]>
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> |
c5ebcedb | 39 | #include <linux/mm_types.h> |
0793a61d | 40 | |
76369139 FW |
41 | #include "internal.h" |
42 | ||
4e193bd4 TB |
43 | #include <asm/irq_regs.h> |
44 | ||
fe4b04fa | 45 | struct remote_function_call { |
e7e7ee2e IM |
46 | struct task_struct *p; |
47 | int (*func)(void *info); | |
48 | void *info; | |
49 | int ret; | |
fe4b04fa PZ |
50 | }; |
51 | ||
52 | static void remote_function(void *data) | |
53 | { | |
54 | struct remote_function_call *tfc = data; | |
55 | struct task_struct *p = tfc->p; | |
56 | ||
57 | if (p) { | |
58 | tfc->ret = -EAGAIN; | |
59 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
60 | return; | |
61 | } | |
62 | ||
63 | tfc->ret = tfc->func(tfc->info); | |
64 | } | |
65 | ||
66 | /** | |
67 | * task_function_call - call a function on the cpu on which a task runs | |
68 | * @p: the task to evaluate | |
69 | * @func: the function to be called | |
70 | * @info: the function call argument | |
71 | * | |
72 | * Calls the function @func when the task is currently running. This might | |
73 | * be on the current CPU, which just calls the function directly | |
74 | * | |
75 | * returns: @func return value, or | |
76 | * -ESRCH - when the process isn't running | |
77 | * -EAGAIN - when the process moved away | |
78 | */ | |
79 | static int | |
80 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
81 | { | |
82 | struct remote_function_call data = { | |
e7e7ee2e IM |
83 | .p = p, |
84 | .func = func, | |
85 | .info = info, | |
86 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
87 | }; |
88 | ||
89 | if (task_curr(p)) | |
90 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
91 | ||
92 | return data.ret; | |
93 | } | |
94 | ||
95 | /** | |
96 | * cpu_function_call - call a function on the cpu | |
97 | * @func: the function to be called | |
98 | * @info: the function call argument | |
99 | * | |
100 | * Calls the function @func on the remote cpu. | |
101 | * | |
102 | * returns: @func return value or -ENXIO when the cpu is offline | |
103 | */ | |
104 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
105 | { | |
106 | struct remote_function_call data = { | |
e7e7ee2e IM |
107 | .p = NULL, |
108 | .func = func, | |
109 | .info = info, | |
110 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
111 | }; |
112 | ||
113 | smp_call_function_single(cpu, remote_function, &data, 1); | |
114 | ||
115 | return data.ret; | |
116 | } | |
117 | ||
e5d1367f SE |
118 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
119 | PERF_FLAG_FD_OUTPUT |\ | |
120 | PERF_FLAG_PID_CGROUP) | |
121 | ||
bce38cd5 SE |
122 | /* |
123 | * branch priv levels that need permission checks | |
124 | */ | |
125 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
126 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
127 | PERF_SAMPLE_BRANCH_HV) | |
128 | ||
0b3fcf17 SE |
129 | enum event_type_t { |
130 | EVENT_FLEXIBLE = 0x1, | |
131 | EVENT_PINNED = 0x2, | |
132 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
133 | }; | |
134 | ||
e5d1367f SE |
135 | /* |
136 | * perf_sched_events : >0 events exist | |
137 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
138 | */ | |
c5905afb | 139 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 140 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 141 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 142 | |
cdd6c482 IM |
143 | static atomic_t nr_mmap_events __read_mostly; |
144 | static atomic_t nr_comm_events __read_mostly; | |
145 | static atomic_t nr_task_events __read_mostly; | |
9ee318a7 | 146 | |
108b02cf PZ |
147 | static LIST_HEAD(pmus); |
148 | static DEFINE_MUTEX(pmus_lock); | |
149 | static struct srcu_struct pmus_srcu; | |
150 | ||
0764771d | 151 | /* |
cdd6c482 | 152 | * perf event paranoia level: |
0fbdea19 IM |
153 | * -1 - not paranoid at all |
154 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 155 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 156 | * 2 - disallow kernel profiling for unpriv |
0764771d | 157 | */ |
cdd6c482 | 158 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 159 | |
20443384 FW |
160 | /* Minimum for 512 kiB + 1 user control page */ |
161 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
162 | |
163 | /* | |
cdd6c482 | 164 | * max perf event sample rate |
df58ab24 | 165 | */ |
163ec435 PZ |
166 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
167 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
168 | static int max_samples_per_tick __read_mostly = | |
169 | DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
170 | ||
171 | int perf_proc_update_handler(struct ctl_table *table, int write, | |
172 | void __user *buffer, size_t *lenp, | |
173 | loff_t *ppos) | |
174 | { | |
175 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
176 | ||
177 | if (ret || !write) | |
178 | return ret; | |
179 | ||
180 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
181 | ||
182 | return 0; | |
183 | } | |
1ccd1549 | 184 | |
cdd6c482 | 185 | static atomic64_t perf_event_id; |
a96bbc16 | 186 | |
0b3fcf17 SE |
187 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
188 | enum event_type_t event_type); | |
189 | ||
190 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
191 | enum event_type_t event_type, |
192 | struct task_struct *task); | |
193 | ||
194 | static void update_context_time(struct perf_event_context *ctx); | |
195 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 196 | |
10c6db11 PZ |
197 | static void ring_buffer_attach(struct perf_event *event, |
198 | struct ring_buffer *rb); | |
199 | ||
cdd6c482 | 200 | void __weak perf_event_print_debug(void) { } |
0793a61d | 201 | |
84c79910 | 202 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 203 | { |
84c79910 | 204 | return "pmu"; |
0793a61d TG |
205 | } |
206 | ||
0b3fcf17 SE |
207 | static inline u64 perf_clock(void) |
208 | { | |
209 | return local_clock(); | |
210 | } | |
211 | ||
e5d1367f SE |
212 | static inline struct perf_cpu_context * |
213 | __get_cpu_context(struct perf_event_context *ctx) | |
214 | { | |
215 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
216 | } | |
217 | ||
facc4307 PZ |
218 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
219 | struct perf_event_context *ctx) | |
220 | { | |
221 | raw_spin_lock(&cpuctx->ctx.lock); | |
222 | if (ctx) | |
223 | raw_spin_lock(&ctx->lock); | |
224 | } | |
225 | ||
226 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
227 | struct perf_event_context *ctx) | |
228 | { | |
229 | if (ctx) | |
230 | raw_spin_unlock(&ctx->lock); | |
231 | raw_spin_unlock(&cpuctx->ctx.lock); | |
232 | } | |
233 | ||
e5d1367f SE |
234 | #ifdef CONFIG_CGROUP_PERF |
235 | ||
3f7cce3c SE |
236 | /* |
237 | * Must ensure cgroup is pinned (css_get) before calling | |
238 | * this function. In other words, we cannot call this function | |
239 | * if there is no cgroup event for the current CPU context. | |
240 | */ | |
e5d1367f SE |
241 | static inline struct perf_cgroup * |
242 | perf_cgroup_from_task(struct task_struct *task) | |
243 | { | |
244 | return container_of(task_subsys_state(task, perf_subsys_id), | |
245 | struct perf_cgroup, css); | |
246 | } | |
247 | ||
248 | static inline bool | |
249 | perf_cgroup_match(struct perf_event *event) | |
250 | { | |
251 | struct perf_event_context *ctx = event->ctx; | |
252 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
253 | ||
254 | return !event->cgrp || event->cgrp == cpuctx->cgrp; | |
255 | } | |
256 | ||
9c5da09d | 257 | static inline bool perf_tryget_cgroup(struct perf_event *event) |
e5d1367f | 258 | { |
9c5da09d | 259 | return css_tryget(&event->cgrp->css); |
e5d1367f SE |
260 | } |
261 | ||
262 | static inline void perf_put_cgroup(struct perf_event *event) | |
263 | { | |
264 | css_put(&event->cgrp->css); | |
265 | } | |
266 | ||
267 | static inline void perf_detach_cgroup(struct perf_event *event) | |
268 | { | |
269 | perf_put_cgroup(event); | |
270 | event->cgrp = NULL; | |
271 | } | |
272 | ||
273 | static inline int is_cgroup_event(struct perf_event *event) | |
274 | { | |
275 | return event->cgrp != NULL; | |
276 | } | |
277 | ||
278 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
279 | { | |
280 | struct perf_cgroup_info *t; | |
281 | ||
282 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
283 | return t->time; | |
284 | } | |
285 | ||
286 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
287 | { | |
288 | struct perf_cgroup_info *info; | |
289 | u64 now; | |
290 | ||
291 | now = perf_clock(); | |
292 | ||
293 | info = this_cpu_ptr(cgrp->info); | |
294 | ||
295 | info->time += now - info->timestamp; | |
296 | info->timestamp = now; | |
297 | } | |
298 | ||
299 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
300 | { | |
301 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
302 | if (cgrp_out) | |
303 | __update_cgrp_time(cgrp_out); | |
304 | } | |
305 | ||
306 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
307 | { | |
3f7cce3c SE |
308 | struct perf_cgroup *cgrp; |
309 | ||
e5d1367f | 310 | /* |
3f7cce3c SE |
311 | * ensure we access cgroup data only when needed and |
312 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 313 | */ |
3f7cce3c | 314 | if (!is_cgroup_event(event)) |
e5d1367f SE |
315 | return; |
316 | ||
3f7cce3c SE |
317 | cgrp = perf_cgroup_from_task(current); |
318 | /* | |
319 | * Do not update time when cgroup is not active | |
320 | */ | |
321 | if (cgrp == event->cgrp) | |
322 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
323 | } |
324 | ||
325 | static inline void | |
3f7cce3c SE |
326 | perf_cgroup_set_timestamp(struct task_struct *task, |
327 | struct perf_event_context *ctx) | |
e5d1367f SE |
328 | { |
329 | struct perf_cgroup *cgrp; | |
330 | struct perf_cgroup_info *info; | |
331 | ||
3f7cce3c SE |
332 | /* |
333 | * ctx->lock held by caller | |
334 | * ensure we do not access cgroup data | |
335 | * unless we have the cgroup pinned (css_get) | |
336 | */ | |
337 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
338 | return; |
339 | ||
340 | cgrp = perf_cgroup_from_task(task); | |
341 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 342 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
343 | } |
344 | ||
345 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
346 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
347 | ||
348 | /* | |
349 | * reschedule events based on the cgroup constraint of task. | |
350 | * | |
351 | * mode SWOUT : schedule out everything | |
352 | * mode SWIN : schedule in based on cgroup for next | |
353 | */ | |
354 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
355 | { | |
356 | struct perf_cpu_context *cpuctx; | |
357 | struct pmu *pmu; | |
358 | unsigned long flags; | |
359 | ||
360 | /* | |
361 | * disable interrupts to avoid geting nr_cgroup | |
362 | * changes via __perf_event_disable(). Also | |
363 | * avoids preemption. | |
364 | */ | |
365 | local_irq_save(flags); | |
366 | ||
367 | /* | |
368 | * we reschedule only in the presence of cgroup | |
369 | * constrained events. | |
370 | */ | |
371 | rcu_read_lock(); | |
372 | ||
373 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f SE |
374 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
375 | ||
e5d1367f SE |
376 | /* |
377 | * perf_cgroup_events says at least one | |
378 | * context on this CPU has cgroup events. | |
379 | * | |
380 | * ctx->nr_cgroups reports the number of cgroup | |
381 | * events for a context. | |
382 | */ | |
383 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
384 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
385 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
386 | |
387 | if (mode & PERF_CGROUP_SWOUT) { | |
388 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
389 | /* | |
390 | * must not be done before ctxswout due | |
391 | * to event_filter_match() in event_sched_out() | |
392 | */ | |
393 | cpuctx->cgrp = NULL; | |
394 | } | |
395 | ||
396 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 397 | WARN_ON_ONCE(cpuctx->cgrp); |
e5d1367f SE |
398 | /* set cgrp before ctxsw in to |
399 | * allow event_filter_match() to not | |
400 | * have to pass task around | |
401 | */ | |
402 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
403 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
404 | } | |
facc4307 PZ |
405 | perf_pmu_enable(cpuctx->ctx.pmu); |
406 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 407 | } |
e5d1367f SE |
408 | } |
409 | ||
410 | rcu_read_unlock(); | |
411 | ||
412 | local_irq_restore(flags); | |
413 | } | |
414 | ||
a8d757ef SE |
415 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
416 | struct task_struct *next) | |
e5d1367f | 417 | { |
a8d757ef SE |
418 | struct perf_cgroup *cgrp1; |
419 | struct perf_cgroup *cgrp2 = NULL; | |
420 | ||
421 | /* | |
422 | * we come here when we know perf_cgroup_events > 0 | |
423 | */ | |
424 | cgrp1 = perf_cgroup_from_task(task); | |
425 | ||
426 | /* | |
427 | * next is NULL when called from perf_event_enable_on_exec() | |
428 | * that will systematically cause a cgroup_switch() | |
429 | */ | |
430 | if (next) | |
431 | cgrp2 = perf_cgroup_from_task(next); | |
432 | ||
433 | /* | |
434 | * only schedule out current cgroup events if we know | |
435 | * that we are switching to a different cgroup. Otherwise, | |
436 | * do no touch the cgroup events. | |
437 | */ | |
438 | if (cgrp1 != cgrp2) | |
439 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
440 | } |
441 | ||
a8d757ef SE |
442 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
443 | struct task_struct *task) | |
e5d1367f | 444 | { |
a8d757ef SE |
445 | struct perf_cgroup *cgrp1; |
446 | struct perf_cgroup *cgrp2 = NULL; | |
447 | ||
448 | /* | |
449 | * we come here when we know perf_cgroup_events > 0 | |
450 | */ | |
451 | cgrp1 = perf_cgroup_from_task(task); | |
452 | ||
453 | /* prev can never be NULL */ | |
454 | cgrp2 = perf_cgroup_from_task(prev); | |
455 | ||
456 | /* | |
457 | * only need to schedule in cgroup events if we are changing | |
458 | * cgroup during ctxsw. Cgroup events were not scheduled | |
459 | * out of ctxsw out if that was not the case. | |
460 | */ | |
461 | if (cgrp1 != cgrp2) | |
462 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
463 | } |
464 | ||
465 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
466 | struct perf_event_attr *attr, | |
467 | struct perf_event *group_leader) | |
468 | { | |
469 | struct perf_cgroup *cgrp; | |
470 | struct cgroup_subsys_state *css; | |
471 | struct file *file; | |
472 | int ret = 0, fput_needed; | |
473 | ||
474 | file = fget_light(fd, &fput_needed); | |
475 | if (!file) | |
476 | return -EBADF; | |
477 | ||
478 | css = cgroup_css_from_dir(file, perf_subsys_id); | |
3db272c0 LZ |
479 | if (IS_ERR(css)) { |
480 | ret = PTR_ERR(css); | |
481 | goto out; | |
482 | } | |
e5d1367f SE |
483 | |
484 | cgrp = container_of(css, struct perf_cgroup, css); | |
485 | event->cgrp = cgrp; | |
486 | ||
f75e18cb | 487 | /* must be done before we fput() the file */ |
9c5da09d SQ |
488 | if (!perf_tryget_cgroup(event)) { |
489 | event->cgrp = NULL; | |
490 | ret = -ENOENT; | |
491 | goto out; | |
492 | } | |
f75e18cb | 493 | |
e5d1367f SE |
494 | /* |
495 | * all events in a group must monitor | |
496 | * the same cgroup because a task belongs | |
497 | * to only one perf cgroup at a time | |
498 | */ | |
499 | if (group_leader && group_leader->cgrp != cgrp) { | |
500 | perf_detach_cgroup(event); | |
501 | ret = -EINVAL; | |
e5d1367f | 502 | } |
3db272c0 | 503 | out: |
e5d1367f SE |
504 | fput_light(file, fput_needed); |
505 | return ret; | |
506 | } | |
507 | ||
508 | static inline void | |
509 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
510 | { | |
511 | struct perf_cgroup_info *t; | |
512 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
513 | event->shadow_ctx_time = now - t->timestamp; | |
514 | } | |
515 | ||
516 | static inline void | |
517 | perf_cgroup_defer_enabled(struct perf_event *event) | |
518 | { | |
519 | /* | |
520 | * when the current task's perf cgroup does not match | |
521 | * the event's, we need to remember to call the | |
522 | * perf_mark_enable() function the first time a task with | |
523 | * a matching perf cgroup is scheduled in. | |
524 | */ | |
525 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
526 | event->cgrp_defer_enabled = 1; | |
527 | } | |
528 | ||
529 | static inline void | |
530 | perf_cgroup_mark_enabled(struct perf_event *event, | |
531 | struct perf_event_context *ctx) | |
532 | { | |
533 | struct perf_event *sub; | |
534 | u64 tstamp = perf_event_time(event); | |
535 | ||
536 | if (!event->cgrp_defer_enabled) | |
537 | return; | |
538 | ||
539 | event->cgrp_defer_enabled = 0; | |
540 | ||
541 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
542 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
543 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
544 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
545 | sub->cgrp_defer_enabled = 0; | |
546 | } | |
547 | } | |
548 | } | |
549 | #else /* !CONFIG_CGROUP_PERF */ | |
550 | ||
551 | static inline bool | |
552 | perf_cgroup_match(struct perf_event *event) | |
553 | { | |
554 | return true; | |
555 | } | |
556 | ||
557 | static inline void perf_detach_cgroup(struct perf_event *event) | |
558 | {} | |
559 | ||
560 | static inline int is_cgroup_event(struct perf_event *event) | |
561 | { | |
562 | return 0; | |
563 | } | |
564 | ||
565 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
566 | { | |
567 | return 0; | |
568 | } | |
569 | ||
570 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
571 | { | |
572 | } | |
573 | ||
574 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
575 | { | |
576 | } | |
577 | ||
a8d757ef SE |
578 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
579 | struct task_struct *next) | |
e5d1367f SE |
580 | { |
581 | } | |
582 | ||
a8d757ef SE |
583 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
584 | struct task_struct *task) | |
e5d1367f SE |
585 | { |
586 | } | |
587 | ||
588 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
589 | struct perf_event_attr *attr, | |
590 | struct perf_event *group_leader) | |
591 | { | |
592 | return -EINVAL; | |
593 | } | |
594 | ||
595 | static inline void | |
3f7cce3c SE |
596 | perf_cgroup_set_timestamp(struct task_struct *task, |
597 | struct perf_event_context *ctx) | |
e5d1367f SE |
598 | { |
599 | } | |
600 | ||
601 | void | |
602 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
603 | { | |
604 | } | |
605 | ||
606 | static inline void | |
607 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
608 | { | |
609 | } | |
610 | ||
611 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
612 | { | |
613 | return 0; | |
614 | } | |
615 | ||
616 | static inline void | |
617 | perf_cgroup_defer_enabled(struct perf_event *event) | |
618 | { | |
619 | } | |
620 | ||
621 | static inline void | |
622 | perf_cgroup_mark_enabled(struct perf_event *event, | |
623 | struct perf_event_context *ctx) | |
624 | { | |
625 | } | |
626 | #endif | |
627 | ||
33696fc0 | 628 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 629 | { |
33696fc0 PZ |
630 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
631 | if (!(*count)++) | |
632 | pmu->pmu_disable(pmu); | |
9e35ad38 | 633 | } |
9e35ad38 | 634 | |
33696fc0 | 635 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 636 | { |
33696fc0 PZ |
637 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
638 | if (!--(*count)) | |
639 | pmu->pmu_enable(pmu); | |
9e35ad38 | 640 | } |
9e35ad38 | 641 | |
e9d2b064 PZ |
642 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
643 | ||
644 | /* | |
645 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
646 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
647 | * disabled, while rotate_context is called from IRQ context. | |
648 | */ | |
108b02cf | 649 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 650 | { |
108b02cf | 651 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 652 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 653 | |
e9d2b064 | 654 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 655 | |
e9d2b064 PZ |
656 | if (list_empty(&cpuctx->rotation_list)) |
657 | list_add(&cpuctx->rotation_list, head); | |
9e35ad38 | 658 | } |
9e35ad38 | 659 | |
cdd6c482 | 660 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 661 | { |
e5289d4a | 662 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
663 | } |
664 | ||
cdd6c482 | 665 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 666 | { |
564c2b21 PM |
667 | if (atomic_dec_and_test(&ctx->refcount)) { |
668 | if (ctx->parent_ctx) | |
669 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
670 | if (ctx->task) |
671 | put_task_struct(ctx->task); | |
cb796ff3 | 672 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 673 | } |
a63eaf34 PM |
674 | } |
675 | ||
cdd6c482 | 676 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
677 | { |
678 | if (ctx->parent_ctx) { | |
679 | put_ctx(ctx->parent_ctx); | |
680 | ctx->parent_ctx = NULL; | |
681 | } | |
682 | } | |
683 | ||
6844c09d ACM |
684 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
685 | { | |
686 | /* | |
687 | * only top level events have the pid namespace they were created in | |
688 | */ | |
689 | if (event->parent) | |
690 | event = event->parent; | |
691 | ||
692 | return task_tgid_nr_ns(p, event->ns); | |
693 | } | |
694 | ||
695 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
696 | { | |
697 | /* | |
698 | * only top level events have the pid namespace they were created in | |
699 | */ | |
700 | if (event->parent) | |
701 | event = event->parent; | |
702 | ||
703 | return task_pid_nr_ns(p, event->ns); | |
704 | } | |
705 | ||
7f453c24 | 706 | /* |
cdd6c482 | 707 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
708 | * to userspace. |
709 | */ | |
cdd6c482 | 710 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 711 | { |
cdd6c482 | 712 | u64 id = event->id; |
7f453c24 | 713 | |
cdd6c482 IM |
714 | if (event->parent) |
715 | id = event->parent->id; | |
7f453c24 PZ |
716 | |
717 | return id; | |
718 | } | |
719 | ||
25346b93 | 720 | /* |
cdd6c482 | 721 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
722 | * This has to cope with with the fact that until it is locked, |
723 | * the context could get moved to another task. | |
724 | */ | |
cdd6c482 | 725 | static struct perf_event_context * |
8dc85d54 | 726 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 727 | { |
cdd6c482 | 728 | struct perf_event_context *ctx; |
25346b93 PM |
729 | |
730 | rcu_read_lock(); | |
9ed6060d | 731 | retry: |
8dc85d54 | 732 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
733 | if (ctx) { |
734 | /* | |
735 | * If this context is a clone of another, it might | |
736 | * get swapped for another underneath us by | |
cdd6c482 | 737 | * perf_event_task_sched_out, though the |
25346b93 PM |
738 | * rcu_read_lock() protects us from any context |
739 | * getting freed. Lock the context and check if it | |
740 | * got swapped before we could get the lock, and retry | |
741 | * if so. If we locked the right context, then it | |
742 | * can't get swapped on us any more. | |
743 | */ | |
e625cce1 | 744 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 745 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 746 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
25346b93 PM |
747 | goto retry; |
748 | } | |
b49a9e7e PZ |
749 | |
750 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 751 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
752 | ctx = NULL; |
753 | } | |
25346b93 PM |
754 | } |
755 | rcu_read_unlock(); | |
756 | return ctx; | |
757 | } | |
758 | ||
759 | /* | |
760 | * Get the context for a task and increment its pin_count so it | |
761 | * can't get swapped to another task. This also increments its | |
762 | * reference count so that the context can't get freed. | |
763 | */ | |
8dc85d54 PZ |
764 | static struct perf_event_context * |
765 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 766 | { |
cdd6c482 | 767 | struct perf_event_context *ctx; |
25346b93 PM |
768 | unsigned long flags; |
769 | ||
8dc85d54 | 770 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
771 | if (ctx) { |
772 | ++ctx->pin_count; | |
e625cce1 | 773 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
774 | } |
775 | return ctx; | |
776 | } | |
777 | ||
cdd6c482 | 778 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
779 | { |
780 | unsigned long flags; | |
781 | ||
e625cce1 | 782 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 783 | --ctx->pin_count; |
e625cce1 | 784 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
785 | } |
786 | ||
f67218c3 PZ |
787 | /* |
788 | * Update the record of the current time in a context. | |
789 | */ | |
790 | static void update_context_time(struct perf_event_context *ctx) | |
791 | { | |
792 | u64 now = perf_clock(); | |
793 | ||
794 | ctx->time += now - ctx->timestamp; | |
795 | ctx->timestamp = now; | |
796 | } | |
797 | ||
4158755d SE |
798 | static u64 perf_event_time(struct perf_event *event) |
799 | { | |
800 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
801 | |
802 | if (is_cgroup_event(event)) | |
803 | return perf_cgroup_event_time(event); | |
804 | ||
4158755d SE |
805 | return ctx ? ctx->time : 0; |
806 | } | |
807 | ||
f67218c3 PZ |
808 | /* |
809 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 810 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
811 | */ |
812 | static void update_event_times(struct perf_event *event) | |
813 | { | |
814 | struct perf_event_context *ctx = event->ctx; | |
815 | u64 run_end; | |
816 | ||
817 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
818 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
819 | return; | |
e5d1367f SE |
820 | /* |
821 | * in cgroup mode, time_enabled represents | |
822 | * the time the event was enabled AND active | |
823 | * tasks were in the monitored cgroup. This is | |
824 | * independent of the activity of the context as | |
825 | * there may be a mix of cgroup and non-cgroup events. | |
826 | * | |
827 | * That is why we treat cgroup events differently | |
828 | * here. | |
829 | */ | |
830 | if (is_cgroup_event(event)) | |
46cd6a7f | 831 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
832 | else if (ctx->is_active) |
833 | run_end = ctx->time; | |
acd1d7c1 PZ |
834 | else |
835 | run_end = event->tstamp_stopped; | |
836 | ||
837 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
838 | |
839 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
840 | run_end = event->tstamp_stopped; | |
841 | else | |
4158755d | 842 | run_end = perf_event_time(event); |
f67218c3 PZ |
843 | |
844 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 845 | |
f67218c3 PZ |
846 | } |
847 | ||
96c21a46 PZ |
848 | /* |
849 | * Update total_time_enabled and total_time_running for all events in a group. | |
850 | */ | |
851 | static void update_group_times(struct perf_event *leader) | |
852 | { | |
853 | struct perf_event *event; | |
854 | ||
855 | update_event_times(leader); | |
856 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
857 | update_event_times(event); | |
858 | } | |
859 | ||
889ff015 FW |
860 | static struct list_head * |
861 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
862 | { | |
863 | if (event->attr.pinned) | |
864 | return &ctx->pinned_groups; | |
865 | else | |
866 | return &ctx->flexible_groups; | |
867 | } | |
868 | ||
fccc714b | 869 | /* |
cdd6c482 | 870 | * Add a event from the lists for its context. |
fccc714b PZ |
871 | * Must be called with ctx->mutex and ctx->lock held. |
872 | */ | |
04289bb9 | 873 | static void |
cdd6c482 | 874 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 875 | { |
8a49542c PZ |
876 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
877 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
878 | |
879 | /* | |
8a49542c PZ |
880 | * If we're a stand alone event or group leader, we go to the context |
881 | * list, group events are kept attached to the group so that | |
882 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 883 | */ |
8a49542c | 884 | if (event->group_leader == event) { |
889ff015 FW |
885 | struct list_head *list; |
886 | ||
d6f962b5 FW |
887 | if (is_software_event(event)) |
888 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
889 | ||
889ff015 FW |
890 | list = ctx_group_list(event, ctx); |
891 | list_add_tail(&event->group_entry, list); | |
5c148194 | 892 | } |
592903cd | 893 | |
08309379 | 894 | if (is_cgroup_event(event)) |
e5d1367f | 895 | ctx->nr_cgroups++; |
e5d1367f | 896 | |
d010b332 SE |
897 | if (has_branch_stack(event)) |
898 | ctx->nr_branch_stack++; | |
899 | ||
cdd6c482 | 900 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 901 | if (!ctx->nr_events) |
108b02cf | 902 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
903 | ctx->nr_events++; |
904 | if (event->attr.inherit_stat) | |
bfbd3381 | 905 | ctx->nr_stat++; |
04289bb9 IM |
906 | } |
907 | ||
c320c7b7 ACM |
908 | /* |
909 | * Called at perf_event creation and when events are attached/detached from a | |
910 | * group. | |
911 | */ | |
912 | static void perf_event__read_size(struct perf_event *event) | |
913 | { | |
914 | int entry = sizeof(u64); /* value */ | |
915 | int size = 0; | |
916 | int nr = 1; | |
917 | ||
918 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
919 | size += sizeof(u64); | |
920 | ||
921 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
922 | size += sizeof(u64); | |
923 | ||
924 | if (event->attr.read_format & PERF_FORMAT_ID) | |
925 | entry += sizeof(u64); | |
926 | ||
927 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
928 | nr += event->group_leader->nr_siblings; | |
929 | size += sizeof(u64); | |
930 | } | |
931 | ||
932 | size += entry * nr; | |
933 | event->read_size = size; | |
934 | } | |
935 | ||
936 | static void perf_event__header_size(struct perf_event *event) | |
937 | { | |
938 | struct perf_sample_data *data; | |
939 | u64 sample_type = event->attr.sample_type; | |
940 | u16 size = 0; | |
941 | ||
942 | perf_event__read_size(event); | |
943 | ||
944 | if (sample_type & PERF_SAMPLE_IP) | |
945 | size += sizeof(data->ip); | |
946 | ||
6844c09d ACM |
947 | if (sample_type & PERF_SAMPLE_ADDR) |
948 | size += sizeof(data->addr); | |
949 | ||
950 | if (sample_type & PERF_SAMPLE_PERIOD) | |
951 | size += sizeof(data->period); | |
952 | ||
953 | if (sample_type & PERF_SAMPLE_READ) | |
954 | size += event->read_size; | |
955 | ||
956 | event->header_size = size; | |
957 | } | |
958 | ||
959 | static void perf_event__id_header_size(struct perf_event *event) | |
960 | { | |
961 | struct perf_sample_data *data; | |
962 | u64 sample_type = event->attr.sample_type; | |
963 | u16 size = 0; | |
964 | ||
c320c7b7 ACM |
965 | if (sample_type & PERF_SAMPLE_TID) |
966 | size += sizeof(data->tid_entry); | |
967 | ||
968 | if (sample_type & PERF_SAMPLE_TIME) | |
969 | size += sizeof(data->time); | |
970 | ||
c320c7b7 ACM |
971 | if (sample_type & PERF_SAMPLE_ID) |
972 | size += sizeof(data->id); | |
973 | ||
974 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
975 | size += sizeof(data->stream_id); | |
976 | ||
977 | if (sample_type & PERF_SAMPLE_CPU) | |
978 | size += sizeof(data->cpu_entry); | |
979 | ||
6844c09d | 980 | event->id_header_size = size; |
c320c7b7 ACM |
981 | } |
982 | ||
8a49542c PZ |
983 | static void perf_group_attach(struct perf_event *event) |
984 | { | |
c320c7b7 | 985 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 986 | |
74c3337c PZ |
987 | /* |
988 | * We can have double attach due to group movement in perf_event_open. | |
989 | */ | |
990 | if (event->attach_state & PERF_ATTACH_GROUP) | |
991 | return; | |
992 | ||
8a49542c PZ |
993 | event->attach_state |= PERF_ATTACH_GROUP; |
994 | ||
995 | if (group_leader == event) | |
996 | return; | |
997 | ||
998 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
999 | !is_software_event(event)) | |
1000 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1001 | ||
1002 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1003 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1004 | |
1005 | perf_event__header_size(group_leader); | |
1006 | ||
1007 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1008 | perf_event__header_size(pos); | |
8a49542c PZ |
1009 | } |
1010 | ||
a63eaf34 | 1011 | /* |
cdd6c482 | 1012 | * Remove a event from the lists for its context. |
fccc714b | 1013 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1014 | */ |
04289bb9 | 1015 | static void |
cdd6c482 | 1016 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1017 | { |
68cacd29 | 1018 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1019 | /* |
1020 | * We can have double detach due to exit/hot-unplug + close. | |
1021 | */ | |
1022 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1023 | return; |
8a49542c PZ |
1024 | |
1025 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1026 | ||
68cacd29 | 1027 | if (is_cgroup_event(event)) { |
e5d1367f | 1028 | ctx->nr_cgroups--; |
68cacd29 SE |
1029 | cpuctx = __get_cpu_context(ctx); |
1030 | /* | |
1031 | * if there are no more cgroup events | |
1032 | * then cler cgrp to avoid stale pointer | |
1033 | * in update_cgrp_time_from_cpuctx() | |
1034 | */ | |
1035 | if (!ctx->nr_cgroups) | |
1036 | cpuctx->cgrp = NULL; | |
1037 | } | |
e5d1367f | 1038 | |
d010b332 SE |
1039 | if (has_branch_stack(event)) |
1040 | ctx->nr_branch_stack--; | |
1041 | ||
cdd6c482 IM |
1042 | ctx->nr_events--; |
1043 | if (event->attr.inherit_stat) | |
bfbd3381 | 1044 | ctx->nr_stat--; |
8bc20959 | 1045 | |
cdd6c482 | 1046 | list_del_rcu(&event->event_entry); |
04289bb9 | 1047 | |
8a49542c PZ |
1048 | if (event->group_leader == event) |
1049 | list_del_init(&event->group_entry); | |
5c148194 | 1050 | |
96c21a46 | 1051 | update_group_times(event); |
b2e74a26 SE |
1052 | |
1053 | /* | |
1054 | * If event was in error state, then keep it | |
1055 | * that way, otherwise bogus counts will be | |
1056 | * returned on read(). The only way to get out | |
1057 | * of error state is by explicit re-enabling | |
1058 | * of the event | |
1059 | */ | |
1060 | if (event->state > PERF_EVENT_STATE_OFF) | |
1061 | event->state = PERF_EVENT_STATE_OFF; | |
050735b0 PZ |
1062 | } |
1063 | ||
8a49542c | 1064 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1065 | { |
1066 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1067 | struct list_head *list = NULL; |
1068 | ||
1069 | /* | |
1070 | * We can have double detach due to exit/hot-unplug + close. | |
1071 | */ | |
1072 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1073 | return; | |
1074 | ||
1075 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1076 | ||
1077 | /* | |
1078 | * If this is a sibling, remove it from its group. | |
1079 | */ | |
1080 | if (event->group_leader != event) { | |
1081 | list_del_init(&event->group_entry); | |
1082 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1083 | goto out; |
8a49542c PZ |
1084 | } |
1085 | ||
1086 | if (!list_empty(&event->group_entry)) | |
1087 | list = &event->group_entry; | |
2e2af50b | 1088 | |
04289bb9 | 1089 | /* |
cdd6c482 IM |
1090 | * If this was a group event with sibling events then |
1091 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1092 | * to whatever list we are on. |
04289bb9 | 1093 | */ |
cdd6c482 | 1094 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1095 | if (list) |
1096 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1097 | sibling->group_leader = sibling; |
d6f962b5 FW |
1098 | |
1099 | /* Inherit group flags from the previous leader */ | |
1100 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1101 | } |
c320c7b7 ACM |
1102 | |
1103 | out: | |
1104 | perf_event__header_size(event->group_leader); | |
1105 | ||
1106 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1107 | perf_event__header_size(tmp); | |
04289bb9 IM |
1108 | } |
1109 | ||
fa66f07a SE |
1110 | static inline int |
1111 | event_filter_match(struct perf_event *event) | |
1112 | { | |
e5d1367f SE |
1113 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1114 | && perf_cgroup_match(event); | |
fa66f07a SE |
1115 | } |
1116 | ||
9ffcfa6f SE |
1117 | static void |
1118 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1119 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1120 | struct perf_event_context *ctx) |
3b6f9e5c | 1121 | { |
4158755d | 1122 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1123 | u64 delta; |
1124 | /* | |
1125 | * An event which could not be activated because of | |
1126 | * filter mismatch still needs to have its timings | |
1127 | * maintained, otherwise bogus information is return | |
1128 | * via read() for time_enabled, time_running: | |
1129 | */ | |
1130 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1131 | && !event_filter_match(event)) { | |
e5d1367f | 1132 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1133 | event->tstamp_running += delta; |
4158755d | 1134 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1135 | } |
1136 | ||
cdd6c482 | 1137 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1138 | return; |
3b6f9e5c | 1139 | |
cdd6c482 IM |
1140 | event->state = PERF_EVENT_STATE_INACTIVE; |
1141 | if (event->pending_disable) { | |
1142 | event->pending_disable = 0; | |
1143 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1144 | } |
4158755d | 1145 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1146 | event->pmu->del(event, 0); |
cdd6c482 | 1147 | event->oncpu = -1; |
3b6f9e5c | 1148 | |
cdd6c482 | 1149 | if (!is_software_event(event)) |
3b6f9e5c PM |
1150 | cpuctx->active_oncpu--; |
1151 | ctx->nr_active--; | |
0f5a2601 PZ |
1152 | if (event->attr.freq && event->attr.sample_freq) |
1153 | ctx->nr_freq--; | |
cdd6c482 | 1154 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c PM |
1155 | cpuctx->exclusive = 0; |
1156 | } | |
1157 | ||
d859e29f | 1158 | static void |
cdd6c482 | 1159 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1160 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1161 | struct perf_event_context *ctx) |
d859e29f | 1162 | { |
cdd6c482 | 1163 | struct perf_event *event; |
fa66f07a | 1164 | int state = group_event->state; |
d859e29f | 1165 | |
cdd6c482 | 1166 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1167 | |
1168 | /* | |
1169 | * Schedule out siblings (if any): | |
1170 | */ | |
cdd6c482 IM |
1171 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1172 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1173 | |
fa66f07a | 1174 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1175 | cpuctx->exclusive = 0; |
1176 | } | |
1177 | ||
0793a61d | 1178 | /* |
cdd6c482 | 1179 | * Cross CPU call to remove a performance event |
0793a61d | 1180 | * |
cdd6c482 | 1181 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1182 | * remove it from the context list. |
1183 | */ | |
fe4b04fa | 1184 | static int __perf_remove_from_context(void *info) |
0793a61d | 1185 | { |
cdd6c482 IM |
1186 | struct perf_event *event = info; |
1187 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1188 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1189 | |
e625cce1 | 1190 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1191 | event_sched_out(event, cpuctx, ctx); |
cdd6c482 | 1192 | list_del_event(event, ctx); |
64ce3126 PZ |
1193 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1194 | ctx->is_active = 0; | |
1195 | cpuctx->task_ctx = NULL; | |
1196 | } | |
e625cce1 | 1197 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1198 | |
1199 | return 0; | |
0793a61d TG |
1200 | } |
1201 | ||
1202 | ||
1203 | /* | |
cdd6c482 | 1204 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1205 | * |
cdd6c482 | 1206 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1207 | * call when the task is on a CPU. |
c93f7669 | 1208 | * |
cdd6c482 IM |
1209 | * If event->ctx is a cloned context, callers must make sure that |
1210 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1211 | * remains valid. This is OK when called from perf_release since |
1212 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1213 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1214 | * context has been detached from its task. |
0793a61d | 1215 | */ |
fe4b04fa | 1216 | static void perf_remove_from_context(struct perf_event *event) |
0793a61d | 1217 | { |
cdd6c482 | 1218 | struct perf_event_context *ctx = event->ctx; |
0793a61d TG |
1219 | struct task_struct *task = ctx->task; |
1220 | ||
fe4b04fa PZ |
1221 | lockdep_assert_held(&ctx->mutex); |
1222 | ||
0793a61d TG |
1223 | if (!task) { |
1224 | /* | |
cdd6c482 | 1225 | * Per cpu events are removed via an smp call and |
af901ca1 | 1226 | * the removal is always successful. |
0793a61d | 1227 | */ |
fe4b04fa | 1228 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
0793a61d TG |
1229 | return; |
1230 | } | |
1231 | ||
1232 | retry: | |
fe4b04fa PZ |
1233 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1234 | return; | |
0793a61d | 1235 | |
e625cce1 | 1236 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1237 | /* |
fe4b04fa PZ |
1238 | * If we failed to find a running task, but find the context active now |
1239 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1240 | */ |
fe4b04fa | 1241 | if (ctx->is_active) { |
e625cce1 | 1242 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1243 | goto retry; |
1244 | } | |
1245 | ||
1246 | /* | |
fe4b04fa PZ |
1247 | * Since the task isn't running, its safe to remove the event, us |
1248 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1249 | */ |
fe4b04fa | 1250 | list_del_event(event, ctx); |
e625cce1 | 1251 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1252 | } |
1253 | ||
d859e29f | 1254 | /* |
cdd6c482 | 1255 | * Cross CPU call to disable a performance event |
d859e29f | 1256 | */ |
500ad2d8 | 1257 | int __perf_event_disable(void *info) |
d859e29f | 1258 | { |
cdd6c482 | 1259 | struct perf_event *event = info; |
cdd6c482 | 1260 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1261 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1262 | |
1263 | /* | |
cdd6c482 IM |
1264 | * If this is a per-task event, need to check whether this |
1265 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1266 | * |
1267 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1268 | * flipping contexts around. | |
d859e29f | 1269 | */ |
665c2142 | 1270 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1271 | return -EINVAL; |
d859e29f | 1272 | |
e625cce1 | 1273 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1274 | |
1275 | /* | |
cdd6c482 | 1276 | * If the event is on, turn it off. |
d859e29f PM |
1277 | * If it is in error state, leave it in error state. |
1278 | */ | |
cdd6c482 | 1279 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1280 | update_context_time(ctx); |
e5d1367f | 1281 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1282 | update_group_times(event); |
1283 | if (event == event->group_leader) | |
1284 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1285 | else |
cdd6c482 IM |
1286 | event_sched_out(event, cpuctx, ctx); |
1287 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1288 | } |
1289 | ||
e625cce1 | 1290 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1291 | |
1292 | return 0; | |
d859e29f PM |
1293 | } |
1294 | ||
1295 | /* | |
cdd6c482 | 1296 | * Disable a event. |
c93f7669 | 1297 | * |
cdd6c482 IM |
1298 | * If event->ctx is a cloned context, callers must make sure that |
1299 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1300 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1301 | * perf_event_for_each_child or perf_event_for_each because they |
1302 | * hold the top-level event's child_mutex, so any descendant that | |
1303 | * goes to exit will block in sync_child_event. | |
1304 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1305 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1306 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1307 | */ |
44234adc | 1308 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1309 | { |
cdd6c482 | 1310 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1311 | struct task_struct *task = ctx->task; |
1312 | ||
1313 | if (!task) { | |
1314 | /* | |
cdd6c482 | 1315 | * Disable the event on the cpu that it's on |
d859e29f | 1316 | */ |
fe4b04fa | 1317 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1318 | return; |
1319 | } | |
1320 | ||
9ed6060d | 1321 | retry: |
fe4b04fa PZ |
1322 | if (!task_function_call(task, __perf_event_disable, event)) |
1323 | return; | |
d859e29f | 1324 | |
e625cce1 | 1325 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1326 | /* |
cdd6c482 | 1327 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1328 | */ |
cdd6c482 | 1329 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1330 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1331 | /* |
1332 | * Reload the task pointer, it might have been changed by | |
1333 | * a concurrent perf_event_context_sched_out(). | |
1334 | */ | |
1335 | task = ctx->task; | |
d859e29f PM |
1336 | goto retry; |
1337 | } | |
1338 | ||
1339 | /* | |
1340 | * Since we have the lock this context can't be scheduled | |
1341 | * in, so we can change the state safely. | |
1342 | */ | |
cdd6c482 IM |
1343 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1344 | update_group_times(event); | |
1345 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1346 | } |
e625cce1 | 1347 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1348 | } |
dcfce4a0 | 1349 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1350 | |
e5d1367f SE |
1351 | static void perf_set_shadow_time(struct perf_event *event, |
1352 | struct perf_event_context *ctx, | |
1353 | u64 tstamp) | |
1354 | { | |
1355 | /* | |
1356 | * use the correct time source for the time snapshot | |
1357 | * | |
1358 | * We could get by without this by leveraging the | |
1359 | * fact that to get to this function, the caller | |
1360 | * has most likely already called update_context_time() | |
1361 | * and update_cgrp_time_xx() and thus both timestamp | |
1362 | * are identical (or very close). Given that tstamp is, | |
1363 | * already adjusted for cgroup, we could say that: | |
1364 | * tstamp - ctx->timestamp | |
1365 | * is equivalent to | |
1366 | * tstamp - cgrp->timestamp. | |
1367 | * | |
1368 | * Then, in perf_output_read(), the calculation would | |
1369 | * work with no changes because: | |
1370 | * - event is guaranteed scheduled in | |
1371 | * - no scheduled out in between | |
1372 | * - thus the timestamp would be the same | |
1373 | * | |
1374 | * But this is a bit hairy. | |
1375 | * | |
1376 | * So instead, we have an explicit cgroup call to remain | |
1377 | * within the time time source all along. We believe it | |
1378 | * is cleaner and simpler to understand. | |
1379 | */ | |
1380 | if (is_cgroup_event(event)) | |
1381 | perf_cgroup_set_shadow_time(event, tstamp); | |
1382 | else | |
1383 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1384 | } | |
1385 | ||
4fe757dd PZ |
1386 | #define MAX_INTERRUPTS (~0ULL) |
1387 | ||
1388 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1389 | ||
235c7fc7 | 1390 | static int |
9ffcfa6f | 1391 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1392 | struct perf_cpu_context *cpuctx, |
6e37738a | 1393 | struct perf_event_context *ctx) |
235c7fc7 | 1394 | { |
4158755d SE |
1395 | u64 tstamp = perf_event_time(event); |
1396 | ||
cdd6c482 | 1397 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1398 | return 0; |
1399 | ||
cdd6c482 | 1400 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1401 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1402 | |
1403 | /* | |
1404 | * Unthrottle events, since we scheduled we might have missed several | |
1405 | * ticks already, also for a heavily scheduling task there is little | |
1406 | * guarantee it'll get a tick in a timely manner. | |
1407 | */ | |
1408 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1409 | perf_log_throttle(event, 1); | |
1410 | event->hw.interrupts = 0; | |
1411 | } | |
1412 | ||
235c7fc7 IM |
1413 | /* |
1414 | * The new state must be visible before we turn it on in the hardware: | |
1415 | */ | |
1416 | smp_wmb(); | |
1417 | ||
a4eaf7f1 | 1418 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1419 | event->state = PERF_EVENT_STATE_INACTIVE; |
1420 | event->oncpu = -1; | |
235c7fc7 IM |
1421 | return -EAGAIN; |
1422 | } | |
1423 | ||
4158755d | 1424 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1425 | |
e5d1367f | 1426 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1427 | |
cdd6c482 | 1428 | if (!is_software_event(event)) |
3b6f9e5c | 1429 | cpuctx->active_oncpu++; |
235c7fc7 | 1430 | ctx->nr_active++; |
0f5a2601 PZ |
1431 | if (event->attr.freq && event->attr.sample_freq) |
1432 | ctx->nr_freq++; | |
235c7fc7 | 1433 | |
cdd6c482 | 1434 | if (event->attr.exclusive) |
3b6f9e5c PM |
1435 | cpuctx->exclusive = 1; |
1436 | ||
235c7fc7 IM |
1437 | return 0; |
1438 | } | |
1439 | ||
6751b71e | 1440 | static int |
cdd6c482 | 1441 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1442 | struct perf_cpu_context *cpuctx, |
6e37738a | 1443 | struct perf_event_context *ctx) |
6751b71e | 1444 | { |
6bde9b6c | 1445 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1446 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1447 | u64 now = ctx->time; |
1448 | bool simulate = false; | |
6751b71e | 1449 | |
cdd6c482 | 1450 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1451 | return 0; |
1452 | ||
ad5133b7 | 1453 | pmu->start_txn(pmu); |
6bde9b6c | 1454 | |
9ffcfa6f | 1455 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1456 | pmu->cancel_txn(pmu); |
6751b71e | 1457 | return -EAGAIN; |
90151c35 | 1458 | } |
6751b71e PM |
1459 | |
1460 | /* | |
1461 | * Schedule in siblings as one group (if any): | |
1462 | */ | |
cdd6c482 | 1463 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1464 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1465 | partial_group = event; |
6751b71e PM |
1466 | goto group_error; |
1467 | } | |
1468 | } | |
1469 | ||
9ffcfa6f | 1470 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1471 | return 0; |
9ffcfa6f | 1472 | |
6751b71e PM |
1473 | group_error: |
1474 | /* | |
1475 | * Groups can be scheduled in as one unit only, so undo any | |
1476 | * partial group before returning: | |
d7842da4 SE |
1477 | * The events up to the failed event are scheduled out normally, |
1478 | * tstamp_stopped will be updated. | |
1479 | * | |
1480 | * The failed events and the remaining siblings need to have | |
1481 | * their timings updated as if they had gone thru event_sched_in() | |
1482 | * and event_sched_out(). This is required to get consistent timings | |
1483 | * across the group. This also takes care of the case where the group | |
1484 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1485 | * the time the event was actually stopped, such that time delta | |
1486 | * calculation in update_event_times() is correct. | |
6751b71e | 1487 | */ |
cdd6c482 IM |
1488 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1489 | if (event == partial_group) | |
d7842da4 SE |
1490 | simulate = true; |
1491 | ||
1492 | if (simulate) { | |
1493 | event->tstamp_running += now - event->tstamp_stopped; | |
1494 | event->tstamp_stopped = now; | |
1495 | } else { | |
1496 | event_sched_out(event, cpuctx, ctx); | |
1497 | } | |
6751b71e | 1498 | } |
9ffcfa6f | 1499 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1500 | |
ad5133b7 | 1501 | pmu->cancel_txn(pmu); |
90151c35 | 1502 | |
6751b71e PM |
1503 | return -EAGAIN; |
1504 | } | |
1505 | ||
3b6f9e5c | 1506 | /* |
cdd6c482 | 1507 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1508 | */ |
cdd6c482 | 1509 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1510 | struct perf_cpu_context *cpuctx, |
1511 | int can_add_hw) | |
1512 | { | |
1513 | /* | |
cdd6c482 | 1514 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1515 | */ |
d6f962b5 | 1516 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1517 | return 1; |
1518 | /* | |
1519 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1520 | * events can go on. |
3b6f9e5c PM |
1521 | */ |
1522 | if (cpuctx->exclusive) | |
1523 | return 0; | |
1524 | /* | |
1525 | * If this group is exclusive and there are already | |
cdd6c482 | 1526 | * events on the CPU, it can't go on. |
3b6f9e5c | 1527 | */ |
cdd6c482 | 1528 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1529 | return 0; |
1530 | /* | |
1531 | * Otherwise, try to add it if all previous groups were able | |
1532 | * to go on. | |
1533 | */ | |
1534 | return can_add_hw; | |
1535 | } | |
1536 | ||
cdd6c482 IM |
1537 | static void add_event_to_ctx(struct perf_event *event, |
1538 | struct perf_event_context *ctx) | |
53cfbf59 | 1539 | { |
4158755d SE |
1540 | u64 tstamp = perf_event_time(event); |
1541 | ||
cdd6c482 | 1542 | list_add_event(event, ctx); |
8a49542c | 1543 | perf_group_attach(event); |
4158755d SE |
1544 | event->tstamp_enabled = tstamp; |
1545 | event->tstamp_running = tstamp; | |
1546 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1547 | } |
1548 | ||
2c29ef0f PZ |
1549 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1550 | static void | |
1551 | ctx_sched_in(struct perf_event_context *ctx, | |
1552 | struct perf_cpu_context *cpuctx, | |
1553 | enum event_type_t event_type, | |
1554 | struct task_struct *task); | |
fe4b04fa | 1555 | |
dce5855b PZ |
1556 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1557 | struct perf_event_context *ctx, | |
1558 | struct task_struct *task) | |
1559 | { | |
1560 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1561 | if (ctx) | |
1562 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1563 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1564 | if (ctx) | |
1565 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1566 | } | |
1567 | ||
0793a61d | 1568 | /* |
cdd6c482 | 1569 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1570 | * |
1571 | * Must be called with ctx->mutex held | |
0793a61d | 1572 | */ |
fe4b04fa | 1573 | static int __perf_install_in_context(void *info) |
0793a61d | 1574 | { |
cdd6c482 IM |
1575 | struct perf_event *event = info; |
1576 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1577 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1578 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1579 | struct task_struct *task = current; | |
1580 | ||
b58f6b0d | 1581 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1582 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1583 | |
1584 | /* | |
2c29ef0f | 1585 | * If there was an active task_ctx schedule it out. |
0793a61d | 1586 | */ |
b58f6b0d | 1587 | if (task_ctx) |
2c29ef0f | 1588 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1589 | |
1590 | /* | |
1591 | * If the context we're installing events in is not the | |
1592 | * active task_ctx, flip them. | |
1593 | */ | |
1594 | if (ctx->task && task_ctx != ctx) { | |
1595 | if (task_ctx) | |
1596 | raw_spin_unlock(&task_ctx->lock); | |
1597 | raw_spin_lock(&ctx->lock); | |
1598 | task_ctx = ctx; | |
1599 | } | |
1600 | ||
1601 | if (task_ctx) { | |
1602 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1603 | task = task_ctx->task; |
1604 | } | |
b58f6b0d | 1605 | |
2c29ef0f | 1606 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1607 | |
4af4998b | 1608 | update_context_time(ctx); |
e5d1367f SE |
1609 | /* |
1610 | * update cgrp time only if current cgrp | |
1611 | * matches event->cgrp. Must be done before | |
1612 | * calling add_event_to_ctx() | |
1613 | */ | |
1614 | update_cgrp_time_from_event(event); | |
0793a61d | 1615 | |
cdd6c482 | 1616 | add_event_to_ctx(event, ctx); |
0793a61d | 1617 | |
d859e29f | 1618 | /* |
2c29ef0f | 1619 | * Schedule everything back in |
d859e29f | 1620 | */ |
dce5855b | 1621 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1622 | |
1623 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1624 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1625 | |
1626 | return 0; | |
0793a61d TG |
1627 | } |
1628 | ||
1629 | /* | |
cdd6c482 | 1630 | * Attach a performance event to a context |
0793a61d | 1631 | * |
cdd6c482 IM |
1632 | * First we add the event to the list with the hardware enable bit |
1633 | * in event->hw_config cleared. | |
0793a61d | 1634 | * |
cdd6c482 | 1635 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1636 | * call to enable it in the task context. The task might have been |
1637 | * scheduled away, but we check this in the smp call again. | |
1638 | */ | |
1639 | static void | |
cdd6c482 IM |
1640 | perf_install_in_context(struct perf_event_context *ctx, |
1641 | struct perf_event *event, | |
0793a61d TG |
1642 | int cpu) |
1643 | { | |
1644 | struct task_struct *task = ctx->task; | |
1645 | ||
fe4b04fa PZ |
1646 | lockdep_assert_held(&ctx->mutex); |
1647 | ||
c3f00c70 | 1648 | event->ctx = ctx; |
0cda4c02 YZ |
1649 | if (event->cpu != -1) |
1650 | event->cpu = cpu; | |
c3f00c70 | 1651 | |
0793a61d TG |
1652 | if (!task) { |
1653 | /* | |
cdd6c482 | 1654 | * Per cpu events are installed via an smp call and |
af901ca1 | 1655 | * the install is always successful. |
0793a61d | 1656 | */ |
fe4b04fa | 1657 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1658 | return; |
1659 | } | |
1660 | ||
0793a61d | 1661 | retry: |
fe4b04fa PZ |
1662 | if (!task_function_call(task, __perf_install_in_context, event)) |
1663 | return; | |
0793a61d | 1664 | |
e625cce1 | 1665 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1666 | /* |
fe4b04fa PZ |
1667 | * If we failed to find a running task, but find the context active now |
1668 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1669 | */ |
fe4b04fa | 1670 | if (ctx->is_active) { |
e625cce1 | 1671 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1672 | goto retry; |
1673 | } | |
1674 | ||
1675 | /* | |
fe4b04fa PZ |
1676 | * Since the task isn't running, its safe to add the event, us holding |
1677 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1678 | */ |
fe4b04fa | 1679 | add_event_to_ctx(event, ctx); |
e625cce1 | 1680 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1681 | } |
1682 | ||
fa289bec | 1683 | /* |
cdd6c482 | 1684 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1685 | * Enabling the leader of a group effectively enables all |
1686 | * the group members that aren't explicitly disabled, so we | |
1687 | * have to update their ->tstamp_enabled also. | |
1688 | * Note: this works for group members as well as group leaders | |
1689 | * since the non-leader members' sibling_lists will be empty. | |
1690 | */ | |
1d9b482e | 1691 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1692 | { |
cdd6c482 | 1693 | struct perf_event *sub; |
4158755d | 1694 | u64 tstamp = perf_event_time(event); |
fa289bec | 1695 | |
cdd6c482 | 1696 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1697 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1698 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1699 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1700 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1701 | } |
fa289bec PM |
1702 | } |
1703 | ||
d859e29f | 1704 | /* |
cdd6c482 | 1705 | * Cross CPU call to enable a performance event |
d859e29f | 1706 | */ |
fe4b04fa | 1707 | static int __perf_event_enable(void *info) |
04289bb9 | 1708 | { |
cdd6c482 | 1709 | struct perf_event *event = info; |
cdd6c482 IM |
1710 | struct perf_event_context *ctx = event->ctx; |
1711 | struct perf_event *leader = event->group_leader; | |
108b02cf | 1712 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 1713 | int err; |
04289bb9 | 1714 | |
fe4b04fa PZ |
1715 | if (WARN_ON_ONCE(!ctx->is_active)) |
1716 | return -EINVAL; | |
3cbed429 | 1717 | |
e625cce1 | 1718 | raw_spin_lock(&ctx->lock); |
4af4998b | 1719 | update_context_time(ctx); |
d859e29f | 1720 | |
cdd6c482 | 1721 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 1722 | goto unlock; |
e5d1367f SE |
1723 | |
1724 | /* | |
1725 | * set current task's cgroup time reference point | |
1726 | */ | |
3f7cce3c | 1727 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 1728 | |
1d9b482e | 1729 | __perf_event_mark_enabled(event); |
04289bb9 | 1730 | |
e5d1367f SE |
1731 | if (!event_filter_match(event)) { |
1732 | if (is_cgroup_event(event)) | |
1733 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 1734 | goto unlock; |
e5d1367f | 1735 | } |
f4c4176f | 1736 | |
04289bb9 | 1737 | /* |
cdd6c482 | 1738 | * If the event is in a group and isn't the group leader, |
d859e29f | 1739 | * then don't put it on unless the group is on. |
04289bb9 | 1740 | */ |
cdd6c482 | 1741 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 1742 | goto unlock; |
3b6f9e5c | 1743 | |
cdd6c482 | 1744 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 1745 | err = -EEXIST; |
e758a33d | 1746 | } else { |
cdd6c482 | 1747 | if (event == leader) |
6e37738a | 1748 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 1749 | else |
6e37738a | 1750 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 1751 | } |
d859e29f PM |
1752 | |
1753 | if (err) { | |
1754 | /* | |
cdd6c482 | 1755 | * If this event can't go on and it's part of a |
d859e29f PM |
1756 | * group, then the whole group has to come off. |
1757 | */ | |
cdd6c482 | 1758 | if (leader != event) |
d859e29f | 1759 | group_sched_out(leader, cpuctx, ctx); |
0d48696f | 1760 | if (leader->attr.pinned) { |
53cfbf59 | 1761 | update_group_times(leader); |
cdd6c482 | 1762 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 1763 | } |
d859e29f PM |
1764 | } |
1765 | ||
9ed6060d | 1766 | unlock: |
e625cce1 | 1767 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1768 | |
1769 | return 0; | |
d859e29f PM |
1770 | } |
1771 | ||
1772 | /* | |
cdd6c482 | 1773 | * Enable a event. |
c93f7669 | 1774 | * |
cdd6c482 IM |
1775 | * If event->ctx is a cloned context, callers must make sure that |
1776 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1777 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
1778 | * perf_event_for_each_child or perf_event_for_each as described |
1779 | * for perf_event_disable. | |
d859e29f | 1780 | */ |
44234adc | 1781 | void perf_event_enable(struct perf_event *event) |
d859e29f | 1782 | { |
cdd6c482 | 1783 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1784 | struct task_struct *task = ctx->task; |
1785 | ||
1786 | if (!task) { | |
1787 | /* | |
cdd6c482 | 1788 | * Enable the event on the cpu that it's on |
d859e29f | 1789 | */ |
fe4b04fa | 1790 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
1791 | return; |
1792 | } | |
1793 | ||
e625cce1 | 1794 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 1795 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
1796 | goto out; |
1797 | ||
1798 | /* | |
cdd6c482 IM |
1799 | * If the event is in error state, clear that first. |
1800 | * That way, if we see the event in error state below, we | |
d859e29f PM |
1801 | * know that it has gone back into error state, as distinct |
1802 | * from the task having been scheduled away before the | |
1803 | * cross-call arrived. | |
1804 | */ | |
cdd6c482 IM |
1805 | if (event->state == PERF_EVENT_STATE_ERROR) |
1806 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 1807 | |
9ed6060d | 1808 | retry: |
fe4b04fa | 1809 | if (!ctx->is_active) { |
1d9b482e | 1810 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
1811 | goto out; |
1812 | } | |
1813 | ||
e625cce1 | 1814 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1815 | |
1816 | if (!task_function_call(task, __perf_event_enable, event)) | |
1817 | return; | |
d859e29f | 1818 | |
e625cce1 | 1819 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
1820 | |
1821 | /* | |
cdd6c482 | 1822 | * If the context is active and the event is still off, |
d859e29f PM |
1823 | * we need to retry the cross-call. |
1824 | */ | |
fe4b04fa PZ |
1825 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
1826 | /* | |
1827 | * task could have been flipped by a concurrent | |
1828 | * perf_event_context_sched_out() | |
1829 | */ | |
1830 | task = ctx->task; | |
d859e29f | 1831 | goto retry; |
fe4b04fa | 1832 | } |
fa289bec | 1833 | |
9ed6060d | 1834 | out: |
e625cce1 | 1835 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1836 | } |
dcfce4a0 | 1837 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 1838 | |
26ca5c11 | 1839 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 1840 | { |
2023b359 | 1841 | /* |
cdd6c482 | 1842 | * not supported on inherited events |
2023b359 | 1843 | */ |
2e939d1d | 1844 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
1845 | return -EINVAL; |
1846 | ||
cdd6c482 IM |
1847 | atomic_add(refresh, &event->event_limit); |
1848 | perf_event_enable(event); | |
2023b359 PZ |
1849 | |
1850 | return 0; | |
79f14641 | 1851 | } |
26ca5c11 | 1852 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 1853 | |
5b0311e1 FW |
1854 | static void ctx_sched_out(struct perf_event_context *ctx, |
1855 | struct perf_cpu_context *cpuctx, | |
1856 | enum event_type_t event_type) | |
235c7fc7 | 1857 | { |
cdd6c482 | 1858 | struct perf_event *event; |
db24d33e | 1859 | int is_active = ctx->is_active; |
235c7fc7 | 1860 | |
db24d33e | 1861 | ctx->is_active &= ~event_type; |
cdd6c482 | 1862 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
1863 | return; |
1864 | ||
4af4998b | 1865 | update_context_time(ctx); |
e5d1367f | 1866 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 1867 | if (!ctx->nr_active) |
facc4307 | 1868 | return; |
5b0311e1 | 1869 | |
075e0b00 | 1870 | perf_pmu_disable(ctx->pmu); |
db24d33e | 1871 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
1872 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
1873 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 1874 | } |
889ff015 | 1875 | |
db24d33e | 1876 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 1877 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 1878 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 1879 | } |
1b9a644f | 1880 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
1881 | } |
1882 | ||
564c2b21 PM |
1883 | /* |
1884 | * Test whether two contexts are equivalent, i.e. whether they | |
1885 | * have both been cloned from the same version of the same context | |
cdd6c482 IM |
1886 | * and they both have the same number of enabled events. |
1887 | * If the number of enabled events is the same, then the set | |
1888 | * of enabled events should be the same, because these are both | |
1889 | * inherited contexts, therefore we can't access individual events | |
564c2b21 | 1890 | * in them directly with an fd; we can only enable/disable all |
cdd6c482 | 1891 | * events via prctl, or enable/disable all events in a family |
564c2b21 PM |
1892 | * via ioctl, which will have the same effect on both contexts. |
1893 | */ | |
cdd6c482 IM |
1894 | static int context_equiv(struct perf_event_context *ctx1, |
1895 | struct perf_event_context *ctx2) | |
564c2b21 PM |
1896 | { |
1897 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx | |
ad3a37de | 1898 | && ctx1->parent_gen == ctx2->parent_gen |
25346b93 | 1899 | && !ctx1->pin_count && !ctx2->pin_count; |
564c2b21 PM |
1900 | } |
1901 | ||
cdd6c482 IM |
1902 | static void __perf_event_sync_stat(struct perf_event *event, |
1903 | struct perf_event *next_event) | |
bfbd3381 PZ |
1904 | { |
1905 | u64 value; | |
1906 | ||
cdd6c482 | 1907 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
1908 | return; |
1909 | ||
1910 | /* | |
cdd6c482 | 1911 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
1912 | * because we're in the middle of a context switch and have IRQs |
1913 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 1914 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
1915 | * don't need to use it. |
1916 | */ | |
cdd6c482 IM |
1917 | switch (event->state) { |
1918 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
1919 | event->pmu->read(event); |
1920 | /* fall-through */ | |
bfbd3381 | 1921 | |
cdd6c482 IM |
1922 | case PERF_EVENT_STATE_INACTIVE: |
1923 | update_event_times(event); | |
bfbd3381 PZ |
1924 | break; |
1925 | ||
1926 | default: | |
1927 | break; | |
1928 | } | |
1929 | ||
1930 | /* | |
cdd6c482 | 1931 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
1932 | * values when we flip the contexts. |
1933 | */ | |
e7850595 PZ |
1934 | value = local64_read(&next_event->count); |
1935 | value = local64_xchg(&event->count, value); | |
1936 | local64_set(&next_event->count, value); | |
bfbd3381 | 1937 | |
cdd6c482 IM |
1938 | swap(event->total_time_enabled, next_event->total_time_enabled); |
1939 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 1940 | |
bfbd3381 | 1941 | /* |
19d2e755 | 1942 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 1943 | */ |
cdd6c482 IM |
1944 | perf_event_update_userpage(event); |
1945 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
1946 | } |
1947 | ||
1948 | #define list_next_entry(pos, member) \ | |
1949 | list_entry(pos->member.next, typeof(*pos), member) | |
1950 | ||
cdd6c482 IM |
1951 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
1952 | struct perf_event_context *next_ctx) | |
bfbd3381 | 1953 | { |
cdd6c482 | 1954 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
1955 | |
1956 | if (!ctx->nr_stat) | |
1957 | return; | |
1958 | ||
02ffdbc8 PZ |
1959 | update_context_time(ctx); |
1960 | ||
cdd6c482 IM |
1961 | event = list_first_entry(&ctx->event_list, |
1962 | struct perf_event, event_entry); | |
bfbd3381 | 1963 | |
cdd6c482 IM |
1964 | next_event = list_first_entry(&next_ctx->event_list, |
1965 | struct perf_event, event_entry); | |
bfbd3381 | 1966 | |
cdd6c482 IM |
1967 | while (&event->event_entry != &ctx->event_list && |
1968 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 1969 | |
cdd6c482 | 1970 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 1971 | |
cdd6c482 IM |
1972 | event = list_next_entry(event, event_entry); |
1973 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
1974 | } |
1975 | } | |
1976 | ||
fe4b04fa PZ |
1977 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
1978 | struct task_struct *next) | |
0793a61d | 1979 | { |
8dc85d54 | 1980 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 IM |
1981 | struct perf_event_context *next_ctx; |
1982 | struct perf_event_context *parent; | |
108b02cf | 1983 | struct perf_cpu_context *cpuctx; |
c93f7669 | 1984 | int do_switch = 1; |
0793a61d | 1985 | |
108b02cf PZ |
1986 | if (likely(!ctx)) |
1987 | return; | |
10989fb2 | 1988 | |
108b02cf PZ |
1989 | cpuctx = __get_cpu_context(ctx); |
1990 | if (!cpuctx->task_ctx) | |
0793a61d TG |
1991 | return; |
1992 | ||
c93f7669 PM |
1993 | rcu_read_lock(); |
1994 | parent = rcu_dereference(ctx->parent_ctx); | |
8dc85d54 | 1995 | next_ctx = next->perf_event_ctxp[ctxn]; |
c93f7669 PM |
1996 | if (parent && next_ctx && |
1997 | rcu_dereference(next_ctx->parent_ctx) == parent) { | |
1998 | /* | |
1999 | * Looks like the two contexts are clones, so we might be | |
2000 | * able to optimize the context switch. We lock both | |
2001 | * contexts and check that they are clones under the | |
2002 | * lock (including re-checking that neither has been | |
2003 | * uncloned in the meantime). It doesn't matter which | |
2004 | * order we take the locks because no other cpu could | |
2005 | * be trying to lock both of these tasks. | |
2006 | */ | |
e625cce1 TG |
2007 | raw_spin_lock(&ctx->lock); |
2008 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2009 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2010 | /* |
2011 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2012 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2013 | */ |
8dc85d54 PZ |
2014 | task->perf_event_ctxp[ctxn] = next_ctx; |
2015 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2016 | ctx->task = next; |
2017 | next_ctx->task = task; | |
2018 | do_switch = 0; | |
bfbd3381 | 2019 | |
cdd6c482 | 2020 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2021 | } |
e625cce1 TG |
2022 | raw_spin_unlock(&next_ctx->lock); |
2023 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2024 | } |
c93f7669 | 2025 | rcu_read_unlock(); |
564c2b21 | 2026 | |
c93f7669 | 2027 | if (do_switch) { |
facc4307 | 2028 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2029 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2030 | cpuctx->task_ctx = NULL; |
facc4307 | 2031 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2032 | } |
0793a61d TG |
2033 | } |
2034 | ||
8dc85d54 PZ |
2035 | #define for_each_task_context_nr(ctxn) \ |
2036 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2037 | ||
2038 | /* | |
2039 | * Called from scheduler to remove the events of the current task, | |
2040 | * with interrupts disabled. | |
2041 | * | |
2042 | * We stop each event and update the event value in event->count. | |
2043 | * | |
2044 | * This does not protect us against NMI, but disable() | |
2045 | * sets the disabled bit in the control field of event _before_ | |
2046 | * accessing the event control register. If a NMI hits, then it will | |
2047 | * not restart the event. | |
2048 | */ | |
ab0cce56 JO |
2049 | void __perf_event_task_sched_out(struct task_struct *task, |
2050 | struct task_struct *next) | |
8dc85d54 PZ |
2051 | { |
2052 | int ctxn; | |
2053 | ||
8dc85d54 PZ |
2054 | for_each_task_context_nr(ctxn) |
2055 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2056 | |
2057 | /* | |
2058 | * if cgroup events exist on this CPU, then we need | |
2059 | * to check if we have to switch out PMU state. | |
2060 | * cgroup event are system-wide mode only | |
2061 | */ | |
2062 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2063 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2064 | } |
2065 | ||
04dc2dbb | 2066 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2067 | { |
108b02cf | 2068 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2069 | |
a63eaf34 PM |
2070 | if (!cpuctx->task_ctx) |
2071 | return; | |
012b84da IM |
2072 | |
2073 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2074 | return; | |
2075 | ||
04dc2dbb | 2076 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2077 | cpuctx->task_ctx = NULL; |
2078 | } | |
2079 | ||
5b0311e1 FW |
2080 | /* |
2081 | * Called with IRQs disabled | |
2082 | */ | |
2083 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2084 | enum event_type_t event_type) | |
2085 | { | |
2086 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2087 | } |
2088 | ||
235c7fc7 | 2089 | static void |
5b0311e1 | 2090 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2091 | struct perf_cpu_context *cpuctx) |
0793a61d | 2092 | { |
cdd6c482 | 2093 | struct perf_event *event; |
0793a61d | 2094 | |
889ff015 FW |
2095 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2096 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2097 | continue; |
5632ab12 | 2098 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2099 | continue; |
2100 | ||
e5d1367f SE |
2101 | /* may need to reset tstamp_enabled */ |
2102 | if (is_cgroup_event(event)) | |
2103 | perf_cgroup_mark_enabled(event, ctx); | |
2104 | ||
8c9ed8e1 | 2105 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2106 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2107 | |
2108 | /* | |
2109 | * If this pinned group hasn't been scheduled, | |
2110 | * put it in error state. | |
2111 | */ | |
cdd6c482 IM |
2112 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2113 | update_group_times(event); | |
2114 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2115 | } |
3b6f9e5c | 2116 | } |
5b0311e1 FW |
2117 | } |
2118 | ||
2119 | static void | |
2120 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2121 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2122 | { |
2123 | struct perf_event *event; | |
2124 | int can_add_hw = 1; | |
3b6f9e5c | 2125 | |
889ff015 FW |
2126 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2127 | /* Ignore events in OFF or ERROR state */ | |
2128 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2129 | continue; |
04289bb9 IM |
2130 | /* |
2131 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2132 | * of events: |
04289bb9 | 2133 | */ |
5632ab12 | 2134 | if (!event_filter_match(event)) |
0793a61d TG |
2135 | continue; |
2136 | ||
e5d1367f SE |
2137 | /* may need to reset tstamp_enabled */ |
2138 | if (is_cgroup_event(event)) | |
2139 | perf_cgroup_mark_enabled(event, ctx); | |
2140 | ||
9ed6060d | 2141 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2142 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2143 | can_add_hw = 0; |
9ed6060d | 2144 | } |
0793a61d | 2145 | } |
5b0311e1 FW |
2146 | } |
2147 | ||
2148 | static void | |
2149 | ctx_sched_in(struct perf_event_context *ctx, | |
2150 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2151 | enum event_type_t event_type, |
2152 | struct task_struct *task) | |
5b0311e1 | 2153 | { |
e5d1367f | 2154 | u64 now; |
db24d33e | 2155 | int is_active = ctx->is_active; |
e5d1367f | 2156 | |
db24d33e | 2157 | ctx->is_active |= event_type; |
5b0311e1 | 2158 | if (likely(!ctx->nr_events)) |
facc4307 | 2159 | return; |
5b0311e1 | 2160 | |
e5d1367f SE |
2161 | now = perf_clock(); |
2162 | ctx->timestamp = now; | |
3f7cce3c | 2163 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2164 | /* |
2165 | * First go through the list and put on any pinned groups | |
2166 | * in order to give them the best chance of going on. | |
2167 | */ | |
db24d33e | 2168 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2169 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2170 | |
2171 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2172 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2173 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2174 | } |
2175 | ||
329c0e01 | 2176 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2177 | enum event_type_t event_type, |
2178 | struct task_struct *task) | |
329c0e01 FW |
2179 | { |
2180 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2181 | ||
e5d1367f | 2182 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2183 | } |
2184 | ||
e5d1367f SE |
2185 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2186 | struct task_struct *task) | |
235c7fc7 | 2187 | { |
108b02cf | 2188 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2189 | |
108b02cf | 2190 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2191 | if (cpuctx->task_ctx == ctx) |
2192 | return; | |
2193 | ||
facc4307 | 2194 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2195 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2196 | /* |
2197 | * We want to keep the following priority order: | |
2198 | * cpu pinned (that don't need to move), task pinned, | |
2199 | * cpu flexible, task flexible. | |
2200 | */ | |
2201 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2202 | ||
1d5f003f GN |
2203 | if (ctx->nr_events) |
2204 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2205 | |
86b47c25 GN |
2206 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2207 | ||
facc4307 PZ |
2208 | perf_pmu_enable(ctx->pmu); |
2209 | perf_ctx_unlock(cpuctx, ctx); | |
2210 | ||
b5ab4cd5 PZ |
2211 | /* |
2212 | * Since these rotations are per-cpu, we need to ensure the | |
2213 | * cpu-context we got scheduled on is actually rotating. | |
2214 | */ | |
108b02cf | 2215 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2216 | } |
2217 | ||
d010b332 SE |
2218 | /* |
2219 | * When sampling the branck stack in system-wide, it may be necessary | |
2220 | * to flush the stack on context switch. This happens when the branch | |
2221 | * stack does not tag its entries with the pid of the current task. | |
2222 | * Otherwise it becomes impossible to associate a branch entry with a | |
2223 | * task. This ambiguity is more likely to appear when the branch stack | |
2224 | * supports priv level filtering and the user sets it to monitor only | |
2225 | * at the user level (which could be a useful measurement in system-wide | |
2226 | * mode). In that case, the risk is high of having a branch stack with | |
2227 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2228 | * entries or stashing them somewhere in the PMU specific code layer. | |
2229 | * | |
2230 | * This function provides the context switch callback to the lower code | |
2231 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2232 | * with at least one active event using taken branch sampling. | |
2233 | */ | |
2234 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2235 | struct task_struct *task) | |
2236 | { | |
2237 | struct perf_cpu_context *cpuctx; | |
2238 | struct pmu *pmu; | |
2239 | unsigned long flags; | |
2240 | ||
2241 | /* no need to flush branch stack if not changing task */ | |
2242 | if (prev == task) | |
2243 | return; | |
2244 | ||
2245 | local_irq_save(flags); | |
2246 | ||
2247 | rcu_read_lock(); | |
2248 | ||
2249 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2250 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2251 | ||
2252 | /* | |
2253 | * check if the context has at least one | |
2254 | * event using PERF_SAMPLE_BRANCH_STACK | |
2255 | */ | |
2256 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2257 | && pmu->flush_branch_stack) { | |
2258 | ||
2259 | pmu = cpuctx->ctx.pmu; | |
2260 | ||
2261 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2262 | ||
2263 | perf_pmu_disable(pmu); | |
2264 | ||
2265 | pmu->flush_branch_stack(); | |
2266 | ||
2267 | perf_pmu_enable(pmu); | |
2268 | ||
2269 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2270 | } | |
2271 | } | |
2272 | ||
2273 | rcu_read_unlock(); | |
2274 | ||
2275 | local_irq_restore(flags); | |
2276 | } | |
2277 | ||
8dc85d54 PZ |
2278 | /* |
2279 | * Called from scheduler to add the events of the current task | |
2280 | * with interrupts disabled. | |
2281 | * | |
2282 | * We restore the event value and then enable it. | |
2283 | * | |
2284 | * This does not protect us against NMI, but enable() | |
2285 | * sets the enabled bit in the control field of event _before_ | |
2286 | * accessing the event control register. If a NMI hits, then it will | |
2287 | * keep the event running. | |
2288 | */ | |
ab0cce56 JO |
2289 | void __perf_event_task_sched_in(struct task_struct *prev, |
2290 | struct task_struct *task) | |
8dc85d54 PZ |
2291 | { |
2292 | struct perf_event_context *ctx; | |
2293 | int ctxn; | |
2294 | ||
2295 | for_each_task_context_nr(ctxn) { | |
2296 | ctx = task->perf_event_ctxp[ctxn]; | |
2297 | if (likely(!ctx)) | |
2298 | continue; | |
2299 | ||
e5d1367f | 2300 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2301 | } |
e5d1367f SE |
2302 | /* |
2303 | * if cgroup events exist on this CPU, then we need | |
2304 | * to check if we have to switch in PMU state. | |
2305 | * cgroup event are system-wide mode only | |
2306 | */ | |
2307 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2308 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2309 | |
2310 | /* check for system-wide branch_stack events */ | |
2311 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2312 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2313 | } |
2314 | ||
abd50713 PZ |
2315 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2316 | { | |
2317 | u64 frequency = event->attr.sample_freq; | |
2318 | u64 sec = NSEC_PER_SEC; | |
2319 | u64 divisor, dividend; | |
2320 | ||
2321 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2322 | ||
2323 | count_fls = fls64(count); | |
2324 | nsec_fls = fls64(nsec); | |
2325 | frequency_fls = fls64(frequency); | |
2326 | sec_fls = 30; | |
2327 | ||
2328 | /* | |
2329 | * We got @count in @nsec, with a target of sample_freq HZ | |
2330 | * the target period becomes: | |
2331 | * | |
2332 | * @count * 10^9 | |
2333 | * period = ------------------- | |
2334 | * @nsec * sample_freq | |
2335 | * | |
2336 | */ | |
2337 | ||
2338 | /* | |
2339 | * Reduce accuracy by one bit such that @a and @b converge | |
2340 | * to a similar magnitude. | |
2341 | */ | |
fe4b04fa | 2342 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2343 | do { \ |
2344 | if (a##_fls > b##_fls) { \ | |
2345 | a >>= 1; \ | |
2346 | a##_fls--; \ | |
2347 | } else { \ | |
2348 | b >>= 1; \ | |
2349 | b##_fls--; \ | |
2350 | } \ | |
2351 | } while (0) | |
2352 | ||
2353 | /* | |
2354 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2355 | * the other, so that finally we can do a u64/u64 division. | |
2356 | */ | |
2357 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2358 | REDUCE_FLS(nsec, frequency); | |
2359 | REDUCE_FLS(sec, count); | |
2360 | } | |
2361 | ||
2362 | if (count_fls + sec_fls > 64) { | |
2363 | divisor = nsec * frequency; | |
2364 | ||
2365 | while (count_fls + sec_fls > 64) { | |
2366 | REDUCE_FLS(count, sec); | |
2367 | divisor >>= 1; | |
2368 | } | |
2369 | ||
2370 | dividend = count * sec; | |
2371 | } else { | |
2372 | dividend = count * sec; | |
2373 | ||
2374 | while (nsec_fls + frequency_fls > 64) { | |
2375 | REDUCE_FLS(nsec, frequency); | |
2376 | dividend >>= 1; | |
2377 | } | |
2378 | ||
2379 | divisor = nsec * frequency; | |
2380 | } | |
2381 | ||
f6ab91ad PZ |
2382 | if (!divisor) |
2383 | return dividend; | |
2384 | ||
abd50713 PZ |
2385 | return div64_u64(dividend, divisor); |
2386 | } | |
2387 | ||
e050e3f0 SE |
2388 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2389 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2390 | ||
f39d47ff | 2391 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2392 | { |
cdd6c482 | 2393 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2394 | s64 period, sample_period; |
bd2b5b12 PZ |
2395 | s64 delta; |
2396 | ||
abd50713 | 2397 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2398 | |
2399 | delta = (s64)(period - hwc->sample_period); | |
2400 | delta = (delta + 7) / 8; /* low pass filter */ | |
2401 | ||
2402 | sample_period = hwc->sample_period + delta; | |
2403 | ||
2404 | if (!sample_period) | |
2405 | sample_period = 1; | |
2406 | ||
bd2b5b12 | 2407 | hwc->sample_period = sample_period; |
abd50713 | 2408 | |
e7850595 | 2409 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2410 | if (disable) |
2411 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2412 | ||
e7850595 | 2413 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2414 | |
2415 | if (disable) | |
2416 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2417 | } |
bd2b5b12 PZ |
2418 | } |
2419 | ||
e050e3f0 SE |
2420 | /* |
2421 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2422 | * events. At the same time, make sure, having freq events does not change | |
2423 | * the rate of unthrottling as that would introduce bias. | |
2424 | */ | |
2425 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2426 | int needs_unthr) | |
60db5e09 | 2427 | { |
cdd6c482 IM |
2428 | struct perf_event *event; |
2429 | struct hw_perf_event *hwc; | |
e050e3f0 | 2430 | u64 now, period = TICK_NSEC; |
abd50713 | 2431 | s64 delta; |
60db5e09 | 2432 | |
e050e3f0 SE |
2433 | /* |
2434 | * only need to iterate over all events iff: | |
2435 | * - context have events in frequency mode (needs freq adjust) | |
2436 | * - there are events to unthrottle on this cpu | |
2437 | */ | |
2438 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2439 | return; |
2440 | ||
e050e3f0 | 2441 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2442 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2443 | |
03541f8b | 2444 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2445 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2446 | continue; |
2447 | ||
5632ab12 | 2448 | if (!event_filter_match(event)) |
5d27c23d PZ |
2449 | continue; |
2450 | ||
cdd6c482 | 2451 | hwc = &event->hw; |
6a24ed6c | 2452 | |
e050e3f0 SE |
2453 | if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) { |
2454 | hwc->interrupts = 0; | |
cdd6c482 | 2455 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2456 | event->pmu->start(event, 0); |
a78ac325 PZ |
2457 | } |
2458 | ||
cdd6c482 | 2459 | if (!event->attr.freq || !event->attr.sample_freq) |
60db5e09 PZ |
2460 | continue; |
2461 | ||
e050e3f0 SE |
2462 | /* |
2463 | * stop the event and update event->count | |
2464 | */ | |
2465 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2466 | ||
e7850595 | 2467 | now = local64_read(&event->count); |
abd50713 PZ |
2468 | delta = now - hwc->freq_count_stamp; |
2469 | hwc->freq_count_stamp = now; | |
60db5e09 | 2470 | |
e050e3f0 SE |
2471 | /* |
2472 | * restart the event | |
2473 | * reload only if value has changed | |
f39d47ff SE |
2474 | * we have stopped the event so tell that |
2475 | * to perf_adjust_period() to avoid stopping it | |
2476 | * twice. | |
e050e3f0 | 2477 | */ |
abd50713 | 2478 | if (delta > 0) |
f39d47ff | 2479 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2480 | |
2481 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
60db5e09 | 2482 | } |
e050e3f0 | 2483 | |
f39d47ff | 2484 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2485 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2486 | } |
2487 | ||
235c7fc7 | 2488 | /* |
cdd6c482 | 2489 | * Round-robin a context's events: |
235c7fc7 | 2490 | */ |
cdd6c482 | 2491 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2492 | { |
dddd3379 TG |
2493 | /* |
2494 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2495 | * disabled by the inheritance code. | |
2496 | */ | |
2497 | if (!ctx->rotate_disable) | |
2498 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2499 | } |
2500 | ||
b5ab4cd5 | 2501 | /* |
e9d2b064 PZ |
2502 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2503 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2504 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2505 | */ |
e9d2b064 | 2506 | static void perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2507 | { |
8dc85d54 | 2508 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2509 | int rotate = 0, remove = 1; |
7fc23a53 | 2510 | |
b5ab4cd5 | 2511 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2512 | remove = 0; |
b5ab4cd5 PZ |
2513 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2514 | rotate = 1; | |
2515 | } | |
235c7fc7 | 2516 | |
8dc85d54 | 2517 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2518 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2519 | remove = 0; |
b5ab4cd5 PZ |
2520 | if (ctx->nr_events != ctx->nr_active) |
2521 | rotate = 1; | |
2522 | } | |
9717e6cd | 2523 | |
e050e3f0 | 2524 | if (!rotate) |
0f5a2601 PZ |
2525 | goto done; |
2526 | ||
facc4307 | 2527 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2528 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2529 | |
e050e3f0 SE |
2530 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2531 | if (ctx) | |
2532 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2533 | |
e050e3f0 SE |
2534 | rotate_ctx(&cpuctx->ctx); |
2535 | if (ctx) | |
2536 | rotate_ctx(ctx); | |
235c7fc7 | 2537 | |
e050e3f0 | 2538 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2539 | |
0f5a2601 PZ |
2540 | perf_pmu_enable(cpuctx->ctx.pmu); |
2541 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2542 | done: |
e9d2b064 PZ |
2543 | if (remove) |
2544 | list_del_init(&cpuctx->rotation_list); | |
e9d2b064 PZ |
2545 | } |
2546 | ||
2547 | void perf_event_task_tick(void) | |
2548 | { | |
2549 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2550 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2551 | struct perf_event_context *ctx; |
2552 | int throttled; | |
b5ab4cd5 | 2553 | |
e9d2b064 PZ |
2554 | WARN_ON(!irqs_disabled()); |
2555 | ||
e050e3f0 SE |
2556 | __this_cpu_inc(perf_throttled_seq); |
2557 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2558 | ||
e9d2b064 | 2559 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2560 | ctx = &cpuctx->ctx; |
2561 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2562 | ||
2563 | ctx = cpuctx->task_ctx; | |
2564 | if (ctx) | |
2565 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2566 | ||
e9d2b064 PZ |
2567 | if (cpuctx->jiffies_interval == 1 || |
2568 | !(jiffies % cpuctx->jiffies_interval)) | |
2569 | perf_rotate_context(cpuctx); | |
2570 | } | |
0793a61d TG |
2571 | } |
2572 | ||
889ff015 FW |
2573 | static int event_enable_on_exec(struct perf_event *event, |
2574 | struct perf_event_context *ctx) | |
2575 | { | |
2576 | if (!event->attr.enable_on_exec) | |
2577 | return 0; | |
2578 | ||
2579 | event->attr.enable_on_exec = 0; | |
2580 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2581 | return 0; | |
2582 | ||
1d9b482e | 2583 | __perf_event_mark_enabled(event); |
889ff015 FW |
2584 | |
2585 | return 1; | |
2586 | } | |
2587 | ||
57e7986e | 2588 | /* |
cdd6c482 | 2589 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2590 | * This expects task == current. |
2591 | */ | |
8dc85d54 | 2592 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2593 | { |
cdd6c482 | 2594 | struct perf_event *event; |
57e7986e PM |
2595 | unsigned long flags; |
2596 | int enabled = 0; | |
889ff015 | 2597 | int ret; |
57e7986e PM |
2598 | |
2599 | local_irq_save(flags); | |
cdd6c482 | 2600 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2601 | goto out; |
2602 | ||
e566b76e SE |
2603 | /* |
2604 | * We must ctxsw out cgroup events to avoid conflict | |
2605 | * when invoking perf_task_event_sched_in() later on | |
2606 | * in this function. Otherwise we end up trying to | |
2607 | * ctxswin cgroup events which are already scheduled | |
2608 | * in. | |
2609 | */ | |
a8d757ef | 2610 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2611 | |
e625cce1 | 2612 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2613 | task_ctx_sched_out(ctx); |
57e7986e | 2614 | |
b79387ef | 2615 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2616 | ret = event_enable_on_exec(event, ctx); |
2617 | if (ret) | |
2618 | enabled = 1; | |
57e7986e PM |
2619 | } |
2620 | ||
2621 | /* | |
cdd6c482 | 2622 | * Unclone this context if we enabled any event. |
57e7986e | 2623 | */ |
71a851b4 PZ |
2624 | if (enabled) |
2625 | unclone_ctx(ctx); | |
57e7986e | 2626 | |
e625cce1 | 2627 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2628 | |
e566b76e SE |
2629 | /* |
2630 | * Also calls ctxswin for cgroup events, if any: | |
2631 | */ | |
e5d1367f | 2632 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2633 | out: |
57e7986e PM |
2634 | local_irq_restore(flags); |
2635 | } | |
2636 | ||
0793a61d | 2637 | /* |
cdd6c482 | 2638 | * Cross CPU call to read the hardware event |
0793a61d | 2639 | */ |
cdd6c482 | 2640 | static void __perf_event_read(void *info) |
0793a61d | 2641 | { |
cdd6c482 IM |
2642 | struct perf_event *event = info; |
2643 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2644 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2645 | |
e1ac3614 PM |
2646 | /* |
2647 | * If this is a task context, we need to check whether it is | |
2648 | * the current task context of this cpu. If not it has been | |
2649 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2650 | * event->count would have been updated to a recent sample |
2651 | * when the event was scheduled out. | |
e1ac3614 PM |
2652 | */ |
2653 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2654 | return; | |
2655 | ||
e625cce1 | 2656 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2657 | if (ctx->is_active) { |
542e72fc | 2658 | update_context_time(ctx); |
e5d1367f SE |
2659 | update_cgrp_time_from_event(event); |
2660 | } | |
cdd6c482 | 2661 | update_event_times(event); |
542e72fc PZ |
2662 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2663 | event->pmu->read(event); | |
e625cce1 | 2664 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2665 | } |
2666 | ||
b5e58793 PZ |
2667 | static inline u64 perf_event_count(struct perf_event *event) |
2668 | { | |
e7850595 | 2669 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
2670 | } |
2671 | ||
cdd6c482 | 2672 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
2673 | { |
2674 | /* | |
cdd6c482 IM |
2675 | * If event is enabled and currently active on a CPU, update the |
2676 | * value in the event structure: | |
0793a61d | 2677 | */ |
cdd6c482 IM |
2678 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2679 | smp_call_function_single(event->oncpu, | |
2680 | __perf_event_read, event, 1); | |
2681 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
2682 | struct perf_event_context *ctx = event->ctx; |
2683 | unsigned long flags; | |
2684 | ||
e625cce1 | 2685 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
2686 | /* |
2687 | * may read while context is not active | |
2688 | * (e.g., thread is blocked), in that case | |
2689 | * we cannot update context time | |
2690 | */ | |
e5d1367f | 2691 | if (ctx->is_active) { |
c530ccd9 | 2692 | update_context_time(ctx); |
e5d1367f SE |
2693 | update_cgrp_time_from_event(event); |
2694 | } | |
cdd6c482 | 2695 | update_event_times(event); |
e625cce1 | 2696 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
2697 | } |
2698 | ||
b5e58793 | 2699 | return perf_event_count(event); |
0793a61d TG |
2700 | } |
2701 | ||
a63eaf34 | 2702 | /* |
cdd6c482 | 2703 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 2704 | */ |
eb184479 | 2705 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 2706 | { |
e625cce1 | 2707 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 2708 | mutex_init(&ctx->mutex); |
889ff015 FW |
2709 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2710 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
2711 | INIT_LIST_HEAD(&ctx->event_list); |
2712 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
2713 | } |
2714 | ||
2715 | static struct perf_event_context * | |
2716 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
2717 | { | |
2718 | struct perf_event_context *ctx; | |
2719 | ||
2720 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
2721 | if (!ctx) | |
2722 | return NULL; | |
2723 | ||
2724 | __perf_event_init_context(ctx); | |
2725 | if (task) { | |
2726 | ctx->task = task; | |
2727 | get_task_struct(task); | |
0793a61d | 2728 | } |
eb184479 PZ |
2729 | ctx->pmu = pmu; |
2730 | ||
2731 | return ctx; | |
a63eaf34 PM |
2732 | } |
2733 | ||
2ebd4ffb MH |
2734 | static struct task_struct * |
2735 | find_lively_task_by_vpid(pid_t vpid) | |
2736 | { | |
2737 | struct task_struct *task; | |
2738 | int err; | |
0793a61d TG |
2739 | |
2740 | rcu_read_lock(); | |
2ebd4ffb | 2741 | if (!vpid) |
0793a61d TG |
2742 | task = current; |
2743 | else | |
2ebd4ffb | 2744 | task = find_task_by_vpid(vpid); |
0793a61d TG |
2745 | if (task) |
2746 | get_task_struct(task); | |
2747 | rcu_read_unlock(); | |
2748 | ||
2749 | if (!task) | |
2750 | return ERR_PTR(-ESRCH); | |
2751 | ||
0793a61d | 2752 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
2753 | err = -EACCES; |
2754 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
2755 | goto errout; | |
2756 | ||
2ebd4ffb MH |
2757 | return task; |
2758 | errout: | |
2759 | put_task_struct(task); | |
2760 | return ERR_PTR(err); | |
2761 | ||
2762 | } | |
2763 | ||
fe4b04fa PZ |
2764 | /* |
2765 | * Returns a matching context with refcount and pincount. | |
2766 | */ | |
108b02cf | 2767 | static struct perf_event_context * |
38a81da2 | 2768 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 2769 | { |
cdd6c482 | 2770 | struct perf_event_context *ctx; |
22a4f650 | 2771 | struct perf_cpu_context *cpuctx; |
25346b93 | 2772 | unsigned long flags; |
8dc85d54 | 2773 | int ctxn, err; |
0793a61d | 2774 | |
22a4ec72 | 2775 | if (!task) { |
cdd6c482 | 2776 | /* Must be root to operate on a CPU event: */ |
0764771d | 2777 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
2778 | return ERR_PTR(-EACCES); |
2779 | ||
0793a61d | 2780 | /* |
cdd6c482 | 2781 | * We could be clever and allow to attach a event to an |
0793a61d TG |
2782 | * offline CPU and activate it when the CPU comes up, but |
2783 | * that's for later. | |
2784 | */ | |
f6325e30 | 2785 | if (!cpu_online(cpu)) |
0793a61d TG |
2786 | return ERR_PTR(-ENODEV); |
2787 | ||
108b02cf | 2788 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 2789 | ctx = &cpuctx->ctx; |
c93f7669 | 2790 | get_ctx(ctx); |
fe4b04fa | 2791 | ++ctx->pin_count; |
0793a61d | 2792 | |
0793a61d TG |
2793 | return ctx; |
2794 | } | |
2795 | ||
8dc85d54 PZ |
2796 | err = -EINVAL; |
2797 | ctxn = pmu->task_ctx_nr; | |
2798 | if (ctxn < 0) | |
2799 | goto errout; | |
2800 | ||
9ed6060d | 2801 | retry: |
8dc85d54 | 2802 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 2803 | if (ctx) { |
71a851b4 | 2804 | unclone_ctx(ctx); |
fe4b04fa | 2805 | ++ctx->pin_count; |
e625cce1 | 2806 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 2807 | } else { |
eb184479 | 2808 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
2809 | err = -ENOMEM; |
2810 | if (!ctx) | |
2811 | goto errout; | |
eb184479 | 2812 | |
dbe08d82 ON |
2813 | err = 0; |
2814 | mutex_lock(&task->perf_event_mutex); | |
2815 | /* | |
2816 | * If it has already passed perf_event_exit_task(). | |
2817 | * we must see PF_EXITING, it takes this mutex too. | |
2818 | */ | |
2819 | if (task->flags & PF_EXITING) | |
2820 | err = -ESRCH; | |
2821 | else if (task->perf_event_ctxp[ctxn]) | |
2822 | err = -EAGAIN; | |
fe4b04fa | 2823 | else { |
9137fb28 | 2824 | get_ctx(ctx); |
fe4b04fa | 2825 | ++ctx->pin_count; |
dbe08d82 | 2826 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 2827 | } |
dbe08d82 ON |
2828 | mutex_unlock(&task->perf_event_mutex); |
2829 | ||
2830 | if (unlikely(err)) { | |
9137fb28 | 2831 | put_ctx(ctx); |
dbe08d82 ON |
2832 | |
2833 | if (err == -EAGAIN) | |
2834 | goto retry; | |
2835 | goto errout; | |
a63eaf34 PM |
2836 | } |
2837 | } | |
2838 | ||
0793a61d | 2839 | return ctx; |
c93f7669 | 2840 | |
9ed6060d | 2841 | errout: |
c93f7669 | 2842 | return ERR_PTR(err); |
0793a61d TG |
2843 | } |
2844 | ||
6fb2915d LZ |
2845 | static void perf_event_free_filter(struct perf_event *event); |
2846 | ||
cdd6c482 | 2847 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 2848 | { |
cdd6c482 | 2849 | struct perf_event *event; |
592903cd | 2850 | |
cdd6c482 IM |
2851 | event = container_of(head, struct perf_event, rcu_head); |
2852 | if (event->ns) | |
2853 | put_pid_ns(event->ns); | |
6fb2915d | 2854 | perf_event_free_filter(event); |
cdd6c482 | 2855 | kfree(event); |
592903cd PZ |
2856 | } |
2857 | ||
76369139 | 2858 | static void ring_buffer_put(struct ring_buffer *rb); |
925d519a | 2859 | |
cdd6c482 | 2860 | static void free_event(struct perf_event *event) |
f1600952 | 2861 | { |
e360adbe | 2862 | irq_work_sync(&event->pending); |
925d519a | 2863 | |
cdd6c482 | 2864 | if (!event->parent) { |
82cd6def | 2865 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 2866 | static_key_slow_dec_deferred(&perf_sched_events); |
3af9e859 | 2867 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
2868 | atomic_dec(&nr_mmap_events); |
2869 | if (event->attr.comm) | |
2870 | atomic_dec(&nr_comm_events); | |
2871 | if (event->attr.task) | |
2872 | atomic_dec(&nr_task_events); | |
927c7a9e FW |
2873 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
2874 | put_callchain_buffers(); | |
08309379 PZ |
2875 | if (is_cgroup_event(event)) { |
2876 | atomic_dec(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 2877 | static_key_slow_dec_deferred(&perf_sched_events); |
08309379 | 2878 | } |
d010b332 SE |
2879 | |
2880 | if (has_branch_stack(event)) { | |
2881 | static_key_slow_dec_deferred(&perf_sched_events); | |
2882 | /* is system-wide event */ | |
2883 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
2884 | atomic_dec(&per_cpu(perf_branch_stack_events, | |
2885 | event->cpu)); | |
2886 | } | |
f344011c | 2887 | } |
9ee318a7 | 2888 | |
76369139 FW |
2889 | if (event->rb) { |
2890 | ring_buffer_put(event->rb); | |
2891 | event->rb = NULL; | |
a4be7c27 PZ |
2892 | } |
2893 | ||
e5d1367f SE |
2894 | if (is_cgroup_event(event)) |
2895 | perf_detach_cgroup(event); | |
2896 | ||
cdd6c482 IM |
2897 | if (event->destroy) |
2898 | event->destroy(event); | |
e077df4f | 2899 | |
0c67b408 PZ |
2900 | if (event->ctx) |
2901 | put_ctx(event->ctx); | |
2902 | ||
cdd6c482 | 2903 | call_rcu(&event->rcu_head, free_event_rcu); |
f1600952 PZ |
2904 | } |
2905 | ||
a66a3052 | 2906 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 2907 | { |
cdd6c482 | 2908 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 2909 | |
ad3a37de | 2910 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
2911 | /* |
2912 | * There are two ways this annotation is useful: | |
2913 | * | |
2914 | * 1) there is a lock recursion from perf_event_exit_task | |
2915 | * see the comment there. | |
2916 | * | |
2917 | * 2) there is a lock-inversion with mmap_sem through | |
2918 | * perf_event_read_group(), which takes faults while | |
2919 | * holding ctx->mutex, however this is called after | |
2920 | * the last filedesc died, so there is no possibility | |
2921 | * to trigger the AB-BA case. | |
2922 | */ | |
2923 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
050735b0 | 2924 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 2925 | perf_group_detach(event); |
050735b0 | 2926 | raw_spin_unlock_irq(&ctx->lock); |
e03a9a55 | 2927 | perf_remove_from_context(event); |
d859e29f | 2928 | mutex_unlock(&ctx->mutex); |
0793a61d | 2929 | |
cdd6c482 | 2930 | free_event(event); |
0793a61d TG |
2931 | |
2932 | return 0; | |
2933 | } | |
a66a3052 | 2934 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 2935 | |
a66a3052 PZ |
2936 | /* |
2937 | * Called when the last reference to the file is gone. | |
2938 | */ | |
a6fa941d | 2939 | static void put_event(struct perf_event *event) |
fb0459d7 | 2940 | { |
8882135b | 2941 | struct task_struct *owner; |
fb0459d7 | 2942 | |
a6fa941d AV |
2943 | if (!atomic_long_dec_and_test(&event->refcount)) |
2944 | return; | |
fb0459d7 | 2945 | |
8882135b PZ |
2946 | rcu_read_lock(); |
2947 | owner = ACCESS_ONCE(event->owner); | |
2948 | /* | |
2949 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
2950 | * !owner it means the list deletion is complete and we can indeed | |
2951 | * free this event, otherwise we need to serialize on | |
2952 | * owner->perf_event_mutex. | |
2953 | */ | |
2954 | smp_read_barrier_depends(); | |
2955 | if (owner) { | |
2956 | /* | |
2957 | * Since delayed_put_task_struct() also drops the last | |
2958 | * task reference we can safely take a new reference | |
2959 | * while holding the rcu_read_lock(). | |
2960 | */ | |
2961 | get_task_struct(owner); | |
2962 | } | |
2963 | rcu_read_unlock(); | |
2964 | ||
2965 | if (owner) { | |
2966 | mutex_lock(&owner->perf_event_mutex); | |
2967 | /* | |
2968 | * We have to re-check the event->owner field, if it is cleared | |
2969 | * we raced with perf_event_exit_task(), acquiring the mutex | |
2970 | * ensured they're done, and we can proceed with freeing the | |
2971 | * event. | |
2972 | */ | |
2973 | if (event->owner) | |
2974 | list_del_init(&event->owner_entry); | |
2975 | mutex_unlock(&owner->perf_event_mutex); | |
2976 | put_task_struct(owner); | |
2977 | } | |
2978 | ||
a6fa941d AV |
2979 | perf_event_release_kernel(event); |
2980 | } | |
2981 | ||
2982 | static int perf_release(struct inode *inode, struct file *file) | |
2983 | { | |
2984 | put_event(file->private_data); | |
2985 | return 0; | |
fb0459d7 | 2986 | } |
fb0459d7 | 2987 | |
59ed446f | 2988 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 2989 | { |
cdd6c482 | 2990 | struct perf_event *child; |
e53c0994 PZ |
2991 | u64 total = 0; |
2992 | ||
59ed446f PZ |
2993 | *enabled = 0; |
2994 | *running = 0; | |
2995 | ||
6f10581a | 2996 | mutex_lock(&event->child_mutex); |
cdd6c482 | 2997 | total += perf_event_read(event); |
59ed446f PZ |
2998 | *enabled += event->total_time_enabled + |
2999 | atomic64_read(&event->child_total_time_enabled); | |
3000 | *running += event->total_time_running + | |
3001 | atomic64_read(&event->child_total_time_running); | |
3002 | ||
3003 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3004 | total += perf_event_read(child); |
59ed446f PZ |
3005 | *enabled += child->total_time_enabled; |
3006 | *running += child->total_time_running; | |
3007 | } | |
6f10581a | 3008 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3009 | |
3010 | return total; | |
3011 | } | |
fb0459d7 | 3012 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3013 | |
cdd6c482 | 3014 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3015 | u64 read_format, char __user *buf) |
3016 | { | |
cdd6c482 | 3017 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3018 | int n = 0, size = 0, ret = -EFAULT; |
3019 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3020 | u64 values[5]; |
59ed446f | 3021 | u64 count, enabled, running; |
abf4868b | 3022 | |
6f10581a | 3023 | mutex_lock(&ctx->mutex); |
59ed446f | 3024 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3025 | |
3026 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3027 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3028 | values[n++] = enabled; | |
3029 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3030 | values[n++] = running; | |
abf4868b PZ |
3031 | values[n++] = count; |
3032 | if (read_format & PERF_FORMAT_ID) | |
3033 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3034 | |
3035 | size = n * sizeof(u64); | |
3036 | ||
3037 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3038 | goto unlock; |
3dab77fb | 3039 | |
6f10581a | 3040 | ret = size; |
3dab77fb | 3041 | |
65abc865 | 3042 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3043 | n = 0; |
3dab77fb | 3044 | |
59ed446f | 3045 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3046 | if (read_format & PERF_FORMAT_ID) |
3047 | values[n++] = primary_event_id(sub); | |
3048 | ||
3049 | size = n * sizeof(u64); | |
3050 | ||
184d3da8 | 3051 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3052 | ret = -EFAULT; |
3053 | goto unlock; | |
3054 | } | |
abf4868b PZ |
3055 | |
3056 | ret += size; | |
3dab77fb | 3057 | } |
6f10581a PZ |
3058 | unlock: |
3059 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3060 | |
abf4868b | 3061 | return ret; |
3dab77fb PZ |
3062 | } |
3063 | ||
cdd6c482 | 3064 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3065 | u64 read_format, char __user *buf) |
3066 | { | |
59ed446f | 3067 | u64 enabled, running; |
3dab77fb PZ |
3068 | u64 values[4]; |
3069 | int n = 0; | |
3070 | ||
59ed446f PZ |
3071 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3072 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3073 | values[n++] = enabled; | |
3074 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3075 | values[n++] = running; | |
3dab77fb | 3076 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3077 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3078 | |
3079 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3080 | return -EFAULT; | |
3081 | ||
3082 | return n * sizeof(u64); | |
3083 | } | |
3084 | ||
0793a61d | 3085 | /* |
cdd6c482 | 3086 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3087 | */ |
3088 | static ssize_t | |
cdd6c482 | 3089 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3090 | { |
cdd6c482 | 3091 | u64 read_format = event->attr.read_format; |
3dab77fb | 3092 | int ret; |
0793a61d | 3093 | |
3b6f9e5c | 3094 | /* |
cdd6c482 | 3095 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3096 | * error state (i.e. because it was pinned but it couldn't be |
3097 | * scheduled on to the CPU at some point). | |
3098 | */ | |
cdd6c482 | 3099 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3100 | return 0; |
3101 | ||
c320c7b7 | 3102 | if (count < event->read_size) |
3dab77fb PZ |
3103 | return -ENOSPC; |
3104 | ||
cdd6c482 | 3105 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3106 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3107 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3108 | else |
cdd6c482 | 3109 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3110 | |
3dab77fb | 3111 | return ret; |
0793a61d TG |
3112 | } |
3113 | ||
0793a61d TG |
3114 | static ssize_t |
3115 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3116 | { | |
cdd6c482 | 3117 | struct perf_event *event = file->private_data; |
0793a61d | 3118 | |
cdd6c482 | 3119 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3120 | } |
3121 | ||
3122 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3123 | { | |
cdd6c482 | 3124 | struct perf_event *event = file->private_data; |
76369139 | 3125 | struct ring_buffer *rb; |
c33a0bc4 | 3126 | unsigned int events = POLL_HUP; |
c7138f37 | 3127 | |
10c6db11 PZ |
3128 | /* |
3129 | * Race between perf_event_set_output() and perf_poll(): perf_poll() | |
3130 | * grabs the rb reference but perf_event_set_output() overrides it. | |
3131 | * Here is the timeline for two threads T1, T2: | |
3132 | * t0: T1, rb = rcu_dereference(event->rb) | |
3133 | * t1: T2, old_rb = event->rb | |
3134 | * t2: T2, event->rb = new rb | |
3135 | * t3: T2, ring_buffer_detach(old_rb) | |
3136 | * t4: T1, ring_buffer_attach(rb1) | |
3137 | * t5: T1, poll_wait(event->waitq) | |
3138 | * | |
3139 | * To avoid this problem, we grab mmap_mutex in perf_poll() | |
3140 | * thereby ensuring that the assignment of the new ring buffer | |
3141 | * and the detachment of the old buffer appear atomic to perf_poll() | |
3142 | */ | |
3143 | mutex_lock(&event->mmap_mutex); | |
3144 | ||
c7138f37 | 3145 | rcu_read_lock(); |
76369139 | 3146 | rb = rcu_dereference(event->rb); |
10c6db11 PZ |
3147 | if (rb) { |
3148 | ring_buffer_attach(event, rb); | |
76369139 | 3149 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3150 | } |
c7138f37 | 3151 | rcu_read_unlock(); |
0793a61d | 3152 | |
10c6db11 PZ |
3153 | mutex_unlock(&event->mmap_mutex); |
3154 | ||
cdd6c482 | 3155 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3156 | |
0793a61d TG |
3157 | return events; |
3158 | } | |
3159 | ||
cdd6c482 | 3160 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3161 | { |
cdd6c482 | 3162 | (void)perf_event_read(event); |
e7850595 | 3163 | local64_set(&event->count, 0); |
cdd6c482 | 3164 | perf_event_update_userpage(event); |
3df5edad PZ |
3165 | } |
3166 | ||
c93f7669 | 3167 | /* |
cdd6c482 IM |
3168 | * Holding the top-level event's child_mutex means that any |
3169 | * descendant process that has inherited this event will block | |
3170 | * in sync_child_event if it goes to exit, thus satisfying the | |
3171 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3172 | */ |
cdd6c482 IM |
3173 | static void perf_event_for_each_child(struct perf_event *event, |
3174 | void (*func)(struct perf_event *)) | |
3df5edad | 3175 | { |
cdd6c482 | 3176 | struct perf_event *child; |
3df5edad | 3177 | |
cdd6c482 IM |
3178 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3179 | mutex_lock(&event->child_mutex); | |
3180 | func(event); | |
3181 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3182 | func(child); |
cdd6c482 | 3183 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3184 | } |
3185 | ||
cdd6c482 IM |
3186 | static void perf_event_for_each(struct perf_event *event, |
3187 | void (*func)(struct perf_event *)) | |
3df5edad | 3188 | { |
cdd6c482 IM |
3189 | struct perf_event_context *ctx = event->ctx; |
3190 | struct perf_event *sibling; | |
3df5edad | 3191 | |
75f937f2 PZ |
3192 | WARN_ON_ONCE(ctx->parent_ctx); |
3193 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3194 | event = event->group_leader; |
75f937f2 | 3195 | |
cdd6c482 | 3196 | perf_event_for_each_child(event, func); |
cdd6c482 | 3197 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3198 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3199 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3200 | } |
3201 | ||
cdd6c482 | 3202 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3203 | { |
cdd6c482 | 3204 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3205 | int ret = 0; |
3206 | u64 value; | |
3207 | ||
6c7e550f | 3208 | if (!is_sampling_event(event)) |
08247e31 PZ |
3209 | return -EINVAL; |
3210 | ||
ad0cf347 | 3211 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3212 | return -EFAULT; |
3213 | ||
3214 | if (!value) | |
3215 | return -EINVAL; | |
3216 | ||
e625cce1 | 3217 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3218 | if (event->attr.freq) { |
3219 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3220 | ret = -EINVAL; |
3221 | goto unlock; | |
3222 | } | |
3223 | ||
cdd6c482 | 3224 | event->attr.sample_freq = value; |
08247e31 | 3225 | } else { |
cdd6c482 IM |
3226 | event->attr.sample_period = value; |
3227 | event->hw.sample_period = value; | |
08247e31 PZ |
3228 | } |
3229 | unlock: | |
e625cce1 | 3230 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3231 | |
3232 | return ret; | |
3233 | } | |
3234 | ||
ac9721f3 PZ |
3235 | static const struct file_operations perf_fops; |
3236 | ||
a6fa941d | 3237 | static struct file *perf_fget_light(int fd, int *fput_needed) |
ac9721f3 PZ |
3238 | { |
3239 | struct file *file; | |
3240 | ||
3241 | file = fget_light(fd, fput_needed); | |
3242 | if (!file) | |
3243 | return ERR_PTR(-EBADF); | |
3244 | ||
3245 | if (file->f_op != &perf_fops) { | |
3246 | fput_light(file, *fput_needed); | |
3247 | *fput_needed = 0; | |
3248 | return ERR_PTR(-EBADF); | |
3249 | } | |
3250 | ||
a6fa941d | 3251 | return file; |
ac9721f3 PZ |
3252 | } |
3253 | ||
3254 | static int perf_event_set_output(struct perf_event *event, | |
3255 | struct perf_event *output_event); | |
6fb2915d | 3256 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3257 | |
d859e29f PM |
3258 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3259 | { | |
cdd6c482 IM |
3260 | struct perf_event *event = file->private_data; |
3261 | void (*func)(struct perf_event *); | |
3df5edad | 3262 | u32 flags = arg; |
d859e29f PM |
3263 | |
3264 | switch (cmd) { | |
cdd6c482 IM |
3265 | case PERF_EVENT_IOC_ENABLE: |
3266 | func = perf_event_enable; | |
d859e29f | 3267 | break; |
cdd6c482 IM |
3268 | case PERF_EVENT_IOC_DISABLE: |
3269 | func = perf_event_disable; | |
79f14641 | 3270 | break; |
cdd6c482 IM |
3271 | case PERF_EVENT_IOC_RESET: |
3272 | func = perf_event_reset; | |
6de6a7b9 | 3273 | break; |
3df5edad | 3274 | |
cdd6c482 IM |
3275 | case PERF_EVENT_IOC_REFRESH: |
3276 | return perf_event_refresh(event, arg); | |
08247e31 | 3277 | |
cdd6c482 IM |
3278 | case PERF_EVENT_IOC_PERIOD: |
3279 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3280 | |
cdd6c482 | 3281 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3282 | { |
a6fa941d | 3283 | struct file *output_file = NULL; |
ac9721f3 PZ |
3284 | struct perf_event *output_event = NULL; |
3285 | int fput_needed = 0; | |
3286 | int ret; | |
3287 | ||
3288 | if (arg != -1) { | |
a6fa941d AV |
3289 | output_file = perf_fget_light(arg, &fput_needed); |
3290 | if (IS_ERR(output_file)) | |
3291 | return PTR_ERR(output_file); | |
3292 | output_event = output_file->private_data; | |
ac9721f3 PZ |
3293 | } |
3294 | ||
3295 | ret = perf_event_set_output(event, output_event); | |
3296 | if (output_event) | |
a6fa941d | 3297 | fput_light(output_file, fput_needed); |
ac9721f3 PZ |
3298 | |
3299 | return ret; | |
3300 | } | |
a4be7c27 | 3301 | |
6fb2915d LZ |
3302 | case PERF_EVENT_IOC_SET_FILTER: |
3303 | return perf_event_set_filter(event, (void __user *)arg); | |
3304 | ||
d859e29f | 3305 | default: |
3df5edad | 3306 | return -ENOTTY; |
d859e29f | 3307 | } |
3df5edad PZ |
3308 | |
3309 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3310 | perf_event_for_each(event, func); |
3df5edad | 3311 | else |
cdd6c482 | 3312 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3313 | |
3314 | return 0; | |
d859e29f PM |
3315 | } |
3316 | ||
cdd6c482 | 3317 | int perf_event_task_enable(void) |
771d7cde | 3318 | { |
cdd6c482 | 3319 | struct perf_event *event; |
771d7cde | 3320 | |
cdd6c482 IM |
3321 | mutex_lock(¤t->perf_event_mutex); |
3322 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3323 | perf_event_for_each_child(event, perf_event_enable); | |
3324 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3325 | |
3326 | return 0; | |
3327 | } | |
3328 | ||
cdd6c482 | 3329 | int perf_event_task_disable(void) |
771d7cde | 3330 | { |
cdd6c482 | 3331 | struct perf_event *event; |
771d7cde | 3332 | |
cdd6c482 IM |
3333 | mutex_lock(¤t->perf_event_mutex); |
3334 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3335 | perf_event_for_each_child(event, perf_event_disable); | |
3336 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3337 | |
3338 | return 0; | |
3339 | } | |
3340 | ||
cdd6c482 | 3341 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3342 | { |
a4eaf7f1 PZ |
3343 | if (event->hw.state & PERF_HES_STOPPED) |
3344 | return 0; | |
3345 | ||
cdd6c482 | 3346 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3347 | return 0; |
3348 | ||
35edc2a5 | 3349 | return event->pmu->event_idx(event); |
194002b2 PZ |
3350 | } |
3351 | ||
c4794295 | 3352 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3353 | u64 *now, |
7f310a5d EM |
3354 | u64 *enabled, |
3355 | u64 *running) | |
c4794295 | 3356 | { |
e3f3541c | 3357 | u64 ctx_time; |
c4794295 | 3358 | |
e3f3541c PZ |
3359 | *now = perf_clock(); |
3360 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3361 | *enabled = ctx_time - event->tstamp_enabled; |
3362 | *running = ctx_time - event->tstamp_running; | |
3363 | } | |
3364 | ||
c7206205 | 3365 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3366 | { |
3367 | } | |
3368 | ||
38ff667b PZ |
3369 | /* |
3370 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3371 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3372 | * code calls this from NMI context. | |
3373 | */ | |
cdd6c482 | 3374 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3375 | { |
cdd6c482 | 3376 | struct perf_event_mmap_page *userpg; |
76369139 | 3377 | struct ring_buffer *rb; |
e3f3541c | 3378 | u64 enabled, running, now; |
38ff667b PZ |
3379 | |
3380 | rcu_read_lock(); | |
0d641208 EM |
3381 | /* |
3382 | * compute total_time_enabled, total_time_running | |
3383 | * based on snapshot values taken when the event | |
3384 | * was last scheduled in. | |
3385 | * | |
3386 | * we cannot simply called update_context_time() | |
3387 | * because of locking issue as we can be called in | |
3388 | * NMI context | |
3389 | */ | |
e3f3541c | 3390 | calc_timer_values(event, &now, &enabled, &running); |
76369139 FW |
3391 | rb = rcu_dereference(event->rb); |
3392 | if (!rb) | |
38ff667b PZ |
3393 | goto unlock; |
3394 | ||
76369139 | 3395 | userpg = rb->user_page; |
37d81828 | 3396 | |
7b732a75 PZ |
3397 | /* |
3398 | * Disable preemption so as to not let the corresponding user-space | |
3399 | * spin too long if we get preempted. | |
3400 | */ | |
3401 | preempt_disable(); | |
37d81828 | 3402 | ++userpg->lock; |
92f22a38 | 3403 | barrier(); |
cdd6c482 | 3404 | userpg->index = perf_event_index(event); |
b5e58793 | 3405 | userpg->offset = perf_event_count(event); |
365a4038 | 3406 | if (userpg->index) |
e7850595 | 3407 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3408 | |
0d641208 | 3409 | userpg->time_enabled = enabled + |
cdd6c482 | 3410 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3411 | |
0d641208 | 3412 | userpg->time_running = running + |
cdd6c482 | 3413 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3414 | |
c7206205 | 3415 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3416 | |
92f22a38 | 3417 | barrier(); |
37d81828 | 3418 | ++userpg->lock; |
7b732a75 | 3419 | preempt_enable(); |
38ff667b | 3420 | unlock: |
7b732a75 | 3421 | rcu_read_unlock(); |
37d81828 PM |
3422 | } |
3423 | ||
906010b2 PZ |
3424 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3425 | { | |
3426 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3427 | struct ring_buffer *rb; |
906010b2 PZ |
3428 | int ret = VM_FAULT_SIGBUS; |
3429 | ||
3430 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3431 | if (vmf->pgoff == 0) | |
3432 | ret = 0; | |
3433 | return ret; | |
3434 | } | |
3435 | ||
3436 | rcu_read_lock(); | |
76369139 FW |
3437 | rb = rcu_dereference(event->rb); |
3438 | if (!rb) | |
906010b2 PZ |
3439 | goto unlock; |
3440 | ||
3441 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3442 | goto unlock; | |
3443 | ||
76369139 | 3444 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3445 | if (!vmf->page) |
3446 | goto unlock; | |
3447 | ||
3448 | get_page(vmf->page); | |
3449 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3450 | vmf->page->index = vmf->pgoff; | |
3451 | ||
3452 | ret = 0; | |
3453 | unlock: | |
3454 | rcu_read_unlock(); | |
3455 | ||
3456 | return ret; | |
3457 | } | |
3458 | ||
10c6db11 PZ |
3459 | static void ring_buffer_attach(struct perf_event *event, |
3460 | struct ring_buffer *rb) | |
3461 | { | |
3462 | unsigned long flags; | |
3463 | ||
3464 | if (!list_empty(&event->rb_entry)) | |
3465 | return; | |
3466 | ||
3467 | spin_lock_irqsave(&rb->event_lock, flags); | |
3468 | if (!list_empty(&event->rb_entry)) | |
3469 | goto unlock; | |
3470 | ||
3471 | list_add(&event->rb_entry, &rb->event_list); | |
3472 | unlock: | |
3473 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3474 | } | |
3475 | ||
3476 | static void ring_buffer_detach(struct perf_event *event, | |
3477 | struct ring_buffer *rb) | |
3478 | { | |
3479 | unsigned long flags; | |
3480 | ||
3481 | if (list_empty(&event->rb_entry)) | |
3482 | return; | |
3483 | ||
3484 | spin_lock_irqsave(&rb->event_lock, flags); | |
3485 | list_del_init(&event->rb_entry); | |
3486 | wake_up_all(&event->waitq); | |
3487 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3488 | } | |
3489 | ||
3490 | static void ring_buffer_wakeup(struct perf_event *event) | |
3491 | { | |
3492 | struct ring_buffer *rb; | |
3493 | ||
3494 | rcu_read_lock(); | |
3495 | rb = rcu_dereference(event->rb); | |
44b7f4b9 WD |
3496 | if (!rb) |
3497 | goto unlock; | |
3498 | ||
3499 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
10c6db11 | 3500 | wake_up_all(&event->waitq); |
44b7f4b9 WD |
3501 | |
3502 | unlock: | |
10c6db11 PZ |
3503 | rcu_read_unlock(); |
3504 | } | |
3505 | ||
76369139 | 3506 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3507 | { |
76369139 | 3508 | struct ring_buffer *rb; |
906010b2 | 3509 | |
76369139 FW |
3510 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3511 | rb_free(rb); | |
7b732a75 PZ |
3512 | } |
3513 | ||
76369139 | 3514 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3515 | { |
76369139 | 3516 | struct ring_buffer *rb; |
7b732a75 | 3517 | |
ac9721f3 | 3518 | rcu_read_lock(); |
76369139 FW |
3519 | rb = rcu_dereference(event->rb); |
3520 | if (rb) { | |
3521 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3522 | rb = NULL; | |
ac9721f3 PZ |
3523 | } |
3524 | rcu_read_unlock(); | |
3525 | ||
76369139 | 3526 | return rb; |
ac9721f3 PZ |
3527 | } |
3528 | ||
76369139 | 3529 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3530 | { |
10c6db11 PZ |
3531 | struct perf_event *event, *n; |
3532 | unsigned long flags; | |
3533 | ||
76369139 | 3534 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3535 | return; |
7b732a75 | 3536 | |
10c6db11 PZ |
3537 | spin_lock_irqsave(&rb->event_lock, flags); |
3538 | list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) { | |
3539 | list_del_init(&event->rb_entry); | |
3540 | wake_up_all(&event->waitq); | |
3541 | } | |
3542 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3543 | ||
76369139 | 3544 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3545 | } |
3546 | ||
3547 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3548 | { | |
cdd6c482 | 3549 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3550 | |
cdd6c482 | 3551 | atomic_inc(&event->mmap_count); |
7b732a75 PZ |
3552 | } |
3553 | ||
3554 | static void perf_mmap_close(struct vm_area_struct *vma) | |
3555 | { | |
cdd6c482 | 3556 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3557 | |
cdd6c482 | 3558 | if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) { |
76369139 | 3559 | unsigned long size = perf_data_size(event->rb); |
ac9721f3 | 3560 | struct user_struct *user = event->mmap_user; |
76369139 | 3561 | struct ring_buffer *rb = event->rb; |
789f90fc | 3562 | |
906010b2 | 3563 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm); |
bc3e53f6 | 3564 | vma->vm_mm->pinned_vm -= event->mmap_locked; |
76369139 | 3565 | rcu_assign_pointer(event->rb, NULL); |
10c6db11 | 3566 | ring_buffer_detach(event, rb); |
cdd6c482 | 3567 | mutex_unlock(&event->mmap_mutex); |
ac9721f3 | 3568 | |
76369139 | 3569 | ring_buffer_put(rb); |
ac9721f3 | 3570 | free_uid(user); |
7b732a75 | 3571 | } |
37d81828 PM |
3572 | } |
3573 | ||
f0f37e2f | 3574 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3575 | .open = perf_mmap_open, |
3576 | .close = perf_mmap_close, | |
3577 | .fault = perf_mmap_fault, | |
3578 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3579 | }; |
3580 | ||
3581 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3582 | { | |
cdd6c482 | 3583 | struct perf_event *event = file->private_data; |
22a4f650 | 3584 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3585 | struct user_struct *user = current_user(); |
22a4f650 | 3586 | unsigned long locked, lock_limit; |
76369139 | 3587 | struct ring_buffer *rb; |
7b732a75 PZ |
3588 | unsigned long vma_size; |
3589 | unsigned long nr_pages; | |
789f90fc | 3590 | long user_extra, extra; |
d57e34fd | 3591 | int ret = 0, flags = 0; |
37d81828 | 3592 | |
c7920614 PZ |
3593 | /* |
3594 | * Don't allow mmap() of inherited per-task counters. This would | |
3595 | * create a performance issue due to all children writing to the | |
76369139 | 3596 | * same rb. |
c7920614 PZ |
3597 | */ |
3598 | if (event->cpu == -1 && event->attr.inherit) | |
3599 | return -EINVAL; | |
3600 | ||
43a21ea8 | 3601 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3602 | return -EINVAL; |
7b732a75 PZ |
3603 | |
3604 | vma_size = vma->vm_end - vma->vm_start; | |
3605 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3606 | ||
7730d865 | 3607 | /* |
76369139 | 3608 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3609 | * can do bitmasks instead of modulo. |
3610 | */ | |
3611 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3612 | return -EINVAL; |
3613 | ||
7b732a75 | 3614 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3615 | return -EINVAL; |
3616 | ||
7b732a75 PZ |
3617 | if (vma->vm_pgoff != 0) |
3618 | return -EINVAL; | |
37d81828 | 3619 | |
cdd6c482 IM |
3620 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3621 | mutex_lock(&event->mmap_mutex); | |
76369139 FW |
3622 | if (event->rb) { |
3623 | if (event->rb->nr_pages == nr_pages) | |
3624 | atomic_inc(&event->rb->refcount); | |
ac9721f3 | 3625 | else |
ebb3c4c4 PZ |
3626 | ret = -EINVAL; |
3627 | goto unlock; | |
3628 | } | |
3629 | ||
789f90fc | 3630 | user_extra = nr_pages + 1; |
cdd6c482 | 3631 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
3632 | |
3633 | /* | |
3634 | * Increase the limit linearly with more CPUs: | |
3635 | */ | |
3636 | user_lock_limit *= num_online_cpus(); | |
3637 | ||
789f90fc | 3638 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 3639 | |
789f90fc PZ |
3640 | extra = 0; |
3641 | if (user_locked > user_lock_limit) | |
3642 | extra = user_locked - user_lock_limit; | |
7b732a75 | 3643 | |
78d7d407 | 3644 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 3645 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 3646 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 3647 | |
459ec28a IM |
3648 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
3649 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
3650 | ret = -EPERM; |
3651 | goto unlock; | |
3652 | } | |
7b732a75 | 3653 | |
76369139 | 3654 | WARN_ON(event->rb); |
906010b2 | 3655 | |
d57e34fd | 3656 | if (vma->vm_flags & VM_WRITE) |
76369139 | 3657 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 3658 | |
4ec8363d VW |
3659 | rb = rb_alloc(nr_pages, |
3660 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
3661 | event->cpu, flags); | |
3662 | ||
76369139 | 3663 | if (!rb) { |
ac9721f3 | 3664 | ret = -ENOMEM; |
ebb3c4c4 | 3665 | goto unlock; |
ac9721f3 | 3666 | } |
76369139 | 3667 | rcu_assign_pointer(event->rb, rb); |
43a21ea8 | 3668 | |
ac9721f3 PZ |
3669 | atomic_long_add(user_extra, &user->locked_vm); |
3670 | event->mmap_locked = extra; | |
3671 | event->mmap_user = get_current_user(); | |
bc3e53f6 | 3672 | vma->vm_mm->pinned_vm += event->mmap_locked; |
ac9721f3 | 3673 | |
9a0f05cb PZ |
3674 | perf_event_update_userpage(event); |
3675 | ||
ebb3c4c4 | 3676 | unlock: |
ac9721f3 PZ |
3677 | if (!ret) |
3678 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 3679 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 3680 | |
37d81828 PM |
3681 | vma->vm_flags |= VM_RESERVED; |
3682 | vma->vm_ops = &perf_mmap_vmops; | |
7b732a75 PZ |
3683 | |
3684 | return ret; | |
37d81828 PM |
3685 | } |
3686 | ||
3c446b3d PZ |
3687 | static int perf_fasync(int fd, struct file *filp, int on) |
3688 | { | |
3c446b3d | 3689 | struct inode *inode = filp->f_path.dentry->d_inode; |
cdd6c482 | 3690 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
3691 | int retval; |
3692 | ||
3693 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 3694 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
3695 | mutex_unlock(&inode->i_mutex); |
3696 | ||
3697 | if (retval < 0) | |
3698 | return retval; | |
3699 | ||
3700 | return 0; | |
3701 | } | |
3702 | ||
0793a61d | 3703 | static const struct file_operations perf_fops = { |
3326c1ce | 3704 | .llseek = no_llseek, |
0793a61d TG |
3705 | .release = perf_release, |
3706 | .read = perf_read, | |
3707 | .poll = perf_poll, | |
d859e29f PM |
3708 | .unlocked_ioctl = perf_ioctl, |
3709 | .compat_ioctl = perf_ioctl, | |
37d81828 | 3710 | .mmap = perf_mmap, |
3c446b3d | 3711 | .fasync = perf_fasync, |
0793a61d TG |
3712 | }; |
3713 | ||
925d519a | 3714 | /* |
cdd6c482 | 3715 | * Perf event wakeup |
925d519a PZ |
3716 | * |
3717 | * If there's data, ensure we set the poll() state and publish everything | |
3718 | * to user-space before waking everybody up. | |
3719 | */ | |
3720 | ||
cdd6c482 | 3721 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 3722 | { |
10c6db11 | 3723 | ring_buffer_wakeup(event); |
4c9e2542 | 3724 | |
cdd6c482 IM |
3725 | if (event->pending_kill) { |
3726 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
3727 | event->pending_kill = 0; | |
4c9e2542 | 3728 | } |
925d519a PZ |
3729 | } |
3730 | ||
e360adbe | 3731 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 3732 | { |
cdd6c482 IM |
3733 | struct perf_event *event = container_of(entry, |
3734 | struct perf_event, pending); | |
79f14641 | 3735 | |
cdd6c482 IM |
3736 | if (event->pending_disable) { |
3737 | event->pending_disable = 0; | |
3738 | __perf_event_disable(event); | |
79f14641 PZ |
3739 | } |
3740 | ||
cdd6c482 IM |
3741 | if (event->pending_wakeup) { |
3742 | event->pending_wakeup = 0; | |
3743 | perf_event_wakeup(event); | |
79f14641 PZ |
3744 | } |
3745 | } | |
3746 | ||
39447b38 ZY |
3747 | /* |
3748 | * We assume there is only KVM supporting the callbacks. | |
3749 | * Later on, we might change it to a list if there is | |
3750 | * another virtualization implementation supporting the callbacks. | |
3751 | */ | |
3752 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
3753 | ||
3754 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3755 | { | |
3756 | perf_guest_cbs = cbs; | |
3757 | return 0; | |
3758 | } | |
3759 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
3760 | ||
3761 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3762 | { | |
3763 | perf_guest_cbs = NULL; | |
3764 | return 0; | |
3765 | } | |
3766 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
3767 | ||
4018994f JO |
3768 | static void |
3769 | perf_output_sample_regs(struct perf_output_handle *handle, | |
3770 | struct pt_regs *regs, u64 mask) | |
3771 | { | |
3772 | int bit; | |
3773 | ||
3774 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
3775 | sizeof(mask) * BITS_PER_BYTE) { | |
3776 | u64 val; | |
3777 | ||
3778 | val = perf_reg_value(regs, bit); | |
3779 | perf_output_put(handle, val); | |
3780 | } | |
3781 | } | |
3782 | ||
3783 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
3784 | struct pt_regs *regs) | |
3785 | { | |
3786 | if (!user_mode(regs)) { | |
3787 | if (current->mm) | |
3788 | regs = task_pt_regs(current); | |
3789 | else | |
3790 | regs = NULL; | |
3791 | } | |
3792 | ||
3793 | if (regs) { | |
3794 | regs_user->regs = regs; | |
3795 | regs_user->abi = perf_reg_abi(current); | |
3796 | } | |
3797 | } | |
3798 | ||
c5ebcedb JO |
3799 | /* |
3800 | * Get remaining task size from user stack pointer. | |
3801 | * | |
3802 | * It'd be better to take stack vma map and limit this more | |
3803 | * precisly, but there's no way to get it safely under interrupt, | |
3804 | * so using TASK_SIZE as limit. | |
3805 | */ | |
3806 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
3807 | { | |
3808 | unsigned long addr = perf_user_stack_pointer(regs); | |
3809 | ||
3810 | if (!addr || addr >= TASK_SIZE) | |
3811 | return 0; | |
3812 | ||
3813 | return TASK_SIZE - addr; | |
3814 | } | |
3815 | ||
3816 | static u16 | |
3817 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
3818 | struct pt_regs *regs) | |
3819 | { | |
3820 | u64 task_size; | |
3821 | ||
3822 | /* No regs, no stack pointer, no dump. */ | |
3823 | if (!regs) | |
3824 | return 0; | |
3825 | ||
3826 | /* | |
3827 | * Check if we fit in with the requested stack size into the: | |
3828 | * - TASK_SIZE | |
3829 | * If we don't, we limit the size to the TASK_SIZE. | |
3830 | * | |
3831 | * - remaining sample size | |
3832 | * If we don't, we customize the stack size to | |
3833 | * fit in to the remaining sample size. | |
3834 | */ | |
3835 | ||
3836 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
3837 | stack_size = min(stack_size, (u16) task_size); | |
3838 | ||
3839 | /* Current header size plus static size and dynamic size. */ | |
3840 | header_size += 2 * sizeof(u64); | |
3841 | ||
3842 | /* Do we fit in with the current stack dump size? */ | |
3843 | if ((u16) (header_size + stack_size) < header_size) { | |
3844 | /* | |
3845 | * If we overflow the maximum size for the sample, | |
3846 | * we customize the stack dump size to fit in. | |
3847 | */ | |
3848 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
3849 | stack_size = round_up(stack_size, sizeof(u64)); | |
3850 | } | |
3851 | ||
3852 | return stack_size; | |
3853 | } | |
3854 | ||
3855 | static void | |
3856 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
3857 | struct pt_regs *regs) | |
3858 | { | |
3859 | /* Case of a kernel thread, nothing to dump */ | |
3860 | if (!regs) { | |
3861 | u64 size = 0; | |
3862 | perf_output_put(handle, size); | |
3863 | } else { | |
3864 | unsigned long sp; | |
3865 | unsigned int rem; | |
3866 | u64 dyn_size; | |
3867 | ||
3868 | /* | |
3869 | * We dump: | |
3870 | * static size | |
3871 | * - the size requested by user or the best one we can fit | |
3872 | * in to the sample max size | |
3873 | * data | |
3874 | * - user stack dump data | |
3875 | * dynamic size | |
3876 | * - the actual dumped size | |
3877 | */ | |
3878 | ||
3879 | /* Static size. */ | |
3880 | perf_output_put(handle, dump_size); | |
3881 | ||
3882 | /* Data. */ | |
3883 | sp = perf_user_stack_pointer(regs); | |
3884 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
3885 | dyn_size = dump_size - rem; | |
3886 | ||
3887 | perf_output_skip(handle, rem); | |
3888 | ||
3889 | /* Dynamic size. */ | |
3890 | perf_output_put(handle, dyn_size); | |
3891 | } | |
3892 | } | |
3893 | ||
c980d109 ACM |
3894 | static void __perf_event_header__init_id(struct perf_event_header *header, |
3895 | struct perf_sample_data *data, | |
3896 | struct perf_event *event) | |
6844c09d ACM |
3897 | { |
3898 | u64 sample_type = event->attr.sample_type; | |
3899 | ||
3900 | data->type = sample_type; | |
3901 | header->size += event->id_header_size; | |
3902 | ||
3903 | if (sample_type & PERF_SAMPLE_TID) { | |
3904 | /* namespace issues */ | |
3905 | data->tid_entry.pid = perf_event_pid(event, current); | |
3906 | data->tid_entry.tid = perf_event_tid(event, current); | |
3907 | } | |
3908 | ||
3909 | if (sample_type & PERF_SAMPLE_TIME) | |
3910 | data->time = perf_clock(); | |
3911 | ||
3912 | if (sample_type & PERF_SAMPLE_ID) | |
3913 | data->id = primary_event_id(event); | |
3914 | ||
3915 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3916 | data->stream_id = event->id; | |
3917 | ||
3918 | if (sample_type & PERF_SAMPLE_CPU) { | |
3919 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
3920 | data->cpu_entry.reserved = 0; | |
3921 | } | |
3922 | } | |
3923 | ||
76369139 FW |
3924 | void perf_event_header__init_id(struct perf_event_header *header, |
3925 | struct perf_sample_data *data, | |
3926 | struct perf_event *event) | |
c980d109 ACM |
3927 | { |
3928 | if (event->attr.sample_id_all) | |
3929 | __perf_event_header__init_id(header, data, event); | |
3930 | } | |
3931 | ||
3932 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
3933 | struct perf_sample_data *data) | |
3934 | { | |
3935 | u64 sample_type = data->type; | |
3936 | ||
3937 | if (sample_type & PERF_SAMPLE_TID) | |
3938 | perf_output_put(handle, data->tid_entry); | |
3939 | ||
3940 | if (sample_type & PERF_SAMPLE_TIME) | |
3941 | perf_output_put(handle, data->time); | |
3942 | ||
3943 | if (sample_type & PERF_SAMPLE_ID) | |
3944 | perf_output_put(handle, data->id); | |
3945 | ||
3946 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3947 | perf_output_put(handle, data->stream_id); | |
3948 | ||
3949 | if (sample_type & PERF_SAMPLE_CPU) | |
3950 | perf_output_put(handle, data->cpu_entry); | |
3951 | } | |
3952 | ||
76369139 FW |
3953 | void perf_event__output_id_sample(struct perf_event *event, |
3954 | struct perf_output_handle *handle, | |
3955 | struct perf_sample_data *sample) | |
c980d109 ACM |
3956 | { |
3957 | if (event->attr.sample_id_all) | |
3958 | __perf_event__output_id_sample(handle, sample); | |
3959 | } | |
3960 | ||
3dab77fb | 3961 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
3962 | struct perf_event *event, |
3963 | u64 enabled, u64 running) | |
3dab77fb | 3964 | { |
cdd6c482 | 3965 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
3966 | u64 values[4]; |
3967 | int n = 0; | |
3968 | ||
b5e58793 | 3969 | values[n++] = perf_event_count(event); |
3dab77fb | 3970 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 3971 | values[n++] = enabled + |
cdd6c482 | 3972 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
3973 | } |
3974 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 3975 | values[n++] = running + |
cdd6c482 | 3976 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
3977 | } |
3978 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 3979 | values[n++] = primary_event_id(event); |
3dab77fb | 3980 | |
76369139 | 3981 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
3982 | } |
3983 | ||
3984 | /* | |
cdd6c482 | 3985 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
3986 | */ |
3987 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
3988 | struct perf_event *event, |
3989 | u64 enabled, u64 running) | |
3dab77fb | 3990 | { |
cdd6c482 IM |
3991 | struct perf_event *leader = event->group_leader, *sub; |
3992 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
3993 | u64 values[5]; |
3994 | int n = 0; | |
3995 | ||
3996 | values[n++] = 1 + leader->nr_siblings; | |
3997 | ||
3998 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 3999 | values[n++] = enabled; |
3dab77fb PZ |
4000 | |
4001 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4002 | values[n++] = running; |
3dab77fb | 4003 | |
cdd6c482 | 4004 | if (leader != event) |
3dab77fb PZ |
4005 | leader->pmu->read(leader); |
4006 | ||
b5e58793 | 4007 | values[n++] = perf_event_count(leader); |
3dab77fb | 4008 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4009 | values[n++] = primary_event_id(leader); |
3dab77fb | 4010 | |
76369139 | 4011 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4012 | |
65abc865 | 4013 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4014 | n = 0; |
4015 | ||
cdd6c482 | 4016 | if (sub != event) |
3dab77fb PZ |
4017 | sub->pmu->read(sub); |
4018 | ||
b5e58793 | 4019 | values[n++] = perf_event_count(sub); |
3dab77fb | 4020 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4021 | values[n++] = primary_event_id(sub); |
3dab77fb | 4022 | |
76369139 | 4023 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4024 | } |
4025 | } | |
4026 | ||
eed01528 SE |
4027 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4028 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4029 | ||
3dab77fb | 4030 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4031 | struct perf_event *event) |
3dab77fb | 4032 | { |
e3f3541c | 4033 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4034 | u64 read_format = event->attr.read_format; |
4035 | ||
4036 | /* | |
4037 | * compute total_time_enabled, total_time_running | |
4038 | * based on snapshot values taken when the event | |
4039 | * was last scheduled in. | |
4040 | * | |
4041 | * we cannot simply called update_context_time() | |
4042 | * because of locking issue as we are called in | |
4043 | * NMI context | |
4044 | */ | |
c4794295 | 4045 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4046 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4047 | |
cdd6c482 | 4048 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4049 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4050 | else |
eed01528 | 4051 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4052 | } |
4053 | ||
5622f295 MM |
4054 | void perf_output_sample(struct perf_output_handle *handle, |
4055 | struct perf_event_header *header, | |
4056 | struct perf_sample_data *data, | |
cdd6c482 | 4057 | struct perf_event *event) |
5622f295 MM |
4058 | { |
4059 | u64 sample_type = data->type; | |
4060 | ||
4061 | perf_output_put(handle, *header); | |
4062 | ||
4063 | if (sample_type & PERF_SAMPLE_IP) | |
4064 | perf_output_put(handle, data->ip); | |
4065 | ||
4066 | if (sample_type & PERF_SAMPLE_TID) | |
4067 | perf_output_put(handle, data->tid_entry); | |
4068 | ||
4069 | if (sample_type & PERF_SAMPLE_TIME) | |
4070 | perf_output_put(handle, data->time); | |
4071 | ||
4072 | if (sample_type & PERF_SAMPLE_ADDR) | |
4073 | perf_output_put(handle, data->addr); | |
4074 | ||
4075 | if (sample_type & PERF_SAMPLE_ID) | |
4076 | perf_output_put(handle, data->id); | |
4077 | ||
4078 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4079 | perf_output_put(handle, data->stream_id); | |
4080 | ||
4081 | if (sample_type & PERF_SAMPLE_CPU) | |
4082 | perf_output_put(handle, data->cpu_entry); | |
4083 | ||
4084 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4085 | perf_output_put(handle, data->period); | |
4086 | ||
4087 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4088 | perf_output_read(handle, event); |
5622f295 MM |
4089 | |
4090 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4091 | if (data->callchain) { | |
4092 | int size = 1; | |
4093 | ||
4094 | if (data->callchain) | |
4095 | size += data->callchain->nr; | |
4096 | ||
4097 | size *= sizeof(u64); | |
4098 | ||
76369139 | 4099 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4100 | } else { |
4101 | u64 nr = 0; | |
4102 | perf_output_put(handle, nr); | |
4103 | } | |
4104 | } | |
4105 | ||
4106 | if (sample_type & PERF_SAMPLE_RAW) { | |
4107 | if (data->raw) { | |
4108 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4109 | __output_copy(handle, data->raw->data, |
4110 | data->raw->size); | |
5622f295 MM |
4111 | } else { |
4112 | struct { | |
4113 | u32 size; | |
4114 | u32 data; | |
4115 | } raw = { | |
4116 | .size = sizeof(u32), | |
4117 | .data = 0, | |
4118 | }; | |
4119 | perf_output_put(handle, raw); | |
4120 | } | |
4121 | } | |
a7ac67ea PZ |
4122 | |
4123 | if (!event->attr.watermark) { | |
4124 | int wakeup_events = event->attr.wakeup_events; | |
4125 | ||
4126 | if (wakeup_events) { | |
4127 | struct ring_buffer *rb = handle->rb; | |
4128 | int events = local_inc_return(&rb->events); | |
4129 | ||
4130 | if (events >= wakeup_events) { | |
4131 | local_sub(wakeup_events, &rb->events); | |
4132 | local_inc(&rb->wakeup); | |
4133 | } | |
4134 | } | |
4135 | } | |
bce38cd5 SE |
4136 | |
4137 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4138 | if (data->br_stack) { | |
4139 | size_t size; | |
4140 | ||
4141 | size = data->br_stack->nr | |
4142 | * sizeof(struct perf_branch_entry); | |
4143 | ||
4144 | perf_output_put(handle, data->br_stack->nr); | |
4145 | perf_output_copy(handle, data->br_stack->entries, size); | |
4146 | } else { | |
4147 | /* | |
4148 | * we always store at least the value of nr | |
4149 | */ | |
4150 | u64 nr = 0; | |
4151 | perf_output_put(handle, nr); | |
4152 | } | |
4153 | } | |
4018994f JO |
4154 | |
4155 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4156 | u64 abi = data->regs_user.abi; | |
4157 | ||
4158 | /* | |
4159 | * If there are no regs to dump, notice it through | |
4160 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4161 | */ | |
4162 | perf_output_put(handle, abi); | |
4163 | ||
4164 | if (abi) { | |
4165 | u64 mask = event->attr.sample_regs_user; | |
4166 | perf_output_sample_regs(handle, | |
4167 | data->regs_user.regs, | |
4168 | mask); | |
4169 | } | |
4170 | } | |
c5ebcedb JO |
4171 | |
4172 | if (sample_type & PERF_SAMPLE_STACK_USER) | |
4173 | perf_output_sample_ustack(handle, | |
4174 | data->stack_user_size, | |
4175 | data->regs_user.regs); | |
5622f295 MM |
4176 | } |
4177 | ||
4178 | void perf_prepare_sample(struct perf_event_header *header, | |
4179 | struct perf_sample_data *data, | |
cdd6c482 | 4180 | struct perf_event *event, |
5622f295 | 4181 | struct pt_regs *regs) |
7b732a75 | 4182 | { |
cdd6c482 | 4183 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4184 | |
cdd6c482 | 4185 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4186 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4187 | |
4188 | header->misc = 0; | |
4189 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4190 | |
c980d109 | 4191 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4192 | |
c320c7b7 | 4193 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4194 | data->ip = perf_instruction_pointer(regs); |
4195 | ||
b23f3325 | 4196 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4197 | int size = 1; |
394ee076 | 4198 | |
e6dab5ff | 4199 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4200 | |
4201 | if (data->callchain) | |
4202 | size += data->callchain->nr; | |
4203 | ||
4204 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4205 | } |
4206 | ||
3a43ce68 | 4207 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4208 | int size = sizeof(u32); |
4209 | ||
4210 | if (data->raw) | |
4211 | size += data->raw->size; | |
4212 | else | |
4213 | size += sizeof(u32); | |
4214 | ||
4215 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4216 | header->size += size; |
7f453c24 | 4217 | } |
bce38cd5 SE |
4218 | |
4219 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4220 | int size = sizeof(u64); /* nr */ | |
4221 | if (data->br_stack) { | |
4222 | size += data->br_stack->nr | |
4223 | * sizeof(struct perf_branch_entry); | |
4224 | } | |
4225 | header->size += size; | |
4226 | } | |
4018994f JO |
4227 | |
4228 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4229 | /* regs dump ABI info */ | |
4230 | int size = sizeof(u64); | |
4231 | ||
4232 | perf_sample_regs_user(&data->regs_user, regs); | |
4233 | ||
4234 | if (data->regs_user.regs) { | |
4235 | u64 mask = event->attr.sample_regs_user; | |
4236 | size += hweight64(mask) * sizeof(u64); | |
4237 | } | |
4238 | ||
4239 | header->size += size; | |
4240 | } | |
c5ebcedb JO |
4241 | |
4242 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4243 | /* | |
4244 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4245 | * processed as the last one or have additional check added | |
4246 | * in case new sample type is added, because we could eat | |
4247 | * up the rest of the sample size. | |
4248 | */ | |
4249 | struct perf_regs_user *uregs = &data->regs_user; | |
4250 | u16 stack_size = event->attr.sample_stack_user; | |
4251 | u16 size = sizeof(u64); | |
4252 | ||
4253 | if (!uregs->abi) | |
4254 | perf_sample_regs_user(uregs, regs); | |
4255 | ||
4256 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4257 | uregs->regs); | |
4258 | ||
4259 | /* | |
4260 | * If there is something to dump, add space for the dump | |
4261 | * itself and for the field that tells the dynamic size, | |
4262 | * which is how many have been actually dumped. | |
4263 | */ | |
4264 | if (stack_size) | |
4265 | size += sizeof(u64) + stack_size; | |
4266 | ||
4267 | data->stack_user_size = stack_size; | |
4268 | header->size += size; | |
4269 | } | |
5622f295 | 4270 | } |
7f453c24 | 4271 | |
a8b0ca17 | 4272 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4273 | struct perf_sample_data *data, |
4274 | struct pt_regs *regs) | |
4275 | { | |
4276 | struct perf_output_handle handle; | |
4277 | struct perf_event_header header; | |
689802b2 | 4278 | |
927c7a9e FW |
4279 | /* protect the callchain buffers */ |
4280 | rcu_read_lock(); | |
4281 | ||
cdd6c482 | 4282 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4283 | |
a7ac67ea | 4284 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4285 | goto exit; |
0322cd6e | 4286 | |
cdd6c482 | 4287 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4288 | |
8a057d84 | 4289 | perf_output_end(&handle); |
927c7a9e FW |
4290 | |
4291 | exit: | |
4292 | rcu_read_unlock(); | |
0322cd6e PZ |
4293 | } |
4294 | ||
38b200d6 | 4295 | /* |
cdd6c482 | 4296 | * read event_id |
38b200d6 PZ |
4297 | */ |
4298 | ||
4299 | struct perf_read_event { | |
4300 | struct perf_event_header header; | |
4301 | ||
4302 | u32 pid; | |
4303 | u32 tid; | |
38b200d6 PZ |
4304 | }; |
4305 | ||
4306 | static void | |
cdd6c482 | 4307 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4308 | struct task_struct *task) |
4309 | { | |
4310 | struct perf_output_handle handle; | |
c980d109 | 4311 | struct perf_sample_data sample; |
dfc65094 | 4312 | struct perf_read_event read_event = { |
38b200d6 | 4313 | .header = { |
cdd6c482 | 4314 | .type = PERF_RECORD_READ, |
38b200d6 | 4315 | .misc = 0, |
c320c7b7 | 4316 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4317 | }, |
cdd6c482 IM |
4318 | .pid = perf_event_pid(event, task), |
4319 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4320 | }; |
3dab77fb | 4321 | int ret; |
38b200d6 | 4322 | |
c980d109 | 4323 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4324 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4325 | if (ret) |
4326 | return; | |
4327 | ||
dfc65094 | 4328 | perf_output_put(&handle, read_event); |
cdd6c482 | 4329 | perf_output_read(&handle, event); |
c980d109 | 4330 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4331 | |
38b200d6 PZ |
4332 | perf_output_end(&handle); |
4333 | } | |
4334 | ||
60313ebe | 4335 | /* |
9f498cc5 PZ |
4336 | * task tracking -- fork/exit |
4337 | * | |
3af9e859 | 4338 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
60313ebe PZ |
4339 | */ |
4340 | ||
9f498cc5 | 4341 | struct perf_task_event { |
3a80b4a3 | 4342 | struct task_struct *task; |
cdd6c482 | 4343 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4344 | |
4345 | struct { | |
4346 | struct perf_event_header header; | |
4347 | ||
4348 | u32 pid; | |
4349 | u32 ppid; | |
9f498cc5 PZ |
4350 | u32 tid; |
4351 | u32 ptid; | |
393b2ad8 | 4352 | u64 time; |
cdd6c482 | 4353 | } event_id; |
60313ebe PZ |
4354 | }; |
4355 | ||
cdd6c482 | 4356 | static void perf_event_task_output(struct perf_event *event, |
9f498cc5 | 4357 | struct perf_task_event *task_event) |
60313ebe PZ |
4358 | { |
4359 | struct perf_output_handle handle; | |
c980d109 | 4360 | struct perf_sample_data sample; |
9f498cc5 | 4361 | struct task_struct *task = task_event->task; |
c980d109 | 4362 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4363 | |
c980d109 | 4364 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4365 | |
c980d109 | 4366 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4367 | task_event->event_id.header.size); |
ef60777c | 4368 | if (ret) |
c980d109 | 4369 | goto out; |
60313ebe | 4370 | |
cdd6c482 IM |
4371 | task_event->event_id.pid = perf_event_pid(event, task); |
4372 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4373 | |
cdd6c482 IM |
4374 | task_event->event_id.tid = perf_event_tid(event, task); |
4375 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4376 | |
cdd6c482 | 4377 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4378 | |
c980d109 ACM |
4379 | perf_event__output_id_sample(event, &handle, &sample); |
4380 | ||
60313ebe | 4381 | perf_output_end(&handle); |
c980d109 ACM |
4382 | out: |
4383 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4384 | } |
4385 | ||
cdd6c482 | 4386 | static int perf_event_task_match(struct perf_event *event) |
60313ebe | 4387 | { |
6f93d0a7 | 4388 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4389 | return 0; |
4390 | ||
5632ab12 | 4391 | if (!event_filter_match(event)) |
5d27c23d PZ |
4392 | return 0; |
4393 | ||
3af9e859 EM |
4394 | if (event->attr.comm || event->attr.mmap || |
4395 | event->attr.mmap_data || event->attr.task) | |
60313ebe PZ |
4396 | return 1; |
4397 | ||
4398 | return 0; | |
4399 | } | |
4400 | ||
cdd6c482 | 4401 | static void perf_event_task_ctx(struct perf_event_context *ctx, |
9f498cc5 | 4402 | struct perf_task_event *task_event) |
60313ebe | 4403 | { |
cdd6c482 | 4404 | struct perf_event *event; |
60313ebe | 4405 | |
cdd6c482 IM |
4406 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4407 | if (perf_event_task_match(event)) | |
4408 | perf_event_task_output(event, task_event); | |
60313ebe | 4409 | } |
60313ebe PZ |
4410 | } |
4411 | ||
cdd6c482 | 4412 | static void perf_event_task_event(struct perf_task_event *task_event) |
60313ebe PZ |
4413 | { |
4414 | struct perf_cpu_context *cpuctx; | |
8dc85d54 | 4415 | struct perf_event_context *ctx; |
108b02cf | 4416 | struct pmu *pmu; |
8dc85d54 | 4417 | int ctxn; |
60313ebe | 4418 | |
d6ff86cf | 4419 | rcu_read_lock(); |
108b02cf | 4420 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4421 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4422 | if (cpuctx->active_pmu != pmu) |
4423 | goto next; | |
108b02cf | 4424 | perf_event_task_ctx(&cpuctx->ctx, task_event); |
8dc85d54 PZ |
4425 | |
4426 | ctx = task_event->task_ctx; | |
4427 | if (!ctx) { | |
4428 | ctxn = pmu->task_ctx_nr; | |
4429 | if (ctxn < 0) | |
41945f6c | 4430 | goto next; |
8dc85d54 PZ |
4431 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); |
4432 | } | |
4433 | if (ctx) | |
4434 | perf_event_task_ctx(ctx, task_event); | |
41945f6c PZ |
4435 | next: |
4436 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4437 | } |
60313ebe PZ |
4438 | rcu_read_unlock(); |
4439 | } | |
4440 | ||
cdd6c482 IM |
4441 | static void perf_event_task(struct task_struct *task, |
4442 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4443 | int new) |
60313ebe | 4444 | { |
9f498cc5 | 4445 | struct perf_task_event task_event; |
60313ebe | 4446 | |
cdd6c482 IM |
4447 | if (!atomic_read(&nr_comm_events) && |
4448 | !atomic_read(&nr_mmap_events) && | |
4449 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4450 | return; |
4451 | ||
9f498cc5 | 4452 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4453 | .task = task, |
4454 | .task_ctx = task_ctx, | |
cdd6c482 | 4455 | .event_id = { |
60313ebe | 4456 | .header = { |
cdd6c482 | 4457 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4458 | .misc = 0, |
cdd6c482 | 4459 | .size = sizeof(task_event.event_id), |
60313ebe | 4460 | }, |
573402db PZ |
4461 | /* .pid */ |
4462 | /* .ppid */ | |
9f498cc5 PZ |
4463 | /* .tid */ |
4464 | /* .ptid */ | |
6f93d0a7 | 4465 | .time = perf_clock(), |
60313ebe PZ |
4466 | }, |
4467 | }; | |
4468 | ||
cdd6c482 | 4469 | perf_event_task_event(&task_event); |
9f498cc5 PZ |
4470 | } |
4471 | ||
cdd6c482 | 4472 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4473 | { |
cdd6c482 | 4474 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4475 | } |
4476 | ||
8d1b2d93 PZ |
4477 | /* |
4478 | * comm tracking | |
4479 | */ | |
4480 | ||
4481 | struct perf_comm_event { | |
22a4f650 IM |
4482 | struct task_struct *task; |
4483 | char *comm; | |
8d1b2d93 PZ |
4484 | int comm_size; |
4485 | ||
4486 | struct { | |
4487 | struct perf_event_header header; | |
4488 | ||
4489 | u32 pid; | |
4490 | u32 tid; | |
cdd6c482 | 4491 | } event_id; |
8d1b2d93 PZ |
4492 | }; |
4493 | ||
cdd6c482 | 4494 | static void perf_event_comm_output(struct perf_event *event, |
8d1b2d93 PZ |
4495 | struct perf_comm_event *comm_event) |
4496 | { | |
4497 | struct perf_output_handle handle; | |
c980d109 | 4498 | struct perf_sample_data sample; |
cdd6c482 | 4499 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4500 | int ret; |
4501 | ||
4502 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); | |
4503 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4504 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4505 | |
4506 | if (ret) | |
c980d109 | 4507 | goto out; |
8d1b2d93 | 4508 | |
cdd6c482 IM |
4509 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4510 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4511 | |
cdd6c482 | 4512 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4513 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4514 | comm_event->comm_size); |
c980d109 ACM |
4515 | |
4516 | perf_event__output_id_sample(event, &handle, &sample); | |
4517 | ||
8d1b2d93 | 4518 | perf_output_end(&handle); |
c980d109 ACM |
4519 | out: |
4520 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4521 | } |
4522 | ||
cdd6c482 | 4523 | static int perf_event_comm_match(struct perf_event *event) |
8d1b2d93 | 4524 | { |
6f93d0a7 | 4525 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4526 | return 0; |
4527 | ||
5632ab12 | 4528 | if (!event_filter_match(event)) |
5d27c23d PZ |
4529 | return 0; |
4530 | ||
cdd6c482 | 4531 | if (event->attr.comm) |
8d1b2d93 PZ |
4532 | return 1; |
4533 | ||
4534 | return 0; | |
4535 | } | |
4536 | ||
cdd6c482 | 4537 | static void perf_event_comm_ctx(struct perf_event_context *ctx, |
8d1b2d93 PZ |
4538 | struct perf_comm_event *comm_event) |
4539 | { | |
cdd6c482 | 4540 | struct perf_event *event; |
8d1b2d93 | 4541 | |
cdd6c482 IM |
4542 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4543 | if (perf_event_comm_match(event)) | |
4544 | perf_event_comm_output(event, comm_event); | |
8d1b2d93 | 4545 | } |
8d1b2d93 PZ |
4546 | } |
4547 | ||
cdd6c482 | 4548 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 PZ |
4549 | { |
4550 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4551 | struct perf_event_context *ctx; |
413ee3b4 | 4552 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4553 | unsigned int size; |
108b02cf | 4554 | struct pmu *pmu; |
8dc85d54 | 4555 | int ctxn; |
8d1b2d93 | 4556 | |
413ee3b4 | 4557 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4558 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4559 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4560 | |
4561 | comm_event->comm = comm; | |
4562 | comm_event->comm_size = size; | |
4563 | ||
cdd6c482 | 4564 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
f6595f3a | 4565 | rcu_read_lock(); |
108b02cf | 4566 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4567 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4568 | if (cpuctx->active_pmu != pmu) |
4569 | goto next; | |
108b02cf | 4570 | perf_event_comm_ctx(&cpuctx->ctx, comm_event); |
8dc85d54 PZ |
4571 | |
4572 | ctxn = pmu->task_ctx_nr; | |
4573 | if (ctxn < 0) | |
41945f6c | 4574 | goto next; |
8dc85d54 PZ |
4575 | |
4576 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4577 | if (ctx) | |
4578 | perf_event_comm_ctx(ctx, comm_event); | |
41945f6c PZ |
4579 | next: |
4580 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4581 | } |
665c2142 | 4582 | rcu_read_unlock(); |
8d1b2d93 PZ |
4583 | } |
4584 | ||
cdd6c482 | 4585 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4586 | { |
9ee318a7 | 4587 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4588 | struct perf_event_context *ctx; |
4589 | int ctxn; | |
9ee318a7 | 4590 | |
8dc85d54 PZ |
4591 | for_each_task_context_nr(ctxn) { |
4592 | ctx = task->perf_event_ctxp[ctxn]; | |
4593 | if (!ctx) | |
4594 | continue; | |
9ee318a7 | 4595 | |
8dc85d54 PZ |
4596 | perf_event_enable_on_exec(ctx); |
4597 | } | |
9ee318a7 | 4598 | |
cdd6c482 | 4599 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4600 | return; |
a63eaf34 | 4601 | |
9ee318a7 | 4602 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4603 | .task = task, |
573402db PZ |
4604 | /* .comm */ |
4605 | /* .comm_size */ | |
cdd6c482 | 4606 | .event_id = { |
573402db | 4607 | .header = { |
cdd6c482 | 4608 | .type = PERF_RECORD_COMM, |
573402db PZ |
4609 | .misc = 0, |
4610 | /* .size */ | |
4611 | }, | |
4612 | /* .pid */ | |
4613 | /* .tid */ | |
8d1b2d93 PZ |
4614 | }, |
4615 | }; | |
4616 | ||
cdd6c482 | 4617 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
4618 | } |
4619 | ||
0a4a9391 PZ |
4620 | /* |
4621 | * mmap tracking | |
4622 | */ | |
4623 | ||
4624 | struct perf_mmap_event { | |
089dd79d PZ |
4625 | struct vm_area_struct *vma; |
4626 | ||
4627 | const char *file_name; | |
4628 | int file_size; | |
0a4a9391 PZ |
4629 | |
4630 | struct { | |
4631 | struct perf_event_header header; | |
4632 | ||
4633 | u32 pid; | |
4634 | u32 tid; | |
4635 | u64 start; | |
4636 | u64 len; | |
4637 | u64 pgoff; | |
cdd6c482 | 4638 | } event_id; |
0a4a9391 PZ |
4639 | }; |
4640 | ||
cdd6c482 | 4641 | static void perf_event_mmap_output(struct perf_event *event, |
0a4a9391 PZ |
4642 | struct perf_mmap_event *mmap_event) |
4643 | { | |
4644 | struct perf_output_handle handle; | |
c980d109 | 4645 | struct perf_sample_data sample; |
cdd6c482 | 4646 | int size = mmap_event->event_id.header.size; |
c980d109 | 4647 | int ret; |
0a4a9391 | 4648 | |
c980d109 ACM |
4649 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
4650 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4651 | mmap_event->event_id.header.size); |
0a4a9391 | 4652 | if (ret) |
c980d109 | 4653 | goto out; |
0a4a9391 | 4654 | |
cdd6c482 IM |
4655 | mmap_event->event_id.pid = perf_event_pid(event, current); |
4656 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 4657 | |
cdd6c482 | 4658 | perf_output_put(&handle, mmap_event->event_id); |
76369139 | 4659 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 4660 | mmap_event->file_size); |
c980d109 ACM |
4661 | |
4662 | perf_event__output_id_sample(event, &handle, &sample); | |
4663 | ||
78d613eb | 4664 | perf_output_end(&handle); |
c980d109 ACM |
4665 | out: |
4666 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
4667 | } |
4668 | ||
cdd6c482 | 4669 | static int perf_event_mmap_match(struct perf_event *event, |
3af9e859 EM |
4670 | struct perf_mmap_event *mmap_event, |
4671 | int executable) | |
0a4a9391 | 4672 | { |
6f93d0a7 | 4673 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4674 | return 0; |
4675 | ||
5632ab12 | 4676 | if (!event_filter_match(event)) |
5d27c23d PZ |
4677 | return 0; |
4678 | ||
3af9e859 EM |
4679 | if ((!executable && event->attr.mmap_data) || |
4680 | (executable && event->attr.mmap)) | |
0a4a9391 PZ |
4681 | return 1; |
4682 | ||
4683 | return 0; | |
4684 | } | |
4685 | ||
cdd6c482 | 4686 | static void perf_event_mmap_ctx(struct perf_event_context *ctx, |
3af9e859 EM |
4687 | struct perf_mmap_event *mmap_event, |
4688 | int executable) | |
0a4a9391 | 4689 | { |
cdd6c482 | 4690 | struct perf_event *event; |
0a4a9391 | 4691 | |
cdd6c482 | 4692 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
3af9e859 | 4693 | if (perf_event_mmap_match(event, mmap_event, executable)) |
cdd6c482 | 4694 | perf_event_mmap_output(event, mmap_event); |
0a4a9391 | 4695 | } |
0a4a9391 PZ |
4696 | } |
4697 | ||
cdd6c482 | 4698 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 PZ |
4699 | { |
4700 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4701 | struct perf_event_context *ctx; |
089dd79d PZ |
4702 | struct vm_area_struct *vma = mmap_event->vma; |
4703 | struct file *file = vma->vm_file; | |
0a4a9391 PZ |
4704 | unsigned int size; |
4705 | char tmp[16]; | |
4706 | char *buf = NULL; | |
089dd79d | 4707 | const char *name; |
108b02cf | 4708 | struct pmu *pmu; |
8dc85d54 | 4709 | int ctxn; |
0a4a9391 | 4710 | |
413ee3b4 AB |
4711 | memset(tmp, 0, sizeof(tmp)); |
4712 | ||
0a4a9391 | 4713 | if (file) { |
413ee3b4 | 4714 | /* |
76369139 | 4715 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
4716 | * need to add enough zero bytes after the string to handle |
4717 | * the 64bit alignment we do later. | |
4718 | */ | |
4719 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
4720 | if (!buf) { |
4721 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
4722 | goto got_name; | |
4723 | } | |
d3d21c41 | 4724 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
4725 | if (IS_ERR(name)) { |
4726 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
4727 | goto got_name; | |
4728 | } | |
4729 | } else { | |
413ee3b4 AB |
4730 | if (arch_vma_name(mmap_event->vma)) { |
4731 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
4732 | sizeof(tmp)); | |
089dd79d | 4733 | goto got_name; |
413ee3b4 | 4734 | } |
089dd79d PZ |
4735 | |
4736 | if (!vma->vm_mm) { | |
4737 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
4738 | goto got_name; | |
3af9e859 EM |
4739 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
4740 | vma->vm_end >= vma->vm_mm->brk) { | |
4741 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
4742 | goto got_name; | |
4743 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
4744 | vma->vm_end >= vma->vm_mm->start_stack) { | |
4745 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
4746 | goto got_name; | |
089dd79d PZ |
4747 | } |
4748 | ||
0a4a9391 PZ |
4749 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
4750 | goto got_name; | |
4751 | } | |
4752 | ||
4753 | got_name: | |
888fcee0 | 4754 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
4755 | |
4756 | mmap_event->file_name = name; | |
4757 | mmap_event->file_size = size; | |
4758 | ||
cdd6c482 | 4759 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 4760 | |
f6d9dd23 | 4761 | rcu_read_lock(); |
108b02cf | 4762 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4763 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
51676957 PZ |
4764 | if (cpuctx->active_pmu != pmu) |
4765 | goto next; | |
108b02cf PZ |
4766 | perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, |
4767 | vma->vm_flags & VM_EXEC); | |
8dc85d54 PZ |
4768 | |
4769 | ctxn = pmu->task_ctx_nr; | |
4770 | if (ctxn < 0) | |
41945f6c | 4771 | goto next; |
8dc85d54 PZ |
4772 | |
4773 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4774 | if (ctx) { | |
4775 | perf_event_mmap_ctx(ctx, mmap_event, | |
4776 | vma->vm_flags & VM_EXEC); | |
4777 | } | |
41945f6c PZ |
4778 | next: |
4779 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4780 | } |
665c2142 PZ |
4781 | rcu_read_unlock(); |
4782 | ||
0a4a9391 PZ |
4783 | kfree(buf); |
4784 | } | |
4785 | ||
3af9e859 | 4786 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 4787 | { |
9ee318a7 PZ |
4788 | struct perf_mmap_event mmap_event; |
4789 | ||
cdd6c482 | 4790 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
4791 | return; |
4792 | ||
4793 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 4794 | .vma = vma, |
573402db PZ |
4795 | /* .file_name */ |
4796 | /* .file_size */ | |
cdd6c482 | 4797 | .event_id = { |
573402db | 4798 | .header = { |
cdd6c482 | 4799 | .type = PERF_RECORD_MMAP, |
39447b38 | 4800 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
4801 | /* .size */ |
4802 | }, | |
4803 | /* .pid */ | |
4804 | /* .tid */ | |
089dd79d PZ |
4805 | .start = vma->vm_start, |
4806 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 4807 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 PZ |
4808 | }, |
4809 | }; | |
4810 | ||
cdd6c482 | 4811 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
4812 | } |
4813 | ||
a78ac325 PZ |
4814 | /* |
4815 | * IRQ throttle logging | |
4816 | */ | |
4817 | ||
cdd6c482 | 4818 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
4819 | { |
4820 | struct perf_output_handle handle; | |
c980d109 | 4821 | struct perf_sample_data sample; |
a78ac325 PZ |
4822 | int ret; |
4823 | ||
4824 | struct { | |
4825 | struct perf_event_header header; | |
4826 | u64 time; | |
cca3f454 | 4827 | u64 id; |
7f453c24 | 4828 | u64 stream_id; |
a78ac325 PZ |
4829 | } throttle_event = { |
4830 | .header = { | |
cdd6c482 | 4831 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
4832 | .misc = 0, |
4833 | .size = sizeof(throttle_event), | |
4834 | }, | |
def0a9b2 | 4835 | .time = perf_clock(), |
cdd6c482 IM |
4836 | .id = primary_event_id(event), |
4837 | .stream_id = event->id, | |
a78ac325 PZ |
4838 | }; |
4839 | ||
966ee4d6 | 4840 | if (enable) |
cdd6c482 | 4841 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 4842 | |
c980d109 ACM |
4843 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
4844 | ||
4845 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4846 | throttle_event.header.size); |
a78ac325 PZ |
4847 | if (ret) |
4848 | return; | |
4849 | ||
4850 | perf_output_put(&handle, throttle_event); | |
c980d109 | 4851 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
4852 | perf_output_end(&handle); |
4853 | } | |
4854 | ||
f6c7d5fe | 4855 | /* |
cdd6c482 | 4856 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
4857 | */ |
4858 | ||
a8b0ca17 | 4859 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4860 | int throttle, struct perf_sample_data *data, |
4861 | struct pt_regs *regs) | |
f6c7d5fe | 4862 | { |
cdd6c482 IM |
4863 | int events = atomic_read(&event->event_limit); |
4864 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 4865 | u64 seq; |
79f14641 PZ |
4866 | int ret = 0; |
4867 | ||
96398826 PZ |
4868 | /* |
4869 | * Non-sampling counters might still use the PMI to fold short | |
4870 | * hardware counters, ignore those. | |
4871 | */ | |
4872 | if (unlikely(!is_sampling_event(event))) | |
4873 | return 0; | |
4874 | ||
e050e3f0 SE |
4875 | seq = __this_cpu_read(perf_throttled_seq); |
4876 | if (seq != hwc->interrupts_seq) { | |
4877 | hwc->interrupts_seq = seq; | |
4878 | hwc->interrupts = 1; | |
4879 | } else { | |
4880 | hwc->interrupts++; | |
4881 | if (unlikely(throttle | |
4882 | && hwc->interrupts >= max_samples_per_tick)) { | |
4883 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
4884 | hwc->interrupts = MAX_INTERRUPTS; |
4885 | perf_log_throttle(event, 0); | |
a78ac325 PZ |
4886 | ret = 1; |
4887 | } | |
e050e3f0 | 4888 | } |
60db5e09 | 4889 | |
cdd6c482 | 4890 | if (event->attr.freq) { |
def0a9b2 | 4891 | u64 now = perf_clock(); |
abd50713 | 4892 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 4893 | |
abd50713 | 4894 | hwc->freq_time_stamp = now; |
bd2b5b12 | 4895 | |
abd50713 | 4896 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 4897 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
4898 | } |
4899 | ||
2023b359 PZ |
4900 | /* |
4901 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 4902 | * events |
2023b359 PZ |
4903 | */ |
4904 | ||
cdd6c482 IM |
4905 | event->pending_kill = POLL_IN; |
4906 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 4907 | ret = 1; |
cdd6c482 | 4908 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
4909 | event->pending_disable = 1; |
4910 | irq_work_queue(&event->pending); | |
79f14641 PZ |
4911 | } |
4912 | ||
453f19ee | 4913 | if (event->overflow_handler) |
a8b0ca17 | 4914 | event->overflow_handler(event, data, regs); |
453f19ee | 4915 | else |
a8b0ca17 | 4916 | perf_event_output(event, data, regs); |
453f19ee | 4917 | |
f506b3dc | 4918 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
4919 | event->pending_wakeup = 1; |
4920 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
4921 | } |
4922 | ||
79f14641 | 4923 | return ret; |
f6c7d5fe PZ |
4924 | } |
4925 | ||
a8b0ca17 | 4926 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4927 | struct perf_sample_data *data, |
4928 | struct pt_regs *regs) | |
850bc73f | 4929 | { |
a8b0ca17 | 4930 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
4931 | } |
4932 | ||
15dbf27c | 4933 | /* |
cdd6c482 | 4934 | * Generic software event infrastructure |
15dbf27c PZ |
4935 | */ |
4936 | ||
b28ab83c PZ |
4937 | struct swevent_htable { |
4938 | struct swevent_hlist *swevent_hlist; | |
4939 | struct mutex hlist_mutex; | |
4940 | int hlist_refcount; | |
4941 | ||
4942 | /* Recursion avoidance in each contexts */ | |
4943 | int recursion[PERF_NR_CONTEXTS]; | |
4944 | }; | |
4945 | ||
4946 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
4947 | ||
7b4b6658 | 4948 | /* |
cdd6c482 IM |
4949 | * We directly increment event->count and keep a second value in |
4950 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
4951 | * is kept in the range [-sample_period, 0] so that we can use the |
4952 | * sign as trigger. | |
4953 | */ | |
4954 | ||
cdd6c482 | 4955 | static u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 4956 | { |
cdd6c482 | 4957 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
4958 | u64 period = hwc->last_period; |
4959 | u64 nr, offset; | |
4960 | s64 old, val; | |
4961 | ||
4962 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
4963 | |
4964 | again: | |
e7850595 | 4965 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
4966 | if (val < 0) |
4967 | return 0; | |
15dbf27c | 4968 | |
7b4b6658 PZ |
4969 | nr = div64_u64(period + val, period); |
4970 | offset = nr * period; | |
4971 | val -= offset; | |
e7850595 | 4972 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 4973 | goto again; |
15dbf27c | 4974 | |
7b4b6658 | 4975 | return nr; |
15dbf27c PZ |
4976 | } |
4977 | ||
0cff784a | 4978 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 4979 | struct perf_sample_data *data, |
5622f295 | 4980 | struct pt_regs *regs) |
15dbf27c | 4981 | { |
cdd6c482 | 4982 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 4983 | int throttle = 0; |
15dbf27c | 4984 | |
0cff784a PZ |
4985 | if (!overflow) |
4986 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 4987 | |
7b4b6658 PZ |
4988 | if (hwc->interrupts == MAX_INTERRUPTS) |
4989 | return; | |
15dbf27c | 4990 | |
7b4b6658 | 4991 | for (; overflow; overflow--) { |
a8b0ca17 | 4992 | if (__perf_event_overflow(event, throttle, |
5622f295 | 4993 | data, regs)) { |
7b4b6658 PZ |
4994 | /* |
4995 | * We inhibit the overflow from happening when | |
4996 | * hwc->interrupts == MAX_INTERRUPTS. | |
4997 | */ | |
4998 | break; | |
4999 | } | |
cf450a73 | 5000 | throttle = 1; |
7b4b6658 | 5001 | } |
15dbf27c PZ |
5002 | } |
5003 | ||
a4eaf7f1 | 5004 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5005 | struct perf_sample_data *data, |
5622f295 | 5006 | struct pt_regs *regs) |
7b4b6658 | 5007 | { |
cdd6c482 | 5008 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5009 | |
e7850595 | 5010 | local64_add(nr, &event->count); |
d6d020e9 | 5011 | |
0cff784a PZ |
5012 | if (!regs) |
5013 | return; | |
5014 | ||
6c7e550f | 5015 | if (!is_sampling_event(event)) |
7b4b6658 | 5016 | return; |
d6d020e9 | 5017 | |
5d81e5cf AV |
5018 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5019 | data->period = nr; | |
5020 | return perf_swevent_overflow(event, 1, data, regs); | |
5021 | } else | |
5022 | data->period = event->hw.last_period; | |
5023 | ||
0cff784a | 5024 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5025 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5026 | |
e7850595 | 5027 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5028 | return; |
df1a132b | 5029 | |
a8b0ca17 | 5030 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5031 | } |
5032 | ||
f5ffe02e FW |
5033 | static int perf_exclude_event(struct perf_event *event, |
5034 | struct pt_regs *regs) | |
5035 | { | |
a4eaf7f1 | 5036 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5037 | return 1; |
a4eaf7f1 | 5038 | |
f5ffe02e FW |
5039 | if (regs) { |
5040 | if (event->attr.exclude_user && user_mode(regs)) | |
5041 | return 1; | |
5042 | ||
5043 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5044 | return 1; | |
5045 | } | |
5046 | ||
5047 | return 0; | |
5048 | } | |
5049 | ||
cdd6c482 | 5050 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5051 | enum perf_type_id type, |
6fb2915d LZ |
5052 | u32 event_id, |
5053 | struct perf_sample_data *data, | |
5054 | struct pt_regs *regs) | |
15dbf27c | 5055 | { |
cdd6c482 | 5056 | if (event->attr.type != type) |
a21ca2ca | 5057 | return 0; |
f5ffe02e | 5058 | |
cdd6c482 | 5059 | if (event->attr.config != event_id) |
15dbf27c PZ |
5060 | return 0; |
5061 | ||
f5ffe02e FW |
5062 | if (perf_exclude_event(event, regs)) |
5063 | return 0; | |
15dbf27c PZ |
5064 | |
5065 | return 1; | |
5066 | } | |
5067 | ||
76e1d904 FW |
5068 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5069 | { | |
5070 | u64 val = event_id | (type << 32); | |
5071 | ||
5072 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5073 | } | |
5074 | ||
49f135ed FW |
5075 | static inline struct hlist_head * |
5076 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5077 | { |
49f135ed FW |
5078 | u64 hash = swevent_hash(type, event_id); |
5079 | ||
5080 | return &hlist->heads[hash]; | |
5081 | } | |
76e1d904 | 5082 | |
49f135ed FW |
5083 | /* For the read side: events when they trigger */ |
5084 | static inline struct hlist_head * | |
b28ab83c | 5085 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5086 | { |
5087 | struct swevent_hlist *hlist; | |
76e1d904 | 5088 | |
b28ab83c | 5089 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5090 | if (!hlist) |
5091 | return NULL; | |
5092 | ||
49f135ed FW |
5093 | return __find_swevent_head(hlist, type, event_id); |
5094 | } | |
5095 | ||
5096 | /* For the event head insertion and removal in the hlist */ | |
5097 | static inline struct hlist_head * | |
b28ab83c | 5098 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5099 | { |
5100 | struct swevent_hlist *hlist; | |
5101 | u32 event_id = event->attr.config; | |
5102 | u64 type = event->attr.type; | |
5103 | ||
5104 | /* | |
5105 | * Event scheduling is always serialized against hlist allocation | |
5106 | * and release. Which makes the protected version suitable here. | |
5107 | * The context lock guarantees that. | |
5108 | */ | |
b28ab83c | 5109 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5110 | lockdep_is_held(&event->ctx->lock)); |
5111 | if (!hlist) | |
5112 | return NULL; | |
5113 | ||
5114 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5115 | } |
5116 | ||
5117 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5118 | u64 nr, |
76e1d904 FW |
5119 | struct perf_sample_data *data, |
5120 | struct pt_regs *regs) | |
15dbf27c | 5121 | { |
b28ab83c | 5122 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5123 | struct perf_event *event; |
76e1d904 FW |
5124 | struct hlist_node *node; |
5125 | struct hlist_head *head; | |
15dbf27c | 5126 | |
76e1d904 | 5127 | rcu_read_lock(); |
b28ab83c | 5128 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5129 | if (!head) |
5130 | goto end; | |
5131 | ||
5132 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { | |
6fb2915d | 5133 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5134 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5135 | } |
76e1d904 FW |
5136 | end: |
5137 | rcu_read_unlock(); | |
15dbf27c PZ |
5138 | } |
5139 | ||
4ed7c92d | 5140 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5141 | { |
b28ab83c | 5142 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5143 | |
b28ab83c | 5144 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5145 | } |
645e8cc0 | 5146 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5147 | |
fa9f90be | 5148 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5149 | { |
b28ab83c | 5150 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5151 | |
b28ab83c | 5152 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5153 | } |
15dbf27c | 5154 | |
a8b0ca17 | 5155 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5156 | { |
a4234bfc | 5157 | struct perf_sample_data data; |
4ed7c92d PZ |
5158 | int rctx; |
5159 | ||
1c024eca | 5160 | preempt_disable_notrace(); |
4ed7c92d PZ |
5161 | rctx = perf_swevent_get_recursion_context(); |
5162 | if (rctx < 0) | |
5163 | return; | |
a4234bfc | 5164 | |
fd0d000b | 5165 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5166 | |
a8b0ca17 | 5167 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5168 | |
5169 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5170 | preempt_enable_notrace(); |
b8e83514 PZ |
5171 | } |
5172 | ||
cdd6c482 | 5173 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5174 | { |
15dbf27c PZ |
5175 | } |
5176 | ||
a4eaf7f1 | 5177 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5178 | { |
b28ab83c | 5179 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5180 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5181 | struct hlist_head *head; |
5182 | ||
6c7e550f | 5183 | if (is_sampling_event(event)) { |
7b4b6658 | 5184 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5185 | perf_swevent_set_period(event); |
7b4b6658 | 5186 | } |
76e1d904 | 5187 | |
a4eaf7f1 PZ |
5188 | hwc->state = !(flags & PERF_EF_START); |
5189 | ||
b28ab83c | 5190 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
5191 | if (WARN_ON_ONCE(!head)) |
5192 | return -EINVAL; | |
5193 | ||
5194 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5195 | ||
15dbf27c PZ |
5196 | return 0; |
5197 | } | |
5198 | ||
a4eaf7f1 | 5199 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5200 | { |
76e1d904 | 5201 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5202 | } |
5203 | ||
a4eaf7f1 | 5204 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5205 | { |
a4eaf7f1 | 5206 | event->hw.state = 0; |
d6d020e9 | 5207 | } |
aa9c4c0f | 5208 | |
a4eaf7f1 | 5209 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5210 | { |
a4eaf7f1 | 5211 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5212 | } |
5213 | ||
49f135ed FW |
5214 | /* Deref the hlist from the update side */ |
5215 | static inline struct swevent_hlist * | |
b28ab83c | 5216 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5217 | { |
b28ab83c PZ |
5218 | return rcu_dereference_protected(swhash->swevent_hlist, |
5219 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5220 | } |
5221 | ||
b28ab83c | 5222 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5223 | { |
b28ab83c | 5224 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5225 | |
49f135ed | 5226 | if (!hlist) |
76e1d904 FW |
5227 | return; |
5228 | ||
b28ab83c | 5229 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5230 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5231 | } |
5232 | ||
5233 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5234 | { | |
b28ab83c | 5235 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5236 | |
b28ab83c | 5237 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5238 | |
b28ab83c PZ |
5239 | if (!--swhash->hlist_refcount) |
5240 | swevent_hlist_release(swhash); | |
76e1d904 | 5241 | |
b28ab83c | 5242 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5243 | } |
5244 | ||
5245 | static void swevent_hlist_put(struct perf_event *event) | |
5246 | { | |
5247 | int cpu; | |
5248 | ||
5249 | if (event->cpu != -1) { | |
5250 | swevent_hlist_put_cpu(event, event->cpu); | |
5251 | return; | |
5252 | } | |
5253 | ||
5254 | for_each_possible_cpu(cpu) | |
5255 | swevent_hlist_put_cpu(event, cpu); | |
5256 | } | |
5257 | ||
5258 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5259 | { | |
b28ab83c | 5260 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5261 | int err = 0; |
5262 | ||
b28ab83c | 5263 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5264 | |
b28ab83c | 5265 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5266 | struct swevent_hlist *hlist; |
5267 | ||
5268 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5269 | if (!hlist) { | |
5270 | err = -ENOMEM; | |
5271 | goto exit; | |
5272 | } | |
b28ab83c | 5273 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5274 | } |
b28ab83c | 5275 | swhash->hlist_refcount++; |
9ed6060d | 5276 | exit: |
b28ab83c | 5277 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5278 | |
5279 | return err; | |
5280 | } | |
5281 | ||
5282 | static int swevent_hlist_get(struct perf_event *event) | |
5283 | { | |
5284 | int err; | |
5285 | int cpu, failed_cpu; | |
5286 | ||
5287 | if (event->cpu != -1) | |
5288 | return swevent_hlist_get_cpu(event, event->cpu); | |
5289 | ||
5290 | get_online_cpus(); | |
5291 | for_each_possible_cpu(cpu) { | |
5292 | err = swevent_hlist_get_cpu(event, cpu); | |
5293 | if (err) { | |
5294 | failed_cpu = cpu; | |
5295 | goto fail; | |
5296 | } | |
5297 | } | |
5298 | put_online_cpus(); | |
5299 | ||
5300 | return 0; | |
9ed6060d | 5301 | fail: |
76e1d904 FW |
5302 | for_each_possible_cpu(cpu) { |
5303 | if (cpu == failed_cpu) | |
5304 | break; | |
5305 | swevent_hlist_put_cpu(event, cpu); | |
5306 | } | |
5307 | ||
5308 | put_online_cpus(); | |
5309 | return err; | |
5310 | } | |
5311 | ||
c5905afb | 5312 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5313 | |
b0a873eb PZ |
5314 | static void sw_perf_event_destroy(struct perf_event *event) |
5315 | { | |
5316 | u64 event_id = event->attr.config; | |
95476b64 | 5317 | |
b0a873eb PZ |
5318 | WARN_ON(event->parent); |
5319 | ||
c5905afb | 5320 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5321 | swevent_hlist_put(event); |
5322 | } | |
5323 | ||
5324 | static int perf_swevent_init(struct perf_event *event) | |
5325 | { | |
5326 | int event_id = event->attr.config; | |
5327 | ||
5328 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5329 | return -ENOENT; | |
5330 | ||
2481c5fa SE |
5331 | /* |
5332 | * no branch sampling for software events | |
5333 | */ | |
5334 | if (has_branch_stack(event)) | |
5335 | return -EOPNOTSUPP; | |
5336 | ||
b0a873eb PZ |
5337 | switch (event_id) { |
5338 | case PERF_COUNT_SW_CPU_CLOCK: | |
5339 | case PERF_COUNT_SW_TASK_CLOCK: | |
5340 | return -ENOENT; | |
5341 | ||
5342 | default: | |
5343 | break; | |
5344 | } | |
5345 | ||
ce677831 | 5346 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5347 | return -ENOENT; |
5348 | ||
5349 | if (!event->parent) { | |
5350 | int err; | |
5351 | ||
5352 | err = swevent_hlist_get(event); | |
5353 | if (err) | |
5354 | return err; | |
5355 | ||
c5905afb | 5356 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5357 | event->destroy = sw_perf_event_destroy; |
5358 | } | |
5359 | ||
5360 | return 0; | |
5361 | } | |
5362 | ||
35edc2a5 PZ |
5363 | static int perf_swevent_event_idx(struct perf_event *event) |
5364 | { | |
5365 | return 0; | |
5366 | } | |
5367 | ||
b0a873eb | 5368 | static struct pmu perf_swevent = { |
89a1e187 | 5369 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5370 | |
b0a873eb | 5371 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5372 | .add = perf_swevent_add, |
5373 | .del = perf_swevent_del, | |
5374 | .start = perf_swevent_start, | |
5375 | .stop = perf_swevent_stop, | |
1c024eca | 5376 | .read = perf_swevent_read, |
35edc2a5 PZ |
5377 | |
5378 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5379 | }; |
5380 | ||
b0a873eb PZ |
5381 | #ifdef CONFIG_EVENT_TRACING |
5382 | ||
1c024eca PZ |
5383 | static int perf_tp_filter_match(struct perf_event *event, |
5384 | struct perf_sample_data *data) | |
5385 | { | |
5386 | void *record = data->raw->data; | |
5387 | ||
5388 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5389 | return 1; | |
5390 | return 0; | |
5391 | } | |
5392 | ||
5393 | static int perf_tp_event_match(struct perf_event *event, | |
5394 | struct perf_sample_data *data, | |
5395 | struct pt_regs *regs) | |
5396 | { | |
a0f7d0f7 FW |
5397 | if (event->hw.state & PERF_HES_STOPPED) |
5398 | return 0; | |
580d607c PZ |
5399 | /* |
5400 | * All tracepoints are from kernel-space. | |
5401 | */ | |
5402 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5403 | return 0; |
5404 | ||
5405 | if (!perf_tp_filter_match(event, data)) | |
5406 | return 0; | |
5407 | ||
5408 | return 1; | |
5409 | } | |
5410 | ||
5411 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5412 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5413 | struct task_struct *task) | |
95476b64 FW |
5414 | { |
5415 | struct perf_sample_data data; | |
1c024eca PZ |
5416 | struct perf_event *event; |
5417 | struct hlist_node *node; | |
5418 | ||
95476b64 FW |
5419 | struct perf_raw_record raw = { |
5420 | .size = entry_size, | |
5421 | .data = record, | |
5422 | }; | |
5423 | ||
fd0d000b | 5424 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5425 | data.raw = &raw; |
5426 | ||
1c024eca PZ |
5427 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { |
5428 | if (perf_tp_event_match(event, &data, regs)) | |
a8b0ca17 | 5429 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5430 | } |
ecc55f84 | 5431 | |
e6dab5ff AV |
5432 | /* |
5433 | * If we got specified a target task, also iterate its context and | |
5434 | * deliver this event there too. | |
5435 | */ | |
5436 | if (task && task != current) { | |
5437 | struct perf_event_context *ctx; | |
5438 | struct trace_entry *entry = record; | |
5439 | ||
5440 | rcu_read_lock(); | |
5441 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5442 | if (!ctx) | |
5443 | goto unlock; | |
5444 | ||
5445 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5446 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5447 | continue; | |
5448 | if (event->attr.config != entry->type) | |
5449 | continue; | |
5450 | if (perf_tp_event_match(event, &data, regs)) | |
5451 | perf_swevent_event(event, count, &data, regs); | |
5452 | } | |
5453 | unlock: | |
5454 | rcu_read_unlock(); | |
5455 | } | |
5456 | ||
ecc55f84 | 5457 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5458 | } |
5459 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5460 | ||
cdd6c482 | 5461 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5462 | { |
1c024eca | 5463 | perf_trace_destroy(event); |
e077df4f PZ |
5464 | } |
5465 | ||
b0a873eb | 5466 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5467 | { |
76e1d904 FW |
5468 | int err; |
5469 | ||
b0a873eb PZ |
5470 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5471 | return -ENOENT; | |
5472 | ||
2481c5fa SE |
5473 | /* |
5474 | * no branch sampling for tracepoint events | |
5475 | */ | |
5476 | if (has_branch_stack(event)) | |
5477 | return -EOPNOTSUPP; | |
5478 | ||
1c024eca PZ |
5479 | err = perf_trace_init(event); |
5480 | if (err) | |
b0a873eb | 5481 | return err; |
e077df4f | 5482 | |
cdd6c482 | 5483 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5484 | |
b0a873eb PZ |
5485 | return 0; |
5486 | } | |
5487 | ||
5488 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5489 | .task_ctx_nr = perf_sw_context, |
5490 | ||
b0a873eb | 5491 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5492 | .add = perf_trace_add, |
5493 | .del = perf_trace_del, | |
5494 | .start = perf_swevent_start, | |
5495 | .stop = perf_swevent_stop, | |
b0a873eb | 5496 | .read = perf_swevent_read, |
35edc2a5 PZ |
5497 | |
5498 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5499 | }; |
5500 | ||
5501 | static inline void perf_tp_register(void) | |
5502 | { | |
2e80a82a | 5503 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5504 | } |
6fb2915d LZ |
5505 | |
5506 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5507 | { | |
5508 | char *filter_str; | |
5509 | int ret; | |
5510 | ||
5511 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5512 | return -EINVAL; | |
5513 | ||
5514 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5515 | if (IS_ERR(filter_str)) | |
5516 | return PTR_ERR(filter_str); | |
5517 | ||
5518 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5519 | ||
5520 | kfree(filter_str); | |
5521 | return ret; | |
5522 | } | |
5523 | ||
5524 | static void perf_event_free_filter(struct perf_event *event) | |
5525 | { | |
5526 | ftrace_profile_free_filter(event); | |
5527 | } | |
5528 | ||
e077df4f | 5529 | #else |
6fb2915d | 5530 | |
b0a873eb | 5531 | static inline void perf_tp_register(void) |
e077df4f | 5532 | { |
e077df4f | 5533 | } |
6fb2915d LZ |
5534 | |
5535 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5536 | { | |
5537 | return -ENOENT; | |
5538 | } | |
5539 | ||
5540 | static void perf_event_free_filter(struct perf_event *event) | |
5541 | { | |
5542 | } | |
5543 | ||
07b139c8 | 5544 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5545 | |
24f1e32c | 5546 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5547 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5548 | { |
f5ffe02e FW |
5549 | struct perf_sample_data sample; |
5550 | struct pt_regs *regs = data; | |
5551 | ||
fd0d000b | 5552 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 5553 | |
a4eaf7f1 | 5554 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5555 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5556 | } |
5557 | #endif | |
5558 | ||
b0a873eb PZ |
5559 | /* |
5560 | * hrtimer based swevent callback | |
5561 | */ | |
f29ac756 | 5562 | |
b0a873eb | 5563 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5564 | { |
b0a873eb PZ |
5565 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5566 | struct perf_sample_data data; | |
5567 | struct pt_regs *regs; | |
5568 | struct perf_event *event; | |
5569 | u64 period; | |
f29ac756 | 5570 | |
b0a873eb | 5571 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5572 | |
5573 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5574 | return HRTIMER_NORESTART; | |
5575 | ||
b0a873eb | 5576 | event->pmu->read(event); |
f344011c | 5577 | |
fd0d000b | 5578 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
5579 | regs = get_irq_regs(); |
5580 | ||
5581 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5582 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 5583 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
5584 | ret = HRTIMER_NORESTART; |
5585 | } | |
24f1e32c | 5586 | |
b0a873eb PZ |
5587 | period = max_t(u64, 10000, event->hw.sample_period); |
5588 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5589 | |
b0a873eb | 5590 | return ret; |
f29ac756 PZ |
5591 | } |
5592 | ||
b0a873eb | 5593 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5594 | { |
b0a873eb | 5595 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5596 | s64 period; |
5597 | ||
5598 | if (!is_sampling_event(event)) | |
5599 | return; | |
f5ffe02e | 5600 | |
5d508e82 FBH |
5601 | period = local64_read(&hwc->period_left); |
5602 | if (period) { | |
5603 | if (period < 0) | |
5604 | period = 10000; | |
fa407f35 | 5605 | |
5d508e82 FBH |
5606 | local64_set(&hwc->period_left, 0); |
5607 | } else { | |
5608 | period = max_t(u64, 10000, hwc->sample_period); | |
5609 | } | |
5610 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 5611 | ns_to_ktime(period), 0, |
b5ab4cd5 | 5612 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 5613 | } |
b0a873eb PZ |
5614 | |
5615 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 5616 | { |
b0a873eb PZ |
5617 | struct hw_perf_event *hwc = &event->hw; |
5618 | ||
6c7e550f | 5619 | if (is_sampling_event(event)) { |
b0a873eb | 5620 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 5621 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
5622 | |
5623 | hrtimer_cancel(&hwc->hrtimer); | |
5624 | } | |
24f1e32c FW |
5625 | } |
5626 | ||
ba3dd36c PZ |
5627 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5628 | { | |
5629 | struct hw_perf_event *hwc = &event->hw; | |
5630 | ||
5631 | if (!is_sampling_event(event)) | |
5632 | return; | |
5633 | ||
5634 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5635 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
5636 | ||
5637 | /* | |
5638 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
5639 | * mapping and avoid the whole period adjust feedback stuff. | |
5640 | */ | |
5641 | if (event->attr.freq) { | |
5642 | long freq = event->attr.sample_freq; | |
5643 | ||
5644 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
5645 | hwc->sample_period = event->attr.sample_period; | |
5646 | local64_set(&hwc->period_left, hwc->sample_period); | |
5647 | event->attr.freq = 0; | |
5648 | } | |
5649 | } | |
5650 | ||
b0a873eb PZ |
5651 | /* |
5652 | * Software event: cpu wall time clock | |
5653 | */ | |
5654 | ||
5655 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 5656 | { |
b0a873eb PZ |
5657 | s64 prev; |
5658 | u64 now; | |
5659 | ||
a4eaf7f1 | 5660 | now = local_clock(); |
b0a873eb PZ |
5661 | prev = local64_xchg(&event->hw.prev_count, now); |
5662 | local64_add(now - prev, &event->count); | |
24f1e32c | 5663 | } |
24f1e32c | 5664 | |
a4eaf7f1 | 5665 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5666 | { |
a4eaf7f1 | 5667 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 5668 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5669 | } |
5670 | ||
a4eaf7f1 | 5671 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 5672 | { |
b0a873eb PZ |
5673 | perf_swevent_cancel_hrtimer(event); |
5674 | cpu_clock_event_update(event); | |
5675 | } | |
f29ac756 | 5676 | |
a4eaf7f1 PZ |
5677 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
5678 | { | |
5679 | if (flags & PERF_EF_START) | |
5680 | cpu_clock_event_start(event, flags); | |
5681 | ||
5682 | return 0; | |
5683 | } | |
5684 | ||
5685 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
5686 | { | |
5687 | cpu_clock_event_stop(event, flags); | |
5688 | } | |
5689 | ||
b0a873eb PZ |
5690 | static void cpu_clock_event_read(struct perf_event *event) |
5691 | { | |
5692 | cpu_clock_event_update(event); | |
5693 | } | |
f344011c | 5694 | |
b0a873eb PZ |
5695 | static int cpu_clock_event_init(struct perf_event *event) |
5696 | { | |
5697 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5698 | return -ENOENT; | |
5699 | ||
5700 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
5701 | return -ENOENT; | |
5702 | ||
2481c5fa SE |
5703 | /* |
5704 | * no branch sampling for software events | |
5705 | */ | |
5706 | if (has_branch_stack(event)) | |
5707 | return -EOPNOTSUPP; | |
5708 | ||
ba3dd36c PZ |
5709 | perf_swevent_init_hrtimer(event); |
5710 | ||
b0a873eb | 5711 | return 0; |
f29ac756 PZ |
5712 | } |
5713 | ||
b0a873eb | 5714 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
5715 | .task_ctx_nr = perf_sw_context, |
5716 | ||
b0a873eb | 5717 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
5718 | .add = cpu_clock_event_add, |
5719 | .del = cpu_clock_event_del, | |
5720 | .start = cpu_clock_event_start, | |
5721 | .stop = cpu_clock_event_stop, | |
b0a873eb | 5722 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
5723 | |
5724 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5725 | }; |
5726 | ||
5727 | /* | |
5728 | * Software event: task time clock | |
5729 | */ | |
5730 | ||
5731 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 5732 | { |
b0a873eb PZ |
5733 | u64 prev; |
5734 | s64 delta; | |
5c92d124 | 5735 | |
b0a873eb PZ |
5736 | prev = local64_xchg(&event->hw.prev_count, now); |
5737 | delta = now - prev; | |
5738 | local64_add(delta, &event->count); | |
5739 | } | |
5c92d124 | 5740 | |
a4eaf7f1 | 5741 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5742 | { |
a4eaf7f1 | 5743 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 5744 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5745 | } |
5746 | ||
a4eaf7f1 | 5747 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
5748 | { |
5749 | perf_swevent_cancel_hrtimer(event); | |
5750 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
5751 | } |
5752 | ||
5753 | static int task_clock_event_add(struct perf_event *event, int flags) | |
5754 | { | |
5755 | if (flags & PERF_EF_START) | |
5756 | task_clock_event_start(event, flags); | |
b0a873eb | 5757 | |
a4eaf7f1 PZ |
5758 | return 0; |
5759 | } | |
5760 | ||
5761 | static void task_clock_event_del(struct perf_event *event, int flags) | |
5762 | { | |
5763 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
5764 | } |
5765 | ||
5766 | static void task_clock_event_read(struct perf_event *event) | |
5767 | { | |
768a06e2 PZ |
5768 | u64 now = perf_clock(); |
5769 | u64 delta = now - event->ctx->timestamp; | |
5770 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
5771 | |
5772 | task_clock_event_update(event, time); | |
5773 | } | |
5774 | ||
5775 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 5776 | { |
b0a873eb PZ |
5777 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5778 | return -ENOENT; | |
5779 | ||
5780 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
5781 | return -ENOENT; | |
5782 | ||
2481c5fa SE |
5783 | /* |
5784 | * no branch sampling for software events | |
5785 | */ | |
5786 | if (has_branch_stack(event)) | |
5787 | return -EOPNOTSUPP; | |
5788 | ||
ba3dd36c PZ |
5789 | perf_swevent_init_hrtimer(event); |
5790 | ||
b0a873eb | 5791 | return 0; |
6fb2915d LZ |
5792 | } |
5793 | ||
b0a873eb | 5794 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
5795 | .task_ctx_nr = perf_sw_context, |
5796 | ||
b0a873eb | 5797 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
5798 | .add = task_clock_event_add, |
5799 | .del = task_clock_event_del, | |
5800 | .start = task_clock_event_start, | |
5801 | .stop = task_clock_event_stop, | |
b0a873eb | 5802 | .read = task_clock_event_read, |
35edc2a5 PZ |
5803 | |
5804 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 5805 | }; |
6fb2915d | 5806 | |
ad5133b7 | 5807 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 5808 | { |
e077df4f | 5809 | } |
6fb2915d | 5810 | |
ad5133b7 | 5811 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 5812 | { |
ad5133b7 | 5813 | return 0; |
6fb2915d LZ |
5814 | } |
5815 | ||
ad5133b7 | 5816 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 5817 | { |
ad5133b7 | 5818 | perf_pmu_disable(pmu); |
6fb2915d LZ |
5819 | } |
5820 | ||
ad5133b7 PZ |
5821 | static int perf_pmu_commit_txn(struct pmu *pmu) |
5822 | { | |
5823 | perf_pmu_enable(pmu); | |
5824 | return 0; | |
5825 | } | |
e077df4f | 5826 | |
ad5133b7 | 5827 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 5828 | { |
ad5133b7 | 5829 | perf_pmu_enable(pmu); |
24f1e32c FW |
5830 | } |
5831 | ||
35edc2a5 PZ |
5832 | static int perf_event_idx_default(struct perf_event *event) |
5833 | { | |
5834 | return event->hw.idx + 1; | |
5835 | } | |
5836 | ||
8dc85d54 PZ |
5837 | /* |
5838 | * Ensures all contexts with the same task_ctx_nr have the same | |
5839 | * pmu_cpu_context too. | |
5840 | */ | |
5841 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 5842 | { |
8dc85d54 | 5843 | struct pmu *pmu; |
b326e956 | 5844 | |
8dc85d54 PZ |
5845 | if (ctxn < 0) |
5846 | return NULL; | |
24f1e32c | 5847 | |
8dc85d54 PZ |
5848 | list_for_each_entry(pmu, &pmus, entry) { |
5849 | if (pmu->task_ctx_nr == ctxn) | |
5850 | return pmu->pmu_cpu_context; | |
5851 | } | |
24f1e32c | 5852 | |
8dc85d54 | 5853 | return NULL; |
24f1e32c FW |
5854 | } |
5855 | ||
51676957 | 5856 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 5857 | { |
51676957 PZ |
5858 | int cpu; |
5859 | ||
5860 | for_each_possible_cpu(cpu) { | |
5861 | struct perf_cpu_context *cpuctx; | |
5862 | ||
5863 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
5864 | ||
5865 | if (cpuctx->active_pmu == old_pmu) | |
5866 | cpuctx->active_pmu = pmu; | |
5867 | } | |
5868 | } | |
5869 | ||
5870 | static void free_pmu_context(struct pmu *pmu) | |
5871 | { | |
5872 | struct pmu *i; | |
f5ffe02e | 5873 | |
8dc85d54 | 5874 | mutex_lock(&pmus_lock); |
0475f9ea | 5875 | /* |
8dc85d54 | 5876 | * Like a real lame refcount. |
0475f9ea | 5877 | */ |
51676957 PZ |
5878 | list_for_each_entry(i, &pmus, entry) { |
5879 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
5880 | update_pmu_context(i, pmu); | |
8dc85d54 | 5881 | goto out; |
51676957 | 5882 | } |
8dc85d54 | 5883 | } |
d6d020e9 | 5884 | |
51676957 | 5885 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
5886 | out: |
5887 | mutex_unlock(&pmus_lock); | |
24f1e32c | 5888 | } |
2e80a82a | 5889 | static struct idr pmu_idr; |
d6d020e9 | 5890 | |
abe43400 PZ |
5891 | static ssize_t |
5892 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
5893 | { | |
5894 | struct pmu *pmu = dev_get_drvdata(dev); | |
5895 | ||
5896 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
5897 | } | |
5898 | ||
5899 | static struct device_attribute pmu_dev_attrs[] = { | |
5900 | __ATTR_RO(type), | |
5901 | __ATTR_NULL, | |
5902 | }; | |
5903 | ||
5904 | static int pmu_bus_running; | |
5905 | static struct bus_type pmu_bus = { | |
5906 | .name = "event_source", | |
5907 | .dev_attrs = pmu_dev_attrs, | |
5908 | }; | |
5909 | ||
5910 | static void pmu_dev_release(struct device *dev) | |
5911 | { | |
5912 | kfree(dev); | |
5913 | } | |
5914 | ||
5915 | static int pmu_dev_alloc(struct pmu *pmu) | |
5916 | { | |
5917 | int ret = -ENOMEM; | |
5918 | ||
5919 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
5920 | if (!pmu->dev) | |
5921 | goto out; | |
5922 | ||
0c9d42ed | 5923 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
5924 | device_initialize(pmu->dev); |
5925 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
5926 | if (ret) | |
5927 | goto free_dev; | |
5928 | ||
5929 | dev_set_drvdata(pmu->dev, pmu); | |
5930 | pmu->dev->bus = &pmu_bus; | |
5931 | pmu->dev->release = pmu_dev_release; | |
5932 | ret = device_add(pmu->dev); | |
5933 | if (ret) | |
5934 | goto free_dev; | |
5935 | ||
5936 | out: | |
5937 | return ret; | |
5938 | ||
5939 | free_dev: | |
5940 | put_device(pmu->dev); | |
5941 | goto out; | |
5942 | } | |
5943 | ||
547e9fd7 | 5944 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 5945 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 5946 | |
2e80a82a | 5947 | int perf_pmu_register(struct pmu *pmu, char *name, int type) |
24f1e32c | 5948 | { |
108b02cf | 5949 | int cpu, ret; |
24f1e32c | 5950 | |
b0a873eb | 5951 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
5952 | ret = -ENOMEM; |
5953 | pmu->pmu_disable_count = alloc_percpu(int); | |
5954 | if (!pmu->pmu_disable_count) | |
5955 | goto unlock; | |
f29ac756 | 5956 | |
2e80a82a PZ |
5957 | pmu->type = -1; |
5958 | if (!name) | |
5959 | goto skip_type; | |
5960 | pmu->name = name; | |
5961 | ||
5962 | if (type < 0) { | |
5963 | int err = idr_pre_get(&pmu_idr, GFP_KERNEL); | |
5964 | if (!err) | |
5965 | goto free_pdc; | |
5966 | ||
5967 | err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type); | |
5968 | if (err) { | |
5969 | ret = err; | |
5970 | goto free_pdc; | |
5971 | } | |
5972 | } | |
5973 | pmu->type = type; | |
5974 | ||
abe43400 PZ |
5975 | if (pmu_bus_running) { |
5976 | ret = pmu_dev_alloc(pmu); | |
5977 | if (ret) | |
5978 | goto free_idr; | |
5979 | } | |
5980 | ||
2e80a82a | 5981 | skip_type: |
8dc85d54 PZ |
5982 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
5983 | if (pmu->pmu_cpu_context) | |
5984 | goto got_cpu_context; | |
f29ac756 | 5985 | |
108b02cf PZ |
5986 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
5987 | if (!pmu->pmu_cpu_context) | |
abe43400 | 5988 | goto free_dev; |
f344011c | 5989 | |
108b02cf PZ |
5990 | for_each_possible_cpu(cpu) { |
5991 | struct perf_cpu_context *cpuctx; | |
5992 | ||
5993 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 5994 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 5995 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 5996 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 5997 | cpuctx->ctx.type = cpu_context; |
108b02cf | 5998 | cpuctx->ctx.pmu = pmu; |
e9d2b064 PZ |
5999 | cpuctx->jiffies_interval = 1; |
6000 | INIT_LIST_HEAD(&cpuctx->rotation_list); | |
51676957 | 6001 | cpuctx->active_pmu = pmu; |
108b02cf | 6002 | } |
76e1d904 | 6003 | |
8dc85d54 | 6004 | got_cpu_context: |
ad5133b7 PZ |
6005 | if (!pmu->start_txn) { |
6006 | if (pmu->pmu_enable) { | |
6007 | /* | |
6008 | * If we have pmu_enable/pmu_disable calls, install | |
6009 | * transaction stubs that use that to try and batch | |
6010 | * hardware accesses. | |
6011 | */ | |
6012 | pmu->start_txn = perf_pmu_start_txn; | |
6013 | pmu->commit_txn = perf_pmu_commit_txn; | |
6014 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6015 | } else { | |
6016 | pmu->start_txn = perf_pmu_nop_void; | |
6017 | pmu->commit_txn = perf_pmu_nop_int; | |
6018 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6019 | } |
5c92d124 | 6020 | } |
15dbf27c | 6021 | |
ad5133b7 PZ |
6022 | if (!pmu->pmu_enable) { |
6023 | pmu->pmu_enable = perf_pmu_nop_void; | |
6024 | pmu->pmu_disable = perf_pmu_nop_void; | |
6025 | } | |
6026 | ||
35edc2a5 PZ |
6027 | if (!pmu->event_idx) |
6028 | pmu->event_idx = perf_event_idx_default; | |
6029 | ||
b0a873eb | 6030 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6031 | ret = 0; |
6032 | unlock: | |
b0a873eb PZ |
6033 | mutex_unlock(&pmus_lock); |
6034 | ||
33696fc0 | 6035 | return ret; |
108b02cf | 6036 | |
abe43400 PZ |
6037 | free_dev: |
6038 | device_del(pmu->dev); | |
6039 | put_device(pmu->dev); | |
6040 | ||
2e80a82a PZ |
6041 | free_idr: |
6042 | if (pmu->type >= PERF_TYPE_MAX) | |
6043 | idr_remove(&pmu_idr, pmu->type); | |
6044 | ||
108b02cf PZ |
6045 | free_pdc: |
6046 | free_percpu(pmu->pmu_disable_count); | |
6047 | goto unlock; | |
f29ac756 PZ |
6048 | } |
6049 | ||
b0a873eb | 6050 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6051 | { |
b0a873eb PZ |
6052 | mutex_lock(&pmus_lock); |
6053 | list_del_rcu(&pmu->entry); | |
6054 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6055 | |
0475f9ea | 6056 | /* |
cde8e884 PZ |
6057 | * We dereference the pmu list under both SRCU and regular RCU, so |
6058 | * synchronize against both of those. | |
0475f9ea | 6059 | */ |
b0a873eb | 6060 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6061 | synchronize_rcu(); |
d6d020e9 | 6062 | |
33696fc0 | 6063 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6064 | if (pmu->type >= PERF_TYPE_MAX) |
6065 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6066 | device_del(pmu->dev); |
6067 | put_device(pmu->dev); | |
51676957 | 6068 | free_pmu_context(pmu); |
b0a873eb | 6069 | } |
d6d020e9 | 6070 | |
b0a873eb PZ |
6071 | struct pmu *perf_init_event(struct perf_event *event) |
6072 | { | |
6073 | struct pmu *pmu = NULL; | |
6074 | int idx; | |
940c5b29 | 6075 | int ret; |
b0a873eb PZ |
6076 | |
6077 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6078 | |
6079 | rcu_read_lock(); | |
6080 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6081 | rcu_read_unlock(); | |
940c5b29 | 6082 | if (pmu) { |
7e5b2a01 | 6083 | event->pmu = pmu; |
940c5b29 LM |
6084 | ret = pmu->event_init(event); |
6085 | if (ret) | |
6086 | pmu = ERR_PTR(ret); | |
2e80a82a | 6087 | goto unlock; |
940c5b29 | 6088 | } |
2e80a82a | 6089 | |
b0a873eb | 6090 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 6091 | event->pmu = pmu; |
940c5b29 | 6092 | ret = pmu->event_init(event); |
b0a873eb | 6093 | if (!ret) |
e5f4d339 | 6094 | goto unlock; |
76e1d904 | 6095 | |
b0a873eb PZ |
6096 | if (ret != -ENOENT) { |
6097 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6098 | goto unlock; |
f344011c | 6099 | } |
5c92d124 | 6100 | } |
e5f4d339 PZ |
6101 | pmu = ERR_PTR(-ENOENT); |
6102 | unlock: | |
b0a873eb | 6103 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6104 | |
4aeb0b42 | 6105 | return pmu; |
5c92d124 IM |
6106 | } |
6107 | ||
0793a61d | 6108 | /* |
cdd6c482 | 6109 | * Allocate and initialize a event structure |
0793a61d | 6110 | */ |
cdd6c482 | 6111 | static struct perf_event * |
c3f00c70 | 6112 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6113 | struct task_struct *task, |
6114 | struct perf_event *group_leader, | |
6115 | struct perf_event *parent_event, | |
4dc0da86 AK |
6116 | perf_overflow_handler_t overflow_handler, |
6117 | void *context) | |
0793a61d | 6118 | { |
51b0fe39 | 6119 | struct pmu *pmu; |
cdd6c482 IM |
6120 | struct perf_event *event; |
6121 | struct hw_perf_event *hwc; | |
d5d2bc0d | 6122 | long err; |
0793a61d | 6123 | |
66832eb4 ON |
6124 | if ((unsigned)cpu >= nr_cpu_ids) { |
6125 | if (!task || cpu != -1) | |
6126 | return ERR_PTR(-EINVAL); | |
6127 | } | |
6128 | ||
c3f00c70 | 6129 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6130 | if (!event) |
d5d2bc0d | 6131 | return ERR_PTR(-ENOMEM); |
0793a61d | 6132 | |
04289bb9 | 6133 | /* |
cdd6c482 | 6134 | * Single events are their own group leaders, with an |
04289bb9 IM |
6135 | * empty sibling list: |
6136 | */ | |
6137 | if (!group_leader) | |
cdd6c482 | 6138 | group_leader = event; |
04289bb9 | 6139 | |
cdd6c482 IM |
6140 | mutex_init(&event->child_mutex); |
6141 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6142 | |
cdd6c482 IM |
6143 | INIT_LIST_HEAD(&event->group_entry); |
6144 | INIT_LIST_HEAD(&event->event_entry); | |
6145 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
6146 | INIT_LIST_HEAD(&event->rb_entry); |
6147 | ||
cdd6c482 | 6148 | init_waitqueue_head(&event->waitq); |
e360adbe | 6149 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6150 | |
cdd6c482 | 6151 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6152 | |
a6fa941d | 6153 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6154 | event->cpu = cpu; |
6155 | event->attr = *attr; | |
6156 | event->group_leader = group_leader; | |
6157 | event->pmu = NULL; | |
cdd6c482 | 6158 | event->oncpu = -1; |
a96bbc16 | 6159 | |
cdd6c482 | 6160 | event->parent = parent_event; |
b84fbc9f | 6161 | |
cdd6c482 IM |
6162 | event->ns = get_pid_ns(current->nsproxy->pid_ns); |
6163 | event->id = atomic64_inc_return(&perf_event_id); | |
a96bbc16 | 6164 | |
cdd6c482 | 6165 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6166 | |
d580ff86 PZ |
6167 | if (task) { |
6168 | event->attach_state = PERF_ATTACH_TASK; | |
6169 | #ifdef CONFIG_HAVE_HW_BREAKPOINT | |
6170 | /* | |
6171 | * hw_breakpoint is a bit difficult here.. | |
6172 | */ | |
6173 | if (attr->type == PERF_TYPE_BREAKPOINT) | |
6174 | event->hw.bp_target = task; | |
6175 | #endif | |
6176 | } | |
6177 | ||
4dc0da86 | 6178 | if (!overflow_handler && parent_event) { |
b326e956 | 6179 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6180 | context = parent_event->overflow_handler_context; |
6181 | } | |
66832eb4 | 6182 | |
b326e956 | 6183 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6184 | event->overflow_handler_context = context; |
97eaf530 | 6185 | |
0d48696f | 6186 | if (attr->disabled) |
cdd6c482 | 6187 | event->state = PERF_EVENT_STATE_OFF; |
a86ed508 | 6188 | |
4aeb0b42 | 6189 | pmu = NULL; |
b8e83514 | 6190 | |
cdd6c482 | 6191 | hwc = &event->hw; |
bd2b5b12 | 6192 | hwc->sample_period = attr->sample_period; |
0d48696f | 6193 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6194 | hwc->sample_period = 1; |
eced1dfc | 6195 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6196 | |
e7850595 | 6197 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6198 | |
2023b359 | 6199 | /* |
cdd6c482 | 6200 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6201 | */ |
3dab77fb | 6202 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
2023b359 PZ |
6203 | goto done; |
6204 | ||
b0a873eb | 6205 | pmu = perf_init_event(event); |
974802ea | 6206 | |
d5d2bc0d PM |
6207 | done: |
6208 | err = 0; | |
4aeb0b42 | 6209 | if (!pmu) |
d5d2bc0d | 6210 | err = -EINVAL; |
4aeb0b42 RR |
6211 | else if (IS_ERR(pmu)) |
6212 | err = PTR_ERR(pmu); | |
5c92d124 | 6213 | |
d5d2bc0d | 6214 | if (err) { |
cdd6c482 IM |
6215 | if (event->ns) |
6216 | put_pid_ns(event->ns); | |
6217 | kfree(event); | |
d5d2bc0d | 6218 | return ERR_PTR(err); |
621a01ea | 6219 | } |
d5d2bc0d | 6220 | |
cdd6c482 | 6221 | if (!event->parent) { |
82cd6def | 6222 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 6223 | static_key_slow_inc(&perf_sched_events.key); |
3af9e859 | 6224 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
6225 | atomic_inc(&nr_mmap_events); |
6226 | if (event->attr.comm) | |
6227 | atomic_inc(&nr_comm_events); | |
6228 | if (event->attr.task) | |
6229 | atomic_inc(&nr_task_events); | |
927c7a9e FW |
6230 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6231 | err = get_callchain_buffers(); | |
6232 | if (err) { | |
6233 | free_event(event); | |
6234 | return ERR_PTR(err); | |
6235 | } | |
6236 | } | |
d010b332 SE |
6237 | if (has_branch_stack(event)) { |
6238 | static_key_slow_inc(&perf_sched_events.key); | |
6239 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6240 | atomic_inc(&per_cpu(perf_branch_stack_events, | |
6241 | event->cpu)); | |
6242 | } | |
f344011c | 6243 | } |
9ee318a7 | 6244 | |
cdd6c482 | 6245 | return event; |
0793a61d TG |
6246 | } |
6247 | ||
cdd6c482 IM |
6248 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6249 | struct perf_event_attr *attr) | |
974802ea | 6250 | { |
974802ea | 6251 | u32 size; |
cdf8073d | 6252 | int ret; |
974802ea PZ |
6253 | |
6254 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6255 | return -EFAULT; | |
6256 | ||
6257 | /* | |
6258 | * zero the full structure, so that a short copy will be nice. | |
6259 | */ | |
6260 | memset(attr, 0, sizeof(*attr)); | |
6261 | ||
6262 | ret = get_user(size, &uattr->size); | |
6263 | if (ret) | |
6264 | return ret; | |
6265 | ||
6266 | if (size > PAGE_SIZE) /* silly large */ | |
6267 | goto err_size; | |
6268 | ||
6269 | if (!size) /* abi compat */ | |
6270 | size = PERF_ATTR_SIZE_VER0; | |
6271 | ||
6272 | if (size < PERF_ATTR_SIZE_VER0) | |
6273 | goto err_size; | |
6274 | ||
6275 | /* | |
6276 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6277 | * ensure all the unknown bits are 0 - i.e. new |
6278 | * user-space does not rely on any kernel feature | |
6279 | * extensions we dont know about yet. | |
974802ea PZ |
6280 | */ |
6281 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6282 | unsigned char __user *addr; |
6283 | unsigned char __user *end; | |
6284 | unsigned char val; | |
974802ea | 6285 | |
cdf8073d IS |
6286 | addr = (void __user *)uattr + sizeof(*attr); |
6287 | end = (void __user *)uattr + size; | |
974802ea | 6288 | |
cdf8073d | 6289 | for (; addr < end; addr++) { |
974802ea PZ |
6290 | ret = get_user(val, addr); |
6291 | if (ret) | |
6292 | return ret; | |
6293 | if (val) | |
6294 | goto err_size; | |
6295 | } | |
b3e62e35 | 6296 | size = sizeof(*attr); |
974802ea PZ |
6297 | } |
6298 | ||
6299 | ret = copy_from_user(attr, uattr, size); | |
6300 | if (ret) | |
6301 | return -EFAULT; | |
6302 | ||
cd757645 | 6303 | if (attr->__reserved_1) |
974802ea PZ |
6304 | return -EINVAL; |
6305 | ||
6306 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6307 | return -EINVAL; | |
6308 | ||
6309 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6310 | return -EINVAL; | |
6311 | ||
bce38cd5 SE |
6312 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6313 | u64 mask = attr->branch_sample_type; | |
6314 | ||
6315 | /* only using defined bits */ | |
6316 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6317 | return -EINVAL; | |
6318 | ||
6319 | /* at least one branch bit must be set */ | |
6320 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6321 | return -EINVAL; | |
6322 | ||
6323 | /* kernel level capture: check permissions */ | |
6324 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
6325 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6326 | return -EACCES; | |
6327 | ||
6328 | /* propagate priv level, when not set for branch */ | |
6329 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6330 | ||
6331 | /* exclude_kernel checked on syscall entry */ | |
6332 | if (!attr->exclude_kernel) | |
6333 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6334 | ||
6335 | if (!attr->exclude_user) | |
6336 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6337 | ||
6338 | if (!attr->exclude_hv) | |
6339 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6340 | /* | |
6341 | * adjust user setting (for HW filter setup) | |
6342 | */ | |
6343 | attr->branch_sample_type = mask; | |
6344 | } | |
6345 | } | |
4018994f | 6346 | |
c5ebcedb | 6347 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 6348 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
6349 | if (ret) |
6350 | return ret; | |
6351 | } | |
6352 | ||
6353 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
6354 | if (!arch_perf_have_user_stack_dump()) | |
6355 | return -ENOSYS; | |
6356 | ||
6357 | /* | |
6358 | * We have __u32 type for the size, but so far | |
6359 | * we can only use __u16 as maximum due to the | |
6360 | * __u16 sample size limit. | |
6361 | */ | |
6362 | if (attr->sample_stack_user >= USHRT_MAX) | |
6363 | ret = -EINVAL; | |
6364 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
6365 | ret = -EINVAL; | |
6366 | } | |
4018994f | 6367 | |
974802ea PZ |
6368 | out: |
6369 | return ret; | |
6370 | ||
6371 | err_size: | |
6372 | put_user(sizeof(*attr), &uattr->size); | |
6373 | ret = -E2BIG; | |
6374 | goto out; | |
6375 | } | |
6376 | ||
ac9721f3 PZ |
6377 | static int |
6378 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6379 | { |
76369139 | 6380 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6381 | int ret = -EINVAL; |
6382 | ||
ac9721f3 | 6383 | if (!output_event) |
a4be7c27 PZ |
6384 | goto set; |
6385 | ||
ac9721f3 PZ |
6386 | /* don't allow circular references */ |
6387 | if (event == output_event) | |
a4be7c27 PZ |
6388 | goto out; |
6389 | ||
0f139300 PZ |
6390 | /* |
6391 | * Don't allow cross-cpu buffers | |
6392 | */ | |
6393 | if (output_event->cpu != event->cpu) | |
6394 | goto out; | |
6395 | ||
6396 | /* | |
76369139 | 6397 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6398 | */ |
6399 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6400 | goto out; | |
6401 | ||
a4be7c27 | 6402 | set: |
cdd6c482 | 6403 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6404 | /* Can't redirect output if we've got an active mmap() */ |
6405 | if (atomic_read(&event->mmap_count)) | |
6406 | goto unlock; | |
a4be7c27 | 6407 | |
ac9721f3 | 6408 | if (output_event) { |
76369139 FW |
6409 | /* get the rb we want to redirect to */ |
6410 | rb = ring_buffer_get(output_event); | |
6411 | if (!rb) | |
ac9721f3 | 6412 | goto unlock; |
a4be7c27 PZ |
6413 | } |
6414 | ||
76369139 FW |
6415 | old_rb = event->rb; |
6416 | rcu_assign_pointer(event->rb, rb); | |
10c6db11 PZ |
6417 | if (old_rb) |
6418 | ring_buffer_detach(event, old_rb); | |
a4be7c27 | 6419 | ret = 0; |
ac9721f3 PZ |
6420 | unlock: |
6421 | mutex_unlock(&event->mmap_mutex); | |
6422 | ||
76369139 FW |
6423 | if (old_rb) |
6424 | ring_buffer_put(old_rb); | |
a4be7c27 | 6425 | out: |
a4be7c27 PZ |
6426 | return ret; |
6427 | } | |
6428 | ||
0793a61d | 6429 | /** |
cdd6c482 | 6430 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6431 | * |
cdd6c482 | 6432 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6433 | * @pid: target pid |
9f66a381 | 6434 | * @cpu: target cpu |
cdd6c482 | 6435 | * @group_fd: group leader event fd |
0793a61d | 6436 | */ |
cdd6c482 IM |
6437 | SYSCALL_DEFINE5(perf_event_open, |
6438 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 6439 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 6440 | { |
b04243ef PZ |
6441 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6442 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
6443 | struct perf_event_attr attr; |
6444 | struct perf_event_context *ctx; | |
6445 | struct file *event_file = NULL; | |
04289bb9 | 6446 | struct file *group_file = NULL; |
38a81da2 | 6447 | struct task_struct *task = NULL; |
89a1e187 | 6448 | struct pmu *pmu; |
ea635c64 | 6449 | int event_fd; |
b04243ef | 6450 | int move_group = 0; |
04289bb9 | 6451 | int fput_needed = 0; |
dc86cabe | 6452 | int err; |
0793a61d | 6453 | |
2743a5b0 | 6454 | /* for future expandability... */ |
e5d1367f | 6455 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
6456 | return -EINVAL; |
6457 | ||
dc86cabe IM |
6458 | err = perf_copy_attr(attr_uptr, &attr); |
6459 | if (err) | |
6460 | return err; | |
eab656ae | 6461 | |
0764771d PZ |
6462 | if (!attr.exclude_kernel) { |
6463 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6464 | return -EACCES; | |
6465 | } | |
6466 | ||
df58ab24 | 6467 | if (attr.freq) { |
cdd6c482 | 6468 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
6469 | return -EINVAL; |
6470 | } | |
6471 | ||
e5d1367f SE |
6472 | /* |
6473 | * In cgroup mode, the pid argument is used to pass the fd | |
6474 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6475 | * designates the cpu on which to monitor threads from that | |
6476 | * cgroup. | |
6477 | */ | |
6478 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6479 | return -EINVAL; | |
6480 | ||
ea635c64 AV |
6481 | event_fd = get_unused_fd_flags(O_RDWR); |
6482 | if (event_fd < 0) | |
6483 | return event_fd; | |
6484 | ||
ac9721f3 | 6485 | if (group_fd != -1) { |
a6fa941d AV |
6486 | group_file = perf_fget_light(group_fd, &fput_needed); |
6487 | if (IS_ERR(group_file)) { | |
6488 | err = PTR_ERR(group_file); | |
d14b12d7 | 6489 | goto err_fd; |
ac9721f3 | 6490 | } |
a6fa941d | 6491 | group_leader = group_file->private_data; |
ac9721f3 PZ |
6492 | if (flags & PERF_FLAG_FD_OUTPUT) |
6493 | output_event = group_leader; | |
6494 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6495 | group_leader = NULL; | |
6496 | } | |
6497 | ||
e5d1367f | 6498 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6499 | task = find_lively_task_by_vpid(pid); |
6500 | if (IS_ERR(task)) { | |
6501 | err = PTR_ERR(task); | |
6502 | goto err_group_fd; | |
6503 | } | |
6504 | } | |
6505 | ||
fbfc623f YZ |
6506 | get_online_cpus(); |
6507 | ||
4dc0da86 AK |
6508 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6509 | NULL, NULL); | |
d14b12d7 SE |
6510 | if (IS_ERR(event)) { |
6511 | err = PTR_ERR(event); | |
c6be5a5c | 6512 | goto err_task; |
d14b12d7 SE |
6513 | } |
6514 | ||
e5d1367f SE |
6515 | if (flags & PERF_FLAG_PID_CGROUP) { |
6516 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
6517 | if (err) | |
6518 | goto err_alloc; | |
08309379 PZ |
6519 | /* |
6520 | * one more event: | |
6521 | * - that has cgroup constraint on event->cpu | |
6522 | * - that may need work on context switch | |
6523 | */ | |
6524 | atomic_inc(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 6525 | static_key_slow_inc(&perf_sched_events.key); |
e5d1367f SE |
6526 | } |
6527 | ||
89a1e187 PZ |
6528 | /* |
6529 | * Special case software events and allow them to be part of | |
6530 | * any hardware group. | |
6531 | */ | |
6532 | pmu = event->pmu; | |
b04243ef PZ |
6533 | |
6534 | if (group_leader && | |
6535 | (is_software_event(event) != is_software_event(group_leader))) { | |
6536 | if (is_software_event(event)) { | |
6537 | /* | |
6538 | * If event and group_leader are not both a software | |
6539 | * event, and event is, then group leader is not. | |
6540 | * | |
6541 | * Allow the addition of software events to !software | |
6542 | * groups, this is safe because software events never | |
6543 | * fail to schedule. | |
6544 | */ | |
6545 | pmu = group_leader->pmu; | |
6546 | } else if (is_software_event(group_leader) && | |
6547 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
6548 | /* | |
6549 | * In case the group is a pure software group, and we | |
6550 | * try to add a hardware event, move the whole group to | |
6551 | * the hardware context. | |
6552 | */ | |
6553 | move_group = 1; | |
6554 | } | |
6555 | } | |
89a1e187 PZ |
6556 | |
6557 | /* | |
6558 | * Get the target context (task or percpu): | |
6559 | */ | |
e2d37cd2 | 6560 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
6561 | if (IS_ERR(ctx)) { |
6562 | err = PTR_ERR(ctx); | |
c6be5a5c | 6563 | goto err_alloc; |
89a1e187 PZ |
6564 | } |
6565 | ||
fd1edb3a PZ |
6566 | if (task) { |
6567 | put_task_struct(task); | |
6568 | task = NULL; | |
6569 | } | |
6570 | ||
ccff286d | 6571 | /* |
cdd6c482 | 6572 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 6573 | */ |
ac9721f3 | 6574 | if (group_leader) { |
dc86cabe | 6575 | err = -EINVAL; |
04289bb9 | 6576 | |
04289bb9 | 6577 | /* |
ccff286d IM |
6578 | * Do not allow a recursive hierarchy (this new sibling |
6579 | * becoming part of another group-sibling): | |
6580 | */ | |
6581 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 6582 | goto err_context; |
ccff286d IM |
6583 | /* |
6584 | * Do not allow to attach to a group in a different | |
6585 | * task or CPU context: | |
04289bb9 | 6586 | */ |
b04243ef PZ |
6587 | if (move_group) { |
6588 | if (group_leader->ctx->type != ctx->type) | |
6589 | goto err_context; | |
6590 | } else { | |
6591 | if (group_leader->ctx != ctx) | |
6592 | goto err_context; | |
6593 | } | |
6594 | ||
3b6f9e5c PM |
6595 | /* |
6596 | * Only a group leader can be exclusive or pinned | |
6597 | */ | |
0d48696f | 6598 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 6599 | goto err_context; |
ac9721f3 PZ |
6600 | } |
6601 | ||
6602 | if (output_event) { | |
6603 | err = perf_event_set_output(event, output_event); | |
6604 | if (err) | |
c3f00c70 | 6605 | goto err_context; |
ac9721f3 | 6606 | } |
0793a61d | 6607 | |
ea635c64 AV |
6608 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
6609 | if (IS_ERR(event_file)) { | |
6610 | err = PTR_ERR(event_file); | |
c3f00c70 | 6611 | goto err_context; |
ea635c64 | 6612 | } |
9b51f66d | 6613 | |
b04243ef PZ |
6614 | if (move_group) { |
6615 | struct perf_event_context *gctx = group_leader->ctx; | |
6616 | ||
6617 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 6618 | perf_remove_from_context(group_leader); |
b04243ef PZ |
6619 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6620 | group_entry) { | |
fe4b04fa | 6621 | perf_remove_from_context(sibling); |
b04243ef PZ |
6622 | put_ctx(gctx); |
6623 | } | |
6624 | mutex_unlock(&gctx->mutex); | |
6625 | put_ctx(gctx); | |
ea635c64 | 6626 | } |
9b51f66d | 6627 | |
ad3a37de | 6628 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 6629 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
6630 | |
6631 | if (move_group) { | |
0cda4c02 | 6632 | synchronize_rcu(); |
e2d37cd2 | 6633 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
6634 | get_ctx(ctx); |
6635 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
6636 | group_entry) { | |
e2d37cd2 | 6637 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
6638 | get_ctx(ctx); |
6639 | } | |
6640 | } | |
6641 | ||
e2d37cd2 | 6642 | perf_install_in_context(ctx, event, event->cpu); |
ad3a37de | 6643 | ++ctx->generation; |
fe4b04fa | 6644 | perf_unpin_context(ctx); |
d859e29f | 6645 | mutex_unlock(&ctx->mutex); |
9b51f66d | 6646 | |
fbfc623f YZ |
6647 | put_online_cpus(); |
6648 | ||
cdd6c482 | 6649 | event->owner = current; |
8882135b | 6650 | |
cdd6c482 IM |
6651 | mutex_lock(¤t->perf_event_mutex); |
6652 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
6653 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 6654 | |
c320c7b7 ACM |
6655 | /* |
6656 | * Precalculate sample_data sizes | |
6657 | */ | |
6658 | perf_event__header_size(event); | |
6844c09d | 6659 | perf_event__id_header_size(event); |
c320c7b7 | 6660 | |
8a49542c PZ |
6661 | /* |
6662 | * Drop the reference on the group_event after placing the | |
6663 | * new event on the sibling_list. This ensures destruction | |
6664 | * of the group leader will find the pointer to itself in | |
6665 | * perf_group_detach(). | |
6666 | */ | |
ea635c64 AV |
6667 | fput_light(group_file, fput_needed); |
6668 | fd_install(event_fd, event_file); | |
6669 | return event_fd; | |
0793a61d | 6670 | |
c3f00c70 | 6671 | err_context: |
fe4b04fa | 6672 | perf_unpin_context(ctx); |
ea635c64 | 6673 | put_ctx(ctx); |
c6be5a5c | 6674 | err_alloc: |
ea635c64 | 6675 | free_event(event); |
e7d0bc04 | 6676 | err_task: |
fbfc623f | 6677 | put_online_cpus(); |
e7d0bc04 PZ |
6678 | if (task) |
6679 | put_task_struct(task); | |
89a1e187 | 6680 | err_group_fd: |
dc86cabe | 6681 | fput_light(group_file, fput_needed); |
ea635c64 AV |
6682 | err_fd: |
6683 | put_unused_fd(event_fd); | |
dc86cabe | 6684 | return err; |
0793a61d TG |
6685 | } |
6686 | ||
fb0459d7 AV |
6687 | /** |
6688 | * perf_event_create_kernel_counter | |
6689 | * | |
6690 | * @attr: attributes of the counter to create | |
6691 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 6692 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
6693 | */ |
6694 | struct perf_event * | |
6695 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 6696 | struct task_struct *task, |
4dc0da86 AK |
6697 | perf_overflow_handler_t overflow_handler, |
6698 | void *context) | |
fb0459d7 | 6699 | { |
fb0459d7 | 6700 | struct perf_event_context *ctx; |
c3f00c70 | 6701 | struct perf_event *event; |
fb0459d7 | 6702 | int err; |
d859e29f | 6703 | |
fb0459d7 AV |
6704 | /* |
6705 | * Get the target context (task or percpu): | |
6706 | */ | |
d859e29f | 6707 | |
4dc0da86 AK |
6708 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
6709 | overflow_handler, context); | |
c3f00c70 PZ |
6710 | if (IS_ERR(event)) { |
6711 | err = PTR_ERR(event); | |
6712 | goto err; | |
6713 | } | |
d859e29f | 6714 | |
38a81da2 | 6715 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
6716 | if (IS_ERR(ctx)) { |
6717 | err = PTR_ERR(ctx); | |
c3f00c70 | 6718 | goto err_free; |
d859e29f | 6719 | } |
fb0459d7 | 6720 | |
fb0459d7 AV |
6721 | WARN_ON_ONCE(ctx->parent_ctx); |
6722 | mutex_lock(&ctx->mutex); | |
6723 | perf_install_in_context(ctx, event, cpu); | |
6724 | ++ctx->generation; | |
fe4b04fa | 6725 | perf_unpin_context(ctx); |
fb0459d7 AV |
6726 | mutex_unlock(&ctx->mutex); |
6727 | ||
fb0459d7 AV |
6728 | return event; |
6729 | ||
c3f00c70 PZ |
6730 | err_free: |
6731 | free_event(event); | |
6732 | err: | |
c6567f64 | 6733 | return ERR_PTR(err); |
9b51f66d | 6734 | } |
fb0459d7 | 6735 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 6736 | |
0cda4c02 YZ |
6737 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
6738 | { | |
6739 | struct perf_event_context *src_ctx; | |
6740 | struct perf_event_context *dst_ctx; | |
6741 | struct perf_event *event, *tmp; | |
6742 | LIST_HEAD(events); | |
6743 | ||
6744 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
6745 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
6746 | ||
6747 | mutex_lock(&src_ctx->mutex); | |
6748 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
6749 | event_entry) { | |
6750 | perf_remove_from_context(event); | |
6751 | put_ctx(src_ctx); | |
6752 | list_add(&event->event_entry, &events); | |
6753 | } | |
6754 | mutex_unlock(&src_ctx->mutex); | |
6755 | ||
6756 | synchronize_rcu(); | |
6757 | ||
6758 | mutex_lock(&dst_ctx->mutex); | |
6759 | list_for_each_entry_safe(event, tmp, &events, event_entry) { | |
6760 | list_del(&event->event_entry); | |
6761 | if (event->state >= PERF_EVENT_STATE_OFF) | |
6762 | event->state = PERF_EVENT_STATE_INACTIVE; | |
6763 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
6764 | get_ctx(dst_ctx); | |
6765 | } | |
6766 | mutex_unlock(&dst_ctx->mutex); | |
6767 | } | |
6768 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
6769 | ||
cdd6c482 | 6770 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 6771 | struct task_struct *child) |
d859e29f | 6772 | { |
cdd6c482 | 6773 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 6774 | u64 child_val; |
d859e29f | 6775 | |
cdd6c482 IM |
6776 | if (child_event->attr.inherit_stat) |
6777 | perf_event_read_event(child_event, child); | |
38b200d6 | 6778 | |
b5e58793 | 6779 | child_val = perf_event_count(child_event); |
d859e29f PM |
6780 | |
6781 | /* | |
6782 | * Add back the child's count to the parent's count: | |
6783 | */ | |
a6e6dea6 | 6784 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
6785 | atomic64_add(child_event->total_time_enabled, |
6786 | &parent_event->child_total_time_enabled); | |
6787 | atomic64_add(child_event->total_time_running, | |
6788 | &parent_event->child_total_time_running); | |
d859e29f PM |
6789 | |
6790 | /* | |
cdd6c482 | 6791 | * Remove this event from the parent's list |
d859e29f | 6792 | */ |
cdd6c482 IM |
6793 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
6794 | mutex_lock(&parent_event->child_mutex); | |
6795 | list_del_init(&child_event->child_list); | |
6796 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
6797 | |
6798 | /* | |
cdd6c482 | 6799 | * Release the parent event, if this was the last |
d859e29f PM |
6800 | * reference to it. |
6801 | */ | |
a6fa941d | 6802 | put_event(parent_event); |
d859e29f PM |
6803 | } |
6804 | ||
9b51f66d | 6805 | static void |
cdd6c482 IM |
6806 | __perf_event_exit_task(struct perf_event *child_event, |
6807 | struct perf_event_context *child_ctx, | |
38b200d6 | 6808 | struct task_struct *child) |
9b51f66d | 6809 | { |
38b435b1 PZ |
6810 | if (child_event->parent) { |
6811 | raw_spin_lock_irq(&child_ctx->lock); | |
6812 | perf_group_detach(child_event); | |
6813 | raw_spin_unlock_irq(&child_ctx->lock); | |
6814 | } | |
9b51f66d | 6815 | |
fe4b04fa | 6816 | perf_remove_from_context(child_event); |
0cc0c027 | 6817 | |
9b51f66d | 6818 | /* |
38b435b1 | 6819 | * It can happen that the parent exits first, and has events |
9b51f66d | 6820 | * that are still around due to the child reference. These |
38b435b1 | 6821 | * events need to be zapped. |
9b51f66d | 6822 | */ |
38b435b1 | 6823 | if (child_event->parent) { |
cdd6c482 IM |
6824 | sync_child_event(child_event, child); |
6825 | free_event(child_event); | |
4bcf349a | 6826 | } |
9b51f66d IM |
6827 | } |
6828 | ||
8dc85d54 | 6829 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 6830 | { |
cdd6c482 IM |
6831 | struct perf_event *child_event, *tmp; |
6832 | struct perf_event_context *child_ctx; | |
a63eaf34 | 6833 | unsigned long flags; |
9b51f66d | 6834 | |
8dc85d54 | 6835 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 6836 | perf_event_task(child, NULL, 0); |
9b51f66d | 6837 | return; |
9f498cc5 | 6838 | } |
9b51f66d | 6839 | |
a63eaf34 | 6840 | local_irq_save(flags); |
ad3a37de PM |
6841 | /* |
6842 | * We can't reschedule here because interrupts are disabled, | |
6843 | * and either child is current or it is a task that can't be | |
6844 | * scheduled, so we are now safe from rescheduling changing | |
6845 | * our context. | |
6846 | */ | |
806839b2 | 6847 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
6848 | |
6849 | /* | |
6850 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 6851 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
6852 | * incremented the context's refcount before we do put_ctx below. |
6853 | */ | |
e625cce1 | 6854 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 6855 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 6856 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
6857 | /* |
6858 | * If this context is a clone; unclone it so it can't get | |
6859 | * swapped to another process while we're removing all | |
cdd6c482 | 6860 | * the events from it. |
71a851b4 PZ |
6861 | */ |
6862 | unclone_ctx(child_ctx); | |
5e942bb3 | 6863 | update_context_time(child_ctx); |
e625cce1 | 6864 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
6865 | |
6866 | /* | |
cdd6c482 IM |
6867 | * Report the task dead after unscheduling the events so that we |
6868 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
6869 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 6870 | */ |
cdd6c482 | 6871 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 6872 | |
66fff224 PZ |
6873 | /* |
6874 | * We can recurse on the same lock type through: | |
6875 | * | |
cdd6c482 IM |
6876 | * __perf_event_exit_task() |
6877 | * sync_child_event() | |
a6fa941d AV |
6878 | * put_event() |
6879 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
6880 | * |
6881 | * But since its the parent context it won't be the same instance. | |
6882 | */ | |
a0507c84 | 6883 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 6884 | |
8bc20959 | 6885 | again: |
889ff015 FW |
6886 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
6887 | group_entry) | |
6888 | __perf_event_exit_task(child_event, child_ctx, child); | |
6889 | ||
6890 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 6891 | group_entry) |
cdd6c482 | 6892 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
6893 | |
6894 | /* | |
cdd6c482 | 6895 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
6896 | * its siblings to the list, but we obtained 'tmp' before that which |
6897 | * will still point to the list head terminating the iteration. | |
6898 | */ | |
889ff015 FW |
6899 | if (!list_empty(&child_ctx->pinned_groups) || |
6900 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 6901 | goto again; |
a63eaf34 PM |
6902 | |
6903 | mutex_unlock(&child_ctx->mutex); | |
6904 | ||
6905 | put_ctx(child_ctx); | |
9b51f66d IM |
6906 | } |
6907 | ||
8dc85d54 PZ |
6908 | /* |
6909 | * When a child task exits, feed back event values to parent events. | |
6910 | */ | |
6911 | void perf_event_exit_task(struct task_struct *child) | |
6912 | { | |
8882135b | 6913 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
6914 | int ctxn; |
6915 | ||
8882135b PZ |
6916 | mutex_lock(&child->perf_event_mutex); |
6917 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
6918 | owner_entry) { | |
6919 | list_del_init(&event->owner_entry); | |
6920 | ||
6921 | /* | |
6922 | * Ensure the list deletion is visible before we clear | |
6923 | * the owner, closes a race against perf_release() where | |
6924 | * we need to serialize on the owner->perf_event_mutex. | |
6925 | */ | |
6926 | smp_wmb(); | |
6927 | event->owner = NULL; | |
6928 | } | |
6929 | mutex_unlock(&child->perf_event_mutex); | |
6930 | ||
8dc85d54 PZ |
6931 | for_each_task_context_nr(ctxn) |
6932 | perf_event_exit_task_context(child, ctxn); | |
6933 | } | |
6934 | ||
889ff015 FW |
6935 | static void perf_free_event(struct perf_event *event, |
6936 | struct perf_event_context *ctx) | |
6937 | { | |
6938 | struct perf_event *parent = event->parent; | |
6939 | ||
6940 | if (WARN_ON_ONCE(!parent)) | |
6941 | return; | |
6942 | ||
6943 | mutex_lock(&parent->child_mutex); | |
6944 | list_del_init(&event->child_list); | |
6945 | mutex_unlock(&parent->child_mutex); | |
6946 | ||
a6fa941d | 6947 | put_event(parent); |
889ff015 | 6948 | |
8a49542c | 6949 | perf_group_detach(event); |
889ff015 FW |
6950 | list_del_event(event, ctx); |
6951 | free_event(event); | |
6952 | } | |
6953 | ||
bbbee908 PZ |
6954 | /* |
6955 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 6956 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 6957 | */ |
cdd6c482 | 6958 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 6959 | { |
8dc85d54 | 6960 | struct perf_event_context *ctx; |
cdd6c482 | 6961 | struct perf_event *event, *tmp; |
8dc85d54 | 6962 | int ctxn; |
bbbee908 | 6963 | |
8dc85d54 PZ |
6964 | for_each_task_context_nr(ctxn) { |
6965 | ctx = task->perf_event_ctxp[ctxn]; | |
6966 | if (!ctx) | |
6967 | continue; | |
bbbee908 | 6968 | |
8dc85d54 | 6969 | mutex_lock(&ctx->mutex); |
bbbee908 | 6970 | again: |
8dc85d54 PZ |
6971 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
6972 | group_entry) | |
6973 | perf_free_event(event, ctx); | |
bbbee908 | 6974 | |
8dc85d54 PZ |
6975 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
6976 | group_entry) | |
6977 | perf_free_event(event, ctx); | |
bbbee908 | 6978 | |
8dc85d54 PZ |
6979 | if (!list_empty(&ctx->pinned_groups) || |
6980 | !list_empty(&ctx->flexible_groups)) | |
6981 | goto again; | |
bbbee908 | 6982 | |
8dc85d54 | 6983 | mutex_unlock(&ctx->mutex); |
bbbee908 | 6984 | |
8dc85d54 PZ |
6985 | put_ctx(ctx); |
6986 | } | |
889ff015 FW |
6987 | } |
6988 | ||
4e231c79 PZ |
6989 | void perf_event_delayed_put(struct task_struct *task) |
6990 | { | |
6991 | int ctxn; | |
6992 | ||
6993 | for_each_task_context_nr(ctxn) | |
6994 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
6995 | } | |
6996 | ||
97dee4f3 PZ |
6997 | /* |
6998 | * inherit a event from parent task to child task: | |
6999 | */ | |
7000 | static struct perf_event * | |
7001 | inherit_event(struct perf_event *parent_event, | |
7002 | struct task_struct *parent, | |
7003 | struct perf_event_context *parent_ctx, | |
7004 | struct task_struct *child, | |
7005 | struct perf_event *group_leader, | |
7006 | struct perf_event_context *child_ctx) | |
7007 | { | |
7008 | struct perf_event *child_event; | |
cee010ec | 7009 | unsigned long flags; |
97dee4f3 PZ |
7010 | |
7011 | /* | |
7012 | * Instead of creating recursive hierarchies of events, | |
7013 | * we link inherited events back to the original parent, | |
7014 | * which has a filp for sure, which we use as the reference | |
7015 | * count: | |
7016 | */ | |
7017 | if (parent_event->parent) | |
7018 | parent_event = parent_event->parent; | |
7019 | ||
7020 | child_event = perf_event_alloc(&parent_event->attr, | |
7021 | parent_event->cpu, | |
d580ff86 | 7022 | child, |
97dee4f3 | 7023 | group_leader, parent_event, |
4dc0da86 | 7024 | NULL, NULL); |
97dee4f3 PZ |
7025 | if (IS_ERR(child_event)) |
7026 | return child_event; | |
a6fa941d AV |
7027 | |
7028 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7029 | free_event(child_event); | |
7030 | return NULL; | |
7031 | } | |
7032 | ||
97dee4f3 PZ |
7033 | get_ctx(child_ctx); |
7034 | ||
7035 | /* | |
7036 | * Make the child state follow the state of the parent event, | |
7037 | * not its attr.disabled bit. We hold the parent's mutex, | |
7038 | * so we won't race with perf_event_{en, dis}able_family. | |
7039 | */ | |
7040 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7041 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7042 | else | |
7043 | child_event->state = PERF_EVENT_STATE_OFF; | |
7044 | ||
7045 | if (parent_event->attr.freq) { | |
7046 | u64 sample_period = parent_event->hw.sample_period; | |
7047 | struct hw_perf_event *hwc = &child_event->hw; | |
7048 | ||
7049 | hwc->sample_period = sample_period; | |
7050 | hwc->last_period = sample_period; | |
7051 | ||
7052 | local64_set(&hwc->period_left, sample_period); | |
7053 | } | |
7054 | ||
7055 | child_event->ctx = child_ctx; | |
7056 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7057 | child_event->overflow_handler_context |
7058 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7059 | |
614b6780 TG |
7060 | /* |
7061 | * Precalculate sample_data sizes | |
7062 | */ | |
7063 | perf_event__header_size(child_event); | |
6844c09d | 7064 | perf_event__id_header_size(child_event); |
614b6780 | 7065 | |
97dee4f3 PZ |
7066 | /* |
7067 | * Link it up in the child's context: | |
7068 | */ | |
cee010ec | 7069 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7070 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7071 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7072 | |
97dee4f3 PZ |
7073 | /* |
7074 | * Link this into the parent event's child list | |
7075 | */ | |
7076 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7077 | mutex_lock(&parent_event->child_mutex); | |
7078 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7079 | mutex_unlock(&parent_event->child_mutex); | |
7080 | ||
7081 | return child_event; | |
7082 | } | |
7083 | ||
7084 | static int inherit_group(struct perf_event *parent_event, | |
7085 | struct task_struct *parent, | |
7086 | struct perf_event_context *parent_ctx, | |
7087 | struct task_struct *child, | |
7088 | struct perf_event_context *child_ctx) | |
7089 | { | |
7090 | struct perf_event *leader; | |
7091 | struct perf_event *sub; | |
7092 | struct perf_event *child_ctr; | |
7093 | ||
7094 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7095 | child, NULL, child_ctx); | |
7096 | if (IS_ERR(leader)) | |
7097 | return PTR_ERR(leader); | |
7098 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7099 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7100 | child, leader, child_ctx); | |
7101 | if (IS_ERR(child_ctr)) | |
7102 | return PTR_ERR(child_ctr); | |
7103 | } | |
7104 | return 0; | |
889ff015 FW |
7105 | } |
7106 | ||
7107 | static int | |
7108 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7109 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7110 | struct task_struct *child, int ctxn, |
889ff015 FW |
7111 | int *inherited_all) |
7112 | { | |
7113 | int ret; | |
8dc85d54 | 7114 | struct perf_event_context *child_ctx; |
889ff015 FW |
7115 | |
7116 | if (!event->attr.inherit) { | |
7117 | *inherited_all = 0; | |
7118 | return 0; | |
bbbee908 PZ |
7119 | } |
7120 | ||
fe4b04fa | 7121 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7122 | if (!child_ctx) { |
7123 | /* | |
7124 | * This is executed from the parent task context, so | |
7125 | * inherit events that have been marked for cloning. | |
7126 | * First allocate and initialize a context for the | |
7127 | * child. | |
7128 | */ | |
bbbee908 | 7129 | |
eb184479 | 7130 | child_ctx = alloc_perf_context(event->pmu, child); |
889ff015 FW |
7131 | if (!child_ctx) |
7132 | return -ENOMEM; | |
bbbee908 | 7133 | |
8dc85d54 | 7134 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7135 | } |
7136 | ||
7137 | ret = inherit_group(event, parent, parent_ctx, | |
7138 | child, child_ctx); | |
7139 | ||
7140 | if (ret) | |
7141 | *inherited_all = 0; | |
7142 | ||
7143 | return ret; | |
bbbee908 PZ |
7144 | } |
7145 | ||
9b51f66d | 7146 | /* |
cdd6c482 | 7147 | * Initialize the perf_event context in task_struct |
9b51f66d | 7148 | */ |
8dc85d54 | 7149 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7150 | { |
889ff015 | 7151 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7152 | struct perf_event_context *cloned_ctx; |
7153 | struct perf_event *event; | |
9b51f66d | 7154 | struct task_struct *parent = current; |
564c2b21 | 7155 | int inherited_all = 1; |
dddd3379 | 7156 | unsigned long flags; |
6ab423e0 | 7157 | int ret = 0; |
9b51f66d | 7158 | |
8dc85d54 | 7159 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7160 | return 0; |
7161 | ||
ad3a37de | 7162 | /* |
25346b93 PM |
7163 | * If the parent's context is a clone, pin it so it won't get |
7164 | * swapped under us. | |
ad3a37de | 7165 | */ |
8dc85d54 | 7166 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 7167 | |
ad3a37de PM |
7168 | /* |
7169 | * No need to check if parent_ctx != NULL here; since we saw | |
7170 | * it non-NULL earlier, the only reason for it to become NULL | |
7171 | * is if we exit, and since we're currently in the middle of | |
7172 | * a fork we can't be exiting at the same time. | |
7173 | */ | |
ad3a37de | 7174 | |
9b51f66d IM |
7175 | /* |
7176 | * Lock the parent list. No need to lock the child - not PID | |
7177 | * hashed yet and not running, so nobody can access it. | |
7178 | */ | |
d859e29f | 7179 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7180 | |
7181 | /* | |
7182 | * We dont have to disable NMIs - we are only looking at | |
7183 | * the list, not manipulating it: | |
7184 | */ | |
889ff015 | 7185 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7186 | ret = inherit_task_group(event, parent, parent_ctx, |
7187 | child, ctxn, &inherited_all); | |
889ff015 FW |
7188 | if (ret) |
7189 | break; | |
7190 | } | |
b93f7978 | 7191 | |
dddd3379 TG |
7192 | /* |
7193 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7194 | * to allocations, but we need to prevent rotation because | |
7195 | * rotate_ctx() will change the list from interrupt context. | |
7196 | */ | |
7197 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7198 | parent_ctx->rotate_disable = 1; | |
7199 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7200 | ||
889ff015 | 7201 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7202 | ret = inherit_task_group(event, parent, parent_ctx, |
7203 | child, ctxn, &inherited_all); | |
889ff015 | 7204 | if (ret) |
9b51f66d | 7205 | break; |
564c2b21 PM |
7206 | } |
7207 | ||
dddd3379 TG |
7208 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7209 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7210 | |
8dc85d54 | 7211 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7212 | |
05cbaa28 | 7213 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7214 | /* |
7215 | * Mark the child context as a clone of the parent | |
7216 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7217 | * |
7218 | * Note that if the parent is a clone, the holding of | |
7219 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7220 | */ |
c5ed5145 | 7221 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7222 | if (cloned_ctx) { |
7223 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7224 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7225 | } else { |
7226 | child_ctx->parent_ctx = parent_ctx; | |
7227 | child_ctx->parent_gen = parent_ctx->generation; | |
7228 | } | |
7229 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7230 | } |
7231 | ||
c5ed5145 | 7232 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7233 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7234 | |
25346b93 | 7235 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7236 | put_ctx(parent_ctx); |
ad3a37de | 7237 | |
6ab423e0 | 7238 | return ret; |
9b51f66d IM |
7239 | } |
7240 | ||
8dc85d54 PZ |
7241 | /* |
7242 | * Initialize the perf_event context in task_struct | |
7243 | */ | |
7244 | int perf_event_init_task(struct task_struct *child) | |
7245 | { | |
7246 | int ctxn, ret; | |
7247 | ||
8550d7cb ON |
7248 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7249 | mutex_init(&child->perf_event_mutex); | |
7250 | INIT_LIST_HEAD(&child->perf_event_list); | |
7251 | ||
8dc85d54 PZ |
7252 | for_each_task_context_nr(ctxn) { |
7253 | ret = perf_event_init_context(child, ctxn); | |
7254 | if (ret) | |
7255 | return ret; | |
7256 | } | |
7257 | ||
7258 | return 0; | |
7259 | } | |
7260 | ||
220b140b PM |
7261 | static void __init perf_event_init_all_cpus(void) |
7262 | { | |
b28ab83c | 7263 | struct swevent_htable *swhash; |
220b140b | 7264 | int cpu; |
220b140b PM |
7265 | |
7266 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7267 | swhash = &per_cpu(swevent_htable, cpu); |
7268 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7269 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7270 | } |
7271 | } | |
7272 | ||
cdd6c482 | 7273 | static void __cpuinit perf_event_init_cpu(int cpu) |
0793a61d | 7274 | { |
108b02cf | 7275 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7276 | |
b28ab83c | 7277 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7278 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7279 | struct swevent_hlist *hlist; |
7280 | ||
b28ab83c PZ |
7281 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7282 | WARN_ON(!hlist); | |
7283 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7284 | } |
b28ab83c | 7285 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7286 | } |
7287 | ||
c277443c | 7288 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7289 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7290 | { |
e9d2b064 PZ |
7291 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7292 | ||
7293 | WARN_ON(!irqs_disabled()); | |
7294 | ||
7295 | list_del_init(&cpuctx->rotation_list); | |
7296 | } | |
7297 | ||
108b02cf | 7298 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7299 | { |
108b02cf | 7300 | struct perf_event_context *ctx = __info; |
cdd6c482 | 7301 | struct perf_event *event, *tmp; |
0793a61d | 7302 | |
108b02cf | 7303 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7304 | |
889ff015 | 7305 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 7306 | __perf_remove_from_context(event); |
889ff015 | 7307 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 7308 | __perf_remove_from_context(event); |
0793a61d | 7309 | } |
108b02cf PZ |
7310 | |
7311 | static void perf_event_exit_cpu_context(int cpu) | |
7312 | { | |
7313 | struct perf_event_context *ctx; | |
7314 | struct pmu *pmu; | |
7315 | int idx; | |
7316 | ||
7317 | idx = srcu_read_lock(&pmus_srcu); | |
7318 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7319 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7320 | |
7321 | mutex_lock(&ctx->mutex); | |
7322 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7323 | mutex_unlock(&ctx->mutex); | |
7324 | } | |
7325 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7326 | } |
7327 | ||
cdd6c482 | 7328 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7329 | { |
b28ab83c | 7330 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7331 | |
b28ab83c PZ |
7332 | mutex_lock(&swhash->hlist_mutex); |
7333 | swevent_hlist_release(swhash); | |
7334 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 7335 | |
108b02cf | 7336 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7337 | } |
7338 | #else | |
cdd6c482 | 7339 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7340 | #endif |
7341 | ||
c277443c PZ |
7342 | static int |
7343 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7344 | { | |
7345 | int cpu; | |
7346 | ||
7347 | for_each_online_cpu(cpu) | |
7348 | perf_event_exit_cpu(cpu); | |
7349 | ||
7350 | return NOTIFY_OK; | |
7351 | } | |
7352 | ||
7353 | /* | |
7354 | * Run the perf reboot notifier at the very last possible moment so that | |
7355 | * the generic watchdog code runs as long as possible. | |
7356 | */ | |
7357 | static struct notifier_block perf_reboot_notifier = { | |
7358 | .notifier_call = perf_reboot, | |
7359 | .priority = INT_MIN, | |
7360 | }; | |
7361 | ||
0793a61d TG |
7362 | static int __cpuinit |
7363 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
7364 | { | |
7365 | unsigned int cpu = (long)hcpu; | |
7366 | ||
4536e4d1 | 7367 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7368 | |
7369 | case CPU_UP_PREPARE: | |
5e11637e | 7370 | case CPU_DOWN_FAILED: |
cdd6c482 | 7371 | perf_event_init_cpu(cpu); |
0793a61d TG |
7372 | break; |
7373 | ||
5e11637e | 7374 | case CPU_UP_CANCELED: |
0793a61d | 7375 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7376 | perf_event_exit_cpu(cpu); |
0793a61d TG |
7377 | break; |
7378 | ||
7379 | default: | |
7380 | break; | |
7381 | } | |
7382 | ||
7383 | return NOTIFY_OK; | |
7384 | } | |
7385 | ||
cdd6c482 | 7386 | void __init perf_event_init(void) |
0793a61d | 7387 | { |
3c502e7a JW |
7388 | int ret; |
7389 | ||
2e80a82a PZ |
7390 | idr_init(&pmu_idr); |
7391 | ||
220b140b | 7392 | perf_event_init_all_cpus(); |
b0a873eb | 7393 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7394 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7395 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7396 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7397 | perf_tp_register(); |
7398 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7399 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7400 | |
7401 | ret = init_hw_breakpoint(); | |
7402 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7403 | |
7404 | /* do not patch jump label more than once per second */ | |
7405 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7406 | |
7407 | /* | |
7408 | * Build time assertion that we keep the data_head at the intended | |
7409 | * location. IOW, validation we got the __reserved[] size right. | |
7410 | */ | |
7411 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
7412 | != 1024); | |
0793a61d | 7413 | } |
abe43400 PZ |
7414 | |
7415 | static int __init perf_event_sysfs_init(void) | |
7416 | { | |
7417 | struct pmu *pmu; | |
7418 | int ret; | |
7419 | ||
7420 | mutex_lock(&pmus_lock); | |
7421 | ||
7422 | ret = bus_register(&pmu_bus); | |
7423 | if (ret) | |
7424 | goto unlock; | |
7425 | ||
7426 | list_for_each_entry(pmu, &pmus, entry) { | |
7427 | if (!pmu->name || pmu->type < 0) | |
7428 | continue; | |
7429 | ||
7430 | ret = pmu_dev_alloc(pmu); | |
7431 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
7432 | } | |
7433 | pmu_bus_running = 1; | |
7434 | ret = 0; | |
7435 | ||
7436 | unlock: | |
7437 | mutex_unlock(&pmus_lock); | |
7438 | ||
7439 | return ret; | |
7440 | } | |
7441 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
7442 | |
7443 | #ifdef CONFIG_CGROUP_PERF | |
761b3ef5 | 7444 | static struct cgroup_subsys_state *perf_cgroup_create(struct cgroup *cont) |
e5d1367f SE |
7445 | { |
7446 | struct perf_cgroup *jc; | |
e5d1367f | 7447 | |
1b15d055 | 7448 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
7449 | if (!jc) |
7450 | return ERR_PTR(-ENOMEM); | |
7451 | ||
e5d1367f SE |
7452 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7453 | if (!jc->info) { | |
7454 | kfree(jc); | |
7455 | return ERR_PTR(-ENOMEM); | |
7456 | } | |
7457 | ||
e5d1367f SE |
7458 | return &jc->css; |
7459 | } | |
7460 | ||
761b3ef5 | 7461 | static void perf_cgroup_destroy(struct cgroup *cont) |
e5d1367f SE |
7462 | { |
7463 | struct perf_cgroup *jc; | |
7464 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), | |
7465 | struct perf_cgroup, css); | |
7466 | free_percpu(jc->info); | |
7467 | kfree(jc); | |
7468 | } | |
7469 | ||
7470 | static int __perf_cgroup_move(void *info) | |
7471 | { | |
7472 | struct task_struct *task = info; | |
7473 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
7474 | return 0; | |
7475 | } | |
7476 | ||
761b3ef5 | 7477 | static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
e5d1367f | 7478 | { |
bb9d97b6 TH |
7479 | struct task_struct *task; |
7480 | ||
7481 | cgroup_taskset_for_each(task, cgrp, tset) | |
7482 | task_function_call(task, __perf_cgroup_move, task); | |
e5d1367f SE |
7483 | } |
7484 | ||
761b3ef5 LZ |
7485 | static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, |
7486 | struct task_struct *task) | |
e5d1367f SE |
7487 | { |
7488 | /* | |
7489 | * cgroup_exit() is called in the copy_process() failure path. | |
7490 | * Ignore this case since the task hasn't ran yet, this avoids | |
7491 | * trying to poke a half freed task state from generic code. | |
7492 | */ | |
7493 | if (!(task->flags & PF_EXITING)) | |
7494 | return; | |
7495 | ||
bb9d97b6 | 7496 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7497 | } |
7498 | ||
7499 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
7500 | .name = "perf_event", |
7501 | .subsys_id = perf_subsys_id, | |
7502 | .create = perf_cgroup_create, | |
7503 | .destroy = perf_cgroup_destroy, | |
7504 | .exit = perf_cgroup_exit, | |
bb9d97b6 | 7505 | .attach = perf_cgroup_attach, |
e5d1367f SE |
7506 | }; |
7507 | #endif /* CONFIG_CGROUP_PERF */ |