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 | */ |
500ad2d8 | 1256 | 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 | */ | |
a6fa941d | 2938 | static void put_event(struct perf_event *event) |
fb0459d7 | 2939 | { |
8882135b | 2940 | struct task_struct *owner; |
fb0459d7 | 2941 | |
a6fa941d AV |
2942 | if (!atomic_long_dec_and_test(&event->refcount)) |
2943 | return; | |
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 | ||
a6fa941d AV |
2978 | perf_event_release_kernel(event); |
2979 | } | |
2980 | ||
2981 | static int perf_release(struct inode *inode, struct file *file) | |
2982 | { | |
2983 | put_event(file->private_data); | |
2984 | return 0; | |
fb0459d7 | 2985 | } |
fb0459d7 | 2986 | |
59ed446f | 2987 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 2988 | { |
cdd6c482 | 2989 | struct perf_event *child; |
e53c0994 PZ |
2990 | u64 total = 0; |
2991 | ||
59ed446f PZ |
2992 | *enabled = 0; |
2993 | *running = 0; | |
2994 | ||
6f10581a | 2995 | mutex_lock(&event->child_mutex); |
cdd6c482 | 2996 | total += perf_event_read(event); |
59ed446f PZ |
2997 | *enabled += event->total_time_enabled + |
2998 | atomic64_read(&event->child_total_time_enabled); | |
2999 | *running += event->total_time_running + | |
3000 | atomic64_read(&event->child_total_time_running); | |
3001 | ||
3002 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3003 | total += perf_event_read(child); |
59ed446f PZ |
3004 | *enabled += child->total_time_enabled; |
3005 | *running += child->total_time_running; | |
3006 | } | |
6f10581a | 3007 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3008 | |
3009 | return total; | |
3010 | } | |
fb0459d7 | 3011 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3012 | |
cdd6c482 | 3013 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3014 | u64 read_format, char __user *buf) |
3015 | { | |
cdd6c482 | 3016 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3017 | int n = 0, size = 0, ret = -EFAULT; |
3018 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3019 | u64 values[5]; |
59ed446f | 3020 | u64 count, enabled, running; |
abf4868b | 3021 | |
6f10581a | 3022 | mutex_lock(&ctx->mutex); |
59ed446f | 3023 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3024 | |
3025 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3026 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3027 | values[n++] = enabled; | |
3028 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3029 | values[n++] = running; | |
abf4868b PZ |
3030 | values[n++] = count; |
3031 | if (read_format & PERF_FORMAT_ID) | |
3032 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3033 | |
3034 | size = n * sizeof(u64); | |
3035 | ||
3036 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3037 | goto unlock; |
3dab77fb | 3038 | |
6f10581a | 3039 | ret = size; |
3dab77fb | 3040 | |
65abc865 | 3041 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3042 | n = 0; |
3dab77fb | 3043 | |
59ed446f | 3044 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3045 | if (read_format & PERF_FORMAT_ID) |
3046 | values[n++] = primary_event_id(sub); | |
3047 | ||
3048 | size = n * sizeof(u64); | |
3049 | ||
184d3da8 | 3050 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3051 | ret = -EFAULT; |
3052 | goto unlock; | |
3053 | } | |
abf4868b PZ |
3054 | |
3055 | ret += size; | |
3dab77fb | 3056 | } |
6f10581a PZ |
3057 | unlock: |
3058 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3059 | |
abf4868b | 3060 | return ret; |
3dab77fb PZ |
3061 | } |
3062 | ||
cdd6c482 | 3063 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3064 | u64 read_format, char __user *buf) |
3065 | { | |
59ed446f | 3066 | u64 enabled, running; |
3dab77fb PZ |
3067 | u64 values[4]; |
3068 | int n = 0; | |
3069 | ||
59ed446f PZ |
3070 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3071 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3072 | values[n++] = enabled; | |
3073 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3074 | values[n++] = running; | |
3dab77fb | 3075 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3076 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3077 | |
3078 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3079 | return -EFAULT; | |
3080 | ||
3081 | return n * sizeof(u64); | |
3082 | } | |
3083 | ||
0793a61d | 3084 | /* |
cdd6c482 | 3085 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3086 | */ |
3087 | static ssize_t | |
cdd6c482 | 3088 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3089 | { |
cdd6c482 | 3090 | u64 read_format = event->attr.read_format; |
3dab77fb | 3091 | int ret; |
0793a61d | 3092 | |
3b6f9e5c | 3093 | /* |
cdd6c482 | 3094 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3095 | * error state (i.e. because it was pinned but it couldn't be |
3096 | * scheduled on to the CPU at some point). | |
3097 | */ | |
cdd6c482 | 3098 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3099 | return 0; |
3100 | ||
c320c7b7 | 3101 | if (count < event->read_size) |
3dab77fb PZ |
3102 | return -ENOSPC; |
3103 | ||
cdd6c482 | 3104 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3105 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3106 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3107 | else |
cdd6c482 | 3108 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3109 | |
3dab77fb | 3110 | return ret; |
0793a61d TG |
3111 | } |
3112 | ||
0793a61d TG |
3113 | static ssize_t |
3114 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3115 | { | |
cdd6c482 | 3116 | struct perf_event *event = file->private_data; |
0793a61d | 3117 | |
cdd6c482 | 3118 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3119 | } |
3120 | ||
3121 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3122 | { | |
cdd6c482 | 3123 | struct perf_event *event = file->private_data; |
76369139 | 3124 | struct ring_buffer *rb; |
c33a0bc4 | 3125 | unsigned int events = POLL_HUP; |
c7138f37 | 3126 | |
10c6db11 PZ |
3127 | /* |
3128 | * Race between perf_event_set_output() and perf_poll(): perf_poll() | |
3129 | * grabs the rb reference but perf_event_set_output() overrides it. | |
3130 | * Here is the timeline for two threads T1, T2: | |
3131 | * t0: T1, rb = rcu_dereference(event->rb) | |
3132 | * t1: T2, old_rb = event->rb | |
3133 | * t2: T2, event->rb = new rb | |
3134 | * t3: T2, ring_buffer_detach(old_rb) | |
3135 | * t4: T1, ring_buffer_attach(rb1) | |
3136 | * t5: T1, poll_wait(event->waitq) | |
3137 | * | |
3138 | * To avoid this problem, we grab mmap_mutex in perf_poll() | |
3139 | * thereby ensuring that the assignment of the new ring buffer | |
3140 | * and the detachment of the old buffer appear atomic to perf_poll() | |
3141 | */ | |
3142 | mutex_lock(&event->mmap_mutex); | |
3143 | ||
c7138f37 | 3144 | rcu_read_lock(); |
76369139 | 3145 | rb = rcu_dereference(event->rb); |
10c6db11 PZ |
3146 | if (rb) { |
3147 | ring_buffer_attach(event, rb); | |
76369139 | 3148 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3149 | } |
c7138f37 | 3150 | rcu_read_unlock(); |
0793a61d | 3151 | |
10c6db11 PZ |
3152 | mutex_unlock(&event->mmap_mutex); |
3153 | ||
cdd6c482 | 3154 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3155 | |
0793a61d TG |
3156 | return events; |
3157 | } | |
3158 | ||
cdd6c482 | 3159 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3160 | { |
cdd6c482 | 3161 | (void)perf_event_read(event); |
e7850595 | 3162 | local64_set(&event->count, 0); |
cdd6c482 | 3163 | perf_event_update_userpage(event); |
3df5edad PZ |
3164 | } |
3165 | ||
c93f7669 | 3166 | /* |
cdd6c482 IM |
3167 | * Holding the top-level event's child_mutex means that any |
3168 | * descendant process that has inherited this event will block | |
3169 | * in sync_child_event if it goes to exit, thus satisfying the | |
3170 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3171 | */ |
cdd6c482 IM |
3172 | static void perf_event_for_each_child(struct perf_event *event, |
3173 | void (*func)(struct perf_event *)) | |
3df5edad | 3174 | { |
cdd6c482 | 3175 | struct perf_event *child; |
3df5edad | 3176 | |
cdd6c482 IM |
3177 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3178 | mutex_lock(&event->child_mutex); | |
3179 | func(event); | |
3180 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3181 | func(child); |
cdd6c482 | 3182 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3183 | } |
3184 | ||
cdd6c482 IM |
3185 | static void perf_event_for_each(struct perf_event *event, |
3186 | void (*func)(struct perf_event *)) | |
3df5edad | 3187 | { |
cdd6c482 IM |
3188 | struct perf_event_context *ctx = event->ctx; |
3189 | struct perf_event *sibling; | |
3df5edad | 3190 | |
75f937f2 PZ |
3191 | WARN_ON_ONCE(ctx->parent_ctx); |
3192 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3193 | event = event->group_leader; |
75f937f2 | 3194 | |
cdd6c482 | 3195 | perf_event_for_each_child(event, func); |
cdd6c482 | 3196 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3197 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3198 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3199 | } |
3200 | ||
cdd6c482 | 3201 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3202 | { |
cdd6c482 | 3203 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3204 | int ret = 0; |
3205 | u64 value; | |
3206 | ||
6c7e550f | 3207 | if (!is_sampling_event(event)) |
08247e31 PZ |
3208 | return -EINVAL; |
3209 | ||
ad0cf347 | 3210 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3211 | return -EFAULT; |
3212 | ||
3213 | if (!value) | |
3214 | return -EINVAL; | |
3215 | ||
e625cce1 | 3216 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3217 | if (event->attr.freq) { |
3218 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3219 | ret = -EINVAL; |
3220 | goto unlock; | |
3221 | } | |
3222 | ||
cdd6c482 | 3223 | event->attr.sample_freq = value; |
08247e31 | 3224 | } else { |
cdd6c482 IM |
3225 | event->attr.sample_period = value; |
3226 | event->hw.sample_period = value; | |
08247e31 PZ |
3227 | } |
3228 | unlock: | |
e625cce1 | 3229 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3230 | |
3231 | return ret; | |
3232 | } | |
3233 | ||
ac9721f3 PZ |
3234 | static const struct file_operations perf_fops; |
3235 | ||
a6fa941d | 3236 | static struct file *perf_fget_light(int fd, int *fput_needed) |
ac9721f3 PZ |
3237 | { |
3238 | struct file *file; | |
3239 | ||
3240 | file = fget_light(fd, fput_needed); | |
3241 | if (!file) | |
3242 | return ERR_PTR(-EBADF); | |
3243 | ||
3244 | if (file->f_op != &perf_fops) { | |
3245 | fput_light(file, *fput_needed); | |
3246 | *fput_needed = 0; | |
3247 | return ERR_PTR(-EBADF); | |
3248 | } | |
3249 | ||
a6fa941d | 3250 | return file; |
ac9721f3 PZ |
3251 | } |
3252 | ||
3253 | static int perf_event_set_output(struct perf_event *event, | |
3254 | struct perf_event *output_event); | |
6fb2915d | 3255 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3256 | |
d859e29f PM |
3257 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3258 | { | |
cdd6c482 IM |
3259 | struct perf_event *event = file->private_data; |
3260 | void (*func)(struct perf_event *); | |
3df5edad | 3261 | u32 flags = arg; |
d859e29f PM |
3262 | |
3263 | switch (cmd) { | |
cdd6c482 IM |
3264 | case PERF_EVENT_IOC_ENABLE: |
3265 | func = perf_event_enable; | |
d859e29f | 3266 | break; |
cdd6c482 IM |
3267 | case PERF_EVENT_IOC_DISABLE: |
3268 | func = perf_event_disable; | |
79f14641 | 3269 | break; |
cdd6c482 IM |
3270 | case PERF_EVENT_IOC_RESET: |
3271 | func = perf_event_reset; | |
6de6a7b9 | 3272 | break; |
3df5edad | 3273 | |
cdd6c482 IM |
3274 | case PERF_EVENT_IOC_REFRESH: |
3275 | return perf_event_refresh(event, arg); | |
08247e31 | 3276 | |
cdd6c482 IM |
3277 | case PERF_EVENT_IOC_PERIOD: |
3278 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3279 | |
cdd6c482 | 3280 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3281 | { |
a6fa941d | 3282 | struct file *output_file = NULL; |
ac9721f3 PZ |
3283 | struct perf_event *output_event = NULL; |
3284 | int fput_needed = 0; | |
3285 | int ret; | |
3286 | ||
3287 | if (arg != -1) { | |
a6fa941d AV |
3288 | output_file = perf_fget_light(arg, &fput_needed); |
3289 | if (IS_ERR(output_file)) | |
3290 | return PTR_ERR(output_file); | |
3291 | output_event = output_file->private_data; | |
ac9721f3 PZ |
3292 | } |
3293 | ||
3294 | ret = perf_event_set_output(event, output_event); | |
3295 | if (output_event) | |
a6fa941d | 3296 | fput_light(output_file, fput_needed); |
ac9721f3 PZ |
3297 | |
3298 | return ret; | |
3299 | } | |
a4be7c27 | 3300 | |
6fb2915d LZ |
3301 | case PERF_EVENT_IOC_SET_FILTER: |
3302 | return perf_event_set_filter(event, (void __user *)arg); | |
3303 | ||
d859e29f | 3304 | default: |
3df5edad | 3305 | return -ENOTTY; |
d859e29f | 3306 | } |
3df5edad PZ |
3307 | |
3308 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3309 | perf_event_for_each(event, func); |
3df5edad | 3310 | else |
cdd6c482 | 3311 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3312 | |
3313 | return 0; | |
d859e29f PM |
3314 | } |
3315 | ||
cdd6c482 | 3316 | int perf_event_task_enable(void) |
771d7cde | 3317 | { |
cdd6c482 | 3318 | struct perf_event *event; |
771d7cde | 3319 | |
cdd6c482 IM |
3320 | mutex_lock(¤t->perf_event_mutex); |
3321 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3322 | perf_event_for_each_child(event, perf_event_enable); | |
3323 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3324 | |
3325 | return 0; | |
3326 | } | |
3327 | ||
cdd6c482 | 3328 | int perf_event_task_disable(void) |
771d7cde | 3329 | { |
cdd6c482 | 3330 | struct perf_event *event; |
771d7cde | 3331 | |
cdd6c482 IM |
3332 | mutex_lock(¤t->perf_event_mutex); |
3333 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3334 | perf_event_for_each_child(event, perf_event_disable); | |
3335 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3336 | |
3337 | return 0; | |
3338 | } | |
3339 | ||
cdd6c482 | 3340 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3341 | { |
a4eaf7f1 PZ |
3342 | if (event->hw.state & PERF_HES_STOPPED) |
3343 | return 0; | |
3344 | ||
cdd6c482 | 3345 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3346 | return 0; |
3347 | ||
35edc2a5 | 3348 | return event->pmu->event_idx(event); |
194002b2 PZ |
3349 | } |
3350 | ||
c4794295 | 3351 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3352 | u64 *now, |
7f310a5d EM |
3353 | u64 *enabled, |
3354 | u64 *running) | |
c4794295 | 3355 | { |
e3f3541c | 3356 | u64 ctx_time; |
c4794295 | 3357 | |
e3f3541c PZ |
3358 | *now = perf_clock(); |
3359 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3360 | *enabled = ctx_time - event->tstamp_enabled; |
3361 | *running = ctx_time - event->tstamp_running; | |
3362 | } | |
3363 | ||
c7206205 | 3364 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3365 | { |
3366 | } | |
3367 | ||
38ff667b PZ |
3368 | /* |
3369 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3370 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3371 | * code calls this from NMI context. | |
3372 | */ | |
cdd6c482 | 3373 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3374 | { |
cdd6c482 | 3375 | struct perf_event_mmap_page *userpg; |
76369139 | 3376 | struct ring_buffer *rb; |
e3f3541c | 3377 | u64 enabled, running, now; |
38ff667b PZ |
3378 | |
3379 | rcu_read_lock(); | |
0d641208 EM |
3380 | /* |
3381 | * compute total_time_enabled, total_time_running | |
3382 | * based on snapshot values taken when the event | |
3383 | * was last scheduled in. | |
3384 | * | |
3385 | * we cannot simply called update_context_time() | |
3386 | * because of locking issue as we can be called in | |
3387 | * NMI context | |
3388 | */ | |
e3f3541c | 3389 | calc_timer_values(event, &now, &enabled, &running); |
76369139 FW |
3390 | rb = rcu_dereference(event->rb); |
3391 | if (!rb) | |
38ff667b PZ |
3392 | goto unlock; |
3393 | ||
76369139 | 3394 | userpg = rb->user_page; |
37d81828 | 3395 | |
7b732a75 PZ |
3396 | /* |
3397 | * Disable preemption so as to not let the corresponding user-space | |
3398 | * spin too long if we get preempted. | |
3399 | */ | |
3400 | preempt_disable(); | |
37d81828 | 3401 | ++userpg->lock; |
92f22a38 | 3402 | barrier(); |
cdd6c482 | 3403 | userpg->index = perf_event_index(event); |
b5e58793 | 3404 | userpg->offset = perf_event_count(event); |
365a4038 | 3405 | if (userpg->index) |
e7850595 | 3406 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3407 | |
0d641208 | 3408 | userpg->time_enabled = enabled + |
cdd6c482 | 3409 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3410 | |
0d641208 | 3411 | userpg->time_running = running + |
cdd6c482 | 3412 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3413 | |
c7206205 | 3414 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3415 | |
92f22a38 | 3416 | barrier(); |
37d81828 | 3417 | ++userpg->lock; |
7b732a75 | 3418 | preempt_enable(); |
38ff667b | 3419 | unlock: |
7b732a75 | 3420 | rcu_read_unlock(); |
37d81828 PM |
3421 | } |
3422 | ||
906010b2 PZ |
3423 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3424 | { | |
3425 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3426 | struct ring_buffer *rb; |
906010b2 PZ |
3427 | int ret = VM_FAULT_SIGBUS; |
3428 | ||
3429 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3430 | if (vmf->pgoff == 0) | |
3431 | ret = 0; | |
3432 | return ret; | |
3433 | } | |
3434 | ||
3435 | rcu_read_lock(); | |
76369139 FW |
3436 | rb = rcu_dereference(event->rb); |
3437 | if (!rb) | |
906010b2 PZ |
3438 | goto unlock; |
3439 | ||
3440 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3441 | goto unlock; | |
3442 | ||
76369139 | 3443 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3444 | if (!vmf->page) |
3445 | goto unlock; | |
3446 | ||
3447 | get_page(vmf->page); | |
3448 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3449 | vmf->page->index = vmf->pgoff; | |
3450 | ||
3451 | ret = 0; | |
3452 | unlock: | |
3453 | rcu_read_unlock(); | |
3454 | ||
3455 | return ret; | |
3456 | } | |
3457 | ||
10c6db11 PZ |
3458 | static void ring_buffer_attach(struct perf_event *event, |
3459 | struct ring_buffer *rb) | |
3460 | { | |
3461 | unsigned long flags; | |
3462 | ||
3463 | if (!list_empty(&event->rb_entry)) | |
3464 | return; | |
3465 | ||
3466 | spin_lock_irqsave(&rb->event_lock, flags); | |
3467 | if (!list_empty(&event->rb_entry)) | |
3468 | goto unlock; | |
3469 | ||
3470 | list_add(&event->rb_entry, &rb->event_list); | |
3471 | unlock: | |
3472 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3473 | } | |
3474 | ||
3475 | static void ring_buffer_detach(struct perf_event *event, | |
3476 | struct ring_buffer *rb) | |
3477 | { | |
3478 | unsigned long flags; | |
3479 | ||
3480 | if (list_empty(&event->rb_entry)) | |
3481 | return; | |
3482 | ||
3483 | spin_lock_irqsave(&rb->event_lock, flags); | |
3484 | list_del_init(&event->rb_entry); | |
3485 | wake_up_all(&event->waitq); | |
3486 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3487 | } | |
3488 | ||
3489 | static void ring_buffer_wakeup(struct perf_event *event) | |
3490 | { | |
3491 | struct ring_buffer *rb; | |
3492 | ||
3493 | rcu_read_lock(); | |
3494 | rb = rcu_dereference(event->rb); | |
44b7f4b9 WD |
3495 | if (!rb) |
3496 | goto unlock; | |
3497 | ||
3498 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
10c6db11 | 3499 | wake_up_all(&event->waitq); |
44b7f4b9 WD |
3500 | |
3501 | unlock: | |
10c6db11 PZ |
3502 | rcu_read_unlock(); |
3503 | } | |
3504 | ||
76369139 | 3505 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3506 | { |
76369139 | 3507 | struct ring_buffer *rb; |
906010b2 | 3508 | |
76369139 FW |
3509 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3510 | rb_free(rb); | |
7b732a75 PZ |
3511 | } |
3512 | ||
76369139 | 3513 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3514 | { |
76369139 | 3515 | struct ring_buffer *rb; |
7b732a75 | 3516 | |
ac9721f3 | 3517 | rcu_read_lock(); |
76369139 FW |
3518 | rb = rcu_dereference(event->rb); |
3519 | if (rb) { | |
3520 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3521 | rb = NULL; | |
ac9721f3 PZ |
3522 | } |
3523 | rcu_read_unlock(); | |
3524 | ||
76369139 | 3525 | return rb; |
ac9721f3 PZ |
3526 | } |
3527 | ||
76369139 | 3528 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3529 | { |
10c6db11 PZ |
3530 | struct perf_event *event, *n; |
3531 | unsigned long flags; | |
3532 | ||
76369139 | 3533 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3534 | return; |
7b732a75 | 3535 | |
10c6db11 PZ |
3536 | spin_lock_irqsave(&rb->event_lock, flags); |
3537 | list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) { | |
3538 | list_del_init(&event->rb_entry); | |
3539 | wake_up_all(&event->waitq); | |
3540 | } | |
3541 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3542 | ||
76369139 | 3543 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3544 | } |
3545 | ||
3546 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3547 | { | |
cdd6c482 | 3548 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3549 | |
cdd6c482 | 3550 | atomic_inc(&event->mmap_count); |
7b732a75 PZ |
3551 | } |
3552 | ||
3553 | static void perf_mmap_close(struct vm_area_struct *vma) | |
3554 | { | |
cdd6c482 | 3555 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3556 | |
cdd6c482 | 3557 | if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) { |
76369139 | 3558 | unsigned long size = perf_data_size(event->rb); |
ac9721f3 | 3559 | struct user_struct *user = event->mmap_user; |
76369139 | 3560 | struct ring_buffer *rb = event->rb; |
789f90fc | 3561 | |
906010b2 | 3562 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm); |
bc3e53f6 | 3563 | vma->vm_mm->pinned_vm -= event->mmap_locked; |
76369139 | 3564 | rcu_assign_pointer(event->rb, NULL); |
10c6db11 | 3565 | ring_buffer_detach(event, rb); |
cdd6c482 | 3566 | mutex_unlock(&event->mmap_mutex); |
ac9721f3 | 3567 | |
76369139 | 3568 | ring_buffer_put(rb); |
ac9721f3 | 3569 | free_uid(user); |
7b732a75 | 3570 | } |
37d81828 PM |
3571 | } |
3572 | ||
f0f37e2f | 3573 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3574 | .open = perf_mmap_open, |
3575 | .close = perf_mmap_close, | |
3576 | .fault = perf_mmap_fault, | |
3577 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3578 | }; |
3579 | ||
3580 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3581 | { | |
cdd6c482 | 3582 | struct perf_event *event = file->private_data; |
22a4f650 | 3583 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3584 | struct user_struct *user = current_user(); |
22a4f650 | 3585 | unsigned long locked, lock_limit; |
76369139 | 3586 | struct ring_buffer *rb; |
7b732a75 PZ |
3587 | unsigned long vma_size; |
3588 | unsigned long nr_pages; | |
789f90fc | 3589 | long user_extra, extra; |
d57e34fd | 3590 | int ret = 0, flags = 0; |
37d81828 | 3591 | |
c7920614 PZ |
3592 | /* |
3593 | * Don't allow mmap() of inherited per-task counters. This would | |
3594 | * create a performance issue due to all children writing to the | |
76369139 | 3595 | * same rb. |
c7920614 PZ |
3596 | */ |
3597 | if (event->cpu == -1 && event->attr.inherit) | |
3598 | return -EINVAL; | |
3599 | ||
43a21ea8 | 3600 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3601 | return -EINVAL; |
7b732a75 PZ |
3602 | |
3603 | vma_size = vma->vm_end - vma->vm_start; | |
3604 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3605 | ||
7730d865 | 3606 | /* |
76369139 | 3607 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3608 | * can do bitmasks instead of modulo. |
3609 | */ | |
3610 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3611 | return -EINVAL; |
3612 | ||
7b732a75 | 3613 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3614 | return -EINVAL; |
3615 | ||
7b732a75 PZ |
3616 | if (vma->vm_pgoff != 0) |
3617 | return -EINVAL; | |
37d81828 | 3618 | |
cdd6c482 IM |
3619 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3620 | mutex_lock(&event->mmap_mutex); | |
76369139 FW |
3621 | if (event->rb) { |
3622 | if (event->rb->nr_pages == nr_pages) | |
3623 | atomic_inc(&event->rb->refcount); | |
ac9721f3 | 3624 | else |
ebb3c4c4 PZ |
3625 | ret = -EINVAL; |
3626 | goto unlock; | |
3627 | } | |
3628 | ||
789f90fc | 3629 | user_extra = nr_pages + 1; |
cdd6c482 | 3630 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
3631 | |
3632 | /* | |
3633 | * Increase the limit linearly with more CPUs: | |
3634 | */ | |
3635 | user_lock_limit *= num_online_cpus(); | |
3636 | ||
789f90fc | 3637 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 3638 | |
789f90fc PZ |
3639 | extra = 0; |
3640 | if (user_locked > user_lock_limit) | |
3641 | extra = user_locked - user_lock_limit; | |
7b732a75 | 3642 | |
78d7d407 | 3643 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 3644 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 3645 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 3646 | |
459ec28a IM |
3647 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
3648 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
3649 | ret = -EPERM; |
3650 | goto unlock; | |
3651 | } | |
7b732a75 | 3652 | |
76369139 | 3653 | WARN_ON(event->rb); |
906010b2 | 3654 | |
d57e34fd | 3655 | if (vma->vm_flags & VM_WRITE) |
76369139 | 3656 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 3657 | |
4ec8363d VW |
3658 | rb = rb_alloc(nr_pages, |
3659 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
3660 | event->cpu, flags); | |
3661 | ||
76369139 | 3662 | if (!rb) { |
ac9721f3 | 3663 | ret = -ENOMEM; |
ebb3c4c4 | 3664 | goto unlock; |
ac9721f3 | 3665 | } |
76369139 | 3666 | rcu_assign_pointer(event->rb, rb); |
43a21ea8 | 3667 | |
ac9721f3 PZ |
3668 | atomic_long_add(user_extra, &user->locked_vm); |
3669 | event->mmap_locked = extra; | |
3670 | event->mmap_user = get_current_user(); | |
bc3e53f6 | 3671 | vma->vm_mm->pinned_vm += event->mmap_locked; |
ac9721f3 | 3672 | |
9a0f05cb PZ |
3673 | perf_event_update_userpage(event); |
3674 | ||
ebb3c4c4 | 3675 | unlock: |
ac9721f3 PZ |
3676 | if (!ret) |
3677 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 3678 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 3679 | |
37d81828 PM |
3680 | vma->vm_flags |= VM_RESERVED; |
3681 | vma->vm_ops = &perf_mmap_vmops; | |
7b732a75 PZ |
3682 | |
3683 | return ret; | |
37d81828 PM |
3684 | } |
3685 | ||
3c446b3d PZ |
3686 | static int perf_fasync(int fd, struct file *filp, int on) |
3687 | { | |
3c446b3d | 3688 | struct inode *inode = filp->f_path.dentry->d_inode; |
cdd6c482 | 3689 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
3690 | int retval; |
3691 | ||
3692 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 3693 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
3694 | mutex_unlock(&inode->i_mutex); |
3695 | ||
3696 | if (retval < 0) | |
3697 | return retval; | |
3698 | ||
3699 | return 0; | |
3700 | } | |
3701 | ||
0793a61d | 3702 | static const struct file_operations perf_fops = { |
3326c1ce | 3703 | .llseek = no_llseek, |
0793a61d TG |
3704 | .release = perf_release, |
3705 | .read = perf_read, | |
3706 | .poll = perf_poll, | |
d859e29f PM |
3707 | .unlocked_ioctl = perf_ioctl, |
3708 | .compat_ioctl = perf_ioctl, | |
37d81828 | 3709 | .mmap = perf_mmap, |
3c446b3d | 3710 | .fasync = perf_fasync, |
0793a61d TG |
3711 | }; |
3712 | ||
925d519a | 3713 | /* |
cdd6c482 | 3714 | * Perf event wakeup |
925d519a PZ |
3715 | * |
3716 | * If there's data, ensure we set the poll() state and publish everything | |
3717 | * to user-space before waking everybody up. | |
3718 | */ | |
3719 | ||
cdd6c482 | 3720 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 3721 | { |
10c6db11 | 3722 | ring_buffer_wakeup(event); |
4c9e2542 | 3723 | |
cdd6c482 IM |
3724 | if (event->pending_kill) { |
3725 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
3726 | event->pending_kill = 0; | |
4c9e2542 | 3727 | } |
925d519a PZ |
3728 | } |
3729 | ||
e360adbe | 3730 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 3731 | { |
cdd6c482 IM |
3732 | struct perf_event *event = container_of(entry, |
3733 | struct perf_event, pending); | |
79f14641 | 3734 | |
cdd6c482 IM |
3735 | if (event->pending_disable) { |
3736 | event->pending_disable = 0; | |
3737 | __perf_event_disable(event); | |
79f14641 PZ |
3738 | } |
3739 | ||
cdd6c482 IM |
3740 | if (event->pending_wakeup) { |
3741 | event->pending_wakeup = 0; | |
3742 | perf_event_wakeup(event); | |
79f14641 PZ |
3743 | } |
3744 | } | |
3745 | ||
39447b38 ZY |
3746 | /* |
3747 | * We assume there is only KVM supporting the callbacks. | |
3748 | * Later on, we might change it to a list if there is | |
3749 | * another virtualization implementation supporting the callbacks. | |
3750 | */ | |
3751 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
3752 | ||
3753 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3754 | { | |
3755 | perf_guest_cbs = cbs; | |
3756 | return 0; | |
3757 | } | |
3758 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
3759 | ||
3760 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3761 | { | |
3762 | perf_guest_cbs = NULL; | |
3763 | return 0; | |
3764 | } | |
3765 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
3766 | ||
c980d109 ACM |
3767 | static void __perf_event_header__init_id(struct perf_event_header *header, |
3768 | struct perf_sample_data *data, | |
3769 | struct perf_event *event) | |
6844c09d ACM |
3770 | { |
3771 | u64 sample_type = event->attr.sample_type; | |
3772 | ||
3773 | data->type = sample_type; | |
3774 | header->size += event->id_header_size; | |
3775 | ||
3776 | if (sample_type & PERF_SAMPLE_TID) { | |
3777 | /* namespace issues */ | |
3778 | data->tid_entry.pid = perf_event_pid(event, current); | |
3779 | data->tid_entry.tid = perf_event_tid(event, current); | |
3780 | } | |
3781 | ||
3782 | if (sample_type & PERF_SAMPLE_TIME) | |
3783 | data->time = perf_clock(); | |
3784 | ||
3785 | if (sample_type & PERF_SAMPLE_ID) | |
3786 | data->id = primary_event_id(event); | |
3787 | ||
3788 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3789 | data->stream_id = event->id; | |
3790 | ||
3791 | if (sample_type & PERF_SAMPLE_CPU) { | |
3792 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
3793 | data->cpu_entry.reserved = 0; | |
3794 | } | |
3795 | } | |
3796 | ||
76369139 FW |
3797 | void perf_event_header__init_id(struct perf_event_header *header, |
3798 | struct perf_sample_data *data, | |
3799 | struct perf_event *event) | |
c980d109 ACM |
3800 | { |
3801 | if (event->attr.sample_id_all) | |
3802 | __perf_event_header__init_id(header, data, event); | |
3803 | } | |
3804 | ||
3805 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
3806 | struct perf_sample_data *data) | |
3807 | { | |
3808 | u64 sample_type = data->type; | |
3809 | ||
3810 | if (sample_type & PERF_SAMPLE_TID) | |
3811 | perf_output_put(handle, data->tid_entry); | |
3812 | ||
3813 | if (sample_type & PERF_SAMPLE_TIME) | |
3814 | perf_output_put(handle, data->time); | |
3815 | ||
3816 | if (sample_type & PERF_SAMPLE_ID) | |
3817 | perf_output_put(handle, data->id); | |
3818 | ||
3819 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3820 | perf_output_put(handle, data->stream_id); | |
3821 | ||
3822 | if (sample_type & PERF_SAMPLE_CPU) | |
3823 | perf_output_put(handle, data->cpu_entry); | |
3824 | } | |
3825 | ||
76369139 FW |
3826 | void perf_event__output_id_sample(struct perf_event *event, |
3827 | struct perf_output_handle *handle, | |
3828 | struct perf_sample_data *sample) | |
c980d109 ACM |
3829 | { |
3830 | if (event->attr.sample_id_all) | |
3831 | __perf_event__output_id_sample(handle, sample); | |
3832 | } | |
3833 | ||
3dab77fb | 3834 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
3835 | struct perf_event *event, |
3836 | u64 enabled, u64 running) | |
3dab77fb | 3837 | { |
cdd6c482 | 3838 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
3839 | u64 values[4]; |
3840 | int n = 0; | |
3841 | ||
b5e58793 | 3842 | values[n++] = perf_event_count(event); |
3dab77fb | 3843 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 3844 | values[n++] = enabled + |
cdd6c482 | 3845 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
3846 | } |
3847 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 3848 | values[n++] = running + |
cdd6c482 | 3849 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
3850 | } |
3851 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 3852 | values[n++] = primary_event_id(event); |
3dab77fb | 3853 | |
76369139 | 3854 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
3855 | } |
3856 | ||
3857 | /* | |
cdd6c482 | 3858 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
3859 | */ |
3860 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
3861 | struct perf_event *event, |
3862 | u64 enabled, u64 running) | |
3dab77fb | 3863 | { |
cdd6c482 IM |
3864 | struct perf_event *leader = event->group_leader, *sub; |
3865 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
3866 | u64 values[5]; |
3867 | int n = 0; | |
3868 | ||
3869 | values[n++] = 1 + leader->nr_siblings; | |
3870 | ||
3871 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 3872 | values[n++] = enabled; |
3dab77fb PZ |
3873 | |
3874 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 3875 | values[n++] = running; |
3dab77fb | 3876 | |
cdd6c482 | 3877 | if (leader != event) |
3dab77fb PZ |
3878 | leader->pmu->read(leader); |
3879 | ||
b5e58793 | 3880 | values[n++] = perf_event_count(leader); |
3dab77fb | 3881 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3882 | values[n++] = primary_event_id(leader); |
3dab77fb | 3883 | |
76369139 | 3884 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 3885 | |
65abc865 | 3886 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
3887 | n = 0; |
3888 | ||
cdd6c482 | 3889 | if (sub != event) |
3dab77fb PZ |
3890 | sub->pmu->read(sub); |
3891 | ||
b5e58793 | 3892 | values[n++] = perf_event_count(sub); |
3dab77fb | 3893 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3894 | values[n++] = primary_event_id(sub); |
3dab77fb | 3895 | |
76369139 | 3896 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
3897 | } |
3898 | } | |
3899 | ||
eed01528 SE |
3900 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
3901 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3902 | ||
3dab77fb | 3903 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 3904 | struct perf_event *event) |
3dab77fb | 3905 | { |
e3f3541c | 3906 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
3907 | u64 read_format = event->attr.read_format; |
3908 | ||
3909 | /* | |
3910 | * compute total_time_enabled, total_time_running | |
3911 | * based on snapshot values taken when the event | |
3912 | * was last scheduled in. | |
3913 | * | |
3914 | * we cannot simply called update_context_time() | |
3915 | * because of locking issue as we are called in | |
3916 | * NMI context | |
3917 | */ | |
c4794295 | 3918 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 3919 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 3920 | |
cdd6c482 | 3921 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 3922 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 3923 | else |
eed01528 | 3924 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
3925 | } |
3926 | ||
5622f295 MM |
3927 | void perf_output_sample(struct perf_output_handle *handle, |
3928 | struct perf_event_header *header, | |
3929 | struct perf_sample_data *data, | |
cdd6c482 | 3930 | struct perf_event *event) |
5622f295 MM |
3931 | { |
3932 | u64 sample_type = data->type; | |
3933 | ||
3934 | perf_output_put(handle, *header); | |
3935 | ||
3936 | if (sample_type & PERF_SAMPLE_IP) | |
3937 | perf_output_put(handle, data->ip); | |
3938 | ||
3939 | if (sample_type & PERF_SAMPLE_TID) | |
3940 | perf_output_put(handle, data->tid_entry); | |
3941 | ||
3942 | if (sample_type & PERF_SAMPLE_TIME) | |
3943 | perf_output_put(handle, data->time); | |
3944 | ||
3945 | if (sample_type & PERF_SAMPLE_ADDR) | |
3946 | perf_output_put(handle, data->addr); | |
3947 | ||
3948 | if (sample_type & PERF_SAMPLE_ID) | |
3949 | perf_output_put(handle, data->id); | |
3950 | ||
3951 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3952 | perf_output_put(handle, data->stream_id); | |
3953 | ||
3954 | if (sample_type & PERF_SAMPLE_CPU) | |
3955 | perf_output_put(handle, data->cpu_entry); | |
3956 | ||
3957 | if (sample_type & PERF_SAMPLE_PERIOD) | |
3958 | perf_output_put(handle, data->period); | |
3959 | ||
3960 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 3961 | perf_output_read(handle, event); |
5622f295 MM |
3962 | |
3963 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
3964 | if (data->callchain) { | |
3965 | int size = 1; | |
3966 | ||
3967 | if (data->callchain) | |
3968 | size += data->callchain->nr; | |
3969 | ||
3970 | size *= sizeof(u64); | |
3971 | ||
76369139 | 3972 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
3973 | } else { |
3974 | u64 nr = 0; | |
3975 | perf_output_put(handle, nr); | |
3976 | } | |
3977 | } | |
3978 | ||
3979 | if (sample_type & PERF_SAMPLE_RAW) { | |
3980 | if (data->raw) { | |
3981 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
3982 | __output_copy(handle, data->raw->data, |
3983 | data->raw->size); | |
5622f295 MM |
3984 | } else { |
3985 | struct { | |
3986 | u32 size; | |
3987 | u32 data; | |
3988 | } raw = { | |
3989 | .size = sizeof(u32), | |
3990 | .data = 0, | |
3991 | }; | |
3992 | perf_output_put(handle, raw); | |
3993 | } | |
3994 | } | |
a7ac67ea PZ |
3995 | |
3996 | if (!event->attr.watermark) { | |
3997 | int wakeup_events = event->attr.wakeup_events; | |
3998 | ||
3999 | if (wakeup_events) { | |
4000 | struct ring_buffer *rb = handle->rb; | |
4001 | int events = local_inc_return(&rb->events); | |
4002 | ||
4003 | if (events >= wakeup_events) { | |
4004 | local_sub(wakeup_events, &rb->events); | |
4005 | local_inc(&rb->wakeup); | |
4006 | } | |
4007 | } | |
4008 | } | |
bce38cd5 SE |
4009 | |
4010 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4011 | if (data->br_stack) { | |
4012 | size_t size; | |
4013 | ||
4014 | size = data->br_stack->nr | |
4015 | * sizeof(struct perf_branch_entry); | |
4016 | ||
4017 | perf_output_put(handle, data->br_stack->nr); | |
4018 | perf_output_copy(handle, data->br_stack->entries, size); | |
4019 | } else { | |
4020 | /* | |
4021 | * we always store at least the value of nr | |
4022 | */ | |
4023 | u64 nr = 0; | |
4024 | perf_output_put(handle, nr); | |
4025 | } | |
4026 | } | |
5622f295 MM |
4027 | } |
4028 | ||
4029 | void perf_prepare_sample(struct perf_event_header *header, | |
4030 | struct perf_sample_data *data, | |
cdd6c482 | 4031 | struct perf_event *event, |
5622f295 | 4032 | struct pt_regs *regs) |
7b732a75 | 4033 | { |
cdd6c482 | 4034 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4035 | |
cdd6c482 | 4036 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4037 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4038 | |
4039 | header->misc = 0; | |
4040 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4041 | |
c980d109 | 4042 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4043 | |
c320c7b7 | 4044 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4045 | data->ip = perf_instruction_pointer(regs); |
4046 | ||
b23f3325 | 4047 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4048 | int size = 1; |
394ee076 | 4049 | |
e6dab5ff | 4050 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4051 | |
4052 | if (data->callchain) | |
4053 | size += data->callchain->nr; | |
4054 | ||
4055 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4056 | } |
4057 | ||
3a43ce68 | 4058 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4059 | int size = sizeof(u32); |
4060 | ||
4061 | if (data->raw) | |
4062 | size += data->raw->size; | |
4063 | else | |
4064 | size += sizeof(u32); | |
4065 | ||
4066 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4067 | header->size += size; |
7f453c24 | 4068 | } |
bce38cd5 SE |
4069 | |
4070 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4071 | int size = sizeof(u64); /* nr */ | |
4072 | if (data->br_stack) { | |
4073 | size += data->br_stack->nr | |
4074 | * sizeof(struct perf_branch_entry); | |
4075 | } | |
4076 | header->size += size; | |
4077 | } | |
5622f295 | 4078 | } |
7f453c24 | 4079 | |
a8b0ca17 | 4080 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4081 | struct perf_sample_data *data, |
4082 | struct pt_regs *regs) | |
4083 | { | |
4084 | struct perf_output_handle handle; | |
4085 | struct perf_event_header header; | |
689802b2 | 4086 | |
927c7a9e FW |
4087 | /* protect the callchain buffers */ |
4088 | rcu_read_lock(); | |
4089 | ||
cdd6c482 | 4090 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4091 | |
a7ac67ea | 4092 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4093 | goto exit; |
0322cd6e | 4094 | |
cdd6c482 | 4095 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4096 | |
8a057d84 | 4097 | perf_output_end(&handle); |
927c7a9e FW |
4098 | |
4099 | exit: | |
4100 | rcu_read_unlock(); | |
0322cd6e PZ |
4101 | } |
4102 | ||
38b200d6 | 4103 | /* |
cdd6c482 | 4104 | * read event_id |
38b200d6 PZ |
4105 | */ |
4106 | ||
4107 | struct perf_read_event { | |
4108 | struct perf_event_header header; | |
4109 | ||
4110 | u32 pid; | |
4111 | u32 tid; | |
38b200d6 PZ |
4112 | }; |
4113 | ||
4114 | static void | |
cdd6c482 | 4115 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4116 | struct task_struct *task) |
4117 | { | |
4118 | struct perf_output_handle handle; | |
c980d109 | 4119 | struct perf_sample_data sample; |
dfc65094 | 4120 | struct perf_read_event read_event = { |
38b200d6 | 4121 | .header = { |
cdd6c482 | 4122 | .type = PERF_RECORD_READ, |
38b200d6 | 4123 | .misc = 0, |
c320c7b7 | 4124 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4125 | }, |
cdd6c482 IM |
4126 | .pid = perf_event_pid(event, task), |
4127 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4128 | }; |
3dab77fb | 4129 | int ret; |
38b200d6 | 4130 | |
c980d109 | 4131 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4132 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4133 | if (ret) |
4134 | return; | |
4135 | ||
dfc65094 | 4136 | perf_output_put(&handle, read_event); |
cdd6c482 | 4137 | perf_output_read(&handle, event); |
c980d109 | 4138 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4139 | |
38b200d6 PZ |
4140 | perf_output_end(&handle); |
4141 | } | |
4142 | ||
60313ebe | 4143 | /* |
9f498cc5 PZ |
4144 | * task tracking -- fork/exit |
4145 | * | |
3af9e859 | 4146 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
60313ebe PZ |
4147 | */ |
4148 | ||
9f498cc5 | 4149 | struct perf_task_event { |
3a80b4a3 | 4150 | struct task_struct *task; |
cdd6c482 | 4151 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4152 | |
4153 | struct { | |
4154 | struct perf_event_header header; | |
4155 | ||
4156 | u32 pid; | |
4157 | u32 ppid; | |
9f498cc5 PZ |
4158 | u32 tid; |
4159 | u32 ptid; | |
393b2ad8 | 4160 | u64 time; |
cdd6c482 | 4161 | } event_id; |
60313ebe PZ |
4162 | }; |
4163 | ||
cdd6c482 | 4164 | static void perf_event_task_output(struct perf_event *event, |
9f498cc5 | 4165 | struct perf_task_event *task_event) |
60313ebe PZ |
4166 | { |
4167 | struct perf_output_handle handle; | |
c980d109 | 4168 | struct perf_sample_data sample; |
9f498cc5 | 4169 | struct task_struct *task = task_event->task; |
c980d109 | 4170 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4171 | |
c980d109 | 4172 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4173 | |
c980d109 | 4174 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4175 | task_event->event_id.header.size); |
ef60777c | 4176 | if (ret) |
c980d109 | 4177 | goto out; |
60313ebe | 4178 | |
cdd6c482 IM |
4179 | task_event->event_id.pid = perf_event_pid(event, task); |
4180 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4181 | |
cdd6c482 IM |
4182 | task_event->event_id.tid = perf_event_tid(event, task); |
4183 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4184 | |
cdd6c482 | 4185 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4186 | |
c980d109 ACM |
4187 | perf_event__output_id_sample(event, &handle, &sample); |
4188 | ||
60313ebe | 4189 | perf_output_end(&handle); |
c980d109 ACM |
4190 | out: |
4191 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4192 | } |
4193 | ||
cdd6c482 | 4194 | static int perf_event_task_match(struct perf_event *event) |
60313ebe | 4195 | { |
6f93d0a7 | 4196 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4197 | return 0; |
4198 | ||
5632ab12 | 4199 | if (!event_filter_match(event)) |
5d27c23d PZ |
4200 | return 0; |
4201 | ||
3af9e859 EM |
4202 | if (event->attr.comm || event->attr.mmap || |
4203 | event->attr.mmap_data || event->attr.task) | |
60313ebe PZ |
4204 | return 1; |
4205 | ||
4206 | return 0; | |
4207 | } | |
4208 | ||
cdd6c482 | 4209 | static void perf_event_task_ctx(struct perf_event_context *ctx, |
9f498cc5 | 4210 | struct perf_task_event *task_event) |
60313ebe | 4211 | { |
cdd6c482 | 4212 | struct perf_event *event; |
60313ebe | 4213 | |
cdd6c482 IM |
4214 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4215 | if (perf_event_task_match(event)) | |
4216 | perf_event_task_output(event, task_event); | |
60313ebe | 4217 | } |
60313ebe PZ |
4218 | } |
4219 | ||
cdd6c482 | 4220 | static void perf_event_task_event(struct perf_task_event *task_event) |
60313ebe PZ |
4221 | { |
4222 | struct perf_cpu_context *cpuctx; | |
8dc85d54 | 4223 | struct perf_event_context *ctx; |
108b02cf | 4224 | struct pmu *pmu; |
8dc85d54 | 4225 | int ctxn; |
60313ebe | 4226 | |
d6ff86cf | 4227 | rcu_read_lock(); |
108b02cf | 4228 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4229 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4230 | if (cpuctx->active_pmu != pmu) |
4231 | goto next; | |
108b02cf | 4232 | perf_event_task_ctx(&cpuctx->ctx, task_event); |
8dc85d54 PZ |
4233 | |
4234 | ctx = task_event->task_ctx; | |
4235 | if (!ctx) { | |
4236 | ctxn = pmu->task_ctx_nr; | |
4237 | if (ctxn < 0) | |
41945f6c | 4238 | goto next; |
8dc85d54 PZ |
4239 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); |
4240 | } | |
4241 | if (ctx) | |
4242 | perf_event_task_ctx(ctx, task_event); | |
41945f6c PZ |
4243 | next: |
4244 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4245 | } |
60313ebe PZ |
4246 | rcu_read_unlock(); |
4247 | } | |
4248 | ||
cdd6c482 IM |
4249 | static void perf_event_task(struct task_struct *task, |
4250 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4251 | int new) |
60313ebe | 4252 | { |
9f498cc5 | 4253 | struct perf_task_event task_event; |
60313ebe | 4254 | |
cdd6c482 IM |
4255 | if (!atomic_read(&nr_comm_events) && |
4256 | !atomic_read(&nr_mmap_events) && | |
4257 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4258 | return; |
4259 | ||
9f498cc5 | 4260 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4261 | .task = task, |
4262 | .task_ctx = task_ctx, | |
cdd6c482 | 4263 | .event_id = { |
60313ebe | 4264 | .header = { |
cdd6c482 | 4265 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4266 | .misc = 0, |
cdd6c482 | 4267 | .size = sizeof(task_event.event_id), |
60313ebe | 4268 | }, |
573402db PZ |
4269 | /* .pid */ |
4270 | /* .ppid */ | |
9f498cc5 PZ |
4271 | /* .tid */ |
4272 | /* .ptid */ | |
6f93d0a7 | 4273 | .time = perf_clock(), |
60313ebe PZ |
4274 | }, |
4275 | }; | |
4276 | ||
cdd6c482 | 4277 | perf_event_task_event(&task_event); |
9f498cc5 PZ |
4278 | } |
4279 | ||
cdd6c482 | 4280 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4281 | { |
cdd6c482 | 4282 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4283 | } |
4284 | ||
8d1b2d93 PZ |
4285 | /* |
4286 | * comm tracking | |
4287 | */ | |
4288 | ||
4289 | struct perf_comm_event { | |
22a4f650 IM |
4290 | struct task_struct *task; |
4291 | char *comm; | |
8d1b2d93 PZ |
4292 | int comm_size; |
4293 | ||
4294 | struct { | |
4295 | struct perf_event_header header; | |
4296 | ||
4297 | u32 pid; | |
4298 | u32 tid; | |
cdd6c482 | 4299 | } event_id; |
8d1b2d93 PZ |
4300 | }; |
4301 | ||
cdd6c482 | 4302 | static void perf_event_comm_output(struct perf_event *event, |
8d1b2d93 PZ |
4303 | struct perf_comm_event *comm_event) |
4304 | { | |
4305 | struct perf_output_handle handle; | |
c980d109 | 4306 | struct perf_sample_data sample; |
cdd6c482 | 4307 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4308 | int ret; |
4309 | ||
4310 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); | |
4311 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4312 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4313 | |
4314 | if (ret) | |
c980d109 | 4315 | goto out; |
8d1b2d93 | 4316 | |
cdd6c482 IM |
4317 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4318 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4319 | |
cdd6c482 | 4320 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4321 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4322 | comm_event->comm_size); |
c980d109 ACM |
4323 | |
4324 | perf_event__output_id_sample(event, &handle, &sample); | |
4325 | ||
8d1b2d93 | 4326 | perf_output_end(&handle); |
c980d109 ACM |
4327 | out: |
4328 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4329 | } |
4330 | ||
cdd6c482 | 4331 | static int perf_event_comm_match(struct perf_event *event) |
8d1b2d93 | 4332 | { |
6f93d0a7 | 4333 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4334 | return 0; |
4335 | ||
5632ab12 | 4336 | if (!event_filter_match(event)) |
5d27c23d PZ |
4337 | return 0; |
4338 | ||
cdd6c482 | 4339 | if (event->attr.comm) |
8d1b2d93 PZ |
4340 | return 1; |
4341 | ||
4342 | return 0; | |
4343 | } | |
4344 | ||
cdd6c482 | 4345 | static void perf_event_comm_ctx(struct perf_event_context *ctx, |
8d1b2d93 PZ |
4346 | struct perf_comm_event *comm_event) |
4347 | { | |
cdd6c482 | 4348 | struct perf_event *event; |
8d1b2d93 | 4349 | |
cdd6c482 IM |
4350 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4351 | if (perf_event_comm_match(event)) | |
4352 | perf_event_comm_output(event, comm_event); | |
8d1b2d93 | 4353 | } |
8d1b2d93 PZ |
4354 | } |
4355 | ||
cdd6c482 | 4356 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 PZ |
4357 | { |
4358 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4359 | struct perf_event_context *ctx; |
413ee3b4 | 4360 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4361 | unsigned int size; |
108b02cf | 4362 | struct pmu *pmu; |
8dc85d54 | 4363 | int ctxn; |
8d1b2d93 | 4364 | |
413ee3b4 | 4365 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4366 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4367 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4368 | |
4369 | comm_event->comm = comm; | |
4370 | comm_event->comm_size = size; | |
4371 | ||
cdd6c482 | 4372 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
f6595f3a | 4373 | rcu_read_lock(); |
108b02cf | 4374 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4375 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4376 | if (cpuctx->active_pmu != pmu) |
4377 | goto next; | |
108b02cf | 4378 | perf_event_comm_ctx(&cpuctx->ctx, comm_event); |
8dc85d54 PZ |
4379 | |
4380 | ctxn = pmu->task_ctx_nr; | |
4381 | if (ctxn < 0) | |
41945f6c | 4382 | goto next; |
8dc85d54 PZ |
4383 | |
4384 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4385 | if (ctx) | |
4386 | perf_event_comm_ctx(ctx, comm_event); | |
41945f6c PZ |
4387 | next: |
4388 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4389 | } |
665c2142 | 4390 | rcu_read_unlock(); |
8d1b2d93 PZ |
4391 | } |
4392 | ||
cdd6c482 | 4393 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4394 | { |
9ee318a7 | 4395 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4396 | struct perf_event_context *ctx; |
4397 | int ctxn; | |
9ee318a7 | 4398 | |
8dc85d54 PZ |
4399 | for_each_task_context_nr(ctxn) { |
4400 | ctx = task->perf_event_ctxp[ctxn]; | |
4401 | if (!ctx) | |
4402 | continue; | |
9ee318a7 | 4403 | |
8dc85d54 PZ |
4404 | perf_event_enable_on_exec(ctx); |
4405 | } | |
9ee318a7 | 4406 | |
cdd6c482 | 4407 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4408 | return; |
a63eaf34 | 4409 | |
9ee318a7 | 4410 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4411 | .task = task, |
573402db PZ |
4412 | /* .comm */ |
4413 | /* .comm_size */ | |
cdd6c482 | 4414 | .event_id = { |
573402db | 4415 | .header = { |
cdd6c482 | 4416 | .type = PERF_RECORD_COMM, |
573402db PZ |
4417 | .misc = 0, |
4418 | /* .size */ | |
4419 | }, | |
4420 | /* .pid */ | |
4421 | /* .tid */ | |
8d1b2d93 PZ |
4422 | }, |
4423 | }; | |
4424 | ||
cdd6c482 | 4425 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
4426 | } |
4427 | ||
0a4a9391 PZ |
4428 | /* |
4429 | * mmap tracking | |
4430 | */ | |
4431 | ||
4432 | struct perf_mmap_event { | |
089dd79d PZ |
4433 | struct vm_area_struct *vma; |
4434 | ||
4435 | const char *file_name; | |
4436 | int file_size; | |
0a4a9391 PZ |
4437 | |
4438 | struct { | |
4439 | struct perf_event_header header; | |
4440 | ||
4441 | u32 pid; | |
4442 | u32 tid; | |
4443 | u64 start; | |
4444 | u64 len; | |
4445 | u64 pgoff; | |
cdd6c482 | 4446 | } event_id; |
0a4a9391 PZ |
4447 | }; |
4448 | ||
cdd6c482 | 4449 | static void perf_event_mmap_output(struct perf_event *event, |
0a4a9391 PZ |
4450 | struct perf_mmap_event *mmap_event) |
4451 | { | |
4452 | struct perf_output_handle handle; | |
c980d109 | 4453 | struct perf_sample_data sample; |
cdd6c482 | 4454 | int size = mmap_event->event_id.header.size; |
c980d109 | 4455 | int ret; |
0a4a9391 | 4456 | |
c980d109 ACM |
4457 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
4458 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4459 | mmap_event->event_id.header.size); |
0a4a9391 | 4460 | if (ret) |
c980d109 | 4461 | goto out; |
0a4a9391 | 4462 | |
cdd6c482 IM |
4463 | mmap_event->event_id.pid = perf_event_pid(event, current); |
4464 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 4465 | |
cdd6c482 | 4466 | perf_output_put(&handle, mmap_event->event_id); |
76369139 | 4467 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 4468 | mmap_event->file_size); |
c980d109 ACM |
4469 | |
4470 | perf_event__output_id_sample(event, &handle, &sample); | |
4471 | ||
78d613eb | 4472 | perf_output_end(&handle); |
c980d109 ACM |
4473 | out: |
4474 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
4475 | } |
4476 | ||
cdd6c482 | 4477 | static int perf_event_mmap_match(struct perf_event *event, |
3af9e859 EM |
4478 | struct perf_mmap_event *mmap_event, |
4479 | int executable) | |
0a4a9391 | 4480 | { |
6f93d0a7 | 4481 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4482 | return 0; |
4483 | ||
5632ab12 | 4484 | if (!event_filter_match(event)) |
5d27c23d PZ |
4485 | return 0; |
4486 | ||
3af9e859 EM |
4487 | if ((!executable && event->attr.mmap_data) || |
4488 | (executable && event->attr.mmap)) | |
0a4a9391 PZ |
4489 | return 1; |
4490 | ||
4491 | return 0; | |
4492 | } | |
4493 | ||
cdd6c482 | 4494 | static void perf_event_mmap_ctx(struct perf_event_context *ctx, |
3af9e859 EM |
4495 | struct perf_mmap_event *mmap_event, |
4496 | int executable) | |
0a4a9391 | 4497 | { |
cdd6c482 | 4498 | struct perf_event *event; |
0a4a9391 | 4499 | |
cdd6c482 | 4500 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
3af9e859 | 4501 | if (perf_event_mmap_match(event, mmap_event, executable)) |
cdd6c482 | 4502 | perf_event_mmap_output(event, mmap_event); |
0a4a9391 | 4503 | } |
0a4a9391 PZ |
4504 | } |
4505 | ||
cdd6c482 | 4506 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 PZ |
4507 | { |
4508 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4509 | struct perf_event_context *ctx; |
089dd79d PZ |
4510 | struct vm_area_struct *vma = mmap_event->vma; |
4511 | struct file *file = vma->vm_file; | |
0a4a9391 PZ |
4512 | unsigned int size; |
4513 | char tmp[16]; | |
4514 | char *buf = NULL; | |
089dd79d | 4515 | const char *name; |
108b02cf | 4516 | struct pmu *pmu; |
8dc85d54 | 4517 | int ctxn; |
0a4a9391 | 4518 | |
413ee3b4 AB |
4519 | memset(tmp, 0, sizeof(tmp)); |
4520 | ||
0a4a9391 | 4521 | if (file) { |
413ee3b4 | 4522 | /* |
76369139 | 4523 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
4524 | * need to add enough zero bytes after the string to handle |
4525 | * the 64bit alignment we do later. | |
4526 | */ | |
4527 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
4528 | if (!buf) { |
4529 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
4530 | goto got_name; | |
4531 | } | |
d3d21c41 | 4532 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
4533 | if (IS_ERR(name)) { |
4534 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
4535 | goto got_name; | |
4536 | } | |
4537 | } else { | |
413ee3b4 AB |
4538 | if (arch_vma_name(mmap_event->vma)) { |
4539 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
4540 | sizeof(tmp)); | |
089dd79d | 4541 | goto got_name; |
413ee3b4 | 4542 | } |
089dd79d PZ |
4543 | |
4544 | if (!vma->vm_mm) { | |
4545 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
4546 | goto got_name; | |
3af9e859 EM |
4547 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
4548 | vma->vm_end >= vma->vm_mm->brk) { | |
4549 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
4550 | goto got_name; | |
4551 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
4552 | vma->vm_end >= vma->vm_mm->start_stack) { | |
4553 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
4554 | goto got_name; | |
089dd79d PZ |
4555 | } |
4556 | ||
0a4a9391 PZ |
4557 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
4558 | goto got_name; | |
4559 | } | |
4560 | ||
4561 | got_name: | |
888fcee0 | 4562 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
4563 | |
4564 | mmap_event->file_name = name; | |
4565 | mmap_event->file_size = size; | |
4566 | ||
cdd6c482 | 4567 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 4568 | |
f6d9dd23 | 4569 | rcu_read_lock(); |
108b02cf | 4570 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4571 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4572 | if (cpuctx->active_pmu != pmu) |
4573 | goto next; | |
108b02cf PZ |
4574 | perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, |
4575 | vma->vm_flags & VM_EXEC); | |
8dc85d54 PZ |
4576 | |
4577 | ctxn = pmu->task_ctx_nr; | |
4578 | if (ctxn < 0) | |
41945f6c | 4579 | goto next; |
8dc85d54 PZ |
4580 | |
4581 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4582 | if (ctx) { | |
4583 | perf_event_mmap_ctx(ctx, mmap_event, | |
4584 | vma->vm_flags & VM_EXEC); | |
4585 | } | |
41945f6c PZ |
4586 | next: |
4587 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4588 | } |
665c2142 PZ |
4589 | rcu_read_unlock(); |
4590 | ||
0a4a9391 PZ |
4591 | kfree(buf); |
4592 | } | |
4593 | ||
3af9e859 | 4594 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 4595 | { |
9ee318a7 PZ |
4596 | struct perf_mmap_event mmap_event; |
4597 | ||
cdd6c482 | 4598 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
4599 | return; |
4600 | ||
4601 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 4602 | .vma = vma, |
573402db PZ |
4603 | /* .file_name */ |
4604 | /* .file_size */ | |
cdd6c482 | 4605 | .event_id = { |
573402db | 4606 | .header = { |
cdd6c482 | 4607 | .type = PERF_RECORD_MMAP, |
39447b38 | 4608 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
4609 | /* .size */ |
4610 | }, | |
4611 | /* .pid */ | |
4612 | /* .tid */ | |
089dd79d PZ |
4613 | .start = vma->vm_start, |
4614 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 4615 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 PZ |
4616 | }, |
4617 | }; | |
4618 | ||
cdd6c482 | 4619 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
4620 | } |
4621 | ||
a78ac325 PZ |
4622 | /* |
4623 | * IRQ throttle logging | |
4624 | */ | |
4625 | ||
cdd6c482 | 4626 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
4627 | { |
4628 | struct perf_output_handle handle; | |
c980d109 | 4629 | struct perf_sample_data sample; |
a78ac325 PZ |
4630 | int ret; |
4631 | ||
4632 | struct { | |
4633 | struct perf_event_header header; | |
4634 | u64 time; | |
cca3f454 | 4635 | u64 id; |
7f453c24 | 4636 | u64 stream_id; |
a78ac325 PZ |
4637 | } throttle_event = { |
4638 | .header = { | |
cdd6c482 | 4639 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
4640 | .misc = 0, |
4641 | .size = sizeof(throttle_event), | |
4642 | }, | |
def0a9b2 | 4643 | .time = perf_clock(), |
cdd6c482 IM |
4644 | .id = primary_event_id(event), |
4645 | .stream_id = event->id, | |
a78ac325 PZ |
4646 | }; |
4647 | ||
966ee4d6 | 4648 | if (enable) |
cdd6c482 | 4649 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 4650 | |
c980d109 ACM |
4651 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
4652 | ||
4653 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4654 | throttle_event.header.size); |
a78ac325 PZ |
4655 | if (ret) |
4656 | return; | |
4657 | ||
4658 | perf_output_put(&handle, throttle_event); | |
c980d109 | 4659 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
4660 | perf_output_end(&handle); |
4661 | } | |
4662 | ||
f6c7d5fe | 4663 | /* |
cdd6c482 | 4664 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
4665 | */ |
4666 | ||
a8b0ca17 | 4667 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4668 | int throttle, struct perf_sample_data *data, |
4669 | struct pt_regs *regs) | |
f6c7d5fe | 4670 | { |
cdd6c482 IM |
4671 | int events = atomic_read(&event->event_limit); |
4672 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 4673 | u64 seq; |
79f14641 PZ |
4674 | int ret = 0; |
4675 | ||
96398826 PZ |
4676 | /* |
4677 | * Non-sampling counters might still use the PMI to fold short | |
4678 | * hardware counters, ignore those. | |
4679 | */ | |
4680 | if (unlikely(!is_sampling_event(event))) | |
4681 | return 0; | |
4682 | ||
e050e3f0 SE |
4683 | seq = __this_cpu_read(perf_throttled_seq); |
4684 | if (seq != hwc->interrupts_seq) { | |
4685 | hwc->interrupts_seq = seq; | |
4686 | hwc->interrupts = 1; | |
4687 | } else { | |
4688 | hwc->interrupts++; | |
4689 | if (unlikely(throttle | |
4690 | && hwc->interrupts >= max_samples_per_tick)) { | |
4691 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
4692 | hwc->interrupts = MAX_INTERRUPTS; |
4693 | perf_log_throttle(event, 0); | |
a78ac325 PZ |
4694 | ret = 1; |
4695 | } | |
e050e3f0 | 4696 | } |
60db5e09 | 4697 | |
cdd6c482 | 4698 | if (event->attr.freq) { |
def0a9b2 | 4699 | u64 now = perf_clock(); |
abd50713 | 4700 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 4701 | |
abd50713 | 4702 | hwc->freq_time_stamp = now; |
bd2b5b12 | 4703 | |
abd50713 | 4704 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 4705 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
4706 | } |
4707 | ||
2023b359 PZ |
4708 | /* |
4709 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 4710 | * events |
2023b359 PZ |
4711 | */ |
4712 | ||
cdd6c482 IM |
4713 | event->pending_kill = POLL_IN; |
4714 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 4715 | ret = 1; |
cdd6c482 | 4716 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
4717 | event->pending_disable = 1; |
4718 | irq_work_queue(&event->pending); | |
79f14641 PZ |
4719 | } |
4720 | ||
453f19ee | 4721 | if (event->overflow_handler) |
a8b0ca17 | 4722 | event->overflow_handler(event, data, regs); |
453f19ee | 4723 | else |
a8b0ca17 | 4724 | perf_event_output(event, data, regs); |
453f19ee | 4725 | |
f506b3dc | 4726 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
4727 | event->pending_wakeup = 1; |
4728 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
4729 | } |
4730 | ||
79f14641 | 4731 | return ret; |
f6c7d5fe PZ |
4732 | } |
4733 | ||
a8b0ca17 | 4734 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4735 | struct perf_sample_data *data, |
4736 | struct pt_regs *regs) | |
850bc73f | 4737 | { |
a8b0ca17 | 4738 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
4739 | } |
4740 | ||
15dbf27c | 4741 | /* |
cdd6c482 | 4742 | * Generic software event infrastructure |
15dbf27c PZ |
4743 | */ |
4744 | ||
b28ab83c PZ |
4745 | struct swevent_htable { |
4746 | struct swevent_hlist *swevent_hlist; | |
4747 | struct mutex hlist_mutex; | |
4748 | int hlist_refcount; | |
4749 | ||
4750 | /* Recursion avoidance in each contexts */ | |
4751 | int recursion[PERF_NR_CONTEXTS]; | |
4752 | }; | |
4753 | ||
4754 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
4755 | ||
7b4b6658 | 4756 | /* |
cdd6c482 IM |
4757 | * We directly increment event->count and keep a second value in |
4758 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
4759 | * is kept in the range [-sample_period, 0] so that we can use the |
4760 | * sign as trigger. | |
4761 | */ | |
4762 | ||
cdd6c482 | 4763 | static u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 4764 | { |
cdd6c482 | 4765 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
4766 | u64 period = hwc->last_period; |
4767 | u64 nr, offset; | |
4768 | s64 old, val; | |
4769 | ||
4770 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
4771 | |
4772 | again: | |
e7850595 | 4773 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
4774 | if (val < 0) |
4775 | return 0; | |
15dbf27c | 4776 | |
7b4b6658 PZ |
4777 | nr = div64_u64(period + val, period); |
4778 | offset = nr * period; | |
4779 | val -= offset; | |
e7850595 | 4780 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 4781 | goto again; |
15dbf27c | 4782 | |
7b4b6658 | 4783 | return nr; |
15dbf27c PZ |
4784 | } |
4785 | ||
0cff784a | 4786 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 4787 | struct perf_sample_data *data, |
5622f295 | 4788 | struct pt_regs *regs) |
15dbf27c | 4789 | { |
cdd6c482 | 4790 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 4791 | int throttle = 0; |
15dbf27c | 4792 | |
0cff784a PZ |
4793 | if (!overflow) |
4794 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 4795 | |
7b4b6658 PZ |
4796 | if (hwc->interrupts == MAX_INTERRUPTS) |
4797 | return; | |
15dbf27c | 4798 | |
7b4b6658 | 4799 | for (; overflow; overflow--) { |
a8b0ca17 | 4800 | if (__perf_event_overflow(event, throttle, |
5622f295 | 4801 | data, regs)) { |
7b4b6658 PZ |
4802 | /* |
4803 | * We inhibit the overflow from happening when | |
4804 | * hwc->interrupts == MAX_INTERRUPTS. | |
4805 | */ | |
4806 | break; | |
4807 | } | |
cf450a73 | 4808 | throttle = 1; |
7b4b6658 | 4809 | } |
15dbf27c PZ |
4810 | } |
4811 | ||
a4eaf7f1 | 4812 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 4813 | struct perf_sample_data *data, |
5622f295 | 4814 | struct pt_regs *regs) |
7b4b6658 | 4815 | { |
cdd6c482 | 4816 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 4817 | |
e7850595 | 4818 | local64_add(nr, &event->count); |
d6d020e9 | 4819 | |
0cff784a PZ |
4820 | if (!regs) |
4821 | return; | |
4822 | ||
6c7e550f | 4823 | if (!is_sampling_event(event)) |
7b4b6658 | 4824 | return; |
d6d020e9 | 4825 | |
5d81e5cf AV |
4826 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
4827 | data->period = nr; | |
4828 | return perf_swevent_overflow(event, 1, data, regs); | |
4829 | } else | |
4830 | data->period = event->hw.last_period; | |
4831 | ||
0cff784a | 4832 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 4833 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 4834 | |
e7850595 | 4835 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 4836 | return; |
df1a132b | 4837 | |
a8b0ca17 | 4838 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
4839 | } |
4840 | ||
f5ffe02e FW |
4841 | static int perf_exclude_event(struct perf_event *event, |
4842 | struct pt_regs *regs) | |
4843 | { | |
a4eaf7f1 | 4844 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 4845 | return 1; |
a4eaf7f1 | 4846 | |
f5ffe02e FW |
4847 | if (regs) { |
4848 | if (event->attr.exclude_user && user_mode(regs)) | |
4849 | return 1; | |
4850 | ||
4851 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
4852 | return 1; | |
4853 | } | |
4854 | ||
4855 | return 0; | |
4856 | } | |
4857 | ||
cdd6c482 | 4858 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 4859 | enum perf_type_id type, |
6fb2915d LZ |
4860 | u32 event_id, |
4861 | struct perf_sample_data *data, | |
4862 | struct pt_regs *regs) | |
15dbf27c | 4863 | { |
cdd6c482 | 4864 | if (event->attr.type != type) |
a21ca2ca | 4865 | return 0; |
f5ffe02e | 4866 | |
cdd6c482 | 4867 | if (event->attr.config != event_id) |
15dbf27c PZ |
4868 | return 0; |
4869 | ||
f5ffe02e FW |
4870 | if (perf_exclude_event(event, regs)) |
4871 | return 0; | |
15dbf27c PZ |
4872 | |
4873 | return 1; | |
4874 | } | |
4875 | ||
76e1d904 FW |
4876 | static inline u64 swevent_hash(u64 type, u32 event_id) |
4877 | { | |
4878 | u64 val = event_id | (type << 32); | |
4879 | ||
4880 | return hash_64(val, SWEVENT_HLIST_BITS); | |
4881 | } | |
4882 | ||
49f135ed FW |
4883 | static inline struct hlist_head * |
4884 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 4885 | { |
49f135ed FW |
4886 | u64 hash = swevent_hash(type, event_id); |
4887 | ||
4888 | return &hlist->heads[hash]; | |
4889 | } | |
76e1d904 | 4890 | |
49f135ed FW |
4891 | /* For the read side: events when they trigger */ |
4892 | static inline struct hlist_head * | |
b28ab83c | 4893 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
4894 | { |
4895 | struct swevent_hlist *hlist; | |
76e1d904 | 4896 | |
b28ab83c | 4897 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
4898 | if (!hlist) |
4899 | return NULL; | |
4900 | ||
49f135ed FW |
4901 | return __find_swevent_head(hlist, type, event_id); |
4902 | } | |
4903 | ||
4904 | /* For the event head insertion and removal in the hlist */ | |
4905 | static inline struct hlist_head * | |
b28ab83c | 4906 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
4907 | { |
4908 | struct swevent_hlist *hlist; | |
4909 | u32 event_id = event->attr.config; | |
4910 | u64 type = event->attr.type; | |
4911 | ||
4912 | /* | |
4913 | * Event scheduling is always serialized against hlist allocation | |
4914 | * and release. Which makes the protected version suitable here. | |
4915 | * The context lock guarantees that. | |
4916 | */ | |
b28ab83c | 4917 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
4918 | lockdep_is_held(&event->ctx->lock)); |
4919 | if (!hlist) | |
4920 | return NULL; | |
4921 | ||
4922 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
4923 | } |
4924 | ||
4925 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 4926 | u64 nr, |
76e1d904 FW |
4927 | struct perf_sample_data *data, |
4928 | struct pt_regs *regs) | |
15dbf27c | 4929 | { |
b28ab83c | 4930 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 4931 | struct perf_event *event; |
76e1d904 FW |
4932 | struct hlist_node *node; |
4933 | struct hlist_head *head; | |
15dbf27c | 4934 | |
76e1d904 | 4935 | rcu_read_lock(); |
b28ab83c | 4936 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
4937 | if (!head) |
4938 | goto end; | |
4939 | ||
4940 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { | |
6fb2915d | 4941 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 4942 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 4943 | } |
76e1d904 FW |
4944 | end: |
4945 | rcu_read_unlock(); | |
15dbf27c PZ |
4946 | } |
4947 | ||
4ed7c92d | 4948 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 4949 | { |
b28ab83c | 4950 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 4951 | |
b28ab83c | 4952 | return get_recursion_context(swhash->recursion); |
96f6d444 | 4953 | } |
645e8cc0 | 4954 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 4955 | |
fa9f90be | 4956 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 4957 | { |
b28ab83c | 4958 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 4959 | |
b28ab83c | 4960 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 4961 | } |
15dbf27c | 4962 | |
a8b0ca17 | 4963 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 4964 | { |
a4234bfc | 4965 | struct perf_sample_data data; |
4ed7c92d PZ |
4966 | int rctx; |
4967 | ||
1c024eca | 4968 | preempt_disable_notrace(); |
4ed7c92d PZ |
4969 | rctx = perf_swevent_get_recursion_context(); |
4970 | if (rctx < 0) | |
4971 | return; | |
a4234bfc | 4972 | |
fd0d000b | 4973 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 4974 | |
a8b0ca17 | 4975 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
4976 | |
4977 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 4978 | preempt_enable_notrace(); |
b8e83514 PZ |
4979 | } |
4980 | ||
cdd6c482 | 4981 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 4982 | { |
15dbf27c PZ |
4983 | } |
4984 | ||
a4eaf7f1 | 4985 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 4986 | { |
b28ab83c | 4987 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 4988 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
4989 | struct hlist_head *head; |
4990 | ||
6c7e550f | 4991 | if (is_sampling_event(event)) { |
7b4b6658 | 4992 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 4993 | perf_swevent_set_period(event); |
7b4b6658 | 4994 | } |
76e1d904 | 4995 | |
a4eaf7f1 PZ |
4996 | hwc->state = !(flags & PERF_EF_START); |
4997 | ||
b28ab83c | 4998 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
4999 | if (WARN_ON_ONCE(!head)) |
5000 | return -EINVAL; | |
5001 | ||
5002 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5003 | ||
15dbf27c PZ |
5004 | return 0; |
5005 | } | |
5006 | ||
a4eaf7f1 | 5007 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5008 | { |
76e1d904 | 5009 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5010 | } |
5011 | ||
a4eaf7f1 | 5012 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5013 | { |
a4eaf7f1 | 5014 | event->hw.state = 0; |
d6d020e9 | 5015 | } |
aa9c4c0f | 5016 | |
a4eaf7f1 | 5017 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5018 | { |
a4eaf7f1 | 5019 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5020 | } |
5021 | ||
49f135ed FW |
5022 | /* Deref the hlist from the update side */ |
5023 | static inline struct swevent_hlist * | |
b28ab83c | 5024 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5025 | { |
b28ab83c PZ |
5026 | return rcu_dereference_protected(swhash->swevent_hlist, |
5027 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5028 | } |
5029 | ||
b28ab83c | 5030 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5031 | { |
b28ab83c | 5032 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5033 | |
49f135ed | 5034 | if (!hlist) |
76e1d904 FW |
5035 | return; |
5036 | ||
b28ab83c | 5037 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5038 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5039 | } |
5040 | ||
5041 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5042 | { | |
b28ab83c | 5043 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5044 | |
b28ab83c | 5045 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5046 | |
b28ab83c PZ |
5047 | if (!--swhash->hlist_refcount) |
5048 | swevent_hlist_release(swhash); | |
76e1d904 | 5049 | |
b28ab83c | 5050 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5051 | } |
5052 | ||
5053 | static void swevent_hlist_put(struct perf_event *event) | |
5054 | { | |
5055 | int cpu; | |
5056 | ||
5057 | if (event->cpu != -1) { | |
5058 | swevent_hlist_put_cpu(event, event->cpu); | |
5059 | return; | |
5060 | } | |
5061 | ||
5062 | for_each_possible_cpu(cpu) | |
5063 | swevent_hlist_put_cpu(event, cpu); | |
5064 | } | |
5065 | ||
5066 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5067 | { | |
b28ab83c | 5068 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5069 | int err = 0; |
5070 | ||
b28ab83c | 5071 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5072 | |
b28ab83c | 5073 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5074 | struct swevent_hlist *hlist; |
5075 | ||
5076 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5077 | if (!hlist) { | |
5078 | err = -ENOMEM; | |
5079 | goto exit; | |
5080 | } | |
b28ab83c | 5081 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5082 | } |
b28ab83c | 5083 | swhash->hlist_refcount++; |
9ed6060d | 5084 | exit: |
b28ab83c | 5085 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5086 | |
5087 | return err; | |
5088 | } | |
5089 | ||
5090 | static int swevent_hlist_get(struct perf_event *event) | |
5091 | { | |
5092 | int err; | |
5093 | int cpu, failed_cpu; | |
5094 | ||
5095 | if (event->cpu != -1) | |
5096 | return swevent_hlist_get_cpu(event, event->cpu); | |
5097 | ||
5098 | get_online_cpus(); | |
5099 | for_each_possible_cpu(cpu) { | |
5100 | err = swevent_hlist_get_cpu(event, cpu); | |
5101 | if (err) { | |
5102 | failed_cpu = cpu; | |
5103 | goto fail; | |
5104 | } | |
5105 | } | |
5106 | put_online_cpus(); | |
5107 | ||
5108 | return 0; | |
9ed6060d | 5109 | fail: |
76e1d904 FW |
5110 | for_each_possible_cpu(cpu) { |
5111 | if (cpu == failed_cpu) | |
5112 | break; | |
5113 | swevent_hlist_put_cpu(event, cpu); | |
5114 | } | |
5115 | ||
5116 | put_online_cpus(); | |
5117 | return err; | |
5118 | } | |
5119 | ||
c5905afb | 5120 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5121 | |
b0a873eb PZ |
5122 | static void sw_perf_event_destroy(struct perf_event *event) |
5123 | { | |
5124 | u64 event_id = event->attr.config; | |
95476b64 | 5125 | |
b0a873eb PZ |
5126 | WARN_ON(event->parent); |
5127 | ||
c5905afb | 5128 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5129 | swevent_hlist_put(event); |
5130 | } | |
5131 | ||
5132 | static int perf_swevent_init(struct perf_event *event) | |
5133 | { | |
5134 | int event_id = event->attr.config; | |
5135 | ||
5136 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5137 | return -ENOENT; | |
5138 | ||
2481c5fa SE |
5139 | /* |
5140 | * no branch sampling for software events | |
5141 | */ | |
5142 | if (has_branch_stack(event)) | |
5143 | return -EOPNOTSUPP; | |
5144 | ||
b0a873eb PZ |
5145 | switch (event_id) { |
5146 | case PERF_COUNT_SW_CPU_CLOCK: | |
5147 | case PERF_COUNT_SW_TASK_CLOCK: | |
5148 | return -ENOENT; | |
5149 | ||
5150 | default: | |
5151 | break; | |
5152 | } | |
5153 | ||
ce677831 | 5154 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5155 | return -ENOENT; |
5156 | ||
5157 | if (!event->parent) { | |
5158 | int err; | |
5159 | ||
5160 | err = swevent_hlist_get(event); | |
5161 | if (err) | |
5162 | return err; | |
5163 | ||
c5905afb | 5164 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5165 | event->destroy = sw_perf_event_destroy; |
5166 | } | |
5167 | ||
5168 | return 0; | |
5169 | } | |
5170 | ||
35edc2a5 PZ |
5171 | static int perf_swevent_event_idx(struct perf_event *event) |
5172 | { | |
5173 | return 0; | |
5174 | } | |
5175 | ||
b0a873eb | 5176 | static struct pmu perf_swevent = { |
89a1e187 | 5177 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5178 | |
b0a873eb | 5179 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5180 | .add = perf_swevent_add, |
5181 | .del = perf_swevent_del, | |
5182 | .start = perf_swevent_start, | |
5183 | .stop = perf_swevent_stop, | |
1c024eca | 5184 | .read = perf_swevent_read, |
35edc2a5 PZ |
5185 | |
5186 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5187 | }; |
5188 | ||
b0a873eb PZ |
5189 | #ifdef CONFIG_EVENT_TRACING |
5190 | ||
1c024eca PZ |
5191 | static int perf_tp_filter_match(struct perf_event *event, |
5192 | struct perf_sample_data *data) | |
5193 | { | |
5194 | void *record = data->raw->data; | |
5195 | ||
5196 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5197 | return 1; | |
5198 | return 0; | |
5199 | } | |
5200 | ||
5201 | static int perf_tp_event_match(struct perf_event *event, | |
5202 | struct perf_sample_data *data, | |
5203 | struct pt_regs *regs) | |
5204 | { | |
a0f7d0f7 FW |
5205 | if (event->hw.state & PERF_HES_STOPPED) |
5206 | return 0; | |
580d607c PZ |
5207 | /* |
5208 | * All tracepoints are from kernel-space. | |
5209 | */ | |
5210 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5211 | return 0; |
5212 | ||
5213 | if (!perf_tp_filter_match(event, data)) | |
5214 | return 0; | |
5215 | ||
5216 | return 1; | |
5217 | } | |
5218 | ||
5219 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5220 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5221 | struct task_struct *task) | |
95476b64 FW |
5222 | { |
5223 | struct perf_sample_data data; | |
1c024eca PZ |
5224 | struct perf_event *event; |
5225 | struct hlist_node *node; | |
5226 | ||
95476b64 FW |
5227 | struct perf_raw_record raw = { |
5228 | .size = entry_size, | |
5229 | .data = record, | |
5230 | }; | |
5231 | ||
fd0d000b | 5232 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5233 | data.raw = &raw; |
5234 | ||
1c024eca PZ |
5235 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { |
5236 | if (perf_tp_event_match(event, &data, regs)) | |
a8b0ca17 | 5237 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5238 | } |
ecc55f84 | 5239 | |
e6dab5ff AV |
5240 | /* |
5241 | * If we got specified a target task, also iterate its context and | |
5242 | * deliver this event there too. | |
5243 | */ | |
5244 | if (task && task != current) { | |
5245 | struct perf_event_context *ctx; | |
5246 | struct trace_entry *entry = record; | |
5247 | ||
5248 | rcu_read_lock(); | |
5249 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5250 | if (!ctx) | |
5251 | goto unlock; | |
5252 | ||
5253 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5254 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5255 | continue; | |
5256 | if (event->attr.config != entry->type) | |
5257 | continue; | |
5258 | if (perf_tp_event_match(event, &data, regs)) | |
5259 | perf_swevent_event(event, count, &data, regs); | |
5260 | } | |
5261 | unlock: | |
5262 | rcu_read_unlock(); | |
5263 | } | |
5264 | ||
ecc55f84 | 5265 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5266 | } |
5267 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5268 | ||
cdd6c482 | 5269 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5270 | { |
1c024eca | 5271 | perf_trace_destroy(event); |
e077df4f PZ |
5272 | } |
5273 | ||
b0a873eb | 5274 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5275 | { |
76e1d904 FW |
5276 | int err; |
5277 | ||
b0a873eb PZ |
5278 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5279 | return -ENOENT; | |
5280 | ||
2481c5fa SE |
5281 | /* |
5282 | * no branch sampling for tracepoint events | |
5283 | */ | |
5284 | if (has_branch_stack(event)) | |
5285 | return -EOPNOTSUPP; | |
5286 | ||
1c024eca PZ |
5287 | err = perf_trace_init(event); |
5288 | if (err) | |
b0a873eb | 5289 | return err; |
e077df4f | 5290 | |
cdd6c482 | 5291 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5292 | |
b0a873eb PZ |
5293 | return 0; |
5294 | } | |
5295 | ||
5296 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5297 | .task_ctx_nr = perf_sw_context, |
5298 | ||
b0a873eb | 5299 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5300 | .add = perf_trace_add, |
5301 | .del = perf_trace_del, | |
5302 | .start = perf_swevent_start, | |
5303 | .stop = perf_swevent_stop, | |
b0a873eb | 5304 | .read = perf_swevent_read, |
35edc2a5 PZ |
5305 | |
5306 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5307 | }; |
5308 | ||
5309 | static inline void perf_tp_register(void) | |
5310 | { | |
2e80a82a | 5311 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5312 | } |
6fb2915d LZ |
5313 | |
5314 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5315 | { | |
5316 | char *filter_str; | |
5317 | int ret; | |
5318 | ||
5319 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5320 | return -EINVAL; | |
5321 | ||
5322 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5323 | if (IS_ERR(filter_str)) | |
5324 | return PTR_ERR(filter_str); | |
5325 | ||
5326 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5327 | ||
5328 | kfree(filter_str); | |
5329 | return ret; | |
5330 | } | |
5331 | ||
5332 | static void perf_event_free_filter(struct perf_event *event) | |
5333 | { | |
5334 | ftrace_profile_free_filter(event); | |
5335 | } | |
5336 | ||
e077df4f | 5337 | #else |
6fb2915d | 5338 | |
b0a873eb | 5339 | static inline void perf_tp_register(void) |
e077df4f | 5340 | { |
e077df4f | 5341 | } |
6fb2915d LZ |
5342 | |
5343 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5344 | { | |
5345 | return -ENOENT; | |
5346 | } | |
5347 | ||
5348 | static void perf_event_free_filter(struct perf_event *event) | |
5349 | { | |
5350 | } | |
5351 | ||
07b139c8 | 5352 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5353 | |
24f1e32c | 5354 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5355 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5356 | { |
f5ffe02e FW |
5357 | struct perf_sample_data sample; |
5358 | struct pt_regs *regs = data; | |
5359 | ||
fd0d000b | 5360 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 5361 | |
a4eaf7f1 | 5362 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5363 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5364 | } |
5365 | #endif | |
5366 | ||
b0a873eb PZ |
5367 | /* |
5368 | * hrtimer based swevent callback | |
5369 | */ | |
f29ac756 | 5370 | |
b0a873eb | 5371 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5372 | { |
b0a873eb PZ |
5373 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5374 | struct perf_sample_data data; | |
5375 | struct pt_regs *regs; | |
5376 | struct perf_event *event; | |
5377 | u64 period; | |
f29ac756 | 5378 | |
b0a873eb | 5379 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5380 | |
5381 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5382 | return HRTIMER_NORESTART; | |
5383 | ||
b0a873eb | 5384 | event->pmu->read(event); |
f344011c | 5385 | |
fd0d000b | 5386 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
5387 | regs = get_irq_regs(); |
5388 | ||
5389 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5390 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 5391 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
5392 | ret = HRTIMER_NORESTART; |
5393 | } | |
24f1e32c | 5394 | |
b0a873eb PZ |
5395 | period = max_t(u64, 10000, event->hw.sample_period); |
5396 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5397 | |
b0a873eb | 5398 | return ret; |
f29ac756 PZ |
5399 | } |
5400 | ||
b0a873eb | 5401 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5402 | { |
b0a873eb | 5403 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5404 | s64 period; |
5405 | ||
5406 | if (!is_sampling_event(event)) | |
5407 | return; | |
f5ffe02e | 5408 | |
5d508e82 FBH |
5409 | period = local64_read(&hwc->period_left); |
5410 | if (period) { | |
5411 | if (period < 0) | |
5412 | period = 10000; | |
fa407f35 | 5413 | |
5d508e82 FBH |
5414 | local64_set(&hwc->period_left, 0); |
5415 | } else { | |
5416 | period = max_t(u64, 10000, hwc->sample_period); | |
5417 | } | |
5418 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 5419 | ns_to_ktime(period), 0, |
b5ab4cd5 | 5420 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 5421 | } |
b0a873eb PZ |
5422 | |
5423 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 5424 | { |
b0a873eb PZ |
5425 | struct hw_perf_event *hwc = &event->hw; |
5426 | ||
6c7e550f | 5427 | if (is_sampling_event(event)) { |
b0a873eb | 5428 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 5429 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
5430 | |
5431 | hrtimer_cancel(&hwc->hrtimer); | |
5432 | } | |
24f1e32c FW |
5433 | } |
5434 | ||
ba3dd36c PZ |
5435 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5436 | { | |
5437 | struct hw_perf_event *hwc = &event->hw; | |
5438 | ||
5439 | if (!is_sampling_event(event)) | |
5440 | return; | |
5441 | ||
5442 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5443 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
5444 | ||
5445 | /* | |
5446 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
5447 | * mapping and avoid the whole period adjust feedback stuff. | |
5448 | */ | |
5449 | if (event->attr.freq) { | |
5450 | long freq = event->attr.sample_freq; | |
5451 | ||
5452 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
5453 | hwc->sample_period = event->attr.sample_period; | |
5454 | local64_set(&hwc->period_left, hwc->sample_period); | |
5455 | event->attr.freq = 0; | |
5456 | } | |
5457 | } | |
5458 | ||
b0a873eb PZ |
5459 | /* |
5460 | * Software event: cpu wall time clock | |
5461 | */ | |
5462 | ||
5463 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 5464 | { |
b0a873eb PZ |
5465 | s64 prev; |
5466 | u64 now; | |
5467 | ||
a4eaf7f1 | 5468 | now = local_clock(); |
b0a873eb PZ |
5469 | prev = local64_xchg(&event->hw.prev_count, now); |
5470 | local64_add(now - prev, &event->count); | |
24f1e32c | 5471 | } |
24f1e32c | 5472 | |
a4eaf7f1 | 5473 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5474 | { |
a4eaf7f1 | 5475 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 5476 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5477 | } |
5478 | ||
a4eaf7f1 | 5479 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 5480 | { |
b0a873eb PZ |
5481 | perf_swevent_cancel_hrtimer(event); |
5482 | cpu_clock_event_update(event); | |
5483 | } | |
f29ac756 | 5484 | |
a4eaf7f1 PZ |
5485 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
5486 | { | |
5487 | if (flags & PERF_EF_START) | |
5488 | cpu_clock_event_start(event, flags); | |
5489 | ||
5490 | return 0; | |
5491 | } | |
5492 | ||
5493 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
5494 | { | |
5495 | cpu_clock_event_stop(event, flags); | |
5496 | } | |
5497 | ||
b0a873eb PZ |
5498 | static void cpu_clock_event_read(struct perf_event *event) |
5499 | { | |
5500 | cpu_clock_event_update(event); | |
5501 | } | |
f344011c | 5502 | |
b0a873eb PZ |
5503 | static int cpu_clock_event_init(struct perf_event *event) |
5504 | { | |
5505 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5506 | return -ENOENT; | |
5507 | ||
5508 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
5509 | return -ENOENT; | |
5510 | ||
2481c5fa SE |
5511 | /* |
5512 | * no branch sampling for software events | |
5513 | */ | |
5514 | if (has_branch_stack(event)) | |
5515 | return -EOPNOTSUPP; | |
5516 | ||
ba3dd36c PZ |
5517 | perf_swevent_init_hrtimer(event); |
5518 | ||
b0a873eb | 5519 | return 0; |
f29ac756 PZ |
5520 | } |
5521 | ||
b0a873eb | 5522 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
5523 | .task_ctx_nr = perf_sw_context, |
5524 | ||
b0a873eb | 5525 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
5526 | .add = cpu_clock_event_add, |
5527 | .del = cpu_clock_event_del, | |
5528 | .start = cpu_clock_event_start, | |
5529 | .stop = cpu_clock_event_stop, | |
b0a873eb | 5530 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
5531 | |
5532 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5533 | }; |
5534 | ||
5535 | /* | |
5536 | * Software event: task time clock | |
5537 | */ | |
5538 | ||
5539 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 5540 | { |
b0a873eb PZ |
5541 | u64 prev; |
5542 | s64 delta; | |
5c92d124 | 5543 | |
b0a873eb PZ |
5544 | prev = local64_xchg(&event->hw.prev_count, now); |
5545 | delta = now - prev; | |
5546 | local64_add(delta, &event->count); | |
5547 | } | |
5c92d124 | 5548 | |
a4eaf7f1 | 5549 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5550 | { |
a4eaf7f1 | 5551 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 5552 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5553 | } |
5554 | ||
a4eaf7f1 | 5555 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
5556 | { |
5557 | perf_swevent_cancel_hrtimer(event); | |
5558 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
5559 | } |
5560 | ||
5561 | static int task_clock_event_add(struct perf_event *event, int flags) | |
5562 | { | |
5563 | if (flags & PERF_EF_START) | |
5564 | task_clock_event_start(event, flags); | |
b0a873eb | 5565 | |
a4eaf7f1 PZ |
5566 | return 0; |
5567 | } | |
5568 | ||
5569 | static void task_clock_event_del(struct perf_event *event, int flags) | |
5570 | { | |
5571 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
5572 | } |
5573 | ||
5574 | static void task_clock_event_read(struct perf_event *event) | |
5575 | { | |
768a06e2 PZ |
5576 | u64 now = perf_clock(); |
5577 | u64 delta = now - event->ctx->timestamp; | |
5578 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
5579 | |
5580 | task_clock_event_update(event, time); | |
5581 | } | |
5582 | ||
5583 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 5584 | { |
b0a873eb PZ |
5585 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5586 | return -ENOENT; | |
5587 | ||
5588 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
5589 | return -ENOENT; | |
5590 | ||
2481c5fa SE |
5591 | /* |
5592 | * no branch sampling for software events | |
5593 | */ | |
5594 | if (has_branch_stack(event)) | |
5595 | return -EOPNOTSUPP; | |
5596 | ||
ba3dd36c PZ |
5597 | perf_swevent_init_hrtimer(event); |
5598 | ||
b0a873eb | 5599 | return 0; |
6fb2915d LZ |
5600 | } |
5601 | ||
b0a873eb | 5602 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
5603 | .task_ctx_nr = perf_sw_context, |
5604 | ||
b0a873eb | 5605 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
5606 | .add = task_clock_event_add, |
5607 | .del = task_clock_event_del, | |
5608 | .start = task_clock_event_start, | |
5609 | .stop = task_clock_event_stop, | |
b0a873eb | 5610 | .read = task_clock_event_read, |
35edc2a5 PZ |
5611 | |
5612 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 5613 | }; |
6fb2915d | 5614 | |
ad5133b7 | 5615 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 5616 | { |
e077df4f | 5617 | } |
6fb2915d | 5618 | |
ad5133b7 | 5619 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 5620 | { |
ad5133b7 | 5621 | return 0; |
6fb2915d LZ |
5622 | } |
5623 | ||
ad5133b7 | 5624 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 5625 | { |
ad5133b7 | 5626 | perf_pmu_disable(pmu); |
6fb2915d LZ |
5627 | } |
5628 | ||
ad5133b7 PZ |
5629 | static int perf_pmu_commit_txn(struct pmu *pmu) |
5630 | { | |
5631 | perf_pmu_enable(pmu); | |
5632 | return 0; | |
5633 | } | |
e077df4f | 5634 | |
ad5133b7 | 5635 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 5636 | { |
ad5133b7 | 5637 | perf_pmu_enable(pmu); |
24f1e32c FW |
5638 | } |
5639 | ||
35edc2a5 PZ |
5640 | static int perf_event_idx_default(struct perf_event *event) |
5641 | { | |
5642 | return event->hw.idx + 1; | |
5643 | } | |
5644 | ||
8dc85d54 PZ |
5645 | /* |
5646 | * Ensures all contexts with the same task_ctx_nr have the same | |
5647 | * pmu_cpu_context too. | |
5648 | */ | |
5649 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 5650 | { |
8dc85d54 | 5651 | struct pmu *pmu; |
b326e956 | 5652 | |
8dc85d54 PZ |
5653 | if (ctxn < 0) |
5654 | return NULL; | |
24f1e32c | 5655 | |
8dc85d54 PZ |
5656 | list_for_each_entry(pmu, &pmus, entry) { |
5657 | if (pmu->task_ctx_nr == ctxn) | |
5658 | return pmu->pmu_cpu_context; | |
5659 | } | |
24f1e32c | 5660 | |
8dc85d54 | 5661 | return NULL; |
24f1e32c FW |
5662 | } |
5663 | ||
51676957 | 5664 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 5665 | { |
51676957 PZ |
5666 | int cpu; |
5667 | ||
5668 | for_each_possible_cpu(cpu) { | |
5669 | struct perf_cpu_context *cpuctx; | |
5670 | ||
5671 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
5672 | ||
5673 | if (cpuctx->active_pmu == old_pmu) | |
5674 | cpuctx->active_pmu = pmu; | |
5675 | } | |
5676 | } | |
5677 | ||
5678 | static void free_pmu_context(struct pmu *pmu) | |
5679 | { | |
5680 | struct pmu *i; | |
f5ffe02e | 5681 | |
8dc85d54 | 5682 | mutex_lock(&pmus_lock); |
0475f9ea | 5683 | /* |
8dc85d54 | 5684 | * Like a real lame refcount. |
0475f9ea | 5685 | */ |
51676957 PZ |
5686 | list_for_each_entry(i, &pmus, entry) { |
5687 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
5688 | update_pmu_context(i, pmu); | |
8dc85d54 | 5689 | goto out; |
51676957 | 5690 | } |
8dc85d54 | 5691 | } |
d6d020e9 | 5692 | |
51676957 | 5693 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
5694 | out: |
5695 | mutex_unlock(&pmus_lock); | |
24f1e32c | 5696 | } |
2e80a82a | 5697 | static struct idr pmu_idr; |
d6d020e9 | 5698 | |
abe43400 PZ |
5699 | static ssize_t |
5700 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
5701 | { | |
5702 | struct pmu *pmu = dev_get_drvdata(dev); | |
5703 | ||
5704 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
5705 | } | |
5706 | ||
5707 | static struct device_attribute pmu_dev_attrs[] = { | |
5708 | __ATTR_RO(type), | |
5709 | __ATTR_NULL, | |
5710 | }; | |
5711 | ||
5712 | static int pmu_bus_running; | |
5713 | static struct bus_type pmu_bus = { | |
5714 | .name = "event_source", | |
5715 | .dev_attrs = pmu_dev_attrs, | |
5716 | }; | |
5717 | ||
5718 | static void pmu_dev_release(struct device *dev) | |
5719 | { | |
5720 | kfree(dev); | |
5721 | } | |
5722 | ||
5723 | static int pmu_dev_alloc(struct pmu *pmu) | |
5724 | { | |
5725 | int ret = -ENOMEM; | |
5726 | ||
5727 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
5728 | if (!pmu->dev) | |
5729 | goto out; | |
5730 | ||
0c9d42ed | 5731 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
5732 | device_initialize(pmu->dev); |
5733 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
5734 | if (ret) | |
5735 | goto free_dev; | |
5736 | ||
5737 | dev_set_drvdata(pmu->dev, pmu); | |
5738 | pmu->dev->bus = &pmu_bus; | |
5739 | pmu->dev->release = pmu_dev_release; | |
5740 | ret = device_add(pmu->dev); | |
5741 | if (ret) | |
5742 | goto free_dev; | |
5743 | ||
5744 | out: | |
5745 | return ret; | |
5746 | ||
5747 | free_dev: | |
5748 | put_device(pmu->dev); | |
5749 | goto out; | |
5750 | } | |
5751 | ||
547e9fd7 | 5752 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 5753 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 5754 | |
2e80a82a | 5755 | int perf_pmu_register(struct pmu *pmu, char *name, int type) |
24f1e32c | 5756 | { |
108b02cf | 5757 | int cpu, ret; |
24f1e32c | 5758 | |
b0a873eb | 5759 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
5760 | ret = -ENOMEM; |
5761 | pmu->pmu_disable_count = alloc_percpu(int); | |
5762 | if (!pmu->pmu_disable_count) | |
5763 | goto unlock; | |
f29ac756 | 5764 | |
2e80a82a PZ |
5765 | pmu->type = -1; |
5766 | if (!name) | |
5767 | goto skip_type; | |
5768 | pmu->name = name; | |
5769 | ||
5770 | if (type < 0) { | |
5771 | int err = idr_pre_get(&pmu_idr, GFP_KERNEL); | |
5772 | if (!err) | |
5773 | goto free_pdc; | |
5774 | ||
5775 | err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type); | |
5776 | if (err) { | |
5777 | ret = err; | |
5778 | goto free_pdc; | |
5779 | } | |
5780 | } | |
5781 | pmu->type = type; | |
5782 | ||
abe43400 PZ |
5783 | if (pmu_bus_running) { |
5784 | ret = pmu_dev_alloc(pmu); | |
5785 | if (ret) | |
5786 | goto free_idr; | |
5787 | } | |
5788 | ||
2e80a82a | 5789 | skip_type: |
8dc85d54 PZ |
5790 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
5791 | if (pmu->pmu_cpu_context) | |
5792 | goto got_cpu_context; | |
f29ac756 | 5793 | |
108b02cf PZ |
5794 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
5795 | if (!pmu->pmu_cpu_context) | |
abe43400 | 5796 | goto free_dev; |
f344011c | 5797 | |
108b02cf PZ |
5798 | for_each_possible_cpu(cpu) { |
5799 | struct perf_cpu_context *cpuctx; | |
5800 | ||
5801 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 5802 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 5803 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 5804 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 5805 | cpuctx->ctx.type = cpu_context; |
108b02cf | 5806 | cpuctx->ctx.pmu = pmu; |
e9d2b064 PZ |
5807 | cpuctx->jiffies_interval = 1; |
5808 | INIT_LIST_HEAD(&cpuctx->rotation_list); | |
51676957 | 5809 | cpuctx->active_pmu = pmu; |
108b02cf | 5810 | } |
76e1d904 | 5811 | |
8dc85d54 | 5812 | got_cpu_context: |
ad5133b7 PZ |
5813 | if (!pmu->start_txn) { |
5814 | if (pmu->pmu_enable) { | |
5815 | /* | |
5816 | * If we have pmu_enable/pmu_disable calls, install | |
5817 | * transaction stubs that use that to try and batch | |
5818 | * hardware accesses. | |
5819 | */ | |
5820 | pmu->start_txn = perf_pmu_start_txn; | |
5821 | pmu->commit_txn = perf_pmu_commit_txn; | |
5822 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
5823 | } else { | |
5824 | pmu->start_txn = perf_pmu_nop_void; | |
5825 | pmu->commit_txn = perf_pmu_nop_int; | |
5826 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 5827 | } |
5c92d124 | 5828 | } |
15dbf27c | 5829 | |
ad5133b7 PZ |
5830 | if (!pmu->pmu_enable) { |
5831 | pmu->pmu_enable = perf_pmu_nop_void; | |
5832 | pmu->pmu_disable = perf_pmu_nop_void; | |
5833 | } | |
5834 | ||
35edc2a5 PZ |
5835 | if (!pmu->event_idx) |
5836 | pmu->event_idx = perf_event_idx_default; | |
5837 | ||
b0a873eb | 5838 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
5839 | ret = 0; |
5840 | unlock: | |
b0a873eb PZ |
5841 | mutex_unlock(&pmus_lock); |
5842 | ||
33696fc0 | 5843 | return ret; |
108b02cf | 5844 | |
abe43400 PZ |
5845 | free_dev: |
5846 | device_del(pmu->dev); | |
5847 | put_device(pmu->dev); | |
5848 | ||
2e80a82a PZ |
5849 | free_idr: |
5850 | if (pmu->type >= PERF_TYPE_MAX) | |
5851 | idr_remove(&pmu_idr, pmu->type); | |
5852 | ||
108b02cf PZ |
5853 | free_pdc: |
5854 | free_percpu(pmu->pmu_disable_count); | |
5855 | goto unlock; | |
f29ac756 PZ |
5856 | } |
5857 | ||
b0a873eb | 5858 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 5859 | { |
b0a873eb PZ |
5860 | mutex_lock(&pmus_lock); |
5861 | list_del_rcu(&pmu->entry); | |
5862 | mutex_unlock(&pmus_lock); | |
5c92d124 | 5863 | |
0475f9ea | 5864 | /* |
cde8e884 PZ |
5865 | * We dereference the pmu list under both SRCU and regular RCU, so |
5866 | * synchronize against both of those. | |
0475f9ea | 5867 | */ |
b0a873eb | 5868 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 5869 | synchronize_rcu(); |
d6d020e9 | 5870 | |
33696fc0 | 5871 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
5872 | if (pmu->type >= PERF_TYPE_MAX) |
5873 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
5874 | device_del(pmu->dev); |
5875 | put_device(pmu->dev); | |
51676957 | 5876 | free_pmu_context(pmu); |
b0a873eb | 5877 | } |
d6d020e9 | 5878 | |
b0a873eb PZ |
5879 | struct pmu *perf_init_event(struct perf_event *event) |
5880 | { | |
5881 | struct pmu *pmu = NULL; | |
5882 | int idx; | |
940c5b29 | 5883 | int ret; |
b0a873eb PZ |
5884 | |
5885 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
5886 | |
5887 | rcu_read_lock(); | |
5888 | pmu = idr_find(&pmu_idr, event->attr.type); | |
5889 | rcu_read_unlock(); | |
940c5b29 | 5890 | if (pmu) { |
7e5b2a01 | 5891 | event->pmu = pmu; |
940c5b29 LM |
5892 | ret = pmu->event_init(event); |
5893 | if (ret) | |
5894 | pmu = ERR_PTR(ret); | |
2e80a82a | 5895 | goto unlock; |
940c5b29 | 5896 | } |
2e80a82a | 5897 | |
b0a873eb | 5898 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 5899 | event->pmu = pmu; |
940c5b29 | 5900 | ret = pmu->event_init(event); |
b0a873eb | 5901 | if (!ret) |
e5f4d339 | 5902 | goto unlock; |
76e1d904 | 5903 | |
b0a873eb PZ |
5904 | if (ret != -ENOENT) { |
5905 | pmu = ERR_PTR(ret); | |
e5f4d339 | 5906 | goto unlock; |
f344011c | 5907 | } |
5c92d124 | 5908 | } |
e5f4d339 PZ |
5909 | pmu = ERR_PTR(-ENOENT); |
5910 | unlock: | |
b0a873eb | 5911 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 5912 | |
4aeb0b42 | 5913 | return pmu; |
5c92d124 IM |
5914 | } |
5915 | ||
0793a61d | 5916 | /* |
cdd6c482 | 5917 | * Allocate and initialize a event structure |
0793a61d | 5918 | */ |
cdd6c482 | 5919 | static struct perf_event * |
c3f00c70 | 5920 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
5921 | struct task_struct *task, |
5922 | struct perf_event *group_leader, | |
5923 | struct perf_event *parent_event, | |
4dc0da86 AK |
5924 | perf_overflow_handler_t overflow_handler, |
5925 | void *context) | |
0793a61d | 5926 | { |
51b0fe39 | 5927 | struct pmu *pmu; |
cdd6c482 IM |
5928 | struct perf_event *event; |
5929 | struct hw_perf_event *hwc; | |
d5d2bc0d | 5930 | long err; |
0793a61d | 5931 | |
66832eb4 ON |
5932 | if ((unsigned)cpu >= nr_cpu_ids) { |
5933 | if (!task || cpu != -1) | |
5934 | return ERR_PTR(-EINVAL); | |
5935 | } | |
5936 | ||
c3f00c70 | 5937 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 5938 | if (!event) |
d5d2bc0d | 5939 | return ERR_PTR(-ENOMEM); |
0793a61d | 5940 | |
04289bb9 | 5941 | /* |
cdd6c482 | 5942 | * Single events are their own group leaders, with an |
04289bb9 IM |
5943 | * empty sibling list: |
5944 | */ | |
5945 | if (!group_leader) | |
cdd6c482 | 5946 | group_leader = event; |
04289bb9 | 5947 | |
cdd6c482 IM |
5948 | mutex_init(&event->child_mutex); |
5949 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 5950 | |
cdd6c482 IM |
5951 | INIT_LIST_HEAD(&event->group_entry); |
5952 | INIT_LIST_HEAD(&event->event_entry); | |
5953 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
5954 | INIT_LIST_HEAD(&event->rb_entry); |
5955 | ||
cdd6c482 | 5956 | init_waitqueue_head(&event->waitq); |
e360adbe | 5957 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 5958 | |
cdd6c482 | 5959 | mutex_init(&event->mmap_mutex); |
7b732a75 | 5960 | |
a6fa941d | 5961 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
5962 | event->cpu = cpu; |
5963 | event->attr = *attr; | |
5964 | event->group_leader = group_leader; | |
5965 | event->pmu = NULL; | |
cdd6c482 | 5966 | event->oncpu = -1; |
a96bbc16 | 5967 | |
cdd6c482 | 5968 | event->parent = parent_event; |
b84fbc9f | 5969 | |
cdd6c482 IM |
5970 | event->ns = get_pid_ns(current->nsproxy->pid_ns); |
5971 | event->id = atomic64_inc_return(&perf_event_id); | |
a96bbc16 | 5972 | |
cdd6c482 | 5973 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 5974 | |
d580ff86 PZ |
5975 | if (task) { |
5976 | event->attach_state = PERF_ATTACH_TASK; | |
5977 | #ifdef CONFIG_HAVE_HW_BREAKPOINT | |
5978 | /* | |
5979 | * hw_breakpoint is a bit difficult here.. | |
5980 | */ | |
5981 | if (attr->type == PERF_TYPE_BREAKPOINT) | |
5982 | event->hw.bp_target = task; | |
5983 | #endif | |
5984 | } | |
5985 | ||
4dc0da86 | 5986 | if (!overflow_handler && parent_event) { |
b326e956 | 5987 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
5988 | context = parent_event->overflow_handler_context; |
5989 | } | |
66832eb4 | 5990 | |
b326e956 | 5991 | event->overflow_handler = overflow_handler; |
4dc0da86 | 5992 | event->overflow_handler_context = context; |
97eaf530 | 5993 | |
0d48696f | 5994 | if (attr->disabled) |
cdd6c482 | 5995 | event->state = PERF_EVENT_STATE_OFF; |
a86ed508 | 5996 | |
4aeb0b42 | 5997 | pmu = NULL; |
b8e83514 | 5998 | |
cdd6c482 | 5999 | hwc = &event->hw; |
bd2b5b12 | 6000 | hwc->sample_period = attr->sample_period; |
0d48696f | 6001 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6002 | hwc->sample_period = 1; |
eced1dfc | 6003 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6004 | |
e7850595 | 6005 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6006 | |
2023b359 | 6007 | /* |
cdd6c482 | 6008 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6009 | */ |
3dab77fb | 6010 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
2023b359 PZ |
6011 | goto done; |
6012 | ||
b0a873eb | 6013 | pmu = perf_init_event(event); |
974802ea | 6014 | |
d5d2bc0d PM |
6015 | done: |
6016 | err = 0; | |
4aeb0b42 | 6017 | if (!pmu) |
d5d2bc0d | 6018 | err = -EINVAL; |
4aeb0b42 RR |
6019 | else if (IS_ERR(pmu)) |
6020 | err = PTR_ERR(pmu); | |
5c92d124 | 6021 | |
d5d2bc0d | 6022 | if (err) { |
cdd6c482 IM |
6023 | if (event->ns) |
6024 | put_pid_ns(event->ns); | |
6025 | kfree(event); | |
d5d2bc0d | 6026 | return ERR_PTR(err); |
621a01ea | 6027 | } |
d5d2bc0d | 6028 | |
cdd6c482 | 6029 | if (!event->parent) { |
82cd6def | 6030 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 6031 | static_key_slow_inc(&perf_sched_events.key); |
3af9e859 | 6032 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
6033 | atomic_inc(&nr_mmap_events); |
6034 | if (event->attr.comm) | |
6035 | atomic_inc(&nr_comm_events); | |
6036 | if (event->attr.task) | |
6037 | atomic_inc(&nr_task_events); | |
927c7a9e FW |
6038 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6039 | err = get_callchain_buffers(); | |
6040 | if (err) { | |
6041 | free_event(event); | |
6042 | return ERR_PTR(err); | |
6043 | } | |
6044 | } | |
d010b332 SE |
6045 | if (has_branch_stack(event)) { |
6046 | static_key_slow_inc(&perf_sched_events.key); | |
6047 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6048 | atomic_inc(&per_cpu(perf_branch_stack_events, | |
6049 | event->cpu)); | |
6050 | } | |
f344011c | 6051 | } |
9ee318a7 | 6052 | |
cdd6c482 | 6053 | return event; |
0793a61d TG |
6054 | } |
6055 | ||
cdd6c482 IM |
6056 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6057 | struct perf_event_attr *attr) | |
974802ea | 6058 | { |
974802ea | 6059 | u32 size; |
cdf8073d | 6060 | int ret; |
974802ea PZ |
6061 | |
6062 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6063 | return -EFAULT; | |
6064 | ||
6065 | /* | |
6066 | * zero the full structure, so that a short copy will be nice. | |
6067 | */ | |
6068 | memset(attr, 0, sizeof(*attr)); | |
6069 | ||
6070 | ret = get_user(size, &uattr->size); | |
6071 | if (ret) | |
6072 | return ret; | |
6073 | ||
6074 | if (size > PAGE_SIZE) /* silly large */ | |
6075 | goto err_size; | |
6076 | ||
6077 | if (!size) /* abi compat */ | |
6078 | size = PERF_ATTR_SIZE_VER0; | |
6079 | ||
6080 | if (size < PERF_ATTR_SIZE_VER0) | |
6081 | goto err_size; | |
6082 | ||
6083 | /* | |
6084 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6085 | * ensure all the unknown bits are 0 - i.e. new |
6086 | * user-space does not rely on any kernel feature | |
6087 | * extensions we dont know about yet. | |
974802ea PZ |
6088 | */ |
6089 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6090 | unsigned char __user *addr; |
6091 | unsigned char __user *end; | |
6092 | unsigned char val; | |
974802ea | 6093 | |
cdf8073d IS |
6094 | addr = (void __user *)uattr + sizeof(*attr); |
6095 | end = (void __user *)uattr + size; | |
974802ea | 6096 | |
cdf8073d | 6097 | for (; addr < end; addr++) { |
974802ea PZ |
6098 | ret = get_user(val, addr); |
6099 | if (ret) | |
6100 | return ret; | |
6101 | if (val) | |
6102 | goto err_size; | |
6103 | } | |
b3e62e35 | 6104 | size = sizeof(*attr); |
974802ea PZ |
6105 | } |
6106 | ||
6107 | ret = copy_from_user(attr, uattr, size); | |
6108 | if (ret) | |
6109 | return -EFAULT; | |
6110 | ||
cd757645 | 6111 | if (attr->__reserved_1) |
974802ea PZ |
6112 | return -EINVAL; |
6113 | ||
6114 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6115 | return -EINVAL; | |
6116 | ||
6117 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6118 | return -EINVAL; | |
6119 | ||
bce38cd5 SE |
6120 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6121 | u64 mask = attr->branch_sample_type; | |
6122 | ||
6123 | /* only using defined bits */ | |
6124 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6125 | return -EINVAL; | |
6126 | ||
6127 | /* at least one branch bit must be set */ | |
6128 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6129 | return -EINVAL; | |
6130 | ||
6131 | /* kernel level capture: check permissions */ | |
6132 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
6133 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6134 | return -EACCES; | |
6135 | ||
6136 | /* propagate priv level, when not set for branch */ | |
6137 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6138 | ||
6139 | /* exclude_kernel checked on syscall entry */ | |
6140 | if (!attr->exclude_kernel) | |
6141 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6142 | ||
6143 | if (!attr->exclude_user) | |
6144 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6145 | ||
6146 | if (!attr->exclude_hv) | |
6147 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6148 | /* | |
6149 | * adjust user setting (for HW filter setup) | |
6150 | */ | |
6151 | attr->branch_sample_type = mask; | |
6152 | } | |
6153 | } | |
974802ea PZ |
6154 | out: |
6155 | return ret; | |
6156 | ||
6157 | err_size: | |
6158 | put_user(sizeof(*attr), &uattr->size); | |
6159 | ret = -E2BIG; | |
6160 | goto out; | |
6161 | } | |
6162 | ||
ac9721f3 PZ |
6163 | static int |
6164 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6165 | { |
76369139 | 6166 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6167 | int ret = -EINVAL; |
6168 | ||
ac9721f3 | 6169 | if (!output_event) |
a4be7c27 PZ |
6170 | goto set; |
6171 | ||
ac9721f3 PZ |
6172 | /* don't allow circular references */ |
6173 | if (event == output_event) | |
a4be7c27 PZ |
6174 | goto out; |
6175 | ||
0f139300 PZ |
6176 | /* |
6177 | * Don't allow cross-cpu buffers | |
6178 | */ | |
6179 | if (output_event->cpu != event->cpu) | |
6180 | goto out; | |
6181 | ||
6182 | /* | |
76369139 | 6183 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6184 | */ |
6185 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6186 | goto out; | |
6187 | ||
a4be7c27 | 6188 | set: |
cdd6c482 | 6189 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6190 | /* Can't redirect output if we've got an active mmap() */ |
6191 | if (atomic_read(&event->mmap_count)) | |
6192 | goto unlock; | |
a4be7c27 | 6193 | |
ac9721f3 | 6194 | if (output_event) { |
76369139 FW |
6195 | /* get the rb we want to redirect to */ |
6196 | rb = ring_buffer_get(output_event); | |
6197 | if (!rb) | |
ac9721f3 | 6198 | goto unlock; |
a4be7c27 PZ |
6199 | } |
6200 | ||
76369139 FW |
6201 | old_rb = event->rb; |
6202 | rcu_assign_pointer(event->rb, rb); | |
10c6db11 PZ |
6203 | if (old_rb) |
6204 | ring_buffer_detach(event, old_rb); | |
a4be7c27 | 6205 | ret = 0; |
ac9721f3 PZ |
6206 | unlock: |
6207 | mutex_unlock(&event->mmap_mutex); | |
6208 | ||
76369139 FW |
6209 | if (old_rb) |
6210 | ring_buffer_put(old_rb); | |
a4be7c27 | 6211 | out: |
a4be7c27 PZ |
6212 | return ret; |
6213 | } | |
6214 | ||
0793a61d | 6215 | /** |
cdd6c482 | 6216 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6217 | * |
cdd6c482 | 6218 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6219 | * @pid: target pid |
9f66a381 | 6220 | * @cpu: target cpu |
cdd6c482 | 6221 | * @group_fd: group leader event fd |
0793a61d | 6222 | */ |
cdd6c482 IM |
6223 | SYSCALL_DEFINE5(perf_event_open, |
6224 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 6225 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 6226 | { |
b04243ef PZ |
6227 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6228 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
6229 | struct perf_event_attr attr; |
6230 | struct perf_event_context *ctx; | |
6231 | struct file *event_file = NULL; | |
04289bb9 | 6232 | struct file *group_file = NULL; |
38a81da2 | 6233 | struct task_struct *task = NULL; |
89a1e187 | 6234 | struct pmu *pmu; |
ea635c64 | 6235 | int event_fd; |
b04243ef | 6236 | int move_group = 0; |
04289bb9 | 6237 | int fput_needed = 0; |
dc86cabe | 6238 | int err; |
0793a61d | 6239 | |
2743a5b0 | 6240 | /* for future expandability... */ |
e5d1367f | 6241 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
6242 | return -EINVAL; |
6243 | ||
dc86cabe IM |
6244 | err = perf_copy_attr(attr_uptr, &attr); |
6245 | if (err) | |
6246 | return err; | |
eab656ae | 6247 | |
0764771d PZ |
6248 | if (!attr.exclude_kernel) { |
6249 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6250 | return -EACCES; | |
6251 | } | |
6252 | ||
df58ab24 | 6253 | if (attr.freq) { |
cdd6c482 | 6254 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
6255 | return -EINVAL; |
6256 | } | |
6257 | ||
e5d1367f SE |
6258 | /* |
6259 | * In cgroup mode, the pid argument is used to pass the fd | |
6260 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6261 | * designates the cpu on which to monitor threads from that | |
6262 | * cgroup. | |
6263 | */ | |
6264 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6265 | return -EINVAL; | |
6266 | ||
ab72a702 | 6267 | event_fd = get_unused_fd(); |
ea635c64 AV |
6268 | if (event_fd < 0) |
6269 | return event_fd; | |
6270 | ||
ac9721f3 | 6271 | if (group_fd != -1) { |
a6fa941d AV |
6272 | group_file = perf_fget_light(group_fd, &fput_needed); |
6273 | if (IS_ERR(group_file)) { | |
6274 | err = PTR_ERR(group_file); | |
d14b12d7 | 6275 | goto err_fd; |
ac9721f3 | 6276 | } |
a6fa941d | 6277 | group_leader = group_file->private_data; |
ac9721f3 PZ |
6278 | if (flags & PERF_FLAG_FD_OUTPUT) |
6279 | output_event = group_leader; | |
6280 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6281 | group_leader = NULL; | |
6282 | } | |
6283 | ||
e5d1367f | 6284 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6285 | task = find_lively_task_by_vpid(pid); |
6286 | if (IS_ERR(task)) { | |
6287 | err = PTR_ERR(task); | |
6288 | goto err_group_fd; | |
6289 | } | |
6290 | } | |
6291 | ||
fbfc623f YZ |
6292 | get_online_cpus(); |
6293 | ||
4dc0da86 AK |
6294 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6295 | NULL, NULL); | |
d14b12d7 SE |
6296 | if (IS_ERR(event)) { |
6297 | err = PTR_ERR(event); | |
c6be5a5c | 6298 | goto err_task; |
d14b12d7 SE |
6299 | } |
6300 | ||
e5d1367f SE |
6301 | if (flags & PERF_FLAG_PID_CGROUP) { |
6302 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
6303 | if (err) | |
6304 | goto err_alloc; | |
08309379 PZ |
6305 | /* |
6306 | * one more event: | |
6307 | * - that has cgroup constraint on event->cpu | |
6308 | * - that may need work on context switch | |
6309 | */ | |
6310 | atomic_inc(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 6311 | static_key_slow_inc(&perf_sched_events.key); |
e5d1367f SE |
6312 | } |
6313 | ||
89a1e187 PZ |
6314 | /* |
6315 | * Special case software events and allow them to be part of | |
6316 | * any hardware group. | |
6317 | */ | |
6318 | pmu = event->pmu; | |
b04243ef PZ |
6319 | |
6320 | if (group_leader && | |
6321 | (is_software_event(event) != is_software_event(group_leader))) { | |
6322 | if (is_software_event(event)) { | |
6323 | /* | |
6324 | * If event and group_leader are not both a software | |
6325 | * event, and event is, then group leader is not. | |
6326 | * | |
6327 | * Allow the addition of software events to !software | |
6328 | * groups, this is safe because software events never | |
6329 | * fail to schedule. | |
6330 | */ | |
6331 | pmu = group_leader->pmu; | |
6332 | } else if (is_software_event(group_leader) && | |
6333 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
6334 | /* | |
6335 | * In case the group is a pure software group, and we | |
6336 | * try to add a hardware event, move the whole group to | |
6337 | * the hardware context. | |
6338 | */ | |
6339 | move_group = 1; | |
6340 | } | |
6341 | } | |
89a1e187 PZ |
6342 | |
6343 | /* | |
6344 | * Get the target context (task or percpu): | |
6345 | */ | |
e2d37cd2 | 6346 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
6347 | if (IS_ERR(ctx)) { |
6348 | err = PTR_ERR(ctx); | |
c6be5a5c | 6349 | goto err_alloc; |
89a1e187 PZ |
6350 | } |
6351 | ||
fd1edb3a PZ |
6352 | if (task) { |
6353 | put_task_struct(task); | |
6354 | task = NULL; | |
6355 | } | |
6356 | ||
ccff286d | 6357 | /* |
cdd6c482 | 6358 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 6359 | */ |
ac9721f3 | 6360 | if (group_leader) { |
dc86cabe | 6361 | err = -EINVAL; |
04289bb9 | 6362 | |
04289bb9 | 6363 | /* |
ccff286d IM |
6364 | * Do not allow a recursive hierarchy (this new sibling |
6365 | * becoming part of another group-sibling): | |
6366 | */ | |
6367 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 6368 | goto err_context; |
ccff286d IM |
6369 | /* |
6370 | * Do not allow to attach to a group in a different | |
6371 | * task or CPU context: | |
04289bb9 | 6372 | */ |
b04243ef PZ |
6373 | if (move_group) { |
6374 | if (group_leader->ctx->type != ctx->type) | |
6375 | goto err_context; | |
6376 | } else { | |
6377 | if (group_leader->ctx != ctx) | |
6378 | goto err_context; | |
6379 | } | |
6380 | ||
3b6f9e5c PM |
6381 | /* |
6382 | * Only a group leader can be exclusive or pinned | |
6383 | */ | |
0d48696f | 6384 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 6385 | goto err_context; |
ac9721f3 PZ |
6386 | } |
6387 | ||
6388 | if (output_event) { | |
6389 | err = perf_event_set_output(event, output_event); | |
6390 | if (err) | |
c3f00c70 | 6391 | goto err_context; |
ac9721f3 | 6392 | } |
0793a61d | 6393 | |
ea635c64 AV |
6394 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
6395 | if (IS_ERR(event_file)) { | |
6396 | err = PTR_ERR(event_file); | |
c3f00c70 | 6397 | goto err_context; |
ea635c64 | 6398 | } |
9b51f66d | 6399 | |
b04243ef PZ |
6400 | if (move_group) { |
6401 | struct perf_event_context *gctx = group_leader->ctx; | |
6402 | ||
6403 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 6404 | perf_remove_from_context(group_leader); |
b04243ef PZ |
6405 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6406 | group_entry) { | |
fe4b04fa | 6407 | perf_remove_from_context(sibling); |
b04243ef PZ |
6408 | put_ctx(gctx); |
6409 | } | |
6410 | mutex_unlock(&gctx->mutex); | |
6411 | put_ctx(gctx); | |
ea635c64 | 6412 | } |
9b51f66d | 6413 | |
ad3a37de | 6414 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 6415 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
6416 | |
6417 | if (move_group) { | |
0cda4c02 | 6418 | synchronize_rcu(); |
e2d37cd2 | 6419 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
6420 | get_ctx(ctx); |
6421 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
6422 | group_entry) { | |
e2d37cd2 | 6423 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
6424 | get_ctx(ctx); |
6425 | } | |
6426 | } | |
6427 | ||
e2d37cd2 | 6428 | perf_install_in_context(ctx, event, event->cpu); |
ad3a37de | 6429 | ++ctx->generation; |
fe4b04fa | 6430 | perf_unpin_context(ctx); |
d859e29f | 6431 | mutex_unlock(&ctx->mutex); |
9b51f66d | 6432 | |
fbfc623f YZ |
6433 | put_online_cpus(); |
6434 | ||
cdd6c482 | 6435 | event->owner = current; |
8882135b | 6436 | |
cdd6c482 IM |
6437 | mutex_lock(¤t->perf_event_mutex); |
6438 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
6439 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 6440 | |
c320c7b7 ACM |
6441 | /* |
6442 | * Precalculate sample_data sizes | |
6443 | */ | |
6444 | perf_event__header_size(event); | |
6844c09d | 6445 | perf_event__id_header_size(event); |
c320c7b7 | 6446 | |
8a49542c PZ |
6447 | /* |
6448 | * Drop the reference on the group_event after placing the | |
6449 | * new event on the sibling_list. This ensures destruction | |
6450 | * of the group leader will find the pointer to itself in | |
6451 | * perf_group_detach(). | |
6452 | */ | |
ea635c64 AV |
6453 | fput_light(group_file, fput_needed); |
6454 | fd_install(event_fd, event_file); | |
6455 | return event_fd; | |
0793a61d | 6456 | |
c3f00c70 | 6457 | err_context: |
fe4b04fa | 6458 | perf_unpin_context(ctx); |
ea635c64 | 6459 | put_ctx(ctx); |
c6be5a5c | 6460 | err_alloc: |
ea635c64 | 6461 | free_event(event); |
e7d0bc04 | 6462 | err_task: |
fbfc623f | 6463 | put_online_cpus(); |
e7d0bc04 PZ |
6464 | if (task) |
6465 | put_task_struct(task); | |
89a1e187 | 6466 | err_group_fd: |
dc86cabe | 6467 | fput_light(group_file, fput_needed); |
ea635c64 AV |
6468 | err_fd: |
6469 | put_unused_fd(event_fd); | |
dc86cabe | 6470 | return err; |
0793a61d TG |
6471 | } |
6472 | ||
fb0459d7 AV |
6473 | /** |
6474 | * perf_event_create_kernel_counter | |
6475 | * | |
6476 | * @attr: attributes of the counter to create | |
6477 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 6478 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
6479 | */ |
6480 | struct perf_event * | |
6481 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 6482 | struct task_struct *task, |
4dc0da86 AK |
6483 | perf_overflow_handler_t overflow_handler, |
6484 | void *context) | |
fb0459d7 | 6485 | { |
fb0459d7 | 6486 | struct perf_event_context *ctx; |
c3f00c70 | 6487 | struct perf_event *event; |
fb0459d7 | 6488 | int err; |
d859e29f | 6489 | |
fb0459d7 AV |
6490 | /* |
6491 | * Get the target context (task or percpu): | |
6492 | */ | |
d859e29f | 6493 | |
4dc0da86 AK |
6494 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
6495 | overflow_handler, context); | |
c3f00c70 PZ |
6496 | if (IS_ERR(event)) { |
6497 | err = PTR_ERR(event); | |
6498 | goto err; | |
6499 | } | |
d859e29f | 6500 | |
38a81da2 | 6501 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
6502 | if (IS_ERR(ctx)) { |
6503 | err = PTR_ERR(ctx); | |
c3f00c70 | 6504 | goto err_free; |
d859e29f | 6505 | } |
fb0459d7 | 6506 | |
fb0459d7 AV |
6507 | WARN_ON_ONCE(ctx->parent_ctx); |
6508 | mutex_lock(&ctx->mutex); | |
6509 | perf_install_in_context(ctx, event, cpu); | |
6510 | ++ctx->generation; | |
fe4b04fa | 6511 | perf_unpin_context(ctx); |
fb0459d7 AV |
6512 | mutex_unlock(&ctx->mutex); |
6513 | ||
fb0459d7 AV |
6514 | return event; |
6515 | ||
c3f00c70 PZ |
6516 | err_free: |
6517 | free_event(event); | |
6518 | err: | |
c6567f64 | 6519 | return ERR_PTR(err); |
9b51f66d | 6520 | } |
fb0459d7 | 6521 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 6522 | |
0cda4c02 YZ |
6523 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
6524 | { | |
6525 | struct perf_event_context *src_ctx; | |
6526 | struct perf_event_context *dst_ctx; | |
6527 | struct perf_event *event, *tmp; | |
6528 | LIST_HEAD(events); | |
6529 | ||
6530 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
6531 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
6532 | ||
6533 | mutex_lock(&src_ctx->mutex); | |
6534 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
6535 | event_entry) { | |
6536 | perf_remove_from_context(event); | |
6537 | put_ctx(src_ctx); | |
6538 | list_add(&event->event_entry, &events); | |
6539 | } | |
6540 | mutex_unlock(&src_ctx->mutex); | |
6541 | ||
6542 | synchronize_rcu(); | |
6543 | ||
6544 | mutex_lock(&dst_ctx->mutex); | |
6545 | list_for_each_entry_safe(event, tmp, &events, event_entry) { | |
6546 | list_del(&event->event_entry); | |
6547 | if (event->state >= PERF_EVENT_STATE_OFF) | |
6548 | event->state = PERF_EVENT_STATE_INACTIVE; | |
6549 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
6550 | get_ctx(dst_ctx); | |
6551 | } | |
6552 | mutex_unlock(&dst_ctx->mutex); | |
6553 | } | |
6554 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
6555 | ||
cdd6c482 | 6556 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 6557 | struct task_struct *child) |
d859e29f | 6558 | { |
cdd6c482 | 6559 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 6560 | u64 child_val; |
d859e29f | 6561 | |
cdd6c482 IM |
6562 | if (child_event->attr.inherit_stat) |
6563 | perf_event_read_event(child_event, child); | |
38b200d6 | 6564 | |
b5e58793 | 6565 | child_val = perf_event_count(child_event); |
d859e29f PM |
6566 | |
6567 | /* | |
6568 | * Add back the child's count to the parent's count: | |
6569 | */ | |
a6e6dea6 | 6570 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
6571 | atomic64_add(child_event->total_time_enabled, |
6572 | &parent_event->child_total_time_enabled); | |
6573 | atomic64_add(child_event->total_time_running, | |
6574 | &parent_event->child_total_time_running); | |
d859e29f PM |
6575 | |
6576 | /* | |
cdd6c482 | 6577 | * Remove this event from the parent's list |
d859e29f | 6578 | */ |
cdd6c482 IM |
6579 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
6580 | mutex_lock(&parent_event->child_mutex); | |
6581 | list_del_init(&child_event->child_list); | |
6582 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
6583 | |
6584 | /* | |
cdd6c482 | 6585 | * Release the parent event, if this was the last |
d859e29f PM |
6586 | * reference to it. |
6587 | */ | |
a6fa941d | 6588 | put_event(parent_event); |
d859e29f PM |
6589 | } |
6590 | ||
9b51f66d | 6591 | static void |
cdd6c482 IM |
6592 | __perf_event_exit_task(struct perf_event *child_event, |
6593 | struct perf_event_context *child_ctx, | |
38b200d6 | 6594 | struct task_struct *child) |
9b51f66d | 6595 | { |
38b435b1 PZ |
6596 | if (child_event->parent) { |
6597 | raw_spin_lock_irq(&child_ctx->lock); | |
6598 | perf_group_detach(child_event); | |
6599 | raw_spin_unlock_irq(&child_ctx->lock); | |
6600 | } | |
9b51f66d | 6601 | |
fe4b04fa | 6602 | perf_remove_from_context(child_event); |
0cc0c027 | 6603 | |
9b51f66d | 6604 | /* |
38b435b1 | 6605 | * It can happen that the parent exits first, and has events |
9b51f66d | 6606 | * that are still around due to the child reference. These |
38b435b1 | 6607 | * events need to be zapped. |
9b51f66d | 6608 | */ |
38b435b1 | 6609 | if (child_event->parent) { |
cdd6c482 IM |
6610 | sync_child_event(child_event, child); |
6611 | free_event(child_event); | |
4bcf349a | 6612 | } |
9b51f66d IM |
6613 | } |
6614 | ||
8dc85d54 | 6615 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 6616 | { |
cdd6c482 IM |
6617 | struct perf_event *child_event, *tmp; |
6618 | struct perf_event_context *child_ctx; | |
a63eaf34 | 6619 | unsigned long flags; |
9b51f66d | 6620 | |
8dc85d54 | 6621 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 6622 | perf_event_task(child, NULL, 0); |
9b51f66d | 6623 | return; |
9f498cc5 | 6624 | } |
9b51f66d | 6625 | |
a63eaf34 | 6626 | local_irq_save(flags); |
ad3a37de PM |
6627 | /* |
6628 | * We can't reschedule here because interrupts are disabled, | |
6629 | * and either child is current or it is a task that can't be | |
6630 | * scheduled, so we are now safe from rescheduling changing | |
6631 | * our context. | |
6632 | */ | |
806839b2 | 6633 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
6634 | |
6635 | /* | |
6636 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 6637 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
6638 | * incremented the context's refcount before we do put_ctx below. |
6639 | */ | |
e625cce1 | 6640 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 6641 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 6642 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
6643 | /* |
6644 | * If this context is a clone; unclone it so it can't get | |
6645 | * swapped to another process while we're removing all | |
cdd6c482 | 6646 | * the events from it. |
71a851b4 PZ |
6647 | */ |
6648 | unclone_ctx(child_ctx); | |
5e942bb3 | 6649 | update_context_time(child_ctx); |
e625cce1 | 6650 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
6651 | |
6652 | /* | |
cdd6c482 IM |
6653 | * Report the task dead after unscheduling the events so that we |
6654 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
6655 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 6656 | */ |
cdd6c482 | 6657 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 6658 | |
66fff224 PZ |
6659 | /* |
6660 | * We can recurse on the same lock type through: | |
6661 | * | |
cdd6c482 IM |
6662 | * __perf_event_exit_task() |
6663 | * sync_child_event() | |
a6fa941d AV |
6664 | * put_event() |
6665 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
6666 | * |
6667 | * But since its the parent context it won't be the same instance. | |
6668 | */ | |
a0507c84 | 6669 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 6670 | |
8bc20959 | 6671 | again: |
889ff015 FW |
6672 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
6673 | group_entry) | |
6674 | __perf_event_exit_task(child_event, child_ctx, child); | |
6675 | ||
6676 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 6677 | group_entry) |
cdd6c482 | 6678 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
6679 | |
6680 | /* | |
cdd6c482 | 6681 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
6682 | * its siblings to the list, but we obtained 'tmp' before that which |
6683 | * will still point to the list head terminating the iteration. | |
6684 | */ | |
889ff015 FW |
6685 | if (!list_empty(&child_ctx->pinned_groups) || |
6686 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 6687 | goto again; |
a63eaf34 PM |
6688 | |
6689 | mutex_unlock(&child_ctx->mutex); | |
6690 | ||
6691 | put_ctx(child_ctx); | |
9b51f66d IM |
6692 | } |
6693 | ||
8dc85d54 PZ |
6694 | /* |
6695 | * When a child task exits, feed back event values to parent events. | |
6696 | */ | |
6697 | void perf_event_exit_task(struct task_struct *child) | |
6698 | { | |
8882135b | 6699 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
6700 | int ctxn; |
6701 | ||
8882135b PZ |
6702 | mutex_lock(&child->perf_event_mutex); |
6703 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
6704 | owner_entry) { | |
6705 | list_del_init(&event->owner_entry); | |
6706 | ||
6707 | /* | |
6708 | * Ensure the list deletion is visible before we clear | |
6709 | * the owner, closes a race against perf_release() where | |
6710 | * we need to serialize on the owner->perf_event_mutex. | |
6711 | */ | |
6712 | smp_wmb(); | |
6713 | event->owner = NULL; | |
6714 | } | |
6715 | mutex_unlock(&child->perf_event_mutex); | |
6716 | ||
8dc85d54 PZ |
6717 | for_each_task_context_nr(ctxn) |
6718 | perf_event_exit_task_context(child, ctxn); | |
6719 | } | |
6720 | ||
889ff015 FW |
6721 | static void perf_free_event(struct perf_event *event, |
6722 | struct perf_event_context *ctx) | |
6723 | { | |
6724 | struct perf_event *parent = event->parent; | |
6725 | ||
6726 | if (WARN_ON_ONCE(!parent)) | |
6727 | return; | |
6728 | ||
6729 | mutex_lock(&parent->child_mutex); | |
6730 | list_del_init(&event->child_list); | |
6731 | mutex_unlock(&parent->child_mutex); | |
6732 | ||
a6fa941d | 6733 | put_event(parent); |
889ff015 | 6734 | |
8a49542c | 6735 | perf_group_detach(event); |
889ff015 FW |
6736 | list_del_event(event, ctx); |
6737 | free_event(event); | |
6738 | } | |
6739 | ||
bbbee908 PZ |
6740 | /* |
6741 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 6742 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 6743 | */ |
cdd6c482 | 6744 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 6745 | { |
8dc85d54 | 6746 | struct perf_event_context *ctx; |
cdd6c482 | 6747 | struct perf_event *event, *tmp; |
8dc85d54 | 6748 | int ctxn; |
bbbee908 | 6749 | |
8dc85d54 PZ |
6750 | for_each_task_context_nr(ctxn) { |
6751 | ctx = task->perf_event_ctxp[ctxn]; | |
6752 | if (!ctx) | |
6753 | continue; | |
bbbee908 | 6754 | |
8dc85d54 | 6755 | mutex_lock(&ctx->mutex); |
bbbee908 | 6756 | again: |
8dc85d54 PZ |
6757 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
6758 | group_entry) | |
6759 | perf_free_event(event, ctx); | |
bbbee908 | 6760 | |
8dc85d54 PZ |
6761 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
6762 | group_entry) | |
6763 | perf_free_event(event, ctx); | |
bbbee908 | 6764 | |
8dc85d54 PZ |
6765 | if (!list_empty(&ctx->pinned_groups) || |
6766 | !list_empty(&ctx->flexible_groups)) | |
6767 | goto again; | |
bbbee908 | 6768 | |
8dc85d54 | 6769 | mutex_unlock(&ctx->mutex); |
bbbee908 | 6770 | |
8dc85d54 PZ |
6771 | put_ctx(ctx); |
6772 | } | |
889ff015 FW |
6773 | } |
6774 | ||
4e231c79 PZ |
6775 | void perf_event_delayed_put(struct task_struct *task) |
6776 | { | |
6777 | int ctxn; | |
6778 | ||
6779 | for_each_task_context_nr(ctxn) | |
6780 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
6781 | } | |
6782 | ||
97dee4f3 PZ |
6783 | /* |
6784 | * inherit a event from parent task to child task: | |
6785 | */ | |
6786 | static struct perf_event * | |
6787 | inherit_event(struct perf_event *parent_event, | |
6788 | struct task_struct *parent, | |
6789 | struct perf_event_context *parent_ctx, | |
6790 | struct task_struct *child, | |
6791 | struct perf_event *group_leader, | |
6792 | struct perf_event_context *child_ctx) | |
6793 | { | |
6794 | struct perf_event *child_event; | |
cee010ec | 6795 | unsigned long flags; |
97dee4f3 PZ |
6796 | |
6797 | /* | |
6798 | * Instead of creating recursive hierarchies of events, | |
6799 | * we link inherited events back to the original parent, | |
6800 | * which has a filp for sure, which we use as the reference | |
6801 | * count: | |
6802 | */ | |
6803 | if (parent_event->parent) | |
6804 | parent_event = parent_event->parent; | |
6805 | ||
6806 | child_event = perf_event_alloc(&parent_event->attr, | |
6807 | parent_event->cpu, | |
d580ff86 | 6808 | child, |
97dee4f3 | 6809 | group_leader, parent_event, |
4dc0da86 | 6810 | NULL, NULL); |
97dee4f3 PZ |
6811 | if (IS_ERR(child_event)) |
6812 | return child_event; | |
a6fa941d AV |
6813 | |
6814 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
6815 | free_event(child_event); | |
6816 | return NULL; | |
6817 | } | |
6818 | ||
97dee4f3 PZ |
6819 | get_ctx(child_ctx); |
6820 | ||
6821 | /* | |
6822 | * Make the child state follow the state of the parent event, | |
6823 | * not its attr.disabled bit. We hold the parent's mutex, | |
6824 | * so we won't race with perf_event_{en, dis}able_family. | |
6825 | */ | |
6826 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
6827 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
6828 | else | |
6829 | child_event->state = PERF_EVENT_STATE_OFF; | |
6830 | ||
6831 | if (parent_event->attr.freq) { | |
6832 | u64 sample_period = parent_event->hw.sample_period; | |
6833 | struct hw_perf_event *hwc = &child_event->hw; | |
6834 | ||
6835 | hwc->sample_period = sample_period; | |
6836 | hwc->last_period = sample_period; | |
6837 | ||
6838 | local64_set(&hwc->period_left, sample_period); | |
6839 | } | |
6840 | ||
6841 | child_event->ctx = child_ctx; | |
6842 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
6843 | child_event->overflow_handler_context |
6844 | = parent_event->overflow_handler_context; | |
97dee4f3 | 6845 | |
614b6780 TG |
6846 | /* |
6847 | * Precalculate sample_data sizes | |
6848 | */ | |
6849 | perf_event__header_size(child_event); | |
6844c09d | 6850 | perf_event__id_header_size(child_event); |
614b6780 | 6851 | |
97dee4f3 PZ |
6852 | /* |
6853 | * Link it up in the child's context: | |
6854 | */ | |
cee010ec | 6855 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 6856 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 6857 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 6858 | |
97dee4f3 PZ |
6859 | /* |
6860 | * Link this into the parent event's child list | |
6861 | */ | |
6862 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
6863 | mutex_lock(&parent_event->child_mutex); | |
6864 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
6865 | mutex_unlock(&parent_event->child_mutex); | |
6866 | ||
6867 | return child_event; | |
6868 | } | |
6869 | ||
6870 | static int inherit_group(struct perf_event *parent_event, | |
6871 | struct task_struct *parent, | |
6872 | struct perf_event_context *parent_ctx, | |
6873 | struct task_struct *child, | |
6874 | struct perf_event_context *child_ctx) | |
6875 | { | |
6876 | struct perf_event *leader; | |
6877 | struct perf_event *sub; | |
6878 | struct perf_event *child_ctr; | |
6879 | ||
6880 | leader = inherit_event(parent_event, parent, parent_ctx, | |
6881 | child, NULL, child_ctx); | |
6882 | if (IS_ERR(leader)) | |
6883 | return PTR_ERR(leader); | |
6884 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
6885 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
6886 | child, leader, child_ctx); | |
6887 | if (IS_ERR(child_ctr)) | |
6888 | return PTR_ERR(child_ctr); | |
6889 | } | |
6890 | return 0; | |
889ff015 FW |
6891 | } |
6892 | ||
6893 | static int | |
6894 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
6895 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 6896 | struct task_struct *child, int ctxn, |
889ff015 FW |
6897 | int *inherited_all) |
6898 | { | |
6899 | int ret; | |
8dc85d54 | 6900 | struct perf_event_context *child_ctx; |
889ff015 FW |
6901 | |
6902 | if (!event->attr.inherit) { | |
6903 | *inherited_all = 0; | |
6904 | return 0; | |
bbbee908 PZ |
6905 | } |
6906 | ||
fe4b04fa | 6907 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
6908 | if (!child_ctx) { |
6909 | /* | |
6910 | * This is executed from the parent task context, so | |
6911 | * inherit events that have been marked for cloning. | |
6912 | * First allocate and initialize a context for the | |
6913 | * child. | |
6914 | */ | |
bbbee908 | 6915 | |
eb184479 | 6916 | child_ctx = alloc_perf_context(event->pmu, child); |
889ff015 FW |
6917 | if (!child_ctx) |
6918 | return -ENOMEM; | |
bbbee908 | 6919 | |
8dc85d54 | 6920 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
6921 | } |
6922 | ||
6923 | ret = inherit_group(event, parent, parent_ctx, | |
6924 | child, child_ctx); | |
6925 | ||
6926 | if (ret) | |
6927 | *inherited_all = 0; | |
6928 | ||
6929 | return ret; | |
bbbee908 PZ |
6930 | } |
6931 | ||
9b51f66d | 6932 | /* |
cdd6c482 | 6933 | * Initialize the perf_event context in task_struct |
9b51f66d | 6934 | */ |
8dc85d54 | 6935 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 6936 | { |
889ff015 | 6937 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
6938 | struct perf_event_context *cloned_ctx; |
6939 | struct perf_event *event; | |
9b51f66d | 6940 | struct task_struct *parent = current; |
564c2b21 | 6941 | int inherited_all = 1; |
dddd3379 | 6942 | unsigned long flags; |
6ab423e0 | 6943 | int ret = 0; |
9b51f66d | 6944 | |
8dc85d54 | 6945 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
6946 | return 0; |
6947 | ||
ad3a37de | 6948 | /* |
25346b93 PM |
6949 | * If the parent's context is a clone, pin it so it won't get |
6950 | * swapped under us. | |
ad3a37de | 6951 | */ |
8dc85d54 | 6952 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 6953 | |
ad3a37de PM |
6954 | /* |
6955 | * No need to check if parent_ctx != NULL here; since we saw | |
6956 | * it non-NULL earlier, the only reason for it to become NULL | |
6957 | * is if we exit, and since we're currently in the middle of | |
6958 | * a fork we can't be exiting at the same time. | |
6959 | */ | |
ad3a37de | 6960 | |
9b51f66d IM |
6961 | /* |
6962 | * Lock the parent list. No need to lock the child - not PID | |
6963 | * hashed yet and not running, so nobody can access it. | |
6964 | */ | |
d859e29f | 6965 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
6966 | |
6967 | /* | |
6968 | * We dont have to disable NMIs - we are only looking at | |
6969 | * the list, not manipulating it: | |
6970 | */ | |
889ff015 | 6971 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
6972 | ret = inherit_task_group(event, parent, parent_ctx, |
6973 | child, ctxn, &inherited_all); | |
889ff015 FW |
6974 | if (ret) |
6975 | break; | |
6976 | } | |
b93f7978 | 6977 | |
dddd3379 TG |
6978 | /* |
6979 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
6980 | * to allocations, but we need to prevent rotation because | |
6981 | * rotate_ctx() will change the list from interrupt context. | |
6982 | */ | |
6983 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
6984 | parent_ctx->rotate_disable = 1; | |
6985 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
6986 | ||
889ff015 | 6987 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
6988 | ret = inherit_task_group(event, parent, parent_ctx, |
6989 | child, ctxn, &inherited_all); | |
889ff015 | 6990 | if (ret) |
9b51f66d | 6991 | break; |
564c2b21 PM |
6992 | } |
6993 | ||
dddd3379 TG |
6994 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
6995 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 6996 | |
8dc85d54 | 6997 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 6998 | |
05cbaa28 | 6999 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7000 | /* |
7001 | * Mark the child context as a clone of the parent | |
7002 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7003 | * |
7004 | * Note that if the parent is a clone, the holding of | |
7005 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7006 | */ |
c5ed5145 | 7007 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7008 | if (cloned_ctx) { |
7009 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7010 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7011 | } else { |
7012 | child_ctx->parent_ctx = parent_ctx; | |
7013 | child_ctx->parent_gen = parent_ctx->generation; | |
7014 | } | |
7015 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7016 | } |
7017 | ||
c5ed5145 | 7018 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7019 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7020 | |
25346b93 | 7021 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7022 | put_ctx(parent_ctx); |
ad3a37de | 7023 | |
6ab423e0 | 7024 | return ret; |
9b51f66d IM |
7025 | } |
7026 | ||
8dc85d54 PZ |
7027 | /* |
7028 | * Initialize the perf_event context in task_struct | |
7029 | */ | |
7030 | int perf_event_init_task(struct task_struct *child) | |
7031 | { | |
7032 | int ctxn, ret; | |
7033 | ||
8550d7cb ON |
7034 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7035 | mutex_init(&child->perf_event_mutex); | |
7036 | INIT_LIST_HEAD(&child->perf_event_list); | |
7037 | ||
8dc85d54 PZ |
7038 | for_each_task_context_nr(ctxn) { |
7039 | ret = perf_event_init_context(child, ctxn); | |
7040 | if (ret) | |
7041 | return ret; | |
7042 | } | |
7043 | ||
7044 | return 0; | |
7045 | } | |
7046 | ||
220b140b PM |
7047 | static void __init perf_event_init_all_cpus(void) |
7048 | { | |
b28ab83c | 7049 | struct swevent_htable *swhash; |
220b140b | 7050 | int cpu; |
220b140b PM |
7051 | |
7052 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7053 | swhash = &per_cpu(swevent_htable, cpu); |
7054 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7055 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7056 | } |
7057 | } | |
7058 | ||
cdd6c482 | 7059 | static void __cpuinit perf_event_init_cpu(int cpu) |
0793a61d | 7060 | { |
108b02cf | 7061 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7062 | |
b28ab83c | 7063 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7064 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7065 | struct swevent_hlist *hlist; |
7066 | ||
b28ab83c PZ |
7067 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7068 | WARN_ON(!hlist); | |
7069 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7070 | } |
b28ab83c | 7071 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7072 | } |
7073 | ||
c277443c | 7074 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7075 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7076 | { |
e9d2b064 PZ |
7077 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7078 | ||
7079 | WARN_ON(!irqs_disabled()); | |
7080 | ||
7081 | list_del_init(&cpuctx->rotation_list); | |
7082 | } | |
7083 | ||
108b02cf | 7084 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7085 | { |
108b02cf | 7086 | struct perf_event_context *ctx = __info; |
cdd6c482 | 7087 | struct perf_event *event, *tmp; |
0793a61d | 7088 | |
108b02cf | 7089 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7090 | |
889ff015 | 7091 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 7092 | __perf_remove_from_context(event); |
889ff015 | 7093 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 7094 | __perf_remove_from_context(event); |
0793a61d | 7095 | } |
108b02cf PZ |
7096 | |
7097 | static void perf_event_exit_cpu_context(int cpu) | |
7098 | { | |
7099 | struct perf_event_context *ctx; | |
7100 | struct pmu *pmu; | |
7101 | int idx; | |
7102 | ||
7103 | idx = srcu_read_lock(&pmus_srcu); | |
7104 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7105 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7106 | |
7107 | mutex_lock(&ctx->mutex); | |
7108 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7109 | mutex_unlock(&ctx->mutex); | |
7110 | } | |
7111 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7112 | } |
7113 | ||
cdd6c482 | 7114 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7115 | { |
b28ab83c | 7116 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7117 | |
b28ab83c PZ |
7118 | mutex_lock(&swhash->hlist_mutex); |
7119 | swevent_hlist_release(swhash); | |
7120 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 7121 | |
108b02cf | 7122 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7123 | } |
7124 | #else | |
cdd6c482 | 7125 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7126 | #endif |
7127 | ||
c277443c PZ |
7128 | static int |
7129 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7130 | { | |
7131 | int cpu; | |
7132 | ||
7133 | for_each_online_cpu(cpu) | |
7134 | perf_event_exit_cpu(cpu); | |
7135 | ||
7136 | return NOTIFY_OK; | |
7137 | } | |
7138 | ||
7139 | /* | |
7140 | * Run the perf reboot notifier at the very last possible moment so that | |
7141 | * the generic watchdog code runs as long as possible. | |
7142 | */ | |
7143 | static struct notifier_block perf_reboot_notifier = { | |
7144 | .notifier_call = perf_reboot, | |
7145 | .priority = INT_MIN, | |
7146 | }; | |
7147 | ||
0793a61d TG |
7148 | static int __cpuinit |
7149 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
7150 | { | |
7151 | unsigned int cpu = (long)hcpu; | |
7152 | ||
4536e4d1 | 7153 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7154 | |
7155 | case CPU_UP_PREPARE: | |
5e11637e | 7156 | case CPU_DOWN_FAILED: |
cdd6c482 | 7157 | perf_event_init_cpu(cpu); |
0793a61d TG |
7158 | break; |
7159 | ||
5e11637e | 7160 | case CPU_UP_CANCELED: |
0793a61d | 7161 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7162 | perf_event_exit_cpu(cpu); |
0793a61d TG |
7163 | break; |
7164 | ||
7165 | default: | |
7166 | break; | |
7167 | } | |
7168 | ||
7169 | return NOTIFY_OK; | |
7170 | } | |
7171 | ||
cdd6c482 | 7172 | void __init perf_event_init(void) |
0793a61d | 7173 | { |
3c502e7a JW |
7174 | int ret; |
7175 | ||
2e80a82a PZ |
7176 | idr_init(&pmu_idr); |
7177 | ||
220b140b | 7178 | perf_event_init_all_cpus(); |
b0a873eb | 7179 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7180 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7181 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7182 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7183 | perf_tp_register(); |
7184 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7185 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7186 | |
7187 | ret = init_hw_breakpoint(); | |
7188 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7189 | |
7190 | /* do not patch jump label more than once per second */ | |
7191 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7192 | |
7193 | /* | |
7194 | * Build time assertion that we keep the data_head at the intended | |
7195 | * location. IOW, validation we got the __reserved[] size right. | |
7196 | */ | |
7197 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
7198 | != 1024); | |
0793a61d | 7199 | } |
abe43400 PZ |
7200 | |
7201 | static int __init perf_event_sysfs_init(void) | |
7202 | { | |
7203 | struct pmu *pmu; | |
7204 | int ret; | |
7205 | ||
7206 | mutex_lock(&pmus_lock); | |
7207 | ||
7208 | ret = bus_register(&pmu_bus); | |
7209 | if (ret) | |
7210 | goto unlock; | |
7211 | ||
7212 | list_for_each_entry(pmu, &pmus, entry) { | |
7213 | if (!pmu->name || pmu->type < 0) | |
7214 | continue; | |
7215 | ||
7216 | ret = pmu_dev_alloc(pmu); | |
7217 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
7218 | } | |
7219 | pmu_bus_running = 1; | |
7220 | ret = 0; | |
7221 | ||
7222 | unlock: | |
7223 | mutex_unlock(&pmus_lock); | |
7224 | ||
7225 | return ret; | |
7226 | } | |
7227 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
7228 | |
7229 | #ifdef CONFIG_CGROUP_PERF | |
761b3ef5 | 7230 | static struct cgroup_subsys_state *perf_cgroup_create(struct cgroup *cont) |
e5d1367f SE |
7231 | { |
7232 | struct perf_cgroup *jc; | |
e5d1367f | 7233 | |
1b15d055 | 7234 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
7235 | if (!jc) |
7236 | return ERR_PTR(-ENOMEM); | |
7237 | ||
e5d1367f SE |
7238 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7239 | if (!jc->info) { | |
7240 | kfree(jc); | |
7241 | return ERR_PTR(-ENOMEM); | |
7242 | } | |
7243 | ||
e5d1367f SE |
7244 | return &jc->css; |
7245 | } | |
7246 | ||
761b3ef5 | 7247 | static void perf_cgroup_destroy(struct cgroup *cont) |
e5d1367f SE |
7248 | { |
7249 | struct perf_cgroup *jc; | |
7250 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), | |
7251 | struct perf_cgroup, css); | |
7252 | free_percpu(jc->info); | |
7253 | kfree(jc); | |
7254 | } | |
7255 | ||
7256 | static int __perf_cgroup_move(void *info) | |
7257 | { | |
7258 | struct task_struct *task = info; | |
7259 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
7260 | return 0; | |
7261 | } | |
7262 | ||
761b3ef5 | 7263 | static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
e5d1367f | 7264 | { |
bb9d97b6 TH |
7265 | struct task_struct *task; |
7266 | ||
7267 | cgroup_taskset_for_each(task, cgrp, tset) | |
7268 | task_function_call(task, __perf_cgroup_move, task); | |
e5d1367f SE |
7269 | } |
7270 | ||
761b3ef5 LZ |
7271 | static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, |
7272 | struct task_struct *task) | |
e5d1367f SE |
7273 | { |
7274 | /* | |
7275 | * cgroup_exit() is called in the copy_process() failure path. | |
7276 | * Ignore this case since the task hasn't ran yet, this avoids | |
7277 | * trying to poke a half freed task state from generic code. | |
7278 | */ | |
7279 | if (!(task->flags & PF_EXITING)) | |
7280 | return; | |
7281 | ||
bb9d97b6 | 7282 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7283 | } |
7284 | ||
7285 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
7286 | .name = "perf_event", |
7287 | .subsys_id = perf_subsys_id, | |
7288 | .create = perf_cgroup_create, | |
7289 | .destroy = perf_cgroup_destroy, | |
7290 | .exit = perf_cgroup_exit, | |
bb9d97b6 | 7291 | .attach = perf_cgroup_attach, |
e5d1367f SE |
7292 | }; |
7293 | #endif /* CONFIG_CGROUP_PERF */ |