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