]>
Commit | Line | Data |
---|---|---|
0793a61d | 1 | /* |
57c0c15b | 2 | * Performance events core code: |
0793a61d | 3 | * |
98144511 | 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
e7e7ee2e IM |
5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
7b732a75 | 8 | * |
57c0c15b | 9 | * For licensing details see kernel-base/COPYING |
0793a61d TG |
10 | */ |
11 | ||
12 | #include <linux/fs.h> | |
b9cacc7b | 13 | #include <linux/mm.h> |
0793a61d TG |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> | |
2e80a82a | 16 | #include <linux/idr.h> |
04289bb9 | 17 | #include <linux/file.h> |
0793a61d | 18 | #include <linux/poll.h> |
5a0e3ad6 | 19 | #include <linux/slab.h> |
76e1d904 | 20 | #include <linux/hash.h> |
0793a61d | 21 | #include <linux/sysfs.h> |
22a4f650 | 22 | #include <linux/dcache.h> |
0793a61d | 23 | #include <linux/percpu.h> |
22a4f650 | 24 | #include <linux/ptrace.h> |
c277443c | 25 | #include <linux/reboot.h> |
b9cacc7b | 26 | #include <linux/vmstat.h> |
abe43400 | 27 | #include <linux/device.h> |
6e5fdeed | 28 | #include <linux/export.h> |
906010b2 | 29 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
30 | #include <linux/hardirq.h> |
31 | #include <linux/rculist.h> | |
0793a61d TG |
32 | #include <linux/uaccess.h> |
33 | #include <linux/syscalls.h> | |
34 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 35 | #include <linux/kernel_stat.h> |
cdd6c482 | 36 | #include <linux/perf_event.h> |
6fb2915d | 37 | #include <linux/ftrace_event.h> |
3c502e7a | 38 | #include <linux/hw_breakpoint.h> |
0793a61d | 39 | |
76369139 FW |
40 | #include "internal.h" |
41 | ||
4e193bd4 TB |
42 | #include <asm/irq_regs.h> |
43 | ||
fe4b04fa | 44 | struct remote_function_call { |
e7e7ee2e IM |
45 | struct task_struct *p; |
46 | int (*func)(void *info); | |
47 | void *info; | |
48 | int ret; | |
fe4b04fa PZ |
49 | }; |
50 | ||
51 | static void remote_function(void *data) | |
52 | { | |
53 | struct remote_function_call *tfc = data; | |
54 | struct task_struct *p = tfc->p; | |
55 | ||
56 | if (p) { | |
57 | tfc->ret = -EAGAIN; | |
58 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
59 | return; | |
60 | } | |
61 | ||
62 | tfc->ret = tfc->func(tfc->info); | |
63 | } | |
64 | ||
65 | /** | |
66 | * task_function_call - call a function on the cpu on which a task runs | |
67 | * @p: the task to evaluate | |
68 | * @func: the function to be called | |
69 | * @info: the function call argument | |
70 | * | |
71 | * Calls the function @func when the task is currently running. This might | |
72 | * be on the current CPU, which just calls the function directly | |
73 | * | |
74 | * returns: @func return value, or | |
75 | * -ESRCH - when the process isn't running | |
76 | * -EAGAIN - when the process moved away | |
77 | */ | |
78 | static int | |
79 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
80 | { | |
81 | struct remote_function_call data = { | |
e7e7ee2e IM |
82 | .p = p, |
83 | .func = func, | |
84 | .info = info, | |
85 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
86 | }; |
87 | ||
88 | if (task_curr(p)) | |
89 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
90 | ||
91 | return data.ret; | |
92 | } | |
93 | ||
94 | /** | |
95 | * cpu_function_call - call a function on the cpu | |
96 | * @func: the function to be called | |
97 | * @info: the function call argument | |
98 | * | |
99 | * Calls the function @func on the remote cpu. | |
100 | * | |
101 | * returns: @func return value or -ENXIO when the cpu is offline | |
102 | */ | |
103 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
104 | { | |
105 | struct remote_function_call data = { | |
e7e7ee2e IM |
106 | .p = NULL, |
107 | .func = func, | |
108 | .info = info, | |
109 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
110 | }; |
111 | ||
112 | smp_call_function_single(cpu, remote_function, &data, 1); | |
113 | ||
114 | return data.ret; | |
115 | } | |
116 | ||
e5d1367f SE |
117 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
118 | PERF_FLAG_FD_OUTPUT |\ | |
119 | PERF_FLAG_PID_CGROUP) | |
120 | ||
bce38cd5 SE |
121 | /* |
122 | * branch priv levels that need permission checks | |
123 | */ | |
124 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
125 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
126 | PERF_SAMPLE_BRANCH_HV) | |
127 | ||
0b3fcf17 SE |
128 | enum event_type_t { |
129 | EVENT_FLEXIBLE = 0x1, | |
130 | EVENT_PINNED = 0x2, | |
131 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
132 | }; | |
133 | ||
e5d1367f SE |
134 | /* |
135 | * perf_sched_events : >0 events exist | |
136 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
137 | */ | |
c5905afb | 138 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 139 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 140 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 141 | |
cdd6c482 IM |
142 | static atomic_t nr_mmap_events __read_mostly; |
143 | static atomic_t nr_comm_events __read_mostly; | |
144 | static atomic_t nr_task_events __read_mostly; | |
9ee318a7 | 145 | |
108b02cf PZ |
146 | static LIST_HEAD(pmus); |
147 | static DEFINE_MUTEX(pmus_lock); | |
148 | static struct srcu_struct pmus_srcu; | |
149 | ||
0764771d | 150 | /* |
cdd6c482 | 151 | * perf event paranoia level: |
0fbdea19 IM |
152 | * -1 - not paranoid at all |
153 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 154 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 155 | * 2 - disallow kernel profiling for unpriv |
0764771d | 156 | */ |
cdd6c482 | 157 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 158 | |
20443384 FW |
159 | /* Minimum for 512 kiB + 1 user control page */ |
160 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
161 | |
162 | /* | |
cdd6c482 | 163 | * max perf event sample rate |
df58ab24 | 164 | */ |
163ec435 PZ |
165 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
166 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
167 | static int max_samples_per_tick __read_mostly = | |
168 | DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
169 | ||
170 | int perf_proc_update_handler(struct ctl_table *table, int write, | |
171 | void __user *buffer, size_t *lenp, | |
172 | loff_t *ppos) | |
173 | { | |
174 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
175 | ||
176 | if (ret || !write) | |
177 | return ret; | |
178 | ||
179 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
180 | ||
181 | return 0; | |
182 | } | |
1ccd1549 | 183 | |
cdd6c482 | 184 | static atomic64_t perf_event_id; |
a96bbc16 | 185 | |
0b3fcf17 SE |
186 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
187 | enum event_type_t event_type); | |
188 | ||
189 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
190 | enum event_type_t event_type, |
191 | struct task_struct *task); | |
192 | ||
193 | static void update_context_time(struct perf_event_context *ctx); | |
194 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 195 | |
10c6db11 PZ |
196 | static void ring_buffer_attach(struct perf_event *event, |
197 | struct ring_buffer *rb); | |
198 | ||
cdd6c482 | 199 | void __weak perf_event_print_debug(void) { } |
0793a61d | 200 | |
84c79910 | 201 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 202 | { |
84c79910 | 203 | return "pmu"; |
0793a61d TG |
204 | } |
205 | ||
0b3fcf17 SE |
206 | static inline u64 perf_clock(void) |
207 | { | |
208 | return local_clock(); | |
209 | } | |
210 | ||
e5d1367f SE |
211 | static inline struct perf_cpu_context * |
212 | __get_cpu_context(struct perf_event_context *ctx) | |
213 | { | |
214 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
215 | } | |
216 | ||
facc4307 PZ |
217 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
218 | struct perf_event_context *ctx) | |
219 | { | |
220 | raw_spin_lock(&cpuctx->ctx.lock); | |
221 | if (ctx) | |
222 | raw_spin_lock(&ctx->lock); | |
223 | } | |
224 | ||
225 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
226 | struct perf_event_context *ctx) | |
227 | { | |
228 | if (ctx) | |
229 | raw_spin_unlock(&ctx->lock); | |
230 | raw_spin_unlock(&cpuctx->ctx.lock); | |
231 | } | |
232 | ||
e5d1367f SE |
233 | #ifdef CONFIG_CGROUP_PERF |
234 | ||
3f7cce3c SE |
235 | /* |
236 | * Must ensure cgroup is pinned (css_get) before calling | |
237 | * this function. In other words, we cannot call this function | |
238 | * if there is no cgroup event for the current CPU context. | |
239 | */ | |
e5d1367f SE |
240 | static inline struct perf_cgroup * |
241 | perf_cgroup_from_task(struct task_struct *task) | |
242 | { | |
243 | return container_of(task_subsys_state(task, perf_subsys_id), | |
244 | struct perf_cgroup, css); | |
245 | } | |
246 | ||
247 | static inline bool | |
248 | perf_cgroup_match(struct perf_event *event) | |
249 | { | |
250 | struct perf_event_context *ctx = event->ctx; | |
251 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
252 | ||
253 | return !event->cgrp || event->cgrp == cpuctx->cgrp; | |
254 | } | |
255 | ||
9c5da09d | 256 | static inline bool perf_tryget_cgroup(struct perf_event *event) |
e5d1367f | 257 | { |
9c5da09d | 258 | return css_tryget(&event->cgrp->css); |
e5d1367f SE |
259 | } |
260 | ||
261 | static inline void perf_put_cgroup(struct perf_event *event) | |
262 | { | |
263 | css_put(&event->cgrp->css); | |
264 | } | |
265 | ||
266 | static inline void perf_detach_cgroup(struct perf_event *event) | |
267 | { | |
268 | perf_put_cgroup(event); | |
269 | event->cgrp = NULL; | |
270 | } | |
271 | ||
272 | static inline int is_cgroup_event(struct perf_event *event) | |
273 | { | |
274 | return event->cgrp != NULL; | |
275 | } | |
276 | ||
277 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
278 | { | |
279 | struct perf_cgroup_info *t; | |
280 | ||
281 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
282 | return t->time; | |
283 | } | |
284 | ||
285 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
286 | { | |
287 | struct perf_cgroup_info *info; | |
288 | u64 now; | |
289 | ||
290 | now = perf_clock(); | |
291 | ||
292 | info = this_cpu_ptr(cgrp->info); | |
293 | ||
294 | info->time += now - info->timestamp; | |
295 | info->timestamp = now; | |
296 | } | |
297 | ||
298 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
299 | { | |
300 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
301 | if (cgrp_out) | |
302 | __update_cgrp_time(cgrp_out); | |
303 | } | |
304 | ||
305 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
306 | { | |
3f7cce3c SE |
307 | struct perf_cgroup *cgrp; |
308 | ||
e5d1367f | 309 | /* |
3f7cce3c SE |
310 | * ensure we access cgroup data only when needed and |
311 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 312 | */ |
3f7cce3c | 313 | if (!is_cgroup_event(event)) |
e5d1367f SE |
314 | return; |
315 | ||
3f7cce3c SE |
316 | cgrp = perf_cgroup_from_task(current); |
317 | /* | |
318 | * Do not update time when cgroup is not active | |
319 | */ | |
320 | if (cgrp == event->cgrp) | |
321 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
322 | } |
323 | ||
324 | static inline void | |
3f7cce3c SE |
325 | perf_cgroup_set_timestamp(struct task_struct *task, |
326 | struct perf_event_context *ctx) | |
e5d1367f SE |
327 | { |
328 | struct perf_cgroup *cgrp; | |
329 | struct perf_cgroup_info *info; | |
330 | ||
3f7cce3c SE |
331 | /* |
332 | * ctx->lock held by caller | |
333 | * ensure we do not access cgroup data | |
334 | * unless we have the cgroup pinned (css_get) | |
335 | */ | |
336 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
337 | return; |
338 | ||
339 | cgrp = perf_cgroup_from_task(task); | |
340 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 341 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
342 | } |
343 | ||
344 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
345 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
346 | ||
347 | /* | |
348 | * reschedule events based on the cgroup constraint of task. | |
349 | * | |
350 | * mode SWOUT : schedule out everything | |
351 | * mode SWIN : schedule in based on cgroup for next | |
352 | */ | |
353 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
354 | { | |
355 | struct perf_cpu_context *cpuctx; | |
356 | struct pmu *pmu; | |
357 | unsigned long flags; | |
358 | ||
359 | /* | |
360 | * disable interrupts to avoid geting nr_cgroup | |
361 | * changes via __perf_event_disable(). Also | |
362 | * avoids preemption. | |
363 | */ | |
364 | local_irq_save(flags); | |
365 | ||
366 | /* | |
367 | * we reschedule only in the presence of cgroup | |
368 | * constrained events. | |
369 | */ | |
370 | rcu_read_lock(); | |
371 | ||
372 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f SE |
373 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
374 | ||
e5d1367f SE |
375 | /* |
376 | * perf_cgroup_events says at least one | |
377 | * context on this CPU has cgroup events. | |
378 | * | |
379 | * ctx->nr_cgroups reports the number of cgroup | |
380 | * events for a context. | |
381 | */ | |
382 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
383 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
384 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
385 | |
386 | if (mode & PERF_CGROUP_SWOUT) { | |
387 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
388 | /* | |
389 | * must not be done before ctxswout due | |
390 | * to event_filter_match() in event_sched_out() | |
391 | */ | |
392 | cpuctx->cgrp = NULL; | |
393 | } | |
394 | ||
395 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 396 | WARN_ON_ONCE(cpuctx->cgrp); |
e5d1367f SE |
397 | /* set cgrp before ctxsw in to |
398 | * allow event_filter_match() to not | |
399 | * have to pass task around | |
400 | */ | |
401 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
402 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
403 | } | |
facc4307 PZ |
404 | perf_pmu_enable(cpuctx->ctx.pmu); |
405 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 406 | } |
e5d1367f SE |
407 | } |
408 | ||
409 | rcu_read_unlock(); | |
410 | ||
411 | local_irq_restore(flags); | |
412 | } | |
413 | ||
a8d757ef SE |
414 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
415 | struct task_struct *next) | |
e5d1367f | 416 | { |
a8d757ef SE |
417 | struct perf_cgroup *cgrp1; |
418 | struct perf_cgroup *cgrp2 = NULL; | |
419 | ||
420 | /* | |
421 | * we come here when we know perf_cgroup_events > 0 | |
422 | */ | |
423 | cgrp1 = perf_cgroup_from_task(task); | |
424 | ||
425 | /* | |
426 | * next is NULL when called from perf_event_enable_on_exec() | |
427 | * that will systematically cause a cgroup_switch() | |
428 | */ | |
429 | if (next) | |
430 | cgrp2 = perf_cgroup_from_task(next); | |
431 | ||
432 | /* | |
433 | * only schedule out current cgroup events if we know | |
434 | * that we are switching to a different cgroup. Otherwise, | |
435 | * do no touch the cgroup events. | |
436 | */ | |
437 | if (cgrp1 != cgrp2) | |
438 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
439 | } |
440 | ||
a8d757ef SE |
441 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
442 | struct task_struct *task) | |
e5d1367f | 443 | { |
a8d757ef SE |
444 | struct perf_cgroup *cgrp1; |
445 | struct perf_cgroup *cgrp2 = NULL; | |
446 | ||
447 | /* | |
448 | * we come here when we know perf_cgroup_events > 0 | |
449 | */ | |
450 | cgrp1 = perf_cgroup_from_task(task); | |
451 | ||
452 | /* prev can never be NULL */ | |
453 | cgrp2 = perf_cgroup_from_task(prev); | |
454 | ||
455 | /* | |
456 | * only need to schedule in cgroup events if we are changing | |
457 | * cgroup during ctxsw. Cgroup events were not scheduled | |
458 | * out of ctxsw out if that was not the case. | |
459 | */ | |
460 | if (cgrp1 != cgrp2) | |
461 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
462 | } |
463 | ||
464 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
465 | struct perf_event_attr *attr, | |
466 | struct perf_event *group_leader) | |
467 | { | |
468 | struct perf_cgroup *cgrp; | |
469 | struct cgroup_subsys_state *css; | |
470 | struct file *file; | |
471 | int ret = 0, fput_needed; | |
472 | ||
473 | file = fget_light(fd, &fput_needed); | |
474 | if (!file) | |
475 | return -EBADF; | |
476 | ||
477 | css = cgroup_css_from_dir(file, perf_subsys_id); | |
3db272c0 LZ |
478 | if (IS_ERR(css)) { |
479 | ret = PTR_ERR(css); | |
480 | goto out; | |
481 | } | |
e5d1367f SE |
482 | |
483 | cgrp = container_of(css, struct perf_cgroup, css); | |
484 | event->cgrp = cgrp; | |
485 | ||
f75e18cb | 486 | /* must be done before we fput() the file */ |
9c5da09d SQ |
487 | if (!perf_tryget_cgroup(event)) { |
488 | event->cgrp = NULL; | |
489 | ret = -ENOENT; | |
490 | goto out; | |
491 | } | |
f75e18cb | 492 | |
e5d1367f SE |
493 | /* |
494 | * all events in a group must monitor | |
495 | * the same cgroup because a task belongs | |
496 | * to only one perf cgroup at a time | |
497 | */ | |
498 | if (group_leader && group_leader->cgrp != cgrp) { | |
499 | perf_detach_cgroup(event); | |
500 | ret = -EINVAL; | |
e5d1367f | 501 | } |
3db272c0 | 502 | out: |
e5d1367f SE |
503 | fput_light(file, fput_needed); |
504 | return ret; | |
505 | } | |
506 | ||
507 | static inline void | |
508 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
509 | { | |
510 | struct perf_cgroup_info *t; | |
511 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
512 | event->shadow_ctx_time = now - t->timestamp; | |
513 | } | |
514 | ||
515 | static inline void | |
516 | perf_cgroup_defer_enabled(struct perf_event *event) | |
517 | { | |
518 | /* | |
519 | * when the current task's perf cgroup does not match | |
520 | * the event's, we need to remember to call the | |
521 | * perf_mark_enable() function the first time a task with | |
522 | * a matching perf cgroup is scheduled in. | |
523 | */ | |
524 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
525 | event->cgrp_defer_enabled = 1; | |
526 | } | |
527 | ||
528 | static inline void | |
529 | perf_cgroup_mark_enabled(struct perf_event *event, | |
530 | struct perf_event_context *ctx) | |
531 | { | |
532 | struct perf_event *sub; | |
533 | u64 tstamp = perf_event_time(event); | |
534 | ||
535 | if (!event->cgrp_defer_enabled) | |
536 | return; | |
537 | ||
538 | event->cgrp_defer_enabled = 0; | |
539 | ||
540 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
541 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
542 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
543 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
544 | sub->cgrp_defer_enabled = 0; | |
545 | } | |
546 | } | |
547 | } | |
548 | #else /* !CONFIG_CGROUP_PERF */ | |
549 | ||
550 | static inline bool | |
551 | perf_cgroup_match(struct perf_event *event) | |
552 | { | |
553 | return true; | |
554 | } | |
555 | ||
556 | static inline void perf_detach_cgroup(struct perf_event *event) | |
557 | {} | |
558 | ||
559 | static inline int is_cgroup_event(struct perf_event *event) | |
560 | { | |
561 | return 0; | |
562 | } | |
563 | ||
564 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
565 | { | |
566 | return 0; | |
567 | } | |
568 | ||
569 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
570 | { | |
571 | } | |
572 | ||
573 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
574 | { | |
575 | } | |
576 | ||
a8d757ef SE |
577 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
578 | struct task_struct *next) | |
e5d1367f SE |
579 | { |
580 | } | |
581 | ||
a8d757ef SE |
582 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
583 | struct task_struct *task) | |
e5d1367f SE |
584 | { |
585 | } | |
586 | ||
587 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
588 | struct perf_event_attr *attr, | |
589 | struct perf_event *group_leader) | |
590 | { | |
591 | return -EINVAL; | |
592 | } | |
593 | ||
594 | static inline void | |
3f7cce3c SE |
595 | perf_cgroup_set_timestamp(struct task_struct *task, |
596 | struct perf_event_context *ctx) | |
e5d1367f SE |
597 | { |
598 | } | |
599 | ||
600 | void | |
601 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
602 | { | |
603 | } | |
604 | ||
605 | static inline void | |
606 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
607 | { | |
608 | } | |
609 | ||
610 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
611 | { | |
612 | return 0; | |
613 | } | |
614 | ||
615 | static inline void | |
616 | perf_cgroup_defer_enabled(struct perf_event *event) | |
617 | { | |
618 | } | |
619 | ||
620 | static inline void | |
621 | perf_cgroup_mark_enabled(struct perf_event *event, | |
622 | struct perf_event_context *ctx) | |
623 | { | |
624 | } | |
625 | #endif | |
626 | ||
33696fc0 | 627 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 628 | { |
33696fc0 PZ |
629 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
630 | if (!(*count)++) | |
631 | pmu->pmu_disable(pmu); | |
9e35ad38 | 632 | } |
9e35ad38 | 633 | |
33696fc0 | 634 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 635 | { |
33696fc0 PZ |
636 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
637 | if (!--(*count)) | |
638 | pmu->pmu_enable(pmu); | |
9e35ad38 | 639 | } |
9e35ad38 | 640 | |
e9d2b064 PZ |
641 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
642 | ||
643 | /* | |
644 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
645 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
646 | * disabled, while rotate_context is called from IRQ context. | |
647 | */ | |
108b02cf | 648 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 649 | { |
108b02cf | 650 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 651 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 652 | |
e9d2b064 | 653 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 654 | |
e9d2b064 PZ |
655 | if (list_empty(&cpuctx->rotation_list)) |
656 | list_add(&cpuctx->rotation_list, head); | |
9e35ad38 | 657 | } |
9e35ad38 | 658 | |
cdd6c482 | 659 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 660 | { |
e5289d4a | 661 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
662 | } |
663 | ||
cdd6c482 | 664 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 665 | { |
564c2b21 PM |
666 | if (atomic_dec_and_test(&ctx->refcount)) { |
667 | if (ctx->parent_ctx) | |
668 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
669 | if (ctx->task) |
670 | put_task_struct(ctx->task); | |
cb796ff3 | 671 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 672 | } |
a63eaf34 PM |
673 | } |
674 | ||
cdd6c482 | 675 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
676 | { |
677 | if (ctx->parent_ctx) { | |
678 | put_ctx(ctx->parent_ctx); | |
679 | ctx->parent_ctx = NULL; | |
680 | } | |
681 | } | |
682 | ||
6844c09d ACM |
683 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
684 | { | |
685 | /* | |
686 | * only top level events have the pid namespace they were created in | |
687 | */ | |
688 | if (event->parent) | |
689 | event = event->parent; | |
690 | ||
691 | return task_tgid_nr_ns(p, event->ns); | |
692 | } | |
693 | ||
694 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
695 | { | |
696 | /* | |
697 | * only top level events have the pid namespace they were created in | |
698 | */ | |
699 | if (event->parent) | |
700 | event = event->parent; | |
701 | ||
702 | return task_pid_nr_ns(p, event->ns); | |
703 | } | |
704 | ||
7f453c24 | 705 | /* |
cdd6c482 | 706 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
707 | * to userspace. |
708 | */ | |
cdd6c482 | 709 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 710 | { |
cdd6c482 | 711 | u64 id = event->id; |
7f453c24 | 712 | |
cdd6c482 IM |
713 | if (event->parent) |
714 | id = event->parent->id; | |
7f453c24 PZ |
715 | |
716 | return id; | |
717 | } | |
718 | ||
25346b93 | 719 | /* |
cdd6c482 | 720 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
721 | * This has to cope with with the fact that until it is locked, |
722 | * the context could get moved to another task. | |
723 | */ | |
cdd6c482 | 724 | static struct perf_event_context * |
8dc85d54 | 725 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 726 | { |
cdd6c482 | 727 | struct perf_event_context *ctx; |
25346b93 PM |
728 | |
729 | rcu_read_lock(); | |
9ed6060d | 730 | retry: |
8dc85d54 | 731 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
732 | if (ctx) { |
733 | /* | |
734 | * If this context is a clone of another, it might | |
735 | * get swapped for another underneath us by | |
cdd6c482 | 736 | * perf_event_task_sched_out, though the |
25346b93 PM |
737 | * rcu_read_lock() protects us from any context |
738 | * getting freed. Lock the context and check if it | |
739 | * got swapped before we could get the lock, and retry | |
740 | * if so. If we locked the right context, then it | |
741 | * can't get swapped on us any more. | |
742 | */ | |
e625cce1 | 743 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 744 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 745 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
25346b93 PM |
746 | goto retry; |
747 | } | |
b49a9e7e PZ |
748 | |
749 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 750 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
751 | ctx = NULL; |
752 | } | |
25346b93 PM |
753 | } |
754 | rcu_read_unlock(); | |
755 | return ctx; | |
756 | } | |
757 | ||
758 | /* | |
759 | * Get the context for a task and increment its pin_count so it | |
760 | * can't get swapped to another task. This also increments its | |
761 | * reference count so that the context can't get freed. | |
762 | */ | |
8dc85d54 PZ |
763 | static struct perf_event_context * |
764 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 765 | { |
cdd6c482 | 766 | struct perf_event_context *ctx; |
25346b93 PM |
767 | unsigned long flags; |
768 | ||
8dc85d54 | 769 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
770 | if (ctx) { |
771 | ++ctx->pin_count; | |
e625cce1 | 772 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
773 | } |
774 | return ctx; | |
775 | } | |
776 | ||
cdd6c482 | 777 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
778 | { |
779 | unsigned long flags; | |
780 | ||
e625cce1 | 781 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 782 | --ctx->pin_count; |
e625cce1 | 783 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
784 | } |
785 | ||
f67218c3 PZ |
786 | /* |
787 | * Update the record of the current time in a context. | |
788 | */ | |
789 | static void update_context_time(struct perf_event_context *ctx) | |
790 | { | |
791 | u64 now = perf_clock(); | |
792 | ||
793 | ctx->time += now - ctx->timestamp; | |
794 | ctx->timestamp = now; | |
795 | } | |
796 | ||
4158755d SE |
797 | static u64 perf_event_time(struct perf_event *event) |
798 | { | |
799 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
800 | |
801 | if (is_cgroup_event(event)) | |
802 | return perf_cgroup_event_time(event); | |
803 | ||
4158755d SE |
804 | return ctx ? ctx->time : 0; |
805 | } | |
806 | ||
f67218c3 PZ |
807 | /* |
808 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 809 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
810 | */ |
811 | static void update_event_times(struct perf_event *event) | |
812 | { | |
813 | struct perf_event_context *ctx = event->ctx; | |
814 | u64 run_end; | |
815 | ||
816 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
817 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
818 | return; | |
e5d1367f SE |
819 | /* |
820 | * in cgroup mode, time_enabled represents | |
821 | * the time the event was enabled AND active | |
822 | * tasks were in the monitored cgroup. This is | |
823 | * independent of the activity of the context as | |
824 | * there may be a mix of cgroup and non-cgroup events. | |
825 | * | |
826 | * That is why we treat cgroup events differently | |
827 | * here. | |
828 | */ | |
829 | if (is_cgroup_event(event)) | |
46cd6a7f | 830 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
831 | else if (ctx->is_active) |
832 | run_end = ctx->time; | |
acd1d7c1 PZ |
833 | else |
834 | run_end = event->tstamp_stopped; | |
835 | ||
836 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
837 | |
838 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
839 | run_end = event->tstamp_stopped; | |
840 | else | |
4158755d | 841 | run_end = perf_event_time(event); |
f67218c3 PZ |
842 | |
843 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 844 | |
f67218c3 PZ |
845 | } |
846 | ||
96c21a46 PZ |
847 | /* |
848 | * Update total_time_enabled and total_time_running for all events in a group. | |
849 | */ | |
850 | static void update_group_times(struct perf_event *leader) | |
851 | { | |
852 | struct perf_event *event; | |
853 | ||
854 | update_event_times(leader); | |
855 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
856 | update_event_times(event); | |
857 | } | |
858 | ||
889ff015 FW |
859 | static struct list_head * |
860 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
861 | { | |
862 | if (event->attr.pinned) | |
863 | return &ctx->pinned_groups; | |
864 | else | |
865 | return &ctx->flexible_groups; | |
866 | } | |
867 | ||
fccc714b | 868 | /* |
cdd6c482 | 869 | * Add a event from the lists for its context. |
fccc714b PZ |
870 | * Must be called with ctx->mutex and ctx->lock held. |
871 | */ | |
04289bb9 | 872 | static void |
cdd6c482 | 873 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 874 | { |
8a49542c PZ |
875 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
876 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
877 | |
878 | /* | |
8a49542c PZ |
879 | * If we're a stand alone event or group leader, we go to the context |
880 | * list, group events are kept attached to the group so that | |
881 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 882 | */ |
8a49542c | 883 | if (event->group_leader == event) { |
889ff015 FW |
884 | struct list_head *list; |
885 | ||
d6f962b5 FW |
886 | if (is_software_event(event)) |
887 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
888 | ||
889ff015 FW |
889 | list = ctx_group_list(event, ctx); |
890 | list_add_tail(&event->group_entry, list); | |
5c148194 | 891 | } |
592903cd | 892 | |
08309379 | 893 | if (is_cgroup_event(event)) |
e5d1367f | 894 | ctx->nr_cgroups++; |
e5d1367f | 895 | |
d010b332 SE |
896 | if (has_branch_stack(event)) |
897 | ctx->nr_branch_stack++; | |
898 | ||
cdd6c482 | 899 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 900 | if (!ctx->nr_events) |
108b02cf | 901 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
902 | ctx->nr_events++; |
903 | if (event->attr.inherit_stat) | |
bfbd3381 | 904 | ctx->nr_stat++; |
04289bb9 IM |
905 | } |
906 | ||
c320c7b7 ACM |
907 | /* |
908 | * Called at perf_event creation and when events are attached/detached from a | |
909 | * group. | |
910 | */ | |
911 | static void perf_event__read_size(struct perf_event *event) | |
912 | { | |
913 | int entry = sizeof(u64); /* value */ | |
914 | int size = 0; | |
915 | int nr = 1; | |
916 | ||
917 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
918 | size += sizeof(u64); | |
919 | ||
920 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
921 | size += sizeof(u64); | |
922 | ||
923 | if (event->attr.read_format & PERF_FORMAT_ID) | |
924 | entry += sizeof(u64); | |
925 | ||
926 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
927 | nr += event->group_leader->nr_siblings; | |
928 | size += sizeof(u64); | |
929 | } | |
930 | ||
931 | size += entry * nr; | |
932 | event->read_size = size; | |
933 | } | |
934 | ||
935 | static void perf_event__header_size(struct perf_event *event) | |
936 | { | |
937 | struct perf_sample_data *data; | |
938 | u64 sample_type = event->attr.sample_type; | |
939 | u16 size = 0; | |
940 | ||
941 | perf_event__read_size(event); | |
942 | ||
943 | if (sample_type & PERF_SAMPLE_IP) | |
944 | size += sizeof(data->ip); | |
945 | ||
6844c09d ACM |
946 | if (sample_type & PERF_SAMPLE_ADDR) |
947 | size += sizeof(data->addr); | |
948 | ||
949 | if (sample_type & PERF_SAMPLE_PERIOD) | |
950 | size += sizeof(data->period); | |
951 | ||
952 | if (sample_type & PERF_SAMPLE_READ) | |
953 | size += event->read_size; | |
954 | ||
955 | event->header_size = size; | |
956 | } | |
957 | ||
958 | static void perf_event__id_header_size(struct perf_event *event) | |
959 | { | |
960 | struct perf_sample_data *data; | |
961 | u64 sample_type = event->attr.sample_type; | |
962 | u16 size = 0; | |
963 | ||
c320c7b7 ACM |
964 | if (sample_type & PERF_SAMPLE_TID) |
965 | size += sizeof(data->tid_entry); | |
966 | ||
967 | if (sample_type & PERF_SAMPLE_TIME) | |
968 | size += sizeof(data->time); | |
969 | ||
c320c7b7 ACM |
970 | if (sample_type & PERF_SAMPLE_ID) |
971 | size += sizeof(data->id); | |
972 | ||
973 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
974 | size += sizeof(data->stream_id); | |
975 | ||
976 | if (sample_type & PERF_SAMPLE_CPU) | |
977 | size += sizeof(data->cpu_entry); | |
978 | ||
6844c09d | 979 | event->id_header_size = size; |
c320c7b7 ACM |
980 | } |
981 | ||
8a49542c PZ |
982 | static void perf_group_attach(struct perf_event *event) |
983 | { | |
c320c7b7 | 984 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 985 | |
74c3337c PZ |
986 | /* |
987 | * We can have double attach due to group movement in perf_event_open. | |
988 | */ | |
989 | if (event->attach_state & PERF_ATTACH_GROUP) | |
990 | return; | |
991 | ||
8a49542c PZ |
992 | event->attach_state |= PERF_ATTACH_GROUP; |
993 | ||
994 | if (group_leader == event) | |
995 | return; | |
996 | ||
997 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
998 | !is_software_event(event)) | |
999 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1000 | ||
1001 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1002 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1003 | |
1004 | perf_event__header_size(group_leader); | |
1005 | ||
1006 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1007 | perf_event__header_size(pos); | |
8a49542c PZ |
1008 | } |
1009 | ||
a63eaf34 | 1010 | /* |
cdd6c482 | 1011 | * Remove a event from the lists for its context. |
fccc714b | 1012 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1013 | */ |
04289bb9 | 1014 | static void |
cdd6c482 | 1015 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1016 | { |
68cacd29 | 1017 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1018 | /* |
1019 | * We can have double detach due to exit/hot-unplug + close. | |
1020 | */ | |
1021 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1022 | return; |
8a49542c PZ |
1023 | |
1024 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1025 | ||
68cacd29 | 1026 | if (is_cgroup_event(event)) { |
e5d1367f | 1027 | ctx->nr_cgroups--; |
68cacd29 SE |
1028 | cpuctx = __get_cpu_context(ctx); |
1029 | /* | |
1030 | * if there are no more cgroup events | |
1031 | * then cler cgrp to avoid stale pointer | |
1032 | * in update_cgrp_time_from_cpuctx() | |
1033 | */ | |
1034 | if (!ctx->nr_cgroups) | |
1035 | cpuctx->cgrp = NULL; | |
1036 | } | |
e5d1367f | 1037 | |
d010b332 SE |
1038 | if (has_branch_stack(event)) |
1039 | ctx->nr_branch_stack--; | |
1040 | ||
cdd6c482 IM |
1041 | ctx->nr_events--; |
1042 | if (event->attr.inherit_stat) | |
bfbd3381 | 1043 | ctx->nr_stat--; |
8bc20959 | 1044 | |
cdd6c482 | 1045 | list_del_rcu(&event->event_entry); |
04289bb9 | 1046 | |
8a49542c PZ |
1047 | if (event->group_leader == event) |
1048 | list_del_init(&event->group_entry); | |
5c148194 | 1049 | |
96c21a46 | 1050 | update_group_times(event); |
b2e74a26 SE |
1051 | |
1052 | /* | |
1053 | * If event was in error state, then keep it | |
1054 | * that way, otherwise bogus counts will be | |
1055 | * returned on read(). The only way to get out | |
1056 | * of error state is by explicit re-enabling | |
1057 | * of the event | |
1058 | */ | |
1059 | if (event->state > PERF_EVENT_STATE_OFF) | |
1060 | event->state = PERF_EVENT_STATE_OFF; | |
050735b0 PZ |
1061 | } |
1062 | ||
8a49542c | 1063 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1064 | { |
1065 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1066 | struct list_head *list = NULL; |
1067 | ||
1068 | /* | |
1069 | * We can have double detach due to exit/hot-unplug + close. | |
1070 | */ | |
1071 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1072 | return; | |
1073 | ||
1074 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1075 | ||
1076 | /* | |
1077 | * If this is a sibling, remove it from its group. | |
1078 | */ | |
1079 | if (event->group_leader != event) { | |
1080 | list_del_init(&event->group_entry); | |
1081 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1082 | goto out; |
8a49542c PZ |
1083 | } |
1084 | ||
1085 | if (!list_empty(&event->group_entry)) | |
1086 | list = &event->group_entry; | |
2e2af50b | 1087 | |
04289bb9 | 1088 | /* |
cdd6c482 IM |
1089 | * If this was a group event with sibling events then |
1090 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1091 | * to whatever list we are on. |
04289bb9 | 1092 | */ |
cdd6c482 | 1093 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1094 | if (list) |
1095 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1096 | sibling->group_leader = sibling; |
d6f962b5 FW |
1097 | |
1098 | /* Inherit group flags from the previous leader */ | |
1099 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1100 | } |
c320c7b7 ACM |
1101 | |
1102 | out: | |
1103 | perf_event__header_size(event->group_leader); | |
1104 | ||
1105 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1106 | perf_event__header_size(tmp); | |
04289bb9 IM |
1107 | } |
1108 | ||
fa66f07a SE |
1109 | static inline int |
1110 | event_filter_match(struct perf_event *event) | |
1111 | { | |
e5d1367f SE |
1112 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1113 | && perf_cgroup_match(event); | |
fa66f07a SE |
1114 | } |
1115 | ||
9ffcfa6f SE |
1116 | static void |
1117 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1118 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1119 | struct perf_event_context *ctx) |
3b6f9e5c | 1120 | { |
4158755d | 1121 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1122 | u64 delta; |
1123 | /* | |
1124 | * An event which could not be activated because of | |
1125 | * filter mismatch still needs to have its timings | |
1126 | * maintained, otherwise bogus information is return | |
1127 | * via read() for time_enabled, time_running: | |
1128 | */ | |
1129 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1130 | && !event_filter_match(event)) { | |
e5d1367f | 1131 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1132 | event->tstamp_running += delta; |
4158755d | 1133 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1134 | } |
1135 | ||
cdd6c482 | 1136 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1137 | return; |
3b6f9e5c | 1138 | |
cdd6c482 IM |
1139 | event->state = PERF_EVENT_STATE_INACTIVE; |
1140 | if (event->pending_disable) { | |
1141 | event->pending_disable = 0; | |
1142 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1143 | } |
4158755d | 1144 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1145 | event->pmu->del(event, 0); |
cdd6c482 | 1146 | event->oncpu = -1; |
3b6f9e5c | 1147 | |
cdd6c482 | 1148 | if (!is_software_event(event)) |
3b6f9e5c PM |
1149 | cpuctx->active_oncpu--; |
1150 | ctx->nr_active--; | |
0f5a2601 PZ |
1151 | if (event->attr.freq && event->attr.sample_freq) |
1152 | ctx->nr_freq--; | |
cdd6c482 | 1153 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c PM |
1154 | cpuctx->exclusive = 0; |
1155 | } | |
1156 | ||
d859e29f | 1157 | static void |
cdd6c482 | 1158 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1159 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1160 | struct perf_event_context *ctx) |
d859e29f | 1161 | { |
cdd6c482 | 1162 | struct perf_event *event; |
fa66f07a | 1163 | int state = group_event->state; |
d859e29f | 1164 | |
cdd6c482 | 1165 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1166 | |
1167 | /* | |
1168 | * Schedule out siblings (if any): | |
1169 | */ | |
cdd6c482 IM |
1170 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1171 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1172 | |
fa66f07a | 1173 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1174 | cpuctx->exclusive = 0; |
1175 | } | |
1176 | ||
0793a61d | 1177 | /* |
cdd6c482 | 1178 | * Cross CPU call to remove a performance event |
0793a61d | 1179 | * |
cdd6c482 | 1180 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1181 | * remove it from the context list. |
1182 | */ | |
fe4b04fa | 1183 | static int __perf_remove_from_context(void *info) |
0793a61d | 1184 | { |
cdd6c482 IM |
1185 | struct perf_event *event = info; |
1186 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1187 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1188 | |
e625cce1 | 1189 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1190 | event_sched_out(event, cpuctx, ctx); |
cdd6c482 | 1191 | list_del_event(event, ctx); |
64ce3126 PZ |
1192 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1193 | ctx->is_active = 0; | |
1194 | cpuctx->task_ctx = NULL; | |
1195 | } | |
e625cce1 | 1196 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1197 | |
1198 | return 0; | |
0793a61d TG |
1199 | } |
1200 | ||
1201 | ||
1202 | /* | |
cdd6c482 | 1203 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1204 | * |
cdd6c482 | 1205 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1206 | * call when the task is on a CPU. |
c93f7669 | 1207 | * |
cdd6c482 IM |
1208 | * If event->ctx is a cloned context, callers must make sure that |
1209 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1210 | * remains valid. This is OK when called from perf_release since |
1211 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1212 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1213 | * context has been detached from its task. |
0793a61d | 1214 | */ |
fe4b04fa | 1215 | static void perf_remove_from_context(struct perf_event *event) |
0793a61d | 1216 | { |
cdd6c482 | 1217 | struct perf_event_context *ctx = event->ctx; |
0793a61d TG |
1218 | struct task_struct *task = ctx->task; |
1219 | ||
fe4b04fa PZ |
1220 | lockdep_assert_held(&ctx->mutex); |
1221 | ||
0793a61d TG |
1222 | if (!task) { |
1223 | /* | |
cdd6c482 | 1224 | * Per cpu events are removed via an smp call and |
af901ca1 | 1225 | * the removal is always successful. |
0793a61d | 1226 | */ |
fe4b04fa | 1227 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
0793a61d TG |
1228 | return; |
1229 | } | |
1230 | ||
1231 | retry: | |
fe4b04fa PZ |
1232 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1233 | return; | |
0793a61d | 1234 | |
e625cce1 | 1235 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1236 | /* |
fe4b04fa PZ |
1237 | * If we failed to find a running task, but find the context active now |
1238 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1239 | */ |
fe4b04fa | 1240 | if (ctx->is_active) { |
e625cce1 | 1241 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1242 | goto retry; |
1243 | } | |
1244 | ||
1245 | /* | |
fe4b04fa PZ |
1246 | * Since the task isn't running, its safe to remove the event, us |
1247 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1248 | */ |
fe4b04fa | 1249 | list_del_event(event, ctx); |
e625cce1 | 1250 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1251 | } |
1252 | ||
d859e29f | 1253 | /* |
cdd6c482 | 1254 | * Cross CPU call to disable a performance event |
d859e29f | 1255 | */ |
fe4b04fa | 1256 | static int __perf_event_disable(void *info) |
d859e29f | 1257 | { |
cdd6c482 | 1258 | struct perf_event *event = info; |
cdd6c482 | 1259 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1260 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1261 | |
1262 | /* | |
cdd6c482 IM |
1263 | * If this is a per-task event, need to check whether this |
1264 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1265 | * |
1266 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1267 | * flipping contexts around. | |
d859e29f | 1268 | */ |
665c2142 | 1269 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1270 | return -EINVAL; |
d859e29f | 1271 | |
e625cce1 | 1272 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1273 | |
1274 | /* | |
cdd6c482 | 1275 | * If the event is on, turn it off. |
d859e29f PM |
1276 | * If it is in error state, leave it in error state. |
1277 | */ | |
cdd6c482 | 1278 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1279 | update_context_time(ctx); |
e5d1367f | 1280 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1281 | update_group_times(event); |
1282 | if (event == event->group_leader) | |
1283 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1284 | else |
cdd6c482 IM |
1285 | event_sched_out(event, cpuctx, ctx); |
1286 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1287 | } |
1288 | ||
e625cce1 | 1289 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1290 | |
1291 | return 0; | |
d859e29f PM |
1292 | } |
1293 | ||
1294 | /* | |
cdd6c482 | 1295 | * Disable a event. |
c93f7669 | 1296 | * |
cdd6c482 IM |
1297 | * If event->ctx is a cloned context, callers must make sure that |
1298 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1299 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1300 | * perf_event_for_each_child or perf_event_for_each because they |
1301 | * hold the top-level event's child_mutex, so any descendant that | |
1302 | * goes to exit will block in sync_child_event. | |
1303 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1304 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1305 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1306 | */ |
44234adc | 1307 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1308 | { |
cdd6c482 | 1309 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1310 | struct task_struct *task = ctx->task; |
1311 | ||
1312 | if (!task) { | |
1313 | /* | |
cdd6c482 | 1314 | * Disable the event on the cpu that it's on |
d859e29f | 1315 | */ |
fe4b04fa | 1316 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1317 | return; |
1318 | } | |
1319 | ||
9ed6060d | 1320 | retry: |
fe4b04fa PZ |
1321 | if (!task_function_call(task, __perf_event_disable, event)) |
1322 | return; | |
d859e29f | 1323 | |
e625cce1 | 1324 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1325 | /* |
cdd6c482 | 1326 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1327 | */ |
cdd6c482 | 1328 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1329 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1330 | /* |
1331 | * Reload the task pointer, it might have been changed by | |
1332 | * a concurrent perf_event_context_sched_out(). | |
1333 | */ | |
1334 | task = ctx->task; | |
d859e29f PM |
1335 | goto retry; |
1336 | } | |
1337 | ||
1338 | /* | |
1339 | * Since we have the lock this context can't be scheduled | |
1340 | * in, so we can change the state safely. | |
1341 | */ | |
cdd6c482 IM |
1342 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1343 | update_group_times(event); | |
1344 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1345 | } |
e625cce1 | 1346 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1347 | } |
dcfce4a0 | 1348 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1349 | |
e5d1367f SE |
1350 | static void perf_set_shadow_time(struct perf_event *event, |
1351 | struct perf_event_context *ctx, | |
1352 | u64 tstamp) | |
1353 | { | |
1354 | /* | |
1355 | * use the correct time source for the time snapshot | |
1356 | * | |
1357 | * We could get by without this by leveraging the | |
1358 | * fact that to get to this function, the caller | |
1359 | * has most likely already called update_context_time() | |
1360 | * and update_cgrp_time_xx() and thus both timestamp | |
1361 | * are identical (or very close). Given that tstamp is, | |
1362 | * already adjusted for cgroup, we could say that: | |
1363 | * tstamp - ctx->timestamp | |
1364 | * is equivalent to | |
1365 | * tstamp - cgrp->timestamp. | |
1366 | * | |
1367 | * Then, in perf_output_read(), the calculation would | |
1368 | * work with no changes because: | |
1369 | * - event is guaranteed scheduled in | |
1370 | * - no scheduled out in between | |
1371 | * - thus the timestamp would be the same | |
1372 | * | |
1373 | * But this is a bit hairy. | |
1374 | * | |
1375 | * So instead, we have an explicit cgroup call to remain | |
1376 | * within the time time source all along. We believe it | |
1377 | * is cleaner and simpler to understand. | |
1378 | */ | |
1379 | if (is_cgroup_event(event)) | |
1380 | perf_cgroup_set_shadow_time(event, tstamp); | |
1381 | else | |
1382 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1383 | } | |
1384 | ||
4fe757dd PZ |
1385 | #define MAX_INTERRUPTS (~0ULL) |
1386 | ||
1387 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1388 | ||
235c7fc7 | 1389 | static int |
9ffcfa6f | 1390 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1391 | struct perf_cpu_context *cpuctx, |
6e37738a | 1392 | struct perf_event_context *ctx) |
235c7fc7 | 1393 | { |
4158755d SE |
1394 | u64 tstamp = perf_event_time(event); |
1395 | ||
cdd6c482 | 1396 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1397 | return 0; |
1398 | ||
cdd6c482 | 1399 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1400 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1401 | |
1402 | /* | |
1403 | * Unthrottle events, since we scheduled we might have missed several | |
1404 | * ticks already, also for a heavily scheduling task there is little | |
1405 | * guarantee it'll get a tick in a timely manner. | |
1406 | */ | |
1407 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1408 | perf_log_throttle(event, 1); | |
1409 | event->hw.interrupts = 0; | |
1410 | } | |
1411 | ||
235c7fc7 IM |
1412 | /* |
1413 | * The new state must be visible before we turn it on in the hardware: | |
1414 | */ | |
1415 | smp_wmb(); | |
1416 | ||
a4eaf7f1 | 1417 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1418 | event->state = PERF_EVENT_STATE_INACTIVE; |
1419 | event->oncpu = -1; | |
235c7fc7 IM |
1420 | return -EAGAIN; |
1421 | } | |
1422 | ||
4158755d | 1423 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1424 | |
e5d1367f | 1425 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1426 | |
cdd6c482 | 1427 | if (!is_software_event(event)) |
3b6f9e5c | 1428 | cpuctx->active_oncpu++; |
235c7fc7 | 1429 | ctx->nr_active++; |
0f5a2601 PZ |
1430 | if (event->attr.freq && event->attr.sample_freq) |
1431 | ctx->nr_freq++; | |
235c7fc7 | 1432 | |
cdd6c482 | 1433 | if (event->attr.exclusive) |
3b6f9e5c PM |
1434 | cpuctx->exclusive = 1; |
1435 | ||
235c7fc7 IM |
1436 | return 0; |
1437 | } | |
1438 | ||
6751b71e | 1439 | static int |
cdd6c482 | 1440 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1441 | struct perf_cpu_context *cpuctx, |
6e37738a | 1442 | struct perf_event_context *ctx) |
6751b71e | 1443 | { |
6bde9b6c | 1444 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1445 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1446 | u64 now = ctx->time; |
1447 | bool simulate = false; | |
6751b71e | 1448 | |
cdd6c482 | 1449 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1450 | return 0; |
1451 | ||
ad5133b7 | 1452 | pmu->start_txn(pmu); |
6bde9b6c | 1453 | |
9ffcfa6f | 1454 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1455 | pmu->cancel_txn(pmu); |
6751b71e | 1456 | return -EAGAIN; |
90151c35 | 1457 | } |
6751b71e PM |
1458 | |
1459 | /* | |
1460 | * Schedule in siblings as one group (if any): | |
1461 | */ | |
cdd6c482 | 1462 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1463 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1464 | partial_group = event; |
6751b71e PM |
1465 | goto group_error; |
1466 | } | |
1467 | } | |
1468 | ||
9ffcfa6f | 1469 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1470 | return 0; |
9ffcfa6f | 1471 | |
6751b71e PM |
1472 | group_error: |
1473 | /* | |
1474 | * Groups can be scheduled in as one unit only, so undo any | |
1475 | * partial group before returning: | |
d7842da4 SE |
1476 | * The events up to the failed event are scheduled out normally, |
1477 | * tstamp_stopped will be updated. | |
1478 | * | |
1479 | * The failed events and the remaining siblings need to have | |
1480 | * their timings updated as if they had gone thru event_sched_in() | |
1481 | * and event_sched_out(). This is required to get consistent timings | |
1482 | * across the group. This also takes care of the case where the group | |
1483 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1484 | * the time the event was actually stopped, such that time delta | |
1485 | * calculation in update_event_times() is correct. | |
6751b71e | 1486 | */ |
cdd6c482 IM |
1487 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1488 | if (event == partial_group) | |
d7842da4 SE |
1489 | simulate = true; |
1490 | ||
1491 | if (simulate) { | |
1492 | event->tstamp_running += now - event->tstamp_stopped; | |
1493 | event->tstamp_stopped = now; | |
1494 | } else { | |
1495 | event_sched_out(event, cpuctx, ctx); | |
1496 | } | |
6751b71e | 1497 | } |
9ffcfa6f | 1498 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1499 | |
ad5133b7 | 1500 | pmu->cancel_txn(pmu); |
90151c35 | 1501 | |
6751b71e PM |
1502 | return -EAGAIN; |
1503 | } | |
1504 | ||
3b6f9e5c | 1505 | /* |
cdd6c482 | 1506 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1507 | */ |
cdd6c482 | 1508 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1509 | struct perf_cpu_context *cpuctx, |
1510 | int can_add_hw) | |
1511 | { | |
1512 | /* | |
cdd6c482 | 1513 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1514 | */ |
d6f962b5 | 1515 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1516 | return 1; |
1517 | /* | |
1518 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1519 | * events can go on. |
3b6f9e5c PM |
1520 | */ |
1521 | if (cpuctx->exclusive) | |
1522 | return 0; | |
1523 | /* | |
1524 | * If this group is exclusive and there are already | |
cdd6c482 | 1525 | * events on the CPU, it can't go on. |
3b6f9e5c | 1526 | */ |
cdd6c482 | 1527 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1528 | return 0; |
1529 | /* | |
1530 | * Otherwise, try to add it if all previous groups were able | |
1531 | * to go on. | |
1532 | */ | |
1533 | return can_add_hw; | |
1534 | } | |
1535 | ||
cdd6c482 IM |
1536 | static void add_event_to_ctx(struct perf_event *event, |
1537 | struct perf_event_context *ctx) | |
53cfbf59 | 1538 | { |
4158755d SE |
1539 | u64 tstamp = perf_event_time(event); |
1540 | ||
cdd6c482 | 1541 | list_add_event(event, ctx); |
8a49542c | 1542 | perf_group_attach(event); |
4158755d SE |
1543 | event->tstamp_enabled = tstamp; |
1544 | event->tstamp_running = tstamp; | |
1545 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1546 | } |
1547 | ||
2c29ef0f PZ |
1548 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1549 | static void | |
1550 | ctx_sched_in(struct perf_event_context *ctx, | |
1551 | struct perf_cpu_context *cpuctx, | |
1552 | enum event_type_t event_type, | |
1553 | struct task_struct *task); | |
fe4b04fa | 1554 | |
dce5855b PZ |
1555 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1556 | struct perf_event_context *ctx, | |
1557 | struct task_struct *task) | |
1558 | { | |
1559 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1560 | if (ctx) | |
1561 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1562 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1563 | if (ctx) | |
1564 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1565 | } | |
1566 | ||
0793a61d | 1567 | /* |
cdd6c482 | 1568 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1569 | * |
1570 | * Must be called with ctx->mutex held | |
0793a61d | 1571 | */ |
fe4b04fa | 1572 | static int __perf_install_in_context(void *info) |
0793a61d | 1573 | { |
cdd6c482 IM |
1574 | struct perf_event *event = info; |
1575 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1576 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1577 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1578 | struct task_struct *task = current; | |
1579 | ||
b58f6b0d | 1580 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1581 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1582 | |
1583 | /* | |
2c29ef0f | 1584 | * If there was an active task_ctx schedule it out. |
0793a61d | 1585 | */ |
b58f6b0d | 1586 | if (task_ctx) |
2c29ef0f | 1587 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1588 | |
1589 | /* | |
1590 | * If the context we're installing events in is not the | |
1591 | * active task_ctx, flip them. | |
1592 | */ | |
1593 | if (ctx->task && task_ctx != ctx) { | |
1594 | if (task_ctx) | |
1595 | raw_spin_unlock(&task_ctx->lock); | |
1596 | raw_spin_lock(&ctx->lock); | |
1597 | task_ctx = ctx; | |
1598 | } | |
1599 | ||
1600 | if (task_ctx) { | |
1601 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1602 | task = task_ctx->task; |
1603 | } | |
b58f6b0d | 1604 | |
2c29ef0f | 1605 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1606 | |
4af4998b | 1607 | update_context_time(ctx); |
e5d1367f SE |
1608 | /* |
1609 | * update cgrp time only if current cgrp | |
1610 | * matches event->cgrp. Must be done before | |
1611 | * calling add_event_to_ctx() | |
1612 | */ | |
1613 | update_cgrp_time_from_event(event); | |
0793a61d | 1614 | |
cdd6c482 | 1615 | add_event_to_ctx(event, ctx); |
0793a61d | 1616 | |
d859e29f | 1617 | /* |
2c29ef0f | 1618 | * Schedule everything back in |
d859e29f | 1619 | */ |
dce5855b | 1620 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1621 | |
1622 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1623 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1624 | |
1625 | return 0; | |
0793a61d TG |
1626 | } |
1627 | ||
1628 | /* | |
cdd6c482 | 1629 | * Attach a performance event to a context |
0793a61d | 1630 | * |
cdd6c482 IM |
1631 | * First we add the event to the list with the hardware enable bit |
1632 | * in event->hw_config cleared. | |
0793a61d | 1633 | * |
cdd6c482 | 1634 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1635 | * call to enable it in the task context. The task might have been |
1636 | * scheduled away, but we check this in the smp call again. | |
1637 | */ | |
1638 | static void | |
cdd6c482 IM |
1639 | perf_install_in_context(struct perf_event_context *ctx, |
1640 | struct perf_event *event, | |
0793a61d TG |
1641 | int cpu) |
1642 | { | |
1643 | struct task_struct *task = ctx->task; | |
1644 | ||
fe4b04fa PZ |
1645 | lockdep_assert_held(&ctx->mutex); |
1646 | ||
c3f00c70 | 1647 | event->ctx = ctx; |
0cda4c02 YZ |
1648 | if (event->cpu != -1) |
1649 | event->cpu = cpu; | |
c3f00c70 | 1650 | |
0793a61d TG |
1651 | if (!task) { |
1652 | /* | |
cdd6c482 | 1653 | * Per cpu events are installed via an smp call and |
af901ca1 | 1654 | * the install is always successful. |
0793a61d | 1655 | */ |
fe4b04fa | 1656 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1657 | return; |
1658 | } | |
1659 | ||
0793a61d | 1660 | retry: |
fe4b04fa PZ |
1661 | if (!task_function_call(task, __perf_install_in_context, event)) |
1662 | return; | |
0793a61d | 1663 | |
e625cce1 | 1664 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1665 | /* |
fe4b04fa PZ |
1666 | * If we failed to find a running task, but find the context active now |
1667 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1668 | */ |
fe4b04fa | 1669 | if (ctx->is_active) { |
e625cce1 | 1670 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1671 | goto retry; |
1672 | } | |
1673 | ||
1674 | /* | |
fe4b04fa PZ |
1675 | * Since the task isn't running, its safe to add the event, us holding |
1676 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1677 | */ |
fe4b04fa | 1678 | add_event_to_ctx(event, ctx); |
e625cce1 | 1679 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1680 | } |
1681 | ||
fa289bec | 1682 | /* |
cdd6c482 | 1683 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1684 | * Enabling the leader of a group effectively enables all |
1685 | * the group members that aren't explicitly disabled, so we | |
1686 | * have to update their ->tstamp_enabled also. | |
1687 | * Note: this works for group members as well as group leaders | |
1688 | * since the non-leader members' sibling_lists will be empty. | |
1689 | */ | |
1d9b482e | 1690 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1691 | { |
cdd6c482 | 1692 | struct perf_event *sub; |
4158755d | 1693 | u64 tstamp = perf_event_time(event); |
fa289bec | 1694 | |
cdd6c482 | 1695 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1696 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1697 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1698 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1699 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1700 | } |
fa289bec PM |
1701 | } |
1702 | ||
d859e29f | 1703 | /* |
cdd6c482 | 1704 | * Cross CPU call to enable a performance event |
d859e29f | 1705 | */ |
fe4b04fa | 1706 | static int __perf_event_enable(void *info) |
04289bb9 | 1707 | { |
cdd6c482 | 1708 | struct perf_event *event = info; |
cdd6c482 IM |
1709 | struct perf_event_context *ctx = event->ctx; |
1710 | struct perf_event *leader = event->group_leader; | |
108b02cf | 1711 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 1712 | int err; |
04289bb9 | 1713 | |
fe4b04fa PZ |
1714 | if (WARN_ON_ONCE(!ctx->is_active)) |
1715 | return -EINVAL; | |
3cbed429 | 1716 | |
e625cce1 | 1717 | raw_spin_lock(&ctx->lock); |
4af4998b | 1718 | update_context_time(ctx); |
d859e29f | 1719 | |
cdd6c482 | 1720 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 1721 | goto unlock; |
e5d1367f SE |
1722 | |
1723 | /* | |
1724 | * set current task's cgroup time reference point | |
1725 | */ | |
3f7cce3c | 1726 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 1727 | |
1d9b482e | 1728 | __perf_event_mark_enabled(event); |
04289bb9 | 1729 | |
e5d1367f SE |
1730 | if (!event_filter_match(event)) { |
1731 | if (is_cgroup_event(event)) | |
1732 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 1733 | goto unlock; |
e5d1367f | 1734 | } |
f4c4176f | 1735 | |
04289bb9 | 1736 | /* |
cdd6c482 | 1737 | * If the event is in a group and isn't the group leader, |
d859e29f | 1738 | * then don't put it on unless the group is on. |
04289bb9 | 1739 | */ |
cdd6c482 | 1740 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 1741 | goto unlock; |
3b6f9e5c | 1742 | |
cdd6c482 | 1743 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 1744 | err = -EEXIST; |
e758a33d | 1745 | } else { |
cdd6c482 | 1746 | if (event == leader) |
6e37738a | 1747 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 1748 | else |
6e37738a | 1749 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 1750 | } |
d859e29f PM |
1751 | |
1752 | if (err) { | |
1753 | /* | |
cdd6c482 | 1754 | * If this event can't go on and it's part of a |
d859e29f PM |
1755 | * group, then the whole group has to come off. |
1756 | */ | |
cdd6c482 | 1757 | if (leader != event) |
d859e29f | 1758 | group_sched_out(leader, cpuctx, ctx); |
0d48696f | 1759 | if (leader->attr.pinned) { |
53cfbf59 | 1760 | update_group_times(leader); |
cdd6c482 | 1761 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 1762 | } |
d859e29f PM |
1763 | } |
1764 | ||
9ed6060d | 1765 | unlock: |
e625cce1 | 1766 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1767 | |
1768 | return 0; | |
d859e29f PM |
1769 | } |
1770 | ||
1771 | /* | |
cdd6c482 | 1772 | * Enable a event. |
c93f7669 | 1773 | * |
cdd6c482 IM |
1774 | * If event->ctx is a cloned context, callers must make sure that |
1775 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1776 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
1777 | * perf_event_for_each_child or perf_event_for_each as described |
1778 | * for perf_event_disable. | |
d859e29f | 1779 | */ |
44234adc | 1780 | void perf_event_enable(struct perf_event *event) |
d859e29f | 1781 | { |
cdd6c482 | 1782 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1783 | struct task_struct *task = ctx->task; |
1784 | ||
1785 | if (!task) { | |
1786 | /* | |
cdd6c482 | 1787 | * Enable the event on the cpu that it's on |
d859e29f | 1788 | */ |
fe4b04fa | 1789 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
1790 | return; |
1791 | } | |
1792 | ||
e625cce1 | 1793 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 1794 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
1795 | goto out; |
1796 | ||
1797 | /* | |
cdd6c482 IM |
1798 | * If the event is in error state, clear that first. |
1799 | * That way, if we see the event in error state below, we | |
d859e29f PM |
1800 | * know that it has gone back into error state, as distinct |
1801 | * from the task having been scheduled away before the | |
1802 | * cross-call arrived. | |
1803 | */ | |
cdd6c482 IM |
1804 | if (event->state == PERF_EVENT_STATE_ERROR) |
1805 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 1806 | |
9ed6060d | 1807 | retry: |
fe4b04fa | 1808 | if (!ctx->is_active) { |
1d9b482e | 1809 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
1810 | goto out; |
1811 | } | |
1812 | ||
e625cce1 | 1813 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1814 | |
1815 | if (!task_function_call(task, __perf_event_enable, event)) | |
1816 | return; | |
d859e29f | 1817 | |
e625cce1 | 1818 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
1819 | |
1820 | /* | |
cdd6c482 | 1821 | * If the context is active and the event is still off, |
d859e29f PM |
1822 | * we need to retry the cross-call. |
1823 | */ | |
fe4b04fa PZ |
1824 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
1825 | /* | |
1826 | * task could have been flipped by a concurrent | |
1827 | * perf_event_context_sched_out() | |
1828 | */ | |
1829 | task = ctx->task; | |
d859e29f | 1830 | goto retry; |
fe4b04fa | 1831 | } |
fa289bec | 1832 | |
9ed6060d | 1833 | out: |
e625cce1 | 1834 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1835 | } |
dcfce4a0 | 1836 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 1837 | |
26ca5c11 | 1838 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 1839 | { |
2023b359 | 1840 | /* |
cdd6c482 | 1841 | * not supported on inherited events |
2023b359 | 1842 | */ |
2e939d1d | 1843 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
1844 | return -EINVAL; |
1845 | ||
cdd6c482 IM |
1846 | atomic_add(refresh, &event->event_limit); |
1847 | perf_event_enable(event); | |
2023b359 PZ |
1848 | |
1849 | return 0; | |
79f14641 | 1850 | } |
26ca5c11 | 1851 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 1852 | |
5b0311e1 FW |
1853 | static void ctx_sched_out(struct perf_event_context *ctx, |
1854 | struct perf_cpu_context *cpuctx, | |
1855 | enum event_type_t event_type) | |
235c7fc7 | 1856 | { |
cdd6c482 | 1857 | struct perf_event *event; |
db24d33e | 1858 | int is_active = ctx->is_active; |
235c7fc7 | 1859 | |
db24d33e | 1860 | ctx->is_active &= ~event_type; |
cdd6c482 | 1861 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
1862 | return; |
1863 | ||
4af4998b | 1864 | update_context_time(ctx); |
e5d1367f | 1865 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 1866 | if (!ctx->nr_active) |
facc4307 | 1867 | return; |
5b0311e1 | 1868 | |
075e0b00 | 1869 | perf_pmu_disable(ctx->pmu); |
db24d33e | 1870 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
1871 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
1872 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 1873 | } |
889ff015 | 1874 | |
db24d33e | 1875 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 1876 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 1877 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 1878 | } |
1b9a644f | 1879 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
1880 | } |
1881 | ||
564c2b21 PM |
1882 | /* |
1883 | * Test whether two contexts are equivalent, i.e. whether they | |
1884 | * have both been cloned from the same version of the same context | |
cdd6c482 IM |
1885 | * and they both have the same number of enabled events. |
1886 | * If the number of enabled events is the same, then the set | |
1887 | * of enabled events should be the same, because these are both | |
1888 | * inherited contexts, therefore we can't access individual events | |
564c2b21 | 1889 | * in them directly with an fd; we can only enable/disable all |
cdd6c482 | 1890 | * events via prctl, or enable/disable all events in a family |
564c2b21 PM |
1891 | * via ioctl, which will have the same effect on both contexts. |
1892 | */ | |
cdd6c482 IM |
1893 | static int context_equiv(struct perf_event_context *ctx1, |
1894 | struct perf_event_context *ctx2) | |
564c2b21 PM |
1895 | { |
1896 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx | |
ad3a37de | 1897 | && ctx1->parent_gen == ctx2->parent_gen |
25346b93 | 1898 | && !ctx1->pin_count && !ctx2->pin_count; |
564c2b21 PM |
1899 | } |
1900 | ||
cdd6c482 IM |
1901 | static void __perf_event_sync_stat(struct perf_event *event, |
1902 | struct perf_event *next_event) | |
bfbd3381 PZ |
1903 | { |
1904 | u64 value; | |
1905 | ||
cdd6c482 | 1906 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
1907 | return; |
1908 | ||
1909 | /* | |
cdd6c482 | 1910 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
1911 | * because we're in the middle of a context switch and have IRQs |
1912 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 1913 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
1914 | * don't need to use it. |
1915 | */ | |
cdd6c482 IM |
1916 | switch (event->state) { |
1917 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
1918 | event->pmu->read(event); |
1919 | /* fall-through */ | |
bfbd3381 | 1920 | |
cdd6c482 IM |
1921 | case PERF_EVENT_STATE_INACTIVE: |
1922 | update_event_times(event); | |
bfbd3381 PZ |
1923 | break; |
1924 | ||
1925 | default: | |
1926 | break; | |
1927 | } | |
1928 | ||
1929 | /* | |
cdd6c482 | 1930 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
1931 | * values when we flip the contexts. |
1932 | */ | |
e7850595 PZ |
1933 | value = local64_read(&next_event->count); |
1934 | value = local64_xchg(&event->count, value); | |
1935 | local64_set(&next_event->count, value); | |
bfbd3381 | 1936 | |
cdd6c482 IM |
1937 | swap(event->total_time_enabled, next_event->total_time_enabled); |
1938 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 1939 | |
bfbd3381 | 1940 | /* |
19d2e755 | 1941 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 1942 | */ |
cdd6c482 IM |
1943 | perf_event_update_userpage(event); |
1944 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
1945 | } |
1946 | ||
1947 | #define list_next_entry(pos, member) \ | |
1948 | list_entry(pos->member.next, typeof(*pos), member) | |
1949 | ||
cdd6c482 IM |
1950 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
1951 | struct perf_event_context *next_ctx) | |
bfbd3381 | 1952 | { |
cdd6c482 | 1953 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
1954 | |
1955 | if (!ctx->nr_stat) | |
1956 | return; | |
1957 | ||
02ffdbc8 PZ |
1958 | update_context_time(ctx); |
1959 | ||
cdd6c482 IM |
1960 | event = list_first_entry(&ctx->event_list, |
1961 | struct perf_event, event_entry); | |
bfbd3381 | 1962 | |
cdd6c482 IM |
1963 | next_event = list_first_entry(&next_ctx->event_list, |
1964 | struct perf_event, event_entry); | |
bfbd3381 | 1965 | |
cdd6c482 IM |
1966 | while (&event->event_entry != &ctx->event_list && |
1967 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 1968 | |
cdd6c482 | 1969 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 1970 | |
cdd6c482 IM |
1971 | event = list_next_entry(event, event_entry); |
1972 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
1973 | } |
1974 | } | |
1975 | ||
fe4b04fa PZ |
1976 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
1977 | struct task_struct *next) | |
0793a61d | 1978 | { |
8dc85d54 | 1979 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 IM |
1980 | struct perf_event_context *next_ctx; |
1981 | struct perf_event_context *parent; | |
108b02cf | 1982 | struct perf_cpu_context *cpuctx; |
c93f7669 | 1983 | int do_switch = 1; |
0793a61d | 1984 | |
108b02cf PZ |
1985 | if (likely(!ctx)) |
1986 | return; | |
10989fb2 | 1987 | |
108b02cf PZ |
1988 | cpuctx = __get_cpu_context(ctx); |
1989 | if (!cpuctx->task_ctx) | |
0793a61d TG |
1990 | return; |
1991 | ||
c93f7669 PM |
1992 | rcu_read_lock(); |
1993 | parent = rcu_dereference(ctx->parent_ctx); | |
8dc85d54 | 1994 | next_ctx = next->perf_event_ctxp[ctxn]; |
c93f7669 PM |
1995 | if (parent && next_ctx && |
1996 | rcu_dereference(next_ctx->parent_ctx) == parent) { | |
1997 | /* | |
1998 | * Looks like the two contexts are clones, so we might be | |
1999 | * able to optimize the context switch. We lock both | |
2000 | * contexts and check that they are clones under the | |
2001 | * lock (including re-checking that neither has been | |
2002 | * uncloned in the meantime). It doesn't matter which | |
2003 | * order we take the locks because no other cpu could | |
2004 | * be trying to lock both of these tasks. | |
2005 | */ | |
e625cce1 TG |
2006 | raw_spin_lock(&ctx->lock); |
2007 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2008 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2009 | /* |
2010 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2011 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2012 | */ |
8dc85d54 PZ |
2013 | task->perf_event_ctxp[ctxn] = next_ctx; |
2014 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2015 | ctx->task = next; |
2016 | next_ctx->task = task; | |
2017 | do_switch = 0; | |
bfbd3381 | 2018 | |
cdd6c482 | 2019 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2020 | } |
e625cce1 TG |
2021 | raw_spin_unlock(&next_ctx->lock); |
2022 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2023 | } |
c93f7669 | 2024 | rcu_read_unlock(); |
564c2b21 | 2025 | |
c93f7669 | 2026 | if (do_switch) { |
facc4307 | 2027 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2028 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2029 | cpuctx->task_ctx = NULL; |
facc4307 | 2030 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2031 | } |
0793a61d TG |
2032 | } |
2033 | ||
8dc85d54 PZ |
2034 | #define for_each_task_context_nr(ctxn) \ |
2035 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2036 | ||
2037 | /* | |
2038 | * Called from scheduler to remove the events of the current task, | |
2039 | * with interrupts disabled. | |
2040 | * | |
2041 | * We stop each event and update the event value in event->count. | |
2042 | * | |
2043 | * This does not protect us against NMI, but disable() | |
2044 | * sets the disabled bit in the control field of event _before_ | |
2045 | * accessing the event control register. If a NMI hits, then it will | |
2046 | * not restart the event. | |
2047 | */ | |
ab0cce56 JO |
2048 | void __perf_event_task_sched_out(struct task_struct *task, |
2049 | struct task_struct *next) | |
8dc85d54 PZ |
2050 | { |
2051 | int ctxn; | |
2052 | ||
8dc85d54 PZ |
2053 | for_each_task_context_nr(ctxn) |
2054 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2055 | |
2056 | /* | |
2057 | * if cgroup events exist on this CPU, then we need | |
2058 | * to check if we have to switch out PMU state. | |
2059 | * cgroup event are system-wide mode only | |
2060 | */ | |
2061 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2062 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2063 | } |
2064 | ||
04dc2dbb | 2065 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2066 | { |
108b02cf | 2067 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2068 | |
a63eaf34 PM |
2069 | if (!cpuctx->task_ctx) |
2070 | return; | |
012b84da IM |
2071 | |
2072 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2073 | return; | |
2074 | ||
04dc2dbb | 2075 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2076 | cpuctx->task_ctx = NULL; |
2077 | } | |
2078 | ||
5b0311e1 FW |
2079 | /* |
2080 | * Called with IRQs disabled | |
2081 | */ | |
2082 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2083 | enum event_type_t event_type) | |
2084 | { | |
2085 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2086 | } |
2087 | ||
235c7fc7 | 2088 | static void |
5b0311e1 | 2089 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2090 | struct perf_cpu_context *cpuctx) |
0793a61d | 2091 | { |
cdd6c482 | 2092 | struct perf_event *event; |
0793a61d | 2093 | |
889ff015 FW |
2094 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2095 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2096 | continue; |
5632ab12 | 2097 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2098 | continue; |
2099 | ||
e5d1367f SE |
2100 | /* may need to reset tstamp_enabled */ |
2101 | if (is_cgroup_event(event)) | |
2102 | perf_cgroup_mark_enabled(event, ctx); | |
2103 | ||
8c9ed8e1 | 2104 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2105 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2106 | |
2107 | /* | |
2108 | * If this pinned group hasn't been scheduled, | |
2109 | * put it in error state. | |
2110 | */ | |
cdd6c482 IM |
2111 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2112 | update_group_times(event); | |
2113 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2114 | } |
3b6f9e5c | 2115 | } |
5b0311e1 FW |
2116 | } |
2117 | ||
2118 | static void | |
2119 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2120 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2121 | { |
2122 | struct perf_event *event; | |
2123 | int can_add_hw = 1; | |
3b6f9e5c | 2124 | |
889ff015 FW |
2125 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2126 | /* Ignore events in OFF or ERROR state */ | |
2127 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2128 | continue; |
04289bb9 IM |
2129 | /* |
2130 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2131 | * of events: |
04289bb9 | 2132 | */ |
5632ab12 | 2133 | if (!event_filter_match(event)) |
0793a61d TG |
2134 | continue; |
2135 | ||
e5d1367f SE |
2136 | /* may need to reset tstamp_enabled */ |
2137 | if (is_cgroup_event(event)) | |
2138 | perf_cgroup_mark_enabled(event, ctx); | |
2139 | ||
9ed6060d | 2140 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2141 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2142 | can_add_hw = 0; |
9ed6060d | 2143 | } |
0793a61d | 2144 | } |
5b0311e1 FW |
2145 | } |
2146 | ||
2147 | static void | |
2148 | ctx_sched_in(struct perf_event_context *ctx, | |
2149 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2150 | enum event_type_t event_type, |
2151 | struct task_struct *task) | |
5b0311e1 | 2152 | { |
e5d1367f | 2153 | u64 now; |
db24d33e | 2154 | int is_active = ctx->is_active; |
e5d1367f | 2155 | |
db24d33e | 2156 | ctx->is_active |= event_type; |
5b0311e1 | 2157 | if (likely(!ctx->nr_events)) |
facc4307 | 2158 | return; |
5b0311e1 | 2159 | |
e5d1367f SE |
2160 | now = perf_clock(); |
2161 | ctx->timestamp = now; | |
3f7cce3c | 2162 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2163 | /* |
2164 | * First go through the list and put on any pinned groups | |
2165 | * in order to give them the best chance of going on. | |
2166 | */ | |
db24d33e | 2167 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2168 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2169 | |
2170 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2171 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2172 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2173 | } |
2174 | ||
329c0e01 | 2175 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2176 | enum event_type_t event_type, |
2177 | struct task_struct *task) | |
329c0e01 FW |
2178 | { |
2179 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2180 | ||
e5d1367f | 2181 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2182 | } |
2183 | ||
e5d1367f SE |
2184 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2185 | struct task_struct *task) | |
235c7fc7 | 2186 | { |
108b02cf | 2187 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2188 | |
108b02cf | 2189 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2190 | if (cpuctx->task_ctx == ctx) |
2191 | return; | |
2192 | ||
facc4307 | 2193 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2194 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2195 | /* |
2196 | * We want to keep the following priority order: | |
2197 | * cpu pinned (that don't need to move), task pinned, | |
2198 | * cpu flexible, task flexible. | |
2199 | */ | |
2200 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2201 | ||
1d5f003f GN |
2202 | if (ctx->nr_events) |
2203 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2204 | |
86b47c25 GN |
2205 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2206 | ||
facc4307 PZ |
2207 | perf_pmu_enable(ctx->pmu); |
2208 | perf_ctx_unlock(cpuctx, ctx); | |
2209 | ||
b5ab4cd5 PZ |
2210 | /* |
2211 | * Since these rotations are per-cpu, we need to ensure the | |
2212 | * cpu-context we got scheduled on is actually rotating. | |
2213 | */ | |
108b02cf | 2214 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2215 | } |
2216 | ||
d010b332 SE |
2217 | /* |
2218 | * When sampling the branck stack in system-wide, it may be necessary | |
2219 | * to flush the stack on context switch. This happens when the branch | |
2220 | * stack does not tag its entries with the pid of the current task. | |
2221 | * Otherwise it becomes impossible to associate a branch entry with a | |
2222 | * task. This ambiguity is more likely to appear when the branch stack | |
2223 | * supports priv level filtering and the user sets it to monitor only | |
2224 | * at the user level (which could be a useful measurement in system-wide | |
2225 | * mode). In that case, the risk is high of having a branch stack with | |
2226 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2227 | * entries or stashing them somewhere in the PMU specific code layer. | |
2228 | * | |
2229 | * This function provides the context switch callback to the lower code | |
2230 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2231 | * with at least one active event using taken branch sampling. | |
2232 | */ | |
2233 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2234 | struct task_struct *task) | |
2235 | { | |
2236 | struct perf_cpu_context *cpuctx; | |
2237 | struct pmu *pmu; | |
2238 | unsigned long flags; | |
2239 | ||
2240 | /* no need to flush branch stack if not changing task */ | |
2241 | if (prev == task) | |
2242 | return; | |
2243 | ||
2244 | local_irq_save(flags); | |
2245 | ||
2246 | rcu_read_lock(); | |
2247 | ||
2248 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2249 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2250 | ||
2251 | /* | |
2252 | * check if the context has at least one | |
2253 | * event using PERF_SAMPLE_BRANCH_STACK | |
2254 | */ | |
2255 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2256 | && pmu->flush_branch_stack) { | |
2257 | ||
2258 | pmu = cpuctx->ctx.pmu; | |
2259 | ||
2260 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2261 | ||
2262 | perf_pmu_disable(pmu); | |
2263 | ||
2264 | pmu->flush_branch_stack(); | |
2265 | ||
2266 | perf_pmu_enable(pmu); | |
2267 | ||
2268 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2269 | } | |
2270 | } | |
2271 | ||
2272 | rcu_read_unlock(); | |
2273 | ||
2274 | local_irq_restore(flags); | |
2275 | } | |
2276 | ||
8dc85d54 PZ |
2277 | /* |
2278 | * Called from scheduler to add the events of the current task | |
2279 | * with interrupts disabled. | |
2280 | * | |
2281 | * We restore the event value and then enable it. | |
2282 | * | |
2283 | * This does not protect us against NMI, but enable() | |
2284 | * sets the enabled bit in the control field of event _before_ | |
2285 | * accessing the event control register. If a NMI hits, then it will | |
2286 | * keep the event running. | |
2287 | */ | |
ab0cce56 JO |
2288 | void __perf_event_task_sched_in(struct task_struct *prev, |
2289 | struct task_struct *task) | |
8dc85d54 PZ |
2290 | { |
2291 | struct perf_event_context *ctx; | |
2292 | int ctxn; | |
2293 | ||
2294 | for_each_task_context_nr(ctxn) { | |
2295 | ctx = task->perf_event_ctxp[ctxn]; | |
2296 | if (likely(!ctx)) | |
2297 | continue; | |
2298 | ||
e5d1367f | 2299 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2300 | } |
e5d1367f SE |
2301 | /* |
2302 | * if cgroup events exist on this CPU, then we need | |
2303 | * to check if we have to switch in PMU state. | |
2304 | * cgroup event are system-wide mode only | |
2305 | */ | |
2306 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2307 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2308 | |
2309 | /* check for system-wide branch_stack events */ | |
2310 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2311 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2312 | } |
2313 | ||
abd50713 PZ |
2314 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2315 | { | |
2316 | u64 frequency = event->attr.sample_freq; | |
2317 | u64 sec = NSEC_PER_SEC; | |
2318 | u64 divisor, dividend; | |
2319 | ||
2320 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2321 | ||
2322 | count_fls = fls64(count); | |
2323 | nsec_fls = fls64(nsec); | |
2324 | frequency_fls = fls64(frequency); | |
2325 | sec_fls = 30; | |
2326 | ||
2327 | /* | |
2328 | * We got @count in @nsec, with a target of sample_freq HZ | |
2329 | * the target period becomes: | |
2330 | * | |
2331 | * @count * 10^9 | |
2332 | * period = ------------------- | |
2333 | * @nsec * sample_freq | |
2334 | * | |
2335 | */ | |
2336 | ||
2337 | /* | |
2338 | * Reduce accuracy by one bit such that @a and @b converge | |
2339 | * to a similar magnitude. | |
2340 | */ | |
fe4b04fa | 2341 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2342 | do { \ |
2343 | if (a##_fls > b##_fls) { \ | |
2344 | a >>= 1; \ | |
2345 | a##_fls--; \ | |
2346 | } else { \ | |
2347 | b >>= 1; \ | |
2348 | b##_fls--; \ | |
2349 | } \ | |
2350 | } while (0) | |
2351 | ||
2352 | /* | |
2353 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2354 | * the other, so that finally we can do a u64/u64 division. | |
2355 | */ | |
2356 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2357 | REDUCE_FLS(nsec, frequency); | |
2358 | REDUCE_FLS(sec, count); | |
2359 | } | |
2360 | ||
2361 | if (count_fls + sec_fls > 64) { | |
2362 | divisor = nsec * frequency; | |
2363 | ||
2364 | while (count_fls + sec_fls > 64) { | |
2365 | REDUCE_FLS(count, sec); | |
2366 | divisor >>= 1; | |
2367 | } | |
2368 | ||
2369 | dividend = count * sec; | |
2370 | } else { | |
2371 | dividend = count * sec; | |
2372 | ||
2373 | while (nsec_fls + frequency_fls > 64) { | |
2374 | REDUCE_FLS(nsec, frequency); | |
2375 | dividend >>= 1; | |
2376 | } | |
2377 | ||
2378 | divisor = nsec * frequency; | |
2379 | } | |
2380 | ||
f6ab91ad PZ |
2381 | if (!divisor) |
2382 | return dividend; | |
2383 | ||
abd50713 PZ |
2384 | return div64_u64(dividend, divisor); |
2385 | } | |
2386 | ||
e050e3f0 SE |
2387 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2388 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2389 | ||
f39d47ff | 2390 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2391 | { |
cdd6c482 | 2392 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2393 | s64 period, sample_period; |
bd2b5b12 PZ |
2394 | s64 delta; |
2395 | ||
abd50713 | 2396 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2397 | |
2398 | delta = (s64)(period - hwc->sample_period); | |
2399 | delta = (delta + 7) / 8; /* low pass filter */ | |
2400 | ||
2401 | sample_period = hwc->sample_period + delta; | |
2402 | ||
2403 | if (!sample_period) | |
2404 | sample_period = 1; | |
2405 | ||
bd2b5b12 | 2406 | hwc->sample_period = sample_period; |
abd50713 | 2407 | |
e7850595 | 2408 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2409 | if (disable) |
2410 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2411 | ||
e7850595 | 2412 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2413 | |
2414 | if (disable) | |
2415 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2416 | } |
bd2b5b12 PZ |
2417 | } |
2418 | ||
e050e3f0 SE |
2419 | /* |
2420 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2421 | * events. At the same time, make sure, having freq events does not change | |
2422 | * the rate of unthrottling as that would introduce bias. | |
2423 | */ | |
2424 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2425 | int needs_unthr) | |
60db5e09 | 2426 | { |
cdd6c482 IM |
2427 | struct perf_event *event; |
2428 | struct hw_perf_event *hwc; | |
e050e3f0 | 2429 | u64 now, period = TICK_NSEC; |
abd50713 | 2430 | s64 delta; |
60db5e09 | 2431 | |
e050e3f0 SE |
2432 | /* |
2433 | * only need to iterate over all events iff: | |
2434 | * - context have events in frequency mode (needs freq adjust) | |
2435 | * - there are events to unthrottle on this cpu | |
2436 | */ | |
2437 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2438 | return; |
2439 | ||
e050e3f0 | 2440 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2441 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2442 | |
03541f8b | 2443 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2444 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2445 | continue; |
2446 | ||
5632ab12 | 2447 | if (!event_filter_match(event)) |
5d27c23d PZ |
2448 | continue; |
2449 | ||
cdd6c482 | 2450 | hwc = &event->hw; |
6a24ed6c | 2451 | |
e050e3f0 SE |
2452 | if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) { |
2453 | hwc->interrupts = 0; | |
cdd6c482 | 2454 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2455 | event->pmu->start(event, 0); |
a78ac325 PZ |
2456 | } |
2457 | ||
cdd6c482 | 2458 | if (!event->attr.freq || !event->attr.sample_freq) |
60db5e09 PZ |
2459 | continue; |
2460 | ||
e050e3f0 SE |
2461 | /* |
2462 | * stop the event and update event->count | |
2463 | */ | |
2464 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2465 | ||
e7850595 | 2466 | now = local64_read(&event->count); |
abd50713 PZ |
2467 | delta = now - hwc->freq_count_stamp; |
2468 | hwc->freq_count_stamp = now; | |
60db5e09 | 2469 | |
e050e3f0 SE |
2470 | /* |
2471 | * restart the event | |
2472 | * reload only if value has changed | |
f39d47ff SE |
2473 | * we have stopped the event so tell that |
2474 | * to perf_adjust_period() to avoid stopping it | |
2475 | * twice. | |
e050e3f0 | 2476 | */ |
abd50713 | 2477 | if (delta > 0) |
f39d47ff | 2478 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2479 | |
2480 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
60db5e09 | 2481 | } |
e050e3f0 | 2482 | |
f39d47ff | 2483 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2484 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2485 | } |
2486 | ||
235c7fc7 | 2487 | /* |
cdd6c482 | 2488 | * Round-robin a context's events: |
235c7fc7 | 2489 | */ |
cdd6c482 | 2490 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2491 | { |
dddd3379 TG |
2492 | /* |
2493 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2494 | * disabled by the inheritance code. | |
2495 | */ | |
2496 | if (!ctx->rotate_disable) | |
2497 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2498 | } |
2499 | ||
b5ab4cd5 | 2500 | /* |
e9d2b064 PZ |
2501 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2502 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2503 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2504 | */ |
e9d2b064 | 2505 | static void perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2506 | { |
8dc85d54 | 2507 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2508 | int rotate = 0, remove = 1; |
7fc23a53 | 2509 | |
b5ab4cd5 | 2510 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2511 | remove = 0; |
b5ab4cd5 PZ |
2512 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2513 | rotate = 1; | |
2514 | } | |
235c7fc7 | 2515 | |
8dc85d54 | 2516 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2517 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2518 | remove = 0; |
b5ab4cd5 PZ |
2519 | if (ctx->nr_events != ctx->nr_active) |
2520 | rotate = 1; | |
2521 | } | |
9717e6cd | 2522 | |
e050e3f0 | 2523 | if (!rotate) |
0f5a2601 PZ |
2524 | goto done; |
2525 | ||
facc4307 | 2526 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2527 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2528 | |
e050e3f0 SE |
2529 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2530 | if (ctx) | |
2531 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2532 | |
e050e3f0 SE |
2533 | rotate_ctx(&cpuctx->ctx); |
2534 | if (ctx) | |
2535 | rotate_ctx(ctx); | |
235c7fc7 | 2536 | |
e050e3f0 | 2537 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2538 | |
0f5a2601 PZ |
2539 | perf_pmu_enable(cpuctx->ctx.pmu); |
2540 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2541 | done: |
e9d2b064 PZ |
2542 | if (remove) |
2543 | list_del_init(&cpuctx->rotation_list); | |
e9d2b064 PZ |
2544 | } |
2545 | ||
2546 | void perf_event_task_tick(void) | |
2547 | { | |
2548 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2549 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2550 | struct perf_event_context *ctx; |
2551 | int throttled; | |
b5ab4cd5 | 2552 | |
e9d2b064 PZ |
2553 | WARN_ON(!irqs_disabled()); |
2554 | ||
e050e3f0 SE |
2555 | __this_cpu_inc(perf_throttled_seq); |
2556 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2557 | ||
e9d2b064 | 2558 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2559 | ctx = &cpuctx->ctx; |
2560 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2561 | ||
2562 | ctx = cpuctx->task_ctx; | |
2563 | if (ctx) | |
2564 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2565 | ||
e9d2b064 PZ |
2566 | if (cpuctx->jiffies_interval == 1 || |
2567 | !(jiffies % cpuctx->jiffies_interval)) | |
2568 | perf_rotate_context(cpuctx); | |
2569 | } | |
0793a61d TG |
2570 | } |
2571 | ||
889ff015 FW |
2572 | static int event_enable_on_exec(struct perf_event *event, |
2573 | struct perf_event_context *ctx) | |
2574 | { | |
2575 | if (!event->attr.enable_on_exec) | |
2576 | return 0; | |
2577 | ||
2578 | event->attr.enable_on_exec = 0; | |
2579 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2580 | return 0; | |
2581 | ||
1d9b482e | 2582 | __perf_event_mark_enabled(event); |
889ff015 FW |
2583 | |
2584 | return 1; | |
2585 | } | |
2586 | ||
57e7986e | 2587 | /* |
cdd6c482 | 2588 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2589 | * This expects task == current. |
2590 | */ | |
8dc85d54 | 2591 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2592 | { |
cdd6c482 | 2593 | struct perf_event *event; |
57e7986e PM |
2594 | unsigned long flags; |
2595 | int enabled = 0; | |
889ff015 | 2596 | int ret; |
57e7986e PM |
2597 | |
2598 | local_irq_save(flags); | |
cdd6c482 | 2599 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2600 | goto out; |
2601 | ||
e566b76e SE |
2602 | /* |
2603 | * We must ctxsw out cgroup events to avoid conflict | |
2604 | * when invoking perf_task_event_sched_in() later on | |
2605 | * in this function. Otherwise we end up trying to | |
2606 | * ctxswin cgroup events which are already scheduled | |
2607 | * in. | |
2608 | */ | |
a8d757ef | 2609 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2610 | |
e625cce1 | 2611 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2612 | task_ctx_sched_out(ctx); |
57e7986e | 2613 | |
b79387ef | 2614 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2615 | ret = event_enable_on_exec(event, ctx); |
2616 | if (ret) | |
2617 | enabled = 1; | |
57e7986e PM |
2618 | } |
2619 | ||
2620 | /* | |
cdd6c482 | 2621 | * Unclone this context if we enabled any event. |
57e7986e | 2622 | */ |
71a851b4 PZ |
2623 | if (enabled) |
2624 | unclone_ctx(ctx); | |
57e7986e | 2625 | |
e625cce1 | 2626 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2627 | |
e566b76e SE |
2628 | /* |
2629 | * Also calls ctxswin for cgroup events, if any: | |
2630 | */ | |
e5d1367f | 2631 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2632 | out: |
57e7986e PM |
2633 | local_irq_restore(flags); |
2634 | } | |
2635 | ||
0793a61d | 2636 | /* |
cdd6c482 | 2637 | * Cross CPU call to read the hardware event |
0793a61d | 2638 | */ |
cdd6c482 | 2639 | static void __perf_event_read(void *info) |
0793a61d | 2640 | { |
cdd6c482 IM |
2641 | struct perf_event *event = info; |
2642 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2643 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2644 | |
e1ac3614 PM |
2645 | /* |
2646 | * If this is a task context, we need to check whether it is | |
2647 | * the current task context of this cpu. If not it has been | |
2648 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2649 | * event->count would have been updated to a recent sample |
2650 | * when the event was scheduled out. | |
e1ac3614 PM |
2651 | */ |
2652 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2653 | return; | |
2654 | ||
e625cce1 | 2655 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2656 | if (ctx->is_active) { |
542e72fc | 2657 | update_context_time(ctx); |
e5d1367f SE |
2658 | update_cgrp_time_from_event(event); |
2659 | } | |
cdd6c482 | 2660 | update_event_times(event); |
542e72fc PZ |
2661 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2662 | event->pmu->read(event); | |
e625cce1 | 2663 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2664 | } |
2665 | ||
b5e58793 PZ |
2666 | static inline u64 perf_event_count(struct perf_event *event) |
2667 | { | |
e7850595 | 2668 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
2669 | } |
2670 | ||
cdd6c482 | 2671 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
2672 | { |
2673 | /* | |
cdd6c482 IM |
2674 | * If event is enabled and currently active on a CPU, update the |
2675 | * value in the event structure: | |
0793a61d | 2676 | */ |
cdd6c482 IM |
2677 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2678 | smp_call_function_single(event->oncpu, | |
2679 | __perf_event_read, event, 1); | |
2680 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
2681 | struct perf_event_context *ctx = event->ctx; |
2682 | unsigned long flags; | |
2683 | ||
e625cce1 | 2684 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
2685 | /* |
2686 | * may read while context is not active | |
2687 | * (e.g., thread is blocked), in that case | |
2688 | * we cannot update context time | |
2689 | */ | |
e5d1367f | 2690 | if (ctx->is_active) { |
c530ccd9 | 2691 | update_context_time(ctx); |
e5d1367f SE |
2692 | update_cgrp_time_from_event(event); |
2693 | } | |
cdd6c482 | 2694 | update_event_times(event); |
e625cce1 | 2695 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
2696 | } |
2697 | ||
b5e58793 | 2698 | return perf_event_count(event); |
0793a61d TG |
2699 | } |
2700 | ||
a63eaf34 | 2701 | /* |
cdd6c482 | 2702 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 2703 | */ |
eb184479 | 2704 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 2705 | { |
e625cce1 | 2706 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 2707 | mutex_init(&ctx->mutex); |
889ff015 FW |
2708 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2709 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
2710 | INIT_LIST_HEAD(&ctx->event_list); |
2711 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
2712 | } |
2713 | ||
2714 | static struct perf_event_context * | |
2715 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
2716 | { | |
2717 | struct perf_event_context *ctx; | |
2718 | ||
2719 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
2720 | if (!ctx) | |
2721 | return NULL; | |
2722 | ||
2723 | __perf_event_init_context(ctx); | |
2724 | if (task) { | |
2725 | ctx->task = task; | |
2726 | get_task_struct(task); | |
0793a61d | 2727 | } |
eb184479 PZ |
2728 | ctx->pmu = pmu; |
2729 | ||
2730 | return ctx; | |
a63eaf34 PM |
2731 | } |
2732 | ||
2ebd4ffb MH |
2733 | static struct task_struct * |
2734 | find_lively_task_by_vpid(pid_t vpid) | |
2735 | { | |
2736 | struct task_struct *task; | |
2737 | int err; | |
0793a61d TG |
2738 | |
2739 | rcu_read_lock(); | |
2ebd4ffb | 2740 | if (!vpid) |
0793a61d TG |
2741 | task = current; |
2742 | else | |
2ebd4ffb | 2743 | task = find_task_by_vpid(vpid); |
0793a61d TG |
2744 | if (task) |
2745 | get_task_struct(task); | |
2746 | rcu_read_unlock(); | |
2747 | ||
2748 | if (!task) | |
2749 | return ERR_PTR(-ESRCH); | |
2750 | ||
0793a61d | 2751 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
2752 | err = -EACCES; |
2753 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
2754 | goto errout; | |
2755 | ||
2ebd4ffb MH |
2756 | return task; |
2757 | errout: | |
2758 | put_task_struct(task); | |
2759 | return ERR_PTR(err); | |
2760 | ||
2761 | } | |
2762 | ||
fe4b04fa PZ |
2763 | /* |
2764 | * Returns a matching context with refcount and pincount. | |
2765 | */ | |
108b02cf | 2766 | static struct perf_event_context * |
38a81da2 | 2767 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 2768 | { |
cdd6c482 | 2769 | struct perf_event_context *ctx; |
22a4f650 | 2770 | struct perf_cpu_context *cpuctx; |
25346b93 | 2771 | unsigned long flags; |
8dc85d54 | 2772 | int ctxn, err; |
0793a61d | 2773 | |
22a4ec72 | 2774 | if (!task) { |
cdd6c482 | 2775 | /* Must be root to operate on a CPU event: */ |
0764771d | 2776 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
2777 | return ERR_PTR(-EACCES); |
2778 | ||
0793a61d | 2779 | /* |
cdd6c482 | 2780 | * We could be clever and allow to attach a event to an |
0793a61d TG |
2781 | * offline CPU and activate it when the CPU comes up, but |
2782 | * that's for later. | |
2783 | */ | |
f6325e30 | 2784 | if (!cpu_online(cpu)) |
0793a61d TG |
2785 | return ERR_PTR(-ENODEV); |
2786 | ||
108b02cf | 2787 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 2788 | ctx = &cpuctx->ctx; |
c93f7669 | 2789 | get_ctx(ctx); |
fe4b04fa | 2790 | ++ctx->pin_count; |
0793a61d | 2791 | |
0793a61d TG |
2792 | return ctx; |
2793 | } | |
2794 | ||
8dc85d54 PZ |
2795 | err = -EINVAL; |
2796 | ctxn = pmu->task_ctx_nr; | |
2797 | if (ctxn < 0) | |
2798 | goto errout; | |
2799 | ||
9ed6060d | 2800 | retry: |
8dc85d54 | 2801 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 2802 | if (ctx) { |
71a851b4 | 2803 | unclone_ctx(ctx); |
fe4b04fa | 2804 | ++ctx->pin_count; |
e625cce1 | 2805 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 2806 | } else { |
eb184479 | 2807 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
2808 | err = -ENOMEM; |
2809 | if (!ctx) | |
2810 | goto errout; | |
eb184479 | 2811 | |
dbe08d82 ON |
2812 | err = 0; |
2813 | mutex_lock(&task->perf_event_mutex); | |
2814 | /* | |
2815 | * If it has already passed perf_event_exit_task(). | |
2816 | * we must see PF_EXITING, it takes this mutex too. | |
2817 | */ | |
2818 | if (task->flags & PF_EXITING) | |
2819 | err = -ESRCH; | |
2820 | else if (task->perf_event_ctxp[ctxn]) | |
2821 | err = -EAGAIN; | |
fe4b04fa | 2822 | else { |
9137fb28 | 2823 | get_ctx(ctx); |
fe4b04fa | 2824 | ++ctx->pin_count; |
dbe08d82 | 2825 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 2826 | } |
dbe08d82 ON |
2827 | mutex_unlock(&task->perf_event_mutex); |
2828 | ||
2829 | if (unlikely(err)) { | |
9137fb28 | 2830 | put_ctx(ctx); |
dbe08d82 ON |
2831 | |
2832 | if (err == -EAGAIN) | |
2833 | goto retry; | |
2834 | goto errout; | |
a63eaf34 PM |
2835 | } |
2836 | } | |
2837 | ||
0793a61d | 2838 | return ctx; |
c93f7669 | 2839 | |
9ed6060d | 2840 | errout: |
c93f7669 | 2841 | return ERR_PTR(err); |
0793a61d TG |
2842 | } |
2843 | ||
6fb2915d LZ |
2844 | static void perf_event_free_filter(struct perf_event *event); |
2845 | ||
cdd6c482 | 2846 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 2847 | { |
cdd6c482 | 2848 | struct perf_event *event; |
592903cd | 2849 | |
cdd6c482 IM |
2850 | event = container_of(head, struct perf_event, rcu_head); |
2851 | if (event->ns) | |
2852 | put_pid_ns(event->ns); | |
6fb2915d | 2853 | perf_event_free_filter(event); |
cdd6c482 | 2854 | kfree(event); |
592903cd PZ |
2855 | } |
2856 | ||
76369139 | 2857 | static void ring_buffer_put(struct ring_buffer *rb); |
925d519a | 2858 | |
cdd6c482 | 2859 | static void free_event(struct perf_event *event) |
f1600952 | 2860 | { |
e360adbe | 2861 | irq_work_sync(&event->pending); |
925d519a | 2862 | |
cdd6c482 | 2863 | if (!event->parent) { |
82cd6def | 2864 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 2865 | static_key_slow_dec_deferred(&perf_sched_events); |
3af9e859 | 2866 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
2867 | atomic_dec(&nr_mmap_events); |
2868 | if (event->attr.comm) | |
2869 | atomic_dec(&nr_comm_events); | |
2870 | if (event->attr.task) | |
2871 | atomic_dec(&nr_task_events); | |
927c7a9e FW |
2872 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
2873 | put_callchain_buffers(); | |
08309379 PZ |
2874 | if (is_cgroup_event(event)) { |
2875 | atomic_dec(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 2876 | static_key_slow_dec_deferred(&perf_sched_events); |
08309379 | 2877 | } |
d010b332 SE |
2878 | |
2879 | if (has_branch_stack(event)) { | |
2880 | static_key_slow_dec_deferred(&perf_sched_events); | |
2881 | /* is system-wide event */ | |
2882 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
2883 | atomic_dec(&per_cpu(perf_branch_stack_events, | |
2884 | event->cpu)); | |
2885 | } | |
f344011c | 2886 | } |
9ee318a7 | 2887 | |
76369139 FW |
2888 | if (event->rb) { |
2889 | ring_buffer_put(event->rb); | |
2890 | event->rb = NULL; | |
a4be7c27 PZ |
2891 | } |
2892 | ||
e5d1367f SE |
2893 | if (is_cgroup_event(event)) |
2894 | perf_detach_cgroup(event); | |
2895 | ||
cdd6c482 IM |
2896 | if (event->destroy) |
2897 | event->destroy(event); | |
e077df4f | 2898 | |
0c67b408 PZ |
2899 | if (event->ctx) |
2900 | put_ctx(event->ctx); | |
2901 | ||
cdd6c482 | 2902 | call_rcu(&event->rcu_head, free_event_rcu); |
f1600952 PZ |
2903 | } |
2904 | ||
a66a3052 | 2905 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 2906 | { |
cdd6c482 | 2907 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 2908 | |
ad3a37de | 2909 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
2910 | /* |
2911 | * There are two ways this annotation is useful: | |
2912 | * | |
2913 | * 1) there is a lock recursion from perf_event_exit_task | |
2914 | * see the comment there. | |
2915 | * | |
2916 | * 2) there is a lock-inversion with mmap_sem through | |
2917 | * perf_event_read_group(), which takes faults while | |
2918 | * holding ctx->mutex, however this is called after | |
2919 | * the last filedesc died, so there is no possibility | |
2920 | * to trigger the AB-BA case. | |
2921 | */ | |
2922 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
050735b0 | 2923 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 2924 | perf_group_detach(event); |
050735b0 | 2925 | raw_spin_unlock_irq(&ctx->lock); |
e03a9a55 | 2926 | perf_remove_from_context(event); |
d859e29f | 2927 | mutex_unlock(&ctx->mutex); |
0793a61d | 2928 | |
cdd6c482 | 2929 | free_event(event); |
0793a61d TG |
2930 | |
2931 | return 0; | |
2932 | } | |
a66a3052 | 2933 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 2934 | |
a66a3052 PZ |
2935 | /* |
2936 | * Called when the last reference to the file is gone. | |
2937 | */ | |
2938 | static int perf_release(struct inode *inode, struct file *file) | |
fb0459d7 | 2939 | { |
a66a3052 | 2940 | struct perf_event *event = file->private_data; |
8882135b | 2941 | struct task_struct *owner; |
fb0459d7 | 2942 | |
a66a3052 | 2943 | file->private_data = NULL; |
fb0459d7 | 2944 | |
8882135b PZ |
2945 | rcu_read_lock(); |
2946 | owner = ACCESS_ONCE(event->owner); | |
2947 | /* | |
2948 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
2949 | * !owner it means the list deletion is complete and we can indeed | |
2950 | * free this event, otherwise we need to serialize on | |
2951 | * owner->perf_event_mutex. | |
2952 | */ | |
2953 | smp_read_barrier_depends(); | |
2954 | if (owner) { | |
2955 | /* | |
2956 | * Since delayed_put_task_struct() also drops the last | |
2957 | * task reference we can safely take a new reference | |
2958 | * while holding the rcu_read_lock(). | |
2959 | */ | |
2960 | get_task_struct(owner); | |
2961 | } | |
2962 | rcu_read_unlock(); | |
2963 | ||
2964 | if (owner) { | |
2965 | mutex_lock(&owner->perf_event_mutex); | |
2966 | /* | |
2967 | * We have to re-check the event->owner field, if it is cleared | |
2968 | * we raced with perf_event_exit_task(), acquiring the mutex | |
2969 | * ensured they're done, and we can proceed with freeing the | |
2970 | * event. | |
2971 | */ | |
2972 | if (event->owner) | |
2973 | list_del_init(&event->owner_entry); | |
2974 | mutex_unlock(&owner->perf_event_mutex); | |
2975 | put_task_struct(owner); | |
2976 | } | |
2977 | ||
a66a3052 | 2978 | return perf_event_release_kernel(event); |
fb0459d7 | 2979 | } |
fb0459d7 | 2980 | |
59ed446f | 2981 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 2982 | { |
cdd6c482 | 2983 | struct perf_event *child; |
e53c0994 PZ |
2984 | u64 total = 0; |
2985 | ||
59ed446f PZ |
2986 | *enabled = 0; |
2987 | *running = 0; | |
2988 | ||
6f10581a | 2989 | mutex_lock(&event->child_mutex); |
cdd6c482 | 2990 | total += perf_event_read(event); |
59ed446f PZ |
2991 | *enabled += event->total_time_enabled + |
2992 | atomic64_read(&event->child_total_time_enabled); | |
2993 | *running += event->total_time_running + | |
2994 | atomic64_read(&event->child_total_time_running); | |
2995 | ||
2996 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 2997 | total += perf_event_read(child); |
59ed446f PZ |
2998 | *enabled += child->total_time_enabled; |
2999 | *running += child->total_time_running; | |
3000 | } | |
6f10581a | 3001 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3002 | |
3003 | return total; | |
3004 | } | |
fb0459d7 | 3005 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3006 | |
cdd6c482 | 3007 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3008 | u64 read_format, char __user *buf) |
3009 | { | |
cdd6c482 | 3010 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3011 | int n = 0, size = 0, ret = -EFAULT; |
3012 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3013 | u64 values[5]; |
59ed446f | 3014 | u64 count, enabled, running; |
abf4868b | 3015 | |
6f10581a | 3016 | mutex_lock(&ctx->mutex); |
59ed446f | 3017 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3018 | |
3019 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3020 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3021 | values[n++] = enabled; | |
3022 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3023 | values[n++] = running; | |
abf4868b PZ |
3024 | values[n++] = count; |
3025 | if (read_format & PERF_FORMAT_ID) | |
3026 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3027 | |
3028 | size = n * sizeof(u64); | |
3029 | ||
3030 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3031 | goto unlock; |
3dab77fb | 3032 | |
6f10581a | 3033 | ret = size; |
3dab77fb | 3034 | |
65abc865 | 3035 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3036 | n = 0; |
3dab77fb | 3037 | |
59ed446f | 3038 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3039 | if (read_format & PERF_FORMAT_ID) |
3040 | values[n++] = primary_event_id(sub); | |
3041 | ||
3042 | size = n * sizeof(u64); | |
3043 | ||
184d3da8 | 3044 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3045 | ret = -EFAULT; |
3046 | goto unlock; | |
3047 | } | |
abf4868b PZ |
3048 | |
3049 | ret += size; | |
3dab77fb | 3050 | } |
6f10581a PZ |
3051 | unlock: |
3052 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3053 | |
abf4868b | 3054 | return ret; |
3dab77fb PZ |
3055 | } |
3056 | ||
cdd6c482 | 3057 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3058 | u64 read_format, char __user *buf) |
3059 | { | |
59ed446f | 3060 | u64 enabled, running; |
3dab77fb PZ |
3061 | u64 values[4]; |
3062 | int n = 0; | |
3063 | ||
59ed446f PZ |
3064 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3065 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3066 | values[n++] = enabled; | |
3067 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3068 | values[n++] = running; | |
3dab77fb | 3069 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3070 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3071 | |
3072 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3073 | return -EFAULT; | |
3074 | ||
3075 | return n * sizeof(u64); | |
3076 | } | |
3077 | ||
0793a61d | 3078 | /* |
cdd6c482 | 3079 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3080 | */ |
3081 | static ssize_t | |
cdd6c482 | 3082 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3083 | { |
cdd6c482 | 3084 | u64 read_format = event->attr.read_format; |
3dab77fb | 3085 | int ret; |
0793a61d | 3086 | |
3b6f9e5c | 3087 | /* |
cdd6c482 | 3088 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3089 | * error state (i.e. because it was pinned but it couldn't be |
3090 | * scheduled on to the CPU at some point). | |
3091 | */ | |
cdd6c482 | 3092 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3093 | return 0; |
3094 | ||
c320c7b7 | 3095 | if (count < event->read_size) |
3dab77fb PZ |
3096 | return -ENOSPC; |
3097 | ||
cdd6c482 | 3098 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3099 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3100 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3101 | else |
cdd6c482 | 3102 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3103 | |
3dab77fb | 3104 | return ret; |
0793a61d TG |
3105 | } |
3106 | ||
0793a61d TG |
3107 | static ssize_t |
3108 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3109 | { | |
cdd6c482 | 3110 | struct perf_event *event = file->private_data; |
0793a61d | 3111 | |
cdd6c482 | 3112 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3113 | } |
3114 | ||
3115 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3116 | { | |
cdd6c482 | 3117 | struct perf_event *event = file->private_data; |
76369139 | 3118 | struct ring_buffer *rb; |
c33a0bc4 | 3119 | unsigned int events = POLL_HUP; |
c7138f37 | 3120 | |
10c6db11 PZ |
3121 | /* |
3122 | * Race between perf_event_set_output() and perf_poll(): perf_poll() | |
3123 | * grabs the rb reference but perf_event_set_output() overrides it. | |
3124 | * Here is the timeline for two threads T1, T2: | |
3125 | * t0: T1, rb = rcu_dereference(event->rb) | |
3126 | * t1: T2, old_rb = event->rb | |
3127 | * t2: T2, event->rb = new rb | |
3128 | * t3: T2, ring_buffer_detach(old_rb) | |
3129 | * t4: T1, ring_buffer_attach(rb1) | |
3130 | * t5: T1, poll_wait(event->waitq) | |
3131 | * | |
3132 | * To avoid this problem, we grab mmap_mutex in perf_poll() | |
3133 | * thereby ensuring that the assignment of the new ring buffer | |
3134 | * and the detachment of the old buffer appear atomic to perf_poll() | |
3135 | */ | |
3136 | mutex_lock(&event->mmap_mutex); | |
3137 | ||
c7138f37 | 3138 | rcu_read_lock(); |
76369139 | 3139 | rb = rcu_dereference(event->rb); |
10c6db11 PZ |
3140 | if (rb) { |
3141 | ring_buffer_attach(event, rb); | |
76369139 | 3142 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3143 | } |
c7138f37 | 3144 | rcu_read_unlock(); |
0793a61d | 3145 | |
10c6db11 PZ |
3146 | mutex_unlock(&event->mmap_mutex); |
3147 | ||
cdd6c482 | 3148 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3149 | |
0793a61d TG |
3150 | return events; |
3151 | } | |
3152 | ||
cdd6c482 | 3153 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3154 | { |
cdd6c482 | 3155 | (void)perf_event_read(event); |
e7850595 | 3156 | local64_set(&event->count, 0); |
cdd6c482 | 3157 | perf_event_update_userpage(event); |
3df5edad PZ |
3158 | } |
3159 | ||
c93f7669 | 3160 | /* |
cdd6c482 IM |
3161 | * Holding the top-level event's child_mutex means that any |
3162 | * descendant process that has inherited this event will block | |
3163 | * in sync_child_event if it goes to exit, thus satisfying the | |
3164 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3165 | */ |
cdd6c482 IM |
3166 | static void perf_event_for_each_child(struct perf_event *event, |
3167 | void (*func)(struct perf_event *)) | |
3df5edad | 3168 | { |
cdd6c482 | 3169 | struct perf_event *child; |
3df5edad | 3170 | |
cdd6c482 IM |
3171 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3172 | mutex_lock(&event->child_mutex); | |
3173 | func(event); | |
3174 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3175 | func(child); |
cdd6c482 | 3176 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3177 | } |
3178 | ||
cdd6c482 IM |
3179 | static void perf_event_for_each(struct perf_event *event, |
3180 | void (*func)(struct perf_event *)) | |
3df5edad | 3181 | { |
cdd6c482 IM |
3182 | struct perf_event_context *ctx = event->ctx; |
3183 | struct perf_event *sibling; | |
3df5edad | 3184 | |
75f937f2 PZ |
3185 | WARN_ON_ONCE(ctx->parent_ctx); |
3186 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3187 | event = event->group_leader; |
75f937f2 | 3188 | |
cdd6c482 | 3189 | perf_event_for_each_child(event, func); |
cdd6c482 | 3190 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3191 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3192 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3193 | } |
3194 | ||
cdd6c482 | 3195 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3196 | { |
cdd6c482 | 3197 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3198 | int ret = 0; |
3199 | u64 value; | |
3200 | ||
6c7e550f | 3201 | if (!is_sampling_event(event)) |
08247e31 PZ |
3202 | return -EINVAL; |
3203 | ||
ad0cf347 | 3204 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3205 | return -EFAULT; |
3206 | ||
3207 | if (!value) | |
3208 | return -EINVAL; | |
3209 | ||
e625cce1 | 3210 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3211 | if (event->attr.freq) { |
3212 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3213 | ret = -EINVAL; |
3214 | goto unlock; | |
3215 | } | |
3216 | ||
cdd6c482 | 3217 | event->attr.sample_freq = value; |
08247e31 | 3218 | } else { |
cdd6c482 IM |
3219 | event->attr.sample_period = value; |
3220 | event->hw.sample_period = value; | |
08247e31 PZ |
3221 | } |
3222 | unlock: | |
e625cce1 | 3223 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3224 | |
3225 | return ret; | |
3226 | } | |
3227 | ||
ac9721f3 PZ |
3228 | static const struct file_operations perf_fops; |
3229 | ||
3230 | static struct perf_event *perf_fget_light(int fd, int *fput_needed) | |
3231 | { | |
3232 | struct file *file; | |
3233 | ||
3234 | file = fget_light(fd, fput_needed); | |
3235 | if (!file) | |
3236 | return ERR_PTR(-EBADF); | |
3237 | ||
3238 | if (file->f_op != &perf_fops) { | |
3239 | fput_light(file, *fput_needed); | |
3240 | *fput_needed = 0; | |
3241 | return ERR_PTR(-EBADF); | |
3242 | } | |
3243 | ||
3244 | return file->private_data; | |
3245 | } | |
3246 | ||
3247 | static int perf_event_set_output(struct perf_event *event, | |
3248 | struct perf_event *output_event); | |
6fb2915d | 3249 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3250 | |
d859e29f PM |
3251 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3252 | { | |
cdd6c482 IM |
3253 | struct perf_event *event = file->private_data; |
3254 | void (*func)(struct perf_event *); | |
3df5edad | 3255 | u32 flags = arg; |
d859e29f PM |
3256 | |
3257 | switch (cmd) { | |
cdd6c482 IM |
3258 | case PERF_EVENT_IOC_ENABLE: |
3259 | func = perf_event_enable; | |
d859e29f | 3260 | break; |
cdd6c482 IM |
3261 | case PERF_EVENT_IOC_DISABLE: |
3262 | func = perf_event_disable; | |
79f14641 | 3263 | break; |
cdd6c482 IM |
3264 | case PERF_EVENT_IOC_RESET: |
3265 | func = perf_event_reset; | |
6de6a7b9 | 3266 | break; |
3df5edad | 3267 | |
cdd6c482 IM |
3268 | case PERF_EVENT_IOC_REFRESH: |
3269 | return perf_event_refresh(event, arg); | |
08247e31 | 3270 | |
cdd6c482 IM |
3271 | case PERF_EVENT_IOC_PERIOD: |
3272 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3273 | |
cdd6c482 | 3274 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 PZ |
3275 | { |
3276 | struct perf_event *output_event = NULL; | |
3277 | int fput_needed = 0; | |
3278 | int ret; | |
3279 | ||
3280 | if (arg != -1) { | |
3281 | output_event = perf_fget_light(arg, &fput_needed); | |
3282 | if (IS_ERR(output_event)) | |
3283 | return PTR_ERR(output_event); | |
3284 | } | |
3285 | ||
3286 | ret = perf_event_set_output(event, output_event); | |
3287 | if (output_event) | |
3288 | fput_light(output_event->filp, fput_needed); | |
3289 | ||
3290 | return ret; | |
3291 | } | |
a4be7c27 | 3292 | |
6fb2915d LZ |
3293 | case PERF_EVENT_IOC_SET_FILTER: |
3294 | return perf_event_set_filter(event, (void __user *)arg); | |
3295 | ||
d859e29f | 3296 | default: |
3df5edad | 3297 | return -ENOTTY; |
d859e29f | 3298 | } |
3df5edad PZ |
3299 | |
3300 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3301 | perf_event_for_each(event, func); |
3df5edad | 3302 | else |
cdd6c482 | 3303 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3304 | |
3305 | return 0; | |
d859e29f PM |
3306 | } |
3307 | ||
cdd6c482 | 3308 | int perf_event_task_enable(void) |
771d7cde | 3309 | { |
cdd6c482 | 3310 | struct perf_event *event; |
771d7cde | 3311 | |
cdd6c482 IM |
3312 | mutex_lock(¤t->perf_event_mutex); |
3313 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3314 | perf_event_for_each_child(event, perf_event_enable); | |
3315 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3316 | |
3317 | return 0; | |
3318 | } | |
3319 | ||
cdd6c482 | 3320 | int perf_event_task_disable(void) |
771d7cde | 3321 | { |
cdd6c482 | 3322 | struct perf_event *event; |
771d7cde | 3323 | |
cdd6c482 IM |
3324 | mutex_lock(¤t->perf_event_mutex); |
3325 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3326 | perf_event_for_each_child(event, perf_event_disable); | |
3327 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3328 | |
3329 | return 0; | |
3330 | } | |
3331 | ||
cdd6c482 | 3332 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3333 | { |
a4eaf7f1 PZ |
3334 | if (event->hw.state & PERF_HES_STOPPED) |
3335 | return 0; | |
3336 | ||
cdd6c482 | 3337 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3338 | return 0; |
3339 | ||
35edc2a5 | 3340 | return event->pmu->event_idx(event); |
194002b2 PZ |
3341 | } |
3342 | ||
c4794295 | 3343 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3344 | u64 *now, |
7f310a5d EM |
3345 | u64 *enabled, |
3346 | u64 *running) | |
c4794295 | 3347 | { |
e3f3541c | 3348 | u64 ctx_time; |
c4794295 | 3349 | |
e3f3541c PZ |
3350 | *now = perf_clock(); |
3351 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3352 | *enabled = ctx_time - event->tstamp_enabled; |
3353 | *running = ctx_time - event->tstamp_running; | |
3354 | } | |
3355 | ||
c7206205 | 3356 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3357 | { |
3358 | } | |
3359 | ||
38ff667b PZ |
3360 | /* |
3361 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3362 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3363 | * code calls this from NMI context. | |
3364 | */ | |
cdd6c482 | 3365 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3366 | { |
cdd6c482 | 3367 | struct perf_event_mmap_page *userpg; |
76369139 | 3368 | struct ring_buffer *rb; |
e3f3541c | 3369 | u64 enabled, running, now; |
38ff667b PZ |
3370 | |
3371 | rcu_read_lock(); | |
0d641208 EM |
3372 | /* |
3373 | * compute total_time_enabled, total_time_running | |
3374 | * based on snapshot values taken when the event | |
3375 | * was last scheduled in. | |
3376 | * | |
3377 | * we cannot simply called update_context_time() | |
3378 | * because of locking issue as we can be called in | |
3379 | * NMI context | |
3380 | */ | |
e3f3541c | 3381 | calc_timer_values(event, &now, &enabled, &running); |
76369139 FW |
3382 | rb = rcu_dereference(event->rb); |
3383 | if (!rb) | |
38ff667b PZ |
3384 | goto unlock; |
3385 | ||
76369139 | 3386 | userpg = rb->user_page; |
37d81828 | 3387 | |
7b732a75 PZ |
3388 | /* |
3389 | * Disable preemption so as to not let the corresponding user-space | |
3390 | * spin too long if we get preempted. | |
3391 | */ | |
3392 | preempt_disable(); | |
37d81828 | 3393 | ++userpg->lock; |
92f22a38 | 3394 | barrier(); |
cdd6c482 | 3395 | userpg->index = perf_event_index(event); |
b5e58793 | 3396 | userpg->offset = perf_event_count(event); |
365a4038 | 3397 | if (userpg->index) |
e7850595 | 3398 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3399 | |
0d641208 | 3400 | userpg->time_enabled = enabled + |
cdd6c482 | 3401 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3402 | |
0d641208 | 3403 | userpg->time_running = running + |
cdd6c482 | 3404 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3405 | |
c7206205 | 3406 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3407 | |
92f22a38 | 3408 | barrier(); |
37d81828 | 3409 | ++userpg->lock; |
7b732a75 | 3410 | preempt_enable(); |
38ff667b | 3411 | unlock: |
7b732a75 | 3412 | rcu_read_unlock(); |
37d81828 PM |
3413 | } |
3414 | ||
906010b2 PZ |
3415 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3416 | { | |
3417 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3418 | struct ring_buffer *rb; |
906010b2 PZ |
3419 | int ret = VM_FAULT_SIGBUS; |
3420 | ||
3421 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3422 | if (vmf->pgoff == 0) | |
3423 | ret = 0; | |
3424 | return ret; | |
3425 | } | |
3426 | ||
3427 | rcu_read_lock(); | |
76369139 FW |
3428 | rb = rcu_dereference(event->rb); |
3429 | if (!rb) | |
906010b2 PZ |
3430 | goto unlock; |
3431 | ||
3432 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3433 | goto unlock; | |
3434 | ||
76369139 | 3435 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3436 | if (!vmf->page) |
3437 | goto unlock; | |
3438 | ||
3439 | get_page(vmf->page); | |
3440 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3441 | vmf->page->index = vmf->pgoff; | |
3442 | ||
3443 | ret = 0; | |
3444 | unlock: | |
3445 | rcu_read_unlock(); | |
3446 | ||
3447 | return ret; | |
3448 | } | |
3449 | ||
10c6db11 PZ |
3450 | static void ring_buffer_attach(struct perf_event *event, |
3451 | struct ring_buffer *rb) | |
3452 | { | |
3453 | unsigned long flags; | |
3454 | ||
3455 | if (!list_empty(&event->rb_entry)) | |
3456 | return; | |
3457 | ||
3458 | spin_lock_irqsave(&rb->event_lock, flags); | |
3459 | if (!list_empty(&event->rb_entry)) | |
3460 | goto unlock; | |
3461 | ||
3462 | list_add(&event->rb_entry, &rb->event_list); | |
3463 | unlock: | |
3464 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3465 | } | |
3466 | ||
3467 | static void ring_buffer_detach(struct perf_event *event, | |
3468 | struct ring_buffer *rb) | |
3469 | { | |
3470 | unsigned long flags; | |
3471 | ||
3472 | if (list_empty(&event->rb_entry)) | |
3473 | return; | |
3474 | ||
3475 | spin_lock_irqsave(&rb->event_lock, flags); | |
3476 | list_del_init(&event->rb_entry); | |
3477 | wake_up_all(&event->waitq); | |
3478 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3479 | } | |
3480 | ||
3481 | static void ring_buffer_wakeup(struct perf_event *event) | |
3482 | { | |
3483 | struct ring_buffer *rb; | |
3484 | ||
3485 | rcu_read_lock(); | |
3486 | rb = rcu_dereference(event->rb); | |
44b7f4b9 WD |
3487 | if (!rb) |
3488 | goto unlock; | |
3489 | ||
3490 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
10c6db11 | 3491 | wake_up_all(&event->waitq); |
44b7f4b9 WD |
3492 | |
3493 | unlock: | |
10c6db11 PZ |
3494 | rcu_read_unlock(); |
3495 | } | |
3496 | ||
76369139 | 3497 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3498 | { |
76369139 | 3499 | struct ring_buffer *rb; |
906010b2 | 3500 | |
76369139 FW |
3501 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3502 | rb_free(rb); | |
7b732a75 PZ |
3503 | } |
3504 | ||
76369139 | 3505 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3506 | { |
76369139 | 3507 | struct ring_buffer *rb; |
7b732a75 | 3508 | |
ac9721f3 | 3509 | rcu_read_lock(); |
76369139 FW |
3510 | rb = rcu_dereference(event->rb); |
3511 | if (rb) { | |
3512 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3513 | rb = NULL; | |
ac9721f3 PZ |
3514 | } |
3515 | rcu_read_unlock(); | |
3516 | ||
76369139 | 3517 | return rb; |
ac9721f3 PZ |
3518 | } |
3519 | ||
76369139 | 3520 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3521 | { |
10c6db11 PZ |
3522 | struct perf_event *event, *n; |
3523 | unsigned long flags; | |
3524 | ||
76369139 | 3525 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3526 | return; |
7b732a75 | 3527 | |
10c6db11 PZ |
3528 | spin_lock_irqsave(&rb->event_lock, flags); |
3529 | list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) { | |
3530 | list_del_init(&event->rb_entry); | |
3531 | wake_up_all(&event->waitq); | |
3532 | } | |
3533 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3534 | ||
76369139 | 3535 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3536 | } |
3537 | ||
3538 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3539 | { | |
cdd6c482 | 3540 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3541 | |
cdd6c482 | 3542 | atomic_inc(&event->mmap_count); |
7b732a75 PZ |
3543 | } |
3544 | ||
3545 | static void perf_mmap_close(struct vm_area_struct *vma) | |
3546 | { | |
cdd6c482 | 3547 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3548 | |
cdd6c482 | 3549 | if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) { |
76369139 | 3550 | unsigned long size = perf_data_size(event->rb); |
ac9721f3 | 3551 | struct user_struct *user = event->mmap_user; |
76369139 | 3552 | struct ring_buffer *rb = event->rb; |
789f90fc | 3553 | |
906010b2 | 3554 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm); |
bc3e53f6 | 3555 | vma->vm_mm->pinned_vm -= event->mmap_locked; |
76369139 | 3556 | rcu_assign_pointer(event->rb, NULL); |
10c6db11 | 3557 | ring_buffer_detach(event, rb); |
cdd6c482 | 3558 | mutex_unlock(&event->mmap_mutex); |
ac9721f3 | 3559 | |
76369139 | 3560 | ring_buffer_put(rb); |
ac9721f3 | 3561 | free_uid(user); |
7b732a75 | 3562 | } |
37d81828 PM |
3563 | } |
3564 | ||
f0f37e2f | 3565 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3566 | .open = perf_mmap_open, |
3567 | .close = perf_mmap_close, | |
3568 | .fault = perf_mmap_fault, | |
3569 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3570 | }; |
3571 | ||
3572 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3573 | { | |
cdd6c482 | 3574 | struct perf_event *event = file->private_data; |
22a4f650 | 3575 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3576 | struct user_struct *user = current_user(); |
22a4f650 | 3577 | unsigned long locked, lock_limit; |
76369139 | 3578 | struct ring_buffer *rb; |
7b732a75 PZ |
3579 | unsigned long vma_size; |
3580 | unsigned long nr_pages; | |
789f90fc | 3581 | long user_extra, extra; |
d57e34fd | 3582 | int ret = 0, flags = 0; |
37d81828 | 3583 | |
c7920614 PZ |
3584 | /* |
3585 | * Don't allow mmap() of inherited per-task counters. This would | |
3586 | * create a performance issue due to all children writing to the | |
76369139 | 3587 | * same rb. |
c7920614 PZ |
3588 | */ |
3589 | if (event->cpu == -1 && event->attr.inherit) | |
3590 | return -EINVAL; | |
3591 | ||
43a21ea8 | 3592 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3593 | return -EINVAL; |
7b732a75 PZ |
3594 | |
3595 | vma_size = vma->vm_end - vma->vm_start; | |
3596 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3597 | ||
7730d865 | 3598 | /* |
76369139 | 3599 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3600 | * can do bitmasks instead of modulo. |
3601 | */ | |
3602 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3603 | return -EINVAL; |
3604 | ||
7b732a75 | 3605 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3606 | return -EINVAL; |
3607 | ||
7b732a75 PZ |
3608 | if (vma->vm_pgoff != 0) |
3609 | return -EINVAL; | |
37d81828 | 3610 | |
cdd6c482 IM |
3611 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3612 | mutex_lock(&event->mmap_mutex); | |
76369139 FW |
3613 | if (event->rb) { |
3614 | if (event->rb->nr_pages == nr_pages) | |
3615 | atomic_inc(&event->rb->refcount); | |
ac9721f3 | 3616 | else |
ebb3c4c4 PZ |
3617 | ret = -EINVAL; |
3618 | goto unlock; | |
3619 | } | |
3620 | ||
789f90fc | 3621 | user_extra = nr_pages + 1; |
cdd6c482 | 3622 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
3623 | |
3624 | /* | |
3625 | * Increase the limit linearly with more CPUs: | |
3626 | */ | |
3627 | user_lock_limit *= num_online_cpus(); | |
3628 | ||
789f90fc | 3629 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 3630 | |
789f90fc PZ |
3631 | extra = 0; |
3632 | if (user_locked > user_lock_limit) | |
3633 | extra = user_locked - user_lock_limit; | |
7b732a75 | 3634 | |
78d7d407 | 3635 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 3636 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 3637 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 3638 | |
459ec28a IM |
3639 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
3640 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
3641 | ret = -EPERM; |
3642 | goto unlock; | |
3643 | } | |
7b732a75 | 3644 | |
76369139 | 3645 | WARN_ON(event->rb); |
906010b2 | 3646 | |
d57e34fd | 3647 | if (vma->vm_flags & VM_WRITE) |
76369139 | 3648 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 3649 | |
4ec8363d VW |
3650 | rb = rb_alloc(nr_pages, |
3651 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
3652 | event->cpu, flags); | |
3653 | ||
76369139 | 3654 | if (!rb) { |
ac9721f3 | 3655 | ret = -ENOMEM; |
ebb3c4c4 | 3656 | goto unlock; |
ac9721f3 | 3657 | } |
76369139 | 3658 | rcu_assign_pointer(event->rb, rb); |
43a21ea8 | 3659 | |
ac9721f3 PZ |
3660 | atomic_long_add(user_extra, &user->locked_vm); |
3661 | event->mmap_locked = extra; | |
3662 | event->mmap_user = get_current_user(); | |
bc3e53f6 | 3663 | vma->vm_mm->pinned_vm += event->mmap_locked; |
ac9721f3 | 3664 | |
9a0f05cb PZ |
3665 | perf_event_update_userpage(event); |
3666 | ||
ebb3c4c4 | 3667 | unlock: |
ac9721f3 PZ |
3668 | if (!ret) |
3669 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 3670 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 3671 | |
37d81828 PM |
3672 | vma->vm_flags |= VM_RESERVED; |
3673 | vma->vm_ops = &perf_mmap_vmops; | |
7b732a75 PZ |
3674 | |
3675 | return ret; | |
37d81828 PM |
3676 | } |
3677 | ||
3c446b3d PZ |
3678 | static int perf_fasync(int fd, struct file *filp, int on) |
3679 | { | |
3c446b3d | 3680 | struct inode *inode = filp->f_path.dentry->d_inode; |
cdd6c482 | 3681 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
3682 | int retval; |
3683 | ||
3684 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 3685 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
3686 | mutex_unlock(&inode->i_mutex); |
3687 | ||
3688 | if (retval < 0) | |
3689 | return retval; | |
3690 | ||
3691 | return 0; | |
3692 | } | |
3693 | ||
0793a61d | 3694 | static const struct file_operations perf_fops = { |
3326c1ce | 3695 | .llseek = no_llseek, |
0793a61d TG |
3696 | .release = perf_release, |
3697 | .read = perf_read, | |
3698 | .poll = perf_poll, | |
d859e29f PM |
3699 | .unlocked_ioctl = perf_ioctl, |
3700 | .compat_ioctl = perf_ioctl, | |
37d81828 | 3701 | .mmap = perf_mmap, |
3c446b3d | 3702 | .fasync = perf_fasync, |
0793a61d TG |
3703 | }; |
3704 | ||
925d519a | 3705 | /* |
cdd6c482 | 3706 | * Perf event wakeup |
925d519a PZ |
3707 | * |
3708 | * If there's data, ensure we set the poll() state and publish everything | |
3709 | * to user-space before waking everybody up. | |
3710 | */ | |
3711 | ||
cdd6c482 | 3712 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 3713 | { |
10c6db11 | 3714 | ring_buffer_wakeup(event); |
4c9e2542 | 3715 | |
cdd6c482 IM |
3716 | if (event->pending_kill) { |
3717 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
3718 | event->pending_kill = 0; | |
4c9e2542 | 3719 | } |
925d519a PZ |
3720 | } |
3721 | ||
e360adbe | 3722 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 3723 | { |
cdd6c482 IM |
3724 | struct perf_event *event = container_of(entry, |
3725 | struct perf_event, pending); | |
79f14641 | 3726 | |
cdd6c482 IM |
3727 | if (event->pending_disable) { |
3728 | event->pending_disable = 0; | |
3729 | __perf_event_disable(event); | |
79f14641 PZ |
3730 | } |
3731 | ||
cdd6c482 IM |
3732 | if (event->pending_wakeup) { |
3733 | event->pending_wakeup = 0; | |
3734 | perf_event_wakeup(event); | |
79f14641 PZ |
3735 | } |
3736 | } | |
3737 | ||
39447b38 ZY |
3738 | /* |
3739 | * We assume there is only KVM supporting the callbacks. | |
3740 | * Later on, we might change it to a list if there is | |
3741 | * another virtualization implementation supporting the callbacks. | |
3742 | */ | |
3743 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
3744 | ||
3745 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3746 | { | |
3747 | perf_guest_cbs = cbs; | |
3748 | return 0; | |
3749 | } | |
3750 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
3751 | ||
3752 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3753 | { | |
3754 | perf_guest_cbs = NULL; | |
3755 | return 0; | |
3756 | } | |
3757 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
3758 | ||
c980d109 ACM |
3759 | static void __perf_event_header__init_id(struct perf_event_header *header, |
3760 | struct perf_sample_data *data, | |
3761 | struct perf_event *event) | |
6844c09d ACM |
3762 | { |
3763 | u64 sample_type = event->attr.sample_type; | |
3764 | ||
3765 | data->type = sample_type; | |
3766 | header->size += event->id_header_size; | |
3767 | ||
3768 | if (sample_type & PERF_SAMPLE_TID) { | |
3769 | /* namespace issues */ | |
3770 | data->tid_entry.pid = perf_event_pid(event, current); | |
3771 | data->tid_entry.tid = perf_event_tid(event, current); | |
3772 | } | |
3773 | ||
3774 | if (sample_type & PERF_SAMPLE_TIME) | |
3775 | data->time = perf_clock(); | |
3776 | ||
3777 | if (sample_type & PERF_SAMPLE_ID) | |
3778 | data->id = primary_event_id(event); | |
3779 | ||
3780 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3781 | data->stream_id = event->id; | |
3782 | ||
3783 | if (sample_type & PERF_SAMPLE_CPU) { | |
3784 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
3785 | data->cpu_entry.reserved = 0; | |
3786 | } | |
3787 | } | |
3788 | ||
76369139 FW |
3789 | void perf_event_header__init_id(struct perf_event_header *header, |
3790 | struct perf_sample_data *data, | |
3791 | struct perf_event *event) | |
c980d109 ACM |
3792 | { |
3793 | if (event->attr.sample_id_all) | |
3794 | __perf_event_header__init_id(header, data, event); | |
3795 | } | |
3796 | ||
3797 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
3798 | struct perf_sample_data *data) | |
3799 | { | |
3800 | u64 sample_type = data->type; | |
3801 | ||
3802 | if (sample_type & PERF_SAMPLE_TID) | |
3803 | perf_output_put(handle, data->tid_entry); | |
3804 | ||
3805 | if (sample_type & PERF_SAMPLE_TIME) | |
3806 | perf_output_put(handle, data->time); | |
3807 | ||
3808 | if (sample_type & PERF_SAMPLE_ID) | |
3809 | perf_output_put(handle, data->id); | |
3810 | ||
3811 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3812 | perf_output_put(handle, data->stream_id); | |
3813 | ||
3814 | if (sample_type & PERF_SAMPLE_CPU) | |
3815 | perf_output_put(handle, data->cpu_entry); | |
3816 | } | |
3817 | ||
76369139 FW |
3818 | void perf_event__output_id_sample(struct perf_event *event, |
3819 | struct perf_output_handle *handle, | |
3820 | struct perf_sample_data *sample) | |
c980d109 ACM |
3821 | { |
3822 | if (event->attr.sample_id_all) | |
3823 | __perf_event__output_id_sample(handle, sample); | |
3824 | } | |
3825 | ||
3dab77fb | 3826 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
3827 | struct perf_event *event, |
3828 | u64 enabled, u64 running) | |
3dab77fb | 3829 | { |
cdd6c482 | 3830 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
3831 | u64 values[4]; |
3832 | int n = 0; | |
3833 | ||
b5e58793 | 3834 | values[n++] = perf_event_count(event); |
3dab77fb | 3835 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 3836 | values[n++] = enabled + |
cdd6c482 | 3837 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
3838 | } |
3839 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 3840 | values[n++] = running + |
cdd6c482 | 3841 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
3842 | } |
3843 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 3844 | values[n++] = primary_event_id(event); |
3dab77fb | 3845 | |
76369139 | 3846 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
3847 | } |
3848 | ||
3849 | /* | |
cdd6c482 | 3850 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
3851 | */ |
3852 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
3853 | struct perf_event *event, |
3854 | u64 enabled, u64 running) | |
3dab77fb | 3855 | { |
cdd6c482 IM |
3856 | struct perf_event *leader = event->group_leader, *sub; |
3857 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
3858 | u64 values[5]; |
3859 | int n = 0; | |
3860 | ||
3861 | values[n++] = 1 + leader->nr_siblings; | |
3862 | ||
3863 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 3864 | values[n++] = enabled; |
3dab77fb PZ |
3865 | |
3866 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 3867 | values[n++] = running; |
3dab77fb | 3868 | |
cdd6c482 | 3869 | if (leader != event) |
3dab77fb PZ |
3870 | leader->pmu->read(leader); |
3871 | ||
b5e58793 | 3872 | values[n++] = perf_event_count(leader); |
3dab77fb | 3873 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3874 | values[n++] = primary_event_id(leader); |
3dab77fb | 3875 | |
76369139 | 3876 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 3877 | |
65abc865 | 3878 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
3879 | n = 0; |
3880 | ||
cdd6c482 | 3881 | if (sub != event) |
3dab77fb PZ |
3882 | sub->pmu->read(sub); |
3883 | ||
b5e58793 | 3884 | values[n++] = perf_event_count(sub); |
3dab77fb | 3885 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3886 | values[n++] = primary_event_id(sub); |
3dab77fb | 3887 | |
76369139 | 3888 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
3889 | } |
3890 | } | |
3891 | ||
eed01528 SE |
3892 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
3893 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3894 | ||
3dab77fb | 3895 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 3896 | struct perf_event *event) |
3dab77fb | 3897 | { |
e3f3541c | 3898 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
3899 | u64 read_format = event->attr.read_format; |
3900 | ||
3901 | /* | |
3902 | * compute total_time_enabled, total_time_running | |
3903 | * based on snapshot values taken when the event | |
3904 | * was last scheduled in. | |
3905 | * | |
3906 | * we cannot simply called update_context_time() | |
3907 | * because of locking issue as we are called in | |
3908 | * NMI context | |
3909 | */ | |
c4794295 | 3910 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 3911 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 3912 | |
cdd6c482 | 3913 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 3914 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 3915 | else |
eed01528 | 3916 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
3917 | } |
3918 | ||
5622f295 MM |
3919 | void perf_output_sample(struct perf_output_handle *handle, |
3920 | struct perf_event_header *header, | |
3921 | struct perf_sample_data *data, | |
cdd6c482 | 3922 | struct perf_event *event) |
5622f295 MM |
3923 | { |
3924 | u64 sample_type = data->type; | |
3925 | ||
3926 | perf_output_put(handle, *header); | |
3927 | ||
3928 | if (sample_type & PERF_SAMPLE_IP) | |
3929 | perf_output_put(handle, data->ip); | |
3930 | ||
3931 | if (sample_type & PERF_SAMPLE_TID) | |
3932 | perf_output_put(handle, data->tid_entry); | |
3933 | ||
3934 | if (sample_type & PERF_SAMPLE_TIME) | |
3935 | perf_output_put(handle, data->time); | |
3936 | ||
3937 | if (sample_type & PERF_SAMPLE_ADDR) | |
3938 | perf_output_put(handle, data->addr); | |
3939 | ||
3940 | if (sample_type & PERF_SAMPLE_ID) | |
3941 | perf_output_put(handle, data->id); | |
3942 | ||
3943 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3944 | perf_output_put(handle, data->stream_id); | |
3945 | ||
3946 | if (sample_type & PERF_SAMPLE_CPU) | |
3947 | perf_output_put(handle, data->cpu_entry); | |
3948 | ||
3949 | if (sample_type & PERF_SAMPLE_PERIOD) | |
3950 | perf_output_put(handle, data->period); | |
3951 | ||
3952 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 3953 | perf_output_read(handle, event); |
5622f295 MM |
3954 | |
3955 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
3956 | if (data->callchain) { | |
3957 | int size = 1; | |
3958 | ||
3959 | if (data->callchain) | |
3960 | size += data->callchain->nr; | |
3961 | ||
3962 | size *= sizeof(u64); | |
3963 | ||
76369139 | 3964 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
3965 | } else { |
3966 | u64 nr = 0; | |
3967 | perf_output_put(handle, nr); | |
3968 | } | |
3969 | } | |
3970 | ||
3971 | if (sample_type & PERF_SAMPLE_RAW) { | |
3972 | if (data->raw) { | |
3973 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
3974 | __output_copy(handle, data->raw->data, |
3975 | data->raw->size); | |
5622f295 MM |
3976 | } else { |
3977 | struct { | |
3978 | u32 size; | |
3979 | u32 data; | |
3980 | } raw = { | |
3981 | .size = sizeof(u32), | |
3982 | .data = 0, | |
3983 | }; | |
3984 | perf_output_put(handle, raw); | |
3985 | } | |
3986 | } | |
a7ac67ea PZ |
3987 | |
3988 | if (!event->attr.watermark) { | |
3989 | int wakeup_events = event->attr.wakeup_events; | |
3990 | ||
3991 | if (wakeup_events) { | |
3992 | struct ring_buffer *rb = handle->rb; | |
3993 | int events = local_inc_return(&rb->events); | |
3994 | ||
3995 | if (events >= wakeup_events) { | |
3996 | local_sub(wakeup_events, &rb->events); | |
3997 | local_inc(&rb->wakeup); | |
3998 | } | |
3999 | } | |
4000 | } | |
bce38cd5 SE |
4001 | |
4002 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4003 | if (data->br_stack) { | |
4004 | size_t size; | |
4005 | ||
4006 | size = data->br_stack->nr | |
4007 | * sizeof(struct perf_branch_entry); | |
4008 | ||
4009 | perf_output_put(handle, data->br_stack->nr); | |
4010 | perf_output_copy(handle, data->br_stack->entries, size); | |
4011 | } else { | |
4012 | /* | |
4013 | * we always store at least the value of nr | |
4014 | */ | |
4015 | u64 nr = 0; | |
4016 | perf_output_put(handle, nr); | |
4017 | } | |
4018 | } | |
5622f295 MM |
4019 | } |
4020 | ||
4021 | void perf_prepare_sample(struct perf_event_header *header, | |
4022 | struct perf_sample_data *data, | |
cdd6c482 | 4023 | struct perf_event *event, |
5622f295 | 4024 | struct pt_regs *regs) |
7b732a75 | 4025 | { |
cdd6c482 | 4026 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4027 | |
cdd6c482 | 4028 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4029 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4030 | |
4031 | header->misc = 0; | |
4032 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4033 | |
c980d109 | 4034 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4035 | |
c320c7b7 | 4036 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4037 | data->ip = perf_instruction_pointer(regs); |
4038 | ||
b23f3325 | 4039 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4040 | int size = 1; |
394ee076 | 4041 | |
e6dab5ff | 4042 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4043 | |
4044 | if (data->callchain) | |
4045 | size += data->callchain->nr; | |
4046 | ||
4047 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4048 | } |
4049 | ||
3a43ce68 | 4050 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4051 | int size = sizeof(u32); |
4052 | ||
4053 | if (data->raw) | |
4054 | size += data->raw->size; | |
4055 | else | |
4056 | size += sizeof(u32); | |
4057 | ||
4058 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4059 | header->size += size; |
7f453c24 | 4060 | } |
bce38cd5 SE |
4061 | |
4062 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4063 | int size = sizeof(u64); /* nr */ | |
4064 | if (data->br_stack) { | |
4065 | size += data->br_stack->nr | |
4066 | * sizeof(struct perf_branch_entry); | |
4067 | } | |
4068 | header->size += size; | |
4069 | } | |
5622f295 | 4070 | } |
7f453c24 | 4071 | |
a8b0ca17 | 4072 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4073 | struct perf_sample_data *data, |
4074 | struct pt_regs *regs) | |
4075 | { | |
4076 | struct perf_output_handle handle; | |
4077 | struct perf_event_header header; | |
689802b2 | 4078 | |
927c7a9e FW |
4079 | /* protect the callchain buffers */ |
4080 | rcu_read_lock(); | |
4081 | ||
cdd6c482 | 4082 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4083 | |
a7ac67ea | 4084 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4085 | goto exit; |
0322cd6e | 4086 | |
cdd6c482 | 4087 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4088 | |
8a057d84 | 4089 | perf_output_end(&handle); |
927c7a9e FW |
4090 | |
4091 | exit: | |
4092 | rcu_read_unlock(); | |
0322cd6e PZ |
4093 | } |
4094 | ||
38b200d6 | 4095 | /* |
cdd6c482 | 4096 | * read event_id |
38b200d6 PZ |
4097 | */ |
4098 | ||
4099 | struct perf_read_event { | |
4100 | struct perf_event_header header; | |
4101 | ||
4102 | u32 pid; | |
4103 | u32 tid; | |
38b200d6 PZ |
4104 | }; |
4105 | ||
4106 | static void | |
cdd6c482 | 4107 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4108 | struct task_struct *task) |
4109 | { | |
4110 | struct perf_output_handle handle; | |
c980d109 | 4111 | struct perf_sample_data sample; |
dfc65094 | 4112 | struct perf_read_event read_event = { |
38b200d6 | 4113 | .header = { |
cdd6c482 | 4114 | .type = PERF_RECORD_READ, |
38b200d6 | 4115 | .misc = 0, |
c320c7b7 | 4116 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4117 | }, |
cdd6c482 IM |
4118 | .pid = perf_event_pid(event, task), |
4119 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4120 | }; |
3dab77fb | 4121 | int ret; |
38b200d6 | 4122 | |
c980d109 | 4123 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4124 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4125 | if (ret) |
4126 | return; | |
4127 | ||
dfc65094 | 4128 | perf_output_put(&handle, read_event); |
cdd6c482 | 4129 | perf_output_read(&handle, event); |
c980d109 | 4130 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4131 | |
38b200d6 PZ |
4132 | perf_output_end(&handle); |
4133 | } | |
4134 | ||
60313ebe | 4135 | /* |
9f498cc5 PZ |
4136 | * task tracking -- fork/exit |
4137 | * | |
3af9e859 | 4138 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
60313ebe PZ |
4139 | */ |
4140 | ||
9f498cc5 | 4141 | struct perf_task_event { |
3a80b4a3 | 4142 | struct task_struct *task; |
cdd6c482 | 4143 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4144 | |
4145 | struct { | |
4146 | struct perf_event_header header; | |
4147 | ||
4148 | u32 pid; | |
4149 | u32 ppid; | |
9f498cc5 PZ |
4150 | u32 tid; |
4151 | u32 ptid; | |
393b2ad8 | 4152 | u64 time; |
cdd6c482 | 4153 | } event_id; |
60313ebe PZ |
4154 | }; |
4155 | ||
cdd6c482 | 4156 | static void perf_event_task_output(struct perf_event *event, |
9f498cc5 | 4157 | struct perf_task_event *task_event) |
60313ebe PZ |
4158 | { |
4159 | struct perf_output_handle handle; | |
c980d109 | 4160 | struct perf_sample_data sample; |
9f498cc5 | 4161 | struct task_struct *task = task_event->task; |
c980d109 | 4162 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4163 | |
c980d109 | 4164 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4165 | |
c980d109 | 4166 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4167 | task_event->event_id.header.size); |
ef60777c | 4168 | if (ret) |
c980d109 | 4169 | goto out; |
60313ebe | 4170 | |
cdd6c482 IM |
4171 | task_event->event_id.pid = perf_event_pid(event, task); |
4172 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4173 | |
cdd6c482 IM |
4174 | task_event->event_id.tid = perf_event_tid(event, task); |
4175 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4176 | |
cdd6c482 | 4177 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4178 | |
c980d109 ACM |
4179 | perf_event__output_id_sample(event, &handle, &sample); |
4180 | ||
60313ebe | 4181 | perf_output_end(&handle); |
c980d109 ACM |
4182 | out: |
4183 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4184 | } |
4185 | ||
cdd6c482 | 4186 | static int perf_event_task_match(struct perf_event *event) |
60313ebe | 4187 | { |
6f93d0a7 | 4188 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4189 | return 0; |
4190 | ||
5632ab12 | 4191 | if (!event_filter_match(event)) |
5d27c23d PZ |
4192 | return 0; |
4193 | ||
3af9e859 EM |
4194 | if (event->attr.comm || event->attr.mmap || |
4195 | event->attr.mmap_data || event->attr.task) | |
60313ebe PZ |
4196 | return 1; |
4197 | ||
4198 | return 0; | |
4199 | } | |
4200 | ||
cdd6c482 | 4201 | static void perf_event_task_ctx(struct perf_event_context *ctx, |
9f498cc5 | 4202 | struct perf_task_event *task_event) |
60313ebe | 4203 | { |
cdd6c482 | 4204 | struct perf_event *event; |
60313ebe | 4205 | |
cdd6c482 IM |
4206 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4207 | if (perf_event_task_match(event)) | |
4208 | perf_event_task_output(event, task_event); | |
60313ebe | 4209 | } |
60313ebe PZ |
4210 | } |
4211 | ||
cdd6c482 | 4212 | static void perf_event_task_event(struct perf_task_event *task_event) |
60313ebe PZ |
4213 | { |
4214 | struct perf_cpu_context *cpuctx; | |
8dc85d54 | 4215 | struct perf_event_context *ctx; |
108b02cf | 4216 | struct pmu *pmu; |
8dc85d54 | 4217 | int ctxn; |
60313ebe | 4218 | |
d6ff86cf | 4219 | rcu_read_lock(); |
108b02cf | 4220 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4221 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4222 | if (cpuctx->active_pmu != pmu) |
4223 | goto next; | |
108b02cf | 4224 | perf_event_task_ctx(&cpuctx->ctx, task_event); |
8dc85d54 PZ |
4225 | |
4226 | ctx = task_event->task_ctx; | |
4227 | if (!ctx) { | |
4228 | ctxn = pmu->task_ctx_nr; | |
4229 | if (ctxn < 0) | |
41945f6c | 4230 | goto next; |
8dc85d54 PZ |
4231 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); |
4232 | } | |
4233 | if (ctx) | |
4234 | perf_event_task_ctx(ctx, task_event); | |
41945f6c PZ |
4235 | next: |
4236 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4237 | } |
60313ebe PZ |
4238 | rcu_read_unlock(); |
4239 | } | |
4240 | ||
cdd6c482 IM |
4241 | static void perf_event_task(struct task_struct *task, |
4242 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4243 | int new) |
60313ebe | 4244 | { |
9f498cc5 | 4245 | struct perf_task_event task_event; |
60313ebe | 4246 | |
cdd6c482 IM |
4247 | if (!atomic_read(&nr_comm_events) && |
4248 | !atomic_read(&nr_mmap_events) && | |
4249 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4250 | return; |
4251 | ||
9f498cc5 | 4252 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4253 | .task = task, |
4254 | .task_ctx = task_ctx, | |
cdd6c482 | 4255 | .event_id = { |
60313ebe | 4256 | .header = { |
cdd6c482 | 4257 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4258 | .misc = 0, |
cdd6c482 | 4259 | .size = sizeof(task_event.event_id), |
60313ebe | 4260 | }, |
573402db PZ |
4261 | /* .pid */ |
4262 | /* .ppid */ | |
9f498cc5 PZ |
4263 | /* .tid */ |
4264 | /* .ptid */ | |
6f93d0a7 | 4265 | .time = perf_clock(), |
60313ebe PZ |
4266 | }, |
4267 | }; | |
4268 | ||
cdd6c482 | 4269 | perf_event_task_event(&task_event); |
9f498cc5 PZ |
4270 | } |
4271 | ||
cdd6c482 | 4272 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4273 | { |
cdd6c482 | 4274 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4275 | } |
4276 | ||
8d1b2d93 PZ |
4277 | /* |
4278 | * comm tracking | |
4279 | */ | |
4280 | ||
4281 | struct perf_comm_event { | |
22a4f650 IM |
4282 | struct task_struct *task; |
4283 | char *comm; | |
8d1b2d93 PZ |
4284 | int comm_size; |
4285 | ||
4286 | struct { | |
4287 | struct perf_event_header header; | |
4288 | ||
4289 | u32 pid; | |
4290 | u32 tid; | |
cdd6c482 | 4291 | } event_id; |
8d1b2d93 PZ |
4292 | }; |
4293 | ||
cdd6c482 | 4294 | static void perf_event_comm_output(struct perf_event *event, |
8d1b2d93 PZ |
4295 | struct perf_comm_event *comm_event) |
4296 | { | |
4297 | struct perf_output_handle handle; | |
c980d109 | 4298 | struct perf_sample_data sample; |
cdd6c482 | 4299 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4300 | int ret; |
4301 | ||
4302 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); | |
4303 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4304 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4305 | |
4306 | if (ret) | |
c980d109 | 4307 | goto out; |
8d1b2d93 | 4308 | |
cdd6c482 IM |
4309 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4310 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4311 | |
cdd6c482 | 4312 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4313 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4314 | comm_event->comm_size); |
c980d109 ACM |
4315 | |
4316 | perf_event__output_id_sample(event, &handle, &sample); | |
4317 | ||
8d1b2d93 | 4318 | perf_output_end(&handle); |
c980d109 ACM |
4319 | out: |
4320 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4321 | } |
4322 | ||
cdd6c482 | 4323 | static int perf_event_comm_match(struct perf_event *event) |
8d1b2d93 | 4324 | { |
6f93d0a7 | 4325 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4326 | return 0; |
4327 | ||
5632ab12 | 4328 | if (!event_filter_match(event)) |
5d27c23d PZ |
4329 | return 0; |
4330 | ||
cdd6c482 | 4331 | if (event->attr.comm) |
8d1b2d93 PZ |
4332 | return 1; |
4333 | ||
4334 | return 0; | |
4335 | } | |
4336 | ||
cdd6c482 | 4337 | static void perf_event_comm_ctx(struct perf_event_context *ctx, |
8d1b2d93 PZ |
4338 | struct perf_comm_event *comm_event) |
4339 | { | |
cdd6c482 | 4340 | struct perf_event *event; |
8d1b2d93 | 4341 | |
cdd6c482 IM |
4342 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4343 | if (perf_event_comm_match(event)) | |
4344 | perf_event_comm_output(event, comm_event); | |
8d1b2d93 | 4345 | } |
8d1b2d93 PZ |
4346 | } |
4347 | ||
cdd6c482 | 4348 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 PZ |
4349 | { |
4350 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4351 | struct perf_event_context *ctx; |
413ee3b4 | 4352 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4353 | unsigned int size; |
108b02cf | 4354 | struct pmu *pmu; |
8dc85d54 | 4355 | int ctxn; |
8d1b2d93 | 4356 | |
413ee3b4 | 4357 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4358 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4359 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4360 | |
4361 | comm_event->comm = comm; | |
4362 | comm_event->comm_size = size; | |
4363 | ||
cdd6c482 | 4364 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
f6595f3a | 4365 | rcu_read_lock(); |
108b02cf | 4366 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4367 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4368 | if (cpuctx->active_pmu != pmu) |
4369 | goto next; | |
108b02cf | 4370 | perf_event_comm_ctx(&cpuctx->ctx, comm_event); |
8dc85d54 PZ |
4371 | |
4372 | ctxn = pmu->task_ctx_nr; | |
4373 | if (ctxn < 0) | |
41945f6c | 4374 | goto next; |
8dc85d54 PZ |
4375 | |
4376 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4377 | if (ctx) | |
4378 | perf_event_comm_ctx(ctx, comm_event); | |
41945f6c PZ |
4379 | next: |
4380 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4381 | } |
665c2142 | 4382 | rcu_read_unlock(); |
8d1b2d93 PZ |
4383 | } |
4384 | ||
cdd6c482 | 4385 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4386 | { |
9ee318a7 | 4387 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4388 | struct perf_event_context *ctx; |
4389 | int ctxn; | |
9ee318a7 | 4390 | |
8dc85d54 PZ |
4391 | for_each_task_context_nr(ctxn) { |
4392 | ctx = task->perf_event_ctxp[ctxn]; | |
4393 | if (!ctx) | |
4394 | continue; | |
9ee318a7 | 4395 | |
8dc85d54 PZ |
4396 | perf_event_enable_on_exec(ctx); |
4397 | } | |
9ee318a7 | 4398 | |
cdd6c482 | 4399 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4400 | return; |
a63eaf34 | 4401 | |
9ee318a7 | 4402 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4403 | .task = task, |
573402db PZ |
4404 | /* .comm */ |
4405 | /* .comm_size */ | |
cdd6c482 | 4406 | .event_id = { |
573402db | 4407 | .header = { |
cdd6c482 | 4408 | .type = PERF_RECORD_COMM, |
573402db PZ |
4409 | .misc = 0, |
4410 | /* .size */ | |
4411 | }, | |
4412 | /* .pid */ | |
4413 | /* .tid */ | |
8d1b2d93 PZ |
4414 | }, |
4415 | }; | |
4416 | ||
cdd6c482 | 4417 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
4418 | } |
4419 | ||
0a4a9391 PZ |
4420 | /* |
4421 | * mmap tracking | |
4422 | */ | |
4423 | ||
4424 | struct perf_mmap_event { | |
089dd79d PZ |
4425 | struct vm_area_struct *vma; |
4426 | ||
4427 | const char *file_name; | |
4428 | int file_size; | |
0a4a9391 PZ |
4429 | |
4430 | struct { | |
4431 | struct perf_event_header header; | |
4432 | ||
4433 | u32 pid; | |
4434 | u32 tid; | |
4435 | u64 start; | |
4436 | u64 len; | |
4437 | u64 pgoff; | |
cdd6c482 | 4438 | } event_id; |
0a4a9391 PZ |
4439 | }; |
4440 | ||
cdd6c482 | 4441 | static void perf_event_mmap_output(struct perf_event *event, |
0a4a9391 PZ |
4442 | struct perf_mmap_event *mmap_event) |
4443 | { | |
4444 | struct perf_output_handle handle; | |
c980d109 | 4445 | struct perf_sample_data sample; |
cdd6c482 | 4446 | int size = mmap_event->event_id.header.size; |
c980d109 | 4447 | int ret; |
0a4a9391 | 4448 | |
c980d109 ACM |
4449 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
4450 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4451 | mmap_event->event_id.header.size); |
0a4a9391 | 4452 | if (ret) |
c980d109 | 4453 | goto out; |
0a4a9391 | 4454 | |
cdd6c482 IM |
4455 | mmap_event->event_id.pid = perf_event_pid(event, current); |
4456 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 4457 | |
cdd6c482 | 4458 | perf_output_put(&handle, mmap_event->event_id); |
76369139 | 4459 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 4460 | mmap_event->file_size); |
c980d109 ACM |
4461 | |
4462 | perf_event__output_id_sample(event, &handle, &sample); | |
4463 | ||
78d613eb | 4464 | perf_output_end(&handle); |
c980d109 ACM |
4465 | out: |
4466 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
4467 | } |
4468 | ||
cdd6c482 | 4469 | static int perf_event_mmap_match(struct perf_event *event, |
3af9e859 EM |
4470 | struct perf_mmap_event *mmap_event, |
4471 | int executable) | |
0a4a9391 | 4472 | { |
6f93d0a7 | 4473 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4474 | return 0; |
4475 | ||
5632ab12 | 4476 | if (!event_filter_match(event)) |
5d27c23d PZ |
4477 | return 0; |
4478 | ||
3af9e859 EM |
4479 | if ((!executable && event->attr.mmap_data) || |
4480 | (executable && event->attr.mmap)) | |
0a4a9391 PZ |
4481 | return 1; |
4482 | ||
4483 | return 0; | |
4484 | } | |
4485 | ||
cdd6c482 | 4486 | static void perf_event_mmap_ctx(struct perf_event_context *ctx, |
3af9e859 EM |
4487 | struct perf_mmap_event *mmap_event, |
4488 | int executable) | |
0a4a9391 | 4489 | { |
cdd6c482 | 4490 | struct perf_event *event; |
0a4a9391 | 4491 | |
cdd6c482 | 4492 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
3af9e859 | 4493 | if (perf_event_mmap_match(event, mmap_event, executable)) |
cdd6c482 | 4494 | perf_event_mmap_output(event, mmap_event); |
0a4a9391 | 4495 | } |
0a4a9391 PZ |
4496 | } |
4497 | ||
cdd6c482 | 4498 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 PZ |
4499 | { |
4500 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4501 | struct perf_event_context *ctx; |
089dd79d PZ |
4502 | struct vm_area_struct *vma = mmap_event->vma; |
4503 | struct file *file = vma->vm_file; | |
0a4a9391 PZ |
4504 | unsigned int size; |
4505 | char tmp[16]; | |
4506 | char *buf = NULL; | |
089dd79d | 4507 | const char *name; |
108b02cf | 4508 | struct pmu *pmu; |
8dc85d54 | 4509 | int ctxn; |
0a4a9391 | 4510 | |
413ee3b4 AB |
4511 | memset(tmp, 0, sizeof(tmp)); |
4512 | ||
0a4a9391 | 4513 | if (file) { |
413ee3b4 | 4514 | /* |
76369139 | 4515 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
4516 | * need to add enough zero bytes after the string to handle |
4517 | * the 64bit alignment we do later. | |
4518 | */ | |
4519 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
4520 | if (!buf) { |
4521 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
4522 | goto got_name; | |
4523 | } | |
d3d21c41 | 4524 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
4525 | if (IS_ERR(name)) { |
4526 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
4527 | goto got_name; | |
4528 | } | |
4529 | } else { | |
413ee3b4 AB |
4530 | if (arch_vma_name(mmap_event->vma)) { |
4531 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
4532 | sizeof(tmp)); | |
089dd79d | 4533 | goto got_name; |
413ee3b4 | 4534 | } |
089dd79d PZ |
4535 | |
4536 | if (!vma->vm_mm) { | |
4537 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
4538 | goto got_name; | |
3af9e859 EM |
4539 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
4540 | vma->vm_end >= vma->vm_mm->brk) { | |
4541 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
4542 | goto got_name; | |
4543 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
4544 | vma->vm_end >= vma->vm_mm->start_stack) { | |
4545 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
4546 | goto got_name; | |
089dd79d PZ |
4547 | } |
4548 | ||
0a4a9391 PZ |
4549 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
4550 | goto got_name; | |
4551 | } | |
4552 | ||
4553 | got_name: | |
888fcee0 | 4554 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
4555 | |
4556 | mmap_event->file_name = name; | |
4557 | mmap_event->file_size = size; | |
4558 | ||
cdd6c482 | 4559 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 4560 | |
f6d9dd23 | 4561 | rcu_read_lock(); |
108b02cf | 4562 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4563 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4564 | if (cpuctx->active_pmu != pmu) |
4565 | goto next; | |
108b02cf PZ |
4566 | perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, |
4567 | vma->vm_flags & VM_EXEC); | |
8dc85d54 PZ |
4568 | |
4569 | ctxn = pmu->task_ctx_nr; | |
4570 | if (ctxn < 0) | |
41945f6c | 4571 | goto next; |
8dc85d54 PZ |
4572 | |
4573 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4574 | if (ctx) { | |
4575 | perf_event_mmap_ctx(ctx, mmap_event, | |
4576 | vma->vm_flags & VM_EXEC); | |
4577 | } | |
41945f6c PZ |
4578 | next: |
4579 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4580 | } |
665c2142 PZ |
4581 | rcu_read_unlock(); |
4582 | ||
0a4a9391 PZ |
4583 | kfree(buf); |
4584 | } | |
4585 | ||
3af9e859 | 4586 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 4587 | { |
9ee318a7 PZ |
4588 | struct perf_mmap_event mmap_event; |
4589 | ||
cdd6c482 | 4590 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
4591 | return; |
4592 | ||
4593 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 4594 | .vma = vma, |
573402db PZ |
4595 | /* .file_name */ |
4596 | /* .file_size */ | |
cdd6c482 | 4597 | .event_id = { |
573402db | 4598 | .header = { |
cdd6c482 | 4599 | .type = PERF_RECORD_MMAP, |
39447b38 | 4600 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
4601 | /* .size */ |
4602 | }, | |
4603 | /* .pid */ | |
4604 | /* .tid */ | |
089dd79d PZ |
4605 | .start = vma->vm_start, |
4606 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 4607 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 PZ |
4608 | }, |
4609 | }; | |
4610 | ||
cdd6c482 | 4611 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
4612 | } |
4613 | ||
a78ac325 PZ |
4614 | /* |
4615 | * IRQ throttle logging | |
4616 | */ | |
4617 | ||
cdd6c482 | 4618 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
4619 | { |
4620 | struct perf_output_handle handle; | |
c980d109 | 4621 | struct perf_sample_data sample; |
a78ac325 PZ |
4622 | int ret; |
4623 | ||
4624 | struct { | |
4625 | struct perf_event_header header; | |
4626 | u64 time; | |
cca3f454 | 4627 | u64 id; |
7f453c24 | 4628 | u64 stream_id; |
a78ac325 PZ |
4629 | } throttle_event = { |
4630 | .header = { | |
cdd6c482 | 4631 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
4632 | .misc = 0, |
4633 | .size = sizeof(throttle_event), | |
4634 | }, | |
def0a9b2 | 4635 | .time = perf_clock(), |
cdd6c482 IM |
4636 | .id = primary_event_id(event), |
4637 | .stream_id = event->id, | |
a78ac325 PZ |
4638 | }; |
4639 | ||
966ee4d6 | 4640 | if (enable) |
cdd6c482 | 4641 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 4642 | |
c980d109 ACM |
4643 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
4644 | ||
4645 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4646 | throttle_event.header.size); |
a78ac325 PZ |
4647 | if (ret) |
4648 | return; | |
4649 | ||
4650 | perf_output_put(&handle, throttle_event); | |
c980d109 | 4651 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
4652 | perf_output_end(&handle); |
4653 | } | |
4654 | ||
f6c7d5fe | 4655 | /* |
cdd6c482 | 4656 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
4657 | */ |
4658 | ||
a8b0ca17 | 4659 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4660 | int throttle, struct perf_sample_data *data, |
4661 | struct pt_regs *regs) | |
f6c7d5fe | 4662 | { |
cdd6c482 IM |
4663 | int events = atomic_read(&event->event_limit); |
4664 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 4665 | u64 seq; |
79f14641 PZ |
4666 | int ret = 0; |
4667 | ||
96398826 PZ |
4668 | /* |
4669 | * Non-sampling counters might still use the PMI to fold short | |
4670 | * hardware counters, ignore those. | |
4671 | */ | |
4672 | if (unlikely(!is_sampling_event(event))) | |
4673 | return 0; | |
4674 | ||
e050e3f0 SE |
4675 | seq = __this_cpu_read(perf_throttled_seq); |
4676 | if (seq != hwc->interrupts_seq) { | |
4677 | hwc->interrupts_seq = seq; | |
4678 | hwc->interrupts = 1; | |
4679 | } else { | |
4680 | hwc->interrupts++; | |
4681 | if (unlikely(throttle | |
4682 | && hwc->interrupts >= max_samples_per_tick)) { | |
4683 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
4684 | hwc->interrupts = MAX_INTERRUPTS; |
4685 | perf_log_throttle(event, 0); | |
a78ac325 PZ |
4686 | ret = 1; |
4687 | } | |
e050e3f0 | 4688 | } |
60db5e09 | 4689 | |
cdd6c482 | 4690 | if (event->attr.freq) { |
def0a9b2 | 4691 | u64 now = perf_clock(); |
abd50713 | 4692 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 4693 | |
abd50713 | 4694 | hwc->freq_time_stamp = now; |
bd2b5b12 | 4695 | |
abd50713 | 4696 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 4697 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
4698 | } |
4699 | ||
2023b359 PZ |
4700 | /* |
4701 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 4702 | * events |
2023b359 PZ |
4703 | */ |
4704 | ||
cdd6c482 IM |
4705 | event->pending_kill = POLL_IN; |
4706 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 4707 | ret = 1; |
cdd6c482 | 4708 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
4709 | event->pending_disable = 1; |
4710 | irq_work_queue(&event->pending); | |
79f14641 PZ |
4711 | } |
4712 | ||
453f19ee | 4713 | if (event->overflow_handler) |
a8b0ca17 | 4714 | event->overflow_handler(event, data, regs); |
453f19ee | 4715 | else |
a8b0ca17 | 4716 | perf_event_output(event, data, regs); |
453f19ee | 4717 | |
f506b3dc | 4718 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
4719 | event->pending_wakeup = 1; |
4720 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
4721 | } |
4722 | ||
79f14641 | 4723 | return ret; |
f6c7d5fe PZ |
4724 | } |
4725 | ||
a8b0ca17 | 4726 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4727 | struct perf_sample_data *data, |
4728 | struct pt_regs *regs) | |
850bc73f | 4729 | { |
a8b0ca17 | 4730 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
4731 | } |
4732 | ||
15dbf27c | 4733 | /* |
cdd6c482 | 4734 | * Generic software event infrastructure |
15dbf27c PZ |
4735 | */ |
4736 | ||
b28ab83c PZ |
4737 | struct swevent_htable { |
4738 | struct swevent_hlist *swevent_hlist; | |
4739 | struct mutex hlist_mutex; | |
4740 | int hlist_refcount; | |
4741 | ||
4742 | /* Recursion avoidance in each contexts */ | |
4743 | int recursion[PERF_NR_CONTEXTS]; | |
4744 | }; | |
4745 | ||
4746 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
4747 | ||
7b4b6658 | 4748 | /* |
cdd6c482 IM |
4749 | * We directly increment event->count and keep a second value in |
4750 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
4751 | * is kept in the range [-sample_period, 0] so that we can use the |
4752 | * sign as trigger. | |
4753 | */ | |
4754 | ||
cdd6c482 | 4755 | static u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 4756 | { |
cdd6c482 | 4757 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
4758 | u64 period = hwc->last_period; |
4759 | u64 nr, offset; | |
4760 | s64 old, val; | |
4761 | ||
4762 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
4763 | |
4764 | again: | |
e7850595 | 4765 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
4766 | if (val < 0) |
4767 | return 0; | |
15dbf27c | 4768 | |
7b4b6658 PZ |
4769 | nr = div64_u64(period + val, period); |
4770 | offset = nr * period; | |
4771 | val -= offset; | |
e7850595 | 4772 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 4773 | goto again; |
15dbf27c | 4774 | |
7b4b6658 | 4775 | return nr; |
15dbf27c PZ |
4776 | } |
4777 | ||
0cff784a | 4778 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 4779 | struct perf_sample_data *data, |
5622f295 | 4780 | struct pt_regs *regs) |
15dbf27c | 4781 | { |
cdd6c482 | 4782 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 4783 | int throttle = 0; |
15dbf27c | 4784 | |
0cff784a PZ |
4785 | if (!overflow) |
4786 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 4787 | |
7b4b6658 PZ |
4788 | if (hwc->interrupts == MAX_INTERRUPTS) |
4789 | return; | |
15dbf27c | 4790 | |
7b4b6658 | 4791 | for (; overflow; overflow--) { |
a8b0ca17 | 4792 | if (__perf_event_overflow(event, throttle, |
5622f295 | 4793 | data, regs)) { |
7b4b6658 PZ |
4794 | /* |
4795 | * We inhibit the overflow from happening when | |
4796 | * hwc->interrupts == MAX_INTERRUPTS. | |
4797 | */ | |
4798 | break; | |
4799 | } | |
cf450a73 | 4800 | throttle = 1; |
7b4b6658 | 4801 | } |
15dbf27c PZ |
4802 | } |
4803 | ||
a4eaf7f1 | 4804 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 4805 | struct perf_sample_data *data, |
5622f295 | 4806 | struct pt_regs *regs) |
7b4b6658 | 4807 | { |
cdd6c482 | 4808 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 4809 | |
e7850595 | 4810 | local64_add(nr, &event->count); |
d6d020e9 | 4811 | |
0cff784a PZ |
4812 | if (!regs) |
4813 | return; | |
4814 | ||
6c7e550f | 4815 | if (!is_sampling_event(event)) |
7b4b6658 | 4816 | return; |
d6d020e9 | 4817 | |
5d81e5cf AV |
4818 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
4819 | data->period = nr; | |
4820 | return perf_swevent_overflow(event, 1, data, regs); | |
4821 | } else | |
4822 | data->period = event->hw.last_period; | |
4823 | ||
0cff784a | 4824 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 4825 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 4826 | |
e7850595 | 4827 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 4828 | return; |
df1a132b | 4829 | |
a8b0ca17 | 4830 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
4831 | } |
4832 | ||
f5ffe02e FW |
4833 | static int perf_exclude_event(struct perf_event *event, |
4834 | struct pt_regs *regs) | |
4835 | { | |
a4eaf7f1 | 4836 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 4837 | return 1; |
a4eaf7f1 | 4838 | |
f5ffe02e FW |
4839 | if (regs) { |
4840 | if (event->attr.exclude_user && user_mode(regs)) | |
4841 | return 1; | |
4842 | ||
4843 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
4844 | return 1; | |
4845 | } | |
4846 | ||
4847 | return 0; | |
4848 | } | |
4849 | ||
cdd6c482 | 4850 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 4851 | enum perf_type_id type, |
6fb2915d LZ |
4852 | u32 event_id, |
4853 | struct perf_sample_data *data, | |
4854 | struct pt_regs *regs) | |
15dbf27c | 4855 | { |
cdd6c482 | 4856 | if (event->attr.type != type) |
a21ca2ca | 4857 | return 0; |
f5ffe02e | 4858 | |
cdd6c482 | 4859 | if (event->attr.config != event_id) |
15dbf27c PZ |
4860 | return 0; |
4861 | ||
f5ffe02e FW |
4862 | if (perf_exclude_event(event, regs)) |
4863 | return 0; | |
15dbf27c PZ |
4864 | |
4865 | return 1; | |
4866 | } | |
4867 | ||
76e1d904 FW |
4868 | static inline u64 swevent_hash(u64 type, u32 event_id) |
4869 | { | |
4870 | u64 val = event_id | (type << 32); | |
4871 | ||
4872 | return hash_64(val, SWEVENT_HLIST_BITS); | |
4873 | } | |
4874 | ||
49f135ed FW |
4875 | static inline struct hlist_head * |
4876 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 4877 | { |
49f135ed FW |
4878 | u64 hash = swevent_hash(type, event_id); |
4879 | ||
4880 | return &hlist->heads[hash]; | |
4881 | } | |
76e1d904 | 4882 | |
49f135ed FW |
4883 | /* For the read side: events when they trigger */ |
4884 | static inline struct hlist_head * | |
b28ab83c | 4885 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
4886 | { |
4887 | struct swevent_hlist *hlist; | |
76e1d904 | 4888 | |
b28ab83c | 4889 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
4890 | if (!hlist) |
4891 | return NULL; | |
4892 | ||
49f135ed FW |
4893 | return __find_swevent_head(hlist, type, event_id); |
4894 | } | |
4895 | ||
4896 | /* For the event head insertion and removal in the hlist */ | |
4897 | static inline struct hlist_head * | |
b28ab83c | 4898 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
4899 | { |
4900 | struct swevent_hlist *hlist; | |
4901 | u32 event_id = event->attr.config; | |
4902 | u64 type = event->attr.type; | |
4903 | ||
4904 | /* | |
4905 | * Event scheduling is always serialized against hlist allocation | |
4906 | * and release. Which makes the protected version suitable here. | |
4907 | * The context lock guarantees that. | |
4908 | */ | |
b28ab83c | 4909 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
4910 | lockdep_is_held(&event->ctx->lock)); |
4911 | if (!hlist) | |
4912 | return NULL; | |
4913 | ||
4914 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
4915 | } |
4916 | ||
4917 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 4918 | u64 nr, |
76e1d904 FW |
4919 | struct perf_sample_data *data, |
4920 | struct pt_regs *regs) | |
15dbf27c | 4921 | { |
b28ab83c | 4922 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 4923 | struct perf_event *event; |
76e1d904 FW |
4924 | struct hlist_node *node; |
4925 | struct hlist_head *head; | |
15dbf27c | 4926 | |
76e1d904 | 4927 | rcu_read_lock(); |
b28ab83c | 4928 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
4929 | if (!head) |
4930 | goto end; | |
4931 | ||
4932 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { | |
6fb2915d | 4933 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 4934 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 4935 | } |
76e1d904 FW |
4936 | end: |
4937 | rcu_read_unlock(); | |
15dbf27c PZ |
4938 | } |
4939 | ||
4ed7c92d | 4940 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 4941 | { |
b28ab83c | 4942 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 4943 | |
b28ab83c | 4944 | return get_recursion_context(swhash->recursion); |
96f6d444 | 4945 | } |
645e8cc0 | 4946 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 4947 | |
fa9f90be | 4948 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 4949 | { |
b28ab83c | 4950 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 4951 | |
b28ab83c | 4952 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 4953 | } |
15dbf27c | 4954 | |
a8b0ca17 | 4955 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 4956 | { |
a4234bfc | 4957 | struct perf_sample_data data; |
4ed7c92d PZ |
4958 | int rctx; |
4959 | ||
1c024eca | 4960 | preempt_disable_notrace(); |
4ed7c92d PZ |
4961 | rctx = perf_swevent_get_recursion_context(); |
4962 | if (rctx < 0) | |
4963 | return; | |
a4234bfc | 4964 | |
fd0d000b | 4965 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 4966 | |
a8b0ca17 | 4967 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
4968 | |
4969 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 4970 | preempt_enable_notrace(); |
b8e83514 PZ |
4971 | } |
4972 | ||
cdd6c482 | 4973 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 4974 | { |
15dbf27c PZ |
4975 | } |
4976 | ||
a4eaf7f1 | 4977 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 4978 | { |
b28ab83c | 4979 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 4980 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
4981 | struct hlist_head *head; |
4982 | ||
6c7e550f | 4983 | if (is_sampling_event(event)) { |
7b4b6658 | 4984 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 4985 | perf_swevent_set_period(event); |
7b4b6658 | 4986 | } |
76e1d904 | 4987 | |
a4eaf7f1 PZ |
4988 | hwc->state = !(flags & PERF_EF_START); |
4989 | ||
b28ab83c | 4990 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
4991 | if (WARN_ON_ONCE(!head)) |
4992 | return -EINVAL; | |
4993 | ||
4994 | hlist_add_head_rcu(&event->hlist_entry, head); | |
4995 | ||
15dbf27c PZ |
4996 | return 0; |
4997 | } | |
4998 | ||
a4eaf7f1 | 4999 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5000 | { |
76e1d904 | 5001 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5002 | } |
5003 | ||
a4eaf7f1 | 5004 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5005 | { |
a4eaf7f1 | 5006 | event->hw.state = 0; |
d6d020e9 | 5007 | } |
aa9c4c0f | 5008 | |
a4eaf7f1 | 5009 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5010 | { |
a4eaf7f1 | 5011 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5012 | } |
5013 | ||
49f135ed FW |
5014 | /* Deref the hlist from the update side */ |
5015 | static inline struct swevent_hlist * | |
b28ab83c | 5016 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5017 | { |
b28ab83c PZ |
5018 | return rcu_dereference_protected(swhash->swevent_hlist, |
5019 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5020 | } |
5021 | ||
b28ab83c | 5022 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5023 | { |
b28ab83c | 5024 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5025 | |
49f135ed | 5026 | if (!hlist) |
76e1d904 FW |
5027 | return; |
5028 | ||
b28ab83c | 5029 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5030 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5031 | } |
5032 | ||
5033 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5034 | { | |
b28ab83c | 5035 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5036 | |
b28ab83c | 5037 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5038 | |
b28ab83c PZ |
5039 | if (!--swhash->hlist_refcount) |
5040 | swevent_hlist_release(swhash); | |
76e1d904 | 5041 | |
b28ab83c | 5042 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5043 | } |
5044 | ||
5045 | static void swevent_hlist_put(struct perf_event *event) | |
5046 | { | |
5047 | int cpu; | |
5048 | ||
5049 | if (event->cpu != -1) { | |
5050 | swevent_hlist_put_cpu(event, event->cpu); | |
5051 | return; | |
5052 | } | |
5053 | ||
5054 | for_each_possible_cpu(cpu) | |
5055 | swevent_hlist_put_cpu(event, cpu); | |
5056 | } | |
5057 | ||
5058 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5059 | { | |
b28ab83c | 5060 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5061 | int err = 0; |
5062 | ||
b28ab83c | 5063 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5064 | |
b28ab83c | 5065 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5066 | struct swevent_hlist *hlist; |
5067 | ||
5068 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5069 | if (!hlist) { | |
5070 | err = -ENOMEM; | |
5071 | goto exit; | |
5072 | } | |
b28ab83c | 5073 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5074 | } |
b28ab83c | 5075 | swhash->hlist_refcount++; |
9ed6060d | 5076 | exit: |
b28ab83c | 5077 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5078 | |
5079 | return err; | |
5080 | } | |
5081 | ||
5082 | static int swevent_hlist_get(struct perf_event *event) | |
5083 | { | |
5084 | int err; | |
5085 | int cpu, failed_cpu; | |
5086 | ||
5087 | if (event->cpu != -1) | |
5088 | return swevent_hlist_get_cpu(event, event->cpu); | |
5089 | ||
5090 | get_online_cpus(); | |
5091 | for_each_possible_cpu(cpu) { | |
5092 | err = swevent_hlist_get_cpu(event, cpu); | |
5093 | if (err) { | |
5094 | failed_cpu = cpu; | |
5095 | goto fail; | |
5096 | } | |
5097 | } | |
5098 | put_online_cpus(); | |
5099 | ||
5100 | return 0; | |
9ed6060d | 5101 | fail: |
76e1d904 FW |
5102 | for_each_possible_cpu(cpu) { |
5103 | if (cpu == failed_cpu) | |
5104 | break; | |
5105 | swevent_hlist_put_cpu(event, cpu); | |
5106 | } | |
5107 | ||
5108 | put_online_cpus(); | |
5109 | return err; | |
5110 | } | |
5111 | ||
c5905afb | 5112 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5113 | |
b0a873eb PZ |
5114 | static void sw_perf_event_destroy(struct perf_event *event) |
5115 | { | |
5116 | u64 event_id = event->attr.config; | |
95476b64 | 5117 | |
b0a873eb PZ |
5118 | WARN_ON(event->parent); |
5119 | ||
c5905afb | 5120 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5121 | swevent_hlist_put(event); |
5122 | } | |
5123 | ||
5124 | static int perf_swevent_init(struct perf_event *event) | |
5125 | { | |
5126 | int event_id = event->attr.config; | |
5127 | ||
5128 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5129 | return -ENOENT; | |
5130 | ||
2481c5fa SE |
5131 | /* |
5132 | * no branch sampling for software events | |
5133 | */ | |
5134 | if (has_branch_stack(event)) | |
5135 | return -EOPNOTSUPP; | |
5136 | ||
b0a873eb PZ |
5137 | switch (event_id) { |
5138 | case PERF_COUNT_SW_CPU_CLOCK: | |
5139 | case PERF_COUNT_SW_TASK_CLOCK: | |
5140 | return -ENOENT; | |
5141 | ||
5142 | default: | |
5143 | break; | |
5144 | } | |
5145 | ||
ce677831 | 5146 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5147 | return -ENOENT; |
5148 | ||
5149 | if (!event->parent) { | |
5150 | int err; | |
5151 | ||
5152 | err = swevent_hlist_get(event); | |
5153 | if (err) | |
5154 | return err; | |
5155 | ||
c5905afb | 5156 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5157 | event->destroy = sw_perf_event_destroy; |
5158 | } | |
5159 | ||
5160 | return 0; | |
5161 | } | |
5162 | ||
35edc2a5 PZ |
5163 | static int perf_swevent_event_idx(struct perf_event *event) |
5164 | { | |
5165 | return 0; | |
5166 | } | |
5167 | ||
b0a873eb | 5168 | static struct pmu perf_swevent = { |
89a1e187 | 5169 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5170 | |
b0a873eb | 5171 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5172 | .add = perf_swevent_add, |
5173 | .del = perf_swevent_del, | |
5174 | .start = perf_swevent_start, | |
5175 | .stop = perf_swevent_stop, | |
1c024eca | 5176 | .read = perf_swevent_read, |
35edc2a5 PZ |
5177 | |
5178 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5179 | }; |
5180 | ||
b0a873eb PZ |
5181 | #ifdef CONFIG_EVENT_TRACING |
5182 | ||
1c024eca PZ |
5183 | static int perf_tp_filter_match(struct perf_event *event, |
5184 | struct perf_sample_data *data) | |
5185 | { | |
5186 | void *record = data->raw->data; | |
5187 | ||
5188 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5189 | return 1; | |
5190 | return 0; | |
5191 | } | |
5192 | ||
5193 | static int perf_tp_event_match(struct perf_event *event, | |
5194 | struct perf_sample_data *data, | |
5195 | struct pt_regs *regs) | |
5196 | { | |
a0f7d0f7 FW |
5197 | if (event->hw.state & PERF_HES_STOPPED) |
5198 | return 0; | |
580d607c PZ |
5199 | /* |
5200 | * All tracepoints are from kernel-space. | |
5201 | */ | |
5202 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5203 | return 0; |
5204 | ||
5205 | if (!perf_tp_filter_match(event, data)) | |
5206 | return 0; | |
5207 | ||
5208 | return 1; | |
5209 | } | |
5210 | ||
5211 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5212 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5213 | struct task_struct *task) | |
95476b64 FW |
5214 | { |
5215 | struct perf_sample_data data; | |
1c024eca PZ |
5216 | struct perf_event *event; |
5217 | struct hlist_node *node; | |
5218 | ||
95476b64 FW |
5219 | struct perf_raw_record raw = { |
5220 | .size = entry_size, | |
5221 | .data = record, | |
5222 | }; | |
5223 | ||
fd0d000b | 5224 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5225 | data.raw = &raw; |
5226 | ||
1c024eca PZ |
5227 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { |
5228 | if (perf_tp_event_match(event, &data, regs)) | |
a8b0ca17 | 5229 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5230 | } |
ecc55f84 | 5231 | |
e6dab5ff AV |
5232 | /* |
5233 | * If we got specified a target task, also iterate its context and | |
5234 | * deliver this event there too. | |
5235 | */ | |
5236 | if (task && task != current) { | |
5237 | struct perf_event_context *ctx; | |
5238 | struct trace_entry *entry = record; | |
5239 | ||
5240 | rcu_read_lock(); | |
5241 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5242 | if (!ctx) | |
5243 | goto unlock; | |
5244 | ||
5245 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5246 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5247 | continue; | |
5248 | if (event->attr.config != entry->type) | |
5249 | continue; | |
5250 | if (perf_tp_event_match(event, &data, regs)) | |
5251 | perf_swevent_event(event, count, &data, regs); | |
5252 | } | |
5253 | unlock: | |
5254 | rcu_read_unlock(); | |
5255 | } | |
5256 | ||
ecc55f84 | 5257 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5258 | } |
5259 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5260 | ||
cdd6c482 | 5261 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5262 | { |
1c024eca | 5263 | perf_trace_destroy(event); |
e077df4f PZ |
5264 | } |
5265 | ||
b0a873eb | 5266 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5267 | { |
76e1d904 FW |
5268 | int err; |
5269 | ||
b0a873eb PZ |
5270 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5271 | return -ENOENT; | |
5272 | ||
2481c5fa SE |
5273 | /* |
5274 | * no branch sampling for tracepoint events | |
5275 | */ | |
5276 | if (has_branch_stack(event)) | |
5277 | return -EOPNOTSUPP; | |
5278 | ||
1c024eca PZ |
5279 | err = perf_trace_init(event); |
5280 | if (err) | |
b0a873eb | 5281 | return err; |
e077df4f | 5282 | |
cdd6c482 | 5283 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5284 | |
b0a873eb PZ |
5285 | return 0; |
5286 | } | |
5287 | ||
5288 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5289 | .task_ctx_nr = perf_sw_context, |
5290 | ||
b0a873eb | 5291 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5292 | .add = perf_trace_add, |
5293 | .del = perf_trace_del, | |
5294 | .start = perf_swevent_start, | |
5295 | .stop = perf_swevent_stop, | |
b0a873eb | 5296 | .read = perf_swevent_read, |
35edc2a5 PZ |
5297 | |
5298 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5299 | }; |
5300 | ||
5301 | static inline void perf_tp_register(void) | |
5302 | { | |
2e80a82a | 5303 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5304 | } |
6fb2915d LZ |
5305 | |
5306 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5307 | { | |
5308 | char *filter_str; | |
5309 | int ret; | |
5310 | ||
5311 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5312 | return -EINVAL; | |
5313 | ||
5314 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5315 | if (IS_ERR(filter_str)) | |
5316 | return PTR_ERR(filter_str); | |
5317 | ||
5318 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5319 | ||
5320 | kfree(filter_str); | |
5321 | return ret; | |
5322 | } | |
5323 | ||
5324 | static void perf_event_free_filter(struct perf_event *event) | |
5325 | { | |
5326 | ftrace_profile_free_filter(event); | |
5327 | } | |
5328 | ||
e077df4f | 5329 | #else |
6fb2915d | 5330 | |
b0a873eb | 5331 | static inline void perf_tp_register(void) |
e077df4f | 5332 | { |
e077df4f | 5333 | } |
6fb2915d LZ |
5334 | |
5335 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5336 | { | |
5337 | return -ENOENT; | |
5338 | } | |
5339 | ||
5340 | static void perf_event_free_filter(struct perf_event *event) | |
5341 | { | |
5342 | } | |
5343 | ||
07b139c8 | 5344 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5345 | |
24f1e32c | 5346 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5347 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5348 | { |
f5ffe02e FW |
5349 | struct perf_sample_data sample; |
5350 | struct pt_regs *regs = data; | |
5351 | ||
fd0d000b | 5352 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 5353 | |
a4eaf7f1 | 5354 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5355 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5356 | } |
5357 | #endif | |
5358 | ||
b0a873eb PZ |
5359 | /* |
5360 | * hrtimer based swevent callback | |
5361 | */ | |
f29ac756 | 5362 | |
b0a873eb | 5363 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5364 | { |
b0a873eb PZ |
5365 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5366 | struct perf_sample_data data; | |
5367 | struct pt_regs *regs; | |
5368 | struct perf_event *event; | |
5369 | u64 period; | |
f29ac756 | 5370 | |
b0a873eb | 5371 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5372 | |
5373 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5374 | return HRTIMER_NORESTART; | |
5375 | ||
b0a873eb | 5376 | event->pmu->read(event); |
f344011c | 5377 | |
fd0d000b | 5378 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
5379 | regs = get_irq_regs(); |
5380 | ||
5381 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5382 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 5383 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
5384 | ret = HRTIMER_NORESTART; |
5385 | } | |
24f1e32c | 5386 | |
b0a873eb PZ |
5387 | period = max_t(u64, 10000, event->hw.sample_period); |
5388 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5389 | |
b0a873eb | 5390 | return ret; |
f29ac756 PZ |
5391 | } |
5392 | ||
b0a873eb | 5393 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5394 | { |
b0a873eb | 5395 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5396 | s64 period; |
5397 | ||
5398 | if (!is_sampling_event(event)) | |
5399 | return; | |
f5ffe02e | 5400 | |
5d508e82 FBH |
5401 | period = local64_read(&hwc->period_left); |
5402 | if (period) { | |
5403 | if (period < 0) | |
5404 | period = 10000; | |
fa407f35 | 5405 | |
5d508e82 FBH |
5406 | local64_set(&hwc->period_left, 0); |
5407 | } else { | |
5408 | period = max_t(u64, 10000, hwc->sample_period); | |
5409 | } | |
5410 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 5411 | ns_to_ktime(period), 0, |
b5ab4cd5 | 5412 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 5413 | } |
b0a873eb PZ |
5414 | |
5415 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 5416 | { |
b0a873eb PZ |
5417 | struct hw_perf_event *hwc = &event->hw; |
5418 | ||
6c7e550f | 5419 | if (is_sampling_event(event)) { |
b0a873eb | 5420 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 5421 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
5422 | |
5423 | hrtimer_cancel(&hwc->hrtimer); | |
5424 | } | |
24f1e32c FW |
5425 | } |
5426 | ||
ba3dd36c PZ |
5427 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5428 | { | |
5429 | struct hw_perf_event *hwc = &event->hw; | |
5430 | ||
5431 | if (!is_sampling_event(event)) | |
5432 | return; | |
5433 | ||
5434 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5435 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
5436 | ||
5437 | /* | |
5438 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
5439 | * mapping and avoid the whole period adjust feedback stuff. | |
5440 | */ | |
5441 | if (event->attr.freq) { | |
5442 | long freq = event->attr.sample_freq; | |
5443 | ||
5444 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
5445 | hwc->sample_period = event->attr.sample_period; | |
5446 | local64_set(&hwc->period_left, hwc->sample_period); | |
5447 | event->attr.freq = 0; | |
5448 | } | |
5449 | } | |
5450 | ||
b0a873eb PZ |
5451 | /* |
5452 | * Software event: cpu wall time clock | |
5453 | */ | |
5454 | ||
5455 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 5456 | { |
b0a873eb PZ |
5457 | s64 prev; |
5458 | u64 now; | |
5459 | ||
a4eaf7f1 | 5460 | now = local_clock(); |
b0a873eb PZ |
5461 | prev = local64_xchg(&event->hw.prev_count, now); |
5462 | local64_add(now - prev, &event->count); | |
24f1e32c | 5463 | } |
24f1e32c | 5464 | |
a4eaf7f1 | 5465 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5466 | { |
a4eaf7f1 | 5467 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 5468 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5469 | } |
5470 | ||
a4eaf7f1 | 5471 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 5472 | { |
b0a873eb PZ |
5473 | perf_swevent_cancel_hrtimer(event); |
5474 | cpu_clock_event_update(event); | |
5475 | } | |
f29ac756 | 5476 | |
a4eaf7f1 PZ |
5477 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
5478 | { | |
5479 | if (flags & PERF_EF_START) | |
5480 | cpu_clock_event_start(event, flags); | |
5481 | ||
5482 | return 0; | |
5483 | } | |
5484 | ||
5485 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
5486 | { | |
5487 | cpu_clock_event_stop(event, flags); | |
5488 | } | |
5489 | ||
b0a873eb PZ |
5490 | static void cpu_clock_event_read(struct perf_event *event) |
5491 | { | |
5492 | cpu_clock_event_update(event); | |
5493 | } | |
f344011c | 5494 | |
b0a873eb PZ |
5495 | static int cpu_clock_event_init(struct perf_event *event) |
5496 | { | |
5497 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5498 | return -ENOENT; | |
5499 | ||
5500 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
5501 | return -ENOENT; | |
5502 | ||
2481c5fa SE |
5503 | /* |
5504 | * no branch sampling for software events | |
5505 | */ | |
5506 | if (has_branch_stack(event)) | |
5507 | return -EOPNOTSUPP; | |
5508 | ||
ba3dd36c PZ |
5509 | perf_swevent_init_hrtimer(event); |
5510 | ||
b0a873eb | 5511 | return 0; |
f29ac756 PZ |
5512 | } |
5513 | ||
b0a873eb | 5514 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
5515 | .task_ctx_nr = perf_sw_context, |
5516 | ||
b0a873eb | 5517 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
5518 | .add = cpu_clock_event_add, |
5519 | .del = cpu_clock_event_del, | |
5520 | .start = cpu_clock_event_start, | |
5521 | .stop = cpu_clock_event_stop, | |
b0a873eb | 5522 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
5523 | |
5524 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5525 | }; |
5526 | ||
5527 | /* | |
5528 | * Software event: task time clock | |
5529 | */ | |
5530 | ||
5531 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 5532 | { |
b0a873eb PZ |
5533 | u64 prev; |
5534 | s64 delta; | |
5c92d124 | 5535 | |
b0a873eb PZ |
5536 | prev = local64_xchg(&event->hw.prev_count, now); |
5537 | delta = now - prev; | |
5538 | local64_add(delta, &event->count); | |
5539 | } | |
5c92d124 | 5540 | |
a4eaf7f1 | 5541 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5542 | { |
a4eaf7f1 | 5543 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 5544 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5545 | } |
5546 | ||
a4eaf7f1 | 5547 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
5548 | { |
5549 | perf_swevent_cancel_hrtimer(event); | |
5550 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
5551 | } |
5552 | ||
5553 | static int task_clock_event_add(struct perf_event *event, int flags) | |
5554 | { | |
5555 | if (flags & PERF_EF_START) | |
5556 | task_clock_event_start(event, flags); | |
b0a873eb | 5557 | |
a4eaf7f1 PZ |
5558 | return 0; |
5559 | } | |
5560 | ||
5561 | static void task_clock_event_del(struct perf_event *event, int flags) | |
5562 | { | |
5563 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
5564 | } |
5565 | ||
5566 | static void task_clock_event_read(struct perf_event *event) | |
5567 | { | |
768a06e2 PZ |
5568 | u64 now = perf_clock(); |
5569 | u64 delta = now - event->ctx->timestamp; | |
5570 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
5571 | |
5572 | task_clock_event_update(event, time); | |
5573 | } | |
5574 | ||
5575 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 5576 | { |
b0a873eb PZ |
5577 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5578 | return -ENOENT; | |
5579 | ||
5580 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
5581 | return -ENOENT; | |
5582 | ||
2481c5fa SE |
5583 | /* |
5584 | * no branch sampling for software events | |
5585 | */ | |
5586 | if (has_branch_stack(event)) | |
5587 | return -EOPNOTSUPP; | |
5588 | ||
ba3dd36c PZ |
5589 | perf_swevent_init_hrtimer(event); |
5590 | ||
b0a873eb | 5591 | return 0; |
6fb2915d LZ |
5592 | } |
5593 | ||
b0a873eb | 5594 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
5595 | .task_ctx_nr = perf_sw_context, |
5596 | ||
b0a873eb | 5597 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
5598 | .add = task_clock_event_add, |
5599 | .del = task_clock_event_del, | |
5600 | .start = task_clock_event_start, | |
5601 | .stop = task_clock_event_stop, | |
b0a873eb | 5602 | .read = task_clock_event_read, |
35edc2a5 PZ |
5603 | |
5604 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 5605 | }; |
6fb2915d | 5606 | |
ad5133b7 | 5607 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 5608 | { |
e077df4f | 5609 | } |
6fb2915d | 5610 | |
ad5133b7 | 5611 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 5612 | { |
ad5133b7 | 5613 | return 0; |
6fb2915d LZ |
5614 | } |
5615 | ||
ad5133b7 | 5616 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 5617 | { |
ad5133b7 | 5618 | perf_pmu_disable(pmu); |
6fb2915d LZ |
5619 | } |
5620 | ||
ad5133b7 PZ |
5621 | static int perf_pmu_commit_txn(struct pmu *pmu) |
5622 | { | |
5623 | perf_pmu_enable(pmu); | |
5624 | return 0; | |
5625 | } | |
e077df4f | 5626 | |
ad5133b7 | 5627 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 5628 | { |
ad5133b7 | 5629 | perf_pmu_enable(pmu); |
24f1e32c FW |
5630 | } |
5631 | ||
35edc2a5 PZ |
5632 | static int perf_event_idx_default(struct perf_event *event) |
5633 | { | |
5634 | return event->hw.idx + 1; | |
5635 | } | |
5636 | ||
8dc85d54 PZ |
5637 | /* |
5638 | * Ensures all contexts with the same task_ctx_nr have the same | |
5639 | * pmu_cpu_context too. | |
5640 | */ | |
5641 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 5642 | { |
8dc85d54 | 5643 | struct pmu *pmu; |
b326e956 | 5644 | |
8dc85d54 PZ |
5645 | if (ctxn < 0) |
5646 | return NULL; | |
24f1e32c | 5647 | |
8dc85d54 PZ |
5648 | list_for_each_entry(pmu, &pmus, entry) { |
5649 | if (pmu->task_ctx_nr == ctxn) | |
5650 | return pmu->pmu_cpu_context; | |
5651 | } | |
24f1e32c | 5652 | |
8dc85d54 | 5653 | return NULL; |
24f1e32c FW |
5654 | } |
5655 | ||
51676957 | 5656 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 5657 | { |
51676957 PZ |
5658 | int cpu; |
5659 | ||
5660 | for_each_possible_cpu(cpu) { | |
5661 | struct perf_cpu_context *cpuctx; | |
5662 | ||
5663 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
5664 | ||
5665 | if (cpuctx->active_pmu == old_pmu) | |
5666 | cpuctx->active_pmu = pmu; | |
5667 | } | |
5668 | } | |
5669 | ||
5670 | static void free_pmu_context(struct pmu *pmu) | |
5671 | { | |
5672 | struct pmu *i; | |
f5ffe02e | 5673 | |
8dc85d54 | 5674 | mutex_lock(&pmus_lock); |
0475f9ea | 5675 | /* |
8dc85d54 | 5676 | * Like a real lame refcount. |
0475f9ea | 5677 | */ |
51676957 PZ |
5678 | list_for_each_entry(i, &pmus, entry) { |
5679 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
5680 | update_pmu_context(i, pmu); | |
8dc85d54 | 5681 | goto out; |
51676957 | 5682 | } |
8dc85d54 | 5683 | } |
d6d020e9 | 5684 | |
51676957 | 5685 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
5686 | out: |
5687 | mutex_unlock(&pmus_lock); | |
24f1e32c | 5688 | } |
2e80a82a | 5689 | static struct idr pmu_idr; |
d6d020e9 | 5690 | |
abe43400 PZ |
5691 | static ssize_t |
5692 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
5693 | { | |
5694 | struct pmu *pmu = dev_get_drvdata(dev); | |
5695 | ||
5696 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
5697 | } | |
5698 | ||
5699 | static struct device_attribute pmu_dev_attrs[] = { | |
5700 | __ATTR_RO(type), | |
5701 | __ATTR_NULL, | |
5702 | }; | |
5703 | ||
5704 | static int pmu_bus_running; | |
5705 | static struct bus_type pmu_bus = { | |
5706 | .name = "event_source", | |
5707 | .dev_attrs = pmu_dev_attrs, | |
5708 | }; | |
5709 | ||
5710 | static void pmu_dev_release(struct device *dev) | |
5711 | { | |
5712 | kfree(dev); | |
5713 | } | |
5714 | ||
5715 | static int pmu_dev_alloc(struct pmu *pmu) | |
5716 | { | |
5717 | int ret = -ENOMEM; | |
5718 | ||
5719 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
5720 | if (!pmu->dev) | |
5721 | goto out; | |
5722 | ||
0c9d42ed | 5723 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
5724 | device_initialize(pmu->dev); |
5725 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
5726 | if (ret) | |
5727 | goto free_dev; | |
5728 | ||
5729 | dev_set_drvdata(pmu->dev, pmu); | |
5730 | pmu->dev->bus = &pmu_bus; | |
5731 | pmu->dev->release = pmu_dev_release; | |
5732 | ret = device_add(pmu->dev); | |
5733 | if (ret) | |
5734 | goto free_dev; | |
5735 | ||
5736 | out: | |
5737 | return ret; | |
5738 | ||
5739 | free_dev: | |
5740 | put_device(pmu->dev); | |
5741 | goto out; | |
5742 | } | |
5743 | ||
547e9fd7 | 5744 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 5745 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 5746 | |
2e80a82a | 5747 | int perf_pmu_register(struct pmu *pmu, char *name, int type) |
24f1e32c | 5748 | { |
108b02cf | 5749 | int cpu, ret; |
24f1e32c | 5750 | |
b0a873eb | 5751 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
5752 | ret = -ENOMEM; |
5753 | pmu->pmu_disable_count = alloc_percpu(int); | |
5754 | if (!pmu->pmu_disable_count) | |
5755 | goto unlock; | |
f29ac756 | 5756 | |
2e80a82a PZ |
5757 | pmu->type = -1; |
5758 | if (!name) | |
5759 | goto skip_type; | |
5760 | pmu->name = name; | |
5761 | ||
5762 | if (type < 0) { | |
5763 | int err = idr_pre_get(&pmu_idr, GFP_KERNEL); | |
5764 | if (!err) | |
5765 | goto free_pdc; | |
5766 | ||
5767 | err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type); | |
5768 | if (err) { | |
5769 | ret = err; | |
5770 | goto free_pdc; | |
5771 | } | |
5772 | } | |
5773 | pmu->type = type; | |
5774 | ||
abe43400 PZ |
5775 | if (pmu_bus_running) { |
5776 | ret = pmu_dev_alloc(pmu); | |
5777 | if (ret) | |
5778 | goto free_idr; | |
5779 | } | |
5780 | ||
2e80a82a | 5781 | skip_type: |
8dc85d54 PZ |
5782 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
5783 | if (pmu->pmu_cpu_context) | |
5784 | goto got_cpu_context; | |
f29ac756 | 5785 | |
108b02cf PZ |
5786 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
5787 | if (!pmu->pmu_cpu_context) | |
abe43400 | 5788 | goto free_dev; |
f344011c | 5789 | |
108b02cf PZ |
5790 | for_each_possible_cpu(cpu) { |
5791 | struct perf_cpu_context *cpuctx; | |
5792 | ||
5793 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 5794 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 5795 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 5796 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 5797 | cpuctx->ctx.type = cpu_context; |
108b02cf | 5798 | cpuctx->ctx.pmu = pmu; |
e9d2b064 PZ |
5799 | cpuctx->jiffies_interval = 1; |
5800 | INIT_LIST_HEAD(&cpuctx->rotation_list); | |
51676957 | 5801 | cpuctx->active_pmu = pmu; |
108b02cf | 5802 | } |
76e1d904 | 5803 | |
8dc85d54 | 5804 | got_cpu_context: |
ad5133b7 PZ |
5805 | if (!pmu->start_txn) { |
5806 | if (pmu->pmu_enable) { | |
5807 | /* | |
5808 | * If we have pmu_enable/pmu_disable calls, install | |
5809 | * transaction stubs that use that to try and batch | |
5810 | * hardware accesses. | |
5811 | */ | |
5812 | pmu->start_txn = perf_pmu_start_txn; | |
5813 | pmu->commit_txn = perf_pmu_commit_txn; | |
5814 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
5815 | } else { | |
5816 | pmu->start_txn = perf_pmu_nop_void; | |
5817 | pmu->commit_txn = perf_pmu_nop_int; | |
5818 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 5819 | } |
5c92d124 | 5820 | } |
15dbf27c | 5821 | |
ad5133b7 PZ |
5822 | if (!pmu->pmu_enable) { |
5823 | pmu->pmu_enable = perf_pmu_nop_void; | |
5824 | pmu->pmu_disable = perf_pmu_nop_void; | |
5825 | } | |
5826 | ||
35edc2a5 PZ |
5827 | if (!pmu->event_idx) |
5828 | pmu->event_idx = perf_event_idx_default; | |
5829 | ||
b0a873eb | 5830 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
5831 | ret = 0; |
5832 | unlock: | |
b0a873eb PZ |
5833 | mutex_unlock(&pmus_lock); |
5834 | ||
33696fc0 | 5835 | return ret; |
108b02cf | 5836 | |
abe43400 PZ |
5837 | free_dev: |
5838 | device_del(pmu->dev); | |
5839 | put_device(pmu->dev); | |
5840 | ||
2e80a82a PZ |
5841 | free_idr: |
5842 | if (pmu->type >= PERF_TYPE_MAX) | |
5843 | idr_remove(&pmu_idr, pmu->type); | |
5844 | ||
108b02cf PZ |
5845 | free_pdc: |
5846 | free_percpu(pmu->pmu_disable_count); | |
5847 | goto unlock; | |
f29ac756 PZ |
5848 | } |
5849 | ||
b0a873eb | 5850 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 5851 | { |
b0a873eb PZ |
5852 | mutex_lock(&pmus_lock); |
5853 | list_del_rcu(&pmu->entry); | |
5854 | mutex_unlock(&pmus_lock); | |
5c92d124 | 5855 | |
0475f9ea | 5856 | /* |
cde8e884 PZ |
5857 | * We dereference the pmu list under both SRCU and regular RCU, so |
5858 | * synchronize against both of those. | |
0475f9ea | 5859 | */ |
b0a873eb | 5860 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 5861 | synchronize_rcu(); |
d6d020e9 | 5862 | |
33696fc0 | 5863 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
5864 | if (pmu->type >= PERF_TYPE_MAX) |
5865 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
5866 | device_del(pmu->dev); |
5867 | put_device(pmu->dev); | |
51676957 | 5868 | free_pmu_context(pmu); |
b0a873eb | 5869 | } |
d6d020e9 | 5870 | |
b0a873eb PZ |
5871 | struct pmu *perf_init_event(struct perf_event *event) |
5872 | { | |
5873 | struct pmu *pmu = NULL; | |
5874 | int idx; | |
940c5b29 | 5875 | int ret; |
b0a873eb PZ |
5876 | |
5877 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
5878 | |
5879 | rcu_read_lock(); | |
5880 | pmu = idr_find(&pmu_idr, event->attr.type); | |
5881 | rcu_read_unlock(); | |
940c5b29 | 5882 | if (pmu) { |
7e5b2a01 | 5883 | event->pmu = pmu; |
940c5b29 LM |
5884 | ret = pmu->event_init(event); |
5885 | if (ret) | |
5886 | pmu = ERR_PTR(ret); | |
2e80a82a | 5887 | goto unlock; |
940c5b29 | 5888 | } |
2e80a82a | 5889 | |
b0a873eb | 5890 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 5891 | event->pmu = pmu; |
940c5b29 | 5892 | ret = pmu->event_init(event); |
b0a873eb | 5893 | if (!ret) |
e5f4d339 | 5894 | goto unlock; |
76e1d904 | 5895 | |
b0a873eb PZ |
5896 | if (ret != -ENOENT) { |
5897 | pmu = ERR_PTR(ret); | |
e5f4d339 | 5898 | goto unlock; |
f344011c | 5899 | } |
5c92d124 | 5900 | } |
e5f4d339 PZ |
5901 | pmu = ERR_PTR(-ENOENT); |
5902 | unlock: | |
b0a873eb | 5903 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 5904 | |
4aeb0b42 | 5905 | return pmu; |
5c92d124 IM |
5906 | } |
5907 | ||
0793a61d | 5908 | /* |
cdd6c482 | 5909 | * Allocate and initialize a event structure |
0793a61d | 5910 | */ |
cdd6c482 | 5911 | static struct perf_event * |
c3f00c70 | 5912 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
5913 | struct task_struct *task, |
5914 | struct perf_event *group_leader, | |
5915 | struct perf_event *parent_event, | |
4dc0da86 AK |
5916 | perf_overflow_handler_t overflow_handler, |
5917 | void *context) | |
0793a61d | 5918 | { |
51b0fe39 | 5919 | struct pmu *pmu; |
cdd6c482 IM |
5920 | struct perf_event *event; |
5921 | struct hw_perf_event *hwc; | |
d5d2bc0d | 5922 | long err; |
0793a61d | 5923 | |
66832eb4 ON |
5924 | if ((unsigned)cpu >= nr_cpu_ids) { |
5925 | if (!task || cpu != -1) | |
5926 | return ERR_PTR(-EINVAL); | |
5927 | } | |
5928 | ||
c3f00c70 | 5929 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 5930 | if (!event) |
d5d2bc0d | 5931 | return ERR_PTR(-ENOMEM); |
0793a61d | 5932 | |
04289bb9 | 5933 | /* |
cdd6c482 | 5934 | * Single events are their own group leaders, with an |
04289bb9 IM |
5935 | * empty sibling list: |
5936 | */ | |
5937 | if (!group_leader) | |
cdd6c482 | 5938 | group_leader = event; |
04289bb9 | 5939 | |
cdd6c482 IM |
5940 | mutex_init(&event->child_mutex); |
5941 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 5942 | |
cdd6c482 IM |
5943 | INIT_LIST_HEAD(&event->group_entry); |
5944 | INIT_LIST_HEAD(&event->event_entry); | |
5945 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
5946 | INIT_LIST_HEAD(&event->rb_entry); |
5947 | ||
cdd6c482 | 5948 | init_waitqueue_head(&event->waitq); |
e360adbe | 5949 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 5950 | |
cdd6c482 | 5951 | mutex_init(&event->mmap_mutex); |
7b732a75 | 5952 | |
cdd6c482 IM |
5953 | event->cpu = cpu; |
5954 | event->attr = *attr; | |
5955 | event->group_leader = group_leader; | |
5956 | event->pmu = NULL; | |
cdd6c482 | 5957 | event->oncpu = -1; |
a96bbc16 | 5958 | |
cdd6c482 | 5959 | event->parent = parent_event; |
b84fbc9f | 5960 | |
cdd6c482 IM |
5961 | event->ns = get_pid_ns(current->nsproxy->pid_ns); |
5962 | event->id = atomic64_inc_return(&perf_event_id); | |
a96bbc16 | 5963 | |
cdd6c482 | 5964 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 5965 | |
d580ff86 PZ |
5966 | if (task) { |
5967 | event->attach_state = PERF_ATTACH_TASK; | |
5968 | #ifdef CONFIG_HAVE_HW_BREAKPOINT | |
5969 | /* | |
5970 | * hw_breakpoint is a bit difficult here.. | |
5971 | */ | |
5972 | if (attr->type == PERF_TYPE_BREAKPOINT) | |
5973 | event->hw.bp_target = task; | |
5974 | #endif | |
5975 | } | |
5976 | ||
4dc0da86 | 5977 | if (!overflow_handler && parent_event) { |
b326e956 | 5978 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
5979 | context = parent_event->overflow_handler_context; |
5980 | } | |
66832eb4 | 5981 | |
b326e956 | 5982 | event->overflow_handler = overflow_handler; |
4dc0da86 | 5983 | event->overflow_handler_context = context; |
97eaf530 | 5984 | |
0d48696f | 5985 | if (attr->disabled) |
cdd6c482 | 5986 | event->state = PERF_EVENT_STATE_OFF; |
a86ed508 | 5987 | |
4aeb0b42 | 5988 | pmu = NULL; |
b8e83514 | 5989 | |
cdd6c482 | 5990 | hwc = &event->hw; |
bd2b5b12 | 5991 | hwc->sample_period = attr->sample_period; |
0d48696f | 5992 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 5993 | hwc->sample_period = 1; |
eced1dfc | 5994 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 5995 | |
e7850595 | 5996 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 5997 | |
2023b359 | 5998 | /* |
cdd6c482 | 5999 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6000 | */ |
3dab77fb | 6001 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
2023b359 PZ |
6002 | goto done; |
6003 | ||
b0a873eb | 6004 | pmu = perf_init_event(event); |
974802ea | 6005 | |
d5d2bc0d PM |
6006 | done: |
6007 | err = 0; | |
4aeb0b42 | 6008 | if (!pmu) |
d5d2bc0d | 6009 | err = -EINVAL; |
4aeb0b42 RR |
6010 | else if (IS_ERR(pmu)) |
6011 | err = PTR_ERR(pmu); | |
5c92d124 | 6012 | |
d5d2bc0d | 6013 | if (err) { |
cdd6c482 IM |
6014 | if (event->ns) |
6015 | put_pid_ns(event->ns); | |
6016 | kfree(event); | |
d5d2bc0d | 6017 | return ERR_PTR(err); |
621a01ea | 6018 | } |
d5d2bc0d | 6019 | |
cdd6c482 | 6020 | if (!event->parent) { |
82cd6def | 6021 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 6022 | static_key_slow_inc(&perf_sched_events.key); |
3af9e859 | 6023 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
6024 | atomic_inc(&nr_mmap_events); |
6025 | if (event->attr.comm) | |
6026 | atomic_inc(&nr_comm_events); | |
6027 | if (event->attr.task) | |
6028 | atomic_inc(&nr_task_events); | |
927c7a9e FW |
6029 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6030 | err = get_callchain_buffers(); | |
6031 | if (err) { | |
6032 | free_event(event); | |
6033 | return ERR_PTR(err); | |
6034 | } | |
6035 | } | |
d010b332 SE |
6036 | if (has_branch_stack(event)) { |
6037 | static_key_slow_inc(&perf_sched_events.key); | |
6038 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6039 | atomic_inc(&per_cpu(perf_branch_stack_events, | |
6040 | event->cpu)); | |
6041 | } | |
f344011c | 6042 | } |
9ee318a7 | 6043 | |
cdd6c482 | 6044 | return event; |
0793a61d TG |
6045 | } |
6046 | ||
cdd6c482 IM |
6047 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6048 | struct perf_event_attr *attr) | |
974802ea | 6049 | { |
974802ea | 6050 | u32 size; |
cdf8073d | 6051 | int ret; |
974802ea PZ |
6052 | |
6053 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6054 | return -EFAULT; | |
6055 | ||
6056 | /* | |
6057 | * zero the full structure, so that a short copy will be nice. | |
6058 | */ | |
6059 | memset(attr, 0, sizeof(*attr)); | |
6060 | ||
6061 | ret = get_user(size, &uattr->size); | |
6062 | if (ret) | |
6063 | return ret; | |
6064 | ||
6065 | if (size > PAGE_SIZE) /* silly large */ | |
6066 | goto err_size; | |
6067 | ||
6068 | if (!size) /* abi compat */ | |
6069 | size = PERF_ATTR_SIZE_VER0; | |
6070 | ||
6071 | if (size < PERF_ATTR_SIZE_VER0) | |
6072 | goto err_size; | |
6073 | ||
6074 | /* | |
6075 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6076 | * ensure all the unknown bits are 0 - i.e. new |
6077 | * user-space does not rely on any kernel feature | |
6078 | * extensions we dont know about yet. | |
974802ea PZ |
6079 | */ |
6080 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6081 | unsigned char __user *addr; |
6082 | unsigned char __user *end; | |
6083 | unsigned char val; | |
974802ea | 6084 | |
cdf8073d IS |
6085 | addr = (void __user *)uattr + sizeof(*attr); |
6086 | end = (void __user *)uattr + size; | |
974802ea | 6087 | |
cdf8073d | 6088 | for (; addr < end; addr++) { |
974802ea PZ |
6089 | ret = get_user(val, addr); |
6090 | if (ret) | |
6091 | return ret; | |
6092 | if (val) | |
6093 | goto err_size; | |
6094 | } | |
b3e62e35 | 6095 | size = sizeof(*attr); |
974802ea PZ |
6096 | } |
6097 | ||
6098 | ret = copy_from_user(attr, uattr, size); | |
6099 | if (ret) | |
6100 | return -EFAULT; | |
6101 | ||
cd757645 | 6102 | if (attr->__reserved_1) |
974802ea PZ |
6103 | return -EINVAL; |
6104 | ||
6105 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6106 | return -EINVAL; | |
6107 | ||
6108 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6109 | return -EINVAL; | |
6110 | ||
bce38cd5 SE |
6111 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6112 | u64 mask = attr->branch_sample_type; | |
6113 | ||
6114 | /* only using defined bits */ | |
6115 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6116 | return -EINVAL; | |
6117 | ||
6118 | /* at least one branch bit must be set */ | |
6119 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6120 | return -EINVAL; | |
6121 | ||
6122 | /* kernel level capture: check permissions */ | |
6123 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
6124 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6125 | return -EACCES; | |
6126 | ||
6127 | /* propagate priv level, when not set for branch */ | |
6128 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6129 | ||
6130 | /* exclude_kernel checked on syscall entry */ | |
6131 | if (!attr->exclude_kernel) | |
6132 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6133 | ||
6134 | if (!attr->exclude_user) | |
6135 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6136 | ||
6137 | if (!attr->exclude_hv) | |
6138 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6139 | /* | |
6140 | * adjust user setting (for HW filter setup) | |
6141 | */ | |
6142 | attr->branch_sample_type = mask; | |
6143 | } | |
6144 | } | |
974802ea PZ |
6145 | out: |
6146 | return ret; | |
6147 | ||
6148 | err_size: | |
6149 | put_user(sizeof(*attr), &uattr->size); | |
6150 | ret = -E2BIG; | |
6151 | goto out; | |
6152 | } | |
6153 | ||
ac9721f3 PZ |
6154 | static int |
6155 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6156 | { |
76369139 | 6157 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6158 | int ret = -EINVAL; |
6159 | ||
ac9721f3 | 6160 | if (!output_event) |
a4be7c27 PZ |
6161 | goto set; |
6162 | ||
ac9721f3 PZ |
6163 | /* don't allow circular references */ |
6164 | if (event == output_event) | |
a4be7c27 PZ |
6165 | goto out; |
6166 | ||
0f139300 PZ |
6167 | /* |
6168 | * Don't allow cross-cpu buffers | |
6169 | */ | |
6170 | if (output_event->cpu != event->cpu) | |
6171 | goto out; | |
6172 | ||
6173 | /* | |
76369139 | 6174 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6175 | */ |
6176 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6177 | goto out; | |
6178 | ||
a4be7c27 | 6179 | set: |
cdd6c482 | 6180 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6181 | /* Can't redirect output if we've got an active mmap() */ |
6182 | if (atomic_read(&event->mmap_count)) | |
6183 | goto unlock; | |
a4be7c27 | 6184 | |
ac9721f3 | 6185 | if (output_event) { |
76369139 FW |
6186 | /* get the rb we want to redirect to */ |
6187 | rb = ring_buffer_get(output_event); | |
6188 | if (!rb) | |
ac9721f3 | 6189 | goto unlock; |
a4be7c27 PZ |
6190 | } |
6191 | ||
76369139 FW |
6192 | old_rb = event->rb; |
6193 | rcu_assign_pointer(event->rb, rb); | |
10c6db11 PZ |
6194 | if (old_rb) |
6195 | ring_buffer_detach(event, old_rb); | |
a4be7c27 | 6196 | ret = 0; |
ac9721f3 PZ |
6197 | unlock: |
6198 | mutex_unlock(&event->mmap_mutex); | |
6199 | ||
76369139 FW |
6200 | if (old_rb) |
6201 | ring_buffer_put(old_rb); | |
a4be7c27 | 6202 | out: |
a4be7c27 PZ |
6203 | return ret; |
6204 | } | |
6205 | ||
0793a61d | 6206 | /** |
cdd6c482 | 6207 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6208 | * |
cdd6c482 | 6209 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6210 | * @pid: target pid |
9f66a381 | 6211 | * @cpu: target cpu |
cdd6c482 | 6212 | * @group_fd: group leader event fd |
0793a61d | 6213 | */ |
cdd6c482 IM |
6214 | SYSCALL_DEFINE5(perf_event_open, |
6215 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 6216 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 6217 | { |
b04243ef PZ |
6218 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6219 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
6220 | struct perf_event_attr attr; |
6221 | struct perf_event_context *ctx; | |
6222 | struct file *event_file = NULL; | |
04289bb9 | 6223 | struct file *group_file = NULL; |
38a81da2 | 6224 | struct task_struct *task = NULL; |
89a1e187 | 6225 | struct pmu *pmu; |
ea635c64 | 6226 | int event_fd; |
b04243ef | 6227 | int move_group = 0; |
04289bb9 | 6228 | int fput_needed = 0; |
dc86cabe | 6229 | int err; |
0793a61d | 6230 | |
2743a5b0 | 6231 | /* for future expandability... */ |
e5d1367f | 6232 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
6233 | return -EINVAL; |
6234 | ||
dc86cabe IM |
6235 | err = perf_copy_attr(attr_uptr, &attr); |
6236 | if (err) | |
6237 | return err; | |
eab656ae | 6238 | |
0764771d PZ |
6239 | if (!attr.exclude_kernel) { |
6240 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6241 | return -EACCES; | |
6242 | } | |
6243 | ||
df58ab24 | 6244 | if (attr.freq) { |
cdd6c482 | 6245 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
6246 | return -EINVAL; |
6247 | } | |
6248 | ||
e5d1367f SE |
6249 | /* |
6250 | * In cgroup mode, the pid argument is used to pass the fd | |
6251 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6252 | * designates the cpu on which to monitor threads from that | |
6253 | * cgroup. | |
6254 | */ | |
6255 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6256 | return -EINVAL; | |
6257 | ||
ea635c64 AV |
6258 | event_fd = get_unused_fd_flags(O_RDWR); |
6259 | if (event_fd < 0) | |
6260 | return event_fd; | |
6261 | ||
ac9721f3 PZ |
6262 | if (group_fd != -1) { |
6263 | group_leader = perf_fget_light(group_fd, &fput_needed); | |
6264 | if (IS_ERR(group_leader)) { | |
6265 | err = PTR_ERR(group_leader); | |
d14b12d7 | 6266 | goto err_fd; |
ac9721f3 PZ |
6267 | } |
6268 | group_file = group_leader->filp; | |
6269 | if (flags & PERF_FLAG_FD_OUTPUT) | |
6270 | output_event = group_leader; | |
6271 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6272 | group_leader = NULL; | |
6273 | } | |
6274 | ||
e5d1367f | 6275 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6276 | task = find_lively_task_by_vpid(pid); |
6277 | if (IS_ERR(task)) { | |
6278 | err = PTR_ERR(task); | |
6279 | goto err_group_fd; | |
6280 | } | |
6281 | } | |
6282 | ||
fbfc623f YZ |
6283 | get_online_cpus(); |
6284 | ||
4dc0da86 AK |
6285 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6286 | NULL, NULL); | |
d14b12d7 SE |
6287 | if (IS_ERR(event)) { |
6288 | err = PTR_ERR(event); | |
c6be5a5c | 6289 | goto err_task; |
d14b12d7 SE |
6290 | } |
6291 | ||
e5d1367f SE |
6292 | if (flags & PERF_FLAG_PID_CGROUP) { |
6293 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
6294 | if (err) | |
6295 | goto err_alloc; | |
08309379 PZ |
6296 | /* |
6297 | * one more event: | |
6298 | * - that has cgroup constraint on event->cpu | |
6299 | * - that may need work on context switch | |
6300 | */ | |
6301 | atomic_inc(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 6302 | static_key_slow_inc(&perf_sched_events.key); |
e5d1367f SE |
6303 | } |
6304 | ||
89a1e187 PZ |
6305 | /* |
6306 | * Special case software events and allow them to be part of | |
6307 | * any hardware group. | |
6308 | */ | |
6309 | pmu = event->pmu; | |
b04243ef PZ |
6310 | |
6311 | if (group_leader && | |
6312 | (is_software_event(event) != is_software_event(group_leader))) { | |
6313 | if (is_software_event(event)) { | |
6314 | /* | |
6315 | * If event and group_leader are not both a software | |
6316 | * event, and event is, then group leader is not. | |
6317 | * | |
6318 | * Allow the addition of software events to !software | |
6319 | * groups, this is safe because software events never | |
6320 | * fail to schedule. | |
6321 | */ | |
6322 | pmu = group_leader->pmu; | |
6323 | } else if (is_software_event(group_leader) && | |
6324 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
6325 | /* | |
6326 | * In case the group is a pure software group, and we | |
6327 | * try to add a hardware event, move the whole group to | |
6328 | * the hardware context. | |
6329 | */ | |
6330 | move_group = 1; | |
6331 | } | |
6332 | } | |
89a1e187 PZ |
6333 | |
6334 | /* | |
6335 | * Get the target context (task or percpu): | |
6336 | */ | |
e2d37cd2 | 6337 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
6338 | if (IS_ERR(ctx)) { |
6339 | err = PTR_ERR(ctx); | |
c6be5a5c | 6340 | goto err_alloc; |
89a1e187 PZ |
6341 | } |
6342 | ||
fd1edb3a PZ |
6343 | if (task) { |
6344 | put_task_struct(task); | |
6345 | task = NULL; | |
6346 | } | |
6347 | ||
ccff286d | 6348 | /* |
cdd6c482 | 6349 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 6350 | */ |
ac9721f3 | 6351 | if (group_leader) { |
dc86cabe | 6352 | err = -EINVAL; |
04289bb9 | 6353 | |
04289bb9 | 6354 | /* |
ccff286d IM |
6355 | * Do not allow a recursive hierarchy (this new sibling |
6356 | * becoming part of another group-sibling): | |
6357 | */ | |
6358 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 6359 | goto err_context; |
ccff286d IM |
6360 | /* |
6361 | * Do not allow to attach to a group in a different | |
6362 | * task or CPU context: | |
04289bb9 | 6363 | */ |
b04243ef PZ |
6364 | if (move_group) { |
6365 | if (group_leader->ctx->type != ctx->type) | |
6366 | goto err_context; | |
6367 | } else { | |
6368 | if (group_leader->ctx != ctx) | |
6369 | goto err_context; | |
6370 | } | |
6371 | ||
3b6f9e5c PM |
6372 | /* |
6373 | * Only a group leader can be exclusive or pinned | |
6374 | */ | |
0d48696f | 6375 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 6376 | goto err_context; |
ac9721f3 PZ |
6377 | } |
6378 | ||
6379 | if (output_event) { | |
6380 | err = perf_event_set_output(event, output_event); | |
6381 | if (err) | |
c3f00c70 | 6382 | goto err_context; |
ac9721f3 | 6383 | } |
0793a61d | 6384 | |
ea635c64 AV |
6385 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
6386 | if (IS_ERR(event_file)) { | |
6387 | err = PTR_ERR(event_file); | |
c3f00c70 | 6388 | goto err_context; |
ea635c64 | 6389 | } |
9b51f66d | 6390 | |
b04243ef PZ |
6391 | if (move_group) { |
6392 | struct perf_event_context *gctx = group_leader->ctx; | |
6393 | ||
6394 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 6395 | perf_remove_from_context(group_leader); |
b04243ef PZ |
6396 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6397 | group_entry) { | |
fe4b04fa | 6398 | perf_remove_from_context(sibling); |
b04243ef PZ |
6399 | put_ctx(gctx); |
6400 | } | |
6401 | mutex_unlock(&gctx->mutex); | |
6402 | put_ctx(gctx); | |
ea635c64 | 6403 | } |
9b51f66d | 6404 | |
cdd6c482 | 6405 | event->filp = event_file; |
ad3a37de | 6406 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 6407 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
6408 | |
6409 | if (move_group) { | |
0cda4c02 | 6410 | synchronize_rcu(); |
e2d37cd2 | 6411 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
6412 | get_ctx(ctx); |
6413 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
6414 | group_entry) { | |
e2d37cd2 | 6415 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
6416 | get_ctx(ctx); |
6417 | } | |
6418 | } | |
6419 | ||
e2d37cd2 | 6420 | perf_install_in_context(ctx, event, event->cpu); |
ad3a37de | 6421 | ++ctx->generation; |
fe4b04fa | 6422 | perf_unpin_context(ctx); |
d859e29f | 6423 | mutex_unlock(&ctx->mutex); |
9b51f66d | 6424 | |
fbfc623f YZ |
6425 | put_online_cpus(); |
6426 | ||
cdd6c482 | 6427 | event->owner = current; |
8882135b | 6428 | |
cdd6c482 IM |
6429 | mutex_lock(¤t->perf_event_mutex); |
6430 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
6431 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 6432 | |
c320c7b7 ACM |
6433 | /* |
6434 | * Precalculate sample_data sizes | |
6435 | */ | |
6436 | perf_event__header_size(event); | |
6844c09d | 6437 | perf_event__id_header_size(event); |
c320c7b7 | 6438 | |
8a49542c PZ |
6439 | /* |
6440 | * Drop the reference on the group_event after placing the | |
6441 | * new event on the sibling_list. This ensures destruction | |
6442 | * of the group leader will find the pointer to itself in | |
6443 | * perf_group_detach(). | |
6444 | */ | |
ea635c64 AV |
6445 | fput_light(group_file, fput_needed); |
6446 | fd_install(event_fd, event_file); | |
6447 | return event_fd; | |
0793a61d | 6448 | |
c3f00c70 | 6449 | err_context: |
fe4b04fa | 6450 | perf_unpin_context(ctx); |
ea635c64 | 6451 | put_ctx(ctx); |
c6be5a5c | 6452 | err_alloc: |
ea635c64 | 6453 | free_event(event); |
e7d0bc04 | 6454 | err_task: |
fbfc623f | 6455 | put_online_cpus(); |
e7d0bc04 PZ |
6456 | if (task) |
6457 | put_task_struct(task); | |
89a1e187 | 6458 | err_group_fd: |
dc86cabe | 6459 | fput_light(group_file, fput_needed); |
ea635c64 AV |
6460 | err_fd: |
6461 | put_unused_fd(event_fd); | |
dc86cabe | 6462 | return err; |
0793a61d TG |
6463 | } |
6464 | ||
fb0459d7 AV |
6465 | /** |
6466 | * perf_event_create_kernel_counter | |
6467 | * | |
6468 | * @attr: attributes of the counter to create | |
6469 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 6470 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
6471 | */ |
6472 | struct perf_event * | |
6473 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 6474 | struct task_struct *task, |
4dc0da86 AK |
6475 | perf_overflow_handler_t overflow_handler, |
6476 | void *context) | |
fb0459d7 | 6477 | { |
fb0459d7 | 6478 | struct perf_event_context *ctx; |
c3f00c70 | 6479 | struct perf_event *event; |
fb0459d7 | 6480 | int err; |
d859e29f | 6481 | |
fb0459d7 AV |
6482 | /* |
6483 | * Get the target context (task or percpu): | |
6484 | */ | |
d859e29f | 6485 | |
4dc0da86 AK |
6486 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
6487 | overflow_handler, context); | |
c3f00c70 PZ |
6488 | if (IS_ERR(event)) { |
6489 | err = PTR_ERR(event); | |
6490 | goto err; | |
6491 | } | |
d859e29f | 6492 | |
38a81da2 | 6493 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
6494 | if (IS_ERR(ctx)) { |
6495 | err = PTR_ERR(ctx); | |
c3f00c70 | 6496 | goto err_free; |
d859e29f | 6497 | } |
fb0459d7 AV |
6498 | |
6499 | event->filp = NULL; | |
6500 | WARN_ON_ONCE(ctx->parent_ctx); | |
6501 | mutex_lock(&ctx->mutex); | |
6502 | perf_install_in_context(ctx, event, cpu); | |
6503 | ++ctx->generation; | |
fe4b04fa | 6504 | perf_unpin_context(ctx); |
fb0459d7 AV |
6505 | mutex_unlock(&ctx->mutex); |
6506 | ||
fb0459d7 AV |
6507 | return event; |
6508 | ||
c3f00c70 PZ |
6509 | err_free: |
6510 | free_event(event); | |
6511 | err: | |
c6567f64 | 6512 | return ERR_PTR(err); |
9b51f66d | 6513 | } |
fb0459d7 | 6514 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 6515 | |
0cda4c02 YZ |
6516 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
6517 | { | |
6518 | struct perf_event_context *src_ctx; | |
6519 | struct perf_event_context *dst_ctx; | |
6520 | struct perf_event *event, *tmp; | |
6521 | LIST_HEAD(events); | |
6522 | ||
6523 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
6524 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
6525 | ||
6526 | mutex_lock(&src_ctx->mutex); | |
6527 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
6528 | event_entry) { | |
6529 | perf_remove_from_context(event); | |
6530 | put_ctx(src_ctx); | |
6531 | list_add(&event->event_entry, &events); | |
6532 | } | |
6533 | mutex_unlock(&src_ctx->mutex); | |
6534 | ||
6535 | synchronize_rcu(); | |
6536 | ||
6537 | mutex_lock(&dst_ctx->mutex); | |
6538 | list_for_each_entry_safe(event, tmp, &events, event_entry) { | |
6539 | list_del(&event->event_entry); | |
6540 | if (event->state >= PERF_EVENT_STATE_OFF) | |
6541 | event->state = PERF_EVENT_STATE_INACTIVE; | |
6542 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
6543 | get_ctx(dst_ctx); | |
6544 | } | |
6545 | mutex_unlock(&dst_ctx->mutex); | |
6546 | } | |
6547 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
6548 | ||
cdd6c482 | 6549 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 6550 | struct task_struct *child) |
d859e29f | 6551 | { |
cdd6c482 | 6552 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 6553 | u64 child_val; |
d859e29f | 6554 | |
cdd6c482 IM |
6555 | if (child_event->attr.inherit_stat) |
6556 | perf_event_read_event(child_event, child); | |
38b200d6 | 6557 | |
b5e58793 | 6558 | child_val = perf_event_count(child_event); |
d859e29f PM |
6559 | |
6560 | /* | |
6561 | * Add back the child's count to the parent's count: | |
6562 | */ | |
a6e6dea6 | 6563 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
6564 | atomic64_add(child_event->total_time_enabled, |
6565 | &parent_event->child_total_time_enabled); | |
6566 | atomic64_add(child_event->total_time_running, | |
6567 | &parent_event->child_total_time_running); | |
d859e29f PM |
6568 | |
6569 | /* | |
cdd6c482 | 6570 | * Remove this event from the parent's list |
d859e29f | 6571 | */ |
cdd6c482 IM |
6572 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
6573 | mutex_lock(&parent_event->child_mutex); | |
6574 | list_del_init(&child_event->child_list); | |
6575 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
6576 | |
6577 | /* | |
cdd6c482 | 6578 | * Release the parent event, if this was the last |
d859e29f PM |
6579 | * reference to it. |
6580 | */ | |
cdd6c482 | 6581 | fput(parent_event->filp); |
d859e29f PM |
6582 | } |
6583 | ||
9b51f66d | 6584 | static void |
cdd6c482 IM |
6585 | __perf_event_exit_task(struct perf_event *child_event, |
6586 | struct perf_event_context *child_ctx, | |
38b200d6 | 6587 | struct task_struct *child) |
9b51f66d | 6588 | { |
38b435b1 PZ |
6589 | if (child_event->parent) { |
6590 | raw_spin_lock_irq(&child_ctx->lock); | |
6591 | perf_group_detach(child_event); | |
6592 | raw_spin_unlock_irq(&child_ctx->lock); | |
6593 | } | |
9b51f66d | 6594 | |
fe4b04fa | 6595 | perf_remove_from_context(child_event); |
0cc0c027 | 6596 | |
9b51f66d | 6597 | /* |
38b435b1 | 6598 | * It can happen that the parent exits first, and has events |
9b51f66d | 6599 | * that are still around due to the child reference. These |
38b435b1 | 6600 | * events need to be zapped. |
9b51f66d | 6601 | */ |
38b435b1 | 6602 | if (child_event->parent) { |
cdd6c482 IM |
6603 | sync_child_event(child_event, child); |
6604 | free_event(child_event); | |
4bcf349a | 6605 | } |
9b51f66d IM |
6606 | } |
6607 | ||
8dc85d54 | 6608 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 6609 | { |
cdd6c482 IM |
6610 | struct perf_event *child_event, *tmp; |
6611 | struct perf_event_context *child_ctx; | |
a63eaf34 | 6612 | unsigned long flags; |
9b51f66d | 6613 | |
8dc85d54 | 6614 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 6615 | perf_event_task(child, NULL, 0); |
9b51f66d | 6616 | return; |
9f498cc5 | 6617 | } |
9b51f66d | 6618 | |
a63eaf34 | 6619 | local_irq_save(flags); |
ad3a37de PM |
6620 | /* |
6621 | * We can't reschedule here because interrupts are disabled, | |
6622 | * and either child is current or it is a task that can't be | |
6623 | * scheduled, so we are now safe from rescheduling changing | |
6624 | * our context. | |
6625 | */ | |
806839b2 | 6626 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
6627 | |
6628 | /* | |
6629 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 6630 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
6631 | * incremented the context's refcount before we do put_ctx below. |
6632 | */ | |
e625cce1 | 6633 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 6634 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 6635 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
6636 | /* |
6637 | * If this context is a clone; unclone it so it can't get | |
6638 | * swapped to another process while we're removing all | |
cdd6c482 | 6639 | * the events from it. |
71a851b4 PZ |
6640 | */ |
6641 | unclone_ctx(child_ctx); | |
5e942bb3 | 6642 | update_context_time(child_ctx); |
e625cce1 | 6643 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
6644 | |
6645 | /* | |
cdd6c482 IM |
6646 | * Report the task dead after unscheduling the events so that we |
6647 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
6648 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 6649 | */ |
cdd6c482 | 6650 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 6651 | |
66fff224 PZ |
6652 | /* |
6653 | * We can recurse on the same lock type through: | |
6654 | * | |
cdd6c482 IM |
6655 | * __perf_event_exit_task() |
6656 | * sync_child_event() | |
6657 | * fput(parent_event->filp) | |
66fff224 PZ |
6658 | * perf_release() |
6659 | * mutex_lock(&ctx->mutex) | |
6660 | * | |
6661 | * But since its the parent context it won't be the same instance. | |
6662 | */ | |
a0507c84 | 6663 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 6664 | |
8bc20959 | 6665 | again: |
889ff015 FW |
6666 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
6667 | group_entry) | |
6668 | __perf_event_exit_task(child_event, child_ctx, child); | |
6669 | ||
6670 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 6671 | group_entry) |
cdd6c482 | 6672 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
6673 | |
6674 | /* | |
cdd6c482 | 6675 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
6676 | * its siblings to the list, but we obtained 'tmp' before that which |
6677 | * will still point to the list head terminating the iteration. | |
6678 | */ | |
889ff015 FW |
6679 | if (!list_empty(&child_ctx->pinned_groups) || |
6680 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 6681 | goto again; |
a63eaf34 PM |
6682 | |
6683 | mutex_unlock(&child_ctx->mutex); | |
6684 | ||
6685 | put_ctx(child_ctx); | |
9b51f66d IM |
6686 | } |
6687 | ||
8dc85d54 PZ |
6688 | /* |
6689 | * When a child task exits, feed back event values to parent events. | |
6690 | */ | |
6691 | void perf_event_exit_task(struct task_struct *child) | |
6692 | { | |
8882135b | 6693 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
6694 | int ctxn; |
6695 | ||
8882135b PZ |
6696 | mutex_lock(&child->perf_event_mutex); |
6697 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
6698 | owner_entry) { | |
6699 | list_del_init(&event->owner_entry); | |
6700 | ||
6701 | /* | |
6702 | * Ensure the list deletion is visible before we clear | |
6703 | * the owner, closes a race against perf_release() where | |
6704 | * we need to serialize on the owner->perf_event_mutex. | |
6705 | */ | |
6706 | smp_wmb(); | |
6707 | event->owner = NULL; | |
6708 | } | |
6709 | mutex_unlock(&child->perf_event_mutex); | |
6710 | ||
8dc85d54 PZ |
6711 | for_each_task_context_nr(ctxn) |
6712 | perf_event_exit_task_context(child, ctxn); | |
6713 | } | |
6714 | ||
889ff015 FW |
6715 | static void perf_free_event(struct perf_event *event, |
6716 | struct perf_event_context *ctx) | |
6717 | { | |
6718 | struct perf_event *parent = event->parent; | |
6719 | ||
6720 | if (WARN_ON_ONCE(!parent)) | |
6721 | return; | |
6722 | ||
6723 | mutex_lock(&parent->child_mutex); | |
6724 | list_del_init(&event->child_list); | |
6725 | mutex_unlock(&parent->child_mutex); | |
6726 | ||
6727 | fput(parent->filp); | |
6728 | ||
8a49542c | 6729 | perf_group_detach(event); |
889ff015 FW |
6730 | list_del_event(event, ctx); |
6731 | free_event(event); | |
6732 | } | |
6733 | ||
bbbee908 PZ |
6734 | /* |
6735 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 6736 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 6737 | */ |
cdd6c482 | 6738 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 6739 | { |
8dc85d54 | 6740 | struct perf_event_context *ctx; |
cdd6c482 | 6741 | struct perf_event *event, *tmp; |
8dc85d54 | 6742 | int ctxn; |
bbbee908 | 6743 | |
8dc85d54 PZ |
6744 | for_each_task_context_nr(ctxn) { |
6745 | ctx = task->perf_event_ctxp[ctxn]; | |
6746 | if (!ctx) | |
6747 | continue; | |
bbbee908 | 6748 | |
8dc85d54 | 6749 | mutex_lock(&ctx->mutex); |
bbbee908 | 6750 | again: |
8dc85d54 PZ |
6751 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
6752 | group_entry) | |
6753 | perf_free_event(event, ctx); | |
bbbee908 | 6754 | |
8dc85d54 PZ |
6755 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
6756 | group_entry) | |
6757 | perf_free_event(event, ctx); | |
bbbee908 | 6758 | |
8dc85d54 PZ |
6759 | if (!list_empty(&ctx->pinned_groups) || |
6760 | !list_empty(&ctx->flexible_groups)) | |
6761 | goto again; | |
bbbee908 | 6762 | |
8dc85d54 | 6763 | mutex_unlock(&ctx->mutex); |
bbbee908 | 6764 | |
8dc85d54 PZ |
6765 | put_ctx(ctx); |
6766 | } | |
889ff015 FW |
6767 | } |
6768 | ||
4e231c79 PZ |
6769 | void perf_event_delayed_put(struct task_struct *task) |
6770 | { | |
6771 | int ctxn; | |
6772 | ||
6773 | for_each_task_context_nr(ctxn) | |
6774 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
6775 | } | |
6776 | ||
97dee4f3 PZ |
6777 | /* |
6778 | * inherit a event from parent task to child task: | |
6779 | */ | |
6780 | static struct perf_event * | |
6781 | inherit_event(struct perf_event *parent_event, | |
6782 | struct task_struct *parent, | |
6783 | struct perf_event_context *parent_ctx, | |
6784 | struct task_struct *child, | |
6785 | struct perf_event *group_leader, | |
6786 | struct perf_event_context *child_ctx) | |
6787 | { | |
6788 | struct perf_event *child_event; | |
cee010ec | 6789 | unsigned long flags; |
97dee4f3 PZ |
6790 | |
6791 | /* | |
6792 | * Instead of creating recursive hierarchies of events, | |
6793 | * we link inherited events back to the original parent, | |
6794 | * which has a filp for sure, which we use as the reference | |
6795 | * count: | |
6796 | */ | |
6797 | if (parent_event->parent) | |
6798 | parent_event = parent_event->parent; | |
6799 | ||
6800 | child_event = perf_event_alloc(&parent_event->attr, | |
6801 | parent_event->cpu, | |
d580ff86 | 6802 | child, |
97dee4f3 | 6803 | group_leader, parent_event, |
4dc0da86 | 6804 | NULL, NULL); |
97dee4f3 PZ |
6805 | if (IS_ERR(child_event)) |
6806 | return child_event; | |
6807 | get_ctx(child_ctx); | |
6808 | ||
6809 | /* | |
6810 | * Make the child state follow the state of the parent event, | |
6811 | * not its attr.disabled bit. We hold the parent's mutex, | |
6812 | * so we won't race with perf_event_{en, dis}able_family. | |
6813 | */ | |
6814 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
6815 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
6816 | else | |
6817 | child_event->state = PERF_EVENT_STATE_OFF; | |
6818 | ||
6819 | if (parent_event->attr.freq) { | |
6820 | u64 sample_period = parent_event->hw.sample_period; | |
6821 | struct hw_perf_event *hwc = &child_event->hw; | |
6822 | ||
6823 | hwc->sample_period = sample_period; | |
6824 | hwc->last_period = sample_period; | |
6825 | ||
6826 | local64_set(&hwc->period_left, sample_period); | |
6827 | } | |
6828 | ||
6829 | child_event->ctx = child_ctx; | |
6830 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
6831 | child_event->overflow_handler_context |
6832 | = parent_event->overflow_handler_context; | |
97dee4f3 | 6833 | |
614b6780 TG |
6834 | /* |
6835 | * Precalculate sample_data sizes | |
6836 | */ | |
6837 | perf_event__header_size(child_event); | |
6844c09d | 6838 | perf_event__id_header_size(child_event); |
614b6780 | 6839 | |
97dee4f3 PZ |
6840 | /* |
6841 | * Link it up in the child's context: | |
6842 | */ | |
cee010ec | 6843 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 6844 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 6845 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 PZ |
6846 | |
6847 | /* | |
6848 | * Get a reference to the parent filp - we will fput it | |
6849 | * when the child event exits. This is safe to do because | |
6850 | * we are in the parent and we know that the filp still | |
6851 | * exists and has a nonzero count: | |
6852 | */ | |
6853 | atomic_long_inc(&parent_event->filp->f_count); | |
6854 | ||
6855 | /* | |
6856 | * Link this into the parent event's child list | |
6857 | */ | |
6858 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
6859 | mutex_lock(&parent_event->child_mutex); | |
6860 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
6861 | mutex_unlock(&parent_event->child_mutex); | |
6862 | ||
6863 | return child_event; | |
6864 | } | |
6865 | ||
6866 | static int inherit_group(struct perf_event *parent_event, | |
6867 | struct task_struct *parent, | |
6868 | struct perf_event_context *parent_ctx, | |
6869 | struct task_struct *child, | |
6870 | struct perf_event_context *child_ctx) | |
6871 | { | |
6872 | struct perf_event *leader; | |
6873 | struct perf_event *sub; | |
6874 | struct perf_event *child_ctr; | |
6875 | ||
6876 | leader = inherit_event(parent_event, parent, parent_ctx, | |
6877 | child, NULL, child_ctx); | |
6878 | if (IS_ERR(leader)) | |
6879 | return PTR_ERR(leader); | |
6880 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
6881 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
6882 | child, leader, child_ctx); | |
6883 | if (IS_ERR(child_ctr)) | |
6884 | return PTR_ERR(child_ctr); | |
6885 | } | |
6886 | return 0; | |
889ff015 FW |
6887 | } |
6888 | ||
6889 | static int | |
6890 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
6891 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 6892 | struct task_struct *child, int ctxn, |
889ff015 FW |
6893 | int *inherited_all) |
6894 | { | |
6895 | int ret; | |
8dc85d54 | 6896 | struct perf_event_context *child_ctx; |
889ff015 FW |
6897 | |
6898 | if (!event->attr.inherit) { | |
6899 | *inherited_all = 0; | |
6900 | return 0; | |
bbbee908 PZ |
6901 | } |
6902 | ||
fe4b04fa | 6903 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
6904 | if (!child_ctx) { |
6905 | /* | |
6906 | * This is executed from the parent task context, so | |
6907 | * inherit events that have been marked for cloning. | |
6908 | * First allocate and initialize a context for the | |
6909 | * child. | |
6910 | */ | |
bbbee908 | 6911 | |
eb184479 | 6912 | child_ctx = alloc_perf_context(event->pmu, child); |
889ff015 FW |
6913 | if (!child_ctx) |
6914 | return -ENOMEM; | |
bbbee908 | 6915 | |
8dc85d54 | 6916 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
6917 | } |
6918 | ||
6919 | ret = inherit_group(event, parent, parent_ctx, | |
6920 | child, child_ctx); | |
6921 | ||
6922 | if (ret) | |
6923 | *inherited_all = 0; | |
6924 | ||
6925 | return ret; | |
bbbee908 PZ |
6926 | } |
6927 | ||
9b51f66d | 6928 | /* |
cdd6c482 | 6929 | * Initialize the perf_event context in task_struct |
9b51f66d | 6930 | */ |
8dc85d54 | 6931 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 6932 | { |
889ff015 | 6933 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
6934 | struct perf_event_context *cloned_ctx; |
6935 | struct perf_event *event; | |
9b51f66d | 6936 | struct task_struct *parent = current; |
564c2b21 | 6937 | int inherited_all = 1; |
dddd3379 | 6938 | unsigned long flags; |
6ab423e0 | 6939 | int ret = 0; |
9b51f66d | 6940 | |
8dc85d54 | 6941 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
6942 | return 0; |
6943 | ||
ad3a37de | 6944 | /* |
25346b93 PM |
6945 | * If the parent's context is a clone, pin it so it won't get |
6946 | * swapped under us. | |
ad3a37de | 6947 | */ |
8dc85d54 | 6948 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 6949 | |
ad3a37de PM |
6950 | /* |
6951 | * No need to check if parent_ctx != NULL here; since we saw | |
6952 | * it non-NULL earlier, the only reason for it to become NULL | |
6953 | * is if we exit, and since we're currently in the middle of | |
6954 | * a fork we can't be exiting at the same time. | |
6955 | */ | |
ad3a37de | 6956 | |
9b51f66d IM |
6957 | /* |
6958 | * Lock the parent list. No need to lock the child - not PID | |
6959 | * hashed yet and not running, so nobody can access it. | |
6960 | */ | |
d859e29f | 6961 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
6962 | |
6963 | /* | |
6964 | * We dont have to disable NMIs - we are only looking at | |
6965 | * the list, not manipulating it: | |
6966 | */ | |
889ff015 | 6967 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
6968 | ret = inherit_task_group(event, parent, parent_ctx, |
6969 | child, ctxn, &inherited_all); | |
889ff015 FW |
6970 | if (ret) |
6971 | break; | |
6972 | } | |
b93f7978 | 6973 | |
dddd3379 TG |
6974 | /* |
6975 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
6976 | * to allocations, but we need to prevent rotation because | |
6977 | * rotate_ctx() will change the list from interrupt context. | |
6978 | */ | |
6979 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
6980 | parent_ctx->rotate_disable = 1; | |
6981 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
6982 | ||
889ff015 | 6983 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
6984 | ret = inherit_task_group(event, parent, parent_ctx, |
6985 | child, ctxn, &inherited_all); | |
889ff015 | 6986 | if (ret) |
9b51f66d | 6987 | break; |
564c2b21 PM |
6988 | } |
6989 | ||
dddd3379 TG |
6990 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
6991 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 6992 | |
8dc85d54 | 6993 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 6994 | |
05cbaa28 | 6995 | if (child_ctx && inherited_all) { |
564c2b21 PM |
6996 | /* |
6997 | * Mark the child context as a clone of the parent | |
6998 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
6999 | * |
7000 | * Note that if the parent is a clone, the holding of | |
7001 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7002 | */ |
c5ed5145 | 7003 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7004 | if (cloned_ctx) { |
7005 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7006 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7007 | } else { |
7008 | child_ctx->parent_ctx = parent_ctx; | |
7009 | child_ctx->parent_gen = parent_ctx->generation; | |
7010 | } | |
7011 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7012 | } |
7013 | ||
c5ed5145 | 7014 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7015 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7016 | |
25346b93 | 7017 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7018 | put_ctx(parent_ctx); |
ad3a37de | 7019 | |
6ab423e0 | 7020 | return ret; |
9b51f66d IM |
7021 | } |
7022 | ||
8dc85d54 PZ |
7023 | /* |
7024 | * Initialize the perf_event context in task_struct | |
7025 | */ | |
7026 | int perf_event_init_task(struct task_struct *child) | |
7027 | { | |
7028 | int ctxn, ret; | |
7029 | ||
8550d7cb ON |
7030 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7031 | mutex_init(&child->perf_event_mutex); | |
7032 | INIT_LIST_HEAD(&child->perf_event_list); | |
7033 | ||
8dc85d54 PZ |
7034 | for_each_task_context_nr(ctxn) { |
7035 | ret = perf_event_init_context(child, ctxn); | |
7036 | if (ret) | |
7037 | return ret; | |
7038 | } | |
7039 | ||
7040 | return 0; | |
7041 | } | |
7042 | ||
220b140b PM |
7043 | static void __init perf_event_init_all_cpus(void) |
7044 | { | |
b28ab83c | 7045 | struct swevent_htable *swhash; |
220b140b | 7046 | int cpu; |
220b140b PM |
7047 | |
7048 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7049 | swhash = &per_cpu(swevent_htable, cpu); |
7050 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7051 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7052 | } |
7053 | } | |
7054 | ||
cdd6c482 | 7055 | static void __cpuinit perf_event_init_cpu(int cpu) |
0793a61d | 7056 | { |
108b02cf | 7057 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7058 | |
b28ab83c | 7059 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7060 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7061 | struct swevent_hlist *hlist; |
7062 | ||
b28ab83c PZ |
7063 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7064 | WARN_ON(!hlist); | |
7065 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7066 | } |
b28ab83c | 7067 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7068 | } |
7069 | ||
c277443c | 7070 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7071 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7072 | { |
e9d2b064 PZ |
7073 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7074 | ||
7075 | WARN_ON(!irqs_disabled()); | |
7076 | ||
7077 | list_del_init(&cpuctx->rotation_list); | |
7078 | } | |
7079 | ||
108b02cf | 7080 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7081 | { |
108b02cf | 7082 | struct perf_event_context *ctx = __info; |
cdd6c482 | 7083 | struct perf_event *event, *tmp; |
0793a61d | 7084 | |
108b02cf | 7085 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7086 | |
889ff015 | 7087 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 7088 | __perf_remove_from_context(event); |
889ff015 | 7089 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 7090 | __perf_remove_from_context(event); |
0793a61d | 7091 | } |
108b02cf PZ |
7092 | |
7093 | static void perf_event_exit_cpu_context(int cpu) | |
7094 | { | |
7095 | struct perf_event_context *ctx; | |
7096 | struct pmu *pmu; | |
7097 | int idx; | |
7098 | ||
7099 | idx = srcu_read_lock(&pmus_srcu); | |
7100 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7101 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7102 | |
7103 | mutex_lock(&ctx->mutex); | |
7104 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7105 | mutex_unlock(&ctx->mutex); | |
7106 | } | |
7107 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7108 | } |
7109 | ||
cdd6c482 | 7110 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7111 | { |
b28ab83c | 7112 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7113 | |
b28ab83c PZ |
7114 | mutex_lock(&swhash->hlist_mutex); |
7115 | swevent_hlist_release(swhash); | |
7116 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 7117 | |
108b02cf | 7118 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7119 | } |
7120 | #else | |
cdd6c482 | 7121 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7122 | #endif |
7123 | ||
c277443c PZ |
7124 | static int |
7125 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7126 | { | |
7127 | int cpu; | |
7128 | ||
7129 | for_each_online_cpu(cpu) | |
7130 | perf_event_exit_cpu(cpu); | |
7131 | ||
7132 | return NOTIFY_OK; | |
7133 | } | |
7134 | ||
7135 | /* | |
7136 | * Run the perf reboot notifier at the very last possible moment so that | |
7137 | * the generic watchdog code runs as long as possible. | |
7138 | */ | |
7139 | static struct notifier_block perf_reboot_notifier = { | |
7140 | .notifier_call = perf_reboot, | |
7141 | .priority = INT_MIN, | |
7142 | }; | |
7143 | ||
0793a61d TG |
7144 | static int __cpuinit |
7145 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
7146 | { | |
7147 | unsigned int cpu = (long)hcpu; | |
7148 | ||
4536e4d1 | 7149 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7150 | |
7151 | case CPU_UP_PREPARE: | |
5e11637e | 7152 | case CPU_DOWN_FAILED: |
cdd6c482 | 7153 | perf_event_init_cpu(cpu); |
0793a61d TG |
7154 | break; |
7155 | ||
5e11637e | 7156 | case CPU_UP_CANCELED: |
0793a61d | 7157 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7158 | perf_event_exit_cpu(cpu); |
0793a61d TG |
7159 | break; |
7160 | ||
7161 | default: | |
7162 | break; | |
7163 | } | |
7164 | ||
7165 | return NOTIFY_OK; | |
7166 | } | |
7167 | ||
cdd6c482 | 7168 | void __init perf_event_init(void) |
0793a61d | 7169 | { |
3c502e7a JW |
7170 | int ret; |
7171 | ||
2e80a82a PZ |
7172 | idr_init(&pmu_idr); |
7173 | ||
220b140b | 7174 | perf_event_init_all_cpus(); |
b0a873eb | 7175 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7176 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7177 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7178 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7179 | perf_tp_register(); |
7180 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7181 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7182 | |
7183 | ret = init_hw_breakpoint(); | |
7184 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7185 | |
7186 | /* do not patch jump label more than once per second */ | |
7187 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7188 | |
7189 | /* | |
7190 | * Build time assertion that we keep the data_head at the intended | |
7191 | * location. IOW, validation we got the __reserved[] size right. | |
7192 | */ | |
7193 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
7194 | != 1024); | |
0793a61d | 7195 | } |
abe43400 PZ |
7196 | |
7197 | static int __init perf_event_sysfs_init(void) | |
7198 | { | |
7199 | struct pmu *pmu; | |
7200 | int ret; | |
7201 | ||
7202 | mutex_lock(&pmus_lock); | |
7203 | ||
7204 | ret = bus_register(&pmu_bus); | |
7205 | if (ret) | |
7206 | goto unlock; | |
7207 | ||
7208 | list_for_each_entry(pmu, &pmus, entry) { | |
7209 | if (!pmu->name || pmu->type < 0) | |
7210 | continue; | |
7211 | ||
7212 | ret = pmu_dev_alloc(pmu); | |
7213 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
7214 | } | |
7215 | pmu_bus_running = 1; | |
7216 | ret = 0; | |
7217 | ||
7218 | unlock: | |
7219 | mutex_unlock(&pmus_lock); | |
7220 | ||
7221 | return ret; | |
7222 | } | |
7223 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
7224 | |
7225 | #ifdef CONFIG_CGROUP_PERF | |
761b3ef5 | 7226 | static struct cgroup_subsys_state *perf_cgroup_create(struct cgroup *cont) |
e5d1367f SE |
7227 | { |
7228 | struct perf_cgroup *jc; | |
e5d1367f | 7229 | |
1b15d055 | 7230 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
7231 | if (!jc) |
7232 | return ERR_PTR(-ENOMEM); | |
7233 | ||
e5d1367f SE |
7234 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7235 | if (!jc->info) { | |
7236 | kfree(jc); | |
7237 | return ERR_PTR(-ENOMEM); | |
7238 | } | |
7239 | ||
e5d1367f SE |
7240 | return &jc->css; |
7241 | } | |
7242 | ||
761b3ef5 | 7243 | static void perf_cgroup_destroy(struct cgroup *cont) |
e5d1367f SE |
7244 | { |
7245 | struct perf_cgroup *jc; | |
7246 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), | |
7247 | struct perf_cgroup, css); | |
7248 | free_percpu(jc->info); | |
7249 | kfree(jc); | |
7250 | } | |
7251 | ||
7252 | static int __perf_cgroup_move(void *info) | |
7253 | { | |
7254 | struct task_struct *task = info; | |
7255 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
7256 | return 0; | |
7257 | } | |
7258 | ||
761b3ef5 | 7259 | static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
e5d1367f | 7260 | { |
bb9d97b6 TH |
7261 | struct task_struct *task; |
7262 | ||
7263 | cgroup_taskset_for_each(task, cgrp, tset) | |
7264 | task_function_call(task, __perf_cgroup_move, task); | |
e5d1367f SE |
7265 | } |
7266 | ||
761b3ef5 LZ |
7267 | static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, |
7268 | struct task_struct *task) | |
e5d1367f SE |
7269 | { |
7270 | /* | |
7271 | * cgroup_exit() is called in the copy_process() failure path. | |
7272 | * Ignore this case since the task hasn't ran yet, this avoids | |
7273 | * trying to poke a half freed task state from generic code. | |
7274 | */ | |
7275 | if (!(task->flags & PF_EXITING)) | |
7276 | return; | |
7277 | ||
bb9d97b6 | 7278 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7279 | } |
7280 | ||
7281 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
7282 | .name = "perf_event", |
7283 | .subsys_id = perf_subsys_id, | |
7284 | .create = perf_cgroup_create, | |
7285 | .destroy = perf_cgroup_destroy, | |
7286 | .exit = perf_cgroup_exit, | |
bb9d97b6 | 7287 | .attach = perf_cgroup_attach, |
e5d1367f SE |
7288 | }; |
7289 | #endif /* CONFIG_CGROUP_PERF */ |