]>
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 | ||
0b3fcf17 SE |
121 | enum event_type_t { |
122 | EVENT_FLEXIBLE = 0x1, | |
123 | EVENT_PINNED = 0x2, | |
124 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
125 | }; | |
126 | ||
e5d1367f SE |
127 | /* |
128 | * perf_sched_events : >0 events exist | |
129 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
130 | */ | |
b2029520 | 131 | struct jump_label_key_deferred perf_sched_events __read_mostly; |
e5d1367f SE |
132 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
133 | ||
cdd6c482 IM |
134 | static atomic_t nr_mmap_events __read_mostly; |
135 | static atomic_t nr_comm_events __read_mostly; | |
136 | static atomic_t nr_task_events __read_mostly; | |
9ee318a7 | 137 | |
108b02cf PZ |
138 | static LIST_HEAD(pmus); |
139 | static DEFINE_MUTEX(pmus_lock); | |
140 | static struct srcu_struct pmus_srcu; | |
141 | ||
0764771d | 142 | /* |
cdd6c482 | 143 | * perf event paranoia level: |
0fbdea19 IM |
144 | * -1 - not paranoid at all |
145 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 146 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 147 | * 2 - disallow kernel profiling for unpriv |
0764771d | 148 | */ |
cdd6c482 | 149 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 150 | |
20443384 FW |
151 | /* Minimum for 512 kiB + 1 user control page */ |
152 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
153 | |
154 | /* | |
cdd6c482 | 155 | * max perf event sample rate |
df58ab24 | 156 | */ |
163ec435 PZ |
157 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
158 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
159 | static int max_samples_per_tick __read_mostly = | |
160 | DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
161 | ||
162 | int perf_proc_update_handler(struct ctl_table *table, int write, | |
163 | void __user *buffer, size_t *lenp, | |
164 | loff_t *ppos) | |
165 | { | |
166 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
167 | ||
168 | if (ret || !write) | |
169 | return ret; | |
170 | ||
171 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
172 | ||
173 | return 0; | |
174 | } | |
1ccd1549 | 175 | |
cdd6c482 | 176 | static atomic64_t perf_event_id; |
a96bbc16 | 177 | |
0b3fcf17 SE |
178 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
179 | enum event_type_t event_type); | |
180 | ||
181 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
182 | enum event_type_t event_type, |
183 | struct task_struct *task); | |
184 | ||
185 | static void update_context_time(struct perf_event_context *ctx); | |
186 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 187 | |
10c6db11 PZ |
188 | static void ring_buffer_attach(struct perf_event *event, |
189 | struct ring_buffer *rb); | |
190 | ||
cdd6c482 | 191 | void __weak perf_event_print_debug(void) { } |
0793a61d | 192 | |
84c79910 | 193 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 194 | { |
84c79910 | 195 | return "pmu"; |
0793a61d TG |
196 | } |
197 | ||
0b3fcf17 SE |
198 | static inline u64 perf_clock(void) |
199 | { | |
200 | return local_clock(); | |
201 | } | |
202 | ||
e5d1367f SE |
203 | static inline struct perf_cpu_context * |
204 | __get_cpu_context(struct perf_event_context *ctx) | |
205 | { | |
206 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
207 | } | |
208 | ||
facc4307 PZ |
209 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
210 | struct perf_event_context *ctx) | |
211 | { | |
212 | raw_spin_lock(&cpuctx->ctx.lock); | |
213 | if (ctx) | |
214 | raw_spin_lock(&ctx->lock); | |
215 | } | |
216 | ||
217 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
218 | struct perf_event_context *ctx) | |
219 | { | |
220 | if (ctx) | |
221 | raw_spin_unlock(&ctx->lock); | |
222 | raw_spin_unlock(&cpuctx->ctx.lock); | |
223 | } | |
224 | ||
e5d1367f SE |
225 | #ifdef CONFIG_CGROUP_PERF |
226 | ||
3f7cce3c SE |
227 | /* |
228 | * Must ensure cgroup is pinned (css_get) before calling | |
229 | * this function. In other words, we cannot call this function | |
230 | * if there is no cgroup event for the current CPU context. | |
231 | */ | |
e5d1367f SE |
232 | static inline struct perf_cgroup * |
233 | perf_cgroup_from_task(struct task_struct *task) | |
234 | { | |
235 | return container_of(task_subsys_state(task, perf_subsys_id), | |
236 | struct perf_cgroup, css); | |
237 | } | |
238 | ||
239 | static inline bool | |
240 | perf_cgroup_match(struct perf_event *event) | |
241 | { | |
242 | struct perf_event_context *ctx = event->ctx; | |
243 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
244 | ||
245 | return !event->cgrp || event->cgrp == cpuctx->cgrp; | |
246 | } | |
247 | ||
248 | static inline void perf_get_cgroup(struct perf_event *event) | |
249 | { | |
250 | css_get(&event->cgrp->css); | |
251 | } | |
252 | ||
253 | static inline void perf_put_cgroup(struct perf_event *event) | |
254 | { | |
255 | css_put(&event->cgrp->css); | |
256 | } | |
257 | ||
258 | static inline void perf_detach_cgroup(struct perf_event *event) | |
259 | { | |
260 | perf_put_cgroup(event); | |
261 | event->cgrp = NULL; | |
262 | } | |
263 | ||
264 | static inline int is_cgroup_event(struct perf_event *event) | |
265 | { | |
266 | return event->cgrp != NULL; | |
267 | } | |
268 | ||
269 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
270 | { | |
271 | struct perf_cgroup_info *t; | |
272 | ||
273 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
274 | return t->time; | |
275 | } | |
276 | ||
277 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
278 | { | |
279 | struct perf_cgroup_info *info; | |
280 | u64 now; | |
281 | ||
282 | now = perf_clock(); | |
283 | ||
284 | info = this_cpu_ptr(cgrp->info); | |
285 | ||
286 | info->time += now - info->timestamp; | |
287 | info->timestamp = now; | |
288 | } | |
289 | ||
290 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
291 | { | |
292 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
293 | if (cgrp_out) | |
294 | __update_cgrp_time(cgrp_out); | |
295 | } | |
296 | ||
297 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
298 | { | |
3f7cce3c SE |
299 | struct perf_cgroup *cgrp; |
300 | ||
e5d1367f | 301 | /* |
3f7cce3c SE |
302 | * ensure we access cgroup data only when needed and |
303 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 304 | */ |
3f7cce3c | 305 | if (!is_cgroup_event(event)) |
e5d1367f SE |
306 | return; |
307 | ||
3f7cce3c SE |
308 | cgrp = perf_cgroup_from_task(current); |
309 | /* | |
310 | * Do not update time when cgroup is not active | |
311 | */ | |
312 | if (cgrp == event->cgrp) | |
313 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
314 | } |
315 | ||
316 | static inline void | |
3f7cce3c SE |
317 | perf_cgroup_set_timestamp(struct task_struct *task, |
318 | struct perf_event_context *ctx) | |
e5d1367f SE |
319 | { |
320 | struct perf_cgroup *cgrp; | |
321 | struct perf_cgroup_info *info; | |
322 | ||
3f7cce3c SE |
323 | /* |
324 | * ctx->lock held by caller | |
325 | * ensure we do not access cgroup data | |
326 | * unless we have the cgroup pinned (css_get) | |
327 | */ | |
328 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
329 | return; |
330 | ||
331 | cgrp = perf_cgroup_from_task(task); | |
332 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 333 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
334 | } |
335 | ||
336 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
337 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
338 | ||
339 | /* | |
340 | * reschedule events based on the cgroup constraint of task. | |
341 | * | |
342 | * mode SWOUT : schedule out everything | |
343 | * mode SWIN : schedule in based on cgroup for next | |
344 | */ | |
345 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
346 | { | |
347 | struct perf_cpu_context *cpuctx; | |
348 | struct pmu *pmu; | |
349 | unsigned long flags; | |
350 | ||
351 | /* | |
352 | * disable interrupts to avoid geting nr_cgroup | |
353 | * changes via __perf_event_disable(). Also | |
354 | * avoids preemption. | |
355 | */ | |
356 | local_irq_save(flags); | |
357 | ||
358 | /* | |
359 | * we reschedule only in the presence of cgroup | |
360 | * constrained events. | |
361 | */ | |
362 | rcu_read_lock(); | |
363 | ||
364 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f SE |
365 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
366 | ||
e5d1367f SE |
367 | /* |
368 | * perf_cgroup_events says at least one | |
369 | * context on this CPU has cgroup events. | |
370 | * | |
371 | * ctx->nr_cgroups reports the number of cgroup | |
372 | * events for a context. | |
373 | */ | |
374 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
375 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
376 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
377 | |
378 | if (mode & PERF_CGROUP_SWOUT) { | |
379 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
380 | /* | |
381 | * must not be done before ctxswout due | |
382 | * to event_filter_match() in event_sched_out() | |
383 | */ | |
384 | cpuctx->cgrp = NULL; | |
385 | } | |
386 | ||
387 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 388 | WARN_ON_ONCE(cpuctx->cgrp); |
e5d1367f SE |
389 | /* set cgrp before ctxsw in to |
390 | * allow event_filter_match() to not | |
391 | * have to pass task around | |
392 | */ | |
393 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
394 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
395 | } | |
facc4307 PZ |
396 | perf_pmu_enable(cpuctx->ctx.pmu); |
397 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 398 | } |
e5d1367f SE |
399 | } |
400 | ||
401 | rcu_read_unlock(); | |
402 | ||
403 | local_irq_restore(flags); | |
404 | } | |
405 | ||
a8d757ef SE |
406 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
407 | struct task_struct *next) | |
e5d1367f | 408 | { |
a8d757ef SE |
409 | struct perf_cgroup *cgrp1; |
410 | struct perf_cgroup *cgrp2 = NULL; | |
411 | ||
412 | /* | |
413 | * we come here when we know perf_cgroup_events > 0 | |
414 | */ | |
415 | cgrp1 = perf_cgroup_from_task(task); | |
416 | ||
417 | /* | |
418 | * next is NULL when called from perf_event_enable_on_exec() | |
419 | * that will systematically cause a cgroup_switch() | |
420 | */ | |
421 | if (next) | |
422 | cgrp2 = perf_cgroup_from_task(next); | |
423 | ||
424 | /* | |
425 | * only schedule out current cgroup events if we know | |
426 | * that we are switching to a different cgroup. Otherwise, | |
427 | * do no touch the cgroup events. | |
428 | */ | |
429 | if (cgrp1 != cgrp2) | |
430 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
431 | } |
432 | ||
a8d757ef SE |
433 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
434 | struct task_struct *task) | |
e5d1367f | 435 | { |
a8d757ef SE |
436 | struct perf_cgroup *cgrp1; |
437 | struct perf_cgroup *cgrp2 = NULL; | |
438 | ||
439 | /* | |
440 | * we come here when we know perf_cgroup_events > 0 | |
441 | */ | |
442 | cgrp1 = perf_cgroup_from_task(task); | |
443 | ||
444 | /* prev can never be NULL */ | |
445 | cgrp2 = perf_cgroup_from_task(prev); | |
446 | ||
447 | /* | |
448 | * only need to schedule in cgroup events if we are changing | |
449 | * cgroup during ctxsw. Cgroup events were not scheduled | |
450 | * out of ctxsw out if that was not the case. | |
451 | */ | |
452 | if (cgrp1 != cgrp2) | |
453 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
454 | } |
455 | ||
456 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
457 | struct perf_event_attr *attr, | |
458 | struct perf_event *group_leader) | |
459 | { | |
460 | struct perf_cgroup *cgrp; | |
461 | struct cgroup_subsys_state *css; | |
462 | struct file *file; | |
463 | int ret = 0, fput_needed; | |
464 | ||
465 | file = fget_light(fd, &fput_needed); | |
466 | if (!file) | |
467 | return -EBADF; | |
468 | ||
469 | css = cgroup_css_from_dir(file, perf_subsys_id); | |
3db272c0 LZ |
470 | if (IS_ERR(css)) { |
471 | ret = PTR_ERR(css); | |
472 | goto out; | |
473 | } | |
e5d1367f SE |
474 | |
475 | cgrp = container_of(css, struct perf_cgroup, css); | |
476 | event->cgrp = cgrp; | |
477 | ||
f75e18cb LZ |
478 | /* must be done before we fput() the file */ |
479 | perf_get_cgroup(event); | |
480 | ||
e5d1367f SE |
481 | /* |
482 | * all events in a group must monitor | |
483 | * the same cgroup because a task belongs | |
484 | * to only one perf cgroup at a time | |
485 | */ | |
486 | if (group_leader && group_leader->cgrp != cgrp) { | |
487 | perf_detach_cgroup(event); | |
488 | ret = -EINVAL; | |
e5d1367f | 489 | } |
3db272c0 | 490 | out: |
e5d1367f SE |
491 | fput_light(file, fput_needed); |
492 | return ret; | |
493 | } | |
494 | ||
495 | static inline void | |
496 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
497 | { | |
498 | struct perf_cgroup_info *t; | |
499 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
500 | event->shadow_ctx_time = now - t->timestamp; | |
501 | } | |
502 | ||
503 | static inline void | |
504 | perf_cgroup_defer_enabled(struct perf_event *event) | |
505 | { | |
506 | /* | |
507 | * when the current task's perf cgroup does not match | |
508 | * the event's, we need to remember to call the | |
509 | * perf_mark_enable() function the first time a task with | |
510 | * a matching perf cgroup is scheduled in. | |
511 | */ | |
512 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
513 | event->cgrp_defer_enabled = 1; | |
514 | } | |
515 | ||
516 | static inline void | |
517 | perf_cgroup_mark_enabled(struct perf_event *event, | |
518 | struct perf_event_context *ctx) | |
519 | { | |
520 | struct perf_event *sub; | |
521 | u64 tstamp = perf_event_time(event); | |
522 | ||
523 | if (!event->cgrp_defer_enabled) | |
524 | return; | |
525 | ||
526 | event->cgrp_defer_enabled = 0; | |
527 | ||
528 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
529 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
530 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
531 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
532 | sub->cgrp_defer_enabled = 0; | |
533 | } | |
534 | } | |
535 | } | |
536 | #else /* !CONFIG_CGROUP_PERF */ | |
537 | ||
538 | static inline bool | |
539 | perf_cgroup_match(struct perf_event *event) | |
540 | { | |
541 | return true; | |
542 | } | |
543 | ||
544 | static inline void perf_detach_cgroup(struct perf_event *event) | |
545 | {} | |
546 | ||
547 | static inline int is_cgroup_event(struct perf_event *event) | |
548 | { | |
549 | return 0; | |
550 | } | |
551 | ||
552 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
553 | { | |
554 | return 0; | |
555 | } | |
556 | ||
557 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
558 | { | |
559 | } | |
560 | ||
561 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
562 | { | |
563 | } | |
564 | ||
a8d757ef SE |
565 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
566 | struct task_struct *next) | |
e5d1367f SE |
567 | { |
568 | } | |
569 | ||
a8d757ef SE |
570 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
571 | struct task_struct *task) | |
e5d1367f SE |
572 | { |
573 | } | |
574 | ||
575 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
576 | struct perf_event_attr *attr, | |
577 | struct perf_event *group_leader) | |
578 | { | |
579 | return -EINVAL; | |
580 | } | |
581 | ||
582 | static inline void | |
3f7cce3c SE |
583 | perf_cgroup_set_timestamp(struct task_struct *task, |
584 | struct perf_event_context *ctx) | |
e5d1367f SE |
585 | { |
586 | } | |
587 | ||
588 | void | |
589 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
590 | { | |
591 | } | |
592 | ||
593 | static inline void | |
594 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
595 | { | |
596 | } | |
597 | ||
598 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
599 | { | |
600 | return 0; | |
601 | } | |
602 | ||
603 | static inline void | |
604 | perf_cgroup_defer_enabled(struct perf_event *event) | |
605 | { | |
606 | } | |
607 | ||
608 | static inline void | |
609 | perf_cgroup_mark_enabled(struct perf_event *event, | |
610 | struct perf_event_context *ctx) | |
611 | { | |
612 | } | |
613 | #endif | |
614 | ||
33696fc0 | 615 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 616 | { |
33696fc0 PZ |
617 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
618 | if (!(*count)++) | |
619 | pmu->pmu_disable(pmu); | |
9e35ad38 | 620 | } |
9e35ad38 | 621 | |
33696fc0 | 622 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 623 | { |
33696fc0 PZ |
624 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
625 | if (!--(*count)) | |
626 | pmu->pmu_enable(pmu); | |
9e35ad38 | 627 | } |
9e35ad38 | 628 | |
e9d2b064 PZ |
629 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
630 | ||
631 | /* | |
632 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
633 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
634 | * disabled, while rotate_context is called from IRQ context. | |
635 | */ | |
108b02cf | 636 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 637 | { |
108b02cf | 638 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 639 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 640 | |
e9d2b064 | 641 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 642 | |
e9d2b064 PZ |
643 | if (list_empty(&cpuctx->rotation_list)) |
644 | list_add(&cpuctx->rotation_list, head); | |
9e35ad38 | 645 | } |
9e35ad38 | 646 | |
cdd6c482 | 647 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 648 | { |
e5289d4a | 649 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
650 | } |
651 | ||
cdd6c482 | 652 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 653 | { |
564c2b21 PM |
654 | if (atomic_dec_and_test(&ctx->refcount)) { |
655 | if (ctx->parent_ctx) | |
656 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
657 | if (ctx->task) |
658 | put_task_struct(ctx->task); | |
cb796ff3 | 659 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 660 | } |
a63eaf34 PM |
661 | } |
662 | ||
cdd6c482 | 663 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
664 | { |
665 | if (ctx->parent_ctx) { | |
666 | put_ctx(ctx->parent_ctx); | |
667 | ctx->parent_ctx = NULL; | |
668 | } | |
669 | } | |
670 | ||
6844c09d ACM |
671 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
672 | { | |
673 | /* | |
674 | * only top level events have the pid namespace they were created in | |
675 | */ | |
676 | if (event->parent) | |
677 | event = event->parent; | |
678 | ||
679 | return task_tgid_nr_ns(p, event->ns); | |
680 | } | |
681 | ||
682 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
683 | { | |
684 | /* | |
685 | * only top level events have the pid namespace they were created in | |
686 | */ | |
687 | if (event->parent) | |
688 | event = event->parent; | |
689 | ||
690 | return task_pid_nr_ns(p, event->ns); | |
691 | } | |
692 | ||
7f453c24 | 693 | /* |
cdd6c482 | 694 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
695 | * to userspace. |
696 | */ | |
cdd6c482 | 697 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 698 | { |
cdd6c482 | 699 | u64 id = event->id; |
7f453c24 | 700 | |
cdd6c482 IM |
701 | if (event->parent) |
702 | id = event->parent->id; | |
7f453c24 PZ |
703 | |
704 | return id; | |
705 | } | |
706 | ||
25346b93 | 707 | /* |
cdd6c482 | 708 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
709 | * This has to cope with with the fact that until it is locked, |
710 | * the context could get moved to another task. | |
711 | */ | |
cdd6c482 | 712 | static struct perf_event_context * |
8dc85d54 | 713 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 714 | { |
cdd6c482 | 715 | struct perf_event_context *ctx; |
25346b93 PM |
716 | |
717 | rcu_read_lock(); | |
9ed6060d | 718 | retry: |
8dc85d54 | 719 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
720 | if (ctx) { |
721 | /* | |
722 | * If this context is a clone of another, it might | |
723 | * get swapped for another underneath us by | |
cdd6c482 | 724 | * perf_event_task_sched_out, though the |
25346b93 PM |
725 | * rcu_read_lock() protects us from any context |
726 | * getting freed. Lock the context and check if it | |
727 | * got swapped before we could get the lock, and retry | |
728 | * if so. If we locked the right context, then it | |
729 | * can't get swapped on us any more. | |
730 | */ | |
e625cce1 | 731 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 732 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 733 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
25346b93 PM |
734 | goto retry; |
735 | } | |
b49a9e7e PZ |
736 | |
737 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 738 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
739 | ctx = NULL; |
740 | } | |
25346b93 PM |
741 | } |
742 | rcu_read_unlock(); | |
743 | return ctx; | |
744 | } | |
745 | ||
746 | /* | |
747 | * Get the context for a task and increment its pin_count so it | |
748 | * can't get swapped to another task. This also increments its | |
749 | * reference count so that the context can't get freed. | |
750 | */ | |
8dc85d54 PZ |
751 | static struct perf_event_context * |
752 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 753 | { |
cdd6c482 | 754 | struct perf_event_context *ctx; |
25346b93 PM |
755 | unsigned long flags; |
756 | ||
8dc85d54 | 757 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
758 | if (ctx) { |
759 | ++ctx->pin_count; | |
e625cce1 | 760 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
761 | } |
762 | return ctx; | |
763 | } | |
764 | ||
cdd6c482 | 765 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
766 | { |
767 | unsigned long flags; | |
768 | ||
e625cce1 | 769 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 770 | --ctx->pin_count; |
e625cce1 | 771 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
772 | } |
773 | ||
f67218c3 PZ |
774 | /* |
775 | * Update the record of the current time in a context. | |
776 | */ | |
777 | static void update_context_time(struct perf_event_context *ctx) | |
778 | { | |
779 | u64 now = perf_clock(); | |
780 | ||
781 | ctx->time += now - ctx->timestamp; | |
782 | ctx->timestamp = now; | |
783 | } | |
784 | ||
4158755d SE |
785 | static u64 perf_event_time(struct perf_event *event) |
786 | { | |
787 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
788 | |
789 | if (is_cgroup_event(event)) | |
790 | return perf_cgroup_event_time(event); | |
791 | ||
4158755d SE |
792 | return ctx ? ctx->time : 0; |
793 | } | |
794 | ||
f67218c3 PZ |
795 | /* |
796 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 797 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
798 | */ |
799 | static void update_event_times(struct perf_event *event) | |
800 | { | |
801 | struct perf_event_context *ctx = event->ctx; | |
802 | u64 run_end; | |
803 | ||
804 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
805 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
806 | return; | |
e5d1367f SE |
807 | /* |
808 | * in cgroup mode, time_enabled represents | |
809 | * the time the event was enabled AND active | |
810 | * tasks were in the monitored cgroup. This is | |
811 | * independent of the activity of the context as | |
812 | * there may be a mix of cgroup and non-cgroup events. | |
813 | * | |
814 | * That is why we treat cgroup events differently | |
815 | * here. | |
816 | */ | |
817 | if (is_cgroup_event(event)) | |
46cd6a7f | 818 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
819 | else if (ctx->is_active) |
820 | run_end = ctx->time; | |
acd1d7c1 PZ |
821 | else |
822 | run_end = event->tstamp_stopped; | |
823 | ||
824 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
825 | |
826 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
827 | run_end = event->tstamp_stopped; | |
828 | else | |
4158755d | 829 | run_end = perf_event_time(event); |
f67218c3 PZ |
830 | |
831 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 832 | |
f67218c3 PZ |
833 | } |
834 | ||
96c21a46 PZ |
835 | /* |
836 | * Update total_time_enabled and total_time_running for all events in a group. | |
837 | */ | |
838 | static void update_group_times(struct perf_event *leader) | |
839 | { | |
840 | struct perf_event *event; | |
841 | ||
842 | update_event_times(leader); | |
843 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
844 | update_event_times(event); | |
845 | } | |
846 | ||
889ff015 FW |
847 | static struct list_head * |
848 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
849 | { | |
850 | if (event->attr.pinned) | |
851 | return &ctx->pinned_groups; | |
852 | else | |
853 | return &ctx->flexible_groups; | |
854 | } | |
855 | ||
fccc714b | 856 | /* |
cdd6c482 | 857 | * Add a event from the lists for its context. |
fccc714b PZ |
858 | * Must be called with ctx->mutex and ctx->lock held. |
859 | */ | |
04289bb9 | 860 | static void |
cdd6c482 | 861 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 862 | { |
8a49542c PZ |
863 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
864 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
865 | |
866 | /* | |
8a49542c PZ |
867 | * If we're a stand alone event or group leader, we go to the context |
868 | * list, group events are kept attached to the group so that | |
869 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 870 | */ |
8a49542c | 871 | if (event->group_leader == event) { |
889ff015 FW |
872 | struct list_head *list; |
873 | ||
d6f962b5 FW |
874 | if (is_software_event(event)) |
875 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
876 | ||
889ff015 FW |
877 | list = ctx_group_list(event, ctx); |
878 | list_add_tail(&event->group_entry, list); | |
5c148194 | 879 | } |
592903cd | 880 | |
08309379 | 881 | if (is_cgroup_event(event)) |
e5d1367f | 882 | ctx->nr_cgroups++; |
e5d1367f | 883 | |
cdd6c482 | 884 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 885 | if (!ctx->nr_events) |
108b02cf | 886 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
887 | ctx->nr_events++; |
888 | if (event->attr.inherit_stat) | |
bfbd3381 | 889 | ctx->nr_stat++; |
04289bb9 IM |
890 | } |
891 | ||
c320c7b7 ACM |
892 | /* |
893 | * Called at perf_event creation and when events are attached/detached from a | |
894 | * group. | |
895 | */ | |
896 | static void perf_event__read_size(struct perf_event *event) | |
897 | { | |
898 | int entry = sizeof(u64); /* value */ | |
899 | int size = 0; | |
900 | int nr = 1; | |
901 | ||
902 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
903 | size += sizeof(u64); | |
904 | ||
905 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
906 | size += sizeof(u64); | |
907 | ||
908 | if (event->attr.read_format & PERF_FORMAT_ID) | |
909 | entry += sizeof(u64); | |
910 | ||
911 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
912 | nr += event->group_leader->nr_siblings; | |
913 | size += sizeof(u64); | |
914 | } | |
915 | ||
916 | size += entry * nr; | |
917 | event->read_size = size; | |
918 | } | |
919 | ||
920 | static void perf_event__header_size(struct perf_event *event) | |
921 | { | |
922 | struct perf_sample_data *data; | |
923 | u64 sample_type = event->attr.sample_type; | |
924 | u16 size = 0; | |
925 | ||
926 | perf_event__read_size(event); | |
927 | ||
928 | if (sample_type & PERF_SAMPLE_IP) | |
929 | size += sizeof(data->ip); | |
930 | ||
6844c09d ACM |
931 | if (sample_type & PERF_SAMPLE_ADDR) |
932 | size += sizeof(data->addr); | |
933 | ||
934 | if (sample_type & PERF_SAMPLE_PERIOD) | |
935 | size += sizeof(data->period); | |
936 | ||
937 | if (sample_type & PERF_SAMPLE_READ) | |
938 | size += event->read_size; | |
939 | ||
940 | event->header_size = size; | |
941 | } | |
942 | ||
943 | static void perf_event__id_header_size(struct perf_event *event) | |
944 | { | |
945 | struct perf_sample_data *data; | |
946 | u64 sample_type = event->attr.sample_type; | |
947 | u16 size = 0; | |
948 | ||
c320c7b7 ACM |
949 | if (sample_type & PERF_SAMPLE_TID) |
950 | size += sizeof(data->tid_entry); | |
951 | ||
952 | if (sample_type & PERF_SAMPLE_TIME) | |
953 | size += sizeof(data->time); | |
954 | ||
c320c7b7 ACM |
955 | if (sample_type & PERF_SAMPLE_ID) |
956 | size += sizeof(data->id); | |
957 | ||
958 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
959 | size += sizeof(data->stream_id); | |
960 | ||
961 | if (sample_type & PERF_SAMPLE_CPU) | |
962 | size += sizeof(data->cpu_entry); | |
963 | ||
6844c09d | 964 | event->id_header_size = size; |
c320c7b7 ACM |
965 | } |
966 | ||
8a49542c PZ |
967 | static void perf_group_attach(struct perf_event *event) |
968 | { | |
c320c7b7 | 969 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 970 | |
74c3337c PZ |
971 | /* |
972 | * We can have double attach due to group movement in perf_event_open. | |
973 | */ | |
974 | if (event->attach_state & PERF_ATTACH_GROUP) | |
975 | return; | |
976 | ||
8a49542c PZ |
977 | event->attach_state |= PERF_ATTACH_GROUP; |
978 | ||
979 | if (group_leader == event) | |
980 | return; | |
981 | ||
982 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
983 | !is_software_event(event)) | |
984 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
985 | ||
986 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
987 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
988 | |
989 | perf_event__header_size(group_leader); | |
990 | ||
991 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
992 | perf_event__header_size(pos); | |
8a49542c PZ |
993 | } |
994 | ||
a63eaf34 | 995 | /* |
cdd6c482 | 996 | * Remove a event from the lists for its context. |
fccc714b | 997 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 998 | */ |
04289bb9 | 999 | static void |
cdd6c482 | 1000 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1001 | { |
68cacd29 | 1002 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1003 | /* |
1004 | * We can have double detach due to exit/hot-unplug + close. | |
1005 | */ | |
1006 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1007 | return; |
8a49542c PZ |
1008 | |
1009 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1010 | ||
68cacd29 | 1011 | if (is_cgroup_event(event)) { |
e5d1367f | 1012 | ctx->nr_cgroups--; |
68cacd29 SE |
1013 | cpuctx = __get_cpu_context(ctx); |
1014 | /* | |
1015 | * if there are no more cgroup events | |
1016 | * then cler cgrp to avoid stale pointer | |
1017 | * in update_cgrp_time_from_cpuctx() | |
1018 | */ | |
1019 | if (!ctx->nr_cgroups) | |
1020 | cpuctx->cgrp = NULL; | |
1021 | } | |
e5d1367f | 1022 | |
cdd6c482 IM |
1023 | ctx->nr_events--; |
1024 | if (event->attr.inherit_stat) | |
bfbd3381 | 1025 | ctx->nr_stat--; |
8bc20959 | 1026 | |
cdd6c482 | 1027 | list_del_rcu(&event->event_entry); |
04289bb9 | 1028 | |
8a49542c PZ |
1029 | if (event->group_leader == event) |
1030 | list_del_init(&event->group_entry); | |
5c148194 | 1031 | |
96c21a46 | 1032 | update_group_times(event); |
b2e74a26 SE |
1033 | |
1034 | /* | |
1035 | * If event was in error state, then keep it | |
1036 | * that way, otherwise bogus counts will be | |
1037 | * returned on read(). The only way to get out | |
1038 | * of error state is by explicit re-enabling | |
1039 | * of the event | |
1040 | */ | |
1041 | if (event->state > PERF_EVENT_STATE_OFF) | |
1042 | event->state = PERF_EVENT_STATE_OFF; | |
050735b0 PZ |
1043 | } |
1044 | ||
8a49542c | 1045 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1046 | { |
1047 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1048 | struct list_head *list = NULL; |
1049 | ||
1050 | /* | |
1051 | * We can have double detach due to exit/hot-unplug + close. | |
1052 | */ | |
1053 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1054 | return; | |
1055 | ||
1056 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1057 | ||
1058 | /* | |
1059 | * If this is a sibling, remove it from its group. | |
1060 | */ | |
1061 | if (event->group_leader != event) { | |
1062 | list_del_init(&event->group_entry); | |
1063 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1064 | goto out; |
8a49542c PZ |
1065 | } |
1066 | ||
1067 | if (!list_empty(&event->group_entry)) | |
1068 | list = &event->group_entry; | |
2e2af50b | 1069 | |
04289bb9 | 1070 | /* |
cdd6c482 IM |
1071 | * If this was a group event with sibling events then |
1072 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1073 | * to whatever list we are on. |
04289bb9 | 1074 | */ |
cdd6c482 | 1075 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1076 | if (list) |
1077 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1078 | sibling->group_leader = sibling; |
d6f962b5 FW |
1079 | |
1080 | /* Inherit group flags from the previous leader */ | |
1081 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1082 | } |
c320c7b7 ACM |
1083 | |
1084 | out: | |
1085 | perf_event__header_size(event->group_leader); | |
1086 | ||
1087 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1088 | perf_event__header_size(tmp); | |
04289bb9 IM |
1089 | } |
1090 | ||
fa66f07a SE |
1091 | static inline int |
1092 | event_filter_match(struct perf_event *event) | |
1093 | { | |
e5d1367f SE |
1094 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1095 | && perf_cgroup_match(event); | |
fa66f07a SE |
1096 | } |
1097 | ||
9ffcfa6f SE |
1098 | static void |
1099 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1100 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1101 | struct perf_event_context *ctx) |
3b6f9e5c | 1102 | { |
4158755d | 1103 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1104 | u64 delta; |
1105 | /* | |
1106 | * An event which could not be activated because of | |
1107 | * filter mismatch still needs to have its timings | |
1108 | * maintained, otherwise bogus information is return | |
1109 | * via read() for time_enabled, time_running: | |
1110 | */ | |
1111 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1112 | && !event_filter_match(event)) { | |
e5d1367f | 1113 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1114 | event->tstamp_running += delta; |
4158755d | 1115 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1116 | } |
1117 | ||
cdd6c482 | 1118 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1119 | return; |
3b6f9e5c | 1120 | |
cdd6c482 IM |
1121 | event->state = PERF_EVENT_STATE_INACTIVE; |
1122 | if (event->pending_disable) { | |
1123 | event->pending_disable = 0; | |
1124 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1125 | } |
4158755d | 1126 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1127 | event->pmu->del(event, 0); |
cdd6c482 | 1128 | event->oncpu = -1; |
3b6f9e5c | 1129 | |
cdd6c482 | 1130 | if (!is_software_event(event)) |
3b6f9e5c PM |
1131 | cpuctx->active_oncpu--; |
1132 | ctx->nr_active--; | |
0f5a2601 PZ |
1133 | if (event->attr.freq && event->attr.sample_freq) |
1134 | ctx->nr_freq--; | |
cdd6c482 | 1135 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c PM |
1136 | cpuctx->exclusive = 0; |
1137 | } | |
1138 | ||
d859e29f | 1139 | static void |
cdd6c482 | 1140 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1141 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1142 | struct perf_event_context *ctx) |
d859e29f | 1143 | { |
cdd6c482 | 1144 | struct perf_event *event; |
fa66f07a | 1145 | int state = group_event->state; |
d859e29f | 1146 | |
cdd6c482 | 1147 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1148 | |
1149 | /* | |
1150 | * Schedule out siblings (if any): | |
1151 | */ | |
cdd6c482 IM |
1152 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1153 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1154 | |
fa66f07a | 1155 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1156 | cpuctx->exclusive = 0; |
1157 | } | |
1158 | ||
0793a61d | 1159 | /* |
cdd6c482 | 1160 | * Cross CPU call to remove a performance event |
0793a61d | 1161 | * |
cdd6c482 | 1162 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1163 | * remove it from the context list. |
1164 | */ | |
fe4b04fa | 1165 | static int __perf_remove_from_context(void *info) |
0793a61d | 1166 | { |
cdd6c482 IM |
1167 | struct perf_event *event = info; |
1168 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1169 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1170 | |
e625cce1 | 1171 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1172 | event_sched_out(event, cpuctx, ctx); |
cdd6c482 | 1173 | list_del_event(event, ctx); |
64ce3126 PZ |
1174 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1175 | ctx->is_active = 0; | |
1176 | cpuctx->task_ctx = NULL; | |
1177 | } | |
e625cce1 | 1178 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1179 | |
1180 | return 0; | |
0793a61d TG |
1181 | } |
1182 | ||
1183 | ||
1184 | /* | |
cdd6c482 | 1185 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1186 | * |
cdd6c482 | 1187 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1188 | * call when the task is on a CPU. |
c93f7669 | 1189 | * |
cdd6c482 IM |
1190 | * If event->ctx is a cloned context, callers must make sure that |
1191 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1192 | * remains valid. This is OK when called from perf_release since |
1193 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1194 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1195 | * context has been detached from its task. |
0793a61d | 1196 | */ |
fe4b04fa | 1197 | static void perf_remove_from_context(struct perf_event *event) |
0793a61d | 1198 | { |
cdd6c482 | 1199 | struct perf_event_context *ctx = event->ctx; |
0793a61d TG |
1200 | struct task_struct *task = ctx->task; |
1201 | ||
fe4b04fa PZ |
1202 | lockdep_assert_held(&ctx->mutex); |
1203 | ||
0793a61d TG |
1204 | if (!task) { |
1205 | /* | |
cdd6c482 | 1206 | * Per cpu events are removed via an smp call and |
af901ca1 | 1207 | * the removal is always successful. |
0793a61d | 1208 | */ |
fe4b04fa | 1209 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
0793a61d TG |
1210 | return; |
1211 | } | |
1212 | ||
1213 | retry: | |
fe4b04fa PZ |
1214 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1215 | return; | |
0793a61d | 1216 | |
e625cce1 | 1217 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1218 | /* |
fe4b04fa PZ |
1219 | * If we failed to find a running task, but find the context active now |
1220 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1221 | */ |
fe4b04fa | 1222 | if (ctx->is_active) { |
e625cce1 | 1223 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1224 | goto retry; |
1225 | } | |
1226 | ||
1227 | /* | |
fe4b04fa PZ |
1228 | * Since the task isn't running, its safe to remove the event, us |
1229 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1230 | */ |
fe4b04fa | 1231 | list_del_event(event, ctx); |
e625cce1 | 1232 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1233 | } |
1234 | ||
d859e29f | 1235 | /* |
cdd6c482 | 1236 | * Cross CPU call to disable a performance event |
d859e29f | 1237 | */ |
fe4b04fa | 1238 | static int __perf_event_disable(void *info) |
d859e29f | 1239 | { |
cdd6c482 | 1240 | struct perf_event *event = info; |
cdd6c482 | 1241 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1242 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1243 | |
1244 | /* | |
cdd6c482 IM |
1245 | * If this is a per-task event, need to check whether this |
1246 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1247 | * |
1248 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1249 | * flipping contexts around. | |
d859e29f | 1250 | */ |
665c2142 | 1251 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1252 | return -EINVAL; |
d859e29f | 1253 | |
e625cce1 | 1254 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1255 | |
1256 | /* | |
cdd6c482 | 1257 | * If the event is on, turn it off. |
d859e29f PM |
1258 | * If it is in error state, leave it in error state. |
1259 | */ | |
cdd6c482 | 1260 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1261 | update_context_time(ctx); |
e5d1367f | 1262 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1263 | update_group_times(event); |
1264 | if (event == event->group_leader) | |
1265 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1266 | else |
cdd6c482 IM |
1267 | event_sched_out(event, cpuctx, ctx); |
1268 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1269 | } |
1270 | ||
e625cce1 | 1271 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1272 | |
1273 | return 0; | |
d859e29f PM |
1274 | } |
1275 | ||
1276 | /* | |
cdd6c482 | 1277 | * Disable a event. |
c93f7669 | 1278 | * |
cdd6c482 IM |
1279 | * If event->ctx is a cloned context, callers must make sure that |
1280 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1281 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1282 | * perf_event_for_each_child or perf_event_for_each because they |
1283 | * hold the top-level event's child_mutex, so any descendant that | |
1284 | * goes to exit will block in sync_child_event. | |
1285 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1286 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1287 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1288 | */ |
44234adc | 1289 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1290 | { |
cdd6c482 | 1291 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1292 | struct task_struct *task = ctx->task; |
1293 | ||
1294 | if (!task) { | |
1295 | /* | |
cdd6c482 | 1296 | * Disable the event on the cpu that it's on |
d859e29f | 1297 | */ |
fe4b04fa | 1298 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1299 | return; |
1300 | } | |
1301 | ||
9ed6060d | 1302 | retry: |
fe4b04fa PZ |
1303 | if (!task_function_call(task, __perf_event_disable, event)) |
1304 | return; | |
d859e29f | 1305 | |
e625cce1 | 1306 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1307 | /* |
cdd6c482 | 1308 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1309 | */ |
cdd6c482 | 1310 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1311 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1312 | /* |
1313 | * Reload the task pointer, it might have been changed by | |
1314 | * a concurrent perf_event_context_sched_out(). | |
1315 | */ | |
1316 | task = ctx->task; | |
d859e29f PM |
1317 | goto retry; |
1318 | } | |
1319 | ||
1320 | /* | |
1321 | * Since we have the lock this context can't be scheduled | |
1322 | * in, so we can change the state safely. | |
1323 | */ | |
cdd6c482 IM |
1324 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1325 | update_group_times(event); | |
1326 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1327 | } |
e625cce1 | 1328 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1329 | } |
dcfce4a0 | 1330 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1331 | |
e5d1367f SE |
1332 | static void perf_set_shadow_time(struct perf_event *event, |
1333 | struct perf_event_context *ctx, | |
1334 | u64 tstamp) | |
1335 | { | |
1336 | /* | |
1337 | * use the correct time source for the time snapshot | |
1338 | * | |
1339 | * We could get by without this by leveraging the | |
1340 | * fact that to get to this function, the caller | |
1341 | * has most likely already called update_context_time() | |
1342 | * and update_cgrp_time_xx() and thus both timestamp | |
1343 | * are identical (or very close). Given that tstamp is, | |
1344 | * already adjusted for cgroup, we could say that: | |
1345 | * tstamp - ctx->timestamp | |
1346 | * is equivalent to | |
1347 | * tstamp - cgrp->timestamp. | |
1348 | * | |
1349 | * Then, in perf_output_read(), the calculation would | |
1350 | * work with no changes because: | |
1351 | * - event is guaranteed scheduled in | |
1352 | * - no scheduled out in between | |
1353 | * - thus the timestamp would be the same | |
1354 | * | |
1355 | * But this is a bit hairy. | |
1356 | * | |
1357 | * So instead, we have an explicit cgroup call to remain | |
1358 | * within the time time source all along. We believe it | |
1359 | * is cleaner and simpler to understand. | |
1360 | */ | |
1361 | if (is_cgroup_event(event)) | |
1362 | perf_cgroup_set_shadow_time(event, tstamp); | |
1363 | else | |
1364 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1365 | } | |
1366 | ||
4fe757dd PZ |
1367 | #define MAX_INTERRUPTS (~0ULL) |
1368 | ||
1369 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1370 | ||
235c7fc7 | 1371 | static int |
9ffcfa6f | 1372 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1373 | struct perf_cpu_context *cpuctx, |
6e37738a | 1374 | struct perf_event_context *ctx) |
235c7fc7 | 1375 | { |
4158755d SE |
1376 | u64 tstamp = perf_event_time(event); |
1377 | ||
cdd6c482 | 1378 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1379 | return 0; |
1380 | ||
cdd6c482 | 1381 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1382 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1383 | |
1384 | /* | |
1385 | * Unthrottle events, since we scheduled we might have missed several | |
1386 | * ticks already, also for a heavily scheduling task there is little | |
1387 | * guarantee it'll get a tick in a timely manner. | |
1388 | */ | |
1389 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1390 | perf_log_throttle(event, 1); | |
1391 | event->hw.interrupts = 0; | |
1392 | } | |
1393 | ||
235c7fc7 IM |
1394 | /* |
1395 | * The new state must be visible before we turn it on in the hardware: | |
1396 | */ | |
1397 | smp_wmb(); | |
1398 | ||
a4eaf7f1 | 1399 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1400 | event->state = PERF_EVENT_STATE_INACTIVE; |
1401 | event->oncpu = -1; | |
235c7fc7 IM |
1402 | return -EAGAIN; |
1403 | } | |
1404 | ||
4158755d | 1405 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1406 | |
e5d1367f | 1407 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1408 | |
cdd6c482 | 1409 | if (!is_software_event(event)) |
3b6f9e5c | 1410 | cpuctx->active_oncpu++; |
235c7fc7 | 1411 | ctx->nr_active++; |
0f5a2601 PZ |
1412 | if (event->attr.freq && event->attr.sample_freq) |
1413 | ctx->nr_freq++; | |
235c7fc7 | 1414 | |
cdd6c482 | 1415 | if (event->attr.exclusive) |
3b6f9e5c PM |
1416 | cpuctx->exclusive = 1; |
1417 | ||
235c7fc7 IM |
1418 | return 0; |
1419 | } | |
1420 | ||
6751b71e | 1421 | static int |
cdd6c482 | 1422 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1423 | struct perf_cpu_context *cpuctx, |
6e37738a | 1424 | struct perf_event_context *ctx) |
6751b71e | 1425 | { |
6bde9b6c | 1426 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1427 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1428 | u64 now = ctx->time; |
1429 | bool simulate = false; | |
6751b71e | 1430 | |
cdd6c482 | 1431 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1432 | return 0; |
1433 | ||
ad5133b7 | 1434 | pmu->start_txn(pmu); |
6bde9b6c | 1435 | |
9ffcfa6f | 1436 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1437 | pmu->cancel_txn(pmu); |
6751b71e | 1438 | return -EAGAIN; |
90151c35 | 1439 | } |
6751b71e PM |
1440 | |
1441 | /* | |
1442 | * Schedule in siblings as one group (if any): | |
1443 | */ | |
cdd6c482 | 1444 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1445 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1446 | partial_group = event; |
6751b71e PM |
1447 | goto group_error; |
1448 | } | |
1449 | } | |
1450 | ||
9ffcfa6f | 1451 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1452 | return 0; |
9ffcfa6f | 1453 | |
6751b71e PM |
1454 | group_error: |
1455 | /* | |
1456 | * Groups can be scheduled in as one unit only, so undo any | |
1457 | * partial group before returning: | |
d7842da4 SE |
1458 | * The events up to the failed event are scheduled out normally, |
1459 | * tstamp_stopped will be updated. | |
1460 | * | |
1461 | * The failed events and the remaining siblings need to have | |
1462 | * their timings updated as if they had gone thru event_sched_in() | |
1463 | * and event_sched_out(). This is required to get consistent timings | |
1464 | * across the group. This also takes care of the case where the group | |
1465 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1466 | * the time the event was actually stopped, such that time delta | |
1467 | * calculation in update_event_times() is correct. | |
6751b71e | 1468 | */ |
cdd6c482 IM |
1469 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1470 | if (event == partial_group) | |
d7842da4 SE |
1471 | simulate = true; |
1472 | ||
1473 | if (simulate) { | |
1474 | event->tstamp_running += now - event->tstamp_stopped; | |
1475 | event->tstamp_stopped = now; | |
1476 | } else { | |
1477 | event_sched_out(event, cpuctx, ctx); | |
1478 | } | |
6751b71e | 1479 | } |
9ffcfa6f | 1480 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1481 | |
ad5133b7 | 1482 | pmu->cancel_txn(pmu); |
90151c35 | 1483 | |
6751b71e PM |
1484 | return -EAGAIN; |
1485 | } | |
1486 | ||
3b6f9e5c | 1487 | /* |
cdd6c482 | 1488 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1489 | */ |
cdd6c482 | 1490 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1491 | struct perf_cpu_context *cpuctx, |
1492 | int can_add_hw) | |
1493 | { | |
1494 | /* | |
cdd6c482 | 1495 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1496 | */ |
d6f962b5 | 1497 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1498 | return 1; |
1499 | /* | |
1500 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1501 | * events can go on. |
3b6f9e5c PM |
1502 | */ |
1503 | if (cpuctx->exclusive) | |
1504 | return 0; | |
1505 | /* | |
1506 | * If this group is exclusive and there are already | |
cdd6c482 | 1507 | * events on the CPU, it can't go on. |
3b6f9e5c | 1508 | */ |
cdd6c482 | 1509 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1510 | return 0; |
1511 | /* | |
1512 | * Otherwise, try to add it if all previous groups were able | |
1513 | * to go on. | |
1514 | */ | |
1515 | return can_add_hw; | |
1516 | } | |
1517 | ||
cdd6c482 IM |
1518 | static void add_event_to_ctx(struct perf_event *event, |
1519 | struct perf_event_context *ctx) | |
53cfbf59 | 1520 | { |
4158755d SE |
1521 | u64 tstamp = perf_event_time(event); |
1522 | ||
cdd6c482 | 1523 | list_add_event(event, ctx); |
8a49542c | 1524 | perf_group_attach(event); |
4158755d SE |
1525 | event->tstamp_enabled = tstamp; |
1526 | event->tstamp_running = tstamp; | |
1527 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1528 | } |
1529 | ||
2c29ef0f PZ |
1530 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1531 | static void | |
1532 | ctx_sched_in(struct perf_event_context *ctx, | |
1533 | struct perf_cpu_context *cpuctx, | |
1534 | enum event_type_t event_type, | |
1535 | struct task_struct *task); | |
fe4b04fa | 1536 | |
dce5855b PZ |
1537 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1538 | struct perf_event_context *ctx, | |
1539 | struct task_struct *task) | |
1540 | { | |
1541 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1542 | if (ctx) | |
1543 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1544 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1545 | if (ctx) | |
1546 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1547 | } | |
1548 | ||
0793a61d | 1549 | /* |
cdd6c482 | 1550 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1551 | * |
1552 | * Must be called with ctx->mutex held | |
0793a61d | 1553 | */ |
fe4b04fa | 1554 | static int __perf_install_in_context(void *info) |
0793a61d | 1555 | { |
cdd6c482 IM |
1556 | struct perf_event *event = info; |
1557 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1558 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1559 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1560 | struct task_struct *task = current; | |
1561 | ||
b58f6b0d | 1562 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1563 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1564 | |
1565 | /* | |
2c29ef0f | 1566 | * If there was an active task_ctx schedule it out. |
0793a61d | 1567 | */ |
b58f6b0d | 1568 | if (task_ctx) |
2c29ef0f | 1569 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1570 | |
1571 | /* | |
1572 | * If the context we're installing events in is not the | |
1573 | * active task_ctx, flip them. | |
1574 | */ | |
1575 | if (ctx->task && task_ctx != ctx) { | |
1576 | if (task_ctx) | |
1577 | raw_spin_unlock(&task_ctx->lock); | |
1578 | raw_spin_lock(&ctx->lock); | |
1579 | task_ctx = ctx; | |
1580 | } | |
1581 | ||
1582 | if (task_ctx) { | |
1583 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1584 | task = task_ctx->task; |
1585 | } | |
b58f6b0d | 1586 | |
2c29ef0f | 1587 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1588 | |
4af4998b | 1589 | update_context_time(ctx); |
e5d1367f SE |
1590 | /* |
1591 | * update cgrp time only if current cgrp | |
1592 | * matches event->cgrp. Must be done before | |
1593 | * calling add_event_to_ctx() | |
1594 | */ | |
1595 | update_cgrp_time_from_event(event); | |
0793a61d | 1596 | |
cdd6c482 | 1597 | add_event_to_ctx(event, ctx); |
0793a61d | 1598 | |
d859e29f | 1599 | /* |
2c29ef0f | 1600 | * Schedule everything back in |
d859e29f | 1601 | */ |
dce5855b | 1602 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1603 | |
1604 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1605 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1606 | |
1607 | return 0; | |
0793a61d TG |
1608 | } |
1609 | ||
1610 | /* | |
cdd6c482 | 1611 | * Attach a performance event to a context |
0793a61d | 1612 | * |
cdd6c482 IM |
1613 | * First we add the event to the list with the hardware enable bit |
1614 | * in event->hw_config cleared. | |
0793a61d | 1615 | * |
cdd6c482 | 1616 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1617 | * call to enable it in the task context. The task might have been |
1618 | * scheduled away, but we check this in the smp call again. | |
1619 | */ | |
1620 | static void | |
cdd6c482 IM |
1621 | perf_install_in_context(struct perf_event_context *ctx, |
1622 | struct perf_event *event, | |
0793a61d TG |
1623 | int cpu) |
1624 | { | |
1625 | struct task_struct *task = ctx->task; | |
1626 | ||
fe4b04fa PZ |
1627 | lockdep_assert_held(&ctx->mutex); |
1628 | ||
c3f00c70 PZ |
1629 | event->ctx = ctx; |
1630 | ||
0793a61d TG |
1631 | if (!task) { |
1632 | /* | |
cdd6c482 | 1633 | * Per cpu events are installed via an smp call and |
af901ca1 | 1634 | * the install is always successful. |
0793a61d | 1635 | */ |
fe4b04fa | 1636 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1637 | return; |
1638 | } | |
1639 | ||
0793a61d | 1640 | retry: |
fe4b04fa PZ |
1641 | if (!task_function_call(task, __perf_install_in_context, event)) |
1642 | return; | |
0793a61d | 1643 | |
e625cce1 | 1644 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1645 | /* |
fe4b04fa PZ |
1646 | * If we failed to find a running task, but find the context active now |
1647 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1648 | */ |
fe4b04fa | 1649 | if (ctx->is_active) { |
e625cce1 | 1650 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1651 | goto retry; |
1652 | } | |
1653 | ||
1654 | /* | |
fe4b04fa PZ |
1655 | * Since the task isn't running, its safe to add the event, us holding |
1656 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1657 | */ |
fe4b04fa | 1658 | add_event_to_ctx(event, ctx); |
e625cce1 | 1659 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1660 | } |
1661 | ||
fa289bec | 1662 | /* |
cdd6c482 | 1663 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1664 | * Enabling the leader of a group effectively enables all |
1665 | * the group members that aren't explicitly disabled, so we | |
1666 | * have to update their ->tstamp_enabled also. | |
1667 | * Note: this works for group members as well as group leaders | |
1668 | * since the non-leader members' sibling_lists will be empty. | |
1669 | */ | |
1d9b482e | 1670 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1671 | { |
cdd6c482 | 1672 | struct perf_event *sub; |
4158755d | 1673 | u64 tstamp = perf_event_time(event); |
fa289bec | 1674 | |
cdd6c482 | 1675 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1676 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1677 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1678 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1679 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1680 | } |
fa289bec PM |
1681 | } |
1682 | ||
d859e29f | 1683 | /* |
cdd6c482 | 1684 | * Cross CPU call to enable a performance event |
d859e29f | 1685 | */ |
fe4b04fa | 1686 | static int __perf_event_enable(void *info) |
04289bb9 | 1687 | { |
cdd6c482 | 1688 | struct perf_event *event = info; |
cdd6c482 IM |
1689 | struct perf_event_context *ctx = event->ctx; |
1690 | struct perf_event *leader = event->group_leader; | |
108b02cf | 1691 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 1692 | int err; |
04289bb9 | 1693 | |
fe4b04fa PZ |
1694 | if (WARN_ON_ONCE(!ctx->is_active)) |
1695 | return -EINVAL; | |
3cbed429 | 1696 | |
e625cce1 | 1697 | raw_spin_lock(&ctx->lock); |
4af4998b | 1698 | update_context_time(ctx); |
d859e29f | 1699 | |
cdd6c482 | 1700 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 1701 | goto unlock; |
e5d1367f SE |
1702 | |
1703 | /* | |
1704 | * set current task's cgroup time reference point | |
1705 | */ | |
3f7cce3c | 1706 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 1707 | |
1d9b482e | 1708 | __perf_event_mark_enabled(event); |
04289bb9 | 1709 | |
e5d1367f SE |
1710 | if (!event_filter_match(event)) { |
1711 | if (is_cgroup_event(event)) | |
1712 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 1713 | goto unlock; |
e5d1367f | 1714 | } |
f4c4176f | 1715 | |
04289bb9 | 1716 | /* |
cdd6c482 | 1717 | * If the event is in a group and isn't the group leader, |
d859e29f | 1718 | * then don't put it on unless the group is on. |
04289bb9 | 1719 | */ |
cdd6c482 | 1720 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 1721 | goto unlock; |
3b6f9e5c | 1722 | |
cdd6c482 | 1723 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 1724 | err = -EEXIST; |
e758a33d | 1725 | } else { |
cdd6c482 | 1726 | if (event == leader) |
6e37738a | 1727 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 1728 | else |
6e37738a | 1729 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 1730 | } |
d859e29f PM |
1731 | |
1732 | if (err) { | |
1733 | /* | |
cdd6c482 | 1734 | * If this event can't go on and it's part of a |
d859e29f PM |
1735 | * group, then the whole group has to come off. |
1736 | */ | |
cdd6c482 | 1737 | if (leader != event) |
d859e29f | 1738 | group_sched_out(leader, cpuctx, ctx); |
0d48696f | 1739 | if (leader->attr.pinned) { |
53cfbf59 | 1740 | update_group_times(leader); |
cdd6c482 | 1741 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 1742 | } |
d859e29f PM |
1743 | } |
1744 | ||
9ed6060d | 1745 | unlock: |
e625cce1 | 1746 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1747 | |
1748 | return 0; | |
d859e29f PM |
1749 | } |
1750 | ||
1751 | /* | |
cdd6c482 | 1752 | * Enable a event. |
c93f7669 | 1753 | * |
cdd6c482 IM |
1754 | * If event->ctx is a cloned context, callers must make sure that |
1755 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1756 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
1757 | * perf_event_for_each_child or perf_event_for_each as described |
1758 | * for perf_event_disable. | |
d859e29f | 1759 | */ |
44234adc | 1760 | void perf_event_enable(struct perf_event *event) |
d859e29f | 1761 | { |
cdd6c482 | 1762 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1763 | struct task_struct *task = ctx->task; |
1764 | ||
1765 | if (!task) { | |
1766 | /* | |
cdd6c482 | 1767 | * Enable the event on the cpu that it's on |
d859e29f | 1768 | */ |
fe4b04fa | 1769 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
1770 | return; |
1771 | } | |
1772 | ||
e625cce1 | 1773 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 1774 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
1775 | goto out; |
1776 | ||
1777 | /* | |
cdd6c482 IM |
1778 | * If the event is in error state, clear that first. |
1779 | * That way, if we see the event in error state below, we | |
d859e29f PM |
1780 | * know that it has gone back into error state, as distinct |
1781 | * from the task having been scheduled away before the | |
1782 | * cross-call arrived. | |
1783 | */ | |
cdd6c482 IM |
1784 | if (event->state == PERF_EVENT_STATE_ERROR) |
1785 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 1786 | |
9ed6060d | 1787 | retry: |
fe4b04fa | 1788 | if (!ctx->is_active) { |
1d9b482e | 1789 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
1790 | goto out; |
1791 | } | |
1792 | ||
e625cce1 | 1793 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1794 | |
1795 | if (!task_function_call(task, __perf_event_enable, event)) | |
1796 | return; | |
d859e29f | 1797 | |
e625cce1 | 1798 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
1799 | |
1800 | /* | |
cdd6c482 | 1801 | * If the context is active and the event is still off, |
d859e29f PM |
1802 | * we need to retry the cross-call. |
1803 | */ | |
fe4b04fa PZ |
1804 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
1805 | /* | |
1806 | * task could have been flipped by a concurrent | |
1807 | * perf_event_context_sched_out() | |
1808 | */ | |
1809 | task = ctx->task; | |
d859e29f | 1810 | goto retry; |
fe4b04fa | 1811 | } |
fa289bec | 1812 | |
9ed6060d | 1813 | out: |
e625cce1 | 1814 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1815 | } |
dcfce4a0 | 1816 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 1817 | |
26ca5c11 | 1818 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 1819 | { |
2023b359 | 1820 | /* |
cdd6c482 | 1821 | * not supported on inherited events |
2023b359 | 1822 | */ |
2e939d1d | 1823 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
1824 | return -EINVAL; |
1825 | ||
cdd6c482 IM |
1826 | atomic_add(refresh, &event->event_limit); |
1827 | perf_event_enable(event); | |
2023b359 PZ |
1828 | |
1829 | return 0; | |
79f14641 | 1830 | } |
26ca5c11 | 1831 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 1832 | |
5b0311e1 FW |
1833 | static void ctx_sched_out(struct perf_event_context *ctx, |
1834 | struct perf_cpu_context *cpuctx, | |
1835 | enum event_type_t event_type) | |
235c7fc7 | 1836 | { |
cdd6c482 | 1837 | struct perf_event *event; |
db24d33e | 1838 | int is_active = ctx->is_active; |
235c7fc7 | 1839 | |
db24d33e | 1840 | ctx->is_active &= ~event_type; |
cdd6c482 | 1841 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
1842 | return; |
1843 | ||
4af4998b | 1844 | update_context_time(ctx); |
e5d1367f | 1845 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 1846 | if (!ctx->nr_active) |
facc4307 | 1847 | return; |
5b0311e1 | 1848 | |
075e0b00 | 1849 | perf_pmu_disable(ctx->pmu); |
db24d33e | 1850 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
1851 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
1852 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 1853 | } |
889ff015 | 1854 | |
db24d33e | 1855 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 1856 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 1857 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 1858 | } |
1b9a644f | 1859 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
1860 | } |
1861 | ||
564c2b21 PM |
1862 | /* |
1863 | * Test whether two contexts are equivalent, i.e. whether they | |
1864 | * have both been cloned from the same version of the same context | |
cdd6c482 IM |
1865 | * and they both have the same number of enabled events. |
1866 | * If the number of enabled events is the same, then the set | |
1867 | * of enabled events should be the same, because these are both | |
1868 | * inherited contexts, therefore we can't access individual events | |
564c2b21 | 1869 | * in them directly with an fd; we can only enable/disable all |
cdd6c482 | 1870 | * events via prctl, or enable/disable all events in a family |
564c2b21 PM |
1871 | * via ioctl, which will have the same effect on both contexts. |
1872 | */ | |
cdd6c482 IM |
1873 | static int context_equiv(struct perf_event_context *ctx1, |
1874 | struct perf_event_context *ctx2) | |
564c2b21 PM |
1875 | { |
1876 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx | |
ad3a37de | 1877 | && ctx1->parent_gen == ctx2->parent_gen |
25346b93 | 1878 | && !ctx1->pin_count && !ctx2->pin_count; |
564c2b21 PM |
1879 | } |
1880 | ||
cdd6c482 IM |
1881 | static void __perf_event_sync_stat(struct perf_event *event, |
1882 | struct perf_event *next_event) | |
bfbd3381 PZ |
1883 | { |
1884 | u64 value; | |
1885 | ||
cdd6c482 | 1886 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
1887 | return; |
1888 | ||
1889 | /* | |
cdd6c482 | 1890 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
1891 | * because we're in the middle of a context switch and have IRQs |
1892 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 1893 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
1894 | * don't need to use it. |
1895 | */ | |
cdd6c482 IM |
1896 | switch (event->state) { |
1897 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
1898 | event->pmu->read(event); |
1899 | /* fall-through */ | |
bfbd3381 | 1900 | |
cdd6c482 IM |
1901 | case PERF_EVENT_STATE_INACTIVE: |
1902 | update_event_times(event); | |
bfbd3381 PZ |
1903 | break; |
1904 | ||
1905 | default: | |
1906 | break; | |
1907 | } | |
1908 | ||
1909 | /* | |
cdd6c482 | 1910 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
1911 | * values when we flip the contexts. |
1912 | */ | |
e7850595 PZ |
1913 | value = local64_read(&next_event->count); |
1914 | value = local64_xchg(&event->count, value); | |
1915 | local64_set(&next_event->count, value); | |
bfbd3381 | 1916 | |
cdd6c482 IM |
1917 | swap(event->total_time_enabled, next_event->total_time_enabled); |
1918 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 1919 | |
bfbd3381 | 1920 | /* |
19d2e755 | 1921 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 1922 | */ |
cdd6c482 IM |
1923 | perf_event_update_userpage(event); |
1924 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
1925 | } |
1926 | ||
1927 | #define list_next_entry(pos, member) \ | |
1928 | list_entry(pos->member.next, typeof(*pos), member) | |
1929 | ||
cdd6c482 IM |
1930 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
1931 | struct perf_event_context *next_ctx) | |
bfbd3381 | 1932 | { |
cdd6c482 | 1933 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
1934 | |
1935 | if (!ctx->nr_stat) | |
1936 | return; | |
1937 | ||
02ffdbc8 PZ |
1938 | update_context_time(ctx); |
1939 | ||
cdd6c482 IM |
1940 | event = list_first_entry(&ctx->event_list, |
1941 | struct perf_event, event_entry); | |
bfbd3381 | 1942 | |
cdd6c482 IM |
1943 | next_event = list_first_entry(&next_ctx->event_list, |
1944 | struct perf_event, event_entry); | |
bfbd3381 | 1945 | |
cdd6c482 IM |
1946 | while (&event->event_entry != &ctx->event_list && |
1947 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 1948 | |
cdd6c482 | 1949 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 1950 | |
cdd6c482 IM |
1951 | event = list_next_entry(event, event_entry); |
1952 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
1953 | } |
1954 | } | |
1955 | ||
fe4b04fa PZ |
1956 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
1957 | struct task_struct *next) | |
0793a61d | 1958 | { |
8dc85d54 | 1959 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 IM |
1960 | struct perf_event_context *next_ctx; |
1961 | struct perf_event_context *parent; | |
108b02cf | 1962 | struct perf_cpu_context *cpuctx; |
c93f7669 | 1963 | int do_switch = 1; |
0793a61d | 1964 | |
108b02cf PZ |
1965 | if (likely(!ctx)) |
1966 | return; | |
10989fb2 | 1967 | |
108b02cf PZ |
1968 | cpuctx = __get_cpu_context(ctx); |
1969 | if (!cpuctx->task_ctx) | |
0793a61d TG |
1970 | return; |
1971 | ||
c93f7669 PM |
1972 | rcu_read_lock(); |
1973 | parent = rcu_dereference(ctx->parent_ctx); | |
8dc85d54 | 1974 | next_ctx = next->perf_event_ctxp[ctxn]; |
c93f7669 PM |
1975 | if (parent && next_ctx && |
1976 | rcu_dereference(next_ctx->parent_ctx) == parent) { | |
1977 | /* | |
1978 | * Looks like the two contexts are clones, so we might be | |
1979 | * able to optimize the context switch. We lock both | |
1980 | * contexts and check that they are clones under the | |
1981 | * lock (including re-checking that neither has been | |
1982 | * uncloned in the meantime). It doesn't matter which | |
1983 | * order we take the locks because no other cpu could | |
1984 | * be trying to lock both of these tasks. | |
1985 | */ | |
e625cce1 TG |
1986 | raw_spin_lock(&ctx->lock); |
1987 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 1988 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
1989 | /* |
1990 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 1991 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 1992 | */ |
8dc85d54 PZ |
1993 | task->perf_event_ctxp[ctxn] = next_ctx; |
1994 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
1995 | ctx->task = next; |
1996 | next_ctx->task = task; | |
1997 | do_switch = 0; | |
bfbd3381 | 1998 | |
cdd6c482 | 1999 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2000 | } |
e625cce1 TG |
2001 | raw_spin_unlock(&next_ctx->lock); |
2002 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2003 | } |
c93f7669 | 2004 | rcu_read_unlock(); |
564c2b21 | 2005 | |
c93f7669 | 2006 | if (do_switch) { |
facc4307 | 2007 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2008 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2009 | cpuctx->task_ctx = NULL; |
facc4307 | 2010 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2011 | } |
0793a61d TG |
2012 | } |
2013 | ||
8dc85d54 PZ |
2014 | #define for_each_task_context_nr(ctxn) \ |
2015 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2016 | ||
2017 | /* | |
2018 | * Called from scheduler to remove the events of the current task, | |
2019 | * with interrupts disabled. | |
2020 | * | |
2021 | * We stop each event and update the event value in event->count. | |
2022 | * | |
2023 | * This does not protect us against NMI, but disable() | |
2024 | * sets the disabled bit in the control field of event _before_ | |
2025 | * accessing the event control register. If a NMI hits, then it will | |
2026 | * not restart the event. | |
2027 | */ | |
82cd6def PZ |
2028 | void __perf_event_task_sched_out(struct task_struct *task, |
2029 | struct task_struct *next) | |
8dc85d54 PZ |
2030 | { |
2031 | int ctxn; | |
2032 | ||
8dc85d54 PZ |
2033 | for_each_task_context_nr(ctxn) |
2034 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2035 | |
2036 | /* | |
2037 | * if cgroup events exist on this CPU, then we need | |
2038 | * to check if we have to switch out PMU state. | |
2039 | * cgroup event are system-wide mode only | |
2040 | */ | |
2041 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2042 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2043 | } |
2044 | ||
04dc2dbb | 2045 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2046 | { |
108b02cf | 2047 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2048 | |
a63eaf34 PM |
2049 | if (!cpuctx->task_ctx) |
2050 | return; | |
012b84da IM |
2051 | |
2052 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2053 | return; | |
2054 | ||
04dc2dbb | 2055 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2056 | cpuctx->task_ctx = NULL; |
2057 | } | |
2058 | ||
5b0311e1 FW |
2059 | /* |
2060 | * Called with IRQs disabled | |
2061 | */ | |
2062 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2063 | enum event_type_t event_type) | |
2064 | { | |
2065 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2066 | } |
2067 | ||
235c7fc7 | 2068 | static void |
5b0311e1 | 2069 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2070 | struct perf_cpu_context *cpuctx) |
0793a61d | 2071 | { |
cdd6c482 | 2072 | struct perf_event *event; |
0793a61d | 2073 | |
889ff015 FW |
2074 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2075 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2076 | continue; |
5632ab12 | 2077 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2078 | continue; |
2079 | ||
e5d1367f SE |
2080 | /* may need to reset tstamp_enabled */ |
2081 | if (is_cgroup_event(event)) | |
2082 | perf_cgroup_mark_enabled(event, ctx); | |
2083 | ||
8c9ed8e1 | 2084 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2085 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2086 | |
2087 | /* | |
2088 | * If this pinned group hasn't been scheduled, | |
2089 | * put it in error state. | |
2090 | */ | |
cdd6c482 IM |
2091 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2092 | update_group_times(event); | |
2093 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2094 | } |
3b6f9e5c | 2095 | } |
5b0311e1 FW |
2096 | } |
2097 | ||
2098 | static void | |
2099 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2100 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2101 | { |
2102 | struct perf_event *event; | |
2103 | int can_add_hw = 1; | |
3b6f9e5c | 2104 | |
889ff015 FW |
2105 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2106 | /* Ignore events in OFF or ERROR state */ | |
2107 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2108 | continue; |
04289bb9 IM |
2109 | /* |
2110 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2111 | * of events: |
04289bb9 | 2112 | */ |
5632ab12 | 2113 | if (!event_filter_match(event)) |
0793a61d TG |
2114 | continue; |
2115 | ||
e5d1367f SE |
2116 | /* may need to reset tstamp_enabled */ |
2117 | if (is_cgroup_event(event)) | |
2118 | perf_cgroup_mark_enabled(event, ctx); | |
2119 | ||
9ed6060d | 2120 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2121 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2122 | can_add_hw = 0; |
9ed6060d | 2123 | } |
0793a61d | 2124 | } |
5b0311e1 FW |
2125 | } |
2126 | ||
2127 | static void | |
2128 | ctx_sched_in(struct perf_event_context *ctx, | |
2129 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2130 | enum event_type_t event_type, |
2131 | struct task_struct *task) | |
5b0311e1 | 2132 | { |
e5d1367f | 2133 | u64 now; |
db24d33e | 2134 | int is_active = ctx->is_active; |
e5d1367f | 2135 | |
db24d33e | 2136 | ctx->is_active |= event_type; |
5b0311e1 | 2137 | if (likely(!ctx->nr_events)) |
facc4307 | 2138 | return; |
5b0311e1 | 2139 | |
e5d1367f SE |
2140 | now = perf_clock(); |
2141 | ctx->timestamp = now; | |
3f7cce3c | 2142 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2143 | /* |
2144 | * First go through the list and put on any pinned groups | |
2145 | * in order to give them the best chance of going on. | |
2146 | */ | |
db24d33e | 2147 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2148 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2149 | |
2150 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2151 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2152 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2153 | } |
2154 | ||
329c0e01 | 2155 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2156 | enum event_type_t event_type, |
2157 | struct task_struct *task) | |
329c0e01 FW |
2158 | { |
2159 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2160 | ||
e5d1367f | 2161 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2162 | } |
2163 | ||
e5d1367f SE |
2164 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2165 | struct task_struct *task) | |
235c7fc7 | 2166 | { |
108b02cf | 2167 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2168 | |
108b02cf | 2169 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2170 | if (cpuctx->task_ctx == ctx) |
2171 | return; | |
2172 | ||
facc4307 | 2173 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2174 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2175 | /* |
2176 | * We want to keep the following priority order: | |
2177 | * cpu pinned (that don't need to move), task pinned, | |
2178 | * cpu flexible, task flexible. | |
2179 | */ | |
2180 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2181 | ||
1d5f003f GN |
2182 | if (ctx->nr_events) |
2183 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2184 | |
86b47c25 GN |
2185 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2186 | ||
facc4307 PZ |
2187 | perf_pmu_enable(ctx->pmu); |
2188 | perf_ctx_unlock(cpuctx, ctx); | |
2189 | ||
b5ab4cd5 PZ |
2190 | /* |
2191 | * Since these rotations are per-cpu, we need to ensure the | |
2192 | * cpu-context we got scheduled on is actually rotating. | |
2193 | */ | |
108b02cf | 2194 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2195 | } |
2196 | ||
8dc85d54 PZ |
2197 | /* |
2198 | * Called from scheduler to add the events of the current task | |
2199 | * with interrupts disabled. | |
2200 | * | |
2201 | * We restore the event value and then enable it. | |
2202 | * | |
2203 | * This does not protect us against NMI, but enable() | |
2204 | * sets the enabled bit in the control field of event _before_ | |
2205 | * accessing the event control register. If a NMI hits, then it will | |
2206 | * keep the event running. | |
2207 | */ | |
a8d757ef SE |
2208 | void __perf_event_task_sched_in(struct task_struct *prev, |
2209 | struct task_struct *task) | |
8dc85d54 PZ |
2210 | { |
2211 | struct perf_event_context *ctx; | |
2212 | int ctxn; | |
2213 | ||
2214 | for_each_task_context_nr(ctxn) { | |
2215 | ctx = task->perf_event_ctxp[ctxn]; | |
2216 | if (likely(!ctx)) | |
2217 | continue; | |
2218 | ||
e5d1367f | 2219 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2220 | } |
e5d1367f SE |
2221 | /* |
2222 | * if cgroup events exist on this CPU, then we need | |
2223 | * to check if we have to switch in PMU state. | |
2224 | * cgroup event are system-wide mode only | |
2225 | */ | |
2226 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2227 | perf_cgroup_sched_in(prev, task); |
235c7fc7 IM |
2228 | } |
2229 | ||
abd50713 PZ |
2230 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2231 | { | |
2232 | u64 frequency = event->attr.sample_freq; | |
2233 | u64 sec = NSEC_PER_SEC; | |
2234 | u64 divisor, dividend; | |
2235 | ||
2236 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2237 | ||
2238 | count_fls = fls64(count); | |
2239 | nsec_fls = fls64(nsec); | |
2240 | frequency_fls = fls64(frequency); | |
2241 | sec_fls = 30; | |
2242 | ||
2243 | /* | |
2244 | * We got @count in @nsec, with a target of sample_freq HZ | |
2245 | * the target period becomes: | |
2246 | * | |
2247 | * @count * 10^9 | |
2248 | * period = ------------------- | |
2249 | * @nsec * sample_freq | |
2250 | * | |
2251 | */ | |
2252 | ||
2253 | /* | |
2254 | * Reduce accuracy by one bit such that @a and @b converge | |
2255 | * to a similar magnitude. | |
2256 | */ | |
fe4b04fa | 2257 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2258 | do { \ |
2259 | if (a##_fls > b##_fls) { \ | |
2260 | a >>= 1; \ | |
2261 | a##_fls--; \ | |
2262 | } else { \ | |
2263 | b >>= 1; \ | |
2264 | b##_fls--; \ | |
2265 | } \ | |
2266 | } while (0) | |
2267 | ||
2268 | /* | |
2269 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2270 | * the other, so that finally we can do a u64/u64 division. | |
2271 | */ | |
2272 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2273 | REDUCE_FLS(nsec, frequency); | |
2274 | REDUCE_FLS(sec, count); | |
2275 | } | |
2276 | ||
2277 | if (count_fls + sec_fls > 64) { | |
2278 | divisor = nsec * frequency; | |
2279 | ||
2280 | while (count_fls + sec_fls > 64) { | |
2281 | REDUCE_FLS(count, sec); | |
2282 | divisor >>= 1; | |
2283 | } | |
2284 | ||
2285 | dividend = count * sec; | |
2286 | } else { | |
2287 | dividend = count * sec; | |
2288 | ||
2289 | while (nsec_fls + frequency_fls > 64) { | |
2290 | REDUCE_FLS(nsec, frequency); | |
2291 | dividend >>= 1; | |
2292 | } | |
2293 | ||
2294 | divisor = nsec * frequency; | |
2295 | } | |
2296 | ||
f6ab91ad PZ |
2297 | if (!divisor) |
2298 | return dividend; | |
2299 | ||
abd50713 PZ |
2300 | return div64_u64(dividend, divisor); |
2301 | } | |
2302 | ||
e050e3f0 SE |
2303 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2304 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2305 | ||
abd50713 | 2306 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count) |
bd2b5b12 | 2307 | { |
cdd6c482 | 2308 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2309 | s64 period, sample_period; |
bd2b5b12 PZ |
2310 | s64 delta; |
2311 | ||
abd50713 | 2312 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2313 | |
2314 | delta = (s64)(period - hwc->sample_period); | |
2315 | delta = (delta + 7) / 8; /* low pass filter */ | |
2316 | ||
2317 | sample_period = hwc->sample_period + delta; | |
2318 | ||
2319 | if (!sample_period) | |
2320 | sample_period = 1; | |
2321 | ||
bd2b5b12 | 2322 | hwc->sample_period = sample_period; |
abd50713 | 2323 | |
e7850595 | 2324 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
a4eaf7f1 | 2325 | event->pmu->stop(event, PERF_EF_UPDATE); |
e7850595 | 2326 | local64_set(&hwc->period_left, 0); |
a4eaf7f1 | 2327 | event->pmu->start(event, PERF_EF_RELOAD); |
abd50713 | 2328 | } |
bd2b5b12 PZ |
2329 | } |
2330 | ||
e050e3f0 SE |
2331 | /* |
2332 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2333 | * events. At the same time, make sure, having freq events does not change | |
2334 | * the rate of unthrottling as that would introduce bias. | |
2335 | */ | |
2336 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2337 | int needs_unthr) | |
60db5e09 | 2338 | { |
cdd6c482 IM |
2339 | struct perf_event *event; |
2340 | struct hw_perf_event *hwc; | |
e050e3f0 | 2341 | u64 now, period = TICK_NSEC; |
abd50713 | 2342 | s64 delta; |
60db5e09 | 2343 | |
e050e3f0 SE |
2344 | /* |
2345 | * only need to iterate over all events iff: | |
2346 | * - context have events in frequency mode (needs freq adjust) | |
2347 | * - there are events to unthrottle on this cpu | |
2348 | */ | |
2349 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2350 | return; |
2351 | ||
e050e3f0 SE |
2352 | raw_spin_lock(&ctx->lock); |
2353 | ||
03541f8b | 2354 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2355 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2356 | continue; |
2357 | ||
5632ab12 | 2358 | if (!event_filter_match(event)) |
5d27c23d PZ |
2359 | continue; |
2360 | ||
cdd6c482 | 2361 | hwc = &event->hw; |
6a24ed6c | 2362 | |
e050e3f0 SE |
2363 | if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) { |
2364 | hwc->interrupts = 0; | |
cdd6c482 | 2365 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2366 | event->pmu->start(event, 0); |
a78ac325 PZ |
2367 | } |
2368 | ||
cdd6c482 | 2369 | if (!event->attr.freq || !event->attr.sample_freq) |
60db5e09 PZ |
2370 | continue; |
2371 | ||
e050e3f0 SE |
2372 | /* |
2373 | * stop the event and update event->count | |
2374 | */ | |
2375 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2376 | ||
e7850595 | 2377 | now = local64_read(&event->count); |
abd50713 PZ |
2378 | delta = now - hwc->freq_count_stamp; |
2379 | hwc->freq_count_stamp = now; | |
60db5e09 | 2380 | |
e050e3f0 SE |
2381 | /* |
2382 | * restart the event | |
2383 | * reload only if value has changed | |
2384 | */ | |
abd50713 | 2385 | if (delta > 0) |
b5ab4cd5 | 2386 | perf_adjust_period(event, period, delta); |
e050e3f0 SE |
2387 | |
2388 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
60db5e09 | 2389 | } |
e050e3f0 SE |
2390 | |
2391 | raw_spin_unlock(&ctx->lock); | |
60db5e09 PZ |
2392 | } |
2393 | ||
235c7fc7 | 2394 | /* |
cdd6c482 | 2395 | * Round-robin a context's events: |
235c7fc7 | 2396 | */ |
cdd6c482 | 2397 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2398 | { |
dddd3379 TG |
2399 | /* |
2400 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2401 | * disabled by the inheritance code. | |
2402 | */ | |
2403 | if (!ctx->rotate_disable) | |
2404 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2405 | } |
2406 | ||
b5ab4cd5 | 2407 | /* |
e9d2b064 PZ |
2408 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2409 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2410 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2411 | */ |
e9d2b064 | 2412 | static void perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2413 | { |
8dc85d54 | 2414 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2415 | int rotate = 0, remove = 1; |
7fc23a53 | 2416 | |
b5ab4cd5 | 2417 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2418 | remove = 0; |
b5ab4cd5 PZ |
2419 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2420 | rotate = 1; | |
2421 | } | |
235c7fc7 | 2422 | |
8dc85d54 | 2423 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2424 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2425 | remove = 0; |
b5ab4cd5 PZ |
2426 | if (ctx->nr_events != ctx->nr_active) |
2427 | rotate = 1; | |
2428 | } | |
9717e6cd | 2429 | |
e050e3f0 | 2430 | if (!rotate) |
0f5a2601 PZ |
2431 | goto done; |
2432 | ||
facc4307 | 2433 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2434 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2435 | |
e050e3f0 SE |
2436 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2437 | if (ctx) | |
2438 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2439 | |
e050e3f0 SE |
2440 | rotate_ctx(&cpuctx->ctx); |
2441 | if (ctx) | |
2442 | rotate_ctx(ctx); | |
235c7fc7 | 2443 | |
e050e3f0 | 2444 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2445 | |
0f5a2601 PZ |
2446 | perf_pmu_enable(cpuctx->ctx.pmu); |
2447 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2448 | done: |
e9d2b064 PZ |
2449 | if (remove) |
2450 | list_del_init(&cpuctx->rotation_list); | |
e9d2b064 PZ |
2451 | } |
2452 | ||
2453 | void perf_event_task_tick(void) | |
2454 | { | |
2455 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2456 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2457 | struct perf_event_context *ctx; |
2458 | int throttled; | |
b5ab4cd5 | 2459 | |
e9d2b064 PZ |
2460 | WARN_ON(!irqs_disabled()); |
2461 | ||
e050e3f0 SE |
2462 | __this_cpu_inc(perf_throttled_seq); |
2463 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2464 | ||
e9d2b064 | 2465 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2466 | ctx = &cpuctx->ctx; |
2467 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2468 | ||
2469 | ctx = cpuctx->task_ctx; | |
2470 | if (ctx) | |
2471 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2472 | ||
e9d2b064 PZ |
2473 | if (cpuctx->jiffies_interval == 1 || |
2474 | !(jiffies % cpuctx->jiffies_interval)) | |
2475 | perf_rotate_context(cpuctx); | |
2476 | } | |
0793a61d TG |
2477 | } |
2478 | ||
889ff015 FW |
2479 | static int event_enable_on_exec(struct perf_event *event, |
2480 | struct perf_event_context *ctx) | |
2481 | { | |
2482 | if (!event->attr.enable_on_exec) | |
2483 | return 0; | |
2484 | ||
2485 | event->attr.enable_on_exec = 0; | |
2486 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2487 | return 0; | |
2488 | ||
1d9b482e | 2489 | __perf_event_mark_enabled(event); |
889ff015 FW |
2490 | |
2491 | return 1; | |
2492 | } | |
2493 | ||
57e7986e | 2494 | /* |
cdd6c482 | 2495 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2496 | * This expects task == current. |
2497 | */ | |
8dc85d54 | 2498 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2499 | { |
cdd6c482 | 2500 | struct perf_event *event; |
57e7986e PM |
2501 | unsigned long flags; |
2502 | int enabled = 0; | |
889ff015 | 2503 | int ret; |
57e7986e PM |
2504 | |
2505 | local_irq_save(flags); | |
cdd6c482 | 2506 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2507 | goto out; |
2508 | ||
e566b76e SE |
2509 | /* |
2510 | * We must ctxsw out cgroup events to avoid conflict | |
2511 | * when invoking perf_task_event_sched_in() later on | |
2512 | * in this function. Otherwise we end up trying to | |
2513 | * ctxswin cgroup events which are already scheduled | |
2514 | * in. | |
2515 | */ | |
a8d757ef | 2516 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2517 | |
e625cce1 | 2518 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2519 | task_ctx_sched_out(ctx); |
57e7986e | 2520 | |
b79387ef | 2521 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2522 | ret = event_enable_on_exec(event, ctx); |
2523 | if (ret) | |
2524 | enabled = 1; | |
57e7986e PM |
2525 | } |
2526 | ||
2527 | /* | |
cdd6c482 | 2528 | * Unclone this context if we enabled any event. |
57e7986e | 2529 | */ |
71a851b4 PZ |
2530 | if (enabled) |
2531 | unclone_ctx(ctx); | |
57e7986e | 2532 | |
e625cce1 | 2533 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2534 | |
e566b76e SE |
2535 | /* |
2536 | * Also calls ctxswin for cgroup events, if any: | |
2537 | */ | |
e5d1367f | 2538 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2539 | out: |
57e7986e PM |
2540 | local_irq_restore(flags); |
2541 | } | |
2542 | ||
0793a61d | 2543 | /* |
cdd6c482 | 2544 | * Cross CPU call to read the hardware event |
0793a61d | 2545 | */ |
cdd6c482 | 2546 | static void __perf_event_read(void *info) |
0793a61d | 2547 | { |
cdd6c482 IM |
2548 | struct perf_event *event = info; |
2549 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2550 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2551 | |
e1ac3614 PM |
2552 | /* |
2553 | * If this is a task context, we need to check whether it is | |
2554 | * the current task context of this cpu. If not it has been | |
2555 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2556 | * event->count would have been updated to a recent sample |
2557 | * when the event was scheduled out. | |
e1ac3614 PM |
2558 | */ |
2559 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2560 | return; | |
2561 | ||
e625cce1 | 2562 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2563 | if (ctx->is_active) { |
542e72fc | 2564 | update_context_time(ctx); |
e5d1367f SE |
2565 | update_cgrp_time_from_event(event); |
2566 | } | |
cdd6c482 | 2567 | update_event_times(event); |
542e72fc PZ |
2568 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2569 | event->pmu->read(event); | |
e625cce1 | 2570 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2571 | } |
2572 | ||
b5e58793 PZ |
2573 | static inline u64 perf_event_count(struct perf_event *event) |
2574 | { | |
e7850595 | 2575 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
2576 | } |
2577 | ||
cdd6c482 | 2578 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
2579 | { |
2580 | /* | |
cdd6c482 IM |
2581 | * If event is enabled and currently active on a CPU, update the |
2582 | * value in the event structure: | |
0793a61d | 2583 | */ |
cdd6c482 IM |
2584 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2585 | smp_call_function_single(event->oncpu, | |
2586 | __perf_event_read, event, 1); | |
2587 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
2588 | struct perf_event_context *ctx = event->ctx; |
2589 | unsigned long flags; | |
2590 | ||
e625cce1 | 2591 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
2592 | /* |
2593 | * may read while context is not active | |
2594 | * (e.g., thread is blocked), in that case | |
2595 | * we cannot update context time | |
2596 | */ | |
e5d1367f | 2597 | if (ctx->is_active) { |
c530ccd9 | 2598 | update_context_time(ctx); |
e5d1367f SE |
2599 | update_cgrp_time_from_event(event); |
2600 | } | |
cdd6c482 | 2601 | update_event_times(event); |
e625cce1 | 2602 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
2603 | } |
2604 | ||
b5e58793 | 2605 | return perf_event_count(event); |
0793a61d TG |
2606 | } |
2607 | ||
a63eaf34 | 2608 | /* |
cdd6c482 | 2609 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 2610 | */ |
eb184479 | 2611 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 2612 | { |
e625cce1 | 2613 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 2614 | mutex_init(&ctx->mutex); |
889ff015 FW |
2615 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2616 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
2617 | INIT_LIST_HEAD(&ctx->event_list); |
2618 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
2619 | } |
2620 | ||
2621 | static struct perf_event_context * | |
2622 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
2623 | { | |
2624 | struct perf_event_context *ctx; | |
2625 | ||
2626 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
2627 | if (!ctx) | |
2628 | return NULL; | |
2629 | ||
2630 | __perf_event_init_context(ctx); | |
2631 | if (task) { | |
2632 | ctx->task = task; | |
2633 | get_task_struct(task); | |
0793a61d | 2634 | } |
eb184479 PZ |
2635 | ctx->pmu = pmu; |
2636 | ||
2637 | return ctx; | |
a63eaf34 PM |
2638 | } |
2639 | ||
2ebd4ffb MH |
2640 | static struct task_struct * |
2641 | find_lively_task_by_vpid(pid_t vpid) | |
2642 | { | |
2643 | struct task_struct *task; | |
2644 | int err; | |
0793a61d TG |
2645 | |
2646 | rcu_read_lock(); | |
2ebd4ffb | 2647 | if (!vpid) |
0793a61d TG |
2648 | task = current; |
2649 | else | |
2ebd4ffb | 2650 | task = find_task_by_vpid(vpid); |
0793a61d TG |
2651 | if (task) |
2652 | get_task_struct(task); | |
2653 | rcu_read_unlock(); | |
2654 | ||
2655 | if (!task) | |
2656 | return ERR_PTR(-ESRCH); | |
2657 | ||
0793a61d | 2658 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
2659 | err = -EACCES; |
2660 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
2661 | goto errout; | |
2662 | ||
2ebd4ffb MH |
2663 | return task; |
2664 | errout: | |
2665 | put_task_struct(task); | |
2666 | return ERR_PTR(err); | |
2667 | ||
2668 | } | |
2669 | ||
fe4b04fa PZ |
2670 | /* |
2671 | * Returns a matching context with refcount and pincount. | |
2672 | */ | |
108b02cf | 2673 | static struct perf_event_context * |
38a81da2 | 2674 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 2675 | { |
cdd6c482 | 2676 | struct perf_event_context *ctx; |
22a4f650 | 2677 | struct perf_cpu_context *cpuctx; |
25346b93 | 2678 | unsigned long flags; |
8dc85d54 | 2679 | int ctxn, err; |
0793a61d | 2680 | |
22a4ec72 | 2681 | if (!task) { |
cdd6c482 | 2682 | /* Must be root to operate on a CPU event: */ |
0764771d | 2683 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
2684 | return ERR_PTR(-EACCES); |
2685 | ||
0793a61d | 2686 | /* |
cdd6c482 | 2687 | * We could be clever and allow to attach a event to an |
0793a61d TG |
2688 | * offline CPU and activate it when the CPU comes up, but |
2689 | * that's for later. | |
2690 | */ | |
f6325e30 | 2691 | if (!cpu_online(cpu)) |
0793a61d TG |
2692 | return ERR_PTR(-ENODEV); |
2693 | ||
108b02cf | 2694 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 2695 | ctx = &cpuctx->ctx; |
c93f7669 | 2696 | get_ctx(ctx); |
fe4b04fa | 2697 | ++ctx->pin_count; |
0793a61d | 2698 | |
0793a61d TG |
2699 | return ctx; |
2700 | } | |
2701 | ||
8dc85d54 PZ |
2702 | err = -EINVAL; |
2703 | ctxn = pmu->task_ctx_nr; | |
2704 | if (ctxn < 0) | |
2705 | goto errout; | |
2706 | ||
9ed6060d | 2707 | retry: |
8dc85d54 | 2708 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 2709 | if (ctx) { |
71a851b4 | 2710 | unclone_ctx(ctx); |
fe4b04fa | 2711 | ++ctx->pin_count; |
e625cce1 | 2712 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 2713 | } else { |
eb184479 | 2714 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
2715 | err = -ENOMEM; |
2716 | if (!ctx) | |
2717 | goto errout; | |
eb184479 | 2718 | |
dbe08d82 ON |
2719 | err = 0; |
2720 | mutex_lock(&task->perf_event_mutex); | |
2721 | /* | |
2722 | * If it has already passed perf_event_exit_task(). | |
2723 | * we must see PF_EXITING, it takes this mutex too. | |
2724 | */ | |
2725 | if (task->flags & PF_EXITING) | |
2726 | err = -ESRCH; | |
2727 | else if (task->perf_event_ctxp[ctxn]) | |
2728 | err = -EAGAIN; | |
fe4b04fa | 2729 | else { |
9137fb28 | 2730 | get_ctx(ctx); |
fe4b04fa | 2731 | ++ctx->pin_count; |
dbe08d82 | 2732 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 2733 | } |
dbe08d82 ON |
2734 | mutex_unlock(&task->perf_event_mutex); |
2735 | ||
2736 | if (unlikely(err)) { | |
9137fb28 | 2737 | put_ctx(ctx); |
dbe08d82 ON |
2738 | |
2739 | if (err == -EAGAIN) | |
2740 | goto retry; | |
2741 | goto errout; | |
a63eaf34 PM |
2742 | } |
2743 | } | |
2744 | ||
0793a61d | 2745 | return ctx; |
c93f7669 | 2746 | |
9ed6060d | 2747 | errout: |
c93f7669 | 2748 | return ERR_PTR(err); |
0793a61d TG |
2749 | } |
2750 | ||
6fb2915d LZ |
2751 | static void perf_event_free_filter(struct perf_event *event); |
2752 | ||
cdd6c482 | 2753 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 2754 | { |
cdd6c482 | 2755 | struct perf_event *event; |
592903cd | 2756 | |
cdd6c482 IM |
2757 | event = container_of(head, struct perf_event, rcu_head); |
2758 | if (event->ns) | |
2759 | put_pid_ns(event->ns); | |
6fb2915d | 2760 | perf_event_free_filter(event); |
cdd6c482 | 2761 | kfree(event); |
592903cd PZ |
2762 | } |
2763 | ||
76369139 | 2764 | static void ring_buffer_put(struct ring_buffer *rb); |
925d519a | 2765 | |
cdd6c482 | 2766 | static void free_event(struct perf_event *event) |
f1600952 | 2767 | { |
e360adbe | 2768 | irq_work_sync(&event->pending); |
925d519a | 2769 | |
cdd6c482 | 2770 | if (!event->parent) { |
82cd6def | 2771 | if (event->attach_state & PERF_ATTACH_TASK) |
b2029520 | 2772 | jump_label_dec_deferred(&perf_sched_events); |
3af9e859 | 2773 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
2774 | atomic_dec(&nr_mmap_events); |
2775 | if (event->attr.comm) | |
2776 | atomic_dec(&nr_comm_events); | |
2777 | if (event->attr.task) | |
2778 | atomic_dec(&nr_task_events); | |
927c7a9e FW |
2779 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
2780 | put_callchain_buffers(); | |
08309379 PZ |
2781 | if (is_cgroup_event(event)) { |
2782 | atomic_dec(&per_cpu(perf_cgroup_events, event->cpu)); | |
b2029520 | 2783 | jump_label_dec_deferred(&perf_sched_events); |
08309379 | 2784 | } |
f344011c | 2785 | } |
9ee318a7 | 2786 | |
76369139 FW |
2787 | if (event->rb) { |
2788 | ring_buffer_put(event->rb); | |
2789 | event->rb = NULL; | |
a4be7c27 PZ |
2790 | } |
2791 | ||
e5d1367f SE |
2792 | if (is_cgroup_event(event)) |
2793 | perf_detach_cgroup(event); | |
2794 | ||
cdd6c482 IM |
2795 | if (event->destroy) |
2796 | event->destroy(event); | |
e077df4f | 2797 | |
0c67b408 PZ |
2798 | if (event->ctx) |
2799 | put_ctx(event->ctx); | |
2800 | ||
cdd6c482 | 2801 | call_rcu(&event->rcu_head, free_event_rcu); |
f1600952 PZ |
2802 | } |
2803 | ||
a66a3052 | 2804 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 2805 | { |
cdd6c482 | 2806 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 2807 | |
ad3a37de | 2808 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
2809 | /* |
2810 | * There are two ways this annotation is useful: | |
2811 | * | |
2812 | * 1) there is a lock recursion from perf_event_exit_task | |
2813 | * see the comment there. | |
2814 | * | |
2815 | * 2) there is a lock-inversion with mmap_sem through | |
2816 | * perf_event_read_group(), which takes faults while | |
2817 | * holding ctx->mutex, however this is called after | |
2818 | * the last filedesc died, so there is no possibility | |
2819 | * to trigger the AB-BA case. | |
2820 | */ | |
2821 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
050735b0 | 2822 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 2823 | perf_group_detach(event); |
050735b0 | 2824 | raw_spin_unlock_irq(&ctx->lock); |
e03a9a55 | 2825 | perf_remove_from_context(event); |
d859e29f | 2826 | mutex_unlock(&ctx->mutex); |
0793a61d | 2827 | |
cdd6c482 | 2828 | free_event(event); |
0793a61d TG |
2829 | |
2830 | return 0; | |
2831 | } | |
a66a3052 | 2832 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 2833 | |
a66a3052 PZ |
2834 | /* |
2835 | * Called when the last reference to the file is gone. | |
2836 | */ | |
2837 | static int perf_release(struct inode *inode, struct file *file) | |
fb0459d7 | 2838 | { |
a66a3052 | 2839 | struct perf_event *event = file->private_data; |
8882135b | 2840 | struct task_struct *owner; |
fb0459d7 | 2841 | |
a66a3052 | 2842 | file->private_data = NULL; |
fb0459d7 | 2843 | |
8882135b PZ |
2844 | rcu_read_lock(); |
2845 | owner = ACCESS_ONCE(event->owner); | |
2846 | /* | |
2847 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
2848 | * !owner it means the list deletion is complete and we can indeed | |
2849 | * free this event, otherwise we need to serialize on | |
2850 | * owner->perf_event_mutex. | |
2851 | */ | |
2852 | smp_read_barrier_depends(); | |
2853 | if (owner) { | |
2854 | /* | |
2855 | * Since delayed_put_task_struct() also drops the last | |
2856 | * task reference we can safely take a new reference | |
2857 | * while holding the rcu_read_lock(). | |
2858 | */ | |
2859 | get_task_struct(owner); | |
2860 | } | |
2861 | rcu_read_unlock(); | |
2862 | ||
2863 | if (owner) { | |
2864 | mutex_lock(&owner->perf_event_mutex); | |
2865 | /* | |
2866 | * We have to re-check the event->owner field, if it is cleared | |
2867 | * we raced with perf_event_exit_task(), acquiring the mutex | |
2868 | * ensured they're done, and we can proceed with freeing the | |
2869 | * event. | |
2870 | */ | |
2871 | if (event->owner) | |
2872 | list_del_init(&event->owner_entry); | |
2873 | mutex_unlock(&owner->perf_event_mutex); | |
2874 | put_task_struct(owner); | |
2875 | } | |
2876 | ||
a66a3052 | 2877 | return perf_event_release_kernel(event); |
fb0459d7 | 2878 | } |
fb0459d7 | 2879 | |
59ed446f | 2880 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 2881 | { |
cdd6c482 | 2882 | struct perf_event *child; |
e53c0994 PZ |
2883 | u64 total = 0; |
2884 | ||
59ed446f PZ |
2885 | *enabled = 0; |
2886 | *running = 0; | |
2887 | ||
6f10581a | 2888 | mutex_lock(&event->child_mutex); |
cdd6c482 | 2889 | total += perf_event_read(event); |
59ed446f PZ |
2890 | *enabled += event->total_time_enabled + |
2891 | atomic64_read(&event->child_total_time_enabled); | |
2892 | *running += event->total_time_running + | |
2893 | atomic64_read(&event->child_total_time_running); | |
2894 | ||
2895 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 2896 | total += perf_event_read(child); |
59ed446f PZ |
2897 | *enabled += child->total_time_enabled; |
2898 | *running += child->total_time_running; | |
2899 | } | |
6f10581a | 2900 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
2901 | |
2902 | return total; | |
2903 | } | |
fb0459d7 | 2904 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 2905 | |
cdd6c482 | 2906 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
2907 | u64 read_format, char __user *buf) |
2908 | { | |
cdd6c482 | 2909 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
2910 | int n = 0, size = 0, ret = -EFAULT; |
2911 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 2912 | u64 values[5]; |
59ed446f | 2913 | u64 count, enabled, running; |
abf4868b | 2914 | |
6f10581a | 2915 | mutex_lock(&ctx->mutex); |
59ed446f | 2916 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
2917 | |
2918 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
2919 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
2920 | values[n++] = enabled; | |
2921 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
2922 | values[n++] = running; | |
abf4868b PZ |
2923 | values[n++] = count; |
2924 | if (read_format & PERF_FORMAT_ID) | |
2925 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
2926 | |
2927 | size = n * sizeof(u64); | |
2928 | ||
2929 | if (copy_to_user(buf, values, size)) | |
6f10581a | 2930 | goto unlock; |
3dab77fb | 2931 | |
6f10581a | 2932 | ret = size; |
3dab77fb | 2933 | |
65abc865 | 2934 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 2935 | n = 0; |
3dab77fb | 2936 | |
59ed446f | 2937 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
2938 | if (read_format & PERF_FORMAT_ID) |
2939 | values[n++] = primary_event_id(sub); | |
2940 | ||
2941 | size = n * sizeof(u64); | |
2942 | ||
184d3da8 | 2943 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
2944 | ret = -EFAULT; |
2945 | goto unlock; | |
2946 | } | |
abf4868b PZ |
2947 | |
2948 | ret += size; | |
3dab77fb | 2949 | } |
6f10581a PZ |
2950 | unlock: |
2951 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 2952 | |
abf4868b | 2953 | return ret; |
3dab77fb PZ |
2954 | } |
2955 | ||
cdd6c482 | 2956 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
2957 | u64 read_format, char __user *buf) |
2958 | { | |
59ed446f | 2959 | u64 enabled, running; |
3dab77fb PZ |
2960 | u64 values[4]; |
2961 | int n = 0; | |
2962 | ||
59ed446f PZ |
2963 | values[n++] = perf_event_read_value(event, &enabled, &running); |
2964 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
2965 | values[n++] = enabled; | |
2966 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
2967 | values[n++] = running; | |
3dab77fb | 2968 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 2969 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
2970 | |
2971 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
2972 | return -EFAULT; | |
2973 | ||
2974 | return n * sizeof(u64); | |
2975 | } | |
2976 | ||
0793a61d | 2977 | /* |
cdd6c482 | 2978 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
2979 | */ |
2980 | static ssize_t | |
cdd6c482 | 2981 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 2982 | { |
cdd6c482 | 2983 | u64 read_format = event->attr.read_format; |
3dab77fb | 2984 | int ret; |
0793a61d | 2985 | |
3b6f9e5c | 2986 | /* |
cdd6c482 | 2987 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
2988 | * error state (i.e. because it was pinned but it couldn't be |
2989 | * scheduled on to the CPU at some point). | |
2990 | */ | |
cdd6c482 | 2991 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
2992 | return 0; |
2993 | ||
c320c7b7 | 2994 | if (count < event->read_size) |
3dab77fb PZ |
2995 | return -ENOSPC; |
2996 | ||
cdd6c482 | 2997 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 2998 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 2999 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3000 | else |
cdd6c482 | 3001 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3002 | |
3dab77fb | 3003 | return ret; |
0793a61d TG |
3004 | } |
3005 | ||
0793a61d TG |
3006 | static ssize_t |
3007 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3008 | { | |
cdd6c482 | 3009 | struct perf_event *event = file->private_data; |
0793a61d | 3010 | |
cdd6c482 | 3011 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3012 | } |
3013 | ||
3014 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3015 | { | |
cdd6c482 | 3016 | struct perf_event *event = file->private_data; |
76369139 | 3017 | struct ring_buffer *rb; |
c33a0bc4 | 3018 | unsigned int events = POLL_HUP; |
c7138f37 | 3019 | |
10c6db11 PZ |
3020 | /* |
3021 | * Race between perf_event_set_output() and perf_poll(): perf_poll() | |
3022 | * grabs the rb reference but perf_event_set_output() overrides it. | |
3023 | * Here is the timeline for two threads T1, T2: | |
3024 | * t0: T1, rb = rcu_dereference(event->rb) | |
3025 | * t1: T2, old_rb = event->rb | |
3026 | * t2: T2, event->rb = new rb | |
3027 | * t3: T2, ring_buffer_detach(old_rb) | |
3028 | * t4: T1, ring_buffer_attach(rb1) | |
3029 | * t5: T1, poll_wait(event->waitq) | |
3030 | * | |
3031 | * To avoid this problem, we grab mmap_mutex in perf_poll() | |
3032 | * thereby ensuring that the assignment of the new ring buffer | |
3033 | * and the detachment of the old buffer appear atomic to perf_poll() | |
3034 | */ | |
3035 | mutex_lock(&event->mmap_mutex); | |
3036 | ||
c7138f37 | 3037 | rcu_read_lock(); |
76369139 | 3038 | rb = rcu_dereference(event->rb); |
10c6db11 PZ |
3039 | if (rb) { |
3040 | ring_buffer_attach(event, rb); | |
76369139 | 3041 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3042 | } |
c7138f37 | 3043 | rcu_read_unlock(); |
0793a61d | 3044 | |
10c6db11 PZ |
3045 | mutex_unlock(&event->mmap_mutex); |
3046 | ||
cdd6c482 | 3047 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3048 | |
0793a61d TG |
3049 | return events; |
3050 | } | |
3051 | ||
cdd6c482 | 3052 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3053 | { |
cdd6c482 | 3054 | (void)perf_event_read(event); |
e7850595 | 3055 | local64_set(&event->count, 0); |
cdd6c482 | 3056 | perf_event_update_userpage(event); |
3df5edad PZ |
3057 | } |
3058 | ||
c93f7669 | 3059 | /* |
cdd6c482 IM |
3060 | * Holding the top-level event's child_mutex means that any |
3061 | * descendant process that has inherited this event will block | |
3062 | * in sync_child_event if it goes to exit, thus satisfying the | |
3063 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3064 | */ |
cdd6c482 IM |
3065 | static void perf_event_for_each_child(struct perf_event *event, |
3066 | void (*func)(struct perf_event *)) | |
3df5edad | 3067 | { |
cdd6c482 | 3068 | struct perf_event *child; |
3df5edad | 3069 | |
cdd6c482 IM |
3070 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3071 | mutex_lock(&event->child_mutex); | |
3072 | func(event); | |
3073 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3074 | func(child); |
cdd6c482 | 3075 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3076 | } |
3077 | ||
cdd6c482 IM |
3078 | static void perf_event_for_each(struct perf_event *event, |
3079 | void (*func)(struct perf_event *)) | |
3df5edad | 3080 | { |
cdd6c482 IM |
3081 | struct perf_event_context *ctx = event->ctx; |
3082 | struct perf_event *sibling; | |
3df5edad | 3083 | |
75f937f2 PZ |
3084 | WARN_ON_ONCE(ctx->parent_ctx); |
3085 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3086 | event = event->group_leader; |
75f937f2 | 3087 | |
cdd6c482 IM |
3088 | perf_event_for_each_child(event, func); |
3089 | func(event); | |
3090 | list_for_each_entry(sibling, &event->sibling_list, group_entry) | |
3091 | perf_event_for_each_child(event, func); | |
75f937f2 | 3092 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3093 | } |
3094 | ||
cdd6c482 | 3095 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3096 | { |
cdd6c482 | 3097 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3098 | int ret = 0; |
3099 | u64 value; | |
3100 | ||
6c7e550f | 3101 | if (!is_sampling_event(event)) |
08247e31 PZ |
3102 | return -EINVAL; |
3103 | ||
ad0cf347 | 3104 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3105 | return -EFAULT; |
3106 | ||
3107 | if (!value) | |
3108 | return -EINVAL; | |
3109 | ||
e625cce1 | 3110 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3111 | if (event->attr.freq) { |
3112 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3113 | ret = -EINVAL; |
3114 | goto unlock; | |
3115 | } | |
3116 | ||
cdd6c482 | 3117 | event->attr.sample_freq = value; |
08247e31 | 3118 | } else { |
cdd6c482 IM |
3119 | event->attr.sample_period = value; |
3120 | event->hw.sample_period = value; | |
08247e31 PZ |
3121 | } |
3122 | unlock: | |
e625cce1 | 3123 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3124 | |
3125 | return ret; | |
3126 | } | |
3127 | ||
ac9721f3 PZ |
3128 | static const struct file_operations perf_fops; |
3129 | ||
3130 | static struct perf_event *perf_fget_light(int fd, int *fput_needed) | |
3131 | { | |
3132 | struct file *file; | |
3133 | ||
3134 | file = fget_light(fd, fput_needed); | |
3135 | if (!file) | |
3136 | return ERR_PTR(-EBADF); | |
3137 | ||
3138 | if (file->f_op != &perf_fops) { | |
3139 | fput_light(file, *fput_needed); | |
3140 | *fput_needed = 0; | |
3141 | return ERR_PTR(-EBADF); | |
3142 | } | |
3143 | ||
3144 | return file->private_data; | |
3145 | } | |
3146 | ||
3147 | static int perf_event_set_output(struct perf_event *event, | |
3148 | struct perf_event *output_event); | |
6fb2915d | 3149 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3150 | |
d859e29f PM |
3151 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3152 | { | |
cdd6c482 IM |
3153 | struct perf_event *event = file->private_data; |
3154 | void (*func)(struct perf_event *); | |
3df5edad | 3155 | u32 flags = arg; |
d859e29f PM |
3156 | |
3157 | switch (cmd) { | |
cdd6c482 IM |
3158 | case PERF_EVENT_IOC_ENABLE: |
3159 | func = perf_event_enable; | |
d859e29f | 3160 | break; |
cdd6c482 IM |
3161 | case PERF_EVENT_IOC_DISABLE: |
3162 | func = perf_event_disable; | |
79f14641 | 3163 | break; |
cdd6c482 IM |
3164 | case PERF_EVENT_IOC_RESET: |
3165 | func = perf_event_reset; | |
6de6a7b9 | 3166 | break; |
3df5edad | 3167 | |
cdd6c482 IM |
3168 | case PERF_EVENT_IOC_REFRESH: |
3169 | return perf_event_refresh(event, arg); | |
08247e31 | 3170 | |
cdd6c482 IM |
3171 | case PERF_EVENT_IOC_PERIOD: |
3172 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3173 | |
cdd6c482 | 3174 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 PZ |
3175 | { |
3176 | struct perf_event *output_event = NULL; | |
3177 | int fput_needed = 0; | |
3178 | int ret; | |
3179 | ||
3180 | if (arg != -1) { | |
3181 | output_event = perf_fget_light(arg, &fput_needed); | |
3182 | if (IS_ERR(output_event)) | |
3183 | return PTR_ERR(output_event); | |
3184 | } | |
3185 | ||
3186 | ret = perf_event_set_output(event, output_event); | |
3187 | if (output_event) | |
3188 | fput_light(output_event->filp, fput_needed); | |
3189 | ||
3190 | return ret; | |
3191 | } | |
a4be7c27 | 3192 | |
6fb2915d LZ |
3193 | case PERF_EVENT_IOC_SET_FILTER: |
3194 | return perf_event_set_filter(event, (void __user *)arg); | |
3195 | ||
d859e29f | 3196 | default: |
3df5edad | 3197 | return -ENOTTY; |
d859e29f | 3198 | } |
3df5edad PZ |
3199 | |
3200 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3201 | perf_event_for_each(event, func); |
3df5edad | 3202 | else |
cdd6c482 | 3203 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3204 | |
3205 | return 0; | |
d859e29f PM |
3206 | } |
3207 | ||
cdd6c482 | 3208 | int perf_event_task_enable(void) |
771d7cde | 3209 | { |
cdd6c482 | 3210 | struct perf_event *event; |
771d7cde | 3211 | |
cdd6c482 IM |
3212 | mutex_lock(¤t->perf_event_mutex); |
3213 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3214 | perf_event_for_each_child(event, perf_event_enable); | |
3215 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3216 | |
3217 | return 0; | |
3218 | } | |
3219 | ||
cdd6c482 | 3220 | int perf_event_task_disable(void) |
771d7cde | 3221 | { |
cdd6c482 | 3222 | struct perf_event *event; |
771d7cde | 3223 | |
cdd6c482 IM |
3224 | mutex_lock(¤t->perf_event_mutex); |
3225 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3226 | perf_event_for_each_child(event, perf_event_disable); | |
3227 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3228 | |
3229 | return 0; | |
3230 | } | |
3231 | ||
cdd6c482 IM |
3232 | #ifndef PERF_EVENT_INDEX_OFFSET |
3233 | # define PERF_EVENT_INDEX_OFFSET 0 | |
f738eb1b IM |
3234 | #endif |
3235 | ||
cdd6c482 | 3236 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3237 | { |
a4eaf7f1 PZ |
3238 | if (event->hw.state & PERF_HES_STOPPED) |
3239 | return 0; | |
3240 | ||
cdd6c482 | 3241 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3242 | return 0; |
3243 | ||
cdd6c482 | 3244 | return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET; |
194002b2 PZ |
3245 | } |
3246 | ||
c4794295 | 3247 | static void calc_timer_values(struct perf_event *event, |
7f310a5d EM |
3248 | u64 *enabled, |
3249 | u64 *running) | |
c4794295 EM |
3250 | { |
3251 | u64 now, ctx_time; | |
3252 | ||
3253 | now = perf_clock(); | |
3254 | ctx_time = event->shadow_ctx_time + now; | |
3255 | *enabled = ctx_time - event->tstamp_enabled; | |
3256 | *running = ctx_time - event->tstamp_running; | |
3257 | } | |
3258 | ||
38ff667b PZ |
3259 | /* |
3260 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3261 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3262 | * code calls this from NMI context. | |
3263 | */ | |
cdd6c482 | 3264 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3265 | { |
cdd6c482 | 3266 | struct perf_event_mmap_page *userpg; |
76369139 | 3267 | struct ring_buffer *rb; |
0d641208 | 3268 | u64 enabled, running; |
38ff667b PZ |
3269 | |
3270 | rcu_read_lock(); | |
0d641208 EM |
3271 | /* |
3272 | * compute total_time_enabled, total_time_running | |
3273 | * based on snapshot values taken when the event | |
3274 | * was last scheduled in. | |
3275 | * | |
3276 | * we cannot simply called update_context_time() | |
3277 | * because of locking issue as we can be called in | |
3278 | * NMI context | |
3279 | */ | |
3280 | calc_timer_values(event, &enabled, &running); | |
76369139 FW |
3281 | rb = rcu_dereference(event->rb); |
3282 | if (!rb) | |
38ff667b PZ |
3283 | goto unlock; |
3284 | ||
76369139 | 3285 | userpg = rb->user_page; |
37d81828 | 3286 | |
7b732a75 PZ |
3287 | /* |
3288 | * Disable preemption so as to not let the corresponding user-space | |
3289 | * spin too long if we get preempted. | |
3290 | */ | |
3291 | preempt_disable(); | |
37d81828 | 3292 | ++userpg->lock; |
92f22a38 | 3293 | barrier(); |
cdd6c482 | 3294 | userpg->index = perf_event_index(event); |
b5e58793 | 3295 | userpg->offset = perf_event_count(event); |
cdd6c482 | 3296 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
e7850595 | 3297 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3298 | |
0d641208 | 3299 | userpg->time_enabled = enabled + |
cdd6c482 | 3300 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3301 | |
0d641208 | 3302 | userpg->time_running = running + |
cdd6c482 | 3303 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3304 | |
92f22a38 | 3305 | barrier(); |
37d81828 | 3306 | ++userpg->lock; |
7b732a75 | 3307 | preempt_enable(); |
38ff667b | 3308 | unlock: |
7b732a75 | 3309 | rcu_read_unlock(); |
37d81828 PM |
3310 | } |
3311 | ||
906010b2 PZ |
3312 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3313 | { | |
3314 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3315 | struct ring_buffer *rb; |
906010b2 PZ |
3316 | int ret = VM_FAULT_SIGBUS; |
3317 | ||
3318 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3319 | if (vmf->pgoff == 0) | |
3320 | ret = 0; | |
3321 | return ret; | |
3322 | } | |
3323 | ||
3324 | rcu_read_lock(); | |
76369139 FW |
3325 | rb = rcu_dereference(event->rb); |
3326 | if (!rb) | |
906010b2 PZ |
3327 | goto unlock; |
3328 | ||
3329 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3330 | goto unlock; | |
3331 | ||
76369139 | 3332 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3333 | if (!vmf->page) |
3334 | goto unlock; | |
3335 | ||
3336 | get_page(vmf->page); | |
3337 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3338 | vmf->page->index = vmf->pgoff; | |
3339 | ||
3340 | ret = 0; | |
3341 | unlock: | |
3342 | rcu_read_unlock(); | |
3343 | ||
3344 | return ret; | |
3345 | } | |
3346 | ||
10c6db11 PZ |
3347 | static void ring_buffer_attach(struct perf_event *event, |
3348 | struct ring_buffer *rb) | |
3349 | { | |
3350 | unsigned long flags; | |
3351 | ||
3352 | if (!list_empty(&event->rb_entry)) | |
3353 | return; | |
3354 | ||
3355 | spin_lock_irqsave(&rb->event_lock, flags); | |
3356 | if (!list_empty(&event->rb_entry)) | |
3357 | goto unlock; | |
3358 | ||
3359 | list_add(&event->rb_entry, &rb->event_list); | |
3360 | unlock: | |
3361 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3362 | } | |
3363 | ||
3364 | static void ring_buffer_detach(struct perf_event *event, | |
3365 | struct ring_buffer *rb) | |
3366 | { | |
3367 | unsigned long flags; | |
3368 | ||
3369 | if (list_empty(&event->rb_entry)) | |
3370 | return; | |
3371 | ||
3372 | spin_lock_irqsave(&rb->event_lock, flags); | |
3373 | list_del_init(&event->rb_entry); | |
3374 | wake_up_all(&event->waitq); | |
3375 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3376 | } | |
3377 | ||
3378 | static void ring_buffer_wakeup(struct perf_event *event) | |
3379 | { | |
3380 | struct ring_buffer *rb; | |
3381 | ||
3382 | rcu_read_lock(); | |
3383 | rb = rcu_dereference(event->rb); | |
44b7f4b9 WD |
3384 | if (!rb) |
3385 | goto unlock; | |
3386 | ||
3387 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
10c6db11 | 3388 | wake_up_all(&event->waitq); |
44b7f4b9 WD |
3389 | |
3390 | unlock: | |
10c6db11 PZ |
3391 | rcu_read_unlock(); |
3392 | } | |
3393 | ||
76369139 | 3394 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3395 | { |
76369139 | 3396 | struct ring_buffer *rb; |
906010b2 | 3397 | |
76369139 FW |
3398 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3399 | rb_free(rb); | |
7b732a75 PZ |
3400 | } |
3401 | ||
76369139 | 3402 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3403 | { |
76369139 | 3404 | struct ring_buffer *rb; |
7b732a75 | 3405 | |
ac9721f3 | 3406 | rcu_read_lock(); |
76369139 FW |
3407 | rb = rcu_dereference(event->rb); |
3408 | if (rb) { | |
3409 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3410 | rb = NULL; | |
ac9721f3 PZ |
3411 | } |
3412 | rcu_read_unlock(); | |
3413 | ||
76369139 | 3414 | return rb; |
ac9721f3 PZ |
3415 | } |
3416 | ||
76369139 | 3417 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3418 | { |
10c6db11 PZ |
3419 | struct perf_event *event, *n; |
3420 | unsigned long flags; | |
3421 | ||
76369139 | 3422 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3423 | return; |
7b732a75 | 3424 | |
10c6db11 PZ |
3425 | spin_lock_irqsave(&rb->event_lock, flags); |
3426 | list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) { | |
3427 | list_del_init(&event->rb_entry); | |
3428 | wake_up_all(&event->waitq); | |
3429 | } | |
3430 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3431 | ||
76369139 | 3432 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3433 | } |
3434 | ||
3435 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3436 | { | |
cdd6c482 | 3437 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3438 | |
cdd6c482 | 3439 | atomic_inc(&event->mmap_count); |
7b732a75 PZ |
3440 | } |
3441 | ||
3442 | static void perf_mmap_close(struct vm_area_struct *vma) | |
3443 | { | |
cdd6c482 | 3444 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3445 | |
cdd6c482 | 3446 | if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) { |
76369139 | 3447 | unsigned long size = perf_data_size(event->rb); |
ac9721f3 | 3448 | struct user_struct *user = event->mmap_user; |
76369139 | 3449 | struct ring_buffer *rb = event->rb; |
789f90fc | 3450 | |
906010b2 | 3451 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm); |
bc3e53f6 | 3452 | vma->vm_mm->pinned_vm -= event->mmap_locked; |
76369139 | 3453 | rcu_assign_pointer(event->rb, NULL); |
10c6db11 | 3454 | ring_buffer_detach(event, rb); |
cdd6c482 | 3455 | mutex_unlock(&event->mmap_mutex); |
ac9721f3 | 3456 | |
76369139 | 3457 | ring_buffer_put(rb); |
ac9721f3 | 3458 | free_uid(user); |
7b732a75 | 3459 | } |
37d81828 PM |
3460 | } |
3461 | ||
f0f37e2f | 3462 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3463 | .open = perf_mmap_open, |
3464 | .close = perf_mmap_close, | |
3465 | .fault = perf_mmap_fault, | |
3466 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3467 | }; |
3468 | ||
3469 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3470 | { | |
cdd6c482 | 3471 | struct perf_event *event = file->private_data; |
22a4f650 | 3472 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3473 | struct user_struct *user = current_user(); |
22a4f650 | 3474 | unsigned long locked, lock_limit; |
76369139 | 3475 | struct ring_buffer *rb; |
7b732a75 PZ |
3476 | unsigned long vma_size; |
3477 | unsigned long nr_pages; | |
789f90fc | 3478 | long user_extra, extra; |
d57e34fd | 3479 | int ret = 0, flags = 0; |
37d81828 | 3480 | |
c7920614 PZ |
3481 | /* |
3482 | * Don't allow mmap() of inherited per-task counters. This would | |
3483 | * create a performance issue due to all children writing to the | |
76369139 | 3484 | * same rb. |
c7920614 PZ |
3485 | */ |
3486 | if (event->cpu == -1 && event->attr.inherit) | |
3487 | return -EINVAL; | |
3488 | ||
43a21ea8 | 3489 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3490 | return -EINVAL; |
7b732a75 PZ |
3491 | |
3492 | vma_size = vma->vm_end - vma->vm_start; | |
3493 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3494 | ||
7730d865 | 3495 | /* |
76369139 | 3496 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3497 | * can do bitmasks instead of modulo. |
3498 | */ | |
3499 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3500 | return -EINVAL; |
3501 | ||
7b732a75 | 3502 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3503 | return -EINVAL; |
3504 | ||
7b732a75 PZ |
3505 | if (vma->vm_pgoff != 0) |
3506 | return -EINVAL; | |
37d81828 | 3507 | |
cdd6c482 IM |
3508 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3509 | mutex_lock(&event->mmap_mutex); | |
76369139 FW |
3510 | if (event->rb) { |
3511 | if (event->rb->nr_pages == nr_pages) | |
3512 | atomic_inc(&event->rb->refcount); | |
ac9721f3 | 3513 | else |
ebb3c4c4 PZ |
3514 | ret = -EINVAL; |
3515 | goto unlock; | |
3516 | } | |
3517 | ||
789f90fc | 3518 | user_extra = nr_pages + 1; |
cdd6c482 | 3519 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
3520 | |
3521 | /* | |
3522 | * Increase the limit linearly with more CPUs: | |
3523 | */ | |
3524 | user_lock_limit *= num_online_cpus(); | |
3525 | ||
789f90fc | 3526 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 3527 | |
789f90fc PZ |
3528 | extra = 0; |
3529 | if (user_locked > user_lock_limit) | |
3530 | extra = user_locked - user_lock_limit; | |
7b732a75 | 3531 | |
78d7d407 | 3532 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 3533 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 3534 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 3535 | |
459ec28a IM |
3536 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
3537 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
3538 | ret = -EPERM; |
3539 | goto unlock; | |
3540 | } | |
7b732a75 | 3541 | |
76369139 | 3542 | WARN_ON(event->rb); |
906010b2 | 3543 | |
d57e34fd | 3544 | if (vma->vm_flags & VM_WRITE) |
76369139 | 3545 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 3546 | |
4ec8363d VW |
3547 | rb = rb_alloc(nr_pages, |
3548 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
3549 | event->cpu, flags); | |
3550 | ||
76369139 | 3551 | if (!rb) { |
ac9721f3 | 3552 | ret = -ENOMEM; |
ebb3c4c4 | 3553 | goto unlock; |
ac9721f3 | 3554 | } |
76369139 | 3555 | rcu_assign_pointer(event->rb, rb); |
43a21ea8 | 3556 | |
ac9721f3 PZ |
3557 | atomic_long_add(user_extra, &user->locked_vm); |
3558 | event->mmap_locked = extra; | |
3559 | event->mmap_user = get_current_user(); | |
bc3e53f6 | 3560 | vma->vm_mm->pinned_vm += event->mmap_locked; |
ac9721f3 | 3561 | |
ebb3c4c4 | 3562 | unlock: |
ac9721f3 PZ |
3563 | if (!ret) |
3564 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 3565 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 3566 | |
37d81828 PM |
3567 | vma->vm_flags |= VM_RESERVED; |
3568 | vma->vm_ops = &perf_mmap_vmops; | |
7b732a75 PZ |
3569 | |
3570 | return ret; | |
37d81828 PM |
3571 | } |
3572 | ||
3c446b3d PZ |
3573 | static int perf_fasync(int fd, struct file *filp, int on) |
3574 | { | |
3c446b3d | 3575 | struct inode *inode = filp->f_path.dentry->d_inode; |
cdd6c482 | 3576 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
3577 | int retval; |
3578 | ||
3579 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 3580 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
3581 | mutex_unlock(&inode->i_mutex); |
3582 | ||
3583 | if (retval < 0) | |
3584 | return retval; | |
3585 | ||
3586 | return 0; | |
3587 | } | |
3588 | ||
0793a61d | 3589 | static const struct file_operations perf_fops = { |
3326c1ce | 3590 | .llseek = no_llseek, |
0793a61d TG |
3591 | .release = perf_release, |
3592 | .read = perf_read, | |
3593 | .poll = perf_poll, | |
d859e29f PM |
3594 | .unlocked_ioctl = perf_ioctl, |
3595 | .compat_ioctl = perf_ioctl, | |
37d81828 | 3596 | .mmap = perf_mmap, |
3c446b3d | 3597 | .fasync = perf_fasync, |
0793a61d TG |
3598 | }; |
3599 | ||
925d519a | 3600 | /* |
cdd6c482 | 3601 | * Perf event wakeup |
925d519a PZ |
3602 | * |
3603 | * If there's data, ensure we set the poll() state and publish everything | |
3604 | * to user-space before waking everybody up. | |
3605 | */ | |
3606 | ||
cdd6c482 | 3607 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 3608 | { |
10c6db11 | 3609 | ring_buffer_wakeup(event); |
4c9e2542 | 3610 | |
cdd6c482 IM |
3611 | if (event->pending_kill) { |
3612 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
3613 | event->pending_kill = 0; | |
4c9e2542 | 3614 | } |
925d519a PZ |
3615 | } |
3616 | ||
e360adbe | 3617 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 3618 | { |
cdd6c482 IM |
3619 | struct perf_event *event = container_of(entry, |
3620 | struct perf_event, pending); | |
79f14641 | 3621 | |
cdd6c482 IM |
3622 | if (event->pending_disable) { |
3623 | event->pending_disable = 0; | |
3624 | __perf_event_disable(event); | |
79f14641 PZ |
3625 | } |
3626 | ||
cdd6c482 IM |
3627 | if (event->pending_wakeup) { |
3628 | event->pending_wakeup = 0; | |
3629 | perf_event_wakeup(event); | |
79f14641 PZ |
3630 | } |
3631 | } | |
3632 | ||
39447b38 ZY |
3633 | /* |
3634 | * We assume there is only KVM supporting the callbacks. | |
3635 | * Later on, we might change it to a list if there is | |
3636 | * another virtualization implementation supporting the callbacks. | |
3637 | */ | |
3638 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
3639 | ||
3640 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3641 | { | |
3642 | perf_guest_cbs = cbs; | |
3643 | return 0; | |
3644 | } | |
3645 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
3646 | ||
3647 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3648 | { | |
3649 | perf_guest_cbs = NULL; | |
3650 | return 0; | |
3651 | } | |
3652 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
3653 | ||
c980d109 ACM |
3654 | static void __perf_event_header__init_id(struct perf_event_header *header, |
3655 | struct perf_sample_data *data, | |
3656 | struct perf_event *event) | |
6844c09d ACM |
3657 | { |
3658 | u64 sample_type = event->attr.sample_type; | |
3659 | ||
3660 | data->type = sample_type; | |
3661 | header->size += event->id_header_size; | |
3662 | ||
3663 | if (sample_type & PERF_SAMPLE_TID) { | |
3664 | /* namespace issues */ | |
3665 | data->tid_entry.pid = perf_event_pid(event, current); | |
3666 | data->tid_entry.tid = perf_event_tid(event, current); | |
3667 | } | |
3668 | ||
3669 | if (sample_type & PERF_SAMPLE_TIME) | |
3670 | data->time = perf_clock(); | |
3671 | ||
3672 | if (sample_type & PERF_SAMPLE_ID) | |
3673 | data->id = primary_event_id(event); | |
3674 | ||
3675 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3676 | data->stream_id = event->id; | |
3677 | ||
3678 | if (sample_type & PERF_SAMPLE_CPU) { | |
3679 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
3680 | data->cpu_entry.reserved = 0; | |
3681 | } | |
3682 | } | |
3683 | ||
76369139 FW |
3684 | void perf_event_header__init_id(struct perf_event_header *header, |
3685 | struct perf_sample_data *data, | |
3686 | struct perf_event *event) | |
c980d109 ACM |
3687 | { |
3688 | if (event->attr.sample_id_all) | |
3689 | __perf_event_header__init_id(header, data, event); | |
3690 | } | |
3691 | ||
3692 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
3693 | struct perf_sample_data *data) | |
3694 | { | |
3695 | u64 sample_type = data->type; | |
3696 | ||
3697 | if (sample_type & PERF_SAMPLE_TID) | |
3698 | perf_output_put(handle, data->tid_entry); | |
3699 | ||
3700 | if (sample_type & PERF_SAMPLE_TIME) | |
3701 | perf_output_put(handle, data->time); | |
3702 | ||
3703 | if (sample_type & PERF_SAMPLE_ID) | |
3704 | perf_output_put(handle, data->id); | |
3705 | ||
3706 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3707 | perf_output_put(handle, data->stream_id); | |
3708 | ||
3709 | if (sample_type & PERF_SAMPLE_CPU) | |
3710 | perf_output_put(handle, data->cpu_entry); | |
3711 | } | |
3712 | ||
76369139 FW |
3713 | void perf_event__output_id_sample(struct perf_event *event, |
3714 | struct perf_output_handle *handle, | |
3715 | struct perf_sample_data *sample) | |
c980d109 ACM |
3716 | { |
3717 | if (event->attr.sample_id_all) | |
3718 | __perf_event__output_id_sample(handle, sample); | |
3719 | } | |
3720 | ||
3dab77fb | 3721 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
3722 | struct perf_event *event, |
3723 | u64 enabled, u64 running) | |
3dab77fb | 3724 | { |
cdd6c482 | 3725 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
3726 | u64 values[4]; |
3727 | int n = 0; | |
3728 | ||
b5e58793 | 3729 | values[n++] = perf_event_count(event); |
3dab77fb | 3730 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 3731 | values[n++] = enabled + |
cdd6c482 | 3732 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
3733 | } |
3734 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 3735 | values[n++] = running + |
cdd6c482 | 3736 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
3737 | } |
3738 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 3739 | values[n++] = primary_event_id(event); |
3dab77fb | 3740 | |
76369139 | 3741 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
3742 | } |
3743 | ||
3744 | /* | |
cdd6c482 | 3745 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
3746 | */ |
3747 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
3748 | struct perf_event *event, |
3749 | u64 enabled, u64 running) | |
3dab77fb | 3750 | { |
cdd6c482 IM |
3751 | struct perf_event *leader = event->group_leader, *sub; |
3752 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
3753 | u64 values[5]; |
3754 | int n = 0; | |
3755 | ||
3756 | values[n++] = 1 + leader->nr_siblings; | |
3757 | ||
3758 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 3759 | values[n++] = enabled; |
3dab77fb PZ |
3760 | |
3761 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 3762 | values[n++] = running; |
3dab77fb | 3763 | |
cdd6c482 | 3764 | if (leader != event) |
3dab77fb PZ |
3765 | leader->pmu->read(leader); |
3766 | ||
b5e58793 | 3767 | values[n++] = perf_event_count(leader); |
3dab77fb | 3768 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3769 | values[n++] = primary_event_id(leader); |
3dab77fb | 3770 | |
76369139 | 3771 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 3772 | |
65abc865 | 3773 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
3774 | n = 0; |
3775 | ||
cdd6c482 | 3776 | if (sub != event) |
3dab77fb PZ |
3777 | sub->pmu->read(sub); |
3778 | ||
b5e58793 | 3779 | values[n++] = perf_event_count(sub); |
3dab77fb | 3780 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3781 | values[n++] = primary_event_id(sub); |
3dab77fb | 3782 | |
76369139 | 3783 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
3784 | } |
3785 | } | |
3786 | ||
eed01528 SE |
3787 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
3788 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3789 | ||
3dab77fb | 3790 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 3791 | struct perf_event *event) |
3dab77fb | 3792 | { |
c4794295 | 3793 | u64 enabled = 0, running = 0; |
eed01528 SE |
3794 | u64 read_format = event->attr.read_format; |
3795 | ||
3796 | /* | |
3797 | * compute total_time_enabled, total_time_running | |
3798 | * based on snapshot values taken when the event | |
3799 | * was last scheduled in. | |
3800 | * | |
3801 | * we cannot simply called update_context_time() | |
3802 | * because of locking issue as we are called in | |
3803 | * NMI context | |
3804 | */ | |
c4794295 EM |
3805 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
3806 | calc_timer_values(event, &enabled, &running); | |
eed01528 | 3807 | |
cdd6c482 | 3808 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 3809 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 3810 | else |
eed01528 | 3811 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
3812 | } |
3813 | ||
5622f295 MM |
3814 | void perf_output_sample(struct perf_output_handle *handle, |
3815 | struct perf_event_header *header, | |
3816 | struct perf_sample_data *data, | |
cdd6c482 | 3817 | struct perf_event *event) |
5622f295 MM |
3818 | { |
3819 | u64 sample_type = data->type; | |
3820 | ||
3821 | perf_output_put(handle, *header); | |
3822 | ||
3823 | if (sample_type & PERF_SAMPLE_IP) | |
3824 | perf_output_put(handle, data->ip); | |
3825 | ||
3826 | if (sample_type & PERF_SAMPLE_TID) | |
3827 | perf_output_put(handle, data->tid_entry); | |
3828 | ||
3829 | if (sample_type & PERF_SAMPLE_TIME) | |
3830 | perf_output_put(handle, data->time); | |
3831 | ||
3832 | if (sample_type & PERF_SAMPLE_ADDR) | |
3833 | perf_output_put(handle, data->addr); | |
3834 | ||
3835 | if (sample_type & PERF_SAMPLE_ID) | |
3836 | perf_output_put(handle, data->id); | |
3837 | ||
3838 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3839 | perf_output_put(handle, data->stream_id); | |
3840 | ||
3841 | if (sample_type & PERF_SAMPLE_CPU) | |
3842 | perf_output_put(handle, data->cpu_entry); | |
3843 | ||
3844 | if (sample_type & PERF_SAMPLE_PERIOD) | |
3845 | perf_output_put(handle, data->period); | |
3846 | ||
3847 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 3848 | perf_output_read(handle, event); |
5622f295 MM |
3849 | |
3850 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
3851 | if (data->callchain) { | |
3852 | int size = 1; | |
3853 | ||
3854 | if (data->callchain) | |
3855 | size += data->callchain->nr; | |
3856 | ||
3857 | size *= sizeof(u64); | |
3858 | ||
76369139 | 3859 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
3860 | } else { |
3861 | u64 nr = 0; | |
3862 | perf_output_put(handle, nr); | |
3863 | } | |
3864 | } | |
3865 | ||
3866 | if (sample_type & PERF_SAMPLE_RAW) { | |
3867 | if (data->raw) { | |
3868 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
3869 | __output_copy(handle, data->raw->data, |
3870 | data->raw->size); | |
5622f295 MM |
3871 | } else { |
3872 | struct { | |
3873 | u32 size; | |
3874 | u32 data; | |
3875 | } raw = { | |
3876 | .size = sizeof(u32), | |
3877 | .data = 0, | |
3878 | }; | |
3879 | perf_output_put(handle, raw); | |
3880 | } | |
3881 | } | |
a7ac67ea PZ |
3882 | |
3883 | if (!event->attr.watermark) { | |
3884 | int wakeup_events = event->attr.wakeup_events; | |
3885 | ||
3886 | if (wakeup_events) { | |
3887 | struct ring_buffer *rb = handle->rb; | |
3888 | int events = local_inc_return(&rb->events); | |
3889 | ||
3890 | if (events >= wakeup_events) { | |
3891 | local_sub(wakeup_events, &rb->events); | |
3892 | local_inc(&rb->wakeup); | |
3893 | } | |
3894 | } | |
3895 | } | |
5622f295 MM |
3896 | } |
3897 | ||
3898 | void perf_prepare_sample(struct perf_event_header *header, | |
3899 | struct perf_sample_data *data, | |
cdd6c482 | 3900 | struct perf_event *event, |
5622f295 | 3901 | struct pt_regs *regs) |
7b732a75 | 3902 | { |
cdd6c482 | 3903 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 3904 | |
cdd6c482 | 3905 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 3906 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
3907 | |
3908 | header->misc = 0; | |
3909 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 3910 | |
c980d109 | 3911 | __perf_event_header__init_id(header, data, event); |
6844c09d | 3912 | |
c320c7b7 | 3913 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
3914 | data->ip = perf_instruction_pointer(regs); |
3915 | ||
b23f3325 | 3916 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 3917 | int size = 1; |
394ee076 | 3918 | |
5622f295 MM |
3919 | data->callchain = perf_callchain(regs); |
3920 | ||
3921 | if (data->callchain) | |
3922 | size += data->callchain->nr; | |
3923 | ||
3924 | header->size += size * sizeof(u64); | |
394ee076 PZ |
3925 | } |
3926 | ||
3a43ce68 | 3927 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
3928 | int size = sizeof(u32); |
3929 | ||
3930 | if (data->raw) | |
3931 | size += data->raw->size; | |
3932 | else | |
3933 | size += sizeof(u32); | |
3934 | ||
3935 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 3936 | header->size += size; |
7f453c24 | 3937 | } |
5622f295 | 3938 | } |
7f453c24 | 3939 | |
a8b0ca17 | 3940 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
3941 | struct perf_sample_data *data, |
3942 | struct pt_regs *regs) | |
3943 | { | |
3944 | struct perf_output_handle handle; | |
3945 | struct perf_event_header header; | |
689802b2 | 3946 | |
927c7a9e FW |
3947 | /* protect the callchain buffers */ |
3948 | rcu_read_lock(); | |
3949 | ||
cdd6c482 | 3950 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 3951 | |
a7ac67ea | 3952 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 3953 | goto exit; |
0322cd6e | 3954 | |
cdd6c482 | 3955 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 3956 | |
8a057d84 | 3957 | perf_output_end(&handle); |
927c7a9e FW |
3958 | |
3959 | exit: | |
3960 | rcu_read_unlock(); | |
0322cd6e PZ |
3961 | } |
3962 | ||
38b200d6 | 3963 | /* |
cdd6c482 | 3964 | * read event_id |
38b200d6 PZ |
3965 | */ |
3966 | ||
3967 | struct perf_read_event { | |
3968 | struct perf_event_header header; | |
3969 | ||
3970 | u32 pid; | |
3971 | u32 tid; | |
38b200d6 PZ |
3972 | }; |
3973 | ||
3974 | static void | |
cdd6c482 | 3975 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
3976 | struct task_struct *task) |
3977 | { | |
3978 | struct perf_output_handle handle; | |
c980d109 | 3979 | struct perf_sample_data sample; |
dfc65094 | 3980 | struct perf_read_event read_event = { |
38b200d6 | 3981 | .header = { |
cdd6c482 | 3982 | .type = PERF_RECORD_READ, |
38b200d6 | 3983 | .misc = 0, |
c320c7b7 | 3984 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 3985 | }, |
cdd6c482 IM |
3986 | .pid = perf_event_pid(event, task), |
3987 | .tid = perf_event_tid(event, task), | |
38b200d6 | 3988 | }; |
3dab77fb | 3989 | int ret; |
38b200d6 | 3990 | |
c980d109 | 3991 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 3992 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
3993 | if (ret) |
3994 | return; | |
3995 | ||
dfc65094 | 3996 | perf_output_put(&handle, read_event); |
cdd6c482 | 3997 | perf_output_read(&handle, event); |
c980d109 | 3998 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 3999 | |
38b200d6 PZ |
4000 | perf_output_end(&handle); |
4001 | } | |
4002 | ||
60313ebe | 4003 | /* |
9f498cc5 PZ |
4004 | * task tracking -- fork/exit |
4005 | * | |
3af9e859 | 4006 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
60313ebe PZ |
4007 | */ |
4008 | ||
9f498cc5 | 4009 | struct perf_task_event { |
3a80b4a3 | 4010 | struct task_struct *task; |
cdd6c482 | 4011 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4012 | |
4013 | struct { | |
4014 | struct perf_event_header header; | |
4015 | ||
4016 | u32 pid; | |
4017 | u32 ppid; | |
9f498cc5 PZ |
4018 | u32 tid; |
4019 | u32 ptid; | |
393b2ad8 | 4020 | u64 time; |
cdd6c482 | 4021 | } event_id; |
60313ebe PZ |
4022 | }; |
4023 | ||
cdd6c482 | 4024 | static void perf_event_task_output(struct perf_event *event, |
9f498cc5 | 4025 | struct perf_task_event *task_event) |
60313ebe PZ |
4026 | { |
4027 | struct perf_output_handle handle; | |
c980d109 | 4028 | struct perf_sample_data sample; |
9f498cc5 | 4029 | struct task_struct *task = task_event->task; |
c980d109 | 4030 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4031 | |
c980d109 | 4032 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4033 | |
c980d109 | 4034 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4035 | task_event->event_id.header.size); |
ef60777c | 4036 | if (ret) |
c980d109 | 4037 | goto out; |
60313ebe | 4038 | |
cdd6c482 IM |
4039 | task_event->event_id.pid = perf_event_pid(event, task); |
4040 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4041 | |
cdd6c482 IM |
4042 | task_event->event_id.tid = perf_event_tid(event, task); |
4043 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4044 | |
cdd6c482 | 4045 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4046 | |
c980d109 ACM |
4047 | perf_event__output_id_sample(event, &handle, &sample); |
4048 | ||
60313ebe | 4049 | perf_output_end(&handle); |
c980d109 ACM |
4050 | out: |
4051 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4052 | } |
4053 | ||
cdd6c482 | 4054 | static int perf_event_task_match(struct perf_event *event) |
60313ebe | 4055 | { |
6f93d0a7 | 4056 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4057 | return 0; |
4058 | ||
5632ab12 | 4059 | if (!event_filter_match(event)) |
5d27c23d PZ |
4060 | return 0; |
4061 | ||
3af9e859 EM |
4062 | if (event->attr.comm || event->attr.mmap || |
4063 | event->attr.mmap_data || event->attr.task) | |
60313ebe PZ |
4064 | return 1; |
4065 | ||
4066 | return 0; | |
4067 | } | |
4068 | ||
cdd6c482 | 4069 | static void perf_event_task_ctx(struct perf_event_context *ctx, |
9f498cc5 | 4070 | struct perf_task_event *task_event) |
60313ebe | 4071 | { |
cdd6c482 | 4072 | struct perf_event *event; |
60313ebe | 4073 | |
cdd6c482 IM |
4074 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4075 | if (perf_event_task_match(event)) | |
4076 | perf_event_task_output(event, task_event); | |
60313ebe | 4077 | } |
60313ebe PZ |
4078 | } |
4079 | ||
cdd6c482 | 4080 | static void perf_event_task_event(struct perf_task_event *task_event) |
60313ebe PZ |
4081 | { |
4082 | struct perf_cpu_context *cpuctx; | |
8dc85d54 | 4083 | struct perf_event_context *ctx; |
108b02cf | 4084 | struct pmu *pmu; |
8dc85d54 | 4085 | int ctxn; |
60313ebe | 4086 | |
d6ff86cf | 4087 | rcu_read_lock(); |
108b02cf | 4088 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4089 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4090 | if (cpuctx->active_pmu != pmu) |
4091 | goto next; | |
108b02cf | 4092 | perf_event_task_ctx(&cpuctx->ctx, task_event); |
8dc85d54 PZ |
4093 | |
4094 | ctx = task_event->task_ctx; | |
4095 | if (!ctx) { | |
4096 | ctxn = pmu->task_ctx_nr; | |
4097 | if (ctxn < 0) | |
41945f6c | 4098 | goto next; |
8dc85d54 PZ |
4099 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); |
4100 | } | |
4101 | if (ctx) | |
4102 | perf_event_task_ctx(ctx, task_event); | |
41945f6c PZ |
4103 | next: |
4104 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4105 | } |
60313ebe PZ |
4106 | rcu_read_unlock(); |
4107 | } | |
4108 | ||
cdd6c482 IM |
4109 | static void perf_event_task(struct task_struct *task, |
4110 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4111 | int new) |
60313ebe | 4112 | { |
9f498cc5 | 4113 | struct perf_task_event task_event; |
60313ebe | 4114 | |
cdd6c482 IM |
4115 | if (!atomic_read(&nr_comm_events) && |
4116 | !atomic_read(&nr_mmap_events) && | |
4117 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4118 | return; |
4119 | ||
9f498cc5 | 4120 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4121 | .task = task, |
4122 | .task_ctx = task_ctx, | |
cdd6c482 | 4123 | .event_id = { |
60313ebe | 4124 | .header = { |
cdd6c482 | 4125 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4126 | .misc = 0, |
cdd6c482 | 4127 | .size = sizeof(task_event.event_id), |
60313ebe | 4128 | }, |
573402db PZ |
4129 | /* .pid */ |
4130 | /* .ppid */ | |
9f498cc5 PZ |
4131 | /* .tid */ |
4132 | /* .ptid */ | |
6f93d0a7 | 4133 | .time = perf_clock(), |
60313ebe PZ |
4134 | }, |
4135 | }; | |
4136 | ||
cdd6c482 | 4137 | perf_event_task_event(&task_event); |
9f498cc5 PZ |
4138 | } |
4139 | ||
cdd6c482 | 4140 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4141 | { |
cdd6c482 | 4142 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4143 | } |
4144 | ||
8d1b2d93 PZ |
4145 | /* |
4146 | * comm tracking | |
4147 | */ | |
4148 | ||
4149 | struct perf_comm_event { | |
22a4f650 IM |
4150 | struct task_struct *task; |
4151 | char *comm; | |
8d1b2d93 PZ |
4152 | int comm_size; |
4153 | ||
4154 | struct { | |
4155 | struct perf_event_header header; | |
4156 | ||
4157 | u32 pid; | |
4158 | u32 tid; | |
cdd6c482 | 4159 | } event_id; |
8d1b2d93 PZ |
4160 | }; |
4161 | ||
cdd6c482 | 4162 | static void perf_event_comm_output(struct perf_event *event, |
8d1b2d93 PZ |
4163 | struct perf_comm_event *comm_event) |
4164 | { | |
4165 | struct perf_output_handle handle; | |
c980d109 | 4166 | struct perf_sample_data sample; |
cdd6c482 | 4167 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4168 | int ret; |
4169 | ||
4170 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); | |
4171 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4172 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4173 | |
4174 | if (ret) | |
c980d109 | 4175 | goto out; |
8d1b2d93 | 4176 | |
cdd6c482 IM |
4177 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4178 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4179 | |
cdd6c482 | 4180 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4181 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4182 | comm_event->comm_size); |
c980d109 ACM |
4183 | |
4184 | perf_event__output_id_sample(event, &handle, &sample); | |
4185 | ||
8d1b2d93 | 4186 | perf_output_end(&handle); |
c980d109 ACM |
4187 | out: |
4188 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4189 | } |
4190 | ||
cdd6c482 | 4191 | static int perf_event_comm_match(struct perf_event *event) |
8d1b2d93 | 4192 | { |
6f93d0a7 | 4193 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4194 | return 0; |
4195 | ||
5632ab12 | 4196 | if (!event_filter_match(event)) |
5d27c23d PZ |
4197 | return 0; |
4198 | ||
cdd6c482 | 4199 | if (event->attr.comm) |
8d1b2d93 PZ |
4200 | return 1; |
4201 | ||
4202 | return 0; | |
4203 | } | |
4204 | ||
cdd6c482 | 4205 | static void perf_event_comm_ctx(struct perf_event_context *ctx, |
8d1b2d93 PZ |
4206 | struct perf_comm_event *comm_event) |
4207 | { | |
cdd6c482 | 4208 | struct perf_event *event; |
8d1b2d93 | 4209 | |
cdd6c482 IM |
4210 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4211 | if (perf_event_comm_match(event)) | |
4212 | perf_event_comm_output(event, comm_event); | |
8d1b2d93 | 4213 | } |
8d1b2d93 PZ |
4214 | } |
4215 | ||
cdd6c482 | 4216 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 PZ |
4217 | { |
4218 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4219 | struct perf_event_context *ctx; |
413ee3b4 | 4220 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4221 | unsigned int size; |
108b02cf | 4222 | struct pmu *pmu; |
8dc85d54 | 4223 | int ctxn; |
8d1b2d93 | 4224 | |
413ee3b4 | 4225 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4226 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4227 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4228 | |
4229 | comm_event->comm = comm; | |
4230 | comm_event->comm_size = size; | |
4231 | ||
cdd6c482 | 4232 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
f6595f3a | 4233 | rcu_read_lock(); |
108b02cf | 4234 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4235 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4236 | if (cpuctx->active_pmu != pmu) |
4237 | goto next; | |
108b02cf | 4238 | perf_event_comm_ctx(&cpuctx->ctx, comm_event); |
8dc85d54 PZ |
4239 | |
4240 | ctxn = pmu->task_ctx_nr; | |
4241 | if (ctxn < 0) | |
41945f6c | 4242 | goto next; |
8dc85d54 PZ |
4243 | |
4244 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4245 | if (ctx) | |
4246 | perf_event_comm_ctx(ctx, comm_event); | |
41945f6c PZ |
4247 | next: |
4248 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4249 | } |
665c2142 | 4250 | rcu_read_unlock(); |
8d1b2d93 PZ |
4251 | } |
4252 | ||
cdd6c482 | 4253 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4254 | { |
9ee318a7 | 4255 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4256 | struct perf_event_context *ctx; |
4257 | int ctxn; | |
9ee318a7 | 4258 | |
8dc85d54 PZ |
4259 | for_each_task_context_nr(ctxn) { |
4260 | ctx = task->perf_event_ctxp[ctxn]; | |
4261 | if (!ctx) | |
4262 | continue; | |
9ee318a7 | 4263 | |
8dc85d54 PZ |
4264 | perf_event_enable_on_exec(ctx); |
4265 | } | |
9ee318a7 | 4266 | |
cdd6c482 | 4267 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4268 | return; |
a63eaf34 | 4269 | |
9ee318a7 | 4270 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4271 | .task = task, |
573402db PZ |
4272 | /* .comm */ |
4273 | /* .comm_size */ | |
cdd6c482 | 4274 | .event_id = { |
573402db | 4275 | .header = { |
cdd6c482 | 4276 | .type = PERF_RECORD_COMM, |
573402db PZ |
4277 | .misc = 0, |
4278 | /* .size */ | |
4279 | }, | |
4280 | /* .pid */ | |
4281 | /* .tid */ | |
8d1b2d93 PZ |
4282 | }, |
4283 | }; | |
4284 | ||
cdd6c482 | 4285 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
4286 | } |
4287 | ||
0a4a9391 PZ |
4288 | /* |
4289 | * mmap tracking | |
4290 | */ | |
4291 | ||
4292 | struct perf_mmap_event { | |
089dd79d PZ |
4293 | struct vm_area_struct *vma; |
4294 | ||
4295 | const char *file_name; | |
4296 | int file_size; | |
0a4a9391 PZ |
4297 | |
4298 | struct { | |
4299 | struct perf_event_header header; | |
4300 | ||
4301 | u32 pid; | |
4302 | u32 tid; | |
4303 | u64 start; | |
4304 | u64 len; | |
4305 | u64 pgoff; | |
cdd6c482 | 4306 | } event_id; |
0a4a9391 PZ |
4307 | }; |
4308 | ||
cdd6c482 | 4309 | static void perf_event_mmap_output(struct perf_event *event, |
0a4a9391 PZ |
4310 | struct perf_mmap_event *mmap_event) |
4311 | { | |
4312 | struct perf_output_handle handle; | |
c980d109 | 4313 | struct perf_sample_data sample; |
cdd6c482 | 4314 | int size = mmap_event->event_id.header.size; |
c980d109 | 4315 | int ret; |
0a4a9391 | 4316 | |
c980d109 ACM |
4317 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
4318 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4319 | mmap_event->event_id.header.size); |
0a4a9391 | 4320 | if (ret) |
c980d109 | 4321 | goto out; |
0a4a9391 | 4322 | |
cdd6c482 IM |
4323 | mmap_event->event_id.pid = perf_event_pid(event, current); |
4324 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 4325 | |
cdd6c482 | 4326 | perf_output_put(&handle, mmap_event->event_id); |
76369139 | 4327 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 4328 | mmap_event->file_size); |
c980d109 ACM |
4329 | |
4330 | perf_event__output_id_sample(event, &handle, &sample); | |
4331 | ||
78d613eb | 4332 | perf_output_end(&handle); |
c980d109 ACM |
4333 | out: |
4334 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
4335 | } |
4336 | ||
cdd6c482 | 4337 | static int perf_event_mmap_match(struct perf_event *event, |
3af9e859 EM |
4338 | struct perf_mmap_event *mmap_event, |
4339 | int executable) | |
0a4a9391 | 4340 | { |
6f93d0a7 | 4341 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4342 | return 0; |
4343 | ||
5632ab12 | 4344 | if (!event_filter_match(event)) |
5d27c23d PZ |
4345 | return 0; |
4346 | ||
3af9e859 EM |
4347 | if ((!executable && event->attr.mmap_data) || |
4348 | (executable && event->attr.mmap)) | |
0a4a9391 PZ |
4349 | return 1; |
4350 | ||
4351 | return 0; | |
4352 | } | |
4353 | ||
cdd6c482 | 4354 | static void perf_event_mmap_ctx(struct perf_event_context *ctx, |
3af9e859 EM |
4355 | struct perf_mmap_event *mmap_event, |
4356 | int executable) | |
0a4a9391 | 4357 | { |
cdd6c482 | 4358 | struct perf_event *event; |
0a4a9391 | 4359 | |
cdd6c482 | 4360 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
3af9e859 | 4361 | if (perf_event_mmap_match(event, mmap_event, executable)) |
cdd6c482 | 4362 | perf_event_mmap_output(event, mmap_event); |
0a4a9391 | 4363 | } |
0a4a9391 PZ |
4364 | } |
4365 | ||
cdd6c482 | 4366 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 PZ |
4367 | { |
4368 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4369 | struct perf_event_context *ctx; |
089dd79d PZ |
4370 | struct vm_area_struct *vma = mmap_event->vma; |
4371 | struct file *file = vma->vm_file; | |
0a4a9391 PZ |
4372 | unsigned int size; |
4373 | char tmp[16]; | |
4374 | char *buf = NULL; | |
089dd79d | 4375 | const char *name; |
108b02cf | 4376 | struct pmu *pmu; |
8dc85d54 | 4377 | int ctxn; |
0a4a9391 | 4378 | |
413ee3b4 AB |
4379 | memset(tmp, 0, sizeof(tmp)); |
4380 | ||
0a4a9391 | 4381 | if (file) { |
413ee3b4 | 4382 | /* |
76369139 | 4383 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
4384 | * need to add enough zero bytes after the string to handle |
4385 | * the 64bit alignment we do later. | |
4386 | */ | |
4387 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
4388 | if (!buf) { |
4389 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
4390 | goto got_name; | |
4391 | } | |
d3d21c41 | 4392 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
4393 | if (IS_ERR(name)) { |
4394 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
4395 | goto got_name; | |
4396 | } | |
4397 | } else { | |
413ee3b4 AB |
4398 | if (arch_vma_name(mmap_event->vma)) { |
4399 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
4400 | sizeof(tmp)); | |
089dd79d | 4401 | goto got_name; |
413ee3b4 | 4402 | } |
089dd79d PZ |
4403 | |
4404 | if (!vma->vm_mm) { | |
4405 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
4406 | goto got_name; | |
3af9e859 EM |
4407 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
4408 | vma->vm_end >= vma->vm_mm->brk) { | |
4409 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
4410 | goto got_name; | |
4411 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
4412 | vma->vm_end >= vma->vm_mm->start_stack) { | |
4413 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
4414 | goto got_name; | |
089dd79d PZ |
4415 | } |
4416 | ||
0a4a9391 PZ |
4417 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
4418 | goto got_name; | |
4419 | } | |
4420 | ||
4421 | got_name: | |
888fcee0 | 4422 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
4423 | |
4424 | mmap_event->file_name = name; | |
4425 | mmap_event->file_size = size; | |
4426 | ||
cdd6c482 | 4427 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 4428 | |
f6d9dd23 | 4429 | rcu_read_lock(); |
108b02cf | 4430 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4431 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4432 | if (cpuctx->active_pmu != pmu) |
4433 | goto next; | |
108b02cf PZ |
4434 | perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, |
4435 | vma->vm_flags & VM_EXEC); | |
8dc85d54 PZ |
4436 | |
4437 | ctxn = pmu->task_ctx_nr; | |
4438 | if (ctxn < 0) | |
41945f6c | 4439 | goto next; |
8dc85d54 PZ |
4440 | |
4441 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4442 | if (ctx) { | |
4443 | perf_event_mmap_ctx(ctx, mmap_event, | |
4444 | vma->vm_flags & VM_EXEC); | |
4445 | } | |
41945f6c PZ |
4446 | next: |
4447 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4448 | } |
665c2142 PZ |
4449 | rcu_read_unlock(); |
4450 | ||
0a4a9391 PZ |
4451 | kfree(buf); |
4452 | } | |
4453 | ||
3af9e859 | 4454 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 4455 | { |
9ee318a7 PZ |
4456 | struct perf_mmap_event mmap_event; |
4457 | ||
cdd6c482 | 4458 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
4459 | return; |
4460 | ||
4461 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 4462 | .vma = vma, |
573402db PZ |
4463 | /* .file_name */ |
4464 | /* .file_size */ | |
cdd6c482 | 4465 | .event_id = { |
573402db | 4466 | .header = { |
cdd6c482 | 4467 | .type = PERF_RECORD_MMAP, |
39447b38 | 4468 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
4469 | /* .size */ |
4470 | }, | |
4471 | /* .pid */ | |
4472 | /* .tid */ | |
089dd79d PZ |
4473 | .start = vma->vm_start, |
4474 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 4475 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 PZ |
4476 | }, |
4477 | }; | |
4478 | ||
cdd6c482 | 4479 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
4480 | } |
4481 | ||
a78ac325 PZ |
4482 | /* |
4483 | * IRQ throttle logging | |
4484 | */ | |
4485 | ||
cdd6c482 | 4486 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
4487 | { |
4488 | struct perf_output_handle handle; | |
c980d109 | 4489 | struct perf_sample_data sample; |
a78ac325 PZ |
4490 | int ret; |
4491 | ||
4492 | struct { | |
4493 | struct perf_event_header header; | |
4494 | u64 time; | |
cca3f454 | 4495 | u64 id; |
7f453c24 | 4496 | u64 stream_id; |
a78ac325 PZ |
4497 | } throttle_event = { |
4498 | .header = { | |
cdd6c482 | 4499 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
4500 | .misc = 0, |
4501 | .size = sizeof(throttle_event), | |
4502 | }, | |
def0a9b2 | 4503 | .time = perf_clock(), |
cdd6c482 IM |
4504 | .id = primary_event_id(event), |
4505 | .stream_id = event->id, | |
a78ac325 PZ |
4506 | }; |
4507 | ||
966ee4d6 | 4508 | if (enable) |
cdd6c482 | 4509 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 4510 | |
c980d109 ACM |
4511 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
4512 | ||
4513 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4514 | throttle_event.header.size); |
a78ac325 PZ |
4515 | if (ret) |
4516 | return; | |
4517 | ||
4518 | perf_output_put(&handle, throttle_event); | |
c980d109 | 4519 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
4520 | perf_output_end(&handle); |
4521 | } | |
4522 | ||
f6c7d5fe | 4523 | /* |
cdd6c482 | 4524 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
4525 | */ |
4526 | ||
a8b0ca17 | 4527 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4528 | int throttle, struct perf_sample_data *data, |
4529 | struct pt_regs *regs) | |
f6c7d5fe | 4530 | { |
cdd6c482 IM |
4531 | int events = atomic_read(&event->event_limit); |
4532 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 4533 | u64 seq; |
79f14641 PZ |
4534 | int ret = 0; |
4535 | ||
96398826 PZ |
4536 | /* |
4537 | * Non-sampling counters might still use the PMI to fold short | |
4538 | * hardware counters, ignore those. | |
4539 | */ | |
4540 | if (unlikely(!is_sampling_event(event))) | |
4541 | return 0; | |
4542 | ||
e050e3f0 SE |
4543 | seq = __this_cpu_read(perf_throttled_seq); |
4544 | if (seq != hwc->interrupts_seq) { | |
4545 | hwc->interrupts_seq = seq; | |
4546 | hwc->interrupts = 1; | |
4547 | } else { | |
4548 | hwc->interrupts++; | |
4549 | if (unlikely(throttle | |
4550 | && hwc->interrupts >= max_samples_per_tick)) { | |
4551 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
4552 | hwc->interrupts = MAX_INTERRUPTS; |
4553 | perf_log_throttle(event, 0); | |
a78ac325 PZ |
4554 | ret = 1; |
4555 | } | |
e050e3f0 | 4556 | } |
60db5e09 | 4557 | |
cdd6c482 | 4558 | if (event->attr.freq) { |
def0a9b2 | 4559 | u64 now = perf_clock(); |
abd50713 | 4560 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 4561 | |
abd50713 | 4562 | hwc->freq_time_stamp = now; |
bd2b5b12 | 4563 | |
abd50713 PZ |
4564 | if (delta > 0 && delta < 2*TICK_NSEC) |
4565 | perf_adjust_period(event, delta, hwc->last_period); | |
bd2b5b12 PZ |
4566 | } |
4567 | ||
2023b359 PZ |
4568 | /* |
4569 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 4570 | * events |
2023b359 PZ |
4571 | */ |
4572 | ||
cdd6c482 IM |
4573 | event->pending_kill = POLL_IN; |
4574 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 4575 | ret = 1; |
cdd6c482 | 4576 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
4577 | event->pending_disable = 1; |
4578 | irq_work_queue(&event->pending); | |
79f14641 PZ |
4579 | } |
4580 | ||
453f19ee | 4581 | if (event->overflow_handler) |
a8b0ca17 | 4582 | event->overflow_handler(event, data, regs); |
453f19ee | 4583 | else |
a8b0ca17 | 4584 | perf_event_output(event, data, regs); |
453f19ee | 4585 | |
f506b3dc | 4586 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
4587 | event->pending_wakeup = 1; |
4588 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
4589 | } |
4590 | ||
79f14641 | 4591 | return ret; |
f6c7d5fe PZ |
4592 | } |
4593 | ||
a8b0ca17 | 4594 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4595 | struct perf_sample_data *data, |
4596 | struct pt_regs *regs) | |
850bc73f | 4597 | { |
a8b0ca17 | 4598 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
4599 | } |
4600 | ||
15dbf27c | 4601 | /* |
cdd6c482 | 4602 | * Generic software event infrastructure |
15dbf27c PZ |
4603 | */ |
4604 | ||
b28ab83c PZ |
4605 | struct swevent_htable { |
4606 | struct swevent_hlist *swevent_hlist; | |
4607 | struct mutex hlist_mutex; | |
4608 | int hlist_refcount; | |
4609 | ||
4610 | /* Recursion avoidance in each contexts */ | |
4611 | int recursion[PERF_NR_CONTEXTS]; | |
4612 | }; | |
4613 | ||
4614 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
4615 | ||
7b4b6658 | 4616 | /* |
cdd6c482 IM |
4617 | * We directly increment event->count and keep a second value in |
4618 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
4619 | * is kept in the range [-sample_period, 0] so that we can use the |
4620 | * sign as trigger. | |
4621 | */ | |
4622 | ||
cdd6c482 | 4623 | static u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 4624 | { |
cdd6c482 | 4625 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
4626 | u64 period = hwc->last_period; |
4627 | u64 nr, offset; | |
4628 | s64 old, val; | |
4629 | ||
4630 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
4631 | |
4632 | again: | |
e7850595 | 4633 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
4634 | if (val < 0) |
4635 | return 0; | |
15dbf27c | 4636 | |
7b4b6658 PZ |
4637 | nr = div64_u64(period + val, period); |
4638 | offset = nr * period; | |
4639 | val -= offset; | |
e7850595 | 4640 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 4641 | goto again; |
15dbf27c | 4642 | |
7b4b6658 | 4643 | return nr; |
15dbf27c PZ |
4644 | } |
4645 | ||
0cff784a | 4646 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 4647 | struct perf_sample_data *data, |
5622f295 | 4648 | struct pt_regs *regs) |
15dbf27c | 4649 | { |
cdd6c482 | 4650 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 4651 | int throttle = 0; |
15dbf27c | 4652 | |
0cff784a PZ |
4653 | if (!overflow) |
4654 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 4655 | |
7b4b6658 PZ |
4656 | if (hwc->interrupts == MAX_INTERRUPTS) |
4657 | return; | |
15dbf27c | 4658 | |
7b4b6658 | 4659 | for (; overflow; overflow--) { |
a8b0ca17 | 4660 | if (__perf_event_overflow(event, throttle, |
5622f295 | 4661 | data, regs)) { |
7b4b6658 PZ |
4662 | /* |
4663 | * We inhibit the overflow from happening when | |
4664 | * hwc->interrupts == MAX_INTERRUPTS. | |
4665 | */ | |
4666 | break; | |
4667 | } | |
cf450a73 | 4668 | throttle = 1; |
7b4b6658 | 4669 | } |
15dbf27c PZ |
4670 | } |
4671 | ||
a4eaf7f1 | 4672 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 4673 | struct perf_sample_data *data, |
5622f295 | 4674 | struct pt_regs *regs) |
7b4b6658 | 4675 | { |
cdd6c482 | 4676 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 4677 | |
e7850595 | 4678 | local64_add(nr, &event->count); |
d6d020e9 | 4679 | |
0cff784a PZ |
4680 | if (!regs) |
4681 | return; | |
4682 | ||
6c7e550f | 4683 | if (!is_sampling_event(event)) |
7b4b6658 | 4684 | return; |
d6d020e9 | 4685 | |
5d81e5cf AV |
4686 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
4687 | data->period = nr; | |
4688 | return perf_swevent_overflow(event, 1, data, regs); | |
4689 | } else | |
4690 | data->period = event->hw.last_period; | |
4691 | ||
0cff784a | 4692 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 4693 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 4694 | |
e7850595 | 4695 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 4696 | return; |
df1a132b | 4697 | |
a8b0ca17 | 4698 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
4699 | } |
4700 | ||
f5ffe02e FW |
4701 | static int perf_exclude_event(struct perf_event *event, |
4702 | struct pt_regs *regs) | |
4703 | { | |
a4eaf7f1 | 4704 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 4705 | return 1; |
a4eaf7f1 | 4706 | |
f5ffe02e FW |
4707 | if (regs) { |
4708 | if (event->attr.exclude_user && user_mode(regs)) | |
4709 | return 1; | |
4710 | ||
4711 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
4712 | return 1; | |
4713 | } | |
4714 | ||
4715 | return 0; | |
4716 | } | |
4717 | ||
cdd6c482 | 4718 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 4719 | enum perf_type_id type, |
6fb2915d LZ |
4720 | u32 event_id, |
4721 | struct perf_sample_data *data, | |
4722 | struct pt_regs *regs) | |
15dbf27c | 4723 | { |
cdd6c482 | 4724 | if (event->attr.type != type) |
a21ca2ca | 4725 | return 0; |
f5ffe02e | 4726 | |
cdd6c482 | 4727 | if (event->attr.config != event_id) |
15dbf27c PZ |
4728 | return 0; |
4729 | ||
f5ffe02e FW |
4730 | if (perf_exclude_event(event, regs)) |
4731 | return 0; | |
15dbf27c PZ |
4732 | |
4733 | return 1; | |
4734 | } | |
4735 | ||
76e1d904 FW |
4736 | static inline u64 swevent_hash(u64 type, u32 event_id) |
4737 | { | |
4738 | u64 val = event_id | (type << 32); | |
4739 | ||
4740 | return hash_64(val, SWEVENT_HLIST_BITS); | |
4741 | } | |
4742 | ||
49f135ed FW |
4743 | static inline struct hlist_head * |
4744 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 4745 | { |
49f135ed FW |
4746 | u64 hash = swevent_hash(type, event_id); |
4747 | ||
4748 | return &hlist->heads[hash]; | |
4749 | } | |
76e1d904 | 4750 | |
49f135ed FW |
4751 | /* For the read side: events when they trigger */ |
4752 | static inline struct hlist_head * | |
b28ab83c | 4753 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
4754 | { |
4755 | struct swevent_hlist *hlist; | |
76e1d904 | 4756 | |
b28ab83c | 4757 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
4758 | if (!hlist) |
4759 | return NULL; | |
4760 | ||
49f135ed FW |
4761 | return __find_swevent_head(hlist, type, event_id); |
4762 | } | |
4763 | ||
4764 | /* For the event head insertion and removal in the hlist */ | |
4765 | static inline struct hlist_head * | |
b28ab83c | 4766 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
4767 | { |
4768 | struct swevent_hlist *hlist; | |
4769 | u32 event_id = event->attr.config; | |
4770 | u64 type = event->attr.type; | |
4771 | ||
4772 | /* | |
4773 | * Event scheduling is always serialized against hlist allocation | |
4774 | * and release. Which makes the protected version suitable here. | |
4775 | * The context lock guarantees that. | |
4776 | */ | |
b28ab83c | 4777 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
4778 | lockdep_is_held(&event->ctx->lock)); |
4779 | if (!hlist) | |
4780 | return NULL; | |
4781 | ||
4782 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
4783 | } |
4784 | ||
4785 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 4786 | u64 nr, |
76e1d904 FW |
4787 | struct perf_sample_data *data, |
4788 | struct pt_regs *regs) | |
15dbf27c | 4789 | { |
b28ab83c | 4790 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 4791 | struct perf_event *event; |
76e1d904 FW |
4792 | struct hlist_node *node; |
4793 | struct hlist_head *head; | |
15dbf27c | 4794 | |
76e1d904 | 4795 | rcu_read_lock(); |
b28ab83c | 4796 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
4797 | if (!head) |
4798 | goto end; | |
4799 | ||
4800 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { | |
6fb2915d | 4801 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 4802 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 4803 | } |
76e1d904 FW |
4804 | end: |
4805 | rcu_read_unlock(); | |
15dbf27c PZ |
4806 | } |
4807 | ||
4ed7c92d | 4808 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 4809 | { |
b28ab83c | 4810 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 4811 | |
b28ab83c | 4812 | return get_recursion_context(swhash->recursion); |
96f6d444 | 4813 | } |
645e8cc0 | 4814 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 4815 | |
fa9f90be | 4816 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 4817 | { |
b28ab83c | 4818 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 4819 | |
b28ab83c | 4820 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 4821 | } |
15dbf27c | 4822 | |
a8b0ca17 | 4823 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 4824 | { |
a4234bfc | 4825 | struct perf_sample_data data; |
4ed7c92d PZ |
4826 | int rctx; |
4827 | ||
1c024eca | 4828 | preempt_disable_notrace(); |
4ed7c92d PZ |
4829 | rctx = perf_swevent_get_recursion_context(); |
4830 | if (rctx < 0) | |
4831 | return; | |
a4234bfc | 4832 | |
dc1d628a | 4833 | perf_sample_data_init(&data, addr); |
92bf309a | 4834 | |
a8b0ca17 | 4835 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
4836 | |
4837 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 4838 | preempt_enable_notrace(); |
b8e83514 PZ |
4839 | } |
4840 | ||
cdd6c482 | 4841 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 4842 | { |
15dbf27c PZ |
4843 | } |
4844 | ||
a4eaf7f1 | 4845 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 4846 | { |
b28ab83c | 4847 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 4848 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
4849 | struct hlist_head *head; |
4850 | ||
6c7e550f | 4851 | if (is_sampling_event(event)) { |
7b4b6658 | 4852 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 4853 | perf_swevent_set_period(event); |
7b4b6658 | 4854 | } |
76e1d904 | 4855 | |
a4eaf7f1 PZ |
4856 | hwc->state = !(flags & PERF_EF_START); |
4857 | ||
b28ab83c | 4858 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
4859 | if (WARN_ON_ONCE(!head)) |
4860 | return -EINVAL; | |
4861 | ||
4862 | hlist_add_head_rcu(&event->hlist_entry, head); | |
4863 | ||
15dbf27c PZ |
4864 | return 0; |
4865 | } | |
4866 | ||
a4eaf7f1 | 4867 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 4868 | { |
76e1d904 | 4869 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
4870 | } |
4871 | ||
a4eaf7f1 | 4872 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 4873 | { |
a4eaf7f1 | 4874 | event->hw.state = 0; |
d6d020e9 | 4875 | } |
aa9c4c0f | 4876 | |
a4eaf7f1 | 4877 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 4878 | { |
a4eaf7f1 | 4879 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
4880 | } |
4881 | ||
49f135ed FW |
4882 | /* Deref the hlist from the update side */ |
4883 | static inline struct swevent_hlist * | |
b28ab83c | 4884 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 4885 | { |
b28ab83c PZ |
4886 | return rcu_dereference_protected(swhash->swevent_hlist, |
4887 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
4888 | } |
4889 | ||
b28ab83c | 4890 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 4891 | { |
b28ab83c | 4892 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 4893 | |
49f135ed | 4894 | if (!hlist) |
76e1d904 FW |
4895 | return; |
4896 | ||
b28ab83c | 4897 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 4898 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
4899 | } |
4900 | ||
4901 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
4902 | { | |
b28ab83c | 4903 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 4904 | |
b28ab83c | 4905 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 4906 | |
b28ab83c PZ |
4907 | if (!--swhash->hlist_refcount) |
4908 | swevent_hlist_release(swhash); | |
76e1d904 | 4909 | |
b28ab83c | 4910 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
4911 | } |
4912 | ||
4913 | static void swevent_hlist_put(struct perf_event *event) | |
4914 | { | |
4915 | int cpu; | |
4916 | ||
4917 | if (event->cpu != -1) { | |
4918 | swevent_hlist_put_cpu(event, event->cpu); | |
4919 | return; | |
4920 | } | |
4921 | ||
4922 | for_each_possible_cpu(cpu) | |
4923 | swevent_hlist_put_cpu(event, cpu); | |
4924 | } | |
4925 | ||
4926 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
4927 | { | |
b28ab83c | 4928 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
4929 | int err = 0; |
4930 | ||
b28ab83c | 4931 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 4932 | |
b28ab83c | 4933 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
4934 | struct swevent_hlist *hlist; |
4935 | ||
4936 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
4937 | if (!hlist) { | |
4938 | err = -ENOMEM; | |
4939 | goto exit; | |
4940 | } | |
b28ab83c | 4941 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 4942 | } |
b28ab83c | 4943 | swhash->hlist_refcount++; |
9ed6060d | 4944 | exit: |
b28ab83c | 4945 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
4946 | |
4947 | return err; | |
4948 | } | |
4949 | ||
4950 | static int swevent_hlist_get(struct perf_event *event) | |
4951 | { | |
4952 | int err; | |
4953 | int cpu, failed_cpu; | |
4954 | ||
4955 | if (event->cpu != -1) | |
4956 | return swevent_hlist_get_cpu(event, event->cpu); | |
4957 | ||
4958 | get_online_cpus(); | |
4959 | for_each_possible_cpu(cpu) { | |
4960 | err = swevent_hlist_get_cpu(event, cpu); | |
4961 | if (err) { | |
4962 | failed_cpu = cpu; | |
4963 | goto fail; | |
4964 | } | |
4965 | } | |
4966 | put_online_cpus(); | |
4967 | ||
4968 | return 0; | |
9ed6060d | 4969 | fail: |
76e1d904 FW |
4970 | for_each_possible_cpu(cpu) { |
4971 | if (cpu == failed_cpu) | |
4972 | break; | |
4973 | swevent_hlist_put_cpu(event, cpu); | |
4974 | } | |
4975 | ||
4976 | put_online_cpus(); | |
4977 | return err; | |
4978 | } | |
4979 | ||
d430d3d7 | 4980 | struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 4981 | |
b0a873eb PZ |
4982 | static void sw_perf_event_destroy(struct perf_event *event) |
4983 | { | |
4984 | u64 event_id = event->attr.config; | |
95476b64 | 4985 | |
b0a873eb PZ |
4986 | WARN_ON(event->parent); |
4987 | ||
7e54a5a0 | 4988 | jump_label_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
4989 | swevent_hlist_put(event); |
4990 | } | |
4991 | ||
4992 | static int perf_swevent_init(struct perf_event *event) | |
4993 | { | |
4994 | int event_id = event->attr.config; | |
4995 | ||
4996 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
4997 | return -ENOENT; | |
4998 | ||
4999 | switch (event_id) { | |
5000 | case PERF_COUNT_SW_CPU_CLOCK: | |
5001 | case PERF_COUNT_SW_TASK_CLOCK: | |
5002 | return -ENOENT; | |
5003 | ||
5004 | default: | |
5005 | break; | |
5006 | } | |
5007 | ||
ce677831 | 5008 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5009 | return -ENOENT; |
5010 | ||
5011 | if (!event->parent) { | |
5012 | int err; | |
5013 | ||
5014 | err = swevent_hlist_get(event); | |
5015 | if (err) | |
5016 | return err; | |
5017 | ||
7e54a5a0 | 5018 | jump_label_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5019 | event->destroy = sw_perf_event_destroy; |
5020 | } | |
5021 | ||
5022 | return 0; | |
5023 | } | |
5024 | ||
5025 | static struct pmu perf_swevent = { | |
89a1e187 | 5026 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5027 | |
b0a873eb | 5028 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5029 | .add = perf_swevent_add, |
5030 | .del = perf_swevent_del, | |
5031 | .start = perf_swevent_start, | |
5032 | .stop = perf_swevent_stop, | |
1c024eca | 5033 | .read = perf_swevent_read, |
1c024eca PZ |
5034 | }; |
5035 | ||
b0a873eb PZ |
5036 | #ifdef CONFIG_EVENT_TRACING |
5037 | ||
1c024eca PZ |
5038 | static int perf_tp_filter_match(struct perf_event *event, |
5039 | struct perf_sample_data *data) | |
5040 | { | |
5041 | void *record = data->raw->data; | |
5042 | ||
5043 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5044 | return 1; | |
5045 | return 0; | |
5046 | } | |
5047 | ||
5048 | static int perf_tp_event_match(struct perf_event *event, | |
5049 | struct perf_sample_data *data, | |
5050 | struct pt_regs *regs) | |
5051 | { | |
a0f7d0f7 FW |
5052 | if (event->hw.state & PERF_HES_STOPPED) |
5053 | return 0; | |
580d607c PZ |
5054 | /* |
5055 | * All tracepoints are from kernel-space. | |
5056 | */ | |
5057 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5058 | return 0; |
5059 | ||
5060 | if (!perf_tp_filter_match(event, data)) | |
5061 | return 0; | |
5062 | ||
5063 | return 1; | |
5064 | } | |
5065 | ||
5066 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
ecc55f84 | 5067 | struct pt_regs *regs, struct hlist_head *head, int rctx) |
95476b64 FW |
5068 | { |
5069 | struct perf_sample_data data; | |
1c024eca PZ |
5070 | struct perf_event *event; |
5071 | struct hlist_node *node; | |
5072 | ||
95476b64 FW |
5073 | struct perf_raw_record raw = { |
5074 | .size = entry_size, | |
5075 | .data = record, | |
5076 | }; | |
5077 | ||
5078 | perf_sample_data_init(&data, addr); | |
5079 | data.raw = &raw; | |
5080 | ||
1c024eca PZ |
5081 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { |
5082 | if (perf_tp_event_match(event, &data, regs)) | |
a8b0ca17 | 5083 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5084 | } |
ecc55f84 PZ |
5085 | |
5086 | perf_swevent_put_recursion_context(rctx); | |
95476b64 FW |
5087 | } |
5088 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5089 | ||
cdd6c482 | 5090 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5091 | { |
1c024eca | 5092 | perf_trace_destroy(event); |
e077df4f PZ |
5093 | } |
5094 | ||
b0a873eb | 5095 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5096 | { |
76e1d904 FW |
5097 | int err; |
5098 | ||
b0a873eb PZ |
5099 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5100 | return -ENOENT; | |
5101 | ||
1c024eca PZ |
5102 | err = perf_trace_init(event); |
5103 | if (err) | |
b0a873eb | 5104 | return err; |
e077df4f | 5105 | |
cdd6c482 | 5106 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5107 | |
b0a873eb PZ |
5108 | return 0; |
5109 | } | |
5110 | ||
5111 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5112 | .task_ctx_nr = perf_sw_context, |
5113 | ||
b0a873eb | 5114 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5115 | .add = perf_trace_add, |
5116 | .del = perf_trace_del, | |
5117 | .start = perf_swevent_start, | |
5118 | .stop = perf_swevent_stop, | |
b0a873eb | 5119 | .read = perf_swevent_read, |
b0a873eb PZ |
5120 | }; |
5121 | ||
5122 | static inline void perf_tp_register(void) | |
5123 | { | |
2e80a82a | 5124 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5125 | } |
6fb2915d LZ |
5126 | |
5127 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5128 | { | |
5129 | char *filter_str; | |
5130 | int ret; | |
5131 | ||
5132 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5133 | return -EINVAL; | |
5134 | ||
5135 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5136 | if (IS_ERR(filter_str)) | |
5137 | return PTR_ERR(filter_str); | |
5138 | ||
5139 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5140 | ||
5141 | kfree(filter_str); | |
5142 | return ret; | |
5143 | } | |
5144 | ||
5145 | static void perf_event_free_filter(struct perf_event *event) | |
5146 | { | |
5147 | ftrace_profile_free_filter(event); | |
5148 | } | |
5149 | ||
e077df4f | 5150 | #else |
6fb2915d | 5151 | |
b0a873eb | 5152 | static inline void perf_tp_register(void) |
e077df4f | 5153 | { |
e077df4f | 5154 | } |
6fb2915d LZ |
5155 | |
5156 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5157 | { | |
5158 | return -ENOENT; | |
5159 | } | |
5160 | ||
5161 | static void perf_event_free_filter(struct perf_event *event) | |
5162 | { | |
5163 | } | |
5164 | ||
07b139c8 | 5165 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5166 | |
24f1e32c | 5167 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5168 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5169 | { |
f5ffe02e FW |
5170 | struct perf_sample_data sample; |
5171 | struct pt_regs *regs = data; | |
5172 | ||
dc1d628a | 5173 | perf_sample_data_init(&sample, bp->attr.bp_addr); |
f5ffe02e | 5174 | |
a4eaf7f1 | 5175 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5176 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5177 | } |
5178 | #endif | |
5179 | ||
b0a873eb PZ |
5180 | /* |
5181 | * hrtimer based swevent callback | |
5182 | */ | |
f29ac756 | 5183 | |
b0a873eb | 5184 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5185 | { |
b0a873eb PZ |
5186 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5187 | struct perf_sample_data data; | |
5188 | struct pt_regs *regs; | |
5189 | struct perf_event *event; | |
5190 | u64 period; | |
f29ac756 | 5191 | |
b0a873eb | 5192 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5193 | |
5194 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5195 | return HRTIMER_NORESTART; | |
5196 | ||
b0a873eb | 5197 | event->pmu->read(event); |
f344011c | 5198 | |
b0a873eb PZ |
5199 | perf_sample_data_init(&data, 0); |
5200 | data.period = event->hw.last_period; | |
5201 | regs = get_irq_regs(); | |
5202 | ||
5203 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5204 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
a8b0ca17 | 5205 | if (perf_event_overflow(event, &data, regs)) |
b0a873eb PZ |
5206 | ret = HRTIMER_NORESTART; |
5207 | } | |
24f1e32c | 5208 | |
b0a873eb PZ |
5209 | period = max_t(u64, 10000, event->hw.sample_period); |
5210 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5211 | |
b0a873eb | 5212 | return ret; |
f29ac756 PZ |
5213 | } |
5214 | ||
b0a873eb | 5215 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5216 | { |
b0a873eb | 5217 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5218 | s64 period; |
5219 | ||
5220 | if (!is_sampling_event(event)) | |
5221 | return; | |
f5ffe02e | 5222 | |
5d508e82 FBH |
5223 | period = local64_read(&hwc->period_left); |
5224 | if (period) { | |
5225 | if (period < 0) | |
5226 | period = 10000; | |
fa407f35 | 5227 | |
5d508e82 FBH |
5228 | local64_set(&hwc->period_left, 0); |
5229 | } else { | |
5230 | period = max_t(u64, 10000, hwc->sample_period); | |
5231 | } | |
5232 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 5233 | ns_to_ktime(period), 0, |
b5ab4cd5 | 5234 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 5235 | } |
b0a873eb PZ |
5236 | |
5237 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 5238 | { |
b0a873eb PZ |
5239 | struct hw_perf_event *hwc = &event->hw; |
5240 | ||
6c7e550f | 5241 | if (is_sampling_event(event)) { |
b0a873eb | 5242 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 5243 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
5244 | |
5245 | hrtimer_cancel(&hwc->hrtimer); | |
5246 | } | |
24f1e32c FW |
5247 | } |
5248 | ||
ba3dd36c PZ |
5249 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5250 | { | |
5251 | struct hw_perf_event *hwc = &event->hw; | |
5252 | ||
5253 | if (!is_sampling_event(event)) | |
5254 | return; | |
5255 | ||
5256 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5257 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
5258 | ||
5259 | /* | |
5260 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
5261 | * mapping and avoid the whole period adjust feedback stuff. | |
5262 | */ | |
5263 | if (event->attr.freq) { | |
5264 | long freq = event->attr.sample_freq; | |
5265 | ||
5266 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
5267 | hwc->sample_period = event->attr.sample_period; | |
5268 | local64_set(&hwc->period_left, hwc->sample_period); | |
5269 | event->attr.freq = 0; | |
5270 | } | |
5271 | } | |
5272 | ||
b0a873eb PZ |
5273 | /* |
5274 | * Software event: cpu wall time clock | |
5275 | */ | |
5276 | ||
5277 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 5278 | { |
b0a873eb PZ |
5279 | s64 prev; |
5280 | u64 now; | |
5281 | ||
a4eaf7f1 | 5282 | now = local_clock(); |
b0a873eb PZ |
5283 | prev = local64_xchg(&event->hw.prev_count, now); |
5284 | local64_add(now - prev, &event->count); | |
24f1e32c | 5285 | } |
24f1e32c | 5286 | |
a4eaf7f1 | 5287 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5288 | { |
a4eaf7f1 | 5289 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 5290 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5291 | } |
5292 | ||
a4eaf7f1 | 5293 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 5294 | { |
b0a873eb PZ |
5295 | perf_swevent_cancel_hrtimer(event); |
5296 | cpu_clock_event_update(event); | |
5297 | } | |
f29ac756 | 5298 | |
a4eaf7f1 PZ |
5299 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
5300 | { | |
5301 | if (flags & PERF_EF_START) | |
5302 | cpu_clock_event_start(event, flags); | |
5303 | ||
5304 | return 0; | |
5305 | } | |
5306 | ||
5307 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
5308 | { | |
5309 | cpu_clock_event_stop(event, flags); | |
5310 | } | |
5311 | ||
b0a873eb PZ |
5312 | static void cpu_clock_event_read(struct perf_event *event) |
5313 | { | |
5314 | cpu_clock_event_update(event); | |
5315 | } | |
f344011c | 5316 | |
b0a873eb PZ |
5317 | static int cpu_clock_event_init(struct perf_event *event) |
5318 | { | |
5319 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5320 | return -ENOENT; | |
5321 | ||
5322 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
5323 | return -ENOENT; | |
5324 | ||
ba3dd36c PZ |
5325 | perf_swevent_init_hrtimer(event); |
5326 | ||
b0a873eb | 5327 | return 0; |
f29ac756 PZ |
5328 | } |
5329 | ||
b0a873eb | 5330 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
5331 | .task_ctx_nr = perf_sw_context, |
5332 | ||
b0a873eb | 5333 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
5334 | .add = cpu_clock_event_add, |
5335 | .del = cpu_clock_event_del, | |
5336 | .start = cpu_clock_event_start, | |
5337 | .stop = cpu_clock_event_stop, | |
b0a873eb PZ |
5338 | .read = cpu_clock_event_read, |
5339 | }; | |
5340 | ||
5341 | /* | |
5342 | * Software event: task time clock | |
5343 | */ | |
5344 | ||
5345 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 5346 | { |
b0a873eb PZ |
5347 | u64 prev; |
5348 | s64 delta; | |
5c92d124 | 5349 | |
b0a873eb PZ |
5350 | prev = local64_xchg(&event->hw.prev_count, now); |
5351 | delta = now - prev; | |
5352 | local64_add(delta, &event->count); | |
5353 | } | |
5c92d124 | 5354 | |
a4eaf7f1 | 5355 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5356 | { |
a4eaf7f1 | 5357 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 5358 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5359 | } |
5360 | ||
a4eaf7f1 | 5361 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
5362 | { |
5363 | perf_swevent_cancel_hrtimer(event); | |
5364 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
5365 | } |
5366 | ||
5367 | static int task_clock_event_add(struct perf_event *event, int flags) | |
5368 | { | |
5369 | if (flags & PERF_EF_START) | |
5370 | task_clock_event_start(event, flags); | |
b0a873eb | 5371 | |
a4eaf7f1 PZ |
5372 | return 0; |
5373 | } | |
5374 | ||
5375 | static void task_clock_event_del(struct perf_event *event, int flags) | |
5376 | { | |
5377 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
5378 | } |
5379 | ||
5380 | static void task_clock_event_read(struct perf_event *event) | |
5381 | { | |
768a06e2 PZ |
5382 | u64 now = perf_clock(); |
5383 | u64 delta = now - event->ctx->timestamp; | |
5384 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
5385 | |
5386 | task_clock_event_update(event, time); | |
5387 | } | |
5388 | ||
5389 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 5390 | { |
b0a873eb PZ |
5391 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5392 | return -ENOENT; | |
5393 | ||
5394 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
5395 | return -ENOENT; | |
5396 | ||
ba3dd36c PZ |
5397 | perf_swevent_init_hrtimer(event); |
5398 | ||
b0a873eb | 5399 | return 0; |
6fb2915d LZ |
5400 | } |
5401 | ||
b0a873eb | 5402 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
5403 | .task_ctx_nr = perf_sw_context, |
5404 | ||
b0a873eb | 5405 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
5406 | .add = task_clock_event_add, |
5407 | .del = task_clock_event_del, | |
5408 | .start = task_clock_event_start, | |
5409 | .stop = task_clock_event_stop, | |
b0a873eb PZ |
5410 | .read = task_clock_event_read, |
5411 | }; | |
6fb2915d | 5412 | |
ad5133b7 | 5413 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 5414 | { |
e077df4f | 5415 | } |
6fb2915d | 5416 | |
ad5133b7 | 5417 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 5418 | { |
ad5133b7 | 5419 | return 0; |
6fb2915d LZ |
5420 | } |
5421 | ||
ad5133b7 | 5422 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 5423 | { |
ad5133b7 | 5424 | perf_pmu_disable(pmu); |
6fb2915d LZ |
5425 | } |
5426 | ||
ad5133b7 PZ |
5427 | static int perf_pmu_commit_txn(struct pmu *pmu) |
5428 | { | |
5429 | perf_pmu_enable(pmu); | |
5430 | return 0; | |
5431 | } | |
e077df4f | 5432 | |
ad5133b7 | 5433 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 5434 | { |
ad5133b7 | 5435 | perf_pmu_enable(pmu); |
24f1e32c FW |
5436 | } |
5437 | ||
8dc85d54 PZ |
5438 | /* |
5439 | * Ensures all contexts with the same task_ctx_nr have the same | |
5440 | * pmu_cpu_context too. | |
5441 | */ | |
5442 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 5443 | { |
8dc85d54 | 5444 | struct pmu *pmu; |
b326e956 | 5445 | |
8dc85d54 PZ |
5446 | if (ctxn < 0) |
5447 | return NULL; | |
24f1e32c | 5448 | |
8dc85d54 PZ |
5449 | list_for_each_entry(pmu, &pmus, entry) { |
5450 | if (pmu->task_ctx_nr == ctxn) | |
5451 | return pmu->pmu_cpu_context; | |
5452 | } | |
24f1e32c | 5453 | |
8dc85d54 | 5454 | return NULL; |
24f1e32c FW |
5455 | } |
5456 | ||
51676957 | 5457 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 5458 | { |
51676957 PZ |
5459 | int cpu; |
5460 | ||
5461 | for_each_possible_cpu(cpu) { | |
5462 | struct perf_cpu_context *cpuctx; | |
5463 | ||
5464 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
5465 | ||
5466 | if (cpuctx->active_pmu == old_pmu) | |
5467 | cpuctx->active_pmu = pmu; | |
5468 | } | |
5469 | } | |
5470 | ||
5471 | static void free_pmu_context(struct pmu *pmu) | |
5472 | { | |
5473 | struct pmu *i; | |
f5ffe02e | 5474 | |
8dc85d54 | 5475 | mutex_lock(&pmus_lock); |
0475f9ea | 5476 | /* |
8dc85d54 | 5477 | * Like a real lame refcount. |
0475f9ea | 5478 | */ |
51676957 PZ |
5479 | list_for_each_entry(i, &pmus, entry) { |
5480 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
5481 | update_pmu_context(i, pmu); | |
8dc85d54 | 5482 | goto out; |
51676957 | 5483 | } |
8dc85d54 | 5484 | } |
d6d020e9 | 5485 | |
51676957 | 5486 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
5487 | out: |
5488 | mutex_unlock(&pmus_lock); | |
24f1e32c | 5489 | } |
2e80a82a | 5490 | static struct idr pmu_idr; |
d6d020e9 | 5491 | |
abe43400 PZ |
5492 | static ssize_t |
5493 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
5494 | { | |
5495 | struct pmu *pmu = dev_get_drvdata(dev); | |
5496 | ||
5497 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
5498 | } | |
5499 | ||
5500 | static struct device_attribute pmu_dev_attrs[] = { | |
5501 | __ATTR_RO(type), | |
5502 | __ATTR_NULL, | |
5503 | }; | |
5504 | ||
5505 | static int pmu_bus_running; | |
5506 | static struct bus_type pmu_bus = { | |
5507 | .name = "event_source", | |
5508 | .dev_attrs = pmu_dev_attrs, | |
5509 | }; | |
5510 | ||
5511 | static void pmu_dev_release(struct device *dev) | |
5512 | { | |
5513 | kfree(dev); | |
5514 | } | |
5515 | ||
5516 | static int pmu_dev_alloc(struct pmu *pmu) | |
5517 | { | |
5518 | int ret = -ENOMEM; | |
5519 | ||
5520 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
5521 | if (!pmu->dev) | |
5522 | goto out; | |
5523 | ||
5524 | device_initialize(pmu->dev); | |
5525 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
5526 | if (ret) | |
5527 | goto free_dev; | |
5528 | ||
5529 | dev_set_drvdata(pmu->dev, pmu); | |
5530 | pmu->dev->bus = &pmu_bus; | |
5531 | pmu->dev->release = pmu_dev_release; | |
5532 | ret = device_add(pmu->dev); | |
5533 | if (ret) | |
5534 | goto free_dev; | |
5535 | ||
5536 | out: | |
5537 | return ret; | |
5538 | ||
5539 | free_dev: | |
5540 | put_device(pmu->dev); | |
5541 | goto out; | |
5542 | } | |
5543 | ||
547e9fd7 | 5544 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 5545 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 5546 | |
2e80a82a | 5547 | int perf_pmu_register(struct pmu *pmu, char *name, int type) |
24f1e32c | 5548 | { |
108b02cf | 5549 | int cpu, ret; |
24f1e32c | 5550 | |
b0a873eb | 5551 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
5552 | ret = -ENOMEM; |
5553 | pmu->pmu_disable_count = alloc_percpu(int); | |
5554 | if (!pmu->pmu_disable_count) | |
5555 | goto unlock; | |
f29ac756 | 5556 | |
2e80a82a PZ |
5557 | pmu->type = -1; |
5558 | if (!name) | |
5559 | goto skip_type; | |
5560 | pmu->name = name; | |
5561 | ||
5562 | if (type < 0) { | |
5563 | int err = idr_pre_get(&pmu_idr, GFP_KERNEL); | |
5564 | if (!err) | |
5565 | goto free_pdc; | |
5566 | ||
5567 | err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type); | |
5568 | if (err) { | |
5569 | ret = err; | |
5570 | goto free_pdc; | |
5571 | } | |
5572 | } | |
5573 | pmu->type = type; | |
5574 | ||
abe43400 PZ |
5575 | if (pmu_bus_running) { |
5576 | ret = pmu_dev_alloc(pmu); | |
5577 | if (ret) | |
5578 | goto free_idr; | |
5579 | } | |
5580 | ||
2e80a82a | 5581 | skip_type: |
8dc85d54 PZ |
5582 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
5583 | if (pmu->pmu_cpu_context) | |
5584 | goto got_cpu_context; | |
f29ac756 | 5585 | |
108b02cf PZ |
5586 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
5587 | if (!pmu->pmu_cpu_context) | |
abe43400 | 5588 | goto free_dev; |
f344011c | 5589 | |
108b02cf PZ |
5590 | for_each_possible_cpu(cpu) { |
5591 | struct perf_cpu_context *cpuctx; | |
5592 | ||
5593 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 5594 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 5595 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 5596 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 5597 | cpuctx->ctx.type = cpu_context; |
108b02cf | 5598 | cpuctx->ctx.pmu = pmu; |
e9d2b064 PZ |
5599 | cpuctx->jiffies_interval = 1; |
5600 | INIT_LIST_HEAD(&cpuctx->rotation_list); | |
51676957 | 5601 | cpuctx->active_pmu = pmu; |
108b02cf | 5602 | } |
76e1d904 | 5603 | |
8dc85d54 | 5604 | got_cpu_context: |
ad5133b7 PZ |
5605 | if (!pmu->start_txn) { |
5606 | if (pmu->pmu_enable) { | |
5607 | /* | |
5608 | * If we have pmu_enable/pmu_disable calls, install | |
5609 | * transaction stubs that use that to try and batch | |
5610 | * hardware accesses. | |
5611 | */ | |
5612 | pmu->start_txn = perf_pmu_start_txn; | |
5613 | pmu->commit_txn = perf_pmu_commit_txn; | |
5614 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
5615 | } else { | |
5616 | pmu->start_txn = perf_pmu_nop_void; | |
5617 | pmu->commit_txn = perf_pmu_nop_int; | |
5618 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 5619 | } |
5c92d124 | 5620 | } |
15dbf27c | 5621 | |
ad5133b7 PZ |
5622 | if (!pmu->pmu_enable) { |
5623 | pmu->pmu_enable = perf_pmu_nop_void; | |
5624 | pmu->pmu_disable = perf_pmu_nop_void; | |
5625 | } | |
5626 | ||
b0a873eb | 5627 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
5628 | ret = 0; |
5629 | unlock: | |
b0a873eb PZ |
5630 | mutex_unlock(&pmus_lock); |
5631 | ||
33696fc0 | 5632 | return ret; |
108b02cf | 5633 | |
abe43400 PZ |
5634 | free_dev: |
5635 | device_del(pmu->dev); | |
5636 | put_device(pmu->dev); | |
5637 | ||
2e80a82a PZ |
5638 | free_idr: |
5639 | if (pmu->type >= PERF_TYPE_MAX) | |
5640 | idr_remove(&pmu_idr, pmu->type); | |
5641 | ||
108b02cf PZ |
5642 | free_pdc: |
5643 | free_percpu(pmu->pmu_disable_count); | |
5644 | goto unlock; | |
f29ac756 PZ |
5645 | } |
5646 | ||
b0a873eb | 5647 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 5648 | { |
b0a873eb PZ |
5649 | mutex_lock(&pmus_lock); |
5650 | list_del_rcu(&pmu->entry); | |
5651 | mutex_unlock(&pmus_lock); | |
5c92d124 | 5652 | |
0475f9ea | 5653 | /* |
cde8e884 PZ |
5654 | * We dereference the pmu list under both SRCU and regular RCU, so |
5655 | * synchronize against both of those. | |
0475f9ea | 5656 | */ |
b0a873eb | 5657 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 5658 | synchronize_rcu(); |
d6d020e9 | 5659 | |
33696fc0 | 5660 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
5661 | if (pmu->type >= PERF_TYPE_MAX) |
5662 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
5663 | device_del(pmu->dev); |
5664 | put_device(pmu->dev); | |
51676957 | 5665 | free_pmu_context(pmu); |
b0a873eb | 5666 | } |
d6d020e9 | 5667 | |
b0a873eb PZ |
5668 | struct pmu *perf_init_event(struct perf_event *event) |
5669 | { | |
5670 | struct pmu *pmu = NULL; | |
5671 | int idx; | |
940c5b29 | 5672 | int ret; |
b0a873eb PZ |
5673 | |
5674 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
5675 | |
5676 | rcu_read_lock(); | |
5677 | pmu = idr_find(&pmu_idr, event->attr.type); | |
5678 | rcu_read_unlock(); | |
940c5b29 | 5679 | if (pmu) { |
7e5b2a01 | 5680 | event->pmu = pmu; |
940c5b29 LM |
5681 | ret = pmu->event_init(event); |
5682 | if (ret) | |
5683 | pmu = ERR_PTR(ret); | |
2e80a82a | 5684 | goto unlock; |
940c5b29 | 5685 | } |
2e80a82a | 5686 | |
b0a873eb | 5687 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 5688 | event->pmu = pmu; |
940c5b29 | 5689 | ret = pmu->event_init(event); |
b0a873eb | 5690 | if (!ret) |
e5f4d339 | 5691 | goto unlock; |
76e1d904 | 5692 | |
b0a873eb PZ |
5693 | if (ret != -ENOENT) { |
5694 | pmu = ERR_PTR(ret); | |
e5f4d339 | 5695 | goto unlock; |
f344011c | 5696 | } |
5c92d124 | 5697 | } |
e5f4d339 PZ |
5698 | pmu = ERR_PTR(-ENOENT); |
5699 | unlock: | |
b0a873eb | 5700 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 5701 | |
4aeb0b42 | 5702 | return pmu; |
5c92d124 IM |
5703 | } |
5704 | ||
0793a61d | 5705 | /* |
cdd6c482 | 5706 | * Allocate and initialize a event structure |
0793a61d | 5707 | */ |
cdd6c482 | 5708 | static struct perf_event * |
c3f00c70 | 5709 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
5710 | struct task_struct *task, |
5711 | struct perf_event *group_leader, | |
5712 | struct perf_event *parent_event, | |
4dc0da86 AK |
5713 | perf_overflow_handler_t overflow_handler, |
5714 | void *context) | |
0793a61d | 5715 | { |
51b0fe39 | 5716 | struct pmu *pmu; |
cdd6c482 IM |
5717 | struct perf_event *event; |
5718 | struct hw_perf_event *hwc; | |
d5d2bc0d | 5719 | long err; |
0793a61d | 5720 | |
66832eb4 ON |
5721 | if ((unsigned)cpu >= nr_cpu_ids) { |
5722 | if (!task || cpu != -1) | |
5723 | return ERR_PTR(-EINVAL); | |
5724 | } | |
5725 | ||
c3f00c70 | 5726 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 5727 | if (!event) |
d5d2bc0d | 5728 | return ERR_PTR(-ENOMEM); |
0793a61d | 5729 | |
04289bb9 | 5730 | /* |
cdd6c482 | 5731 | * Single events are their own group leaders, with an |
04289bb9 IM |
5732 | * empty sibling list: |
5733 | */ | |
5734 | if (!group_leader) | |
cdd6c482 | 5735 | group_leader = event; |
04289bb9 | 5736 | |
cdd6c482 IM |
5737 | mutex_init(&event->child_mutex); |
5738 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 5739 | |
cdd6c482 IM |
5740 | INIT_LIST_HEAD(&event->group_entry); |
5741 | INIT_LIST_HEAD(&event->event_entry); | |
5742 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
5743 | INIT_LIST_HEAD(&event->rb_entry); |
5744 | ||
cdd6c482 | 5745 | init_waitqueue_head(&event->waitq); |
e360adbe | 5746 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 5747 | |
cdd6c482 | 5748 | mutex_init(&event->mmap_mutex); |
7b732a75 | 5749 | |
cdd6c482 IM |
5750 | event->cpu = cpu; |
5751 | event->attr = *attr; | |
5752 | event->group_leader = group_leader; | |
5753 | event->pmu = NULL; | |
cdd6c482 | 5754 | event->oncpu = -1; |
a96bbc16 | 5755 | |
cdd6c482 | 5756 | event->parent = parent_event; |
b84fbc9f | 5757 | |
cdd6c482 IM |
5758 | event->ns = get_pid_ns(current->nsproxy->pid_ns); |
5759 | event->id = atomic64_inc_return(&perf_event_id); | |
a96bbc16 | 5760 | |
cdd6c482 | 5761 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 5762 | |
d580ff86 PZ |
5763 | if (task) { |
5764 | event->attach_state = PERF_ATTACH_TASK; | |
5765 | #ifdef CONFIG_HAVE_HW_BREAKPOINT | |
5766 | /* | |
5767 | * hw_breakpoint is a bit difficult here.. | |
5768 | */ | |
5769 | if (attr->type == PERF_TYPE_BREAKPOINT) | |
5770 | event->hw.bp_target = task; | |
5771 | #endif | |
5772 | } | |
5773 | ||
4dc0da86 | 5774 | if (!overflow_handler && parent_event) { |
b326e956 | 5775 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
5776 | context = parent_event->overflow_handler_context; |
5777 | } | |
66832eb4 | 5778 | |
b326e956 | 5779 | event->overflow_handler = overflow_handler; |
4dc0da86 | 5780 | event->overflow_handler_context = context; |
97eaf530 | 5781 | |
0d48696f | 5782 | if (attr->disabled) |
cdd6c482 | 5783 | event->state = PERF_EVENT_STATE_OFF; |
a86ed508 | 5784 | |
4aeb0b42 | 5785 | pmu = NULL; |
b8e83514 | 5786 | |
cdd6c482 | 5787 | hwc = &event->hw; |
bd2b5b12 | 5788 | hwc->sample_period = attr->sample_period; |
0d48696f | 5789 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 5790 | hwc->sample_period = 1; |
eced1dfc | 5791 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 5792 | |
e7850595 | 5793 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 5794 | |
2023b359 | 5795 | /* |
cdd6c482 | 5796 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 5797 | */ |
3dab77fb | 5798 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
2023b359 PZ |
5799 | goto done; |
5800 | ||
b0a873eb | 5801 | pmu = perf_init_event(event); |
974802ea | 5802 | |
d5d2bc0d PM |
5803 | done: |
5804 | err = 0; | |
4aeb0b42 | 5805 | if (!pmu) |
d5d2bc0d | 5806 | err = -EINVAL; |
4aeb0b42 RR |
5807 | else if (IS_ERR(pmu)) |
5808 | err = PTR_ERR(pmu); | |
5c92d124 | 5809 | |
d5d2bc0d | 5810 | if (err) { |
cdd6c482 IM |
5811 | if (event->ns) |
5812 | put_pid_ns(event->ns); | |
5813 | kfree(event); | |
d5d2bc0d | 5814 | return ERR_PTR(err); |
621a01ea | 5815 | } |
d5d2bc0d | 5816 | |
cdd6c482 | 5817 | if (!event->parent) { |
82cd6def | 5818 | if (event->attach_state & PERF_ATTACH_TASK) |
b2029520 | 5819 | jump_label_inc(&perf_sched_events.key); |
3af9e859 | 5820 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
5821 | atomic_inc(&nr_mmap_events); |
5822 | if (event->attr.comm) | |
5823 | atomic_inc(&nr_comm_events); | |
5824 | if (event->attr.task) | |
5825 | atomic_inc(&nr_task_events); | |
927c7a9e FW |
5826 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
5827 | err = get_callchain_buffers(); | |
5828 | if (err) { | |
5829 | free_event(event); | |
5830 | return ERR_PTR(err); | |
5831 | } | |
5832 | } | |
f344011c | 5833 | } |
9ee318a7 | 5834 | |
cdd6c482 | 5835 | return event; |
0793a61d TG |
5836 | } |
5837 | ||
cdd6c482 IM |
5838 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
5839 | struct perf_event_attr *attr) | |
974802ea | 5840 | { |
974802ea | 5841 | u32 size; |
cdf8073d | 5842 | int ret; |
974802ea PZ |
5843 | |
5844 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
5845 | return -EFAULT; | |
5846 | ||
5847 | /* | |
5848 | * zero the full structure, so that a short copy will be nice. | |
5849 | */ | |
5850 | memset(attr, 0, sizeof(*attr)); | |
5851 | ||
5852 | ret = get_user(size, &uattr->size); | |
5853 | if (ret) | |
5854 | return ret; | |
5855 | ||
5856 | if (size > PAGE_SIZE) /* silly large */ | |
5857 | goto err_size; | |
5858 | ||
5859 | if (!size) /* abi compat */ | |
5860 | size = PERF_ATTR_SIZE_VER0; | |
5861 | ||
5862 | if (size < PERF_ATTR_SIZE_VER0) | |
5863 | goto err_size; | |
5864 | ||
5865 | /* | |
5866 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
5867 | * ensure all the unknown bits are 0 - i.e. new |
5868 | * user-space does not rely on any kernel feature | |
5869 | * extensions we dont know about yet. | |
974802ea PZ |
5870 | */ |
5871 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
5872 | unsigned char __user *addr; |
5873 | unsigned char __user *end; | |
5874 | unsigned char val; | |
974802ea | 5875 | |
cdf8073d IS |
5876 | addr = (void __user *)uattr + sizeof(*attr); |
5877 | end = (void __user *)uattr + size; | |
974802ea | 5878 | |
cdf8073d | 5879 | for (; addr < end; addr++) { |
974802ea PZ |
5880 | ret = get_user(val, addr); |
5881 | if (ret) | |
5882 | return ret; | |
5883 | if (val) | |
5884 | goto err_size; | |
5885 | } | |
b3e62e35 | 5886 | size = sizeof(*attr); |
974802ea PZ |
5887 | } |
5888 | ||
5889 | ret = copy_from_user(attr, uattr, size); | |
5890 | if (ret) | |
5891 | return -EFAULT; | |
5892 | ||
cd757645 | 5893 | if (attr->__reserved_1) |
974802ea PZ |
5894 | return -EINVAL; |
5895 | ||
5896 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
5897 | return -EINVAL; | |
5898 | ||
5899 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
5900 | return -EINVAL; | |
5901 | ||
5902 | out: | |
5903 | return ret; | |
5904 | ||
5905 | err_size: | |
5906 | put_user(sizeof(*attr), &uattr->size); | |
5907 | ret = -E2BIG; | |
5908 | goto out; | |
5909 | } | |
5910 | ||
ac9721f3 PZ |
5911 | static int |
5912 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 5913 | { |
76369139 | 5914 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
5915 | int ret = -EINVAL; |
5916 | ||
ac9721f3 | 5917 | if (!output_event) |
a4be7c27 PZ |
5918 | goto set; |
5919 | ||
ac9721f3 PZ |
5920 | /* don't allow circular references */ |
5921 | if (event == output_event) | |
a4be7c27 PZ |
5922 | goto out; |
5923 | ||
0f139300 PZ |
5924 | /* |
5925 | * Don't allow cross-cpu buffers | |
5926 | */ | |
5927 | if (output_event->cpu != event->cpu) | |
5928 | goto out; | |
5929 | ||
5930 | /* | |
76369139 | 5931 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
5932 | */ |
5933 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
5934 | goto out; | |
5935 | ||
a4be7c27 | 5936 | set: |
cdd6c482 | 5937 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
5938 | /* Can't redirect output if we've got an active mmap() */ |
5939 | if (atomic_read(&event->mmap_count)) | |
5940 | goto unlock; | |
a4be7c27 | 5941 | |
ac9721f3 | 5942 | if (output_event) { |
76369139 FW |
5943 | /* get the rb we want to redirect to */ |
5944 | rb = ring_buffer_get(output_event); | |
5945 | if (!rb) | |
ac9721f3 | 5946 | goto unlock; |
a4be7c27 PZ |
5947 | } |
5948 | ||
76369139 FW |
5949 | old_rb = event->rb; |
5950 | rcu_assign_pointer(event->rb, rb); | |
10c6db11 PZ |
5951 | if (old_rb) |
5952 | ring_buffer_detach(event, old_rb); | |
a4be7c27 | 5953 | ret = 0; |
ac9721f3 PZ |
5954 | unlock: |
5955 | mutex_unlock(&event->mmap_mutex); | |
5956 | ||
76369139 FW |
5957 | if (old_rb) |
5958 | ring_buffer_put(old_rb); | |
a4be7c27 | 5959 | out: |
a4be7c27 PZ |
5960 | return ret; |
5961 | } | |
5962 | ||
0793a61d | 5963 | /** |
cdd6c482 | 5964 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 5965 | * |
cdd6c482 | 5966 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 5967 | * @pid: target pid |
9f66a381 | 5968 | * @cpu: target cpu |
cdd6c482 | 5969 | * @group_fd: group leader event fd |
0793a61d | 5970 | */ |
cdd6c482 IM |
5971 | SYSCALL_DEFINE5(perf_event_open, |
5972 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 5973 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 5974 | { |
b04243ef PZ |
5975 | struct perf_event *group_leader = NULL, *output_event = NULL; |
5976 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
5977 | struct perf_event_attr attr; |
5978 | struct perf_event_context *ctx; | |
5979 | struct file *event_file = NULL; | |
04289bb9 | 5980 | struct file *group_file = NULL; |
38a81da2 | 5981 | struct task_struct *task = NULL; |
89a1e187 | 5982 | struct pmu *pmu; |
ea635c64 | 5983 | int event_fd; |
b04243ef | 5984 | int move_group = 0; |
04289bb9 | 5985 | int fput_needed = 0; |
dc86cabe | 5986 | int err; |
0793a61d | 5987 | |
2743a5b0 | 5988 | /* for future expandability... */ |
e5d1367f | 5989 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
5990 | return -EINVAL; |
5991 | ||
dc86cabe IM |
5992 | err = perf_copy_attr(attr_uptr, &attr); |
5993 | if (err) | |
5994 | return err; | |
eab656ae | 5995 | |
0764771d PZ |
5996 | if (!attr.exclude_kernel) { |
5997 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
5998 | return -EACCES; | |
5999 | } | |
6000 | ||
df58ab24 | 6001 | if (attr.freq) { |
cdd6c482 | 6002 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
6003 | return -EINVAL; |
6004 | } | |
6005 | ||
e5d1367f SE |
6006 | /* |
6007 | * In cgroup mode, the pid argument is used to pass the fd | |
6008 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6009 | * designates the cpu on which to monitor threads from that | |
6010 | * cgroup. | |
6011 | */ | |
6012 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6013 | return -EINVAL; | |
6014 | ||
ea635c64 AV |
6015 | event_fd = get_unused_fd_flags(O_RDWR); |
6016 | if (event_fd < 0) | |
6017 | return event_fd; | |
6018 | ||
ac9721f3 PZ |
6019 | if (group_fd != -1) { |
6020 | group_leader = perf_fget_light(group_fd, &fput_needed); | |
6021 | if (IS_ERR(group_leader)) { | |
6022 | err = PTR_ERR(group_leader); | |
d14b12d7 | 6023 | goto err_fd; |
ac9721f3 PZ |
6024 | } |
6025 | group_file = group_leader->filp; | |
6026 | if (flags & PERF_FLAG_FD_OUTPUT) | |
6027 | output_event = group_leader; | |
6028 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6029 | group_leader = NULL; | |
6030 | } | |
6031 | ||
e5d1367f | 6032 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6033 | task = find_lively_task_by_vpid(pid); |
6034 | if (IS_ERR(task)) { | |
6035 | err = PTR_ERR(task); | |
6036 | goto err_group_fd; | |
6037 | } | |
6038 | } | |
6039 | ||
4dc0da86 AK |
6040 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6041 | NULL, NULL); | |
d14b12d7 SE |
6042 | if (IS_ERR(event)) { |
6043 | err = PTR_ERR(event); | |
c6be5a5c | 6044 | goto err_task; |
d14b12d7 SE |
6045 | } |
6046 | ||
e5d1367f SE |
6047 | if (flags & PERF_FLAG_PID_CGROUP) { |
6048 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
6049 | if (err) | |
6050 | goto err_alloc; | |
08309379 PZ |
6051 | /* |
6052 | * one more event: | |
6053 | * - that has cgroup constraint on event->cpu | |
6054 | * - that may need work on context switch | |
6055 | */ | |
6056 | atomic_inc(&per_cpu(perf_cgroup_events, event->cpu)); | |
b2029520 | 6057 | jump_label_inc(&perf_sched_events.key); |
e5d1367f SE |
6058 | } |
6059 | ||
89a1e187 PZ |
6060 | /* |
6061 | * Special case software events and allow them to be part of | |
6062 | * any hardware group. | |
6063 | */ | |
6064 | pmu = event->pmu; | |
b04243ef PZ |
6065 | |
6066 | if (group_leader && | |
6067 | (is_software_event(event) != is_software_event(group_leader))) { | |
6068 | if (is_software_event(event)) { | |
6069 | /* | |
6070 | * If event and group_leader are not both a software | |
6071 | * event, and event is, then group leader is not. | |
6072 | * | |
6073 | * Allow the addition of software events to !software | |
6074 | * groups, this is safe because software events never | |
6075 | * fail to schedule. | |
6076 | */ | |
6077 | pmu = group_leader->pmu; | |
6078 | } else if (is_software_event(group_leader) && | |
6079 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
6080 | /* | |
6081 | * In case the group is a pure software group, and we | |
6082 | * try to add a hardware event, move the whole group to | |
6083 | * the hardware context. | |
6084 | */ | |
6085 | move_group = 1; | |
6086 | } | |
6087 | } | |
89a1e187 PZ |
6088 | |
6089 | /* | |
6090 | * Get the target context (task or percpu): | |
6091 | */ | |
38a81da2 | 6092 | ctx = find_get_context(pmu, task, cpu); |
89a1e187 PZ |
6093 | if (IS_ERR(ctx)) { |
6094 | err = PTR_ERR(ctx); | |
c6be5a5c | 6095 | goto err_alloc; |
89a1e187 PZ |
6096 | } |
6097 | ||
fd1edb3a PZ |
6098 | if (task) { |
6099 | put_task_struct(task); | |
6100 | task = NULL; | |
6101 | } | |
6102 | ||
ccff286d | 6103 | /* |
cdd6c482 | 6104 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 6105 | */ |
ac9721f3 | 6106 | if (group_leader) { |
dc86cabe | 6107 | err = -EINVAL; |
04289bb9 | 6108 | |
04289bb9 | 6109 | /* |
ccff286d IM |
6110 | * Do not allow a recursive hierarchy (this new sibling |
6111 | * becoming part of another group-sibling): | |
6112 | */ | |
6113 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 6114 | goto err_context; |
ccff286d IM |
6115 | /* |
6116 | * Do not allow to attach to a group in a different | |
6117 | * task or CPU context: | |
04289bb9 | 6118 | */ |
b04243ef PZ |
6119 | if (move_group) { |
6120 | if (group_leader->ctx->type != ctx->type) | |
6121 | goto err_context; | |
6122 | } else { | |
6123 | if (group_leader->ctx != ctx) | |
6124 | goto err_context; | |
6125 | } | |
6126 | ||
3b6f9e5c PM |
6127 | /* |
6128 | * Only a group leader can be exclusive or pinned | |
6129 | */ | |
0d48696f | 6130 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 6131 | goto err_context; |
ac9721f3 PZ |
6132 | } |
6133 | ||
6134 | if (output_event) { | |
6135 | err = perf_event_set_output(event, output_event); | |
6136 | if (err) | |
c3f00c70 | 6137 | goto err_context; |
ac9721f3 | 6138 | } |
0793a61d | 6139 | |
ea635c64 AV |
6140 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
6141 | if (IS_ERR(event_file)) { | |
6142 | err = PTR_ERR(event_file); | |
c3f00c70 | 6143 | goto err_context; |
ea635c64 | 6144 | } |
9b51f66d | 6145 | |
b04243ef PZ |
6146 | if (move_group) { |
6147 | struct perf_event_context *gctx = group_leader->ctx; | |
6148 | ||
6149 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 6150 | perf_remove_from_context(group_leader); |
b04243ef PZ |
6151 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6152 | group_entry) { | |
fe4b04fa | 6153 | perf_remove_from_context(sibling); |
b04243ef PZ |
6154 | put_ctx(gctx); |
6155 | } | |
6156 | mutex_unlock(&gctx->mutex); | |
6157 | put_ctx(gctx); | |
ea635c64 | 6158 | } |
9b51f66d | 6159 | |
cdd6c482 | 6160 | event->filp = event_file; |
ad3a37de | 6161 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 6162 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
6163 | |
6164 | if (move_group) { | |
6165 | perf_install_in_context(ctx, group_leader, cpu); | |
6166 | get_ctx(ctx); | |
6167 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
6168 | group_entry) { | |
6169 | perf_install_in_context(ctx, sibling, cpu); | |
6170 | get_ctx(ctx); | |
6171 | } | |
6172 | } | |
6173 | ||
cdd6c482 | 6174 | perf_install_in_context(ctx, event, cpu); |
ad3a37de | 6175 | ++ctx->generation; |
fe4b04fa | 6176 | perf_unpin_context(ctx); |
d859e29f | 6177 | mutex_unlock(&ctx->mutex); |
9b51f66d | 6178 | |
cdd6c482 | 6179 | event->owner = current; |
8882135b | 6180 | |
cdd6c482 IM |
6181 | mutex_lock(¤t->perf_event_mutex); |
6182 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
6183 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 6184 | |
c320c7b7 ACM |
6185 | /* |
6186 | * Precalculate sample_data sizes | |
6187 | */ | |
6188 | perf_event__header_size(event); | |
6844c09d | 6189 | perf_event__id_header_size(event); |
c320c7b7 | 6190 | |
8a49542c PZ |
6191 | /* |
6192 | * Drop the reference on the group_event after placing the | |
6193 | * new event on the sibling_list. This ensures destruction | |
6194 | * of the group leader will find the pointer to itself in | |
6195 | * perf_group_detach(). | |
6196 | */ | |
ea635c64 AV |
6197 | fput_light(group_file, fput_needed); |
6198 | fd_install(event_fd, event_file); | |
6199 | return event_fd; | |
0793a61d | 6200 | |
c3f00c70 | 6201 | err_context: |
fe4b04fa | 6202 | perf_unpin_context(ctx); |
ea635c64 | 6203 | put_ctx(ctx); |
c6be5a5c | 6204 | err_alloc: |
ea635c64 | 6205 | free_event(event); |
e7d0bc04 PZ |
6206 | err_task: |
6207 | if (task) | |
6208 | put_task_struct(task); | |
89a1e187 | 6209 | err_group_fd: |
dc86cabe | 6210 | fput_light(group_file, fput_needed); |
ea635c64 AV |
6211 | err_fd: |
6212 | put_unused_fd(event_fd); | |
dc86cabe | 6213 | return err; |
0793a61d TG |
6214 | } |
6215 | ||
fb0459d7 AV |
6216 | /** |
6217 | * perf_event_create_kernel_counter | |
6218 | * | |
6219 | * @attr: attributes of the counter to create | |
6220 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 6221 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
6222 | */ |
6223 | struct perf_event * | |
6224 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 6225 | struct task_struct *task, |
4dc0da86 AK |
6226 | perf_overflow_handler_t overflow_handler, |
6227 | void *context) | |
fb0459d7 | 6228 | { |
fb0459d7 | 6229 | struct perf_event_context *ctx; |
c3f00c70 | 6230 | struct perf_event *event; |
fb0459d7 | 6231 | int err; |
d859e29f | 6232 | |
fb0459d7 AV |
6233 | /* |
6234 | * Get the target context (task or percpu): | |
6235 | */ | |
d859e29f | 6236 | |
4dc0da86 AK |
6237 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
6238 | overflow_handler, context); | |
c3f00c70 PZ |
6239 | if (IS_ERR(event)) { |
6240 | err = PTR_ERR(event); | |
6241 | goto err; | |
6242 | } | |
d859e29f | 6243 | |
38a81da2 | 6244 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
6245 | if (IS_ERR(ctx)) { |
6246 | err = PTR_ERR(ctx); | |
c3f00c70 | 6247 | goto err_free; |
d859e29f | 6248 | } |
fb0459d7 AV |
6249 | |
6250 | event->filp = NULL; | |
6251 | WARN_ON_ONCE(ctx->parent_ctx); | |
6252 | mutex_lock(&ctx->mutex); | |
6253 | perf_install_in_context(ctx, event, cpu); | |
6254 | ++ctx->generation; | |
fe4b04fa | 6255 | perf_unpin_context(ctx); |
fb0459d7 AV |
6256 | mutex_unlock(&ctx->mutex); |
6257 | ||
fb0459d7 AV |
6258 | return event; |
6259 | ||
c3f00c70 PZ |
6260 | err_free: |
6261 | free_event(event); | |
6262 | err: | |
c6567f64 | 6263 | return ERR_PTR(err); |
9b51f66d | 6264 | } |
fb0459d7 | 6265 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 6266 | |
cdd6c482 | 6267 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 6268 | struct task_struct *child) |
d859e29f | 6269 | { |
cdd6c482 | 6270 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 6271 | u64 child_val; |
d859e29f | 6272 | |
cdd6c482 IM |
6273 | if (child_event->attr.inherit_stat) |
6274 | perf_event_read_event(child_event, child); | |
38b200d6 | 6275 | |
b5e58793 | 6276 | child_val = perf_event_count(child_event); |
d859e29f PM |
6277 | |
6278 | /* | |
6279 | * Add back the child's count to the parent's count: | |
6280 | */ | |
a6e6dea6 | 6281 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
6282 | atomic64_add(child_event->total_time_enabled, |
6283 | &parent_event->child_total_time_enabled); | |
6284 | atomic64_add(child_event->total_time_running, | |
6285 | &parent_event->child_total_time_running); | |
d859e29f PM |
6286 | |
6287 | /* | |
cdd6c482 | 6288 | * Remove this event from the parent's list |
d859e29f | 6289 | */ |
cdd6c482 IM |
6290 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
6291 | mutex_lock(&parent_event->child_mutex); | |
6292 | list_del_init(&child_event->child_list); | |
6293 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
6294 | |
6295 | /* | |
cdd6c482 | 6296 | * Release the parent event, if this was the last |
d859e29f PM |
6297 | * reference to it. |
6298 | */ | |
cdd6c482 | 6299 | fput(parent_event->filp); |
d859e29f PM |
6300 | } |
6301 | ||
9b51f66d | 6302 | static void |
cdd6c482 IM |
6303 | __perf_event_exit_task(struct perf_event *child_event, |
6304 | struct perf_event_context *child_ctx, | |
38b200d6 | 6305 | struct task_struct *child) |
9b51f66d | 6306 | { |
38b435b1 PZ |
6307 | if (child_event->parent) { |
6308 | raw_spin_lock_irq(&child_ctx->lock); | |
6309 | perf_group_detach(child_event); | |
6310 | raw_spin_unlock_irq(&child_ctx->lock); | |
6311 | } | |
9b51f66d | 6312 | |
fe4b04fa | 6313 | perf_remove_from_context(child_event); |
0cc0c027 | 6314 | |
9b51f66d | 6315 | /* |
38b435b1 | 6316 | * It can happen that the parent exits first, and has events |
9b51f66d | 6317 | * that are still around due to the child reference. These |
38b435b1 | 6318 | * events need to be zapped. |
9b51f66d | 6319 | */ |
38b435b1 | 6320 | if (child_event->parent) { |
cdd6c482 IM |
6321 | sync_child_event(child_event, child); |
6322 | free_event(child_event); | |
4bcf349a | 6323 | } |
9b51f66d IM |
6324 | } |
6325 | ||
8dc85d54 | 6326 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 6327 | { |
cdd6c482 IM |
6328 | struct perf_event *child_event, *tmp; |
6329 | struct perf_event_context *child_ctx; | |
a63eaf34 | 6330 | unsigned long flags; |
9b51f66d | 6331 | |
8dc85d54 | 6332 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 6333 | perf_event_task(child, NULL, 0); |
9b51f66d | 6334 | return; |
9f498cc5 | 6335 | } |
9b51f66d | 6336 | |
a63eaf34 | 6337 | local_irq_save(flags); |
ad3a37de PM |
6338 | /* |
6339 | * We can't reschedule here because interrupts are disabled, | |
6340 | * and either child is current or it is a task that can't be | |
6341 | * scheduled, so we are now safe from rescheduling changing | |
6342 | * our context. | |
6343 | */ | |
806839b2 | 6344 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
6345 | |
6346 | /* | |
6347 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 6348 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
6349 | * incremented the context's refcount before we do put_ctx below. |
6350 | */ | |
e625cce1 | 6351 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 6352 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 6353 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
6354 | /* |
6355 | * If this context is a clone; unclone it so it can't get | |
6356 | * swapped to another process while we're removing all | |
cdd6c482 | 6357 | * the events from it. |
71a851b4 PZ |
6358 | */ |
6359 | unclone_ctx(child_ctx); | |
5e942bb3 | 6360 | update_context_time(child_ctx); |
e625cce1 | 6361 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
6362 | |
6363 | /* | |
cdd6c482 IM |
6364 | * Report the task dead after unscheduling the events so that we |
6365 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
6366 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 6367 | */ |
cdd6c482 | 6368 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 6369 | |
66fff224 PZ |
6370 | /* |
6371 | * We can recurse on the same lock type through: | |
6372 | * | |
cdd6c482 IM |
6373 | * __perf_event_exit_task() |
6374 | * sync_child_event() | |
6375 | * fput(parent_event->filp) | |
66fff224 PZ |
6376 | * perf_release() |
6377 | * mutex_lock(&ctx->mutex) | |
6378 | * | |
6379 | * But since its the parent context it won't be the same instance. | |
6380 | */ | |
a0507c84 | 6381 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 6382 | |
8bc20959 | 6383 | again: |
889ff015 FW |
6384 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
6385 | group_entry) | |
6386 | __perf_event_exit_task(child_event, child_ctx, child); | |
6387 | ||
6388 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 6389 | group_entry) |
cdd6c482 | 6390 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
6391 | |
6392 | /* | |
cdd6c482 | 6393 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
6394 | * its siblings to the list, but we obtained 'tmp' before that which |
6395 | * will still point to the list head terminating the iteration. | |
6396 | */ | |
889ff015 FW |
6397 | if (!list_empty(&child_ctx->pinned_groups) || |
6398 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 6399 | goto again; |
a63eaf34 PM |
6400 | |
6401 | mutex_unlock(&child_ctx->mutex); | |
6402 | ||
6403 | put_ctx(child_ctx); | |
9b51f66d IM |
6404 | } |
6405 | ||
8dc85d54 PZ |
6406 | /* |
6407 | * When a child task exits, feed back event values to parent events. | |
6408 | */ | |
6409 | void perf_event_exit_task(struct task_struct *child) | |
6410 | { | |
8882135b | 6411 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
6412 | int ctxn; |
6413 | ||
8882135b PZ |
6414 | mutex_lock(&child->perf_event_mutex); |
6415 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
6416 | owner_entry) { | |
6417 | list_del_init(&event->owner_entry); | |
6418 | ||
6419 | /* | |
6420 | * Ensure the list deletion is visible before we clear | |
6421 | * the owner, closes a race against perf_release() where | |
6422 | * we need to serialize on the owner->perf_event_mutex. | |
6423 | */ | |
6424 | smp_wmb(); | |
6425 | event->owner = NULL; | |
6426 | } | |
6427 | mutex_unlock(&child->perf_event_mutex); | |
6428 | ||
8dc85d54 PZ |
6429 | for_each_task_context_nr(ctxn) |
6430 | perf_event_exit_task_context(child, ctxn); | |
6431 | } | |
6432 | ||
889ff015 FW |
6433 | static void perf_free_event(struct perf_event *event, |
6434 | struct perf_event_context *ctx) | |
6435 | { | |
6436 | struct perf_event *parent = event->parent; | |
6437 | ||
6438 | if (WARN_ON_ONCE(!parent)) | |
6439 | return; | |
6440 | ||
6441 | mutex_lock(&parent->child_mutex); | |
6442 | list_del_init(&event->child_list); | |
6443 | mutex_unlock(&parent->child_mutex); | |
6444 | ||
6445 | fput(parent->filp); | |
6446 | ||
8a49542c | 6447 | perf_group_detach(event); |
889ff015 FW |
6448 | list_del_event(event, ctx); |
6449 | free_event(event); | |
6450 | } | |
6451 | ||
bbbee908 PZ |
6452 | /* |
6453 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 6454 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 6455 | */ |
cdd6c482 | 6456 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 6457 | { |
8dc85d54 | 6458 | struct perf_event_context *ctx; |
cdd6c482 | 6459 | struct perf_event *event, *tmp; |
8dc85d54 | 6460 | int ctxn; |
bbbee908 | 6461 | |
8dc85d54 PZ |
6462 | for_each_task_context_nr(ctxn) { |
6463 | ctx = task->perf_event_ctxp[ctxn]; | |
6464 | if (!ctx) | |
6465 | continue; | |
bbbee908 | 6466 | |
8dc85d54 | 6467 | mutex_lock(&ctx->mutex); |
bbbee908 | 6468 | again: |
8dc85d54 PZ |
6469 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
6470 | group_entry) | |
6471 | perf_free_event(event, ctx); | |
bbbee908 | 6472 | |
8dc85d54 PZ |
6473 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
6474 | group_entry) | |
6475 | perf_free_event(event, ctx); | |
bbbee908 | 6476 | |
8dc85d54 PZ |
6477 | if (!list_empty(&ctx->pinned_groups) || |
6478 | !list_empty(&ctx->flexible_groups)) | |
6479 | goto again; | |
bbbee908 | 6480 | |
8dc85d54 | 6481 | mutex_unlock(&ctx->mutex); |
bbbee908 | 6482 | |
8dc85d54 PZ |
6483 | put_ctx(ctx); |
6484 | } | |
889ff015 FW |
6485 | } |
6486 | ||
4e231c79 PZ |
6487 | void perf_event_delayed_put(struct task_struct *task) |
6488 | { | |
6489 | int ctxn; | |
6490 | ||
6491 | for_each_task_context_nr(ctxn) | |
6492 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
6493 | } | |
6494 | ||
97dee4f3 PZ |
6495 | /* |
6496 | * inherit a event from parent task to child task: | |
6497 | */ | |
6498 | static struct perf_event * | |
6499 | inherit_event(struct perf_event *parent_event, | |
6500 | struct task_struct *parent, | |
6501 | struct perf_event_context *parent_ctx, | |
6502 | struct task_struct *child, | |
6503 | struct perf_event *group_leader, | |
6504 | struct perf_event_context *child_ctx) | |
6505 | { | |
6506 | struct perf_event *child_event; | |
cee010ec | 6507 | unsigned long flags; |
97dee4f3 PZ |
6508 | |
6509 | /* | |
6510 | * Instead of creating recursive hierarchies of events, | |
6511 | * we link inherited events back to the original parent, | |
6512 | * which has a filp for sure, which we use as the reference | |
6513 | * count: | |
6514 | */ | |
6515 | if (parent_event->parent) | |
6516 | parent_event = parent_event->parent; | |
6517 | ||
6518 | child_event = perf_event_alloc(&parent_event->attr, | |
6519 | parent_event->cpu, | |
d580ff86 | 6520 | child, |
97dee4f3 | 6521 | group_leader, parent_event, |
4dc0da86 | 6522 | NULL, NULL); |
97dee4f3 PZ |
6523 | if (IS_ERR(child_event)) |
6524 | return child_event; | |
6525 | get_ctx(child_ctx); | |
6526 | ||
6527 | /* | |
6528 | * Make the child state follow the state of the parent event, | |
6529 | * not its attr.disabled bit. We hold the parent's mutex, | |
6530 | * so we won't race with perf_event_{en, dis}able_family. | |
6531 | */ | |
6532 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
6533 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
6534 | else | |
6535 | child_event->state = PERF_EVENT_STATE_OFF; | |
6536 | ||
6537 | if (parent_event->attr.freq) { | |
6538 | u64 sample_period = parent_event->hw.sample_period; | |
6539 | struct hw_perf_event *hwc = &child_event->hw; | |
6540 | ||
6541 | hwc->sample_period = sample_period; | |
6542 | hwc->last_period = sample_period; | |
6543 | ||
6544 | local64_set(&hwc->period_left, sample_period); | |
6545 | } | |
6546 | ||
6547 | child_event->ctx = child_ctx; | |
6548 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
6549 | child_event->overflow_handler_context |
6550 | = parent_event->overflow_handler_context; | |
97dee4f3 | 6551 | |
614b6780 TG |
6552 | /* |
6553 | * Precalculate sample_data sizes | |
6554 | */ | |
6555 | perf_event__header_size(child_event); | |
6844c09d | 6556 | perf_event__id_header_size(child_event); |
614b6780 | 6557 | |
97dee4f3 PZ |
6558 | /* |
6559 | * Link it up in the child's context: | |
6560 | */ | |
cee010ec | 6561 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 6562 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 6563 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 PZ |
6564 | |
6565 | /* | |
6566 | * Get a reference to the parent filp - we will fput it | |
6567 | * when the child event exits. This is safe to do because | |
6568 | * we are in the parent and we know that the filp still | |
6569 | * exists and has a nonzero count: | |
6570 | */ | |
6571 | atomic_long_inc(&parent_event->filp->f_count); | |
6572 | ||
6573 | /* | |
6574 | * Link this into the parent event's child list | |
6575 | */ | |
6576 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
6577 | mutex_lock(&parent_event->child_mutex); | |
6578 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
6579 | mutex_unlock(&parent_event->child_mutex); | |
6580 | ||
6581 | return child_event; | |
6582 | } | |
6583 | ||
6584 | static int inherit_group(struct perf_event *parent_event, | |
6585 | struct task_struct *parent, | |
6586 | struct perf_event_context *parent_ctx, | |
6587 | struct task_struct *child, | |
6588 | struct perf_event_context *child_ctx) | |
6589 | { | |
6590 | struct perf_event *leader; | |
6591 | struct perf_event *sub; | |
6592 | struct perf_event *child_ctr; | |
6593 | ||
6594 | leader = inherit_event(parent_event, parent, parent_ctx, | |
6595 | child, NULL, child_ctx); | |
6596 | if (IS_ERR(leader)) | |
6597 | return PTR_ERR(leader); | |
6598 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
6599 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
6600 | child, leader, child_ctx); | |
6601 | if (IS_ERR(child_ctr)) | |
6602 | return PTR_ERR(child_ctr); | |
6603 | } | |
6604 | return 0; | |
889ff015 FW |
6605 | } |
6606 | ||
6607 | static int | |
6608 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
6609 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 6610 | struct task_struct *child, int ctxn, |
889ff015 FW |
6611 | int *inherited_all) |
6612 | { | |
6613 | int ret; | |
8dc85d54 | 6614 | struct perf_event_context *child_ctx; |
889ff015 FW |
6615 | |
6616 | if (!event->attr.inherit) { | |
6617 | *inherited_all = 0; | |
6618 | return 0; | |
bbbee908 PZ |
6619 | } |
6620 | ||
fe4b04fa | 6621 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
6622 | if (!child_ctx) { |
6623 | /* | |
6624 | * This is executed from the parent task context, so | |
6625 | * inherit events that have been marked for cloning. | |
6626 | * First allocate and initialize a context for the | |
6627 | * child. | |
6628 | */ | |
bbbee908 | 6629 | |
eb184479 | 6630 | child_ctx = alloc_perf_context(event->pmu, child); |
889ff015 FW |
6631 | if (!child_ctx) |
6632 | return -ENOMEM; | |
bbbee908 | 6633 | |
8dc85d54 | 6634 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
6635 | } |
6636 | ||
6637 | ret = inherit_group(event, parent, parent_ctx, | |
6638 | child, child_ctx); | |
6639 | ||
6640 | if (ret) | |
6641 | *inherited_all = 0; | |
6642 | ||
6643 | return ret; | |
bbbee908 PZ |
6644 | } |
6645 | ||
9b51f66d | 6646 | /* |
cdd6c482 | 6647 | * Initialize the perf_event context in task_struct |
9b51f66d | 6648 | */ |
8dc85d54 | 6649 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 6650 | { |
889ff015 | 6651 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
6652 | struct perf_event_context *cloned_ctx; |
6653 | struct perf_event *event; | |
9b51f66d | 6654 | struct task_struct *parent = current; |
564c2b21 | 6655 | int inherited_all = 1; |
dddd3379 | 6656 | unsigned long flags; |
6ab423e0 | 6657 | int ret = 0; |
9b51f66d | 6658 | |
8dc85d54 | 6659 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
6660 | return 0; |
6661 | ||
ad3a37de | 6662 | /* |
25346b93 PM |
6663 | * If the parent's context is a clone, pin it so it won't get |
6664 | * swapped under us. | |
ad3a37de | 6665 | */ |
8dc85d54 | 6666 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 6667 | |
ad3a37de PM |
6668 | /* |
6669 | * No need to check if parent_ctx != NULL here; since we saw | |
6670 | * it non-NULL earlier, the only reason for it to become NULL | |
6671 | * is if we exit, and since we're currently in the middle of | |
6672 | * a fork we can't be exiting at the same time. | |
6673 | */ | |
ad3a37de | 6674 | |
9b51f66d IM |
6675 | /* |
6676 | * Lock the parent list. No need to lock the child - not PID | |
6677 | * hashed yet and not running, so nobody can access it. | |
6678 | */ | |
d859e29f | 6679 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
6680 | |
6681 | /* | |
6682 | * We dont have to disable NMIs - we are only looking at | |
6683 | * the list, not manipulating it: | |
6684 | */ | |
889ff015 | 6685 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
6686 | ret = inherit_task_group(event, parent, parent_ctx, |
6687 | child, ctxn, &inherited_all); | |
889ff015 FW |
6688 | if (ret) |
6689 | break; | |
6690 | } | |
b93f7978 | 6691 | |
dddd3379 TG |
6692 | /* |
6693 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
6694 | * to allocations, but we need to prevent rotation because | |
6695 | * rotate_ctx() will change the list from interrupt context. | |
6696 | */ | |
6697 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
6698 | parent_ctx->rotate_disable = 1; | |
6699 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
6700 | ||
889ff015 | 6701 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
6702 | ret = inherit_task_group(event, parent, parent_ctx, |
6703 | child, ctxn, &inherited_all); | |
889ff015 | 6704 | if (ret) |
9b51f66d | 6705 | break; |
564c2b21 PM |
6706 | } |
6707 | ||
dddd3379 TG |
6708 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
6709 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 6710 | |
8dc85d54 | 6711 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 6712 | |
05cbaa28 | 6713 | if (child_ctx && inherited_all) { |
564c2b21 PM |
6714 | /* |
6715 | * Mark the child context as a clone of the parent | |
6716 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
6717 | * |
6718 | * Note that if the parent is a clone, the holding of | |
6719 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 6720 | */ |
c5ed5145 | 6721 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
6722 | if (cloned_ctx) { |
6723 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 6724 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
6725 | } else { |
6726 | child_ctx->parent_ctx = parent_ctx; | |
6727 | child_ctx->parent_gen = parent_ctx->generation; | |
6728 | } | |
6729 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
6730 | } |
6731 | ||
c5ed5145 | 6732 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 6733 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 6734 | |
25346b93 | 6735 | perf_unpin_context(parent_ctx); |
fe4b04fa | 6736 | put_ctx(parent_ctx); |
ad3a37de | 6737 | |
6ab423e0 | 6738 | return ret; |
9b51f66d IM |
6739 | } |
6740 | ||
8dc85d54 PZ |
6741 | /* |
6742 | * Initialize the perf_event context in task_struct | |
6743 | */ | |
6744 | int perf_event_init_task(struct task_struct *child) | |
6745 | { | |
6746 | int ctxn, ret; | |
6747 | ||
8550d7cb ON |
6748 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
6749 | mutex_init(&child->perf_event_mutex); | |
6750 | INIT_LIST_HEAD(&child->perf_event_list); | |
6751 | ||
8dc85d54 PZ |
6752 | for_each_task_context_nr(ctxn) { |
6753 | ret = perf_event_init_context(child, ctxn); | |
6754 | if (ret) | |
6755 | return ret; | |
6756 | } | |
6757 | ||
6758 | return 0; | |
6759 | } | |
6760 | ||
220b140b PM |
6761 | static void __init perf_event_init_all_cpus(void) |
6762 | { | |
b28ab83c | 6763 | struct swevent_htable *swhash; |
220b140b | 6764 | int cpu; |
220b140b PM |
6765 | |
6766 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
6767 | swhash = &per_cpu(swevent_htable, cpu); |
6768 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 6769 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
6770 | } |
6771 | } | |
6772 | ||
cdd6c482 | 6773 | static void __cpuinit perf_event_init_cpu(int cpu) |
0793a61d | 6774 | { |
108b02cf | 6775 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 6776 | |
b28ab83c | 6777 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 6778 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
6779 | struct swevent_hlist *hlist; |
6780 | ||
b28ab83c PZ |
6781 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
6782 | WARN_ON(!hlist); | |
6783 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 6784 | } |
b28ab83c | 6785 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
6786 | } |
6787 | ||
c277443c | 6788 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 6789 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 6790 | { |
e9d2b064 PZ |
6791 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
6792 | ||
6793 | WARN_ON(!irqs_disabled()); | |
6794 | ||
6795 | list_del_init(&cpuctx->rotation_list); | |
6796 | } | |
6797 | ||
108b02cf | 6798 | static void __perf_event_exit_context(void *__info) |
0793a61d | 6799 | { |
108b02cf | 6800 | struct perf_event_context *ctx = __info; |
cdd6c482 | 6801 | struct perf_event *event, *tmp; |
0793a61d | 6802 | |
108b02cf | 6803 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 6804 | |
889ff015 | 6805 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 6806 | __perf_remove_from_context(event); |
889ff015 | 6807 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 6808 | __perf_remove_from_context(event); |
0793a61d | 6809 | } |
108b02cf PZ |
6810 | |
6811 | static void perf_event_exit_cpu_context(int cpu) | |
6812 | { | |
6813 | struct perf_event_context *ctx; | |
6814 | struct pmu *pmu; | |
6815 | int idx; | |
6816 | ||
6817 | idx = srcu_read_lock(&pmus_srcu); | |
6818 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 6819 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
6820 | |
6821 | mutex_lock(&ctx->mutex); | |
6822 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
6823 | mutex_unlock(&ctx->mutex); | |
6824 | } | |
6825 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
6826 | } |
6827 | ||
cdd6c482 | 6828 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 6829 | { |
b28ab83c | 6830 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 6831 | |
b28ab83c PZ |
6832 | mutex_lock(&swhash->hlist_mutex); |
6833 | swevent_hlist_release(swhash); | |
6834 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 6835 | |
108b02cf | 6836 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
6837 | } |
6838 | #else | |
cdd6c482 | 6839 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
6840 | #endif |
6841 | ||
c277443c PZ |
6842 | static int |
6843 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
6844 | { | |
6845 | int cpu; | |
6846 | ||
6847 | for_each_online_cpu(cpu) | |
6848 | perf_event_exit_cpu(cpu); | |
6849 | ||
6850 | return NOTIFY_OK; | |
6851 | } | |
6852 | ||
6853 | /* | |
6854 | * Run the perf reboot notifier at the very last possible moment so that | |
6855 | * the generic watchdog code runs as long as possible. | |
6856 | */ | |
6857 | static struct notifier_block perf_reboot_notifier = { | |
6858 | .notifier_call = perf_reboot, | |
6859 | .priority = INT_MIN, | |
6860 | }; | |
6861 | ||
0793a61d TG |
6862 | static int __cpuinit |
6863 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
6864 | { | |
6865 | unsigned int cpu = (long)hcpu; | |
6866 | ||
4536e4d1 | 6867 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
6868 | |
6869 | case CPU_UP_PREPARE: | |
5e11637e | 6870 | case CPU_DOWN_FAILED: |
cdd6c482 | 6871 | perf_event_init_cpu(cpu); |
0793a61d TG |
6872 | break; |
6873 | ||
5e11637e | 6874 | case CPU_UP_CANCELED: |
0793a61d | 6875 | case CPU_DOWN_PREPARE: |
cdd6c482 | 6876 | perf_event_exit_cpu(cpu); |
0793a61d TG |
6877 | break; |
6878 | ||
6879 | default: | |
6880 | break; | |
6881 | } | |
6882 | ||
6883 | return NOTIFY_OK; | |
6884 | } | |
6885 | ||
cdd6c482 | 6886 | void __init perf_event_init(void) |
0793a61d | 6887 | { |
3c502e7a JW |
6888 | int ret; |
6889 | ||
2e80a82a PZ |
6890 | idr_init(&pmu_idr); |
6891 | ||
220b140b | 6892 | perf_event_init_all_cpus(); |
b0a873eb | 6893 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
6894 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
6895 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
6896 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
6897 | perf_tp_register(); |
6898 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 6899 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
6900 | |
6901 | ret = init_hw_breakpoint(); | |
6902 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
6903 | |
6904 | /* do not patch jump label more than once per second */ | |
6905 | jump_label_rate_limit(&perf_sched_events, HZ); | |
0793a61d | 6906 | } |
abe43400 PZ |
6907 | |
6908 | static int __init perf_event_sysfs_init(void) | |
6909 | { | |
6910 | struct pmu *pmu; | |
6911 | int ret; | |
6912 | ||
6913 | mutex_lock(&pmus_lock); | |
6914 | ||
6915 | ret = bus_register(&pmu_bus); | |
6916 | if (ret) | |
6917 | goto unlock; | |
6918 | ||
6919 | list_for_each_entry(pmu, &pmus, entry) { | |
6920 | if (!pmu->name || pmu->type < 0) | |
6921 | continue; | |
6922 | ||
6923 | ret = pmu_dev_alloc(pmu); | |
6924 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
6925 | } | |
6926 | pmu_bus_running = 1; | |
6927 | ret = 0; | |
6928 | ||
6929 | unlock: | |
6930 | mutex_unlock(&pmus_lock); | |
6931 | ||
6932 | return ret; | |
6933 | } | |
6934 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
6935 | |
6936 | #ifdef CONFIG_CGROUP_PERF | |
6937 | static struct cgroup_subsys_state *perf_cgroup_create( | |
6938 | struct cgroup_subsys *ss, struct cgroup *cont) | |
6939 | { | |
6940 | struct perf_cgroup *jc; | |
e5d1367f | 6941 | |
1b15d055 | 6942 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
6943 | if (!jc) |
6944 | return ERR_PTR(-ENOMEM); | |
6945 | ||
e5d1367f SE |
6946 | jc->info = alloc_percpu(struct perf_cgroup_info); |
6947 | if (!jc->info) { | |
6948 | kfree(jc); | |
6949 | return ERR_PTR(-ENOMEM); | |
6950 | } | |
6951 | ||
e5d1367f SE |
6952 | return &jc->css; |
6953 | } | |
6954 | ||
6955 | static void perf_cgroup_destroy(struct cgroup_subsys *ss, | |
6956 | struct cgroup *cont) | |
6957 | { | |
6958 | struct perf_cgroup *jc; | |
6959 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), | |
6960 | struct perf_cgroup, css); | |
6961 | free_percpu(jc->info); | |
6962 | kfree(jc); | |
6963 | } | |
6964 | ||
6965 | static int __perf_cgroup_move(void *info) | |
6966 | { | |
6967 | struct task_struct *task = info; | |
6968 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
6969 | return 0; | |
6970 | } | |
6971 | ||
bb9d97b6 TH |
6972 | static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
6973 | struct cgroup_taskset *tset) | |
e5d1367f | 6974 | { |
bb9d97b6 TH |
6975 | struct task_struct *task; |
6976 | ||
6977 | cgroup_taskset_for_each(task, cgrp, tset) | |
6978 | task_function_call(task, __perf_cgroup_move, task); | |
e5d1367f SE |
6979 | } |
6980 | ||
e5d1367f SE |
6981 | static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp, |
6982 | struct cgroup *old_cgrp, struct task_struct *task) | |
6983 | { | |
6984 | /* | |
6985 | * cgroup_exit() is called in the copy_process() failure path. | |
6986 | * Ignore this case since the task hasn't ran yet, this avoids | |
6987 | * trying to poke a half freed task state from generic code. | |
6988 | */ | |
6989 | if (!(task->flags & PF_EXITING)) | |
6990 | return; | |
6991 | ||
bb9d97b6 | 6992 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
6993 | } |
6994 | ||
6995 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
6996 | .name = "perf_event", |
6997 | .subsys_id = perf_subsys_id, | |
6998 | .create = perf_cgroup_create, | |
6999 | .destroy = perf_cgroup_destroy, | |
7000 | .exit = perf_cgroup_exit, | |
bb9d97b6 | 7001 | .attach = perf_cgroup_attach, |
e5d1367f SE |
7002 | }; |
7003 | #endif /* CONFIG_CGROUP_PERF */ |