5996 | atomic_inc(&nr_comm_events); | |
5997 | if (event->attr.task) | |
5998 | atomic_inc(&nr_task_events); | |
927c7a9e FW |
5999 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6000 | err = get_callchain_buffers(); | |
6001 | if (err) { | |
6002 | free_event(event); | |
6003 | return ERR_PTR(err); | |
6004 | } | |
6005 | } | |
d010b332 SE |
6006 | if (has_branch_stack(event)) { |
6007 | static_key_slow_inc(&perf_sched_events.key); | |
6008 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6009 | atomic_inc(&per_cpu(perf_branch_stack_events, | |
6010 | event->cpu)); | |
6011 | } | |
f344011c | 6012 | } |
9ee318a7 | 6013 | |
cdd6c482 | 6014 | return event; |
0793a61d TG |
6015 | } |
6016 | ||
cdd6c482 IM |
6017 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6018 | struct perf_event_attr *attr) | |
974802ea | 6019 | { |
974802ea | 6020 | u32 size; |
cdf8073d | 6021 | int ret; |
974802ea PZ |
6022 | |
6023 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6024 | return -EFAULT; | |
6025 | ||
6026 | /* | |
6027 | * zero the full structure, so that a short copy will be nice. | |
6028 | */ | |
6029 | memset(attr, 0, sizeof(*attr)); | |
6030 | ||
6031 | ret = get_user(size, &uattr->size); | |
6032 | if (ret) | |
6033 | return ret; | |
6034 | ||
6035 | if (size > PAGE_SIZE) /* silly large */ | |
6036 | goto err_size; | |
6037 | ||
6038 | if (!size) /* abi compat */ | |
6039 | size = PERF_ATTR_SIZE_VER0; | |
6040 | ||
6041 | if (size < PERF_ATTR_SIZE_VER0) | |
6042 | goto err_size; | |
6043 | ||
6044 | /* | |
6045 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6046 | * ensure all the unknown bits are 0 - i.e. new |
6047 | * user-space does not rely on any kernel feature | |
6048 | * extensions we dont know about yet. | |
974802ea PZ |
6049 | */ |
6050 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6051 | unsigned char __user *addr; |
6052 | unsigned char __user *end; | |
6053 | unsigned char val; | |
974802ea | 6054 | |
cdf8073d IS |
6055 | addr = (void __user *)uattr + sizeof(*attr); |
6056 | end = (void __user *)uattr + size; | |
974802ea | 6057 | |
cdf8073d | 6058 | for (; addr < end; addr++) { |
974802ea PZ |
6059 | ret = get_user(val, addr); |
6060 | if (ret) | |
6061 | return ret; | |
6062 | if (val) | |
6063 | goto err_size; | |
6064 | } | |
b3e62e35 | 6065 | size = sizeof(*attr); |
974802ea PZ |
6066 | } |
6067 | ||
6068 | ret = copy_from_user(attr, uattr, size); | |
6069 | if (ret) | |
6070 | return -EFAULT; | |
6071 | ||
cd757645 | 6072 | if (attr->__reserved_1) |
974802ea PZ |
6073 | return -EINVAL; |
6074 | ||
6075 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6076 | return -EINVAL; | |
6077 | ||
6078 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6079 | return -EINVAL; | |
6080 | ||
bce38cd5 SE |
6081 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6082 | u64 mask = attr->branch_sample_type; | |
6083 | ||
6084 | /* only using defined bits */ | |
6085 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6086 | return -EINVAL; | |
6087 | ||
6088 | /* at least one branch bit must be set */ | |
6089 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6090 | return -EINVAL; | |
6091 | ||
6092 | /* kernel level capture: check permissions */ | |
6093 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
6094 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6095 | return -EACCES; | |
6096 | ||
6097 | /* propagate priv level, when not set for branch */ | |
6098 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6099 | ||
6100 | /* exclude_kernel checked on syscall entry */ | |
6101 | if (!attr->exclude_kernel) | |
6102 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6103 | ||
6104 | if (!attr->exclude_user) | |
6105 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6106 | ||
6107 | if (!attr->exclude_hv) | |
6108 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6109 | /* | |
6110 | * adjust user setting (for HW filter setup) | |
6111 | */ | |
6112 | attr->branch_sample_type = mask; | |
6113 | } | |
6114 | } | |
974802ea PZ |
6115 | out: |
6116 | return ret; | |
6117 | ||
6118 | err_size: | |
6119 | put_user(sizeof(*attr), &uattr->size); | |
6120 | ret = -E2BIG; | |
6121 | goto out; | |
6122 | } | |
6123 | ||
ac9721f3 PZ |
6124 | static int |
6125 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6126 | { |
76369139 | 6127 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6128 | int ret = -EINVAL; |
6129 | ||
ac9721f3 | 6130 | if (!output_event) |
a4be7c27 PZ |
6131 | goto set; |
6132 | ||
ac9721f3 PZ |
6133 | /* don't allow circular references */ |
6134 | if (event == output_event) | |
a4be7c27 PZ |
6135 | goto out; |
6136 | ||
0f139300 PZ |
6137 | /* |
6138 | * Don't allow cross-cpu buffers | |
6139 | */ | |
6140 | if (output_event->cpu != event->cpu) | |
6141 | goto out; | |
6142 | ||
6143 | /* | |
76369139 | 6144 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6145 | */ |
6146 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6147 | goto out; | |
6148 | ||
a4be7c27 | 6149 | set: |
cdd6c482 | 6150 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6151 | /* Can't redirect output if we've got an active mmap() */ |
6152 | if (atomic_read(&event->mmap_count)) | |
6153 | goto unlock; | |
a4be7c27 | 6154 | |
ac9721f3 | 6155 | if (output_event) { |
76369139 FW |
6156 | /* get the rb we want to redirect to */ |
6157 | rb = ring_buffer_get(output_event); | |
6158 | if (!rb) | |
ac9721f3 | 6159 | goto unlock; |
a4be7c27 PZ |
6160 | } |
6161 | ||
76369139 FW |
6162 | old_rb = event->rb; |
6163 | rcu_assign_pointer(event->rb, rb); | |
10c6db11 PZ |
6164 | if (old_rb) |
6165 | ring_buffer_detach(event, old_rb); | |
a4be7c27 | 6166 | ret = 0; |
ac9721f3 PZ |
6167 | unlock: |
6168 | mutex_unlock(&event->mmap_mutex); | |
6169 | ||
76369139 FW |
6170 | if (old_rb) |
6171 | ring_buffer_put(old_rb); | |
a4be7c27 | 6172 | out: |
a4be7c27 PZ |
6173 | return ret; |
6174 | } | |
6175 | ||
0793a61d | 6176 | /** |
cdd6c482 | 6177 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6178 | * |
cdd6c482 | 6179 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6180 | * @pid: target pid |
9f66a381 | 6181 | * @cpu: target cpu |
cdd6c482 | 6182 | * @group_fd: group leader event fd |
0793a61d | 6183 | */ |
cdd6c482 IM |
6184 | SYSCALL_DEFINE5(perf_event_open, |
6185 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 6186 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 6187 | { |
b04243ef PZ |
6188 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6189 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
6190 | struct perf_event_attr attr; |
6191 | struct perf_event_context *ctx; | |
6192 | struct file *event_file = NULL; | |
04289bb9 | 6193 | struct file *group_file = NULL; |
38a81da2 | 6194 | struct task_struct *task = NULL; |
89a1e187 | 6195 | struct pmu *pmu; |
ea635c64 | 6196 | int event_fd; |
b04243ef | 6197 | int move_group = 0; |
04289bb9 | 6198 | int fput_needed = 0; |
dc86cabe | 6199 | int err; |
0793a61d | 6200 | |
2743a5b0 | 6201 | /* for future expandability... */ |
e5d1367f | 6202 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
6203 | return -EINVAL; |
6204 | ||
dc86cabe IM |
6205 | err = perf_copy_attr(attr_uptr, &attr); |
6206 | if (err) | |
6207 | return err; | |
eab656ae | 6208 | |
0764771d PZ |
6209 | if (!attr.exclude_kernel) { |
6210 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6211 | return -EACCES; | |
6212 | } | |
6213 | ||
df58ab24 | 6214 | if (attr.freq) { |
cdd6c482 | 6215 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
6216 | return -EINVAL; |
6217 | } | |
6218 | ||
e5d1367f SE |
6219 | /* |
6220 | * In cgroup mode, the pid argument is used to pass the fd | |
6221 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6222 | * designates the cpu on which to monitor threads from that | |
6223 | * cgroup. | |
6224 | */ | |
6225 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6226 | return -EINVAL; | |
6227 | ||
ea635c64 AV |
6228 | event_fd = get_unused_fd_flags(O_RDWR); |
6229 | if (event_fd < 0) | |
6230 | return event_fd; | |
6231 | ||
ac9721f3 PZ |
6232 | if (group_fd != -1) { |
6233 | group_leader = perf_fget_light(group_fd, &fput_needed); | |
6234 | if (IS_ERR(group_leader)) { | |
6235 | err = PTR_ERR(group_leader); | |
d14b12d7 | 6236 | goto err_fd; |
ac9721f3 PZ |
6237 | } |
6238 | group_file = group_leader->filp; | |
6239 | if (flags & PERF_FLAG_FD_OUTPUT) | |
6240 | output_event = group_leader; | |
6241 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6242 | group_leader = NULL; | |
6243 | } | |
6244 | ||
e5d1367f | 6245 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6246 | task = find_lively_task_by_vpid(pid); |
6247 | if (IS_ERR(task)) { | |
6248 | err = PTR_ERR(task); | |
6249 | goto err_group_fd; | |
6250 | } | |
6251 | } | |
6252 | ||
4dc0da86 AK |
6253 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6254 | NULL, NULL); | |
d14b12d7 SE |
6255 | if (IS_ERR(event)) { |
6256 | err = PTR_ERR(event); | |
c6be5a5c | 6257 | goto err_task; |
d14b12d7 SE |
6258 | } |
6259 | ||
e5d1367f SE |
6260 | if (flags & PERF_FLAG_PID_CGROUP) { |
6261 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
6262 | if (err) | |
6263 | goto err_alloc; | |
08309379 PZ |
6264 | /* |
6265 | * one more event: | |
6266 | * - that has cgroup constraint on event->cpu | |
6267 | * - that may need work on context switch | |
6268 | */ | |
6269 | atomic_inc(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 6270 | static_key_slow_inc(&perf_sched_events.key); |
e5d1367f SE |
6271 | } |
6272 | ||
89a1e187 PZ |
6273 | /* |
6274 | * Special case software events and allow them to be part of | |
6275 | * any hardware group. | |
6276 | */ | |
6277 | pmu = event->pmu; | |
b04243ef PZ |
6278 | |
6279 | if (group_leader && | |
6280 | (is_software_event(event) != is_software_event(group_leader))) { | |
6281 | if (is_software_event(event)) { | |
6282 | /* | |
6283 | * If event and group_leader are not both a software | |
6284 | * event, and event is, then group leader is not. | |
6285 | * | |
6286 | * Allow the addition of software events to !software | |
6287 | * groups, this is safe because software events never | |
6288 | * fail to schedule. | |
6289 | */ | |
6290 | pmu = group_leader->pmu; | |
6291 | } else if (is_software_event(group_leader) && | |
6292 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
6293 | /* | |
6294 | * In case the group is a pure software group, and we | |
6295 | * try to add a hardware event, move the whole group to | |
6296 | * the hardware context. | |
6297 | */ | |
6298 | move_group = 1; | |
6299 | } | |
6300 | } | |
89a1e187 PZ |
6301 | |
6302 | /* | |
6303 | * Get the target context (task or percpu): | |
6304 | */ | |
38a81da2 | 6305 | ctx = find_get_context(pmu, task, cpu); |
89a1e187 PZ |
6306 | if (IS_ERR(ctx)) { |
6307 | err = PTR_ERR(ctx); | |
c6be5a5c | 6308 | goto err_alloc; |
89a1e187 PZ |
6309 | } |
6310 | ||
fd1edb3a PZ |
6311 | if (task) { |
6312 | put_task_struct(task); | |
6313 | task = NULL; | |
6314 | } | |
6315 | ||
ccff286d | 6316 | /* |
cdd6c482 | 6317 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 6318 | */ |
ac9721f3 | 6319 | if (group_leader) { |
dc86cabe | 6320 | err = -EINVAL; |
04289bb9 | 6321 | |
04289bb9 | 6322 | /* |
ccff286d IM |
6323 | * Do not allow a recursive hierarchy (this new sibling |
6324 | * becoming part of another group-sibling): | |
6325 | */ | |
6326 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 6327 | goto err_context; |
ccff286d IM |
6328 | /* |
6329 | * Do not allow to attach to a group in a different | |
6330 | * task or CPU context: | |
04289bb9 | 6331 | */ |
b04243ef PZ |
6332 | if (move_group) { |
6333 | if (group_leader->ctx->type != ctx->type) | |
6334 | goto err_context; | |
6335 | } else { | |
6336 | if (group_leader->ctx != ctx) | |
6337 | goto err_context; | |
6338 | } | |
6339 | ||
3b6f9e5c PM |
6340 | /* |
6341 | * Only a group leader can be exclusive or pinned | |
6342 | */ | |
0d48696f | 6343 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 6344 | goto err_context; |
ac9721f3 PZ |
6345 | } |
6346 | ||
6347 | if (output_event) { | |
6348 | err = perf_event_set_output(event, output_event); | |
6349 | if (err) | |
c3f00c70 | 6350 | goto err_context; |
ac9721f3 | 6351 | } |
0793a61d | 6352 | |
ea635c64 AV |
6353 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
6354 | if (IS_ERR(event_file)) { | |
6355 | err = PTR_ERR(event_file); | |
c3f00c70 | 6356 | goto err_context; |
ea635c64 | 6357 | } |
9b51f66d | 6358 | |
b04243ef PZ |
6359 | if (move_group) { |
6360 | struct perf_event_context *gctx = group_leader->ctx; | |
6361 | ||
6362 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 6363 | perf_remove_from_context(group_leader); |
b04243ef PZ |
6364 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6365 | group_entry) { | |
fe4b04fa | 6366 | perf_remove_from_context(sibling); |
b04243ef PZ |
6367 | put_ctx(gctx); |
6368 | } | |
6369 | mutex_unlock(&gctx->mutex); | |
6370 | put_ctx(gctx); | |
ea635c64 | 6371 | } |
9b51f66d | 6372 | |
cdd6c482 | 6373 | event->filp = event_file; |
ad3a37de | 6374 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 6375 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
6376 | |
6377 | if (move_group) { | |
6378 | perf_install_in_context(ctx, group_leader, cpu); | |
6379 | get_ctx(ctx); | |
6380 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
6381 | group_entry) { | |
6382 | perf_install_in_context(ctx, sibling, cpu); | |
6383 | get_ctx(ctx); | |
6384 | } | |
6385 | } | |
6386 | ||
cdd6c482 | 6387 | perf_install_in_context(ctx, event, cpu); |
ad3a37de | 6388 | ++ctx->generation; |
fe4b04fa | 6389 | perf_unpin_context(ctx); |
d859e29f | 6390 | mutex_unlock(&ctx->mutex); |
9b51f66d | 6391 | |
cdd6c482 | 6392 | event->owner = current; |
8882135b | 6393 | |
cdd6c482 IM |
6394 | mutex_lock(¤t->perf_event_mutex); |
6395 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
6396 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 6397 | |
c320c7b7 ACM |
6398 | /* |
6399 | * Precalculate sample_data sizes | |
6400 | */ | |
6401 | perf_event__header_size(event); | |
6844c09d | 6402 | perf_event__id_header_size(event); |
c320c7b7 | 6403 | |
8a49542c PZ |
6404 | /* |
6405 | * Drop the reference on the group_event after placing the | |
6406 | * new event on the sibling_list. This ensures destruction | |
6407 | * of the group leader will find the pointer to itself in | |
6408 | * perf_group_detach(). | |
6409 | */ | |
ea635c64 AV |
6410 | fput_light(group_file, fput_needed); |
6411 | fd_install(event_fd, event_file); | |
6412 | return event_fd; | |
0793a61d | 6413 | |
c3f00c70 | 6414 | err_context: |
fe4b04fa | 6415 | perf_unpin_context(ctx); |
ea635c64 | 6416 | put_ctx(ctx); |
c6be5a5c | 6417 | err_alloc: |
ea635c64 | 6418 | free_event(event); |
e7d0bc04 PZ |
6419 | err_task: |
6420 | if (task) | |
6421 | put_task_struct(task); | |
89a1e187 | 6422 | err_group_fd: |
dc86cabe | 6423 | fput_light(group_file, fput_needed); |
ea635c64 AV |
6424 | err_fd: |
6425 | put_unused_fd(event_fd); | |
dc86cabe | 6426 | return err; |
0793a61d TG |
6427 | } |
6428 | ||
fb0459d7 AV |
6429 | /** |
6430 | * perf_event_create_kernel_counter | |
6431 | * | |
6432 | * @attr: attributes of the counter to create | |
6433 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 6434 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
6435 | */ |
6436 | struct perf_event * | |
6437 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 6438 | struct task_struct *task, |
4dc0da86 AK |
6439 | perf_overflow_handler_t overflow_handler, |
6440 | void *context) | |
fb0459d7 | 6441 | { |
fb0459d7 | 6442 | struct perf_event_context *ctx; |
c3f00c70 | 6443 | struct perf_event *event; |
fb0459d7 | 6444 | int err; |
d859e29f | 6445 | |
fb0459d7 AV |
6446 | /* |
6447 | * Get the target context (task or percpu): | |
6448 | */ | |
d859e29f | 6449 | |
4dc0da86 AK |
6450 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
6451 | overflow_handler, context); | |
c3f00c70 PZ |
6452 | if (IS_ERR(event)) { |
6453 | err = PTR_ERR(event); | |
6454 | goto err; | |
6455 | } | |
d859e29f | 6456 | |
38a81da2 | 6457 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
6458 | if (IS_ERR(ctx)) { |
6459 | err = PTR_ERR(ctx); | |
c3f00c70 | 6460 | goto err_free; |
d859e29f | 6461 | } |
fb0459d7 AV |
6462 | |
6463 | event->filp = NULL; | |
6464 | WARN_ON_ONCE(ctx->parent_ctx); | |
6465 | mutex_lock(&ctx->mutex); | |
6466 | perf_install_in_context(ctx, event, cpu); | |
6467 | ++ctx->generation; | |
fe4b04fa | 6468 | perf_unpin_context(ctx); |
fb0459d7 AV |
6469 | mutex_unlock(&ctx->mutex); |
6470 | ||
fb0459d7 AV |
6471 | return event; |
6472 | ||
c3f00c70 PZ |
6473 | err_free: |
6474 | free_event(event); | |
6475 | err: | |
c6567f64 | 6476 | return ERR_PTR(err); |
9b51f66d | 6477 | } |
fb0459d7 | 6478 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 6479 | |
cdd6c482 | 6480 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 6481 | struct task_struct *child) |
d859e29f | 6482 | { |
cdd6c482 | 6483 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 6484 | u64 child_val; |
d859e29f | 6485 | |
cdd6c482 IM |
6486 | if (child_event->attr.inherit_stat) |
6487 | perf_event_read_event(child_event, child); | |
38b200d6 | 6488 | |
b5e58793 | 6489 | child_val = perf_event_count(child_event); |
d859e29f PM |
6490 | |
6491 | /* | |
6492 | * Add back the child's count to the parent's count: | |
6493 | */ | |
a6e6dea6 | 6494 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
6495 | atomic64_add(child_event->total_time_enabled, |
6496 | &parent_event->child_total_time_enabled); | |
6497 | atomic64_add(child_event->total_time_running, | |
6498 | &parent_event->child_total_time_running); | |
d859e29f PM |
6499 | |
6500 | /* | |
cdd6c482 | 6501 | * Remove this event from the parent's list |
d859e29f | 6502 | */ |
cdd6c482 IM |
6503 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
6504 | mutex_lock(&parent_event->child_mutex); | |
6505 | list_del_init(&child_event->child_list); | |
6506 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
6507 | |
6508 | /* | |
cdd6c482 | 6509 | * Release the parent event, if this was the last |
d859e29f PM |
6510 | * reference to it. |
6511 | */ | |
cdd6c482 | 6512 | fput(parent_event->filp); |
d859e29f PM |
6513 | } |
6514 | ||
9b51f66d | 6515 | static void |
cdd6c482 IM |
6516 | __perf_event_exit_task(struct perf_event *child_event, |
6517 | struct perf_event_context *child_ctx, | |
38b200d6 | 6518 | struct task_struct *child) |
9b51f66d | 6519 | { |
38b435b1 PZ |
6520 | if (child_event->parent) { |
6521 | raw_spin_lock_irq(&child_ctx->lock); | |
6522 | perf_group_detach(child_event); | |
6523 | raw_spin_unlock_irq(&child_ctx->lock); | |
6524 | } | |
9b51f66d | 6525 | |
fe4b04fa | 6526 | perf_remove_from_context(child_event); |
0cc0c027 | 6527 | |
9b51f66d | 6528 | /* |
38b435b1 | 6529 | * It can happen that the parent exits first, and has events |
9b51f66d | 6530 | * that are still around due to the child reference. These |
38b435b1 | 6531 | * events need to be zapped. |
9b51f66d | 6532 | */ |
38b435b1 | 6533 | if (child_event->parent) { |
cdd6c482 IM |
6534 | sync_child_event(child_event, child); |
6535 | free_event(child_event); | |
4bcf349a | 6536 | } |
9b51f66d IM |
6537 | } |
6538 | ||
8dc85d54 | 6539 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 6540 | { |
cdd6c482 IM |
6541 | struct perf_event *child_event, *tmp; |
6542 | struct perf_event_context *child_ctx; | |
a63eaf34 | 6543 | unsigned long flags; |
9b51f66d | 6544 | |
8dc85d54 | 6545 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 6546 | perf_event_task(child, NULL, 0); |
9b51f66d | 6547 | return; |
9f498cc5 | 6548 | } |
9b51f66d | 6549 | |
a63eaf34 | 6550 | local_irq_save(flags); |
ad3a37de PM |
6551 | /* |
6552 | * We can't reschedule here because interrupts are disabled, | |
6553 | * and either child is current or it is a task that can't be | |
6554 | * scheduled, so we are now safe from rescheduling changing | |
6555 | * our context. | |
6556 | */ | |
806839b2 | 6557 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
6558 | |
6559 | /* | |
6560 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 6561 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
6562 | * incremented the context's refcount before we do put_ctx below. |
6563 | */ | |
e625cce1 | 6564 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 6565 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 6566 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
6567 | /* |
6568 | * If this context is a clone; unclone it so it can't get | |
6569 | * swapped to another process while we're removing all | |
cdd6c482 | 6570 | * the events from it. |
71a851b4 PZ |
6571 | */ |
6572 | unclone_ctx(child_ctx); | |
5e942bb3 | 6573 | update_context_time(child_ctx); |
e625cce1 | 6574 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
6575 | |
6576 | /* | |
cdd6c482 IM |
6577 | * Report the task dead after unscheduling the events so that we |
6578 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
6579 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 6580 | */ |
cdd6c482 | 6581 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 6582 | |
66fff224 PZ |
6583 | /* |
6584 | * We can recurse on the same lock type through: | |
6585 | * | |
cdd6c482 IM |
6586 | * __perf_event_exit_task() |
6587 | * sync_child_event() | |
6588 | * fput(parent_event->filp) | |
66fff224 PZ |
6589 | * perf_release() |
6590 | * mutex_lock(&ctx->mutex) | |
6591 | * | |
6592 | * But since its the parent context it won't be the same instance. | |
6593 | */ | |
a0507c84 | 6594 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 6595 | |
8bc20959 | 6596 | again: |
889ff015 FW |
6597 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
6598 | group_entry) | |
6599 | __perf_event_exit_task(child_event, child_ctx, child); | |
6600 | ||
6601 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 6602 | group_entry) |
cdd6c482 | 6603 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
6604 | |
6605 | /* | |
cdd6c482 | 6606 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
6607 | * its siblings to the list, but we obtained 'tmp' before that which |
6608 | * will still point to the list head terminating the iteration. | |
6609 | */ | |
889ff015 FW |
6610 | if (!list_empty(&child_ctx->pinned_groups) || |
6611 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 6612 | goto again; |
a63eaf34 PM |
6613 | |
6614 | mutex_unlock(&child_ctx->mutex); | |
6615 | ||
6616 | put_ctx(child_ctx); | |
9b51f66d IM |
6617 | } |
6618 | ||
8dc85d54 PZ |
6619 | /* |
6620 | * When a child task exits, feed back event values to parent events. | |
6621 | */ | |
6622 | void perf_event_exit_task(struct task_struct *child) | |
6623 | { | |
8882135b | 6624 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
6625 | int ctxn; |
6626 | ||
8882135b PZ |
6627 | mutex_lock(&child->perf_event_mutex); |
6628 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
6629 | owner_entry) { | |
6630 | list_del_init(&event->owner_entry); | |
6631 | ||
6632 | /* | |
6633 | * Ensure the list deletion is visible before we clear | |
6634 | * the owner, closes a race against perf_release() where | |
6635 | * we need to serialize on the owner->perf_event_mutex. | |
6636 | */ | |
6637 | smp_wmb(); | |
6638 | event->owner = NULL; | |
6639 | } | |
6640 | mutex_unlock(&child->perf_event_mutex); | |
6641 | ||
8dc85d54 PZ |
6642 | for_each_task_context_nr(ctxn) |
6643 | perf_event_exit_task_context(child, ctxn); | |
6644 | } | |
6645 | ||
889ff015 FW |
6646 | static void perf_free_event(struct perf_event *event, |
6647 | struct perf_event_context *ctx) | |
6648 | { | |
6649 | struct perf_event *parent = event->parent; | |
6650 | ||
6651 | if (WARN_ON_ONCE(!parent)) | |
6652 | return; | |
6653 | ||
6654 | mutex_lock(&parent->child_mutex); | |
6655 | list_del_init(&event->child_list); | |
6656 | mutex_unlock(&parent->child_mutex); | |
6657 | ||
6658 | fput(parent->filp); | |
6659 | ||
8a49542c | 6660 | perf_group_detach(event); |
889ff015 FW |
6661 | list_del_event(event, ctx); |
6662 | free_event(event); | |
6663 | } | |
6664 | ||
bbbee908 PZ |
6665 | /* |
6666 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 6667 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 6668 | */ |
cdd6c482 | 6669 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 6670 | { |
8dc85d54 | 6671 | struct perf_event_context *ctx; |
cdd6c482 | 6672 | struct perf_event *event, *tmp; |
8dc85d54 | 6673 | int ctxn; |
bbbee908 | 6674 | |
8dc85d54 PZ |
6675 | for_each_task_context_nr(ctxn) { |
6676 | ctx = task->perf_event_ctxp[ctxn]; | |
6677 | if (!ctx) | |
6678 | continue; | |
bbbee908 | 6679 | |
8dc85d54 | 6680 | mutex_lock(&ctx->mutex); |
bbbee908 | 6681 | again: |
8dc85d54 PZ |
6682 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
6683 | group_entry) | |
6684 | perf_free_event(event, ctx); | |
bbbee908 | 6685 | |
8dc85d54 PZ |
6686 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
6687 | group_entry) | |
6688 | perf_free_event(event, ctx); | |
bbbee908 | 6689 | |
8dc85d54 PZ |
6690 | if (!list_empty(&ctx->pinned_groups) || |
6691 | !list_empty(&ctx->flexible_groups)) | |
6692 | goto again; | |
bbbee908 | 6693 | |
8dc85d54 | 6694 | mutex_unlock(&ctx->mutex); |
bbbee908 | 6695 | |
8dc85d54 PZ |
6696 | put_ctx(ctx); |
6697 | } | |
889ff015 FW |
6698 | } |
6699 | ||
4e231c79 PZ |
6700 | void perf_event_delayed_put(struct task_struct *task) |
6701 | { | |
6702 | int ctxn; | |
6703 | ||
6704 | for_each_task_context_nr(ctxn) | |
6705 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
6706 | } | |
6707 | ||
97dee4f3 PZ |
6708 | /* |
6709 | * inherit a event from parent task to child task: | |
6710 | */ | |
6711 | static struct perf_event * | |
6712 | inherit_event(struct perf_event *parent_event, | |
6713 | struct task_struct *parent, | |
6714 | struct perf_event_context *parent_ctx, | |
6715 | struct task_struct *child, | |
6716 | struct perf_event *group_leader, | |
6717 | struct perf_event_context *child_ctx) | |
6718 | { | |
6719 | struct perf_event *child_event; | |
cee010ec | 6720 | unsigned long flags; |
97dee4f3 PZ |
6721 | |
6722 | /* | |
6723 | * Instead of creating recursive hierarchies of events, | |
6724 | * we link inherited events back to the original parent, | |
6725 | * which has a filp for sure, which we use as the reference | |
6726 | * count: | |
6727 | */ | |
6728 | if (parent_event->parent) | |
6729 | parent_event = parent_event->parent; | |
6730 | ||
6731 | child_event = perf_event_alloc(&parent_event->attr, | |
6732 | parent_event->cpu, | |
d580ff86 | 6733 | child, |
97dee4f3 | 6734 | group_leader, parent_event, |
4dc0da86 | 6735 | NULL, NULL); |
97dee4f3 PZ |
6736 | if (IS_ERR(child_event)) |
6737 | return child_event; | |
6738 | get_ctx(child_ctx); | |
6739 | ||
6740 | /* | |
6741 | * Make the child state follow the state of the parent event, | |
6742 | * not its attr.disabled bit. We hold the parent's mutex, | |
6743 | * so we won't race with perf_event_{en, dis}able_family. | |
6744 | */ | |
6745 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
6746 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
6747 | else | |
6748 | child_event->state = PERF_EVENT_STATE_OFF; | |
6749 | ||
6750 | if (parent_event->attr.freq) { | |
6751 | u64 sample_period = parent_event->hw.sample_period; | |
6752 | struct hw_perf_event *hwc = &child_event->hw; | |
6753 | ||
6754 | hwc->sample_period = sample_period; | |
6755 | hwc->last_period = sample_period; | |
6756 | ||
6757 | local64_set(&hwc->period_left, sample_period); | |
6758 | } | |
6759 | ||
6760 | child_event->ctx = child_ctx; | |
6761 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
6762 | child_event->overflow_handler_context |
6763 | = parent_event->overflow_handler_context; | |
97dee4f3 | 6764 | |
614b6780 TG |
6765 | /* |
6766 | * Precalculate sample_data sizes | |
6767 | */ | |
6768 | perf_event__header_size(child_event); | |
6844c09d | 6769 | perf_event__id_header_size(child_event); |
614b6780 | 6770 | |
97dee4f3 PZ |
6771 | /* |
6772 | * Link it up in the child's context: | |
6773 | */ | |
cee010ec | 6774 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 6775 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 6776 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 PZ |
6777 | |
6778 | /* | |
6779 | * Get a reference to the parent filp - we will fput it | |
6780 | * when the child event exits. This is safe to do because | |
6781 | * we are in the parent and we know that the filp still | |
6782 | * exists and has a nonzero count: | |
6783 | */ | |
6784 | atomic_long_inc(&parent_event->filp->f_count); | |
6785 | ||
6786 | /* | |
6787 | * Link this into the parent event's child list | |
6788 | */ | |
6789 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
6790 | mutex_lock(&parent_event->child_mutex); | |
6791 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
6792 | mutex_unlock(&parent_event->child_mutex); | |
6793 | ||
6794 | return child_event; | |
6795 | } | |
6796 | ||
6797 | static int inherit_group(struct perf_event *parent_event, | |
6798 | struct task_struct *parent, | |
6799 | struct perf_event_context *parent_ctx, | |
6800 | struct task_struct *child, | |
6801 | struct perf_event_context *child_ctx) | |
6802 | { | |
6803 | struct perf_event *leader; | |
6804 | struct perf_event *sub; | |
6805 | struct perf_event *child_ctr; | |
6806 | ||
6807 | leader = inherit_event(parent_event, parent, parent_ctx, | |
6808 | child, NULL, child_ctx); | |
6809 | if (IS_ERR(leader)) | |
6810 | return PTR_ERR(leader); | |
6811 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
6812 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
6813 | child, leader, child_ctx); | |
6814 | if (IS_ERR(child_ctr)) | |
6815 | return PTR_ERR(child_ctr); | |
6816 | } | |
6817 | return 0; | |
889ff015 FW |
6818 | } |
6819 | ||
6820 | static int | |
6821 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
6822 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 6823 | struct task_struct *child, int ctxn, |
889ff015 FW |
6824 | int *inherited_all) |
6825 | { | |
6826 | int ret; | |
8dc85d54 | 6827 | struct perf_event_context *child_ctx; |
889ff015 FW |
6828 | |
6829 | if (!event->attr.inherit) { | |
6830 | *inherited_all = 0; | |
6831 | return 0; | |
bbbee908 PZ |
6832 | } |
6833 | ||
fe4b04fa | 6834 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
6835 | if (!child_ctx) { |
6836 | /* | |
6837 | * This is executed from the parent task context, so | |
6838 | * inherit events that have been marked for cloning. | |
6839 | * First allocate and initialize a context for the | |
6840 | * child. | |
6841 | */ | |
bbbee908 | 6842 | |
eb184479 | 6843 | child_ctx = alloc_perf_context(event->pmu, child); |
889ff015 FW |
6844 | if (!child_ctx) |
6845 | return -ENOMEM; | |
bbbee908 | 6846 | |
8dc85d54 | 6847 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
6848 | } |
6849 | ||
6850 | ret = inherit_group(event, parent, parent_ctx, | |
6851 | child, child_ctx); | |
6852 | ||
6853 | if (ret) | |
6854 | *inherited_all = 0; | |
6855 | ||
6856 | return ret; | |
bbbee908 PZ |
6857 | } |
6858 | ||
9b51f66d | 6859 | /* |
cdd6c482 | 6860 | * Initialize the perf_event context in task_struct |
9b51f66d | 6861 | */ |
8dc85d54 | 6862 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 6863 | { |
889ff015 | 6864 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
6865 | struct perf_event_context *cloned_ctx; |
6866 | struct perf_event *event; | |
9b51f66d | 6867 | struct task_struct *parent = current; |
564c2b21 | 6868 | int inherited_all = 1; |
dddd3379 | 6869 | unsigned long flags; |
6ab423e0 | 6870 | int ret = 0; |
9b51f66d | 6871 | |
8dc85d54 | 6872 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
6873 | return 0; |
6874 | ||
ad3a37de | 6875 | /* |
25346b93 PM |
6876 | * If the parent's context is a clone, pin it so it won't get |
6877 | * swapped under us. | |
ad3a37de | 6878 | */ |
8dc85d54 | 6879 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 6880 | |
ad3a37de PM |
6881 | /* |
6882 | * No need to check if parent_ctx != NULL here; since we saw | |
6883 | * it non-NULL earlier, the only reason for it to become NULL | |
6884 | * is if we exit, and since we're currently in the middle of | |
6885 | * a fork we can't be exiting at the same time. | |
6886 | */ | |
ad3a37de | 6887 | |
9b51f66d IM |
6888 | /* |
6889 | * Lock the parent list. No need to lock the child - not PID | |
6890 | * hashed yet and not running, so nobody can access it. | |
6891 | */ | |
d859e29f | 6892 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
6893 | |
6894 | /* | |
6895 | * We dont have to disable NMIs - we are only looking at | |
6896 | * the list, not manipulating it: | |
6897 | */ | |
889ff015 | 6898 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
6899 | ret = inherit_task_group(event, parent, parent_ctx, |
6900 | child, ctxn, &inherited_all); | |
889ff015 FW |
6901 | if (ret) |
6902 | break; | |
6903 | } | |
b93f7978 | 6904 | |
dddd3379 TG |
6905 | /* |
6906 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
6907 | * to allocations, but we need to prevent rotation because | |
6908 | * rotate_ctx() will change the list from interrupt context. | |
6909 | */ | |
6910 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
6911 | parent_ctx->rotate_disable = 1; | |
6912 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
6913 | ||
889ff015 | 6914 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
6915 | ret = inherit_task_group(event, parent, parent_ctx, |
6916 | child, ctxn, &inherited_all); | |
889ff015 | 6917 | if (ret) |
9b51f66d | 6918 | break; |
564c2b21 PM |
6919 | } |
6920 | ||
dddd3379 TG |
6921 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
6922 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 6923 | |
8dc85d54 | 6924 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 6925 | |
05cbaa28 | 6926 | if (child_ctx && inherited_all) { |
564c2b21 PM |
6927 | /* |
6928 | * Mark the child context as a clone of the parent | |
6929 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
6930 | * |
6931 | * Note that if the parent is a clone, the holding of | |
6932 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 6933 | */ |
c5ed5145 | 6934 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
6935 | if (cloned_ctx) { |
6936 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 6937 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
6938 | } else { |
6939 | child_ctx->parent_ctx = parent_ctx; | |
6940 | child_ctx->parent_gen = parent_ctx->generation; | |
6941 | } | |
6942 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
6943 | } |
6944 | ||
c5ed5145 | 6945 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 6946 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 6947 | |
25346b93 | 6948 | perf_unpin_context(parent_ctx); |
fe4b04fa | 6949 | put_ctx(parent_ctx); |
ad3a37de | 6950 | |
6ab423e0 | 6951 | return ret; |
9b51f66d IM |
6952 | } |
6953 | ||
8dc85d54 PZ |
6954 | /* |
6955 | * Initialize the perf_event context in task_struct | |
6956 | */ | |
6957 | int perf_event_init_task(struct task_struct *child) | |
6958 | { | |
6959 | int ctxn, ret; | |
6960 | ||
8550d7cb ON |
6961 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
6962 | mutex_init(&child->perf_event_mutex); | |
6963 | INIT_LIST_HEAD(&child->perf_event_list); | |
6964 | ||
8dc85d54 PZ |
6965 | for_each_task_context_nr(ctxn) { |
6966 | ret = perf_event_init_context(child, ctxn); | |
6967 | if (ret) | |
6968 | return ret; | |
6969 | } | |
6970 | ||
6971 | return 0; | |
6972 | } | |
6973 | ||
220b140b PM |
6974 | static void __init perf_event_init_all_cpus(void) |
6975 | { | |
b28ab83c | 6976 | struct swevent_htable *swhash; |
220b140b | 6977 | int cpu; |
220b140b PM |
6978 | |
6979 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
6980 | swhash = &per_cpu(swevent_htable, cpu); |
6981 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 6982 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
6983 | } |
6984 | } | |
6985 | ||
cdd6c482 | 6986 | static void __cpuinit perf_event_init_cpu(int cpu) |
0793a61d | 6987 | { |
108b02cf | 6988 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 6989 | |
b28ab83c | 6990 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 6991 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
6992 | struct swevent_hlist *hlist; |
6993 | ||
b28ab83c PZ |
6994 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
6995 | WARN_ON(!hlist); | |
6996 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 6997 | } |
b28ab83c | 6998 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
6999 | } |
7000 | ||
c277443c | 7001 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7002 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7003 | { |
e9d2b064 PZ |
7004 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7005 | ||
7006 | WARN_ON(!irqs_disabled()); | |
7007 | ||
7008 | list_del_init(&cpuctx->rotation_list); | |
7009 | } | |
7010 | ||
108b02cf | 7011 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7012 | { |
108b02cf | 7013 | struct perf_event_context *ctx = __info; |
cdd6c482 | 7014 | struct perf_event *event, *tmp; |
0793a61d | 7015 | |
108b02cf | 7016 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7017 | |
889ff015 | 7018 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 7019 | __perf_remove_from_context(event); |
889ff015 | 7020 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 7021 | __perf_remove_from_context(event); |
0793a61d | 7022 | } |
108b02cf PZ |
7023 | |
7024 | static void perf_event_exit_cpu_context(int cpu) | |
7025 | { | |
7026 | struct perf_event_context *ctx; | |
7027 | struct pmu *pmu; | |
7028 | int idx; | |
7029 | ||
7030 | idx = srcu_read_lock(&pmus_srcu); | |
7031 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7032 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7033 | |
7034 | mutex_lock(&ctx->mutex); | |
7035 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7036 | mutex_unlock(&ctx->mutex); | |
7037 | } | |
7038 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7039 | } |
7040 | ||
cdd6c482 | 7041 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7042 | { |
b28ab83c | 7043 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7044 | |
b28ab83c PZ |
7045 | mutex_lock(&swhash->hlist_mutex); |
7046 | swevent_hlist_release(swhash); | |
7047 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 7048 | |
108b02cf | 7049 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7050 | } |
7051 | #else | |
cdd6c482 | 7052 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7053 | #endif |
7054 | ||
c277443c PZ |
7055 | static int |
7056 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7057 | { | |
7058 | int cpu; | |
7059 | ||
7060 | for_each_online_cpu(cpu) | |
7061 | perf_event_exit_cpu(cpu); | |
7062 | ||
7063 | return NOTIFY_OK; | |
7064 | } | |
7065 | ||
7066 | /* | |
7067 | * Run the perf reboot notifier at the very last possible moment so that | |
7068 | * the generic watchdog code runs as long as possible. | |
7069 | */ | |
7070 | static struct notifier_block perf_reboot_notifier = { | |
7071 | .notifier_call = perf_reboot, | |
7072 | .priority = INT_MIN, | |
7073 | }; | |
7074 | ||
0793a61d TG |
7075 | static int __cpuinit |
7076 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
7077 | { | |
7078 | unsigned int cpu = (long)hcpu; | |
7079 | ||
4536e4d1 | 7080 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7081 | |
7082 | case CPU_UP_PREPARE: | |
5e11637e | 7083 | case CPU_DOWN_FAILED: |
cdd6c482 | 7084 | perf_event_init_cpu(cpu); |
0793a61d TG |
7085 | break; |
7086 | ||
5e11637e | 7087 | case CPU_UP_CANCELED: |
0793a61d | 7088 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7089 | perf_event_exit_cpu(cpu); |
0793a61d TG |
7090 | break; |
7091 | ||
7092 | default: | |
7093 | break; | |
7094 | } | |
7095 | ||
7096 | return NOTIFY_OK; | |
7097 | } | |
7098 | ||
cdd6c482 | 7099 | void __init perf_event_init(void) |
0793a61d | 7100 | { |
3c502e7a JW |
7101 | int ret; |
7102 | ||
2e80a82a PZ |
7103 | idr_init(&pmu_idr); |
7104 | ||
220b140b | 7105 | perf_event_init_all_cpus(); |
b0a873eb | 7106 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7107 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7108 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7109 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7110 | perf_tp_register(); |
7111 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7112 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7113 | |
7114 | ret = init_hw_breakpoint(); | |
7115 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7116 | |
7117 | /* do not patch jump label more than once per second */ | |
7118 | jump_label_rate_limit(&perf_sched_events, HZ); | |
0793a61d | 7119 | } |
abe43400 PZ |
7120 | |
7121 | static int __init perf_event_sysfs_init(void) | |
7122 | { | |
7123 | struct pmu *pmu; | |
7124 | int ret; | |
7125 | ||
7126 | mutex_lock(&pmus_lock); | |
7127 | ||
7128 | ret = bus_register(&pmu_bus); | |
7129 | if (ret) | |
7130 | goto unlock; | |
7131 | ||
7132 | list_for_each_entry(pmu, &pmus, entry) { | |
7133 | if (!pmu->name || pmu->type < 0) | |
7134 | continue; | |
7135 | ||
7136 | ret = pmu_dev_alloc(pmu); | |
7137 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
7138 | } | |
7139 | pmu_bus_running = 1; | |
7140 | ret = 0; | |
7141 | ||
7142 | unlock: | |
7143 | mutex_unlock(&pmus_lock); | |
7144 | ||
7145 | return ret; | |
7146 | } | |
7147 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
7148 | |
7149 | #ifdef CONFIG_CGROUP_PERF | |
7150 | static struct cgroup_subsys_state *perf_cgroup_create( | |
7151 | struct cgroup_subsys *ss, struct cgroup *cont) | |
7152 | { | |
7153 | struct perf_cgroup *jc; | |
e5d1367f | 7154 | |
1b15d055 | 7155 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
7156 | if (!jc) |
7157 | return ERR_PTR(-ENOMEM); | |
7158 | ||
e5d1367f SE |
7159 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7160 | if (!jc->info) { | |
7161 | kfree(jc); | |
7162 | return ERR_PTR(-ENOMEM); | |
7163 | } | |
7164 | ||
e5d1367f SE |
7165 | return &jc->css; |
7166 | } | |
7167 | ||
7168 | static void perf_cgroup_destroy(struct cgroup_subsys *ss, | |
7169 | struct cgroup *cont) | |
7170 | { | |
7171 | struct perf_cgroup *jc; | |
7172 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), | |
7173 | struct perf_cgroup, css); | |
7174 | free_percpu(jc->info); | |
7175 | kfree(jc); | |
7176 | } | |
7177 | ||
7178 | static int __perf_cgroup_move(void *info) | |
7179 | { | |
7180 | struct task_struct *task = info; | |
7181 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
7182 | return 0; | |
7183 | } | |
7184 | ||
bb9d97b6 TH |
7185 | static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
7186 | struct cgroup_taskset *tset) | |
e5d1367f | 7187 | { |
bb9d97b6 TH |
7188 | struct task_struct *task; |
7189 | ||
7190 | cgroup_taskset_for_each(task, cgrp, tset) | |
7191 | task_function_call(task, __perf_cgroup_move, task); | |
e5d1367f SE |
7192 | } |
7193 | ||
e5d1367f SE |
7194 | static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp, |
7195 | struct cgroup *old_cgrp, struct task_struct *task) | |
7196 | { | |
7197 | /* | |
7198 | * cgroup_exit() is called in the copy_process() failure path. | |
7199 | * Ignore this case since the task hasn't ran yet, this avoids | |
7200 | * trying to poke a half freed task state from generic code. | |
7201 | */ | |
7202 | if (!(task->flags & PF_EXITING)) | |
7203 | return; | |
7204 | ||
bb9d97b6 | 7205 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7206 | } |
7207 | ||
7208 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
7209 | .name = "perf_event", |
7210 | .subsys_id = perf_subsys_id, | |
7211 | .create = perf_cgroup_create, | |
7212 | .destroy = perf_cgroup_destroy, | |
7213 | .exit = perf_cgroup_exit, | |
bb9d97b6 | 7214 | .attach = perf_cgroup_attach, |
e5d1367f SE |
7215 | }; |
7216 | #endif /* CONFIG_CGROUP_PERF */ |