]>
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> |
12351ef8 | 21 | #include <linux/tick.h> |
0793a61d | 22 | #include <linux/sysfs.h> |
22a4f650 | 23 | #include <linux/dcache.h> |
0793a61d | 24 | #include <linux/percpu.h> |
22a4f650 | 25 | #include <linux/ptrace.h> |
c277443c | 26 | #include <linux/reboot.h> |
b9cacc7b | 27 | #include <linux/vmstat.h> |
abe43400 | 28 | #include <linux/device.h> |
6e5fdeed | 29 | #include <linux/export.h> |
906010b2 | 30 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
31 | #include <linux/hardirq.h> |
32 | #include <linux/rculist.h> | |
0793a61d TG |
33 | #include <linux/uaccess.h> |
34 | #include <linux/syscalls.h> | |
35 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 36 | #include <linux/kernel_stat.h> |
cdd6c482 | 37 | #include <linux/perf_event.h> |
6fb2915d | 38 | #include <linux/ftrace_event.h> |
3c502e7a | 39 | #include <linux/hw_breakpoint.h> |
c5ebcedb | 40 | #include <linux/mm_types.h> |
877c6856 | 41 | #include <linux/cgroup.h> |
0793a61d | 42 | |
76369139 FW |
43 | #include "internal.h" |
44 | ||
4e193bd4 TB |
45 | #include <asm/irq_regs.h> |
46 | ||
fe4b04fa | 47 | struct remote_function_call { |
e7e7ee2e IM |
48 | struct task_struct *p; |
49 | int (*func)(void *info); | |
50 | void *info; | |
51 | int ret; | |
fe4b04fa PZ |
52 | }; |
53 | ||
54 | static void remote_function(void *data) | |
55 | { | |
56 | struct remote_function_call *tfc = data; | |
57 | struct task_struct *p = tfc->p; | |
58 | ||
59 | if (p) { | |
60 | tfc->ret = -EAGAIN; | |
61 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
62 | return; | |
63 | } | |
64 | ||
65 | tfc->ret = tfc->func(tfc->info); | |
66 | } | |
67 | ||
68 | /** | |
69 | * task_function_call - call a function on the cpu on which a task runs | |
70 | * @p: the task to evaluate | |
71 | * @func: the function to be called | |
72 | * @info: the function call argument | |
73 | * | |
74 | * Calls the function @func when the task is currently running. This might | |
75 | * be on the current CPU, which just calls the function directly | |
76 | * | |
77 | * returns: @func return value, or | |
78 | * -ESRCH - when the process isn't running | |
79 | * -EAGAIN - when the process moved away | |
80 | */ | |
81 | static int | |
82 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
83 | { | |
84 | struct remote_function_call data = { | |
e7e7ee2e IM |
85 | .p = p, |
86 | .func = func, | |
87 | .info = info, | |
88 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
89 | }; |
90 | ||
91 | if (task_curr(p)) | |
92 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
93 | ||
94 | return data.ret; | |
95 | } | |
96 | ||
97 | /** | |
98 | * cpu_function_call - call a function on the cpu | |
99 | * @func: the function to be called | |
100 | * @info: the function call argument | |
101 | * | |
102 | * Calls the function @func on the remote cpu. | |
103 | * | |
104 | * returns: @func return value or -ENXIO when the cpu is offline | |
105 | */ | |
106 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
107 | { | |
108 | struct remote_function_call data = { | |
e7e7ee2e IM |
109 | .p = NULL, |
110 | .func = func, | |
111 | .info = info, | |
112 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
113 | }; |
114 | ||
115 | smp_call_function_single(cpu, remote_function, &data, 1); | |
116 | ||
117 | return data.ret; | |
118 | } | |
119 | ||
e5d1367f SE |
120 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
121 | PERF_FLAG_FD_OUTPUT |\ | |
122 | PERF_FLAG_PID_CGROUP) | |
123 | ||
bce38cd5 SE |
124 | /* |
125 | * branch priv levels that need permission checks | |
126 | */ | |
127 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
128 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
129 | PERF_SAMPLE_BRANCH_HV) | |
130 | ||
0b3fcf17 SE |
131 | enum event_type_t { |
132 | EVENT_FLEXIBLE = 0x1, | |
133 | EVENT_PINNED = 0x2, | |
134 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
135 | }; | |
136 | ||
e5d1367f SE |
137 | /* |
138 | * perf_sched_events : >0 events exist | |
139 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
140 | */ | |
c5905afb | 141 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 142 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 143 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 144 | |
cdd6c482 IM |
145 | static atomic_t nr_mmap_events __read_mostly; |
146 | static atomic_t nr_comm_events __read_mostly; | |
147 | static atomic_t nr_task_events __read_mostly; | |
9ee318a7 | 148 | |
108b02cf PZ |
149 | static LIST_HEAD(pmus); |
150 | static DEFINE_MUTEX(pmus_lock); | |
151 | static struct srcu_struct pmus_srcu; | |
152 | ||
0764771d | 153 | /* |
cdd6c482 | 154 | * perf event paranoia level: |
0fbdea19 IM |
155 | * -1 - not paranoid at all |
156 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 157 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 158 | * 2 - disallow kernel profiling for unpriv |
0764771d | 159 | */ |
cdd6c482 | 160 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 161 | |
20443384 FW |
162 | /* Minimum for 512 kiB + 1 user control page */ |
163 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
164 | |
165 | /* | |
cdd6c482 | 166 | * max perf event sample rate |
df58ab24 | 167 | */ |
163ec435 PZ |
168 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
169 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
170 | static int max_samples_per_tick __read_mostly = | |
171 | DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
172 | ||
9e630205 SE |
173 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
174 | ||
163ec435 PZ |
175 | int perf_proc_update_handler(struct ctl_table *table, int write, |
176 | void __user *buffer, size_t *lenp, | |
177 | loff_t *ppos) | |
178 | { | |
179 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
180 | ||
181 | if (ret || !write) | |
182 | return ret; | |
183 | ||
184 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
185 | ||
186 | return 0; | |
187 | } | |
1ccd1549 | 188 | |
cdd6c482 | 189 | static atomic64_t perf_event_id; |
a96bbc16 | 190 | |
0b3fcf17 SE |
191 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
192 | enum event_type_t event_type); | |
193 | ||
194 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
195 | enum event_type_t event_type, |
196 | struct task_struct *task); | |
197 | ||
198 | static void update_context_time(struct perf_event_context *ctx); | |
199 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 200 | |
10c6db11 PZ |
201 | static void ring_buffer_attach(struct perf_event *event, |
202 | struct ring_buffer *rb); | |
203 | ||
cdd6c482 | 204 | void __weak perf_event_print_debug(void) { } |
0793a61d | 205 | |
84c79910 | 206 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 207 | { |
84c79910 | 208 | return "pmu"; |
0793a61d TG |
209 | } |
210 | ||
0b3fcf17 SE |
211 | static inline u64 perf_clock(void) |
212 | { | |
213 | return local_clock(); | |
214 | } | |
215 | ||
e5d1367f SE |
216 | static inline struct perf_cpu_context * |
217 | __get_cpu_context(struct perf_event_context *ctx) | |
218 | { | |
219 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
220 | } | |
221 | ||
facc4307 PZ |
222 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
223 | struct perf_event_context *ctx) | |
224 | { | |
225 | raw_spin_lock(&cpuctx->ctx.lock); | |
226 | if (ctx) | |
227 | raw_spin_lock(&ctx->lock); | |
228 | } | |
229 | ||
230 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
231 | struct perf_event_context *ctx) | |
232 | { | |
233 | if (ctx) | |
234 | raw_spin_unlock(&ctx->lock); | |
235 | raw_spin_unlock(&cpuctx->ctx.lock); | |
236 | } | |
237 | ||
e5d1367f SE |
238 | #ifdef CONFIG_CGROUP_PERF |
239 | ||
877c6856 LZ |
240 | /* |
241 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
242 | * This is a per-cpu dynamically allocated data structure. | |
243 | */ | |
244 | struct perf_cgroup_info { | |
245 | u64 time; | |
246 | u64 timestamp; | |
247 | }; | |
248 | ||
249 | struct perf_cgroup { | |
250 | struct cgroup_subsys_state css; | |
86e213e1 | 251 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
252 | }; |
253 | ||
3f7cce3c SE |
254 | /* |
255 | * Must ensure cgroup is pinned (css_get) before calling | |
256 | * this function. In other words, we cannot call this function | |
257 | * if there is no cgroup event for the current CPU context. | |
258 | */ | |
e5d1367f SE |
259 | static inline struct perf_cgroup * |
260 | perf_cgroup_from_task(struct task_struct *task) | |
261 | { | |
262 | return container_of(task_subsys_state(task, perf_subsys_id), | |
263 | struct perf_cgroup, css); | |
264 | } | |
265 | ||
266 | static inline bool | |
267 | perf_cgroup_match(struct perf_event *event) | |
268 | { | |
269 | struct perf_event_context *ctx = event->ctx; | |
270 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
271 | ||
ef824fa1 TH |
272 | /* @event doesn't care about cgroup */ |
273 | if (!event->cgrp) | |
274 | return true; | |
275 | ||
276 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
277 | if (!cpuctx->cgrp) | |
278 | return false; | |
279 | ||
280 | /* | |
281 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
282 | * also enabled for all its descendant cgroups. If @cpuctx's | |
283 | * cgroup is a descendant of @event's (the test covers identity | |
284 | * case), it's a match. | |
285 | */ | |
286 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
287 | event->cgrp->css.cgroup); | |
e5d1367f SE |
288 | } |
289 | ||
9c5da09d | 290 | static inline bool perf_tryget_cgroup(struct perf_event *event) |
e5d1367f | 291 | { |
9c5da09d | 292 | return css_tryget(&event->cgrp->css); |
e5d1367f SE |
293 | } |
294 | ||
295 | static inline void perf_put_cgroup(struct perf_event *event) | |
296 | { | |
297 | css_put(&event->cgrp->css); | |
298 | } | |
299 | ||
300 | static inline void perf_detach_cgroup(struct perf_event *event) | |
301 | { | |
302 | perf_put_cgroup(event); | |
303 | event->cgrp = NULL; | |
304 | } | |
305 | ||
306 | static inline int is_cgroup_event(struct perf_event *event) | |
307 | { | |
308 | return event->cgrp != NULL; | |
309 | } | |
310 | ||
311 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
312 | { | |
313 | struct perf_cgroup_info *t; | |
314 | ||
315 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
316 | return t->time; | |
317 | } | |
318 | ||
319 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
320 | { | |
321 | struct perf_cgroup_info *info; | |
322 | u64 now; | |
323 | ||
324 | now = perf_clock(); | |
325 | ||
326 | info = this_cpu_ptr(cgrp->info); | |
327 | ||
328 | info->time += now - info->timestamp; | |
329 | info->timestamp = now; | |
330 | } | |
331 | ||
332 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
333 | { | |
334 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
335 | if (cgrp_out) | |
336 | __update_cgrp_time(cgrp_out); | |
337 | } | |
338 | ||
339 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
340 | { | |
3f7cce3c SE |
341 | struct perf_cgroup *cgrp; |
342 | ||
e5d1367f | 343 | /* |
3f7cce3c SE |
344 | * ensure we access cgroup data only when needed and |
345 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 346 | */ |
3f7cce3c | 347 | if (!is_cgroup_event(event)) |
e5d1367f SE |
348 | return; |
349 | ||
3f7cce3c SE |
350 | cgrp = perf_cgroup_from_task(current); |
351 | /* | |
352 | * Do not update time when cgroup is not active | |
353 | */ | |
354 | if (cgrp == event->cgrp) | |
355 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
356 | } |
357 | ||
358 | static inline void | |
3f7cce3c SE |
359 | perf_cgroup_set_timestamp(struct task_struct *task, |
360 | struct perf_event_context *ctx) | |
e5d1367f SE |
361 | { |
362 | struct perf_cgroup *cgrp; | |
363 | struct perf_cgroup_info *info; | |
364 | ||
3f7cce3c SE |
365 | /* |
366 | * ctx->lock held by caller | |
367 | * ensure we do not access cgroup data | |
368 | * unless we have the cgroup pinned (css_get) | |
369 | */ | |
370 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
371 | return; |
372 | ||
373 | cgrp = perf_cgroup_from_task(task); | |
374 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 375 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
376 | } |
377 | ||
378 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
379 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
380 | ||
381 | /* | |
382 | * reschedule events based on the cgroup constraint of task. | |
383 | * | |
384 | * mode SWOUT : schedule out everything | |
385 | * mode SWIN : schedule in based on cgroup for next | |
386 | */ | |
387 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
388 | { | |
389 | struct perf_cpu_context *cpuctx; | |
390 | struct pmu *pmu; | |
391 | unsigned long flags; | |
392 | ||
393 | /* | |
394 | * disable interrupts to avoid geting nr_cgroup | |
395 | * changes via __perf_event_disable(). Also | |
396 | * avoids preemption. | |
397 | */ | |
398 | local_irq_save(flags); | |
399 | ||
400 | /* | |
401 | * we reschedule only in the presence of cgroup | |
402 | * constrained events. | |
403 | */ | |
404 | rcu_read_lock(); | |
405 | ||
406 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 407 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
408 | if (cpuctx->unique_pmu != pmu) |
409 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 410 | |
e5d1367f SE |
411 | /* |
412 | * perf_cgroup_events says at least one | |
413 | * context on this CPU has cgroup events. | |
414 | * | |
415 | * ctx->nr_cgroups reports the number of cgroup | |
416 | * events for a context. | |
417 | */ | |
418 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
419 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
420 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
421 | |
422 | if (mode & PERF_CGROUP_SWOUT) { | |
423 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
424 | /* | |
425 | * must not be done before ctxswout due | |
426 | * to event_filter_match() in event_sched_out() | |
427 | */ | |
428 | cpuctx->cgrp = NULL; | |
429 | } | |
430 | ||
431 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 432 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
433 | /* |
434 | * set cgrp before ctxsw in to allow | |
435 | * event_filter_match() to not have to pass | |
436 | * task around | |
e5d1367f SE |
437 | */ |
438 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
439 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
440 | } | |
facc4307 PZ |
441 | perf_pmu_enable(cpuctx->ctx.pmu); |
442 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 443 | } |
e5d1367f SE |
444 | } |
445 | ||
446 | rcu_read_unlock(); | |
447 | ||
448 | local_irq_restore(flags); | |
449 | } | |
450 | ||
a8d757ef SE |
451 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
452 | struct task_struct *next) | |
e5d1367f | 453 | { |
a8d757ef SE |
454 | struct perf_cgroup *cgrp1; |
455 | struct perf_cgroup *cgrp2 = NULL; | |
456 | ||
457 | /* | |
458 | * we come here when we know perf_cgroup_events > 0 | |
459 | */ | |
460 | cgrp1 = perf_cgroup_from_task(task); | |
461 | ||
462 | /* | |
463 | * next is NULL when called from perf_event_enable_on_exec() | |
464 | * that will systematically cause a cgroup_switch() | |
465 | */ | |
466 | if (next) | |
467 | cgrp2 = perf_cgroup_from_task(next); | |
468 | ||
469 | /* | |
470 | * only schedule out current cgroup events if we know | |
471 | * that we are switching to a different cgroup. Otherwise, | |
472 | * do no touch the cgroup events. | |
473 | */ | |
474 | if (cgrp1 != cgrp2) | |
475 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
476 | } |
477 | ||
a8d757ef SE |
478 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
479 | struct task_struct *task) | |
e5d1367f | 480 | { |
a8d757ef SE |
481 | struct perf_cgroup *cgrp1; |
482 | struct perf_cgroup *cgrp2 = NULL; | |
483 | ||
484 | /* | |
485 | * we come here when we know perf_cgroup_events > 0 | |
486 | */ | |
487 | cgrp1 = perf_cgroup_from_task(task); | |
488 | ||
489 | /* prev can never be NULL */ | |
490 | cgrp2 = perf_cgroup_from_task(prev); | |
491 | ||
492 | /* | |
493 | * only need to schedule in cgroup events if we are changing | |
494 | * cgroup during ctxsw. Cgroup events were not scheduled | |
495 | * out of ctxsw out if that was not the case. | |
496 | */ | |
497 | if (cgrp1 != cgrp2) | |
498 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
499 | } |
500 | ||
501 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
502 | struct perf_event_attr *attr, | |
503 | struct perf_event *group_leader) | |
504 | { | |
505 | struct perf_cgroup *cgrp; | |
506 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
507 | struct fd f = fdget(fd); |
508 | int ret = 0; | |
e5d1367f | 509 | |
2903ff01 | 510 | if (!f.file) |
e5d1367f SE |
511 | return -EBADF; |
512 | ||
2903ff01 | 513 | css = cgroup_css_from_dir(f.file, perf_subsys_id); |
3db272c0 LZ |
514 | if (IS_ERR(css)) { |
515 | ret = PTR_ERR(css); | |
516 | goto out; | |
517 | } | |
e5d1367f SE |
518 | |
519 | cgrp = container_of(css, struct perf_cgroup, css); | |
520 | event->cgrp = cgrp; | |
521 | ||
f75e18cb | 522 | /* must be done before we fput() the file */ |
9c5da09d SQ |
523 | if (!perf_tryget_cgroup(event)) { |
524 | event->cgrp = NULL; | |
525 | ret = -ENOENT; | |
526 | goto out; | |
527 | } | |
f75e18cb | 528 | |
e5d1367f SE |
529 | /* |
530 | * all events in a group must monitor | |
531 | * the same cgroup because a task belongs | |
532 | * to only one perf cgroup at a time | |
533 | */ | |
534 | if (group_leader && group_leader->cgrp != cgrp) { | |
535 | perf_detach_cgroup(event); | |
536 | ret = -EINVAL; | |
e5d1367f | 537 | } |
3db272c0 | 538 | out: |
2903ff01 | 539 | fdput(f); |
e5d1367f SE |
540 | return ret; |
541 | } | |
542 | ||
543 | static inline void | |
544 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
545 | { | |
546 | struct perf_cgroup_info *t; | |
547 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
548 | event->shadow_ctx_time = now - t->timestamp; | |
549 | } | |
550 | ||
551 | static inline void | |
552 | perf_cgroup_defer_enabled(struct perf_event *event) | |
553 | { | |
554 | /* | |
555 | * when the current task's perf cgroup does not match | |
556 | * the event's, we need to remember to call the | |
557 | * perf_mark_enable() function the first time a task with | |
558 | * a matching perf cgroup is scheduled in. | |
559 | */ | |
560 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
561 | event->cgrp_defer_enabled = 1; | |
562 | } | |
563 | ||
564 | static inline void | |
565 | perf_cgroup_mark_enabled(struct perf_event *event, | |
566 | struct perf_event_context *ctx) | |
567 | { | |
568 | struct perf_event *sub; | |
569 | u64 tstamp = perf_event_time(event); | |
570 | ||
571 | if (!event->cgrp_defer_enabled) | |
572 | return; | |
573 | ||
574 | event->cgrp_defer_enabled = 0; | |
575 | ||
576 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
577 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
578 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
579 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
580 | sub->cgrp_defer_enabled = 0; | |
581 | } | |
582 | } | |
583 | } | |
584 | #else /* !CONFIG_CGROUP_PERF */ | |
585 | ||
586 | static inline bool | |
587 | perf_cgroup_match(struct perf_event *event) | |
588 | { | |
589 | return true; | |
590 | } | |
591 | ||
592 | static inline void perf_detach_cgroup(struct perf_event *event) | |
593 | {} | |
594 | ||
595 | static inline int is_cgroup_event(struct perf_event *event) | |
596 | { | |
597 | return 0; | |
598 | } | |
599 | ||
600 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
601 | { | |
602 | return 0; | |
603 | } | |
604 | ||
605 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
606 | { | |
607 | } | |
608 | ||
609 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
610 | { | |
611 | } | |
612 | ||
a8d757ef SE |
613 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
614 | struct task_struct *next) | |
e5d1367f SE |
615 | { |
616 | } | |
617 | ||
a8d757ef SE |
618 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
619 | struct task_struct *task) | |
e5d1367f SE |
620 | { |
621 | } | |
622 | ||
623 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
624 | struct perf_event_attr *attr, | |
625 | struct perf_event *group_leader) | |
626 | { | |
627 | return -EINVAL; | |
628 | } | |
629 | ||
630 | static inline void | |
3f7cce3c SE |
631 | perf_cgroup_set_timestamp(struct task_struct *task, |
632 | struct perf_event_context *ctx) | |
e5d1367f SE |
633 | { |
634 | } | |
635 | ||
636 | void | |
637 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
638 | { | |
639 | } | |
640 | ||
641 | static inline void | |
642 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
643 | { | |
644 | } | |
645 | ||
646 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
647 | { | |
648 | return 0; | |
649 | } | |
650 | ||
651 | static inline void | |
652 | perf_cgroup_defer_enabled(struct perf_event *event) | |
653 | { | |
654 | } | |
655 | ||
656 | static inline void | |
657 | perf_cgroup_mark_enabled(struct perf_event *event, | |
658 | struct perf_event_context *ctx) | |
659 | { | |
660 | } | |
661 | #endif | |
662 | ||
9e630205 SE |
663 | /* |
664 | * set default to be dependent on timer tick just | |
665 | * like original code | |
666 | */ | |
667 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
668 | /* | |
669 | * function must be called with interrupts disbled | |
670 | */ | |
671 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
672 | { | |
673 | struct perf_cpu_context *cpuctx; | |
674 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
675 | int rotations = 0; | |
676 | ||
677 | WARN_ON(!irqs_disabled()); | |
678 | ||
679 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
680 | ||
681 | rotations = perf_rotate_context(cpuctx); | |
682 | ||
683 | /* | |
684 | * arm timer if needed | |
685 | */ | |
686 | if (rotations) { | |
687 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
688 | ret = HRTIMER_RESTART; | |
689 | } | |
690 | ||
691 | return ret; | |
692 | } | |
693 | ||
694 | /* CPU is going down */ | |
695 | void perf_cpu_hrtimer_cancel(int cpu) | |
696 | { | |
697 | struct perf_cpu_context *cpuctx; | |
698 | struct pmu *pmu; | |
699 | unsigned long flags; | |
700 | ||
701 | if (WARN_ON(cpu != smp_processor_id())) | |
702 | return; | |
703 | ||
704 | local_irq_save(flags); | |
705 | ||
706 | rcu_read_lock(); | |
707 | ||
708 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
709 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
710 | ||
711 | if (pmu->task_ctx_nr == perf_sw_context) | |
712 | continue; | |
713 | ||
714 | hrtimer_cancel(&cpuctx->hrtimer); | |
715 | } | |
716 | ||
717 | rcu_read_unlock(); | |
718 | ||
719 | local_irq_restore(flags); | |
720 | } | |
721 | ||
722 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
723 | { | |
724 | struct hrtimer *hr = &cpuctx->hrtimer; | |
725 | struct pmu *pmu = cpuctx->ctx.pmu; | |
726 | ||
727 | /* no multiplexing needed for SW PMU */ | |
728 | if (pmu->task_ctx_nr == perf_sw_context) | |
729 | return; | |
730 | ||
731 | cpuctx->hrtimer_interval = | |
732 | ns_to_ktime(NSEC_PER_MSEC * PERF_CPU_HRTIMER); | |
733 | ||
734 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
735 | hr->function = perf_cpu_hrtimer_handler; | |
736 | } | |
737 | ||
738 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
739 | { | |
740 | struct hrtimer *hr = &cpuctx->hrtimer; | |
741 | struct pmu *pmu = cpuctx->ctx.pmu; | |
742 | ||
743 | /* not for SW PMU */ | |
744 | if (pmu->task_ctx_nr == perf_sw_context) | |
745 | return; | |
746 | ||
747 | if (hrtimer_active(hr)) | |
748 | return; | |
749 | ||
750 | if (!hrtimer_callback_running(hr)) | |
751 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
752 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
753 | } | |
754 | ||
33696fc0 | 755 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 756 | { |
33696fc0 PZ |
757 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
758 | if (!(*count)++) | |
759 | pmu->pmu_disable(pmu); | |
9e35ad38 | 760 | } |
9e35ad38 | 761 | |
33696fc0 | 762 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 763 | { |
33696fc0 PZ |
764 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
765 | if (!--(*count)) | |
766 | pmu->pmu_enable(pmu); | |
9e35ad38 | 767 | } |
9e35ad38 | 768 | |
e9d2b064 PZ |
769 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
770 | ||
771 | /* | |
772 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
773 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
774 | * disabled, while rotate_context is called from IRQ context. | |
775 | */ | |
108b02cf | 776 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 777 | { |
108b02cf | 778 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 779 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 780 | |
e9d2b064 | 781 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 782 | |
12351ef8 FW |
783 | if (list_empty(&cpuctx->rotation_list)) { |
784 | int was_empty = list_empty(head); | |
e9d2b064 | 785 | list_add(&cpuctx->rotation_list, head); |
12351ef8 FW |
786 | if (was_empty) |
787 | tick_nohz_full_kick(); | |
788 | } | |
9e35ad38 | 789 | } |
9e35ad38 | 790 | |
cdd6c482 | 791 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 792 | { |
e5289d4a | 793 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
794 | } |
795 | ||
cdd6c482 | 796 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 797 | { |
564c2b21 PM |
798 | if (atomic_dec_and_test(&ctx->refcount)) { |
799 | if (ctx->parent_ctx) | |
800 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
801 | if (ctx->task) |
802 | put_task_struct(ctx->task); | |
cb796ff3 | 803 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 804 | } |
a63eaf34 PM |
805 | } |
806 | ||
cdd6c482 | 807 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
808 | { |
809 | if (ctx->parent_ctx) { | |
810 | put_ctx(ctx->parent_ctx); | |
811 | ctx->parent_ctx = NULL; | |
812 | } | |
813 | } | |
814 | ||
6844c09d ACM |
815 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
816 | { | |
817 | /* | |
818 | * only top level events have the pid namespace they were created in | |
819 | */ | |
820 | if (event->parent) | |
821 | event = event->parent; | |
822 | ||
823 | return task_tgid_nr_ns(p, event->ns); | |
824 | } | |
825 | ||
826 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
827 | { | |
828 | /* | |
829 | * only top level events have the pid namespace they were created in | |
830 | */ | |
831 | if (event->parent) | |
832 | event = event->parent; | |
833 | ||
834 | return task_pid_nr_ns(p, event->ns); | |
835 | } | |
836 | ||
7f453c24 | 837 | /* |
cdd6c482 | 838 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
839 | * to userspace. |
840 | */ | |
cdd6c482 | 841 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 842 | { |
cdd6c482 | 843 | u64 id = event->id; |
7f453c24 | 844 | |
cdd6c482 IM |
845 | if (event->parent) |
846 | id = event->parent->id; | |
7f453c24 PZ |
847 | |
848 | return id; | |
849 | } | |
850 | ||
25346b93 | 851 | /* |
cdd6c482 | 852 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
853 | * This has to cope with with the fact that until it is locked, |
854 | * the context could get moved to another task. | |
855 | */ | |
cdd6c482 | 856 | static struct perf_event_context * |
8dc85d54 | 857 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 858 | { |
cdd6c482 | 859 | struct perf_event_context *ctx; |
25346b93 PM |
860 | |
861 | rcu_read_lock(); | |
9ed6060d | 862 | retry: |
8dc85d54 | 863 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
864 | if (ctx) { |
865 | /* | |
866 | * If this context is a clone of another, it might | |
867 | * get swapped for another underneath us by | |
cdd6c482 | 868 | * perf_event_task_sched_out, though the |
25346b93 PM |
869 | * rcu_read_lock() protects us from any context |
870 | * getting freed. Lock the context and check if it | |
871 | * got swapped before we could get the lock, and retry | |
872 | * if so. If we locked the right context, then it | |
873 | * can't get swapped on us any more. | |
874 | */ | |
e625cce1 | 875 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 876 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 877 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
25346b93 PM |
878 | goto retry; |
879 | } | |
b49a9e7e PZ |
880 | |
881 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 882 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
883 | ctx = NULL; |
884 | } | |
25346b93 PM |
885 | } |
886 | rcu_read_unlock(); | |
887 | return ctx; | |
888 | } | |
889 | ||
890 | /* | |
891 | * Get the context for a task and increment its pin_count so it | |
892 | * can't get swapped to another task. This also increments its | |
893 | * reference count so that the context can't get freed. | |
894 | */ | |
8dc85d54 PZ |
895 | static struct perf_event_context * |
896 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 897 | { |
cdd6c482 | 898 | struct perf_event_context *ctx; |
25346b93 PM |
899 | unsigned long flags; |
900 | ||
8dc85d54 | 901 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
902 | if (ctx) { |
903 | ++ctx->pin_count; | |
e625cce1 | 904 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
905 | } |
906 | return ctx; | |
907 | } | |
908 | ||
cdd6c482 | 909 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
910 | { |
911 | unsigned long flags; | |
912 | ||
e625cce1 | 913 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 914 | --ctx->pin_count; |
e625cce1 | 915 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
916 | } |
917 | ||
f67218c3 PZ |
918 | /* |
919 | * Update the record of the current time in a context. | |
920 | */ | |
921 | static void update_context_time(struct perf_event_context *ctx) | |
922 | { | |
923 | u64 now = perf_clock(); | |
924 | ||
925 | ctx->time += now - ctx->timestamp; | |
926 | ctx->timestamp = now; | |
927 | } | |
928 | ||
4158755d SE |
929 | static u64 perf_event_time(struct perf_event *event) |
930 | { | |
931 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
932 | |
933 | if (is_cgroup_event(event)) | |
934 | return perf_cgroup_event_time(event); | |
935 | ||
4158755d SE |
936 | return ctx ? ctx->time : 0; |
937 | } | |
938 | ||
f67218c3 PZ |
939 | /* |
940 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 941 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
942 | */ |
943 | static void update_event_times(struct perf_event *event) | |
944 | { | |
945 | struct perf_event_context *ctx = event->ctx; | |
946 | u64 run_end; | |
947 | ||
948 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
949 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
950 | return; | |
e5d1367f SE |
951 | /* |
952 | * in cgroup mode, time_enabled represents | |
953 | * the time the event was enabled AND active | |
954 | * tasks were in the monitored cgroup. This is | |
955 | * independent of the activity of the context as | |
956 | * there may be a mix of cgroup and non-cgroup events. | |
957 | * | |
958 | * That is why we treat cgroup events differently | |
959 | * here. | |
960 | */ | |
961 | if (is_cgroup_event(event)) | |
46cd6a7f | 962 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
963 | else if (ctx->is_active) |
964 | run_end = ctx->time; | |
acd1d7c1 PZ |
965 | else |
966 | run_end = event->tstamp_stopped; | |
967 | ||
968 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
969 | |
970 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
971 | run_end = event->tstamp_stopped; | |
972 | else | |
4158755d | 973 | run_end = perf_event_time(event); |
f67218c3 PZ |
974 | |
975 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 976 | |
f67218c3 PZ |
977 | } |
978 | ||
96c21a46 PZ |
979 | /* |
980 | * Update total_time_enabled and total_time_running for all events in a group. | |
981 | */ | |
982 | static void update_group_times(struct perf_event *leader) | |
983 | { | |
984 | struct perf_event *event; | |
985 | ||
986 | update_event_times(leader); | |
987 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
988 | update_event_times(event); | |
989 | } | |
990 | ||
889ff015 FW |
991 | static struct list_head * |
992 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
993 | { | |
994 | if (event->attr.pinned) | |
995 | return &ctx->pinned_groups; | |
996 | else | |
997 | return &ctx->flexible_groups; | |
998 | } | |
999 | ||
fccc714b | 1000 | /* |
cdd6c482 | 1001 | * Add a event from the lists for its context. |
fccc714b PZ |
1002 | * Must be called with ctx->mutex and ctx->lock held. |
1003 | */ | |
04289bb9 | 1004 | static void |
cdd6c482 | 1005 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1006 | { |
8a49542c PZ |
1007 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1008 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1009 | |
1010 | /* | |
8a49542c PZ |
1011 | * If we're a stand alone event or group leader, we go to the context |
1012 | * list, group events are kept attached to the group so that | |
1013 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1014 | */ |
8a49542c | 1015 | if (event->group_leader == event) { |
889ff015 FW |
1016 | struct list_head *list; |
1017 | ||
d6f962b5 FW |
1018 | if (is_software_event(event)) |
1019 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1020 | ||
889ff015 FW |
1021 | list = ctx_group_list(event, ctx); |
1022 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1023 | } |
592903cd | 1024 | |
08309379 | 1025 | if (is_cgroup_event(event)) |
e5d1367f | 1026 | ctx->nr_cgroups++; |
e5d1367f | 1027 | |
d010b332 SE |
1028 | if (has_branch_stack(event)) |
1029 | ctx->nr_branch_stack++; | |
1030 | ||
cdd6c482 | 1031 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1032 | if (!ctx->nr_events) |
108b02cf | 1033 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1034 | ctx->nr_events++; |
1035 | if (event->attr.inherit_stat) | |
bfbd3381 | 1036 | ctx->nr_stat++; |
04289bb9 IM |
1037 | } |
1038 | ||
0231bb53 JO |
1039 | /* |
1040 | * Initialize event state based on the perf_event_attr::disabled. | |
1041 | */ | |
1042 | static inline void perf_event__state_init(struct perf_event *event) | |
1043 | { | |
1044 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1045 | PERF_EVENT_STATE_INACTIVE; | |
1046 | } | |
1047 | ||
c320c7b7 ACM |
1048 | /* |
1049 | * Called at perf_event creation and when events are attached/detached from a | |
1050 | * group. | |
1051 | */ | |
1052 | static void perf_event__read_size(struct perf_event *event) | |
1053 | { | |
1054 | int entry = sizeof(u64); /* value */ | |
1055 | int size = 0; | |
1056 | int nr = 1; | |
1057 | ||
1058 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1059 | size += sizeof(u64); | |
1060 | ||
1061 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1062 | size += sizeof(u64); | |
1063 | ||
1064 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1065 | entry += sizeof(u64); | |
1066 | ||
1067 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1068 | nr += event->group_leader->nr_siblings; | |
1069 | size += sizeof(u64); | |
1070 | } | |
1071 | ||
1072 | size += entry * nr; | |
1073 | event->read_size = size; | |
1074 | } | |
1075 | ||
1076 | static void perf_event__header_size(struct perf_event *event) | |
1077 | { | |
1078 | struct perf_sample_data *data; | |
1079 | u64 sample_type = event->attr.sample_type; | |
1080 | u16 size = 0; | |
1081 | ||
1082 | perf_event__read_size(event); | |
1083 | ||
1084 | if (sample_type & PERF_SAMPLE_IP) | |
1085 | size += sizeof(data->ip); | |
1086 | ||
6844c09d ACM |
1087 | if (sample_type & PERF_SAMPLE_ADDR) |
1088 | size += sizeof(data->addr); | |
1089 | ||
1090 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1091 | size += sizeof(data->period); | |
1092 | ||
c3feedf2 AK |
1093 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1094 | size += sizeof(data->weight); | |
1095 | ||
6844c09d ACM |
1096 | if (sample_type & PERF_SAMPLE_READ) |
1097 | size += event->read_size; | |
1098 | ||
d6be9ad6 SE |
1099 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1100 | size += sizeof(data->data_src.val); | |
1101 | ||
6844c09d ACM |
1102 | event->header_size = size; |
1103 | } | |
1104 | ||
1105 | static void perf_event__id_header_size(struct perf_event *event) | |
1106 | { | |
1107 | struct perf_sample_data *data; | |
1108 | u64 sample_type = event->attr.sample_type; | |
1109 | u16 size = 0; | |
1110 | ||
c320c7b7 ACM |
1111 | if (sample_type & PERF_SAMPLE_TID) |
1112 | size += sizeof(data->tid_entry); | |
1113 | ||
1114 | if (sample_type & PERF_SAMPLE_TIME) | |
1115 | size += sizeof(data->time); | |
1116 | ||
c320c7b7 ACM |
1117 | if (sample_type & PERF_SAMPLE_ID) |
1118 | size += sizeof(data->id); | |
1119 | ||
1120 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1121 | size += sizeof(data->stream_id); | |
1122 | ||
1123 | if (sample_type & PERF_SAMPLE_CPU) | |
1124 | size += sizeof(data->cpu_entry); | |
1125 | ||
6844c09d | 1126 | event->id_header_size = size; |
c320c7b7 ACM |
1127 | } |
1128 | ||
8a49542c PZ |
1129 | static void perf_group_attach(struct perf_event *event) |
1130 | { | |
c320c7b7 | 1131 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1132 | |
74c3337c PZ |
1133 | /* |
1134 | * We can have double attach due to group movement in perf_event_open. | |
1135 | */ | |
1136 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1137 | return; | |
1138 | ||
8a49542c PZ |
1139 | event->attach_state |= PERF_ATTACH_GROUP; |
1140 | ||
1141 | if (group_leader == event) | |
1142 | return; | |
1143 | ||
1144 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1145 | !is_software_event(event)) | |
1146 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1147 | ||
1148 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1149 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1150 | |
1151 | perf_event__header_size(group_leader); | |
1152 | ||
1153 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1154 | perf_event__header_size(pos); | |
8a49542c PZ |
1155 | } |
1156 | ||
a63eaf34 | 1157 | /* |
cdd6c482 | 1158 | * Remove a event from the lists for its context. |
fccc714b | 1159 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1160 | */ |
04289bb9 | 1161 | static void |
cdd6c482 | 1162 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1163 | { |
68cacd29 | 1164 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1165 | /* |
1166 | * We can have double detach due to exit/hot-unplug + close. | |
1167 | */ | |
1168 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1169 | return; |
8a49542c PZ |
1170 | |
1171 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1172 | ||
68cacd29 | 1173 | if (is_cgroup_event(event)) { |
e5d1367f | 1174 | ctx->nr_cgroups--; |
68cacd29 SE |
1175 | cpuctx = __get_cpu_context(ctx); |
1176 | /* | |
1177 | * if there are no more cgroup events | |
1178 | * then cler cgrp to avoid stale pointer | |
1179 | * in update_cgrp_time_from_cpuctx() | |
1180 | */ | |
1181 | if (!ctx->nr_cgroups) | |
1182 | cpuctx->cgrp = NULL; | |
1183 | } | |
e5d1367f | 1184 | |
d010b332 SE |
1185 | if (has_branch_stack(event)) |
1186 | ctx->nr_branch_stack--; | |
1187 | ||
cdd6c482 IM |
1188 | ctx->nr_events--; |
1189 | if (event->attr.inherit_stat) | |
bfbd3381 | 1190 | ctx->nr_stat--; |
8bc20959 | 1191 | |
cdd6c482 | 1192 | list_del_rcu(&event->event_entry); |
04289bb9 | 1193 | |
8a49542c PZ |
1194 | if (event->group_leader == event) |
1195 | list_del_init(&event->group_entry); | |
5c148194 | 1196 | |
96c21a46 | 1197 | update_group_times(event); |
b2e74a26 SE |
1198 | |
1199 | /* | |
1200 | * If event was in error state, then keep it | |
1201 | * that way, otherwise bogus counts will be | |
1202 | * returned on read(). The only way to get out | |
1203 | * of error state is by explicit re-enabling | |
1204 | * of the event | |
1205 | */ | |
1206 | if (event->state > PERF_EVENT_STATE_OFF) | |
1207 | event->state = PERF_EVENT_STATE_OFF; | |
050735b0 PZ |
1208 | } |
1209 | ||
8a49542c | 1210 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1211 | { |
1212 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1213 | struct list_head *list = NULL; |
1214 | ||
1215 | /* | |
1216 | * We can have double detach due to exit/hot-unplug + close. | |
1217 | */ | |
1218 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1219 | return; | |
1220 | ||
1221 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1222 | ||
1223 | /* | |
1224 | * If this is a sibling, remove it from its group. | |
1225 | */ | |
1226 | if (event->group_leader != event) { | |
1227 | list_del_init(&event->group_entry); | |
1228 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1229 | goto out; |
8a49542c PZ |
1230 | } |
1231 | ||
1232 | if (!list_empty(&event->group_entry)) | |
1233 | list = &event->group_entry; | |
2e2af50b | 1234 | |
04289bb9 | 1235 | /* |
cdd6c482 IM |
1236 | * If this was a group event with sibling events then |
1237 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1238 | * to whatever list we are on. |
04289bb9 | 1239 | */ |
cdd6c482 | 1240 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1241 | if (list) |
1242 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1243 | sibling->group_leader = sibling; |
d6f962b5 FW |
1244 | |
1245 | /* Inherit group flags from the previous leader */ | |
1246 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1247 | } |
c320c7b7 ACM |
1248 | |
1249 | out: | |
1250 | perf_event__header_size(event->group_leader); | |
1251 | ||
1252 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1253 | perf_event__header_size(tmp); | |
04289bb9 IM |
1254 | } |
1255 | ||
fa66f07a SE |
1256 | static inline int |
1257 | event_filter_match(struct perf_event *event) | |
1258 | { | |
e5d1367f SE |
1259 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1260 | && perf_cgroup_match(event); | |
fa66f07a SE |
1261 | } |
1262 | ||
9ffcfa6f SE |
1263 | static void |
1264 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1265 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1266 | struct perf_event_context *ctx) |
3b6f9e5c | 1267 | { |
4158755d | 1268 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1269 | u64 delta; |
1270 | /* | |
1271 | * An event which could not be activated because of | |
1272 | * filter mismatch still needs to have its timings | |
1273 | * maintained, otherwise bogus information is return | |
1274 | * via read() for time_enabled, time_running: | |
1275 | */ | |
1276 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1277 | && !event_filter_match(event)) { | |
e5d1367f | 1278 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1279 | event->tstamp_running += delta; |
4158755d | 1280 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1281 | } |
1282 | ||
cdd6c482 | 1283 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1284 | return; |
3b6f9e5c | 1285 | |
cdd6c482 IM |
1286 | event->state = PERF_EVENT_STATE_INACTIVE; |
1287 | if (event->pending_disable) { | |
1288 | event->pending_disable = 0; | |
1289 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1290 | } |
4158755d | 1291 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1292 | event->pmu->del(event, 0); |
cdd6c482 | 1293 | event->oncpu = -1; |
3b6f9e5c | 1294 | |
cdd6c482 | 1295 | if (!is_software_event(event)) |
3b6f9e5c PM |
1296 | cpuctx->active_oncpu--; |
1297 | ctx->nr_active--; | |
0f5a2601 PZ |
1298 | if (event->attr.freq && event->attr.sample_freq) |
1299 | ctx->nr_freq--; | |
cdd6c482 | 1300 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c PM |
1301 | cpuctx->exclusive = 0; |
1302 | } | |
1303 | ||
d859e29f | 1304 | static void |
cdd6c482 | 1305 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1306 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1307 | struct perf_event_context *ctx) |
d859e29f | 1308 | { |
cdd6c482 | 1309 | struct perf_event *event; |
fa66f07a | 1310 | int state = group_event->state; |
d859e29f | 1311 | |
cdd6c482 | 1312 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1313 | |
1314 | /* | |
1315 | * Schedule out siblings (if any): | |
1316 | */ | |
cdd6c482 IM |
1317 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1318 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1319 | |
fa66f07a | 1320 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1321 | cpuctx->exclusive = 0; |
1322 | } | |
1323 | ||
0793a61d | 1324 | /* |
cdd6c482 | 1325 | * Cross CPU call to remove a performance event |
0793a61d | 1326 | * |
cdd6c482 | 1327 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1328 | * remove it from the context list. |
1329 | */ | |
fe4b04fa | 1330 | static int __perf_remove_from_context(void *info) |
0793a61d | 1331 | { |
cdd6c482 IM |
1332 | struct perf_event *event = info; |
1333 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1334 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1335 | |
e625cce1 | 1336 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1337 | event_sched_out(event, cpuctx, ctx); |
cdd6c482 | 1338 | list_del_event(event, ctx); |
64ce3126 PZ |
1339 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1340 | ctx->is_active = 0; | |
1341 | cpuctx->task_ctx = NULL; | |
1342 | } | |
e625cce1 | 1343 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1344 | |
1345 | return 0; | |
0793a61d TG |
1346 | } |
1347 | ||
1348 | ||
1349 | /* | |
cdd6c482 | 1350 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1351 | * |
cdd6c482 | 1352 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1353 | * call when the task is on a CPU. |
c93f7669 | 1354 | * |
cdd6c482 IM |
1355 | * If event->ctx is a cloned context, callers must make sure that |
1356 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1357 | * remains valid. This is OK when called from perf_release since |
1358 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1359 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1360 | * context has been detached from its task. |
0793a61d | 1361 | */ |
fe4b04fa | 1362 | static void perf_remove_from_context(struct perf_event *event) |
0793a61d | 1363 | { |
cdd6c482 | 1364 | struct perf_event_context *ctx = event->ctx; |
0793a61d TG |
1365 | struct task_struct *task = ctx->task; |
1366 | ||
fe4b04fa PZ |
1367 | lockdep_assert_held(&ctx->mutex); |
1368 | ||
0793a61d TG |
1369 | if (!task) { |
1370 | /* | |
cdd6c482 | 1371 | * Per cpu events are removed via an smp call and |
af901ca1 | 1372 | * the removal is always successful. |
0793a61d | 1373 | */ |
fe4b04fa | 1374 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
0793a61d TG |
1375 | return; |
1376 | } | |
1377 | ||
1378 | retry: | |
fe4b04fa PZ |
1379 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1380 | return; | |
0793a61d | 1381 | |
e625cce1 | 1382 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1383 | /* |
fe4b04fa PZ |
1384 | * If we failed to find a running task, but find the context active now |
1385 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1386 | */ |
fe4b04fa | 1387 | if (ctx->is_active) { |
e625cce1 | 1388 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1389 | goto retry; |
1390 | } | |
1391 | ||
1392 | /* | |
fe4b04fa PZ |
1393 | * Since the task isn't running, its safe to remove the event, us |
1394 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1395 | */ |
fe4b04fa | 1396 | list_del_event(event, ctx); |
e625cce1 | 1397 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1398 | } |
1399 | ||
d859e29f | 1400 | /* |
cdd6c482 | 1401 | * Cross CPU call to disable a performance event |
d859e29f | 1402 | */ |
500ad2d8 | 1403 | int __perf_event_disable(void *info) |
d859e29f | 1404 | { |
cdd6c482 | 1405 | struct perf_event *event = info; |
cdd6c482 | 1406 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1407 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1408 | |
1409 | /* | |
cdd6c482 IM |
1410 | * If this is a per-task event, need to check whether this |
1411 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1412 | * |
1413 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1414 | * flipping contexts around. | |
d859e29f | 1415 | */ |
665c2142 | 1416 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1417 | return -EINVAL; |
d859e29f | 1418 | |
e625cce1 | 1419 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1420 | |
1421 | /* | |
cdd6c482 | 1422 | * If the event is on, turn it off. |
d859e29f PM |
1423 | * If it is in error state, leave it in error state. |
1424 | */ | |
cdd6c482 | 1425 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1426 | update_context_time(ctx); |
e5d1367f | 1427 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1428 | update_group_times(event); |
1429 | if (event == event->group_leader) | |
1430 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1431 | else |
cdd6c482 IM |
1432 | event_sched_out(event, cpuctx, ctx); |
1433 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1434 | } |
1435 | ||
e625cce1 | 1436 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1437 | |
1438 | return 0; | |
d859e29f PM |
1439 | } |
1440 | ||
1441 | /* | |
cdd6c482 | 1442 | * Disable a event. |
c93f7669 | 1443 | * |
cdd6c482 IM |
1444 | * If event->ctx is a cloned context, callers must make sure that |
1445 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1446 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1447 | * perf_event_for_each_child or perf_event_for_each because they |
1448 | * hold the top-level event's child_mutex, so any descendant that | |
1449 | * goes to exit will block in sync_child_event. | |
1450 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1451 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1452 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1453 | */ |
44234adc | 1454 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1455 | { |
cdd6c482 | 1456 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1457 | struct task_struct *task = ctx->task; |
1458 | ||
1459 | if (!task) { | |
1460 | /* | |
cdd6c482 | 1461 | * Disable the event on the cpu that it's on |
d859e29f | 1462 | */ |
fe4b04fa | 1463 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1464 | return; |
1465 | } | |
1466 | ||
9ed6060d | 1467 | retry: |
fe4b04fa PZ |
1468 | if (!task_function_call(task, __perf_event_disable, event)) |
1469 | return; | |
d859e29f | 1470 | |
e625cce1 | 1471 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1472 | /* |
cdd6c482 | 1473 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1474 | */ |
cdd6c482 | 1475 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1476 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1477 | /* |
1478 | * Reload the task pointer, it might have been changed by | |
1479 | * a concurrent perf_event_context_sched_out(). | |
1480 | */ | |
1481 | task = ctx->task; | |
d859e29f PM |
1482 | goto retry; |
1483 | } | |
1484 | ||
1485 | /* | |
1486 | * Since we have the lock this context can't be scheduled | |
1487 | * in, so we can change the state safely. | |
1488 | */ | |
cdd6c482 IM |
1489 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1490 | update_group_times(event); | |
1491 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1492 | } |
e625cce1 | 1493 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1494 | } |
dcfce4a0 | 1495 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1496 | |
e5d1367f SE |
1497 | static void perf_set_shadow_time(struct perf_event *event, |
1498 | struct perf_event_context *ctx, | |
1499 | u64 tstamp) | |
1500 | { | |
1501 | /* | |
1502 | * use the correct time source for the time snapshot | |
1503 | * | |
1504 | * We could get by without this by leveraging the | |
1505 | * fact that to get to this function, the caller | |
1506 | * has most likely already called update_context_time() | |
1507 | * and update_cgrp_time_xx() and thus both timestamp | |
1508 | * are identical (or very close). Given that tstamp is, | |
1509 | * already adjusted for cgroup, we could say that: | |
1510 | * tstamp - ctx->timestamp | |
1511 | * is equivalent to | |
1512 | * tstamp - cgrp->timestamp. | |
1513 | * | |
1514 | * Then, in perf_output_read(), the calculation would | |
1515 | * work with no changes because: | |
1516 | * - event is guaranteed scheduled in | |
1517 | * - no scheduled out in between | |
1518 | * - thus the timestamp would be the same | |
1519 | * | |
1520 | * But this is a bit hairy. | |
1521 | * | |
1522 | * So instead, we have an explicit cgroup call to remain | |
1523 | * within the time time source all along. We believe it | |
1524 | * is cleaner and simpler to understand. | |
1525 | */ | |
1526 | if (is_cgroup_event(event)) | |
1527 | perf_cgroup_set_shadow_time(event, tstamp); | |
1528 | else | |
1529 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1530 | } | |
1531 | ||
4fe757dd PZ |
1532 | #define MAX_INTERRUPTS (~0ULL) |
1533 | ||
1534 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1535 | ||
235c7fc7 | 1536 | static int |
9ffcfa6f | 1537 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1538 | struct perf_cpu_context *cpuctx, |
6e37738a | 1539 | struct perf_event_context *ctx) |
235c7fc7 | 1540 | { |
4158755d SE |
1541 | u64 tstamp = perf_event_time(event); |
1542 | ||
cdd6c482 | 1543 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1544 | return 0; |
1545 | ||
cdd6c482 | 1546 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1547 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1548 | |
1549 | /* | |
1550 | * Unthrottle events, since we scheduled we might have missed several | |
1551 | * ticks already, also for a heavily scheduling task there is little | |
1552 | * guarantee it'll get a tick in a timely manner. | |
1553 | */ | |
1554 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1555 | perf_log_throttle(event, 1); | |
1556 | event->hw.interrupts = 0; | |
1557 | } | |
1558 | ||
235c7fc7 IM |
1559 | /* |
1560 | * The new state must be visible before we turn it on in the hardware: | |
1561 | */ | |
1562 | smp_wmb(); | |
1563 | ||
a4eaf7f1 | 1564 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1565 | event->state = PERF_EVENT_STATE_INACTIVE; |
1566 | event->oncpu = -1; | |
235c7fc7 IM |
1567 | return -EAGAIN; |
1568 | } | |
1569 | ||
4158755d | 1570 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1571 | |
e5d1367f | 1572 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1573 | |
cdd6c482 | 1574 | if (!is_software_event(event)) |
3b6f9e5c | 1575 | cpuctx->active_oncpu++; |
235c7fc7 | 1576 | ctx->nr_active++; |
0f5a2601 PZ |
1577 | if (event->attr.freq && event->attr.sample_freq) |
1578 | ctx->nr_freq++; | |
235c7fc7 | 1579 | |
cdd6c482 | 1580 | if (event->attr.exclusive) |
3b6f9e5c PM |
1581 | cpuctx->exclusive = 1; |
1582 | ||
235c7fc7 IM |
1583 | return 0; |
1584 | } | |
1585 | ||
6751b71e | 1586 | static int |
cdd6c482 | 1587 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1588 | struct perf_cpu_context *cpuctx, |
6e37738a | 1589 | struct perf_event_context *ctx) |
6751b71e | 1590 | { |
6bde9b6c | 1591 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1592 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1593 | u64 now = ctx->time; |
1594 | bool simulate = false; | |
6751b71e | 1595 | |
cdd6c482 | 1596 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1597 | return 0; |
1598 | ||
ad5133b7 | 1599 | pmu->start_txn(pmu); |
6bde9b6c | 1600 | |
9ffcfa6f | 1601 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1602 | pmu->cancel_txn(pmu); |
9e630205 | 1603 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1604 | return -EAGAIN; |
90151c35 | 1605 | } |
6751b71e PM |
1606 | |
1607 | /* | |
1608 | * Schedule in siblings as one group (if any): | |
1609 | */ | |
cdd6c482 | 1610 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1611 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1612 | partial_group = event; |
6751b71e PM |
1613 | goto group_error; |
1614 | } | |
1615 | } | |
1616 | ||
9ffcfa6f | 1617 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1618 | return 0; |
9ffcfa6f | 1619 | |
6751b71e PM |
1620 | group_error: |
1621 | /* | |
1622 | * Groups can be scheduled in as one unit only, so undo any | |
1623 | * partial group before returning: | |
d7842da4 SE |
1624 | * The events up to the failed event are scheduled out normally, |
1625 | * tstamp_stopped will be updated. | |
1626 | * | |
1627 | * The failed events and the remaining siblings need to have | |
1628 | * their timings updated as if they had gone thru event_sched_in() | |
1629 | * and event_sched_out(). This is required to get consistent timings | |
1630 | * across the group. This also takes care of the case where the group | |
1631 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1632 | * the time the event was actually stopped, such that time delta | |
1633 | * calculation in update_event_times() is correct. | |
6751b71e | 1634 | */ |
cdd6c482 IM |
1635 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1636 | if (event == partial_group) | |
d7842da4 SE |
1637 | simulate = true; |
1638 | ||
1639 | if (simulate) { | |
1640 | event->tstamp_running += now - event->tstamp_stopped; | |
1641 | event->tstamp_stopped = now; | |
1642 | } else { | |
1643 | event_sched_out(event, cpuctx, ctx); | |
1644 | } | |
6751b71e | 1645 | } |
9ffcfa6f | 1646 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1647 | |
ad5133b7 | 1648 | pmu->cancel_txn(pmu); |
90151c35 | 1649 | |
9e630205 SE |
1650 | perf_cpu_hrtimer_restart(cpuctx); |
1651 | ||
6751b71e PM |
1652 | return -EAGAIN; |
1653 | } | |
1654 | ||
3b6f9e5c | 1655 | /* |
cdd6c482 | 1656 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1657 | */ |
cdd6c482 | 1658 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1659 | struct perf_cpu_context *cpuctx, |
1660 | int can_add_hw) | |
1661 | { | |
1662 | /* | |
cdd6c482 | 1663 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1664 | */ |
d6f962b5 | 1665 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1666 | return 1; |
1667 | /* | |
1668 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1669 | * events can go on. |
3b6f9e5c PM |
1670 | */ |
1671 | if (cpuctx->exclusive) | |
1672 | return 0; | |
1673 | /* | |
1674 | * If this group is exclusive and there are already | |
cdd6c482 | 1675 | * events on the CPU, it can't go on. |
3b6f9e5c | 1676 | */ |
cdd6c482 | 1677 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1678 | return 0; |
1679 | /* | |
1680 | * Otherwise, try to add it if all previous groups were able | |
1681 | * to go on. | |
1682 | */ | |
1683 | return can_add_hw; | |
1684 | } | |
1685 | ||
cdd6c482 IM |
1686 | static void add_event_to_ctx(struct perf_event *event, |
1687 | struct perf_event_context *ctx) | |
53cfbf59 | 1688 | { |
4158755d SE |
1689 | u64 tstamp = perf_event_time(event); |
1690 | ||
cdd6c482 | 1691 | list_add_event(event, ctx); |
8a49542c | 1692 | perf_group_attach(event); |
4158755d SE |
1693 | event->tstamp_enabled = tstamp; |
1694 | event->tstamp_running = tstamp; | |
1695 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1696 | } |
1697 | ||
2c29ef0f PZ |
1698 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1699 | static void | |
1700 | ctx_sched_in(struct perf_event_context *ctx, | |
1701 | struct perf_cpu_context *cpuctx, | |
1702 | enum event_type_t event_type, | |
1703 | struct task_struct *task); | |
fe4b04fa | 1704 | |
dce5855b PZ |
1705 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1706 | struct perf_event_context *ctx, | |
1707 | struct task_struct *task) | |
1708 | { | |
1709 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1710 | if (ctx) | |
1711 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1712 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1713 | if (ctx) | |
1714 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1715 | } | |
1716 | ||
0793a61d | 1717 | /* |
cdd6c482 | 1718 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1719 | * |
1720 | * Must be called with ctx->mutex held | |
0793a61d | 1721 | */ |
fe4b04fa | 1722 | static int __perf_install_in_context(void *info) |
0793a61d | 1723 | { |
cdd6c482 IM |
1724 | struct perf_event *event = info; |
1725 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1726 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1727 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1728 | struct task_struct *task = current; | |
1729 | ||
b58f6b0d | 1730 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1731 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1732 | |
1733 | /* | |
2c29ef0f | 1734 | * If there was an active task_ctx schedule it out. |
0793a61d | 1735 | */ |
b58f6b0d | 1736 | if (task_ctx) |
2c29ef0f | 1737 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1738 | |
1739 | /* | |
1740 | * If the context we're installing events in is not the | |
1741 | * active task_ctx, flip them. | |
1742 | */ | |
1743 | if (ctx->task && task_ctx != ctx) { | |
1744 | if (task_ctx) | |
1745 | raw_spin_unlock(&task_ctx->lock); | |
1746 | raw_spin_lock(&ctx->lock); | |
1747 | task_ctx = ctx; | |
1748 | } | |
1749 | ||
1750 | if (task_ctx) { | |
1751 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1752 | task = task_ctx->task; |
1753 | } | |
b58f6b0d | 1754 | |
2c29ef0f | 1755 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1756 | |
4af4998b | 1757 | update_context_time(ctx); |
e5d1367f SE |
1758 | /* |
1759 | * update cgrp time only if current cgrp | |
1760 | * matches event->cgrp. Must be done before | |
1761 | * calling add_event_to_ctx() | |
1762 | */ | |
1763 | update_cgrp_time_from_event(event); | |
0793a61d | 1764 | |
cdd6c482 | 1765 | add_event_to_ctx(event, ctx); |
0793a61d | 1766 | |
d859e29f | 1767 | /* |
2c29ef0f | 1768 | * Schedule everything back in |
d859e29f | 1769 | */ |
dce5855b | 1770 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1771 | |
1772 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1773 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1774 | |
1775 | return 0; | |
0793a61d TG |
1776 | } |
1777 | ||
1778 | /* | |
cdd6c482 | 1779 | * Attach a performance event to a context |
0793a61d | 1780 | * |
cdd6c482 IM |
1781 | * First we add the event to the list with the hardware enable bit |
1782 | * in event->hw_config cleared. | |
0793a61d | 1783 | * |
cdd6c482 | 1784 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1785 | * call to enable it in the task context. The task might have been |
1786 | * scheduled away, but we check this in the smp call again. | |
1787 | */ | |
1788 | static void | |
cdd6c482 IM |
1789 | perf_install_in_context(struct perf_event_context *ctx, |
1790 | struct perf_event *event, | |
0793a61d TG |
1791 | int cpu) |
1792 | { | |
1793 | struct task_struct *task = ctx->task; | |
1794 | ||
fe4b04fa PZ |
1795 | lockdep_assert_held(&ctx->mutex); |
1796 | ||
c3f00c70 | 1797 | event->ctx = ctx; |
0cda4c02 YZ |
1798 | if (event->cpu != -1) |
1799 | event->cpu = cpu; | |
c3f00c70 | 1800 | |
0793a61d TG |
1801 | if (!task) { |
1802 | /* | |
cdd6c482 | 1803 | * Per cpu events are installed via an smp call and |
af901ca1 | 1804 | * the install is always successful. |
0793a61d | 1805 | */ |
fe4b04fa | 1806 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1807 | return; |
1808 | } | |
1809 | ||
0793a61d | 1810 | retry: |
fe4b04fa PZ |
1811 | if (!task_function_call(task, __perf_install_in_context, event)) |
1812 | return; | |
0793a61d | 1813 | |
e625cce1 | 1814 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1815 | /* |
fe4b04fa PZ |
1816 | * If we failed to find a running task, but find the context active now |
1817 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1818 | */ |
fe4b04fa | 1819 | if (ctx->is_active) { |
e625cce1 | 1820 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1821 | goto retry; |
1822 | } | |
1823 | ||
1824 | /* | |
fe4b04fa PZ |
1825 | * Since the task isn't running, its safe to add the event, us holding |
1826 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1827 | */ |
fe4b04fa | 1828 | add_event_to_ctx(event, ctx); |
e625cce1 | 1829 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1830 | } |
1831 | ||
fa289bec | 1832 | /* |
cdd6c482 | 1833 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1834 | * Enabling the leader of a group effectively enables all |
1835 | * the group members that aren't explicitly disabled, so we | |
1836 | * have to update their ->tstamp_enabled also. | |
1837 | * Note: this works for group members as well as group leaders | |
1838 | * since the non-leader members' sibling_lists will be empty. | |
1839 | */ | |
1d9b482e | 1840 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1841 | { |
cdd6c482 | 1842 | struct perf_event *sub; |
4158755d | 1843 | u64 tstamp = perf_event_time(event); |
fa289bec | 1844 | |
cdd6c482 | 1845 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1846 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1847 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1848 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1849 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1850 | } |
fa289bec PM |
1851 | } |
1852 | ||
d859e29f | 1853 | /* |
cdd6c482 | 1854 | * Cross CPU call to enable a performance event |
d859e29f | 1855 | */ |
fe4b04fa | 1856 | static int __perf_event_enable(void *info) |
04289bb9 | 1857 | { |
cdd6c482 | 1858 | struct perf_event *event = info; |
cdd6c482 IM |
1859 | struct perf_event_context *ctx = event->ctx; |
1860 | struct perf_event *leader = event->group_leader; | |
108b02cf | 1861 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 1862 | int err; |
04289bb9 | 1863 | |
fe4b04fa PZ |
1864 | if (WARN_ON_ONCE(!ctx->is_active)) |
1865 | return -EINVAL; | |
3cbed429 | 1866 | |
e625cce1 | 1867 | raw_spin_lock(&ctx->lock); |
4af4998b | 1868 | update_context_time(ctx); |
d859e29f | 1869 | |
cdd6c482 | 1870 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 1871 | goto unlock; |
e5d1367f SE |
1872 | |
1873 | /* | |
1874 | * set current task's cgroup time reference point | |
1875 | */ | |
3f7cce3c | 1876 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 1877 | |
1d9b482e | 1878 | __perf_event_mark_enabled(event); |
04289bb9 | 1879 | |
e5d1367f SE |
1880 | if (!event_filter_match(event)) { |
1881 | if (is_cgroup_event(event)) | |
1882 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 1883 | goto unlock; |
e5d1367f | 1884 | } |
f4c4176f | 1885 | |
04289bb9 | 1886 | /* |
cdd6c482 | 1887 | * If the event is in a group and isn't the group leader, |
d859e29f | 1888 | * then don't put it on unless the group is on. |
04289bb9 | 1889 | */ |
cdd6c482 | 1890 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 1891 | goto unlock; |
3b6f9e5c | 1892 | |
cdd6c482 | 1893 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 1894 | err = -EEXIST; |
e758a33d | 1895 | } else { |
cdd6c482 | 1896 | if (event == leader) |
6e37738a | 1897 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 1898 | else |
6e37738a | 1899 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 1900 | } |
d859e29f PM |
1901 | |
1902 | if (err) { | |
1903 | /* | |
cdd6c482 | 1904 | * If this event can't go on and it's part of a |
d859e29f PM |
1905 | * group, then the whole group has to come off. |
1906 | */ | |
9e630205 | 1907 | if (leader != event) { |
d859e29f | 1908 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
1909 | perf_cpu_hrtimer_restart(cpuctx); |
1910 | } | |
0d48696f | 1911 | if (leader->attr.pinned) { |
53cfbf59 | 1912 | update_group_times(leader); |
cdd6c482 | 1913 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 1914 | } |
d859e29f PM |
1915 | } |
1916 | ||
9ed6060d | 1917 | unlock: |
e625cce1 | 1918 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1919 | |
1920 | return 0; | |
d859e29f PM |
1921 | } |
1922 | ||
1923 | /* | |
cdd6c482 | 1924 | * Enable a event. |
c93f7669 | 1925 | * |
cdd6c482 IM |
1926 | * If event->ctx is a cloned context, callers must make sure that |
1927 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1928 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
1929 | * perf_event_for_each_child or perf_event_for_each as described |
1930 | * for perf_event_disable. | |
d859e29f | 1931 | */ |
44234adc | 1932 | void perf_event_enable(struct perf_event *event) |
d859e29f | 1933 | { |
cdd6c482 | 1934 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1935 | struct task_struct *task = ctx->task; |
1936 | ||
1937 | if (!task) { | |
1938 | /* | |
cdd6c482 | 1939 | * Enable the event on the cpu that it's on |
d859e29f | 1940 | */ |
fe4b04fa | 1941 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
1942 | return; |
1943 | } | |
1944 | ||
e625cce1 | 1945 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 1946 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
1947 | goto out; |
1948 | ||
1949 | /* | |
cdd6c482 IM |
1950 | * If the event is in error state, clear that first. |
1951 | * That way, if we see the event in error state below, we | |
d859e29f PM |
1952 | * know that it has gone back into error state, as distinct |
1953 | * from the task having been scheduled away before the | |
1954 | * cross-call arrived. | |
1955 | */ | |
cdd6c482 IM |
1956 | if (event->state == PERF_EVENT_STATE_ERROR) |
1957 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 1958 | |
9ed6060d | 1959 | retry: |
fe4b04fa | 1960 | if (!ctx->is_active) { |
1d9b482e | 1961 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
1962 | goto out; |
1963 | } | |
1964 | ||
e625cce1 | 1965 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1966 | |
1967 | if (!task_function_call(task, __perf_event_enable, event)) | |
1968 | return; | |
d859e29f | 1969 | |
e625cce1 | 1970 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
1971 | |
1972 | /* | |
cdd6c482 | 1973 | * If the context is active and the event is still off, |
d859e29f PM |
1974 | * we need to retry the cross-call. |
1975 | */ | |
fe4b04fa PZ |
1976 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
1977 | /* | |
1978 | * task could have been flipped by a concurrent | |
1979 | * perf_event_context_sched_out() | |
1980 | */ | |
1981 | task = ctx->task; | |
d859e29f | 1982 | goto retry; |
fe4b04fa | 1983 | } |
fa289bec | 1984 | |
9ed6060d | 1985 | out: |
e625cce1 | 1986 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1987 | } |
dcfce4a0 | 1988 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 1989 | |
26ca5c11 | 1990 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 1991 | { |
2023b359 | 1992 | /* |
cdd6c482 | 1993 | * not supported on inherited events |
2023b359 | 1994 | */ |
2e939d1d | 1995 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
1996 | return -EINVAL; |
1997 | ||
cdd6c482 IM |
1998 | atomic_add(refresh, &event->event_limit); |
1999 | perf_event_enable(event); | |
2023b359 PZ |
2000 | |
2001 | return 0; | |
79f14641 | 2002 | } |
26ca5c11 | 2003 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2004 | |
5b0311e1 FW |
2005 | static void ctx_sched_out(struct perf_event_context *ctx, |
2006 | struct perf_cpu_context *cpuctx, | |
2007 | enum event_type_t event_type) | |
235c7fc7 | 2008 | { |
cdd6c482 | 2009 | struct perf_event *event; |
db24d33e | 2010 | int is_active = ctx->is_active; |
235c7fc7 | 2011 | |
db24d33e | 2012 | ctx->is_active &= ~event_type; |
cdd6c482 | 2013 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2014 | return; |
2015 | ||
4af4998b | 2016 | update_context_time(ctx); |
e5d1367f | 2017 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2018 | if (!ctx->nr_active) |
facc4307 | 2019 | return; |
5b0311e1 | 2020 | |
075e0b00 | 2021 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2022 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2023 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2024 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2025 | } |
889ff015 | 2026 | |
db24d33e | 2027 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2028 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2029 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2030 | } |
1b9a644f | 2031 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2032 | } |
2033 | ||
564c2b21 PM |
2034 | /* |
2035 | * Test whether two contexts are equivalent, i.e. whether they | |
2036 | * have both been cloned from the same version of the same context | |
cdd6c482 IM |
2037 | * and they both have the same number of enabled events. |
2038 | * If the number of enabled events is the same, then the set | |
2039 | * of enabled events should be the same, because these are both | |
2040 | * inherited contexts, therefore we can't access individual events | |
564c2b21 | 2041 | * in them directly with an fd; we can only enable/disable all |
cdd6c482 | 2042 | * events via prctl, or enable/disable all events in a family |
564c2b21 PM |
2043 | * via ioctl, which will have the same effect on both contexts. |
2044 | */ | |
cdd6c482 IM |
2045 | static int context_equiv(struct perf_event_context *ctx1, |
2046 | struct perf_event_context *ctx2) | |
564c2b21 PM |
2047 | { |
2048 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx | |
ad3a37de | 2049 | && ctx1->parent_gen == ctx2->parent_gen |
25346b93 | 2050 | && !ctx1->pin_count && !ctx2->pin_count; |
564c2b21 PM |
2051 | } |
2052 | ||
cdd6c482 IM |
2053 | static void __perf_event_sync_stat(struct perf_event *event, |
2054 | struct perf_event *next_event) | |
bfbd3381 PZ |
2055 | { |
2056 | u64 value; | |
2057 | ||
cdd6c482 | 2058 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2059 | return; |
2060 | ||
2061 | /* | |
cdd6c482 | 2062 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2063 | * because we're in the middle of a context switch and have IRQs |
2064 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2065 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2066 | * don't need to use it. |
2067 | */ | |
cdd6c482 IM |
2068 | switch (event->state) { |
2069 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2070 | event->pmu->read(event); |
2071 | /* fall-through */ | |
bfbd3381 | 2072 | |
cdd6c482 IM |
2073 | case PERF_EVENT_STATE_INACTIVE: |
2074 | update_event_times(event); | |
bfbd3381 PZ |
2075 | break; |
2076 | ||
2077 | default: | |
2078 | break; | |
2079 | } | |
2080 | ||
2081 | /* | |
cdd6c482 | 2082 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2083 | * values when we flip the contexts. |
2084 | */ | |
e7850595 PZ |
2085 | value = local64_read(&next_event->count); |
2086 | value = local64_xchg(&event->count, value); | |
2087 | local64_set(&next_event->count, value); | |
bfbd3381 | 2088 | |
cdd6c482 IM |
2089 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2090 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2091 | |
bfbd3381 | 2092 | /* |
19d2e755 | 2093 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2094 | */ |
cdd6c482 IM |
2095 | perf_event_update_userpage(event); |
2096 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2097 | } |
2098 | ||
2099 | #define list_next_entry(pos, member) \ | |
2100 | list_entry(pos->member.next, typeof(*pos), member) | |
2101 | ||
cdd6c482 IM |
2102 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2103 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2104 | { |
cdd6c482 | 2105 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2106 | |
2107 | if (!ctx->nr_stat) | |
2108 | return; | |
2109 | ||
02ffdbc8 PZ |
2110 | update_context_time(ctx); |
2111 | ||
cdd6c482 IM |
2112 | event = list_first_entry(&ctx->event_list, |
2113 | struct perf_event, event_entry); | |
bfbd3381 | 2114 | |
cdd6c482 IM |
2115 | next_event = list_first_entry(&next_ctx->event_list, |
2116 | struct perf_event, event_entry); | |
bfbd3381 | 2117 | |
cdd6c482 IM |
2118 | while (&event->event_entry != &ctx->event_list && |
2119 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2120 | |
cdd6c482 | 2121 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2122 | |
cdd6c482 IM |
2123 | event = list_next_entry(event, event_entry); |
2124 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2125 | } |
2126 | } | |
2127 | ||
fe4b04fa PZ |
2128 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2129 | struct task_struct *next) | |
0793a61d | 2130 | { |
8dc85d54 | 2131 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 IM |
2132 | struct perf_event_context *next_ctx; |
2133 | struct perf_event_context *parent; | |
108b02cf | 2134 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2135 | int do_switch = 1; |
0793a61d | 2136 | |
108b02cf PZ |
2137 | if (likely(!ctx)) |
2138 | return; | |
10989fb2 | 2139 | |
108b02cf PZ |
2140 | cpuctx = __get_cpu_context(ctx); |
2141 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2142 | return; |
2143 | ||
c93f7669 PM |
2144 | rcu_read_lock(); |
2145 | parent = rcu_dereference(ctx->parent_ctx); | |
8dc85d54 | 2146 | next_ctx = next->perf_event_ctxp[ctxn]; |
c93f7669 PM |
2147 | if (parent && next_ctx && |
2148 | rcu_dereference(next_ctx->parent_ctx) == parent) { | |
2149 | /* | |
2150 | * Looks like the two contexts are clones, so we might be | |
2151 | * able to optimize the context switch. We lock both | |
2152 | * contexts and check that they are clones under the | |
2153 | * lock (including re-checking that neither has been | |
2154 | * uncloned in the meantime). It doesn't matter which | |
2155 | * order we take the locks because no other cpu could | |
2156 | * be trying to lock both of these tasks. | |
2157 | */ | |
e625cce1 TG |
2158 | raw_spin_lock(&ctx->lock); |
2159 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2160 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2161 | /* |
2162 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2163 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2164 | */ |
8dc85d54 PZ |
2165 | task->perf_event_ctxp[ctxn] = next_ctx; |
2166 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2167 | ctx->task = next; |
2168 | next_ctx->task = task; | |
2169 | do_switch = 0; | |
bfbd3381 | 2170 | |
cdd6c482 | 2171 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2172 | } |
e625cce1 TG |
2173 | raw_spin_unlock(&next_ctx->lock); |
2174 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2175 | } |
c93f7669 | 2176 | rcu_read_unlock(); |
564c2b21 | 2177 | |
c93f7669 | 2178 | if (do_switch) { |
facc4307 | 2179 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2180 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2181 | cpuctx->task_ctx = NULL; |
facc4307 | 2182 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2183 | } |
0793a61d TG |
2184 | } |
2185 | ||
8dc85d54 PZ |
2186 | #define for_each_task_context_nr(ctxn) \ |
2187 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2188 | ||
2189 | /* | |
2190 | * Called from scheduler to remove the events of the current task, | |
2191 | * with interrupts disabled. | |
2192 | * | |
2193 | * We stop each event and update the event value in event->count. | |
2194 | * | |
2195 | * This does not protect us against NMI, but disable() | |
2196 | * sets the disabled bit in the control field of event _before_ | |
2197 | * accessing the event control register. If a NMI hits, then it will | |
2198 | * not restart the event. | |
2199 | */ | |
ab0cce56 JO |
2200 | void __perf_event_task_sched_out(struct task_struct *task, |
2201 | struct task_struct *next) | |
8dc85d54 PZ |
2202 | { |
2203 | int ctxn; | |
2204 | ||
8dc85d54 PZ |
2205 | for_each_task_context_nr(ctxn) |
2206 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2207 | |
2208 | /* | |
2209 | * if cgroup events exist on this CPU, then we need | |
2210 | * to check if we have to switch out PMU state. | |
2211 | * cgroup event are system-wide mode only | |
2212 | */ | |
2213 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2214 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2215 | } |
2216 | ||
04dc2dbb | 2217 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2218 | { |
108b02cf | 2219 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2220 | |
a63eaf34 PM |
2221 | if (!cpuctx->task_ctx) |
2222 | return; | |
012b84da IM |
2223 | |
2224 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2225 | return; | |
2226 | ||
04dc2dbb | 2227 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2228 | cpuctx->task_ctx = NULL; |
2229 | } | |
2230 | ||
5b0311e1 FW |
2231 | /* |
2232 | * Called with IRQs disabled | |
2233 | */ | |
2234 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2235 | enum event_type_t event_type) | |
2236 | { | |
2237 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2238 | } |
2239 | ||
235c7fc7 | 2240 | static void |
5b0311e1 | 2241 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2242 | struct perf_cpu_context *cpuctx) |
0793a61d | 2243 | { |
cdd6c482 | 2244 | struct perf_event *event; |
0793a61d | 2245 | |
889ff015 FW |
2246 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2247 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2248 | continue; |
5632ab12 | 2249 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2250 | continue; |
2251 | ||
e5d1367f SE |
2252 | /* may need to reset tstamp_enabled */ |
2253 | if (is_cgroup_event(event)) | |
2254 | perf_cgroup_mark_enabled(event, ctx); | |
2255 | ||
8c9ed8e1 | 2256 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2257 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2258 | |
2259 | /* | |
2260 | * If this pinned group hasn't been scheduled, | |
2261 | * put it in error state. | |
2262 | */ | |
cdd6c482 IM |
2263 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2264 | update_group_times(event); | |
2265 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2266 | } |
3b6f9e5c | 2267 | } |
5b0311e1 FW |
2268 | } |
2269 | ||
2270 | static void | |
2271 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2272 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2273 | { |
2274 | struct perf_event *event; | |
2275 | int can_add_hw = 1; | |
3b6f9e5c | 2276 | |
889ff015 FW |
2277 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2278 | /* Ignore events in OFF or ERROR state */ | |
2279 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2280 | continue; |
04289bb9 IM |
2281 | /* |
2282 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2283 | * of events: |
04289bb9 | 2284 | */ |
5632ab12 | 2285 | if (!event_filter_match(event)) |
0793a61d TG |
2286 | continue; |
2287 | ||
e5d1367f SE |
2288 | /* may need to reset tstamp_enabled */ |
2289 | if (is_cgroup_event(event)) | |
2290 | perf_cgroup_mark_enabled(event, ctx); | |
2291 | ||
9ed6060d | 2292 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2293 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2294 | can_add_hw = 0; |
9ed6060d | 2295 | } |
0793a61d | 2296 | } |
5b0311e1 FW |
2297 | } |
2298 | ||
2299 | static void | |
2300 | ctx_sched_in(struct perf_event_context *ctx, | |
2301 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2302 | enum event_type_t event_type, |
2303 | struct task_struct *task) | |
5b0311e1 | 2304 | { |
e5d1367f | 2305 | u64 now; |
db24d33e | 2306 | int is_active = ctx->is_active; |
e5d1367f | 2307 | |
db24d33e | 2308 | ctx->is_active |= event_type; |
5b0311e1 | 2309 | if (likely(!ctx->nr_events)) |
facc4307 | 2310 | return; |
5b0311e1 | 2311 | |
e5d1367f SE |
2312 | now = perf_clock(); |
2313 | ctx->timestamp = now; | |
3f7cce3c | 2314 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2315 | /* |
2316 | * First go through the list and put on any pinned groups | |
2317 | * in order to give them the best chance of going on. | |
2318 | */ | |
db24d33e | 2319 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2320 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2321 | |
2322 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2323 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2324 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2325 | } |
2326 | ||
329c0e01 | 2327 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2328 | enum event_type_t event_type, |
2329 | struct task_struct *task) | |
329c0e01 FW |
2330 | { |
2331 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2332 | ||
e5d1367f | 2333 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2334 | } |
2335 | ||
e5d1367f SE |
2336 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2337 | struct task_struct *task) | |
235c7fc7 | 2338 | { |
108b02cf | 2339 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2340 | |
108b02cf | 2341 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2342 | if (cpuctx->task_ctx == ctx) |
2343 | return; | |
2344 | ||
facc4307 | 2345 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2346 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2347 | /* |
2348 | * We want to keep the following priority order: | |
2349 | * cpu pinned (that don't need to move), task pinned, | |
2350 | * cpu flexible, task flexible. | |
2351 | */ | |
2352 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2353 | ||
1d5f003f GN |
2354 | if (ctx->nr_events) |
2355 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2356 | |
86b47c25 GN |
2357 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2358 | ||
facc4307 PZ |
2359 | perf_pmu_enable(ctx->pmu); |
2360 | perf_ctx_unlock(cpuctx, ctx); | |
2361 | ||
b5ab4cd5 PZ |
2362 | /* |
2363 | * Since these rotations are per-cpu, we need to ensure the | |
2364 | * cpu-context we got scheduled on is actually rotating. | |
2365 | */ | |
108b02cf | 2366 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2367 | } |
2368 | ||
d010b332 SE |
2369 | /* |
2370 | * When sampling the branck stack in system-wide, it may be necessary | |
2371 | * to flush the stack on context switch. This happens when the branch | |
2372 | * stack does not tag its entries with the pid of the current task. | |
2373 | * Otherwise it becomes impossible to associate a branch entry with a | |
2374 | * task. This ambiguity is more likely to appear when the branch stack | |
2375 | * supports priv level filtering and the user sets it to monitor only | |
2376 | * at the user level (which could be a useful measurement in system-wide | |
2377 | * mode). In that case, the risk is high of having a branch stack with | |
2378 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2379 | * entries or stashing them somewhere in the PMU specific code layer. | |
2380 | * | |
2381 | * This function provides the context switch callback to the lower code | |
2382 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2383 | * with at least one active event using taken branch sampling. | |
2384 | */ | |
2385 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2386 | struct task_struct *task) | |
2387 | { | |
2388 | struct perf_cpu_context *cpuctx; | |
2389 | struct pmu *pmu; | |
2390 | unsigned long flags; | |
2391 | ||
2392 | /* no need to flush branch stack if not changing task */ | |
2393 | if (prev == task) | |
2394 | return; | |
2395 | ||
2396 | local_irq_save(flags); | |
2397 | ||
2398 | rcu_read_lock(); | |
2399 | ||
2400 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2401 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2402 | ||
2403 | /* | |
2404 | * check if the context has at least one | |
2405 | * event using PERF_SAMPLE_BRANCH_STACK | |
2406 | */ | |
2407 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2408 | && pmu->flush_branch_stack) { | |
2409 | ||
2410 | pmu = cpuctx->ctx.pmu; | |
2411 | ||
2412 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2413 | ||
2414 | perf_pmu_disable(pmu); | |
2415 | ||
2416 | pmu->flush_branch_stack(); | |
2417 | ||
2418 | perf_pmu_enable(pmu); | |
2419 | ||
2420 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2421 | } | |
2422 | } | |
2423 | ||
2424 | rcu_read_unlock(); | |
2425 | ||
2426 | local_irq_restore(flags); | |
2427 | } | |
2428 | ||
8dc85d54 PZ |
2429 | /* |
2430 | * Called from scheduler to add the events of the current task | |
2431 | * with interrupts disabled. | |
2432 | * | |
2433 | * We restore the event value and then enable it. | |
2434 | * | |
2435 | * This does not protect us against NMI, but enable() | |
2436 | * sets the enabled bit in the control field of event _before_ | |
2437 | * accessing the event control register. If a NMI hits, then it will | |
2438 | * keep the event running. | |
2439 | */ | |
ab0cce56 JO |
2440 | void __perf_event_task_sched_in(struct task_struct *prev, |
2441 | struct task_struct *task) | |
8dc85d54 PZ |
2442 | { |
2443 | struct perf_event_context *ctx; | |
2444 | int ctxn; | |
2445 | ||
2446 | for_each_task_context_nr(ctxn) { | |
2447 | ctx = task->perf_event_ctxp[ctxn]; | |
2448 | if (likely(!ctx)) | |
2449 | continue; | |
2450 | ||
e5d1367f | 2451 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2452 | } |
e5d1367f SE |
2453 | /* |
2454 | * if cgroup events exist on this CPU, then we need | |
2455 | * to check if we have to switch in PMU state. | |
2456 | * cgroup event are system-wide mode only | |
2457 | */ | |
2458 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2459 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2460 | |
2461 | /* check for system-wide branch_stack events */ | |
2462 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2463 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2464 | } |
2465 | ||
abd50713 PZ |
2466 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2467 | { | |
2468 | u64 frequency = event->attr.sample_freq; | |
2469 | u64 sec = NSEC_PER_SEC; | |
2470 | u64 divisor, dividend; | |
2471 | ||
2472 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2473 | ||
2474 | count_fls = fls64(count); | |
2475 | nsec_fls = fls64(nsec); | |
2476 | frequency_fls = fls64(frequency); | |
2477 | sec_fls = 30; | |
2478 | ||
2479 | /* | |
2480 | * We got @count in @nsec, with a target of sample_freq HZ | |
2481 | * the target period becomes: | |
2482 | * | |
2483 | * @count * 10^9 | |
2484 | * period = ------------------- | |
2485 | * @nsec * sample_freq | |
2486 | * | |
2487 | */ | |
2488 | ||
2489 | /* | |
2490 | * Reduce accuracy by one bit such that @a and @b converge | |
2491 | * to a similar magnitude. | |
2492 | */ | |
fe4b04fa | 2493 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2494 | do { \ |
2495 | if (a##_fls > b##_fls) { \ | |
2496 | a >>= 1; \ | |
2497 | a##_fls--; \ | |
2498 | } else { \ | |
2499 | b >>= 1; \ | |
2500 | b##_fls--; \ | |
2501 | } \ | |
2502 | } while (0) | |
2503 | ||
2504 | /* | |
2505 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2506 | * the other, so that finally we can do a u64/u64 division. | |
2507 | */ | |
2508 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2509 | REDUCE_FLS(nsec, frequency); | |
2510 | REDUCE_FLS(sec, count); | |
2511 | } | |
2512 | ||
2513 | if (count_fls + sec_fls > 64) { | |
2514 | divisor = nsec * frequency; | |
2515 | ||
2516 | while (count_fls + sec_fls > 64) { | |
2517 | REDUCE_FLS(count, sec); | |
2518 | divisor >>= 1; | |
2519 | } | |
2520 | ||
2521 | dividend = count * sec; | |
2522 | } else { | |
2523 | dividend = count * sec; | |
2524 | ||
2525 | while (nsec_fls + frequency_fls > 64) { | |
2526 | REDUCE_FLS(nsec, frequency); | |
2527 | dividend >>= 1; | |
2528 | } | |
2529 | ||
2530 | divisor = nsec * frequency; | |
2531 | } | |
2532 | ||
f6ab91ad PZ |
2533 | if (!divisor) |
2534 | return dividend; | |
2535 | ||
abd50713 PZ |
2536 | return div64_u64(dividend, divisor); |
2537 | } | |
2538 | ||
e050e3f0 SE |
2539 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2540 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2541 | ||
f39d47ff | 2542 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2543 | { |
cdd6c482 | 2544 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2545 | s64 period, sample_period; |
bd2b5b12 PZ |
2546 | s64 delta; |
2547 | ||
abd50713 | 2548 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2549 | |
2550 | delta = (s64)(period - hwc->sample_period); | |
2551 | delta = (delta + 7) / 8; /* low pass filter */ | |
2552 | ||
2553 | sample_period = hwc->sample_period + delta; | |
2554 | ||
2555 | if (!sample_period) | |
2556 | sample_period = 1; | |
2557 | ||
bd2b5b12 | 2558 | hwc->sample_period = sample_period; |
abd50713 | 2559 | |
e7850595 | 2560 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2561 | if (disable) |
2562 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2563 | ||
e7850595 | 2564 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2565 | |
2566 | if (disable) | |
2567 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2568 | } |
bd2b5b12 PZ |
2569 | } |
2570 | ||
e050e3f0 SE |
2571 | /* |
2572 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2573 | * events. At the same time, make sure, having freq events does not change | |
2574 | * the rate of unthrottling as that would introduce bias. | |
2575 | */ | |
2576 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2577 | int needs_unthr) | |
60db5e09 | 2578 | { |
cdd6c482 IM |
2579 | struct perf_event *event; |
2580 | struct hw_perf_event *hwc; | |
e050e3f0 | 2581 | u64 now, period = TICK_NSEC; |
abd50713 | 2582 | s64 delta; |
60db5e09 | 2583 | |
e050e3f0 SE |
2584 | /* |
2585 | * only need to iterate over all events iff: | |
2586 | * - context have events in frequency mode (needs freq adjust) | |
2587 | * - there are events to unthrottle on this cpu | |
2588 | */ | |
2589 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2590 | return; |
2591 | ||
e050e3f0 | 2592 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2593 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2594 | |
03541f8b | 2595 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2596 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2597 | continue; |
2598 | ||
5632ab12 | 2599 | if (!event_filter_match(event)) |
5d27c23d PZ |
2600 | continue; |
2601 | ||
cdd6c482 | 2602 | hwc = &event->hw; |
6a24ed6c | 2603 | |
e050e3f0 SE |
2604 | if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) { |
2605 | hwc->interrupts = 0; | |
cdd6c482 | 2606 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2607 | event->pmu->start(event, 0); |
a78ac325 PZ |
2608 | } |
2609 | ||
cdd6c482 | 2610 | if (!event->attr.freq || !event->attr.sample_freq) |
60db5e09 PZ |
2611 | continue; |
2612 | ||
e050e3f0 SE |
2613 | /* |
2614 | * stop the event and update event->count | |
2615 | */ | |
2616 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2617 | ||
e7850595 | 2618 | now = local64_read(&event->count); |
abd50713 PZ |
2619 | delta = now - hwc->freq_count_stamp; |
2620 | hwc->freq_count_stamp = now; | |
60db5e09 | 2621 | |
e050e3f0 SE |
2622 | /* |
2623 | * restart the event | |
2624 | * reload only if value has changed | |
f39d47ff SE |
2625 | * we have stopped the event so tell that |
2626 | * to perf_adjust_period() to avoid stopping it | |
2627 | * twice. | |
e050e3f0 | 2628 | */ |
abd50713 | 2629 | if (delta > 0) |
f39d47ff | 2630 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2631 | |
2632 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
60db5e09 | 2633 | } |
e050e3f0 | 2634 | |
f39d47ff | 2635 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2636 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2637 | } |
2638 | ||
235c7fc7 | 2639 | /* |
cdd6c482 | 2640 | * Round-robin a context's events: |
235c7fc7 | 2641 | */ |
cdd6c482 | 2642 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2643 | { |
dddd3379 TG |
2644 | /* |
2645 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2646 | * disabled by the inheritance code. | |
2647 | */ | |
2648 | if (!ctx->rotate_disable) | |
2649 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2650 | } |
2651 | ||
b5ab4cd5 | 2652 | /* |
e9d2b064 PZ |
2653 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2654 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2655 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2656 | */ |
9e630205 | 2657 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2658 | { |
8dc85d54 | 2659 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2660 | int rotate = 0, remove = 1; |
7fc23a53 | 2661 | |
b5ab4cd5 | 2662 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2663 | remove = 0; |
b5ab4cd5 PZ |
2664 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2665 | rotate = 1; | |
2666 | } | |
235c7fc7 | 2667 | |
8dc85d54 | 2668 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2669 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2670 | remove = 0; |
b5ab4cd5 PZ |
2671 | if (ctx->nr_events != ctx->nr_active) |
2672 | rotate = 1; | |
2673 | } | |
9717e6cd | 2674 | |
e050e3f0 | 2675 | if (!rotate) |
0f5a2601 PZ |
2676 | goto done; |
2677 | ||
facc4307 | 2678 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2679 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2680 | |
e050e3f0 SE |
2681 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2682 | if (ctx) | |
2683 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2684 | |
e050e3f0 SE |
2685 | rotate_ctx(&cpuctx->ctx); |
2686 | if (ctx) | |
2687 | rotate_ctx(ctx); | |
235c7fc7 | 2688 | |
e050e3f0 | 2689 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2690 | |
0f5a2601 PZ |
2691 | perf_pmu_enable(cpuctx->ctx.pmu); |
2692 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2693 | done: |
e9d2b064 PZ |
2694 | if (remove) |
2695 | list_del_init(&cpuctx->rotation_list); | |
9e630205 SE |
2696 | |
2697 | return rotate; | |
e9d2b064 PZ |
2698 | } |
2699 | ||
026249ef FW |
2700 | #ifdef CONFIG_NO_HZ_FULL |
2701 | bool perf_event_can_stop_tick(void) | |
2702 | { | |
2703 | if (list_empty(&__get_cpu_var(rotation_list))) | |
2704 | return true; | |
2705 | else | |
2706 | return false; | |
2707 | } | |
2708 | #endif | |
2709 | ||
e9d2b064 PZ |
2710 | void perf_event_task_tick(void) |
2711 | { | |
2712 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2713 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2714 | struct perf_event_context *ctx; |
2715 | int throttled; | |
b5ab4cd5 | 2716 | |
e9d2b064 PZ |
2717 | WARN_ON(!irqs_disabled()); |
2718 | ||
e050e3f0 SE |
2719 | __this_cpu_inc(perf_throttled_seq); |
2720 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2721 | ||
e9d2b064 | 2722 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2723 | ctx = &cpuctx->ctx; |
2724 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2725 | ||
2726 | ctx = cpuctx->task_ctx; | |
2727 | if (ctx) | |
2728 | perf_adjust_freq_unthr_context(ctx, throttled); | |
e9d2b064 | 2729 | } |
0793a61d TG |
2730 | } |
2731 | ||
889ff015 FW |
2732 | static int event_enable_on_exec(struct perf_event *event, |
2733 | struct perf_event_context *ctx) | |
2734 | { | |
2735 | if (!event->attr.enable_on_exec) | |
2736 | return 0; | |
2737 | ||
2738 | event->attr.enable_on_exec = 0; | |
2739 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2740 | return 0; | |
2741 | ||
1d9b482e | 2742 | __perf_event_mark_enabled(event); |
889ff015 FW |
2743 | |
2744 | return 1; | |
2745 | } | |
2746 | ||
57e7986e | 2747 | /* |
cdd6c482 | 2748 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2749 | * This expects task == current. |
2750 | */ | |
8dc85d54 | 2751 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2752 | { |
cdd6c482 | 2753 | struct perf_event *event; |
57e7986e PM |
2754 | unsigned long flags; |
2755 | int enabled = 0; | |
889ff015 | 2756 | int ret; |
57e7986e PM |
2757 | |
2758 | local_irq_save(flags); | |
cdd6c482 | 2759 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2760 | goto out; |
2761 | ||
e566b76e SE |
2762 | /* |
2763 | * We must ctxsw out cgroup events to avoid conflict | |
2764 | * when invoking perf_task_event_sched_in() later on | |
2765 | * in this function. Otherwise we end up trying to | |
2766 | * ctxswin cgroup events which are already scheduled | |
2767 | * in. | |
2768 | */ | |
a8d757ef | 2769 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2770 | |
e625cce1 | 2771 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2772 | task_ctx_sched_out(ctx); |
57e7986e | 2773 | |
b79387ef | 2774 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2775 | ret = event_enable_on_exec(event, ctx); |
2776 | if (ret) | |
2777 | enabled = 1; | |
57e7986e PM |
2778 | } |
2779 | ||
2780 | /* | |
cdd6c482 | 2781 | * Unclone this context if we enabled any event. |
57e7986e | 2782 | */ |
71a851b4 PZ |
2783 | if (enabled) |
2784 | unclone_ctx(ctx); | |
57e7986e | 2785 | |
e625cce1 | 2786 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2787 | |
e566b76e SE |
2788 | /* |
2789 | * Also calls ctxswin for cgroup events, if any: | |
2790 | */ | |
e5d1367f | 2791 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2792 | out: |
57e7986e PM |
2793 | local_irq_restore(flags); |
2794 | } | |
2795 | ||
0793a61d | 2796 | /* |
cdd6c482 | 2797 | * Cross CPU call to read the hardware event |
0793a61d | 2798 | */ |
cdd6c482 | 2799 | static void __perf_event_read(void *info) |
0793a61d | 2800 | { |
cdd6c482 IM |
2801 | struct perf_event *event = info; |
2802 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2803 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2804 | |
e1ac3614 PM |
2805 | /* |
2806 | * If this is a task context, we need to check whether it is | |
2807 | * the current task context of this cpu. If not it has been | |
2808 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2809 | * event->count would have been updated to a recent sample |
2810 | * when the event was scheduled out. | |
e1ac3614 PM |
2811 | */ |
2812 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2813 | return; | |
2814 | ||
e625cce1 | 2815 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2816 | if (ctx->is_active) { |
542e72fc | 2817 | update_context_time(ctx); |
e5d1367f SE |
2818 | update_cgrp_time_from_event(event); |
2819 | } | |
cdd6c482 | 2820 | update_event_times(event); |
542e72fc PZ |
2821 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2822 | event->pmu->read(event); | |
e625cce1 | 2823 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2824 | } |
2825 | ||
b5e58793 PZ |
2826 | static inline u64 perf_event_count(struct perf_event *event) |
2827 | { | |
e7850595 | 2828 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
2829 | } |
2830 | ||
cdd6c482 | 2831 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
2832 | { |
2833 | /* | |
cdd6c482 IM |
2834 | * If event is enabled and currently active on a CPU, update the |
2835 | * value in the event structure: | |
0793a61d | 2836 | */ |
cdd6c482 IM |
2837 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2838 | smp_call_function_single(event->oncpu, | |
2839 | __perf_event_read, event, 1); | |
2840 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
2841 | struct perf_event_context *ctx = event->ctx; |
2842 | unsigned long flags; | |
2843 | ||
e625cce1 | 2844 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
2845 | /* |
2846 | * may read while context is not active | |
2847 | * (e.g., thread is blocked), in that case | |
2848 | * we cannot update context time | |
2849 | */ | |
e5d1367f | 2850 | if (ctx->is_active) { |
c530ccd9 | 2851 | update_context_time(ctx); |
e5d1367f SE |
2852 | update_cgrp_time_from_event(event); |
2853 | } | |
cdd6c482 | 2854 | update_event_times(event); |
e625cce1 | 2855 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
2856 | } |
2857 | ||
b5e58793 | 2858 | return perf_event_count(event); |
0793a61d TG |
2859 | } |
2860 | ||
a63eaf34 | 2861 | /* |
cdd6c482 | 2862 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 2863 | */ |
eb184479 | 2864 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 2865 | { |
e625cce1 | 2866 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 2867 | mutex_init(&ctx->mutex); |
889ff015 FW |
2868 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2869 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
2870 | INIT_LIST_HEAD(&ctx->event_list); |
2871 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
2872 | } |
2873 | ||
2874 | static struct perf_event_context * | |
2875 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
2876 | { | |
2877 | struct perf_event_context *ctx; | |
2878 | ||
2879 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
2880 | if (!ctx) | |
2881 | return NULL; | |
2882 | ||
2883 | __perf_event_init_context(ctx); | |
2884 | if (task) { | |
2885 | ctx->task = task; | |
2886 | get_task_struct(task); | |
0793a61d | 2887 | } |
eb184479 PZ |
2888 | ctx->pmu = pmu; |
2889 | ||
2890 | return ctx; | |
a63eaf34 PM |
2891 | } |
2892 | ||
2ebd4ffb MH |
2893 | static struct task_struct * |
2894 | find_lively_task_by_vpid(pid_t vpid) | |
2895 | { | |
2896 | struct task_struct *task; | |
2897 | int err; | |
0793a61d TG |
2898 | |
2899 | rcu_read_lock(); | |
2ebd4ffb | 2900 | if (!vpid) |
0793a61d TG |
2901 | task = current; |
2902 | else | |
2ebd4ffb | 2903 | task = find_task_by_vpid(vpid); |
0793a61d TG |
2904 | if (task) |
2905 | get_task_struct(task); | |
2906 | rcu_read_unlock(); | |
2907 | ||
2908 | if (!task) | |
2909 | return ERR_PTR(-ESRCH); | |
2910 | ||
0793a61d | 2911 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
2912 | err = -EACCES; |
2913 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
2914 | goto errout; | |
2915 | ||
2ebd4ffb MH |
2916 | return task; |
2917 | errout: | |
2918 | put_task_struct(task); | |
2919 | return ERR_PTR(err); | |
2920 | ||
2921 | } | |
2922 | ||
fe4b04fa PZ |
2923 | /* |
2924 | * Returns a matching context with refcount and pincount. | |
2925 | */ | |
108b02cf | 2926 | static struct perf_event_context * |
38a81da2 | 2927 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 2928 | { |
cdd6c482 | 2929 | struct perf_event_context *ctx; |
22a4f650 | 2930 | struct perf_cpu_context *cpuctx; |
25346b93 | 2931 | unsigned long flags; |
8dc85d54 | 2932 | int ctxn, err; |
0793a61d | 2933 | |
22a4ec72 | 2934 | if (!task) { |
cdd6c482 | 2935 | /* Must be root to operate on a CPU event: */ |
0764771d | 2936 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
2937 | return ERR_PTR(-EACCES); |
2938 | ||
0793a61d | 2939 | /* |
cdd6c482 | 2940 | * We could be clever and allow to attach a event to an |
0793a61d TG |
2941 | * offline CPU and activate it when the CPU comes up, but |
2942 | * that's for later. | |
2943 | */ | |
f6325e30 | 2944 | if (!cpu_online(cpu)) |
0793a61d TG |
2945 | return ERR_PTR(-ENODEV); |
2946 | ||
108b02cf | 2947 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 2948 | ctx = &cpuctx->ctx; |
c93f7669 | 2949 | get_ctx(ctx); |
fe4b04fa | 2950 | ++ctx->pin_count; |
0793a61d | 2951 | |
0793a61d TG |
2952 | return ctx; |
2953 | } | |
2954 | ||
8dc85d54 PZ |
2955 | err = -EINVAL; |
2956 | ctxn = pmu->task_ctx_nr; | |
2957 | if (ctxn < 0) | |
2958 | goto errout; | |
2959 | ||
9ed6060d | 2960 | retry: |
8dc85d54 | 2961 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 2962 | if (ctx) { |
71a851b4 | 2963 | unclone_ctx(ctx); |
fe4b04fa | 2964 | ++ctx->pin_count; |
e625cce1 | 2965 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 2966 | } else { |
eb184479 | 2967 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
2968 | err = -ENOMEM; |
2969 | if (!ctx) | |
2970 | goto errout; | |
eb184479 | 2971 | |
dbe08d82 ON |
2972 | err = 0; |
2973 | mutex_lock(&task->perf_event_mutex); | |
2974 | /* | |
2975 | * If it has already passed perf_event_exit_task(). | |
2976 | * we must see PF_EXITING, it takes this mutex too. | |
2977 | */ | |
2978 | if (task->flags & PF_EXITING) | |
2979 | err = -ESRCH; | |
2980 | else if (task->perf_event_ctxp[ctxn]) | |
2981 | err = -EAGAIN; | |
fe4b04fa | 2982 | else { |
9137fb28 | 2983 | get_ctx(ctx); |
fe4b04fa | 2984 | ++ctx->pin_count; |
dbe08d82 | 2985 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 2986 | } |
dbe08d82 ON |
2987 | mutex_unlock(&task->perf_event_mutex); |
2988 | ||
2989 | if (unlikely(err)) { | |
9137fb28 | 2990 | put_ctx(ctx); |
dbe08d82 ON |
2991 | |
2992 | if (err == -EAGAIN) | |
2993 | goto retry; | |
2994 | goto errout; | |
a63eaf34 PM |
2995 | } |
2996 | } | |
2997 | ||
0793a61d | 2998 | return ctx; |
c93f7669 | 2999 | |
9ed6060d | 3000 | errout: |
c93f7669 | 3001 | return ERR_PTR(err); |
0793a61d TG |
3002 | } |
3003 | ||
6fb2915d LZ |
3004 | static void perf_event_free_filter(struct perf_event *event); |
3005 | ||
cdd6c482 | 3006 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3007 | { |
cdd6c482 | 3008 | struct perf_event *event; |
592903cd | 3009 | |
cdd6c482 IM |
3010 | event = container_of(head, struct perf_event, rcu_head); |
3011 | if (event->ns) | |
3012 | put_pid_ns(event->ns); | |
6fb2915d | 3013 | perf_event_free_filter(event); |
cdd6c482 | 3014 | kfree(event); |
592903cd PZ |
3015 | } |
3016 | ||
76369139 | 3017 | static void ring_buffer_put(struct ring_buffer *rb); |
925d519a | 3018 | |
cdd6c482 | 3019 | static void free_event(struct perf_event *event) |
f1600952 | 3020 | { |
e360adbe | 3021 | irq_work_sync(&event->pending); |
925d519a | 3022 | |
cdd6c482 | 3023 | if (!event->parent) { |
82cd6def | 3024 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 3025 | static_key_slow_dec_deferred(&perf_sched_events); |
3af9e859 | 3026 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
3027 | atomic_dec(&nr_mmap_events); |
3028 | if (event->attr.comm) | |
3029 | atomic_dec(&nr_comm_events); | |
3030 | if (event->attr.task) | |
3031 | atomic_dec(&nr_task_events); | |
927c7a9e FW |
3032 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3033 | put_callchain_buffers(); | |
08309379 PZ |
3034 | if (is_cgroup_event(event)) { |
3035 | atomic_dec(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 3036 | static_key_slow_dec_deferred(&perf_sched_events); |
08309379 | 3037 | } |
d010b332 SE |
3038 | |
3039 | if (has_branch_stack(event)) { | |
3040 | static_key_slow_dec_deferred(&perf_sched_events); | |
3041 | /* is system-wide event */ | |
3042 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
3043 | atomic_dec(&per_cpu(perf_branch_stack_events, | |
3044 | event->cpu)); | |
3045 | } | |
f344011c | 3046 | } |
9ee318a7 | 3047 | |
76369139 FW |
3048 | if (event->rb) { |
3049 | ring_buffer_put(event->rb); | |
3050 | event->rb = NULL; | |
a4be7c27 PZ |
3051 | } |
3052 | ||
e5d1367f SE |
3053 | if (is_cgroup_event(event)) |
3054 | perf_detach_cgroup(event); | |
3055 | ||
cdd6c482 IM |
3056 | if (event->destroy) |
3057 | event->destroy(event); | |
e077df4f | 3058 | |
0c67b408 PZ |
3059 | if (event->ctx) |
3060 | put_ctx(event->ctx); | |
3061 | ||
cdd6c482 | 3062 | call_rcu(&event->rcu_head, free_event_rcu); |
f1600952 PZ |
3063 | } |
3064 | ||
a66a3052 | 3065 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 3066 | { |
cdd6c482 | 3067 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 3068 | |
ad3a37de | 3069 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
3070 | /* |
3071 | * There are two ways this annotation is useful: | |
3072 | * | |
3073 | * 1) there is a lock recursion from perf_event_exit_task | |
3074 | * see the comment there. | |
3075 | * | |
3076 | * 2) there is a lock-inversion with mmap_sem through | |
3077 | * perf_event_read_group(), which takes faults while | |
3078 | * holding ctx->mutex, however this is called after | |
3079 | * the last filedesc died, so there is no possibility | |
3080 | * to trigger the AB-BA case. | |
3081 | */ | |
3082 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
050735b0 | 3083 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 3084 | perf_group_detach(event); |
050735b0 | 3085 | raw_spin_unlock_irq(&ctx->lock); |
e03a9a55 | 3086 | perf_remove_from_context(event); |
d859e29f | 3087 | mutex_unlock(&ctx->mutex); |
0793a61d | 3088 | |
cdd6c482 | 3089 | free_event(event); |
0793a61d TG |
3090 | |
3091 | return 0; | |
3092 | } | |
a66a3052 | 3093 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 3094 | |
a66a3052 PZ |
3095 | /* |
3096 | * Called when the last reference to the file is gone. | |
3097 | */ | |
a6fa941d | 3098 | static void put_event(struct perf_event *event) |
fb0459d7 | 3099 | { |
8882135b | 3100 | struct task_struct *owner; |
fb0459d7 | 3101 | |
a6fa941d AV |
3102 | if (!atomic_long_dec_and_test(&event->refcount)) |
3103 | return; | |
fb0459d7 | 3104 | |
8882135b PZ |
3105 | rcu_read_lock(); |
3106 | owner = ACCESS_ONCE(event->owner); | |
3107 | /* | |
3108 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3109 | * !owner it means the list deletion is complete and we can indeed | |
3110 | * free this event, otherwise we need to serialize on | |
3111 | * owner->perf_event_mutex. | |
3112 | */ | |
3113 | smp_read_barrier_depends(); | |
3114 | if (owner) { | |
3115 | /* | |
3116 | * Since delayed_put_task_struct() also drops the last | |
3117 | * task reference we can safely take a new reference | |
3118 | * while holding the rcu_read_lock(). | |
3119 | */ | |
3120 | get_task_struct(owner); | |
3121 | } | |
3122 | rcu_read_unlock(); | |
3123 | ||
3124 | if (owner) { | |
3125 | mutex_lock(&owner->perf_event_mutex); | |
3126 | /* | |
3127 | * We have to re-check the event->owner field, if it is cleared | |
3128 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3129 | * ensured they're done, and we can proceed with freeing the | |
3130 | * event. | |
3131 | */ | |
3132 | if (event->owner) | |
3133 | list_del_init(&event->owner_entry); | |
3134 | mutex_unlock(&owner->perf_event_mutex); | |
3135 | put_task_struct(owner); | |
3136 | } | |
3137 | ||
a6fa941d AV |
3138 | perf_event_release_kernel(event); |
3139 | } | |
3140 | ||
3141 | static int perf_release(struct inode *inode, struct file *file) | |
3142 | { | |
3143 | put_event(file->private_data); | |
3144 | return 0; | |
fb0459d7 | 3145 | } |
fb0459d7 | 3146 | |
59ed446f | 3147 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3148 | { |
cdd6c482 | 3149 | struct perf_event *child; |
e53c0994 PZ |
3150 | u64 total = 0; |
3151 | ||
59ed446f PZ |
3152 | *enabled = 0; |
3153 | *running = 0; | |
3154 | ||
6f10581a | 3155 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3156 | total += perf_event_read(event); |
59ed446f PZ |
3157 | *enabled += event->total_time_enabled + |
3158 | atomic64_read(&event->child_total_time_enabled); | |
3159 | *running += event->total_time_running + | |
3160 | atomic64_read(&event->child_total_time_running); | |
3161 | ||
3162 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3163 | total += perf_event_read(child); |
59ed446f PZ |
3164 | *enabled += child->total_time_enabled; |
3165 | *running += child->total_time_running; | |
3166 | } | |
6f10581a | 3167 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3168 | |
3169 | return total; | |
3170 | } | |
fb0459d7 | 3171 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3172 | |
cdd6c482 | 3173 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3174 | u64 read_format, char __user *buf) |
3175 | { | |
cdd6c482 | 3176 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3177 | int n = 0, size = 0, ret = -EFAULT; |
3178 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3179 | u64 values[5]; |
59ed446f | 3180 | u64 count, enabled, running; |
abf4868b | 3181 | |
6f10581a | 3182 | mutex_lock(&ctx->mutex); |
59ed446f | 3183 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3184 | |
3185 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3186 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3187 | values[n++] = enabled; | |
3188 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3189 | values[n++] = running; | |
abf4868b PZ |
3190 | values[n++] = count; |
3191 | if (read_format & PERF_FORMAT_ID) | |
3192 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3193 | |
3194 | size = n * sizeof(u64); | |
3195 | ||
3196 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3197 | goto unlock; |
3dab77fb | 3198 | |
6f10581a | 3199 | ret = size; |
3dab77fb | 3200 | |
65abc865 | 3201 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3202 | n = 0; |
3dab77fb | 3203 | |
59ed446f | 3204 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3205 | if (read_format & PERF_FORMAT_ID) |
3206 | values[n++] = primary_event_id(sub); | |
3207 | ||
3208 | size = n * sizeof(u64); | |
3209 | ||
184d3da8 | 3210 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3211 | ret = -EFAULT; |
3212 | goto unlock; | |
3213 | } | |
abf4868b PZ |
3214 | |
3215 | ret += size; | |
3dab77fb | 3216 | } |
6f10581a PZ |
3217 | unlock: |
3218 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3219 | |
abf4868b | 3220 | return ret; |
3dab77fb PZ |
3221 | } |
3222 | ||
cdd6c482 | 3223 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3224 | u64 read_format, char __user *buf) |
3225 | { | |
59ed446f | 3226 | u64 enabled, running; |
3dab77fb PZ |
3227 | u64 values[4]; |
3228 | int n = 0; | |
3229 | ||
59ed446f PZ |
3230 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3231 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3232 | values[n++] = enabled; | |
3233 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3234 | values[n++] = running; | |
3dab77fb | 3235 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3236 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3237 | |
3238 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3239 | return -EFAULT; | |
3240 | ||
3241 | return n * sizeof(u64); | |
3242 | } | |
3243 | ||
0793a61d | 3244 | /* |
cdd6c482 | 3245 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3246 | */ |
3247 | static ssize_t | |
cdd6c482 | 3248 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3249 | { |
cdd6c482 | 3250 | u64 read_format = event->attr.read_format; |
3dab77fb | 3251 | int ret; |
0793a61d | 3252 | |
3b6f9e5c | 3253 | /* |
cdd6c482 | 3254 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3255 | * error state (i.e. because it was pinned but it couldn't be |
3256 | * scheduled on to the CPU at some point). | |
3257 | */ | |
cdd6c482 | 3258 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3259 | return 0; |
3260 | ||
c320c7b7 | 3261 | if (count < event->read_size) |
3dab77fb PZ |
3262 | return -ENOSPC; |
3263 | ||
cdd6c482 | 3264 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3265 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3266 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3267 | else |
cdd6c482 | 3268 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3269 | |
3dab77fb | 3270 | return ret; |
0793a61d TG |
3271 | } |
3272 | ||
0793a61d TG |
3273 | static ssize_t |
3274 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3275 | { | |
cdd6c482 | 3276 | struct perf_event *event = file->private_data; |
0793a61d | 3277 | |
cdd6c482 | 3278 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3279 | } |
3280 | ||
3281 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3282 | { | |
cdd6c482 | 3283 | struct perf_event *event = file->private_data; |
76369139 | 3284 | struct ring_buffer *rb; |
c33a0bc4 | 3285 | unsigned int events = POLL_HUP; |
c7138f37 | 3286 | |
10c6db11 PZ |
3287 | /* |
3288 | * Race between perf_event_set_output() and perf_poll(): perf_poll() | |
3289 | * grabs the rb reference but perf_event_set_output() overrides it. | |
3290 | * Here is the timeline for two threads T1, T2: | |
3291 | * t0: T1, rb = rcu_dereference(event->rb) | |
3292 | * t1: T2, old_rb = event->rb | |
3293 | * t2: T2, event->rb = new rb | |
3294 | * t3: T2, ring_buffer_detach(old_rb) | |
3295 | * t4: T1, ring_buffer_attach(rb1) | |
3296 | * t5: T1, poll_wait(event->waitq) | |
3297 | * | |
3298 | * To avoid this problem, we grab mmap_mutex in perf_poll() | |
3299 | * thereby ensuring that the assignment of the new ring buffer | |
3300 | * and the detachment of the old buffer appear atomic to perf_poll() | |
3301 | */ | |
3302 | mutex_lock(&event->mmap_mutex); | |
3303 | ||
c7138f37 | 3304 | rcu_read_lock(); |
76369139 | 3305 | rb = rcu_dereference(event->rb); |
10c6db11 PZ |
3306 | if (rb) { |
3307 | ring_buffer_attach(event, rb); | |
76369139 | 3308 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3309 | } |
c7138f37 | 3310 | rcu_read_unlock(); |
0793a61d | 3311 | |
10c6db11 PZ |
3312 | mutex_unlock(&event->mmap_mutex); |
3313 | ||
cdd6c482 | 3314 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3315 | |
0793a61d TG |
3316 | return events; |
3317 | } | |
3318 | ||
cdd6c482 | 3319 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3320 | { |
cdd6c482 | 3321 | (void)perf_event_read(event); |
e7850595 | 3322 | local64_set(&event->count, 0); |
cdd6c482 | 3323 | perf_event_update_userpage(event); |
3df5edad PZ |
3324 | } |
3325 | ||
c93f7669 | 3326 | /* |
cdd6c482 IM |
3327 | * Holding the top-level event's child_mutex means that any |
3328 | * descendant process that has inherited this event will block | |
3329 | * in sync_child_event if it goes to exit, thus satisfying the | |
3330 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3331 | */ |
cdd6c482 IM |
3332 | static void perf_event_for_each_child(struct perf_event *event, |
3333 | void (*func)(struct perf_event *)) | |
3df5edad | 3334 | { |
cdd6c482 | 3335 | struct perf_event *child; |
3df5edad | 3336 | |
cdd6c482 IM |
3337 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3338 | mutex_lock(&event->child_mutex); | |
3339 | func(event); | |
3340 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3341 | func(child); |
cdd6c482 | 3342 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3343 | } |
3344 | ||
cdd6c482 IM |
3345 | static void perf_event_for_each(struct perf_event *event, |
3346 | void (*func)(struct perf_event *)) | |
3df5edad | 3347 | { |
cdd6c482 IM |
3348 | struct perf_event_context *ctx = event->ctx; |
3349 | struct perf_event *sibling; | |
3df5edad | 3350 | |
75f937f2 PZ |
3351 | WARN_ON_ONCE(ctx->parent_ctx); |
3352 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3353 | event = event->group_leader; |
75f937f2 | 3354 | |
cdd6c482 | 3355 | perf_event_for_each_child(event, func); |
cdd6c482 | 3356 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3357 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3358 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3359 | } |
3360 | ||
cdd6c482 | 3361 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3362 | { |
cdd6c482 | 3363 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3364 | int ret = 0; |
3365 | u64 value; | |
3366 | ||
6c7e550f | 3367 | if (!is_sampling_event(event)) |
08247e31 PZ |
3368 | return -EINVAL; |
3369 | ||
ad0cf347 | 3370 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3371 | return -EFAULT; |
3372 | ||
3373 | if (!value) | |
3374 | return -EINVAL; | |
3375 | ||
e625cce1 | 3376 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3377 | if (event->attr.freq) { |
3378 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3379 | ret = -EINVAL; |
3380 | goto unlock; | |
3381 | } | |
3382 | ||
cdd6c482 | 3383 | event->attr.sample_freq = value; |
08247e31 | 3384 | } else { |
cdd6c482 IM |
3385 | event->attr.sample_period = value; |
3386 | event->hw.sample_period = value; | |
08247e31 PZ |
3387 | } |
3388 | unlock: | |
e625cce1 | 3389 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3390 | |
3391 | return ret; | |
3392 | } | |
3393 | ||
ac9721f3 PZ |
3394 | static const struct file_operations perf_fops; |
3395 | ||
2903ff01 | 3396 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3397 | { |
2903ff01 AV |
3398 | struct fd f = fdget(fd); |
3399 | if (!f.file) | |
3400 | return -EBADF; | |
ac9721f3 | 3401 | |
2903ff01 AV |
3402 | if (f.file->f_op != &perf_fops) { |
3403 | fdput(f); | |
3404 | return -EBADF; | |
ac9721f3 | 3405 | } |
2903ff01 AV |
3406 | *p = f; |
3407 | return 0; | |
ac9721f3 PZ |
3408 | } |
3409 | ||
3410 | static int perf_event_set_output(struct perf_event *event, | |
3411 | struct perf_event *output_event); | |
6fb2915d | 3412 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3413 | |
d859e29f PM |
3414 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3415 | { | |
cdd6c482 IM |
3416 | struct perf_event *event = file->private_data; |
3417 | void (*func)(struct perf_event *); | |
3df5edad | 3418 | u32 flags = arg; |
d859e29f PM |
3419 | |
3420 | switch (cmd) { | |
cdd6c482 IM |
3421 | case PERF_EVENT_IOC_ENABLE: |
3422 | func = perf_event_enable; | |
d859e29f | 3423 | break; |
cdd6c482 IM |
3424 | case PERF_EVENT_IOC_DISABLE: |
3425 | func = perf_event_disable; | |
79f14641 | 3426 | break; |
cdd6c482 IM |
3427 | case PERF_EVENT_IOC_RESET: |
3428 | func = perf_event_reset; | |
6de6a7b9 | 3429 | break; |
3df5edad | 3430 | |
cdd6c482 IM |
3431 | case PERF_EVENT_IOC_REFRESH: |
3432 | return perf_event_refresh(event, arg); | |
08247e31 | 3433 | |
cdd6c482 IM |
3434 | case PERF_EVENT_IOC_PERIOD: |
3435 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3436 | |
cdd6c482 | 3437 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3438 | { |
ac9721f3 | 3439 | int ret; |
ac9721f3 | 3440 | if (arg != -1) { |
2903ff01 AV |
3441 | struct perf_event *output_event; |
3442 | struct fd output; | |
3443 | ret = perf_fget_light(arg, &output); | |
3444 | if (ret) | |
3445 | return ret; | |
3446 | output_event = output.file->private_data; | |
3447 | ret = perf_event_set_output(event, output_event); | |
3448 | fdput(output); | |
3449 | } else { | |
3450 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3451 | } |
ac9721f3 PZ |
3452 | return ret; |
3453 | } | |
a4be7c27 | 3454 | |
6fb2915d LZ |
3455 | case PERF_EVENT_IOC_SET_FILTER: |
3456 | return perf_event_set_filter(event, (void __user *)arg); | |
3457 | ||
d859e29f | 3458 | default: |
3df5edad | 3459 | return -ENOTTY; |
d859e29f | 3460 | } |
3df5edad PZ |
3461 | |
3462 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3463 | perf_event_for_each(event, func); |
3df5edad | 3464 | else |
cdd6c482 | 3465 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3466 | |
3467 | return 0; | |
d859e29f PM |
3468 | } |
3469 | ||
cdd6c482 | 3470 | int perf_event_task_enable(void) |
771d7cde | 3471 | { |
cdd6c482 | 3472 | struct perf_event *event; |
771d7cde | 3473 | |
cdd6c482 IM |
3474 | mutex_lock(¤t->perf_event_mutex); |
3475 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3476 | perf_event_for_each_child(event, perf_event_enable); | |
3477 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3478 | |
3479 | return 0; | |
3480 | } | |
3481 | ||
cdd6c482 | 3482 | int perf_event_task_disable(void) |
771d7cde | 3483 | { |
cdd6c482 | 3484 | struct perf_event *event; |
771d7cde | 3485 | |
cdd6c482 IM |
3486 | mutex_lock(¤t->perf_event_mutex); |
3487 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3488 | perf_event_for_each_child(event, perf_event_disable); | |
3489 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3490 | |
3491 | return 0; | |
3492 | } | |
3493 | ||
cdd6c482 | 3494 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3495 | { |
a4eaf7f1 PZ |
3496 | if (event->hw.state & PERF_HES_STOPPED) |
3497 | return 0; | |
3498 | ||
cdd6c482 | 3499 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3500 | return 0; |
3501 | ||
35edc2a5 | 3502 | return event->pmu->event_idx(event); |
194002b2 PZ |
3503 | } |
3504 | ||
c4794295 | 3505 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3506 | u64 *now, |
7f310a5d EM |
3507 | u64 *enabled, |
3508 | u64 *running) | |
c4794295 | 3509 | { |
e3f3541c | 3510 | u64 ctx_time; |
c4794295 | 3511 | |
e3f3541c PZ |
3512 | *now = perf_clock(); |
3513 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3514 | *enabled = ctx_time - event->tstamp_enabled; |
3515 | *running = ctx_time - event->tstamp_running; | |
3516 | } | |
3517 | ||
c7206205 | 3518 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3519 | { |
3520 | } | |
3521 | ||
38ff667b PZ |
3522 | /* |
3523 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3524 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3525 | * code calls this from NMI context. | |
3526 | */ | |
cdd6c482 | 3527 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3528 | { |
cdd6c482 | 3529 | struct perf_event_mmap_page *userpg; |
76369139 | 3530 | struct ring_buffer *rb; |
e3f3541c | 3531 | u64 enabled, running, now; |
38ff667b PZ |
3532 | |
3533 | rcu_read_lock(); | |
0d641208 EM |
3534 | /* |
3535 | * compute total_time_enabled, total_time_running | |
3536 | * based on snapshot values taken when the event | |
3537 | * was last scheduled in. | |
3538 | * | |
3539 | * we cannot simply called update_context_time() | |
3540 | * because of locking issue as we can be called in | |
3541 | * NMI context | |
3542 | */ | |
e3f3541c | 3543 | calc_timer_values(event, &now, &enabled, &running); |
76369139 FW |
3544 | rb = rcu_dereference(event->rb); |
3545 | if (!rb) | |
38ff667b PZ |
3546 | goto unlock; |
3547 | ||
76369139 | 3548 | userpg = rb->user_page; |
37d81828 | 3549 | |
7b732a75 PZ |
3550 | /* |
3551 | * Disable preemption so as to not let the corresponding user-space | |
3552 | * spin too long if we get preempted. | |
3553 | */ | |
3554 | preempt_disable(); | |
37d81828 | 3555 | ++userpg->lock; |
92f22a38 | 3556 | barrier(); |
cdd6c482 | 3557 | userpg->index = perf_event_index(event); |
b5e58793 | 3558 | userpg->offset = perf_event_count(event); |
365a4038 | 3559 | if (userpg->index) |
e7850595 | 3560 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3561 | |
0d641208 | 3562 | userpg->time_enabled = enabled + |
cdd6c482 | 3563 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3564 | |
0d641208 | 3565 | userpg->time_running = running + |
cdd6c482 | 3566 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3567 | |
c7206205 | 3568 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3569 | |
92f22a38 | 3570 | barrier(); |
37d81828 | 3571 | ++userpg->lock; |
7b732a75 | 3572 | preempt_enable(); |
38ff667b | 3573 | unlock: |
7b732a75 | 3574 | rcu_read_unlock(); |
37d81828 PM |
3575 | } |
3576 | ||
906010b2 PZ |
3577 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3578 | { | |
3579 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3580 | struct ring_buffer *rb; |
906010b2 PZ |
3581 | int ret = VM_FAULT_SIGBUS; |
3582 | ||
3583 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3584 | if (vmf->pgoff == 0) | |
3585 | ret = 0; | |
3586 | return ret; | |
3587 | } | |
3588 | ||
3589 | rcu_read_lock(); | |
76369139 FW |
3590 | rb = rcu_dereference(event->rb); |
3591 | if (!rb) | |
906010b2 PZ |
3592 | goto unlock; |
3593 | ||
3594 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3595 | goto unlock; | |
3596 | ||
76369139 | 3597 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3598 | if (!vmf->page) |
3599 | goto unlock; | |
3600 | ||
3601 | get_page(vmf->page); | |
3602 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3603 | vmf->page->index = vmf->pgoff; | |
3604 | ||
3605 | ret = 0; | |
3606 | unlock: | |
3607 | rcu_read_unlock(); | |
3608 | ||
3609 | return ret; | |
3610 | } | |
3611 | ||
10c6db11 PZ |
3612 | static void ring_buffer_attach(struct perf_event *event, |
3613 | struct ring_buffer *rb) | |
3614 | { | |
3615 | unsigned long flags; | |
3616 | ||
3617 | if (!list_empty(&event->rb_entry)) | |
3618 | return; | |
3619 | ||
3620 | spin_lock_irqsave(&rb->event_lock, flags); | |
3621 | if (!list_empty(&event->rb_entry)) | |
3622 | goto unlock; | |
3623 | ||
3624 | list_add(&event->rb_entry, &rb->event_list); | |
3625 | unlock: | |
3626 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3627 | } | |
3628 | ||
3629 | static void ring_buffer_detach(struct perf_event *event, | |
3630 | struct ring_buffer *rb) | |
3631 | { | |
3632 | unsigned long flags; | |
3633 | ||
3634 | if (list_empty(&event->rb_entry)) | |
3635 | return; | |
3636 | ||
3637 | spin_lock_irqsave(&rb->event_lock, flags); | |
3638 | list_del_init(&event->rb_entry); | |
3639 | wake_up_all(&event->waitq); | |
3640 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3641 | } | |
3642 | ||
3643 | static void ring_buffer_wakeup(struct perf_event *event) | |
3644 | { | |
3645 | struct ring_buffer *rb; | |
3646 | ||
3647 | rcu_read_lock(); | |
3648 | rb = rcu_dereference(event->rb); | |
44b7f4b9 WD |
3649 | if (!rb) |
3650 | goto unlock; | |
3651 | ||
3652 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
10c6db11 | 3653 | wake_up_all(&event->waitq); |
44b7f4b9 WD |
3654 | |
3655 | unlock: | |
10c6db11 PZ |
3656 | rcu_read_unlock(); |
3657 | } | |
3658 | ||
76369139 | 3659 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3660 | { |
76369139 | 3661 | struct ring_buffer *rb; |
906010b2 | 3662 | |
76369139 FW |
3663 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3664 | rb_free(rb); | |
7b732a75 PZ |
3665 | } |
3666 | ||
76369139 | 3667 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3668 | { |
76369139 | 3669 | struct ring_buffer *rb; |
7b732a75 | 3670 | |
ac9721f3 | 3671 | rcu_read_lock(); |
76369139 FW |
3672 | rb = rcu_dereference(event->rb); |
3673 | if (rb) { | |
3674 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3675 | rb = NULL; | |
ac9721f3 PZ |
3676 | } |
3677 | rcu_read_unlock(); | |
3678 | ||
76369139 | 3679 | return rb; |
ac9721f3 PZ |
3680 | } |
3681 | ||
76369139 | 3682 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3683 | { |
10c6db11 PZ |
3684 | struct perf_event *event, *n; |
3685 | unsigned long flags; | |
3686 | ||
76369139 | 3687 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3688 | return; |
7b732a75 | 3689 | |
10c6db11 PZ |
3690 | spin_lock_irqsave(&rb->event_lock, flags); |
3691 | list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) { | |
3692 | list_del_init(&event->rb_entry); | |
3693 | wake_up_all(&event->waitq); | |
3694 | } | |
3695 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3696 | ||
76369139 | 3697 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3698 | } |
3699 | ||
3700 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3701 | { | |
cdd6c482 | 3702 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3703 | |
cdd6c482 | 3704 | atomic_inc(&event->mmap_count); |
7b732a75 PZ |
3705 | } |
3706 | ||
3707 | static void perf_mmap_close(struct vm_area_struct *vma) | |
3708 | { | |
cdd6c482 | 3709 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3710 | |
cdd6c482 | 3711 | if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) { |
76369139 | 3712 | unsigned long size = perf_data_size(event->rb); |
ac9721f3 | 3713 | struct user_struct *user = event->mmap_user; |
76369139 | 3714 | struct ring_buffer *rb = event->rb; |
789f90fc | 3715 | |
906010b2 | 3716 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm); |
bc3e53f6 | 3717 | vma->vm_mm->pinned_vm -= event->mmap_locked; |
76369139 | 3718 | rcu_assign_pointer(event->rb, NULL); |
10c6db11 | 3719 | ring_buffer_detach(event, rb); |
cdd6c482 | 3720 | mutex_unlock(&event->mmap_mutex); |
ac9721f3 | 3721 | |
76369139 | 3722 | ring_buffer_put(rb); |
ac9721f3 | 3723 | free_uid(user); |
7b732a75 | 3724 | } |
37d81828 PM |
3725 | } |
3726 | ||
f0f37e2f | 3727 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3728 | .open = perf_mmap_open, |
3729 | .close = perf_mmap_close, | |
3730 | .fault = perf_mmap_fault, | |
3731 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3732 | }; |
3733 | ||
3734 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3735 | { | |
cdd6c482 | 3736 | struct perf_event *event = file->private_data; |
22a4f650 | 3737 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3738 | struct user_struct *user = current_user(); |
22a4f650 | 3739 | unsigned long locked, lock_limit; |
76369139 | 3740 | struct ring_buffer *rb; |
7b732a75 PZ |
3741 | unsigned long vma_size; |
3742 | unsigned long nr_pages; | |
789f90fc | 3743 | long user_extra, extra; |
d57e34fd | 3744 | int ret = 0, flags = 0; |
37d81828 | 3745 | |
c7920614 PZ |
3746 | /* |
3747 | * Don't allow mmap() of inherited per-task counters. This would | |
3748 | * create a performance issue due to all children writing to the | |
76369139 | 3749 | * same rb. |
c7920614 PZ |
3750 | */ |
3751 | if (event->cpu == -1 && event->attr.inherit) | |
3752 | return -EINVAL; | |
3753 | ||
43a21ea8 | 3754 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3755 | return -EINVAL; |
7b732a75 PZ |
3756 | |
3757 | vma_size = vma->vm_end - vma->vm_start; | |
3758 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3759 | ||
7730d865 | 3760 | /* |
76369139 | 3761 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3762 | * can do bitmasks instead of modulo. |
3763 | */ | |
3764 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3765 | return -EINVAL; |
3766 | ||
7b732a75 | 3767 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3768 | return -EINVAL; |
3769 | ||
7b732a75 PZ |
3770 | if (vma->vm_pgoff != 0) |
3771 | return -EINVAL; | |
37d81828 | 3772 | |
cdd6c482 IM |
3773 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3774 | mutex_lock(&event->mmap_mutex); | |
76369139 FW |
3775 | if (event->rb) { |
3776 | if (event->rb->nr_pages == nr_pages) | |
3777 | atomic_inc(&event->rb->refcount); | |
ac9721f3 | 3778 | else |
ebb3c4c4 PZ |
3779 | ret = -EINVAL; |
3780 | goto unlock; | |
3781 | } | |
3782 | ||
789f90fc | 3783 | user_extra = nr_pages + 1; |
cdd6c482 | 3784 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
3785 | |
3786 | /* | |
3787 | * Increase the limit linearly with more CPUs: | |
3788 | */ | |
3789 | user_lock_limit *= num_online_cpus(); | |
3790 | ||
789f90fc | 3791 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 3792 | |
789f90fc PZ |
3793 | extra = 0; |
3794 | if (user_locked > user_lock_limit) | |
3795 | extra = user_locked - user_lock_limit; | |
7b732a75 | 3796 | |
78d7d407 | 3797 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 3798 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 3799 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 3800 | |
459ec28a IM |
3801 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
3802 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
3803 | ret = -EPERM; |
3804 | goto unlock; | |
3805 | } | |
7b732a75 | 3806 | |
76369139 | 3807 | WARN_ON(event->rb); |
906010b2 | 3808 | |
d57e34fd | 3809 | if (vma->vm_flags & VM_WRITE) |
76369139 | 3810 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 3811 | |
4ec8363d VW |
3812 | rb = rb_alloc(nr_pages, |
3813 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
3814 | event->cpu, flags); | |
3815 | ||
76369139 | 3816 | if (!rb) { |
ac9721f3 | 3817 | ret = -ENOMEM; |
ebb3c4c4 | 3818 | goto unlock; |
ac9721f3 | 3819 | } |
76369139 | 3820 | rcu_assign_pointer(event->rb, rb); |
43a21ea8 | 3821 | |
ac9721f3 PZ |
3822 | atomic_long_add(user_extra, &user->locked_vm); |
3823 | event->mmap_locked = extra; | |
3824 | event->mmap_user = get_current_user(); | |
bc3e53f6 | 3825 | vma->vm_mm->pinned_vm += event->mmap_locked; |
ac9721f3 | 3826 | |
9a0f05cb PZ |
3827 | perf_event_update_userpage(event); |
3828 | ||
ebb3c4c4 | 3829 | unlock: |
ac9721f3 PZ |
3830 | if (!ret) |
3831 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 3832 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 3833 | |
314e51b9 | 3834 | vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 3835 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
3836 | |
3837 | return ret; | |
37d81828 PM |
3838 | } |
3839 | ||
3c446b3d PZ |
3840 | static int perf_fasync(int fd, struct file *filp, int on) |
3841 | { | |
496ad9aa | 3842 | struct inode *inode = file_inode(filp); |
cdd6c482 | 3843 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
3844 | int retval; |
3845 | ||
3846 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 3847 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
3848 | mutex_unlock(&inode->i_mutex); |
3849 | ||
3850 | if (retval < 0) | |
3851 | return retval; | |
3852 | ||
3853 | return 0; | |
3854 | } | |
3855 | ||
0793a61d | 3856 | static const struct file_operations perf_fops = { |
3326c1ce | 3857 | .llseek = no_llseek, |
0793a61d TG |
3858 | .release = perf_release, |
3859 | .read = perf_read, | |
3860 | .poll = perf_poll, | |
d859e29f PM |
3861 | .unlocked_ioctl = perf_ioctl, |
3862 | .compat_ioctl = perf_ioctl, | |
37d81828 | 3863 | .mmap = perf_mmap, |
3c446b3d | 3864 | .fasync = perf_fasync, |
0793a61d TG |
3865 | }; |
3866 | ||
925d519a | 3867 | /* |
cdd6c482 | 3868 | * Perf event wakeup |
925d519a PZ |
3869 | * |
3870 | * If there's data, ensure we set the poll() state and publish everything | |
3871 | * to user-space before waking everybody up. | |
3872 | */ | |
3873 | ||
cdd6c482 | 3874 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 3875 | { |
10c6db11 | 3876 | ring_buffer_wakeup(event); |
4c9e2542 | 3877 | |
cdd6c482 IM |
3878 | if (event->pending_kill) { |
3879 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
3880 | event->pending_kill = 0; | |
4c9e2542 | 3881 | } |
925d519a PZ |
3882 | } |
3883 | ||
e360adbe | 3884 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 3885 | { |
cdd6c482 IM |
3886 | struct perf_event *event = container_of(entry, |
3887 | struct perf_event, pending); | |
79f14641 | 3888 | |
cdd6c482 IM |
3889 | if (event->pending_disable) { |
3890 | event->pending_disable = 0; | |
3891 | __perf_event_disable(event); | |
79f14641 PZ |
3892 | } |
3893 | ||
cdd6c482 IM |
3894 | if (event->pending_wakeup) { |
3895 | event->pending_wakeup = 0; | |
3896 | perf_event_wakeup(event); | |
79f14641 PZ |
3897 | } |
3898 | } | |
3899 | ||
39447b38 ZY |
3900 | /* |
3901 | * We assume there is only KVM supporting the callbacks. | |
3902 | * Later on, we might change it to a list if there is | |
3903 | * another virtualization implementation supporting the callbacks. | |
3904 | */ | |
3905 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
3906 | ||
3907 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3908 | { | |
3909 | perf_guest_cbs = cbs; | |
3910 | return 0; | |
3911 | } | |
3912 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
3913 | ||
3914 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3915 | { | |
3916 | perf_guest_cbs = NULL; | |
3917 | return 0; | |
3918 | } | |
3919 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
3920 | ||
4018994f JO |
3921 | static void |
3922 | perf_output_sample_regs(struct perf_output_handle *handle, | |
3923 | struct pt_regs *regs, u64 mask) | |
3924 | { | |
3925 | int bit; | |
3926 | ||
3927 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
3928 | sizeof(mask) * BITS_PER_BYTE) { | |
3929 | u64 val; | |
3930 | ||
3931 | val = perf_reg_value(regs, bit); | |
3932 | perf_output_put(handle, val); | |
3933 | } | |
3934 | } | |
3935 | ||
3936 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
3937 | struct pt_regs *regs) | |
3938 | { | |
3939 | if (!user_mode(regs)) { | |
3940 | if (current->mm) | |
3941 | regs = task_pt_regs(current); | |
3942 | else | |
3943 | regs = NULL; | |
3944 | } | |
3945 | ||
3946 | if (regs) { | |
3947 | regs_user->regs = regs; | |
3948 | regs_user->abi = perf_reg_abi(current); | |
3949 | } | |
3950 | } | |
3951 | ||
c5ebcedb JO |
3952 | /* |
3953 | * Get remaining task size from user stack pointer. | |
3954 | * | |
3955 | * It'd be better to take stack vma map and limit this more | |
3956 | * precisly, but there's no way to get it safely under interrupt, | |
3957 | * so using TASK_SIZE as limit. | |
3958 | */ | |
3959 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
3960 | { | |
3961 | unsigned long addr = perf_user_stack_pointer(regs); | |
3962 | ||
3963 | if (!addr || addr >= TASK_SIZE) | |
3964 | return 0; | |
3965 | ||
3966 | return TASK_SIZE - addr; | |
3967 | } | |
3968 | ||
3969 | static u16 | |
3970 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
3971 | struct pt_regs *regs) | |
3972 | { | |
3973 | u64 task_size; | |
3974 | ||
3975 | /* No regs, no stack pointer, no dump. */ | |
3976 | if (!regs) | |
3977 | return 0; | |
3978 | ||
3979 | /* | |
3980 | * Check if we fit in with the requested stack size into the: | |
3981 | * - TASK_SIZE | |
3982 | * If we don't, we limit the size to the TASK_SIZE. | |
3983 | * | |
3984 | * - remaining sample size | |
3985 | * If we don't, we customize the stack size to | |
3986 | * fit in to the remaining sample size. | |
3987 | */ | |
3988 | ||
3989 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
3990 | stack_size = min(stack_size, (u16) task_size); | |
3991 | ||
3992 | /* Current header size plus static size and dynamic size. */ | |
3993 | header_size += 2 * sizeof(u64); | |
3994 | ||
3995 | /* Do we fit in with the current stack dump size? */ | |
3996 | if ((u16) (header_size + stack_size) < header_size) { | |
3997 | /* | |
3998 | * If we overflow the maximum size for the sample, | |
3999 | * we customize the stack dump size to fit in. | |
4000 | */ | |
4001 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4002 | stack_size = round_up(stack_size, sizeof(u64)); | |
4003 | } | |
4004 | ||
4005 | return stack_size; | |
4006 | } | |
4007 | ||
4008 | static void | |
4009 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4010 | struct pt_regs *regs) | |
4011 | { | |
4012 | /* Case of a kernel thread, nothing to dump */ | |
4013 | if (!regs) { | |
4014 | u64 size = 0; | |
4015 | perf_output_put(handle, size); | |
4016 | } else { | |
4017 | unsigned long sp; | |
4018 | unsigned int rem; | |
4019 | u64 dyn_size; | |
4020 | ||
4021 | /* | |
4022 | * We dump: | |
4023 | * static size | |
4024 | * - the size requested by user or the best one we can fit | |
4025 | * in to the sample max size | |
4026 | * data | |
4027 | * - user stack dump data | |
4028 | * dynamic size | |
4029 | * - the actual dumped size | |
4030 | */ | |
4031 | ||
4032 | /* Static size. */ | |
4033 | perf_output_put(handle, dump_size); | |
4034 | ||
4035 | /* Data. */ | |
4036 | sp = perf_user_stack_pointer(regs); | |
4037 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4038 | dyn_size = dump_size - rem; | |
4039 | ||
4040 | perf_output_skip(handle, rem); | |
4041 | ||
4042 | /* Dynamic size. */ | |
4043 | perf_output_put(handle, dyn_size); | |
4044 | } | |
4045 | } | |
4046 | ||
c980d109 ACM |
4047 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4048 | struct perf_sample_data *data, | |
4049 | struct perf_event *event) | |
6844c09d ACM |
4050 | { |
4051 | u64 sample_type = event->attr.sample_type; | |
4052 | ||
4053 | data->type = sample_type; | |
4054 | header->size += event->id_header_size; | |
4055 | ||
4056 | if (sample_type & PERF_SAMPLE_TID) { | |
4057 | /* namespace issues */ | |
4058 | data->tid_entry.pid = perf_event_pid(event, current); | |
4059 | data->tid_entry.tid = perf_event_tid(event, current); | |
4060 | } | |
4061 | ||
4062 | if (sample_type & PERF_SAMPLE_TIME) | |
4063 | data->time = perf_clock(); | |
4064 | ||
4065 | if (sample_type & PERF_SAMPLE_ID) | |
4066 | data->id = primary_event_id(event); | |
4067 | ||
4068 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4069 | data->stream_id = event->id; | |
4070 | ||
4071 | if (sample_type & PERF_SAMPLE_CPU) { | |
4072 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4073 | data->cpu_entry.reserved = 0; | |
4074 | } | |
4075 | } | |
4076 | ||
76369139 FW |
4077 | void perf_event_header__init_id(struct perf_event_header *header, |
4078 | struct perf_sample_data *data, | |
4079 | struct perf_event *event) | |
c980d109 ACM |
4080 | { |
4081 | if (event->attr.sample_id_all) | |
4082 | __perf_event_header__init_id(header, data, event); | |
4083 | } | |
4084 | ||
4085 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4086 | struct perf_sample_data *data) | |
4087 | { | |
4088 | u64 sample_type = data->type; | |
4089 | ||
4090 | if (sample_type & PERF_SAMPLE_TID) | |
4091 | perf_output_put(handle, data->tid_entry); | |
4092 | ||
4093 | if (sample_type & PERF_SAMPLE_TIME) | |
4094 | perf_output_put(handle, data->time); | |
4095 | ||
4096 | if (sample_type & PERF_SAMPLE_ID) | |
4097 | perf_output_put(handle, data->id); | |
4098 | ||
4099 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4100 | perf_output_put(handle, data->stream_id); | |
4101 | ||
4102 | if (sample_type & PERF_SAMPLE_CPU) | |
4103 | perf_output_put(handle, data->cpu_entry); | |
4104 | } | |
4105 | ||
76369139 FW |
4106 | void perf_event__output_id_sample(struct perf_event *event, |
4107 | struct perf_output_handle *handle, | |
4108 | struct perf_sample_data *sample) | |
c980d109 ACM |
4109 | { |
4110 | if (event->attr.sample_id_all) | |
4111 | __perf_event__output_id_sample(handle, sample); | |
4112 | } | |
4113 | ||
3dab77fb | 4114 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4115 | struct perf_event *event, |
4116 | u64 enabled, u64 running) | |
3dab77fb | 4117 | { |
cdd6c482 | 4118 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4119 | u64 values[4]; |
4120 | int n = 0; | |
4121 | ||
b5e58793 | 4122 | values[n++] = perf_event_count(event); |
3dab77fb | 4123 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4124 | values[n++] = enabled + |
cdd6c482 | 4125 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4126 | } |
4127 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4128 | values[n++] = running + |
cdd6c482 | 4129 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4130 | } |
4131 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4132 | values[n++] = primary_event_id(event); |
3dab77fb | 4133 | |
76369139 | 4134 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4135 | } |
4136 | ||
4137 | /* | |
cdd6c482 | 4138 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4139 | */ |
4140 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4141 | struct perf_event *event, |
4142 | u64 enabled, u64 running) | |
3dab77fb | 4143 | { |
cdd6c482 IM |
4144 | struct perf_event *leader = event->group_leader, *sub; |
4145 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4146 | u64 values[5]; |
4147 | int n = 0; | |
4148 | ||
4149 | values[n++] = 1 + leader->nr_siblings; | |
4150 | ||
4151 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4152 | values[n++] = enabled; |
3dab77fb PZ |
4153 | |
4154 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4155 | values[n++] = running; |
3dab77fb | 4156 | |
cdd6c482 | 4157 | if (leader != event) |
3dab77fb PZ |
4158 | leader->pmu->read(leader); |
4159 | ||
b5e58793 | 4160 | values[n++] = perf_event_count(leader); |
3dab77fb | 4161 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4162 | values[n++] = primary_event_id(leader); |
3dab77fb | 4163 | |
76369139 | 4164 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4165 | |
65abc865 | 4166 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4167 | n = 0; |
4168 | ||
cdd6c482 | 4169 | if (sub != event) |
3dab77fb PZ |
4170 | sub->pmu->read(sub); |
4171 | ||
b5e58793 | 4172 | values[n++] = perf_event_count(sub); |
3dab77fb | 4173 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4174 | values[n++] = primary_event_id(sub); |
3dab77fb | 4175 | |
76369139 | 4176 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4177 | } |
4178 | } | |
4179 | ||
eed01528 SE |
4180 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4181 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4182 | ||
3dab77fb | 4183 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4184 | struct perf_event *event) |
3dab77fb | 4185 | { |
e3f3541c | 4186 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4187 | u64 read_format = event->attr.read_format; |
4188 | ||
4189 | /* | |
4190 | * compute total_time_enabled, total_time_running | |
4191 | * based on snapshot values taken when the event | |
4192 | * was last scheduled in. | |
4193 | * | |
4194 | * we cannot simply called update_context_time() | |
4195 | * because of locking issue as we are called in | |
4196 | * NMI context | |
4197 | */ | |
c4794295 | 4198 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4199 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4200 | |
cdd6c482 | 4201 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4202 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4203 | else |
eed01528 | 4204 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4205 | } |
4206 | ||
5622f295 MM |
4207 | void perf_output_sample(struct perf_output_handle *handle, |
4208 | struct perf_event_header *header, | |
4209 | struct perf_sample_data *data, | |
cdd6c482 | 4210 | struct perf_event *event) |
5622f295 MM |
4211 | { |
4212 | u64 sample_type = data->type; | |
4213 | ||
4214 | perf_output_put(handle, *header); | |
4215 | ||
4216 | if (sample_type & PERF_SAMPLE_IP) | |
4217 | perf_output_put(handle, data->ip); | |
4218 | ||
4219 | if (sample_type & PERF_SAMPLE_TID) | |
4220 | perf_output_put(handle, data->tid_entry); | |
4221 | ||
4222 | if (sample_type & PERF_SAMPLE_TIME) | |
4223 | perf_output_put(handle, data->time); | |
4224 | ||
4225 | if (sample_type & PERF_SAMPLE_ADDR) | |
4226 | perf_output_put(handle, data->addr); | |
4227 | ||
4228 | if (sample_type & PERF_SAMPLE_ID) | |
4229 | perf_output_put(handle, data->id); | |
4230 | ||
4231 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4232 | perf_output_put(handle, data->stream_id); | |
4233 | ||
4234 | if (sample_type & PERF_SAMPLE_CPU) | |
4235 | perf_output_put(handle, data->cpu_entry); | |
4236 | ||
4237 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4238 | perf_output_put(handle, data->period); | |
4239 | ||
4240 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4241 | perf_output_read(handle, event); |
5622f295 MM |
4242 | |
4243 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4244 | if (data->callchain) { | |
4245 | int size = 1; | |
4246 | ||
4247 | if (data->callchain) | |
4248 | size += data->callchain->nr; | |
4249 | ||
4250 | size *= sizeof(u64); | |
4251 | ||
76369139 | 4252 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4253 | } else { |
4254 | u64 nr = 0; | |
4255 | perf_output_put(handle, nr); | |
4256 | } | |
4257 | } | |
4258 | ||
4259 | if (sample_type & PERF_SAMPLE_RAW) { | |
4260 | if (data->raw) { | |
4261 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4262 | __output_copy(handle, data->raw->data, |
4263 | data->raw->size); | |
5622f295 MM |
4264 | } else { |
4265 | struct { | |
4266 | u32 size; | |
4267 | u32 data; | |
4268 | } raw = { | |
4269 | .size = sizeof(u32), | |
4270 | .data = 0, | |
4271 | }; | |
4272 | perf_output_put(handle, raw); | |
4273 | } | |
4274 | } | |
a7ac67ea PZ |
4275 | |
4276 | if (!event->attr.watermark) { | |
4277 | int wakeup_events = event->attr.wakeup_events; | |
4278 | ||
4279 | if (wakeup_events) { | |
4280 | struct ring_buffer *rb = handle->rb; | |
4281 | int events = local_inc_return(&rb->events); | |
4282 | ||
4283 | if (events >= wakeup_events) { | |
4284 | local_sub(wakeup_events, &rb->events); | |
4285 | local_inc(&rb->wakeup); | |
4286 | } | |
4287 | } | |
4288 | } | |
bce38cd5 SE |
4289 | |
4290 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4291 | if (data->br_stack) { | |
4292 | size_t size; | |
4293 | ||
4294 | size = data->br_stack->nr | |
4295 | * sizeof(struct perf_branch_entry); | |
4296 | ||
4297 | perf_output_put(handle, data->br_stack->nr); | |
4298 | perf_output_copy(handle, data->br_stack->entries, size); | |
4299 | } else { | |
4300 | /* | |
4301 | * we always store at least the value of nr | |
4302 | */ | |
4303 | u64 nr = 0; | |
4304 | perf_output_put(handle, nr); | |
4305 | } | |
4306 | } | |
4018994f JO |
4307 | |
4308 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4309 | u64 abi = data->regs_user.abi; | |
4310 | ||
4311 | /* | |
4312 | * If there are no regs to dump, notice it through | |
4313 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4314 | */ | |
4315 | perf_output_put(handle, abi); | |
4316 | ||
4317 | if (abi) { | |
4318 | u64 mask = event->attr.sample_regs_user; | |
4319 | perf_output_sample_regs(handle, | |
4320 | data->regs_user.regs, | |
4321 | mask); | |
4322 | } | |
4323 | } | |
c5ebcedb JO |
4324 | |
4325 | if (sample_type & PERF_SAMPLE_STACK_USER) | |
4326 | perf_output_sample_ustack(handle, | |
4327 | data->stack_user_size, | |
4328 | data->regs_user.regs); | |
c3feedf2 AK |
4329 | |
4330 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
4331 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
4332 | |
4333 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
4334 | perf_output_put(handle, data->data_src.val); | |
5622f295 MM |
4335 | } |
4336 | ||
4337 | void perf_prepare_sample(struct perf_event_header *header, | |
4338 | struct perf_sample_data *data, | |
cdd6c482 | 4339 | struct perf_event *event, |
5622f295 | 4340 | struct pt_regs *regs) |
7b732a75 | 4341 | { |
cdd6c482 | 4342 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4343 | |
cdd6c482 | 4344 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4345 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4346 | |
4347 | header->misc = 0; | |
4348 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4349 | |
c980d109 | 4350 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4351 | |
c320c7b7 | 4352 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4353 | data->ip = perf_instruction_pointer(regs); |
4354 | ||
b23f3325 | 4355 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4356 | int size = 1; |
394ee076 | 4357 | |
e6dab5ff | 4358 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4359 | |
4360 | if (data->callchain) | |
4361 | size += data->callchain->nr; | |
4362 | ||
4363 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4364 | } |
4365 | ||
3a43ce68 | 4366 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4367 | int size = sizeof(u32); |
4368 | ||
4369 | if (data->raw) | |
4370 | size += data->raw->size; | |
4371 | else | |
4372 | size += sizeof(u32); | |
4373 | ||
4374 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4375 | header->size += size; |
7f453c24 | 4376 | } |
bce38cd5 SE |
4377 | |
4378 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4379 | int size = sizeof(u64); /* nr */ | |
4380 | if (data->br_stack) { | |
4381 | size += data->br_stack->nr | |
4382 | * sizeof(struct perf_branch_entry); | |
4383 | } | |
4384 | header->size += size; | |
4385 | } | |
4018994f JO |
4386 | |
4387 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4388 | /* regs dump ABI info */ | |
4389 | int size = sizeof(u64); | |
4390 | ||
4391 | perf_sample_regs_user(&data->regs_user, regs); | |
4392 | ||
4393 | if (data->regs_user.regs) { | |
4394 | u64 mask = event->attr.sample_regs_user; | |
4395 | size += hweight64(mask) * sizeof(u64); | |
4396 | } | |
4397 | ||
4398 | header->size += size; | |
4399 | } | |
c5ebcedb JO |
4400 | |
4401 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4402 | /* | |
4403 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4404 | * processed as the last one or have additional check added | |
4405 | * in case new sample type is added, because we could eat | |
4406 | * up the rest of the sample size. | |
4407 | */ | |
4408 | struct perf_regs_user *uregs = &data->regs_user; | |
4409 | u16 stack_size = event->attr.sample_stack_user; | |
4410 | u16 size = sizeof(u64); | |
4411 | ||
4412 | if (!uregs->abi) | |
4413 | perf_sample_regs_user(uregs, regs); | |
4414 | ||
4415 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4416 | uregs->regs); | |
4417 | ||
4418 | /* | |
4419 | * If there is something to dump, add space for the dump | |
4420 | * itself and for the field that tells the dynamic size, | |
4421 | * which is how many have been actually dumped. | |
4422 | */ | |
4423 | if (stack_size) | |
4424 | size += sizeof(u64) + stack_size; | |
4425 | ||
4426 | data->stack_user_size = stack_size; | |
4427 | header->size += size; | |
4428 | } | |
5622f295 | 4429 | } |
7f453c24 | 4430 | |
a8b0ca17 | 4431 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4432 | struct perf_sample_data *data, |
4433 | struct pt_regs *regs) | |
4434 | { | |
4435 | struct perf_output_handle handle; | |
4436 | struct perf_event_header header; | |
689802b2 | 4437 | |
927c7a9e FW |
4438 | /* protect the callchain buffers */ |
4439 | rcu_read_lock(); | |
4440 | ||
cdd6c482 | 4441 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4442 | |
a7ac67ea | 4443 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4444 | goto exit; |
0322cd6e | 4445 | |
cdd6c482 | 4446 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4447 | |
8a057d84 | 4448 | perf_output_end(&handle); |
927c7a9e FW |
4449 | |
4450 | exit: | |
4451 | rcu_read_unlock(); | |
0322cd6e PZ |
4452 | } |
4453 | ||
38b200d6 | 4454 | /* |
cdd6c482 | 4455 | * read event_id |
38b200d6 PZ |
4456 | */ |
4457 | ||
4458 | struct perf_read_event { | |
4459 | struct perf_event_header header; | |
4460 | ||
4461 | u32 pid; | |
4462 | u32 tid; | |
38b200d6 PZ |
4463 | }; |
4464 | ||
4465 | static void | |
cdd6c482 | 4466 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4467 | struct task_struct *task) |
4468 | { | |
4469 | struct perf_output_handle handle; | |
c980d109 | 4470 | struct perf_sample_data sample; |
dfc65094 | 4471 | struct perf_read_event read_event = { |
38b200d6 | 4472 | .header = { |
cdd6c482 | 4473 | .type = PERF_RECORD_READ, |
38b200d6 | 4474 | .misc = 0, |
c320c7b7 | 4475 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4476 | }, |
cdd6c482 IM |
4477 | .pid = perf_event_pid(event, task), |
4478 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4479 | }; |
3dab77fb | 4480 | int ret; |
38b200d6 | 4481 | |
c980d109 | 4482 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4483 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4484 | if (ret) |
4485 | return; | |
4486 | ||
dfc65094 | 4487 | perf_output_put(&handle, read_event); |
cdd6c482 | 4488 | perf_output_read(&handle, event); |
c980d109 | 4489 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4490 | |
38b200d6 PZ |
4491 | perf_output_end(&handle); |
4492 | } | |
4493 | ||
52d857a8 JO |
4494 | typedef int (perf_event_aux_match_cb)(struct perf_event *event, void *data); |
4495 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); | |
4496 | ||
4497 | static void | |
4498 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
4499 | perf_event_aux_match_cb match, | |
4500 | perf_event_aux_output_cb output, | |
4501 | void *data) | |
4502 | { | |
4503 | struct perf_event *event; | |
4504 | ||
4505 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
4506 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
4507 | continue; | |
4508 | if (!event_filter_match(event)) | |
4509 | continue; | |
4510 | if (match(event, data)) | |
4511 | output(event, data); | |
4512 | } | |
4513 | } | |
4514 | ||
4515 | static void | |
4516 | perf_event_aux(perf_event_aux_match_cb match, | |
4517 | perf_event_aux_output_cb output, | |
4518 | void *data, | |
4519 | struct perf_event_context *task_ctx) | |
4520 | { | |
4521 | struct perf_cpu_context *cpuctx; | |
4522 | struct perf_event_context *ctx; | |
4523 | struct pmu *pmu; | |
4524 | int ctxn; | |
4525 | ||
4526 | rcu_read_lock(); | |
4527 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
4528 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
4529 | if (cpuctx->unique_pmu != pmu) | |
4530 | goto next; | |
4531 | perf_event_aux_ctx(&cpuctx->ctx, match, output, data); | |
4532 | if (task_ctx) | |
4533 | goto next; | |
4534 | ctxn = pmu->task_ctx_nr; | |
4535 | if (ctxn < 0) | |
4536 | goto next; | |
4537 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4538 | if (ctx) | |
4539 | perf_event_aux_ctx(ctx, match, output, data); | |
4540 | next: | |
4541 | put_cpu_ptr(pmu->pmu_cpu_context); | |
4542 | } | |
4543 | ||
4544 | if (task_ctx) { | |
4545 | preempt_disable(); | |
4546 | perf_event_aux_ctx(task_ctx, match, output, data); | |
4547 | preempt_enable(); | |
4548 | } | |
4549 | rcu_read_unlock(); | |
4550 | } | |
4551 | ||
60313ebe | 4552 | /* |
9f498cc5 PZ |
4553 | * task tracking -- fork/exit |
4554 | * | |
3af9e859 | 4555 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
60313ebe PZ |
4556 | */ |
4557 | ||
9f498cc5 | 4558 | struct perf_task_event { |
3a80b4a3 | 4559 | struct task_struct *task; |
cdd6c482 | 4560 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4561 | |
4562 | struct { | |
4563 | struct perf_event_header header; | |
4564 | ||
4565 | u32 pid; | |
4566 | u32 ppid; | |
9f498cc5 PZ |
4567 | u32 tid; |
4568 | u32 ptid; | |
393b2ad8 | 4569 | u64 time; |
cdd6c482 | 4570 | } event_id; |
60313ebe PZ |
4571 | }; |
4572 | ||
cdd6c482 | 4573 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 4574 | void *data) |
60313ebe | 4575 | { |
52d857a8 | 4576 | struct perf_task_event *task_event = data; |
60313ebe | 4577 | struct perf_output_handle handle; |
c980d109 | 4578 | struct perf_sample_data sample; |
9f498cc5 | 4579 | struct task_struct *task = task_event->task; |
c980d109 | 4580 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4581 | |
c980d109 | 4582 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4583 | |
c980d109 | 4584 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4585 | task_event->event_id.header.size); |
ef60777c | 4586 | if (ret) |
c980d109 | 4587 | goto out; |
60313ebe | 4588 | |
cdd6c482 IM |
4589 | task_event->event_id.pid = perf_event_pid(event, task); |
4590 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4591 | |
cdd6c482 IM |
4592 | task_event->event_id.tid = perf_event_tid(event, task); |
4593 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4594 | |
cdd6c482 | 4595 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4596 | |
c980d109 ACM |
4597 | perf_event__output_id_sample(event, &handle, &sample); |
4598 | ||
60313ebe | 4599 | perf_output_end(&handle); |
c980d109 ACM |
4600 | out: |
4601 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4602 | } |
4603 | ||
52d857a8 JO |
4604 | static int perf_event_task_match(struct perf_event *event, |
4605 | void *data __maybe_unused) | |
60313ebe | 4606 | { |
52d857a8 JO |
4607 | return event->attr.comm || event->attr.mmap || |
4608 | event->attr.mmap_data || event->attr.task; | |
60313ebe PZ |
4609 | } |
4610 | ||
cdd6c482 IM |
4611 | static void perf_event_task(struct task_struct *task, |
4612 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4613 | int new) |
60313ebe | 4614 | { |
9f498cc5 | 4615 | struct perf_task_event task_event; |
60313ebe | 4616 | |
cdd6c482 IM |
4617 | if (!atomic_read(&nr_comm_events) && |
4618 | !atomic_read(&nr_mmap_events) && | |
4619 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4620 | return; |
4621 | ||
9f498cc5 | 4622 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4623 | .task = task, |
4624 | .task_ctx = task_ctx, | |
cdd6c482 | 4625 | .event_id = { |
60313ebe | 4626 | .header = { |
cdd6c482 | 4627 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4628 | .misc = 0, |
cdd6c482 | 4629 | .size = sizeof(task_event.event_id), |
60313ebe | 4630 | }, |
573402db PZ |
4631 | /* .pid */ |
4632 | /* .ppid */ | |
9f498cc5 PZ |
4633 | /* .tid */ |
4634 | /* .ptid */ | |
6f93d0a7 | 4635 | .time = perf_clock(), |
60313ebe PZ |
4636 | }, |
4637 | }; | |
4638 | ||
52d857a8 JO |
4639 | perf_event_aux(perf_event_task_match, |
4640 | perf_event_task_output, | |
4641 | &task_event, | |
4642 | task_ctx); | |
9f498cc5 PZ |
4643 | } |
4644 | ||
cdd6c482 | 4645 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4646 | { |
cdd6c482 | 4647 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4648 | } |
4649 | ||
8d1b2d93 PZ |
4650 | /* |
4651 | * comm tracking | |
4652 | */ | |
4653 | ||
4654 | struct perf_comm_event { | |
22a4f650 IM |
4655 | struct task_struct *task; |
4656 | char *comm; | |
8d1b2d93 PZ |
4657 | int comm_size; |
4658 | ||
4659 | struct { | |
4660 | struct perf_event_header header; | |
4661 | ||
4662 | u32 pid; | |
4663 | u32 tid; | |
cdd6c482 | 4664 | } event_id; |
8d1b2d93 PZ |
4665 | }; |
4666 | ||
cdd6c482 | 4667 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 4668 | void *data) |
8d1b2d93 | 4669 | { |
52d857a8 | 4670 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 4671 | struct perf_output_handle handle; |
c980d109 | 4672 | struct perf_sample_data sample; |
cdd6c482 | 4673 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4674 | int ret; |
4675 | ||
4676 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); | |
4677 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4678 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4679 | |
4680 | if (ret) | |
c980d109 | 4681 | goto out; |
8d1b2d93 | 4682 | |
cdd6c482 IM |
4683 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4684 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4685 | |
cdd6c482 | 4686 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4687 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4688 | comm_event->comm_size); |
c980d109 ACM |
4689 | |
4690 | perf_event__output_id_sample(event, &handle, &sample); | |
4691 | ||
8d1b2d93 | 4692 | perf_output_end(&handle); |
c980d109 ACM |
4693 | out: |
4694 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4695 | } |
4696 | ||
52d857a8 JO |
4697 | static int perf_event_comm_match(struct perf_event *event, |
4698 | void *data __maybe_unused) | |
8d1b2d93 | 4699 | { |
52d857a8 | 4700 | return event->attr.comm; |
8d1b2d93 PZ |
4701 | } |
4702 | ||
cdd6c482 | 4703 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 4704 | { |
413ee3b4 | 4705 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4706 | unsigned int size; |
8d1b2d93 | 4707 | |
413ee3b4 | 4708 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4709 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4710 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4711 | |
4712 | comm_event->comm = comm; | |
4713 | comm_event->comm_size = size; | |
4714 | ||
cdd6c482 | 4715 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 4716 | |
52d857a8 JO |
4717 | perf_event_aux(perf_event_comm_match, |
4718 | perf_event_comm_output, | |
4719 | comm_event, | |
4720 | NULL); | |
8d1b2d93 PZ |
4721 | } |
4722 | ||
cdd6c482 | 4723 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4724 | { |
9ee318a7 | 4725 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4726 | struct perf_event_context *ctx; |
4727 | int ctxn; | |
9ee318a7 | 4728 | |
c79aa0d9 | 4729 | rcu_read_lock(); |
8dc85d54 PZ |
4730 | for_each_task_context_nr(ctxn) { |
4731 | ctx = task->perf_event_ctxp[ctxn]; | |
4732 | if (!ctx) | |
4733 | continue; | |
9ee318a7 | 4734 | |
8dc85d54 PZ |
4735 | perf_event_enable_on_exec(ctx); |
4736 | } | |
c79aa0d9 | 4737 | rcu_read_unlock(); |
9ee318a7 | 4738 | |
cdd6c482 | 4739 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4740 | return; |
a63eaf34 | 4741 | |
9ee318a7 | 4742 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4743 | .task = task, |
573402db PZ |
4744 | /* .comm */ |
4745 | /* .comm_size */ | |
cdd6c482 | 4746 | .event_id = { |
573402db | 4747 | .header = { |
cdd6c482 | 4748 | .type = PERF_RECORD_COMM, |
573402db PZ |
4749 | .misc = 0, |
4750 | /* .size */ | |
4751 | }, | |
4752 | /* .pid */ | |
4753 | /* .tid */ | |
8d1b2d93 PZ |
4754 | }, |
4755 | }; | |
4756 | ||
cdd6c482 | 4757 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
4758 | } |
4759 | ||
0a4a9391 PZ |
4760 | /* |
4761 | * mmap tracking | |
4762 | */ | |
4763 | ||
4764 | struct perf_mmap_event { | |
089dd79d PZ |
4765 | struct vm_area_struct *vma; |
4766 | ||
4767 | const char *file_name; | |
4768 | int file_size; | |
0a4a9391 PZ |
4769 | |
4770 | struct { | |
4771 | struct perf_event_header header; | |
4772 | ||
4773 | u32 pid; | |
4774 | u32 tid; | |
4775 | u64 start; | |
4776 | u64 len; | |
4777 | u64 pgoff; | |
cdd6c482 | 4778 | } event_id; |
0a4a9391 PZ |
4779 | }; |
4780 | ||
cdd6c482 | 4781 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 4782 | void *data) |
0a4a9391 | 4783 | { |
52d857a8 | 4784 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 4785 | struct perf_output_handle handle; |
c980d109 | 4786 | struct perf_sample_data sample; |
cdd6c482 | 4787 | int size = mmap_event->event_id.header.size; |
c980d109 | 4788 | int ret; |
0a4a9391 | 4789 | |
c980d109 ACM |
4790 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
4791 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4792 | mmap_event->event_id.header.size); |
0a4a9391 | 4793 | if (ret) |
c980d109 | 4794 | goto out; |
0a4a9391 | 4795 | |
cdd6c482 IM |
4796 | mmap_event->event_id.pid = perf_event_pid(event, current); |
4797 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 4798 | |
cdd6c482 | 4799 | perf_output_put(&handle, mmap_event->event_id); |
76369139 | 4800 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 4801 | mmap_event->file_size); |
c980d109 ACM |
4802 | |
4803 | perf_event__output_id_sample(event, &handle, &sample); | |
4804 | ||
78d613eb | 4805 | perf_output_end(&handle); |
c980d109 ACM |
4806 | out: |
4807 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
4808 | } |
4809 | ||
cdd6c482 | 4810 | static int perf_event_mmap_match(struct perf_event *event, |
52d857a8 | 4811 | void *data) |
0a4a9391 | 4812 | { |
52d857a8 JO |
4813 | struct perf_mmap_event *mmap_event = data; |
4814 | struct vm_area_struct *vma = mmap_event->vma; | |
4815 | int executable = vma->vm_flags & VM_EXEC; | |
0a4a9391 | 4816 | |
52d857a8 JO |
4817 | return (!executable && event->attr.mmap_data) || |
4818 | (executable && event->attr.mmap); | |
0a4a9391 PZ |
4819 | } |
4820 | ||
cdd6c482 | 4821 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 4822 | { |
089dd79d PZ |
4823 | struct vm_area_struct *vma = mmap_event->vma; |
4824 | struct file *file = vma->vm_file; | |
0a4a9391 PZ |
4825 | unsigned int size; |
4826 | char tmp[16]; | |
4827 | char *buf = NULL; | |
089dd79d | 4828 | const char *name; |
0a4a9391 | 4829 | |
413ee3b4 AB |
4830 | memset(tmp, 0, sizeof(tmp)); |
4831 | ||
0a4a9391 | 4832 | if (file) { |
413ee3b4 | 4833 | /* |
76369139 | 4834 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
4835 | * need to add enough zero bytes after the string to handle |
4836 | * the 64bit alignment we do later. | |
4837 | */ | |
4838 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
4839 | if (!buf) { |
4840 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
4841 | goto got_name; | |
4842 | } | |
d3d21c41 | 4843 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
4844 | if (IS_ERR(name)) { |
4845 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
4846 | goto got_name; | |
4847 | } | |
4848 | } else { | |
413ee3b4 AB |
4849 | if (arch_vma_name(mmap_event->vma)) { |
4850 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
c97847d2 CG |
4851 | sizeof(tmp) - 1); |
4852 | tmp[sizeof(tmp) - 1] = '\0'; | |
089dd79d | 4853 | goto got_name; |
413ee3b4 | 4854 | } |
089dd79d PZ |
4855 | |
4856 | if (!vma->vm_mm) { | |
4857 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
4858 | goto got_name; | |
3af9e859 EM |
4859 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
4860 | vma->vm_end >= vma->vm_mm->brk) { | |
4861 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
4862 | goto got_name; | |
4863 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
4864 | vma->vm_end >= vma->vm_mm->start_stack) { | |
4865 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
4866 | goto got_name; | |
089dd79d PZ |
4867 | } |
4868 | ||
0a4a9391 PZ |
4869 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
4870 | goto got_name; | |
4871 | } | |
4872 | ||
4873 | got_name: | |
888fcee0 | 4874 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
4875 | |
4876 | mmap_event->file_name = name; | |
4877 | mmap_event->file_size = size; | |
4878 | ||
2fe85427 SE |
4879 | if (!(vma->vm_flags & VM_EXEC)) |
4880 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
4881 | ||
cdd6c482 | 4882 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 4883 | |
52d857a8 JO |
4884 | perf_event_aux(perf_event_mmap_match, |
4885 | perf_event_mmap_output, | |
4886 | mmap_event, | |
4887 | NULL); | |
665c2142 | 4888 | |
0a4a9391 PZ |
4889 | kfree(buf); |
4890 | } | |
4891 | ||
3af9e859 | 4892 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 4893 | { |
9ee318a7 PZ |
4894 | struct perf_mmap_event mmap_event; |
4895 | ||
cdd6c482 | 4896 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
4897 | return; |
4898 | ||
4899 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 4900 | .vma = vma, |
573402db PZ |
4901 | /* .file_name */ |
4902 | /* .file_size */ | |
cdd6c482 | 4903 | .event_id = { |
573402db | 4904 | .header = { |
cdd6c482 | 4905 | .type = PERF_RECORD_MMAP, |
39447b38 | 4906 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
4907 | /* .size */ |
4908 | }, | |
4909 | /* .pid */ | |
4910 | /* .tid */ | |
089dd79d PZ |
4911 | .start = vma->vm_start, |
4912 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 4913 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 PZ |
4914 | }, |
4915 | }; | |
4916 | ||
cdd6c482 | 4917 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
4918 | } |
4919 | ||
a78ac325 PZ |
4920 | /* |
4921 | * IRQ throttle logging | |
4922 | */ | |
4923 | ||
cdd6c482 | 4924 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
4925 | { |
4926 | struct perf_output_handle handle; | |
c980d109 | 4927 | struct perf_sample_data sample; |
a78ac325 PZ |
4928 | int ret; |
4929 | ||
4930 | struct { | |
4931 | struct perf_event_header header; | |
4932 | u64 time; | |
cca3f454 | 4933 | u64 id; |
7f453c24 | 4934 | u64 stream_id; |
a78ac325 PZ |
4935 | } throttle_event = { |
4936 | .header = { | |
cdd6c482 | 4937 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
4938 | .misc = 0, |
4939 | .size = sizeof(throttle_event), | |
4940 | }, | |
def0a9b2 | 4941 | .time = perf_clock(), |
cdd6c482 IM |
4942 | .id = primary_event_id(event), |
4943 | .stream_id = event->id, | |
a78ac325 PZ |
4944 | }; |
4945 | ||
966ee4d6 | 4946 | if (enable) |
cdd6c482 | 4947 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 4948 | |
c980d109 ACM |
4949 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
4950 | ||
4951 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4952 | throttle_event.header.size); |
a78ac325 PZ |
4953 | if (ret) |
4954 | return; | |
4955 | ||
4956 | perf_output_put(&handle, throttle_event); | |
c980d109 | 4957 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
4958 | perf_output_end(&handle); |
4959 | } | |
4960 | ||
f6c7d5fe | 4961 | /* |
cdd6c482 | 4962 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
4963 | */ |
4964 | ||
a8b0ca17 | 4965 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4966 | int throttle, struct perf_sample_data *data, |
4967 | struct pt_regs *regs) | |
f6c7d5fe | 4968 | { |
cdd6c482 IM |
4969 | int events = atomic_read(&event->event_limit); |
4970 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 4971 | u64 seq; |
79f14641 PZ |
4972 | int ret = 0; |
4973 | ||
96398826 PZ |
4974 | /* |
4975 | * Non-sampling counters might still use the PMI to fold short | |
4976 | * hardware counters, ignore those. | |
4977 | */ | |
4978 | if (unlikely(!is_sampling_event(event))) | |
4979 | return 0; | |
4980 | ||
e050e3f0 SE |
4981 | seq = __this_cpu_read(perf_throttled_seq); |
4982 | if (seq != hwc->interrupts_seq) { | |
4983 | hwc->interrupts_seq = seq; | |
4984 | hwc->interrupts = 1; | |
4985 | } else { | |
4986 | hwc->interrupts++; | |
4987 | if (unlikely(throttle | |
4988 | && hwc->interrupts >= max_samples_per_tick)) { | |
4989 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
4990 | hwc->interrupts = MAX_INTERRUPTS; |
4991 | perf_log_throttle(event, 0); | |
a78ac325 PZ |
4992 | ret = 1; |
4993 | } | |
e050e3f0 | 4994 | } |
60db5e09 | 4995 | |
cdd6c482 | 4996 | if (event->attr.freq) { |
def0a9b2 | 4997 | u64 now = perf_clock(); |
abd50713 | 4998 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 4999 | |
abd50713 | 5000 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5001 | |
abd50713 | 5002 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5003 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5004 | } |
5005 | ||
2023b359 PZ |
5006 | /* |
5007 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5008 | * events |
2023b359 PZ |
5009 | */ |
5010 | ||
cdd6c482 IM |
5011 | event->pending_kill = POLL_IN; |
5012 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5013 | ret = 1; |
cdd6c482 | 5014 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5015 | event->pending_disable = 1; |
5016 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5017 | } |
5018 | ||
453f19ee | 5019 | if (event->overflow_handler) |
a8b0ca17 | 5020 | event->overflow_handler(event, data, regs); |
453f19ee | 5021 | else |
a8b0ca17 | 5022 | perf_event_output(event, data, regs); |
453f19ee | 5023 | |
f506b3dc | 5024 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5025 | event->pending_wakeup = 1; |
5026 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5027 | } |
5028 | ||
79f14641 | 5029 | return ret; |
f6c7d5fe PZ |
5030 | } |
5031 | ||
a8b0ca17 | 5032 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5033 | struct perf_sample_data *data, |
5034 | struct pt_regs *regs) | |
850bc73f | 5035 | { |
a8b0ca17 | 5036 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5037 | } |
5038 | ||
15dbf27c | 5039 | /* |
cdd6c482 | 5040 | * Generic software event infrastructure |
15dbf27c PZ |
5041 | */ |
5042 | ||
b28ab83c PZ |
5043 | struct swevent_htable { |
5044 | struct swevent_hlist *swevent_hlist; | |
5045 | struct mutex hlist_mutex; | |
5046 | int hlist_refcount; | |
5047 | ||
5048 | /* Recursion avoidance in each contexts */ | |
5049 | int recursion[PERF_NR_CONTEXTS]; | |
5050 | }; | |
5051 | ||
5052 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5053 | ||
7b4b6658 | 5054 | /* |
cdd6c482 IM |
5055 | * We directly increment event->count and keep a second value in |
5056 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5057 | * is kept in the range [-sample_period, 0] so that we can use the |
5058 | * sign as trigger. | |
5059 | */ | |
5060 | ||
ab573844 | 5061 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5062 | { |
cdd6c482 | 5063 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5064 | u64 period = hwc->last_period; |
5065 | u64 nr, offset; | |
5066 | s64 old, val; | |
5067 | ||
5068 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5069 | |
5070 | again: | |
e7850595 | 5071 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5072 | if (val < 0) |
5073 | return 0; | |
15dbf27c | 5074 | |
7b4b6658 PZ |
5075 | nr = div64_u64(period + val, period); |
5076 | offset = nr * period; | |
5077 | val -= offset; | |
e7850595 | 5078 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5079 | goto again; |
15dbf27c | 5080 | |
7b4b6658 | 5081 | return nr; |
15dbf27c PZ |
5082 | } |
5083 | ||
0cff784a | 5084 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5085 | struct perf_sample_data *data, |
5622f295 | 5086 | struct pt_regs *regs) |
15dbf27c | 5087 | { |
cdd6c482 | 5088 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5089 | int throttle = 0; |
15dbf27c | 5090 | |
0cff784a PZ |
5091 | if (!overflow) |
5092 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5093 | |
7b4b6658 PZ |
5094 | if (hwc->interrupts == MAX_INTERRUPTS) |
5095 | return; | |
15dbf27c | 5096 | |
7b4b6658 | 5097 | for (; overflow; overflow--) { |
a8b0ca17 | 5098 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5099 | data, regs)) { |
7b4b6658 PZ |
5100 | /* |
5101 | * We inhibit the overflow from happening when | |
5102 | * hwc->interrupts == MAX_INTERRUPTS. | |
5103 | */ | |
5104 | break; | |
5105 | } | |
cf450a73 | 5106 | throttle = 1; |
7b4b6658 | 5107 | } |
15dbf27c PZ |
5108 | } |
5109 | ||
a4eaf7f1 | 5110 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5111 | struct perf_sample_data *data, |
5622f295 | 5112 | struct pt_regs *regs) |
7b4b6658 | 5113 | { |
cdd6c482 | 5114 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5115 | |
e7850595 | 5116 | local64_add(nr, &event->count); |
d6d020e9 | 5117 | |
0cff784a PZ |
5118 | if (!regs) |
5119 | return; | |
5120 | ||
6c7e550f | 5121 | if (!is_sampling_event(event)) |
7b4b6658 | 5122 | return; |
d6d020e9 | 5123 | |
5d81e5cf AV |
5124 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5125 | data->period = nr; | |
5126 | return perf_swevent_overflow(event, 1, data, regs); | |
5127 | } else | |
5128 | data->period = event->hw.last_period; | |
5129 | ||
0cff784a | 5130 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5131 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5132 | |
e7850595 | 5133 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5134 | return; |
df1a132b | 5135 | |
a8b0ca17 | 5136 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5137 | } |
5138 | ||
f5ffe02e FW |
5139 | static int perf_exclude_event(struct perf_event *event, |
5140 | struct pt_regs *regs) | |
5141 | { | |
a4eaf7f1 | 5142 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5143 | return 1; |
a4eaf7f1 | 5144 | |
f5ffe02e FW |
5145 | if (regs) { |
5146 | if (event->attr.exclude_user && user_mode(regs)) | |
5147 | return 1; | |
5148 | ||
5149 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5150 | return 1; | |
5151 | } | |
5152 | ||
5153 | return 0; | |
5154 | } | |
5155 | ||
cdd6c482 | 5156 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5157 | enum perf_type_id type, |
6fb2915d LZ |
5158 | u32 event_id, |
5159 | struct perf_sample_data *data, | |
5160 | struct pt_regs *regs) | |
15dbf27c | 5161 | { |
cdd6c482 | 5162 | if (event->attr.type != type) |
a21ca2ca | 5163 | return 0; |
f5ffe02e | 5164 | |
cdd6c482 | 5165 | if (event->attr.config != event_id) |
15dbf27c PZ |
5166 | return 0; |
5167 | ||
f5ffe02e FW |
5168 | if (perf_exclude_event(event, regs)) |
5169 | return 0; | |
15dbf27c PZ |
5170 | |
5171 | return 1; | |
5172 | } | |
5173 | ||
76e1d904 FW |
5174 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5175 | { | |
5176 | u64 val = event_id | (type << 32); | |
5177 | ||
5178 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5179 | } | |
5180 | ||
49f135ed FW |
5181 | static inline struct hlist_head * |
5182 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5183 | { |
49f135ed FW |
5184 | u64 hash = swevent_hash(type, event_id); |
5185 | ||
5186 | return &hlist->heads[hash]; | |
5187 | } | |
76e1d904 | 5188 | |
49f135ed FW |
5189 | /* For the read side: events when they trigger */ |
5190 | static inline struct hlist_head * | |
b28ab83c | 5191 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5192 | { |
5193 | struct swevent_hlist *hlist; | |
76e1d904 | 5194 | |
b28ab83c | 5195 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5196 | if (!hlist) |
5197 | return NULL; | |
5198 | ||
49f135ed FW |
5199 | return __find_swevent_head(hlist, type, event_id); |
5200 | } | |
5201 | ||
5202 | /* For the event head insertion and removal in the hlist */ | |
5203 | static inline struct hlist_head * | |
b28ab83c | 5204 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5205 | { |
5206 | struct swevent_hlist *hlist; | |
5207 | u32 event_id = event->attr.config; | |
5208 | u64 type = event->attr.type; | |
5209 | ||
5210 | /* | |
5211 | * Event scheduling is always serialized against hlist allocation | |
5212 | * and release. Which makes the protected version suitable here. | |
5213 | * The context lock guarantees that. | |
5214 | */ | |
b28ab83c | 5215 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5216 | lockdep_is_held(&event->ctx->lock)); |
5217 | if (!hlist) | |
5218 | return NULL; | |
5219 | ||
5220 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5221 | } |
5222 | ||
5223 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5224 | u64 nr, |
76e1d904 FW |
5225 | struct perf_sample_data *data, |
5226 | struct pt_regs *regs) | |
15dbf27c | 5227 | { |
b28ab83c | 5228 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5229 | struct perf_event *event; |
76e1d904 | 5230 | struct hlist_head *head; |
15dbf27c | 5231 | |
76e1d904 | 5232 | rcu_read_lock(); |
b28ab83c | 5233 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5234 | if (!head) |
5235 | goto end; | |
5236 | ||
b67bfe0d | 5237 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5238 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5239 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5240 | } |
76e1d904 FW |
5241 | end: |
5242 | rcu_read_unlock(); | |
15dbf27c PZ |
5243 | } |
5244 | ||
4ed7c92d | 5245 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5246 | { |
b28ab83c | 5247 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5248 | |
b28ab83c | 5249 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5250 | } |
645e8cc0 | 5251 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5252 | |
fa9f90be | 5253 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5254 | { |
b28ab83c | 5255 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5256 | |
b28ab83c | 5257 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5258 | } |
15dbf27c | 5259 | |
a8b0ca17 | 5260 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5261 | { |
a4234bfc | 5262 | struct perf_sample_data data; |
4ed7c92d PZ |
5263 | int rctx; |
5264 | ||
1c024eca | 5265 | preempt_disable_notrace(); |
4ed7c92d PZ |
5266 | rctx = perf_swevent_get_recursion_context(); |
5267 | if (rctx < 0) | |
5268 | return; | |
a4234bfc | 5269 | |
fd0d000b | 5270 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5271 | |
a8b0ca17 | 5272 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5273 | |
5274 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5275 | preempt_enable_notrace(); |
b8e83514 PZ |
5276 | } |
5277 | ||
cdd6c482 | 5278 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5279 | { |
15dbf27c PZ |
5280 | } |
5281 | ||
a4eaf7f1 | 5282 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5283 | { |
b28ab83c | 5284 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5285 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5286 | struct hlist_head *head; |
5287 | ||
6c7e550f | 5288 | if (is_sampling_event(event)) { |
7b4b6658 | 5289 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5290 | perf_swevent_set_period(event); |
7b4b6658 | 5291 | } |
76e1d904 | 5292 | |
a4eaf7f1 PZ |
5293 | hwc->state = !(flags & PERF_EF_START); |
5294 | ||
b28ab83c | 5295 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
5296 | if (WARN_ON_ONCE(!head)) |
5297 | return -EINVAL; | |
5298 | ||
5299 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5300 | ||
15dbf27c PZ |
5301 | return 0; |
5302 | } | |
5303 | ||
a4eaf7f1 | 5304 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5305 | { |
76e1d904 | 5306 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5307 | } |
5308 | ||
a4eaf7f1 | 5309 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5310 | { |
a4eaf7f1 | 5311 | event->hw.state = 0; |
d6d020e9 | 5312 | } |
aa9c4c0f | 5313 | |
a4eaf7f1 | 5314 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5315 | { |
a4eaf7f1 | 5316 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5317 | } |
5318 | ||
49f135ed FW |
5319 | /* Deref the hlist from the update side */ |
5320 | static inline struct swevent_hlist * | |
b28ab83c | 5321 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5322 | { |
b28ab83c PZ |
5323 | return rcu_dereference_protected(swhash->swevent_hlist, |
5324 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5325 | } |
5326 | ||
b28ab83c | 5327 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5328 | { |
b28ab83c | 5329 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5330 | |
49f135ed | 5331 | if (!hlist) |
76e1d904 FW |
5332 | return; |
5333 | ||
b28ab83c | 5334 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5335 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5336 | } |
5337 | ||
5338 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5339 | { | |
b28ab83c | 5340 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5341 | |
b28ab83c | 5342 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5343 | |
b28ab83c PZ |
5344 | if (!--swhash->hlist_refcount) |
5345 | swevent_hlist_release(swhash); | |
76e1d904 | 5346 | |
b28ab83c | 5347 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5348 | } |
5349 | ||
5350 | static void swevent_hlist_put(struct perf_event *event) | |
5351 | { | |
5352 | int cpu; | |
5353 | ||
5354 | if (event->cpu != -1) { | |
5355 | swevent_hlist_put_cpu(event, event->cpu); | |
5356 | return; | |
5357 | } | |
5358 | ||
5359 | for_each_possible_cpu(cpu) | |
5360 | swevent_hlist_put_cpu(event, cpu); | |
5361 | } | |
5362 | ||
5363 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5364 | { | |
b28ab83c | 5365 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5366 | int err = 0; |
5367 | ||
b28ab83c | 5368 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5369 | |
b28ab83c | 5370 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5371 | struct swevent_hlist *hlist; |
5372 | ||
5373 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5374 | if (!hlist) { | |
5375 | err = -ENOMEM; | |
5376 | goto exit; | |
5377 | } | |
b28ab83c | 5378 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5379 | } |
b28ab83c | 5380 | swhash->hlist_refcount++; |
9ed6060d | 5381 | exit: |
b28ab83c | 5382 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5383 | |
5384 | return err; | |
5385 | } | |
5386 | ||
5387 | static int swevent_hlist_get(struct perf_event *event) | |
5388 | { | |
5389 | int err; | |
5390 | int cpu, failed_cpu; | |
5391 | ||
5392 | if (event->cpu != -1) | |
5393 | return swevent_hlist_get_cpu(event, event->cpu); | |
5394 | ||
5395 | get_online_cpus(); | |
5396 | for_each_possible_cpu(cpu) { | |
5397 | err = swevent_hlist_get_cpu(event, cpu); | |
5398 | if (err) { | |
5399 | failed_cpu = cpu; | |
5400 | goto fail; | |
5401 | } | |
5402 | } | |
5403 | put_online_cpus(); | |
5404 | ||
5405 | return 0; | |
9ed6060d | 5406 | fail: |
76e1d904 FW |
5407 | for_each_possible_cpu(cpu) { |
5408 | if (cpu == failed_cpu) | |
5409 | break; | |
5410 | swevent_hlist_put_cpu(event, cpu); | |
5411 | } | |
5412 | ||
5413 | put_online_cpus(); | |
5414 | return err; | |
5415 | } | |
5416 | ||
c5905afb | 5417 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5418 | |
b0a873eb PZ |
5419 | static void sw_perf_event_destroy(struct perf_event *event) |
5420 | { | |
5421 | u64 event_id = event->attr.config; | |
95476b64 | 5422 | |
b0a873eb PZ |
5423 | WARN_ON(event->parent); |
5424 | ||
c5905afb | 5425 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5426 | swevent_hlist_put(event); |
5427 | } | |
5428 | ||
5429 | static int perf_swevent_init(struct perf_event *event) | |
5430 | { | |
8176cced | 5431 | u64 event_id = event->attr.config; |
b0a873eb PZ |
5432 | |
5433 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5434 | return -ENOENT; | |
5435 | ||
2481c5fa SE |
5436 | /* |
5437 | * no branch sampling for software events | |
5438 | */ | |
5439 | if (has_branch_stack(event)) | |
5440 | return -EOPNOTSUPP; | |
5441 | ||
b0a873eb PZ |
5442 | switch (event_id) { |
5443 | case PERF_COUNT_SW_CPU_CLOCK: | |
5444 | case PERF_COUNT_SW_TASK_CLOCK: | |
5445 | return -ENOENT; | |
5446 | ||
5447 | default: | |
5448 | break; | |
5449 | } | |
5450 | ||
ce677831 | 5451 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5452 | return -ENOENT; |
5453 | ||
5454 | if (!event->parent) { | |
5455 | int err; | |
5456 | ||
5457 | err = swevent_hlist_get(event); | |
5458 | if (err) | |
5459 | return err; | |
5460 | ||
c5905afb | 5461 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5462 | event->destroy = sw_perf_event_destroy; |
5463 | } | |
5464 | ||
5465 | return 0; | |
5466 | } | |
5467 | ||
35edc2a5 PZ |
5468 | static int perf_swevent_event_idx(struct perf_event *event) |
5469 | { | |
5470 | return 0; | |
5471 | } | |
5472 | ||
b0a873eb | 5473 | static struct pmu perf_swevent = { |
89a1e187 | 5474 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5475 | |
b0a873eb | 5476 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5477 | .add = perf_swevent_add, |
5478 | .del = perf_swevent_del, | |
5479 | .start = perf_swevent_start, | |
5480 | .stop = perf_swevent_stop, | |
1c024eca | 5481 | .read = perf_swevent_read, |
35edc2a5 PZ |
5482 | |
5483 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5484 | }; |
5485 | ||
b0a873eb PZ |
5486 | #ifdef CONFIG_EVENT_TRACING |
5487 | ||
1c024eca PZ |
5488 | static int perf_tp_filter_match(struct perf_event *event, |
5489 | struct perf_sample_data *data) | |
5490 | { | |
5491 | void *record = data->raw->data; | |
5492 | ||
5493 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5494 | return 1; | |
5495 | return 0; | |
5496 | } | |
5497 | ||
5498 | static int perf_tp_event_match(struct perf_event *event, | |
5499 | struct perf_sample_data *data, | |
5500 | struct pt_regs *regs) | |
5501 | { | |
a0f7d0f7 FW |
5502 | if (event->hw.state & PERF_HES_STOPPED) |
5503 | return 0; | |
580d607c PZ |
5504 | /* |
5505 | * All tracepoints are from kernel-space. | |
5506 | */ | |
5507 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5508 | return 0; |
5509 | ||
5510 | if (!perf_tp_filter_match(event, data)) | |
5511 | return 0; | |
5512 | ||
5513 | return 1; | |
5514 | } | |
5515 | ||
5516 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5517 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5518 | struct task_struct *task) | |
95476b64 FW |
5519 | { |
5520 | struct perf_sample_data data; | |
1c024eca | 5521 | struct perf_event *event; |
1c024eca | 5522 | |
95476b64 FW |
5523 | struct perf_raw_record raw = { |
5524 | .size = entry_size, | |
5525 | .data = record, | |
5526 | }; | |
5527 | ||
fd0d000b | 5528 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5529 | data.raw = &raw; |
5530 | ||
b67bfe0d | 5531 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 5532 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 5533 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5534 | } |
ecc55f84 | 5535 | |
e6dab5ff AV |
5536 | /* |
5537 | * If we got specified a target task, also iterate its context and | |
5538 | * deliver this event there too. | |
5539 | */ | |
5540 | if (task && task != current) { | |
5541 | struct perf_event_context *ctx; | |
5542 | struct trace_entry *entry = record; | |
5543 | ||
5544 | rcu_read_lock(); | |
5545 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5546 | if (!ctx) | |
5547 | goto unlock; | |
5548 | ||
5549 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5550 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5551 | continue; | |
5552 | if (event->attr.config != entry->type) | |
5553 | continue; | |
5554 | if (perf_tp_event_match(event, &data, regs)) | |
5555 | perf_swevent_event(event, count, &data, regs); | |
5556 | } | |
5557 | unlock: | |
5558 | rcu_read_unlock(); | |
5559 | } | |
5560 | ||
ecc55f84 | 5561 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5562 | } |
5563 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5564 | ||
cdd6c482 | 5565 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5566 | { |
1c024eca | 5567 | perf_trace_destroy(event); |
e077df4f PZ |
5568 | } |
5569 | ||
b0a873eb | 5570 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5571 | { |
76e1d904 FW |
5572 | int err; |
5573 | ||
b0a873eb PZ |
5574 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5575 | return -ENOENT; | |
5576 | ||
2481c5fa SE |
5577 | /* |
5578 | * no branch sampling for tracepoint events | |
5579 | */ | |
5580 | if (has_branch_stack(event)) | |
5581 | return -EOPNOTSUPP; | |
5582 | ||
1c024eca PZ |
5583 | err = perf_trace_init(event); |
5584 | if (err) | |
b0a873eb | 5585 | return err; |
e077df4f | 5586 | |
cdd6c482 | 5587 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5588 | |
b0a873eb PZ |
5589 | return 0; |
5590 | } | |
5591 | ||
5592 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5593 | .task_ctx_nr = perf_sw_context, |
5594 | ||
b0a873eb | 5595 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5596 | .add = perf_trace_add, |
5597 | .del = perf_trace_del, | |
5598 | .start = perf_swevent_start, | |
5599 | .stop = perf_swevent_stop, | |
b0a873eb | 5600 | .read = perf_swevent_read, |
35edc2a5 PZ |
5601 | |
5602 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5603 | }; |
5604 | ||
5605 | static inline void perf_tp_register(void) | |
5606 | { | |
2e80a82a | 5607 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5608 | } |
6fb2915d LZ |
5609 | |
5610 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5611 | { | |
5612 | char *filter_str; | |
5613 | int ret; | |
5614 | ||
5615 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5616 | return -EINVAL; | |
5617 | ||
5618 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5619 | if (IS_ERR(filter_str)) | |
5620 | return PTR_ERR(filter_str); | |
5621 | ||
5622 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5623 | ||
5624 | kfree(filter_str); | |
5625 | return ret; | |
5626 | } | |
5627 | ||
5628 | static void perf_event_free_filter(struct perf_event *event) | |
5629 | { | |
5630 | ftrace_profile_free_filter(event); | |
5631 | } | |
5632 | ||
e077df4f | 5633 | #else |
6fb2915d | 5634 | |
b0a873eb | 5635 | static inline void perf_tp_register(void) |
e077df4f | 5636 | { |
e077df4f | 5637 | } |
6fb2915d LZ |
5638 | |
5639 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5640 | { | |
5641 | return -ENOENT; | |
5642 | } | |
5643 | ||
5644 | static void perf_event_free_filter(struct perf_event *event) | |
5645 | { | |
5646 | } | |
5647 | ||
07b139c8 | 5648 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5649 | |
24f1e32c | 5650 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5651 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5652 | { |
f5ffe02e FW |
5653 | struct perf_sample_data sample; |
5654 | struct pt_regs *regs = data; | |
5655 | ||
fd0d000b | 5656 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 5657 | |
a4eaf7f1 | 5658 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5659 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5660 | } |
5661 | #endif | |
5662 | ||
b0a873eb PZ |
5663 | /* |
5664 | * hrtimer based swevent callback | |
5665 | */ | |
f29ac756 | 5666 | |
b0a873eb | 5667 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5668 | { |
b0a873eb PZ |
5669 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5670 | struct perf_sample_data data; | |
5671 | struct pt_regs *regs; | |
5672 | struct perf_event *event; | |
5673 | u64 period; | |
f29ac756 | 5674 | |
b0a873eb | 5675 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5676 | |
5677 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5678 | return HRTIMER_NORESTART; | |
5679 | ||
b0a873eb | 5680 | event->pmu->read(event); |
f344011c | 5681 | |
fd0d000b | 5682 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
5683 | regs = get_irq_regs(); |
5684 | ||
5685 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5686 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 5687 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
5688 | ret = HRTIMER_NORESTART; |
5689 | } | |
24f1e32c | 5690 | |
b0a873eb PZ |
5691 | period = max_t(u64, 10000, event->hw.sample_period); |
5692 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5693 | |
b0a873eb | 5694 | return ret; |
f29ac756 PZ |
5695 | } |
5696 | ||
b0a873eb | 5697 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5698 | { |
b0a873eb | 5699 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5700 | s64 period; |
5701 | ||
5702 | if (!is_sampling_event(event)) | |
5703 | return; | |
f5ffe02e | 5704 | |
5d508e82 FBH |
5705 | period = local64_read(&hwc->period_left); |
5706 | if (period) { | |
5707 | if (period < 0) | |
5708 | period = 10000; | |
fa407f35 | 5709 | |
5d508e82 FBH |
5710 | local64_set(&hwc->period_left, 0); |
5711 | } else { | |
5712 | period = max_t(u64, 10000, hwc->sample_period); | |
5713 | } | |
5714 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 5715 | ns_to_ktime(period), 0, |
b5ab4cd5 | 5716 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 5717 | } |
b0a873eb PZ |
5718 | |
5719 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 5720 | { |
b0a873eb PZ |
5721 | struct hw_perf_event *hwc = &event->hw; |
5722 | ||
6c7e550f | 5723 | if (is_sampling_event(event)) { |
b0a873eb | 5724 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 5725 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
5726 | |
5727 | hrtimer_cancel(&hwc->hrtimer); | |
5728 | } | |
24f1e32c FW |
5729 | } |
5730 | ||
ba3dd36c PZ |
5731 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5732 | { | |
5733 | struct hw_perf_event *hwc = &event->hw; | |
5734 | ||
5735 | if (!is_sampling_event(event)) | |
5736 | return; | |
5737 | ||
5738 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5739 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
5740 | ||
5741 | /* | |
5742 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
5743 | * mapping and avoid the whole period adjust feedback stuff. | |
5744 | */ | |
5745 | if (event->attr.freq) { | |
5746 | long freq = event->attr.sample_freq; | |
5747 | ||
5748 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
5749 | hwc->sample_period = event->attr.sample_period; | |
5750 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 5751 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
5752 | event->attr.freq = 0; |
5753 | } | |
5754 | } | |
5755 | ||
b0a873eb PZ |
5756 | /* |
5757 | * Software event: cpu wall time clock | |
5758 | */ | |
5759 | ||
5760 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 5761 | { |
b0a873eb PZ |
5762 | s64 prev; |
5763 | u64 now; | |
5764 | ||
a4eaf7f1 | 5765 | now = local_clock(); |
b0a873eb PZ |
5766 | prev = local64_xchg(&event->hw.prev_count, now); |
5767 | local64_add(now - prev, &event->count); | |
24f1e32c | 5768 | } |
24f1e32c | 5769 | |
a4eaf7f1 | 5770 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5771 | { |
a4eaf7f1 | 5772 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 5773 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5774 | } |
5775 | ||
a4eaf7f1 | 5776 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 5777 | { |
b0a873eb PZ |
5778 | perf_swevent_cancel_hrtimer(event); |
5779 | cpu_clock_event_update(event); | |
5780 | } | |
f29ac756 | 5781 | |
a4eaf7f1 PZ |
5782 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
5783 | { | |
5784 | if (flags & PERF_EF_START) | |
5785 | cpu_clock_event_start(event, flags); | |
5786 | ||
5787 | return 0; | |
5788 | } | |
5789 | ||
5790 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
5791 | { | |
5792 | cpu_clock_event_stop(event, flags); | |
5793 | } | |
5794 | ||
b0a873eb PZ |
5795 | static void cpu_clock_event_read(struct perf_event *event) |
5796 | { | |
5797 | cpu_clock_event_update(event); | |
5798 | } | |
f344011c | 5799 | |
b0a873eb PZ |
5800 | static int cpu_clock_event_init(struct perf_event *event) |
5801 | { | |
5802 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5803 | return -ENOENT; | |
5804 | ||
5805 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
5806 | return -ENOENT; | |
5807 | ||
2481c5fa SE |
5808 | /* |
5809 | * no branch sampling for software events | |
5810 | */ | |
5811 | if (has_branch_stack(event)) | |
5812 | return -EOPNOTSUPP; | |
5813 | ||
ba3dd36c PZ |
5814 | perf_swevent_init_hrtimer(event); |
5815 | ||
b0a873eb | 5816 | return 0; |
f29ac756 PZ |
5817 | } |
5818 | ||
b0a873eb | 5819 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
5820 | .task_ctx_nr = perf_sw_context, |
5821 | ||
b0a873eb | 5822 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
5823 | .add = cpu_clock_event_add, |
5824 | .del = cpu_clock_event_del, | |
5825 | .start = cpu_clock_event_start, | |
5826 | .stop = cpu_clock_event_stop, | |
b0a873eb | 5827 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
5828 | |
5829 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5830 | }; |
5831 | ||
5832 | /* | |
5833 | * Software event: task time clock | |
5834 | */ | |
5835 | ||
5836 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 5837 | { |
b0a873eb PZ |
5838 | u64 prev; |
5839 | s64 delta; | |
5c92d124 | 5840 | |
b0a873eb PZ |
5841 | prev = local64_xchg(&event->hw.prev_count, now); |
5842 | delta = now - prev; | |
5843 | local64_add(delta, &event->count); | |
5844 | } | |
5c92d124 | 5845 | |
a4eaf7f1 | 5846 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5847 | { |
a4eaf7f1 | 5848 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 5849 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5850 | } |
5851 | ||
a4eaf7f1 | 5852 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
5853 | { |
5854 | perf_swevent_cancel_hrtimer(event); | |
5855 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
5856 | } |
5857 | ||
5858 | static int task_clock_event_add(struct perf_event *event, int flags) | |
5859 | { | |
5860 | if (flags & PERF_EF_START) | |
5861 | task_clock_event_start(event, flags); | |
b0a873eb | 5862 | |
a4eaf7f1 PZ |
5863 | return 0; |
5864 | } | |
5865 | ||
5866 | static void task_clock_event_del(struct perf_event *event, int flags) | |
5867 | { | |
5868 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
5869 | } |
5870 | ||
5871 | static void task_clock_event_read(struct perf_event *event) | |
5872 | { | |
768a06e2 PZ |
5873 | u64 now = perf_clock(); |
5874 | u64 delta = now - event->ctx->timestamp; | |
5875 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
5876 | |
5877 | task_clock_event_update(event, time); | |
5878 | } | |
5879 | ||
5880 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 5881 | { |
b0a873eb PZ |
5882 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5883 | return -ENOENT; | |
5884 | ||
5885 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
5886 | return -ENOENT; | |
5887 | ||
2481c5fa SE |
5888 | /* |
5889 | * no branch sampling for software events | |
5890 | */ | |
5891 | if (has_branch_stack(event)) | |
5892 | return -EOPNOTSUPP; | |
5893 | ||
ba3dd36c PZ |
5894 | perf_swevent_init_hrtimer(event); |
5895 | ||
b0a873eb | 5896 | return 0; |
6fb2915d LZ |
5897 | } |
5898 | ||
b0a873eb | 5899 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
5900 | .task_ctx_nr = perf_sw_context, |
5901 | ||
b0a873eb | 5902 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
5903 | .add = task_clock_event_add, |
5904 | .del = task_clock_event_del, | |
5905 | .start = task_clock_event_start, | |
5906 | .stop = task_clock_event_stop, | |
b0a873eb | 5907 | .read = task_clock_event_read, |
35edc2a5 PZ |
5908 | |
5909 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 5910 | }; |
6fb2915d | 5911 | |
ad5133b7 | 5912 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 5913 | { |
e077df4f | 5914 | } |
6fb2915d | 5915 | |
ad5133b7 | 5916 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 5917 | { |
ad5133b7 | 5918 | return 0; |
6fb2915d LZ |
5919 | } |
5920 | ||
ad5133b7 | 5921 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 5922 | { |
ad5133b7 | 5923 | perf_pmu_disable(pmu); |
6fb2915d LZ |
5924 | } |
5925 | ||
ad5133b7 PZ |
5926 | static int perf_pmu_commit_txn(struct pmu *pmu) |
5927 | { | |
5928 | perf_pmu_enable(pmu); | |
5929 | return 0; | |
5930 | } | |
e077df4f | 5931 | |
ad5133b7 | 5932 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 5933 | { |
ad5133b7 | 5934 | perf_pmu_enable(pmu); |
24f1e32c FW |
5935 | } |
5936 | ||
35edc2a5 PZ |
5937 | static int perf_event_idx_default(struct perf_event *event) |
5938 | { | |
5939 | return event->hw.idx + 1; | |
5940 | } | |
5941 | ||
8dc85d54 PZ |
5942 | /* |
5943 | * Ensures all contexts with the same task_ctx_nr have the same | |
5944 | * pmu_cpu_context too. | |
5945 | */ | |
5946 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 5947 | { |
8dc85d54 | 5948 | struct pmu *pmu; |
b326e956 | 5949 | |
8dc85d54 PZ |
5950 | if (ctxn < 0) |
5951 | return NULL; | |
24f1e32c | 5952 | |
8dc85d54 PZ |
5953 | list_for_each_entry(pmu, &pmus, entry) { |
5954 | if (pmu->task_ctx_nr == ctxn) | |
5955 | return pmu->pmu_cpu_context; | |
5956 | } | |
24f1e32c | 5957 | |
8dc85d54 | 5958 | return NULL; |
24f1e32c FW |
5959 | } |
5960 | ||
51676957 | 5961 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 5962 | { |
51676957 PZ |
5963 | int cpu; |
5964 | ||
5965 | for_each_possible_cpu(cpu) { | |
5966 | struct perf_cpu_context *cpuctx; | |
5967 | ||
5968 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
5969 | ||
3f1f3320 PZ |
5970 | if (cpuctx->unique_pmu == old_pmu) |
5971 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
5972 | } |
5973 | } | |
5974 | ||
5975 | static void free_pmu_context(struct pmu *pmu) | |
5976 | { | |
5977 | struct pmu *i; | |
f5ffe02e | 5978 | |
8dc85d54 | 5979 | mutex_lock(&pmus_lock); |
0475f9ea | 5980 | /* |
8dc85d54 | 5981 | * Like a real lame refcount. |
0475f9ea | 5982 | */ |
51676957 PZ |
5983 | list_for_each_entry(i, &pmus, entry) { |
5984 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
5985 | update_pmu_context(i, pmu); | |
8dc85d54 | 5986 | goto out; |
51676957 | 5987 | } |
8dc85d54 | 5988 | } |
d6d020e9 | 5989 | |
51676957 | 5990 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
5991 | out: |
5992 | mutex_unlock(&pmus_lock); | |
24f1e32c | 5993 | } |
2e80a82a | 5994 | static struct idr pmu_idr; |
d6d020e9 | 5995 | |
abe43400 PZ |
5996 | static ssize_t |
5997 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
5998 | { | |
5999 | struct pmu *pmu = dev_get_drvdata(dev); | |
6000 | ||
6001 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6002 | } | |
6003 | ||
6004 | static struct device_attribute pmu_dev_attrs[] = { | |
6005 | __ATTR_RO(type), | |
6006 | __ATTR_NULL, | |
6007 | }; | |
6008 | ||
6009 | static int pmu_bus_running; | |
6010 | static struct bus_type pmu_bus = { | |
6011 | .name = "event_source", | |
6012 | .dev_attrs = pmu_dev_attrs, | |
6013 | }; | |
6014 | ||
6015 | static void pmu_dev_release(struct device *dev) | |
6016 | { | |
6017 | kfree(dev); | |
6018 | } | |
6019 | ||
6020 | static int pmu_dev_alloc(struct pmu *pmu) | |
6021 | { | |
6022 | int ret = -ENOMEM; | |
6023 | ||
6024 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6025 | if (!pmu->dev) | |
6026 | goto out; | |
6027 | ||
0c9d42ed | 6028 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6029 | device_initialize(pmu->dev); |
6030 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6031 | if (ret) | |
6032 | goto free_dev; | |
6033 | ||
6034 | dev_set_drvdata(pmu->dev, pmu); | |
6035 | pmu->dev->bus = &pmu_bus; | |
6036 | pmu->dev->release = pmu_dev_release; | |
6037 | ret = device_add(pmu->dev); | |
6038 | if (ret) | |
6039 | goto free_dev; | |
6040 | ||
6041 | out: | |
6042 | return ret; | |
6043 | ||
6044 | free_dev: | |
6045 | put_device(pmu->dev); | |
6046 | goto out; | |
6047 | } | |
6048 | ||
547e9fd7 | 6049 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6050 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6051 | |
2e80a82a | 6052 | int perf_pmu_register(struct pmu *pmu, char *name, int type) |
24f1e32c | 6053 | { |
108b02cf | 6054 | int cpu, ret; |
24f1e32c | 6055 | |
b0a873eb | 6056 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6057 | ret = -ENOMEM; |
6058 | pmu->pmu_disable_count = alloc_percpu(int); | |
6059 | if (!pmu->pmu_disable_count) | |
6060 | goto unlock; | |
f29ac756 | 6061 | |
2e80a82a PZ |
6062 | pmu->type = -1; |
6063 | if (!name) | |
6064 | goto skip_type; | |
6065 | pmu->name = name; | |
6066 | ||
6067 | if (type < 0) { | |
0e9c3be2 TH |
6068 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6069 | if (type < 0) { | |
6070 | ret = type; | |
2e80a82a PZ |
6071 | goto free_pdc; |
6072 | } | |
6073 | } | |
6074 | pmu->type = type; | |
6075 | ||
abe43400 PZ |
6076 | if (pmu_bus_running) { |
6077 | ret = pmu_dev_alloc(pmu); | |
6078 | if (ret) | |
6079 | goto free_idr; | |
6080 | } | |
6081 | ||
2e80a82a | 6082 | skip_type: |
8dc85d54 PZ |
6083 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6084 | if (pmu->pmu_cpu_context) | |
6085 | goto got_cpu_context; | |
f29ac756 | 6086 | |
c4814202 | 6087 | ret = -ENOMEM; |
108b02cf PZ |
6088 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6089 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6090 | goto free_dev; |
f344011c | 6091 | |
108b02cf PZ |
6092 | for_each_possible_cpu(cpu) { |
6093 | struct perf_cpu_context *cpuctx; | |
6094 | ||
6095 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6096 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6097 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6098 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6099 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6100 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6101 | |
6102 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6103 | ||
e9d2b064 | 6104 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
3f1f3320 | 6105 | cpuctx->unique_pmu = pmu; |
108b02cf | 6106 | } |
76e1d904 | 6107 | |
8dc85d54 | 6108 | got_cpu_context: |
ad5133b7 PZ |
6109 | if (!pmu->start_txn) { |
6110 | if (pmu->pmu_enable) { | |
6111 | /* | |
6112 | * If we have pmu_enable/pmu_disable calls, install | |
6113 | * transaction stubs that use that to try and batch | |
6114 | * hardware accesses. | |
6115 | */ | |
6116 | pmu->start_txn = perf_pmu_start_txn; | |
6117 | pmu->commit_txn = perf_pmu_commit_txn; | |
6118 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6119 | } else { | |
6120 | pmu->start_txn = perf_pmu_nop_void; | |
6121 | pmu->commit_txn = perf_pmu_nop_int; | |
6122 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6123 | } |
5c92d124 | 6124 | } |
15dbf27c | 6125 | |
ad5133b7 PZ |
6126 | if (!pmu->pmu_enable) { |
6127 | pmu->pmu_enable = perf_pmu_nop_void; | |
6128 | pmu->pmu_disable = perf_pmu_nop_void; | |
6129 | } | |
6130 | ||
35edc2a5 PZ |
6131 | if (!pmu->event_idx) |
6132 | pmu->event_idx = perf_event_idx_default; | |
6133 | ||
b0a873eb | 6134 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6135 | ret = 0; |
6136 | unlock: | |
b0a873eb PZ |
6137 | mutex_unlock(&pmus_lock); |
6138 | ||
33696fc0 | 6139 | return ret; |
108b02cf | 6140 | |
abe43400 PZ |
6141 | free_dev: |
6142 | device_del(pmu->dev); | |
6143 | put_device(pmu->dev); | |
6144 | ||
2e80a82a PZ |
6145 | free_idr: |
6146 | if (pmu->type >= PERF_TYPE_MAX) | |
6147 | idr_remove(&pmu_idr, pmu->type); | |
6148 | ||
108b02cf PZ |
6149 | free_pdc: |
6150 | free_percpu(pmu->pmu_disable_count); | |
6151 | goto unlock; | |
f29ac756 PZ |
6152 | } |
6153 | ||
b0a873eb | 6154 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6155 | { |
b0a873eb PZ |
6156 | mutex_lock(&pmus_lock); |
6157 | list_del_rcu(&pmu->entry); | |
6158 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6159 | |
0475f9ea | 6160 | /* |
cde8e884 PZ |
6161 | * We dereference the pmu list under both SRCU and regular RCU, so |
6162 | * synchronize against both of those. | |
0475f9ea | 6163 | */ |
b0a873eb | 6164 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6165 | synchronize_rcu(); |
d6d020e9 | 6166 | |
33696fc0 | 6167 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6168 | if (pmu->type >= PERF_TYPE_MAX) |
6169 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6170 | device_del(pmu->dev); |
6171 | put_device(pmu->dev); | |
51676957 | 6172 | free_pmu_context(pmu); |
b0a873eb | 6173 | } |
d6d020e9 | 6174 | |
b0a873eb PZ |
6175 | struct pmu *perf_init_event(struct perf_event *event) |
6176 | { | |
6177 | struct pmu *pmu = NULL; | |
6178 | int idx; | |
940c5b29 | 6179 | int ret; |
b0a873eb PZ |
6180 | |
6181 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6182 | |
6183 | rcu_read_lock(); | |
6184 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6185 | rcu_read_unlock(); | |
940c5b29 | 6186 | if (pmu) { |
7e5b2a01 | 6187 | event->pmu = pmu; |
940c5b29 LM |
6188 | ret = pmu->event_init(event); |
6189 | if (ret) | |
6190 | pmu = ERR_PTR(ret); | |
2e80a82a | 6191 | goto unlock; |
940c5b29 | 6192 | } |
2e80a82a | 6193 | |
b0a873eb | 6194 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 6195 | event->pmu = pmu; |
940c5b29 | 6196 | ret = pmu->event_init(event); |
b0a873eb | 6197 | if (!ret) |
e5f4d339 | 6198 | goto unlock; |
76e1d904 | 6199 | |
b0a873eb PZ |
6200 | if (ret != -ENOENT) { |
6201 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6202 | goto unlock; |
f344011c | 6203 | } |
5c92d124 | 6204 | } |
e5f4d339 PZ |
6205 | pmu = ERR_PTR(-ENOENT); |
6206 | unlock: | |
b0a873eb | 6207 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6208 | |
4aeb0b42 | 6209 | return pmu; |
5c92d124 IM |
6210 | } |
6211 | ||
0793a61d | 6212 | /* |
cdd6c482 | 6213 | * Allocate and initialize a event structure |
0793a61d | 6214 | */ |
cdd6c482 | 6215 | static struct perf_event * |
c3f00c70 | 6216 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6217 | struct task_struct *task, |
6218 | struct perf_event *group_leader, | |
6219 | struct perf_event *parent_event, | |
4dc0da86 AK |
6220 | perf_overflow_handler_t overflow_handler, |
6221 | void *context) | |
0793a61d | 6222 | { |
51b0fe39 | 6223 | struct pmu *pmu; |
cdd6c482 IM |
6224 | struct perf_event *event; |
6225 | struct hw_perf_event *hwc; | |
d5d2bc0d | 6226 | long err; |
0793a61d | 6227 | |
66832eb4 ON |
6228 | if ((unsigned)cpu >= nr_cpu_ids) { |
6229 | if (!task || cpu != -1) | |
6230 | return ERR_PTR(-EINVAL); | |
6231 | } | |
6232 | ||
c3f00c70 | 6233 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6234 | if (!event) |
d5d2bc0d | 6235 | return ERR_PTR(-ENOMEM); |
0793a61d | 6236 | |
04289bb9 | 6237 | /* |
cdd6c482 | 6238 | * Single events are their own group leaders, with an |
04289bb9 IM |
6239 | * empty sibling list: |
6240 | */ | |
6241 | if (!group_leader) | |
cdd6c482 | 6242 | group_leader = event; |
04289bb9 | 6243 | |
cdd6c482 IM |
6244 | mutex_init(&event->child_mutex); |
6245 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6246 | |
cdd6c482 IM |
6247 | INIT_LIST_HEAD(&event->group_entry); |
6248 | INIT_LIST_HEAD(&event->event_entry); | |
6249 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
6250 | INIT_LIST_HEAD(&event->rb_entry); |
6251 | ||
cdd6c482 | 6252 | init_waitqueue_head(&event->waitq); |
e360adbe | 6253 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6254 | |
cdd6c482 | 6255 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6256 | |
a6fa941d | 6257 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6258 | event->cpu = cpu; |
6259 | event->attr = *attr; | |
6260 | event->group_leader = group_leader; | |
6261 | event->pmu = NULL; | |
cdd6c482 | 6262 | event->oncpu = -1; |
a96bbc16 | 6263 | |
cdd6c482 | 6264 | event->parent = parent_event; |
b84fbc9f | 6265 | |
17cf22c3 | 6266 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6267 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6268 | |
cdd6c482 | 6269 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6270 | |
d580ff86 PZ |
6271 | if (task) { |
6272 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6273 | |
6274 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6275 | event->hw.tp_target = task; | |
d580ff86 PZ |
6276 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6277 | /* | |
6278 | * hw_breakpoint is a bit difficult here.. | |
6279 | */ | |
f22c1bb6 | 6280 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6281 | event->hw.bp_target = task; |
6282 | #endif | |
6283 | } | |
6284 | ||
4dc0da86 | 6285 | if (!overflow_handler && parent_event) { |
b326e956 | 6286 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6287 | context = parent_event->overflow_handler_context; |
6288 | } | |
66832eb4 | 6289 | |
b326e956 | 6290 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6291 | event->overflow_handler_context = context; |
97eaf530 | 6292 | |
0231bb53 | 6293 | perf_event__state_init(event); |
a86ed508 | 6294 | |
4aeb0b42 | 6295 | pmu = NULL; |
b8e83514 | 6296 | |
cdd6c482 | 6297 | hwc = &event->hw; |
bd2b5b12 | 6298 | hwc->sample_period = attr->sample_period; |
0d48696f | 6299 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6300 | hwc->sample_period = 1; |
eced1dfc | 6301 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6302 | |
e7850595 | 6303 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6304 | |
2023b359 | 6305 | /* |
cdd6c482 | 6306 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6307 | */ |
3dab77fb | 6308 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
2023b359 PZ |
6309 | goto done; |
6310 | ||
b0a873eb | 6311 | pmu = perf_init_event(event); |
974802ea | 6312 | |
d5d2bc0d PM |
6313 | done: |
6314 | err = 0; | |
4aeb0b42 | 6315 | if (!pmu) |
d5d2bc0d | 6316 | err = -EINVAL; |
4aeb0b42 RR |
6317 | else if (IS_ERR(pmu)) |
6318 | err = PTR_ERR(pmu); | |
5c92d124 | 6319 | |
d5d2bc0d | 6320 | if (err) { |
cdd6c482 IM |
6321 | if (event->ns) |
6322 | put_pid_ns(event->ns); | |
6323 | kfree(event); | |
d5d2bc0d | 6324 | return ERR_PTR(err); |
621a01ea | 6325 | } |
d5d2bc0d | 6326 | |
cdd6c482 | 6327 | if (!event->parent) { |
82cd6def | 6328 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 6329 | static_key_slow_inc(&perf_sched_events.key); |
3af9e859 | 6330 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
6331 | atomic_inc(&nr_mmap_events); |
6332 | if (event->attr.comm) | |
6333 | atomic_inc(&nr_comm_events); | |
6334 | if (event->attr.task) | |
6335 | atomic_inc(&nr_task_events); | |
927c7a9e FW |
6336 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6337 | err = get_callchain_buffers(); | |
6338 | if (err) { | |
6339 | free_event(event); | |
6340 | return ERR_PTR(err); | |
6341 | } | |
6342 | } | |
d010b332 SE |
6343 | if (has_branch_stack(event)) { |
6344 | static_key_slow_inc(&perf_sched_events.key); | |
6345 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6346 | atomic_inc(&per_cpu(perf_branch_stack_events, | |
6347 | event->cpu)); | |
6348 | } | |
f344011c | 6349 | } |
9ee318a7 | 6350 | |
cdd6c482 | 6351 | return event; |
0793a61d TG |
6352 | } |
6353 | ||
cdd6c482 IM |
6354 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6355 | struct perf_event_attr *attr) | |
974802ea | 6356 | { |
974802ea | 6357 | u32 size; |
cdf8073d | 6358 | int ret; |
974802ea PZ |
6359 | |
6360 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6361 | return -EFAULT; | |
6362 | ||
6363 | /* | |
6364 | * zero the full structure, so that a short copy will be nice. | |
6365 | */ | |
6366 | memset(attr, 0, sizeof(*attr)); | |
6367 | ||
6368 | ret = get_user(size, &uattr->size); | |
6369 | if (ret) | |
6370 | return ret; | |
6371 | ||
6372 | if (size > PAGE_SIZE) /* silly large */ | |
6373 | goto err_size; | |
6374 | ||
6375 | if (!size) /* abi compat */ | |
6376 | size = PERF_ATTR_SIZE_VER0; | |
6377 | ||
6378 | if (size < PERF_ATTR_SIZE_VER0) | |
6379 | goto err_size; | |
6380 | ||
6381 | /* | |
6382 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6383 | * ensure all the unknown bits are 0 - i.e. new |
6384 | * user-space does not rely on any kernel feature | |
6385 | * extensions we dont know about yet. | |
974802ea PZ |
6386 | */ |
6387 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6388 | unsigned char __user *addr; |
6389 | unsigned char __user *end; | |
6390 | unsigned char val; | |
974802ea | 6391 | |
cdf8073d IS |
6392 | addr = (void __user *)uattr + sizeof(*attr); |
6393 | end = (void __user *)uattr + size; | |
974802ea | 6394 | |
cdf8073d | 6395 | for (; addr < end; addr++) { |
974802ea PZ |
6396 | ret = get_user(val, addr); |
6397 | if (ret) | |
6398 | return ret; | |
6399 | if (val) | |
6400 | goto err_size; | |
6401 | } | |
b3e62e35 | 6402 | size = sizeof(*attr); |
974802ea PZ |
6403 | } |
6404 | ||
6405 | ret = copy_from_user(attr, uattr, size); | |
6406 | if (ret) | |
6407 | return -EFAULT; | |
6408 | ||
cd757645 | 6409 | if (attr->__reserved_1) |
974802ea PZ |
6410 | return -EINVAL; |
6411 | ||
6412 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6413 | return -EINVAL; | |
6414 | ||
6415 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6416 | return -EINVAL; | |
6417 | ||
bce38cd5 SE |
6418 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6419 | u64 mask = attr->branch_sample_type; | |
6420 | ||
6421 | /* only using defined bits */ | |
6422 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6423 | return -EINVAL; | |
6424 | ||
6425 | /* at least one branch bit must be set */ | |
6426 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6427 | return -EINVAL; | |
6428 | ||
6429 | /* kernel level capture: check permissions */ | |
6430 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
6431 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6432 | return -EACCES; | |
6433 | ||
6434 | /* propagate priv level, when not set for branch */ | |
6435 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6436 | ||
6437 | /* exclude_kernel checked on syscall entry */ | |
6438 | if (!attr->exclude_kernel) | |
6439 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6440 | ||
6441 | if (!attr->exclude_user) | |
6442 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6443 | ||
6444 | if (!attr->exclude_hv) | |
6445 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6446 | /* | |
6447 | * adjust user setting (for HW filter setup) | |
6448 | */ | |
6449 | attr->branch_sample_type = mask; | |
6450 | } | |
6451 | } | |
4018994f | 6452 | |
c5ebcedb | 6453 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 6454 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
6455 | if (ret) |
6456 | return ret; | |
6457 | } | |
6458 | ||
6459 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
6460 | if (!arch_perf_have_user_stack_dump()) | |
6461 | return -ENOSYS; | |
6462 | ||
6463 | /* | |
6464 | * We have __u32 type for the size, but so far | |
6465 | * we can only use __u16 as maximum due to the | |
6466 | * __u16 sample size limit. | |
6467 | */ | |
6468 | if (attr->sample_stack_user >= USHRT_MAX) | |
6469 | ret = -EINVAL; | |
6470 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
6471 | ret = -EINVAL; | |
6472 | } | |
4018994f | 6473 | |
974802ea PZ |
6474 | out: |
6475 | return ret; | |
6476 | ||
6477 | err_size: | |
6478 | put_user(sizeof(*attr), &uattr->size); | |
6479 | ret = -E2BIG; | |
6480 | goto out; | |
6481 | } | |
6482 | ||
ac9721f3 PZ |
6483 | static int |
6484 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6485 | { |
76369139 | 6486 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6487 | int ret = -EINVAL; |
6488 | ||
ac9721f3 | 6489 | if (!output_event) |
a4be7c27 PZ |
6490 | goto set; |
6491 | ||
ac9721f3 PZ |
6492 | /* don't allow circular references */ |
6493 | if (event == output_event) | |
a4be7c27 PZ |
6494 | goto out; |
6495 | ||
0f139300 PZ |
6496 | /* |
6497 | * Don't allow cross-cpu buffers | |
6498 | */ | |
6499 | if (output_event->cpu != event->cpu) | |
6500 | goto out; | |
6501 | ||
6502 | /* | |
76369139 | 6503 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6504 | */ |
6505 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6506 | goto out; | |
6507 | ||
a4be7c27 | 6508 | set: |
cdd6c482 | 6509 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6510 | /* Can't redirect output if we've got an active mmap() */ |
6511 | if (atomic_read(&event->mmap_count)) | |
6512 | goto unlock; | |
a4be7c27 | 6513 | |
ac9721f3 | 6514 | if (output_event) { |
76369139 FW |
6515 | /* get the rb we want to redirect to */ |
6516 | rb = ring_buffer_get(output_event); | |
6517 | if (!rb) | |
ac9721f3 | 6518 | goto unlock; |
a4be7c27 PZ |
6519 | } |
6520 | ||
76369139 FW |
6521 | old_rb = event->rb; |
6522 | rcu_assign_pointer(event->rb, rb); | |
10c6db11 PZ |
6523 | if (old_rb) |
6524 | ring_buffer_detach(event, old_rb); | |
a4be7c27 | 6525 | ret = 0; |
ac9721f3 PZ |
6526 | unlock: |
6527 | mutex_unlock(&event->mmap_mutex); | |
6528 | ||
76369139 FW |
6529 | if (old_rb) |
6530 | ring_buffer_put(old_rb); | |
a4be7c27 | 6531 | out: |
a4be7c27 PZ |
6532 | return ret; |
6533 | } | |
6534 | ||
0793a61d | 6535 | /** |
cdd6c482 | 6536 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6537 | * |
cdd6c482 | 6538 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6539 | * @pid: target pid |
9f66a381 | 6540 | * @cpu: target cpu |
cdd6c482 | 6541 | * @group_fd: group leader event fd |
0793a61d | 6542 | */ |
cdd6c482 IM |
6543 | SYSCALL_DEFINE5(perf_event_open, |
6544 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 6545 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 6546 | { |
b04243ef PZ |
6547 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6548 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
6549 | struct perf_event_attr attr; |
6550 | struct perf_event_context *ctx; | |
6551 | struct file *event_file = NULL; | |
2903ff01 | 6552 | struct fd group = {NULL, 0}; |
38a81da2 | 6553 | struct task_struct *task = NULL; |
89a1e187 | 6554 | struct pmu *pmu; |
ea635c64 | 6555 | int event_fd; |
b04243ef | 6556 | int move_group = 0; |
dc86cabe | 6557 | int err; |
0793a61d | 6558 | |
2743a5b0 | 6559 | /* for future expandability... */ |
e5d1367f | 6560 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
6561 | return -EINVAL; |
6562 | ||
dc86cabe IM |
6563 | err = perf_copy_attr(attr_uptr, &attr); |
6564 | if (err) | |
6565 | return err; | |
eab656ae | 6566 | |
0764771d PZ |
6567 | if (!attr.exclude_kernel) { |
6568 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6569 | return -EACCES; | |
6570 | } | |
6571 | ||
df58ab24 | 6572 | if (attr.freq) { |
cdd6c482 | 6573 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
6574 | return -EINVAL; |
6575 | } | |
6576 | ||
e5d1367f SE |
6577 | /* |
6578 | * In cgroup mode, the pid argument is used to pass the fd | |
6579 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6580 | * designates the cpu on which to monitor threads from that | |
6581 | * cgroup. | |
6582 | */ | |
6583 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6584 | return -EINVAL; | |
6585 | ||
ab72a702 | 6586 | event_fd = get_unused_fd(); |
ea635c64 AV |
6587 | if (event_fd < 0) |
6588 | return event_fd; | |
6589 | ||
ac9721f3 | 6590 | if (group_fd != -1) { |
2903ff01 AV |
6591 | err = perf_fget_light(group_fd, &group); |
6592 | if (err) | |
d14b12d7 | 6593 | goto err_fd; |
2903ff01 | 6594 | group_leader = group.file->private_data; |
ac9721f3 PZ |
6595 | if (flags & PERF_FLAG_FD_OUTPUT) |
6596 | output_event = group_leader; | |
6597 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6598 | group_leader = NULL; | |
6599 | } | |
6600 | ||
e5d1367f | 6601 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6602 | task = find_lively_task_by_vpid(pid); |
6603 | if (IS_ERR(task)) { | |
6604 | err = PTR_ERR(task); | |
6605 | goto err_group_fd; | |
6606 | } | |
6607 | } | |
6608 | ||
fbfc623f YZ |
6609 | get_online_cpus(); |
6610 | ||
4dc0da86 AK |
6611 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6612 | NULL, NULL); | |
d14b12d7 SE |
6613 | if (IS_ERR(event)) { |
6614 | err = PTR_ERR(event); | |
c6be5a5c | 6615 | goto err_task; |
d14b12d7 SE |
6616 | } |
6617 | ||
e5d1367f SE |
6618 | if (flags & PERF_FLAG_PID_CGROUP) { |
6619 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
6620 | if (err) | |
6621 | goto err_alloc; | |
08309379 PZ |
6622 | /* |
6623 | * one more event: | |
6624 | * - that has cgroup constraint on event->cpu | |
6625 | * - that may need work on context switch | |
6626 | */ | |
6627 | atomic_inc(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 6628 | static_key_slow_inc(&perf_sched_events.key); |
e5d1367f SE |
6629 | } |
6630 | ||
89a1e187 PZ |
6631 | /* |
6632 | * Special case software events and allow them to be part of | |
6633 | * any hardware group. | |
6634 | */ | |
6635 | pmu = event->pmu; | |
b04243ef PZ |
6636 | |
6637 | if (group_leader && | |
6638 | (is_software_event(event) != is_software_event(group_leader))) { | |
6639 | if (is_software_event(event)) { | |
6640 | /* | |
6641 | * If event and group_leader are not both a software | |
6642 | * event, and event is, then group leader is not. | |
6643 | * | |
6644 | * Allow the addition of software events to !software | |
6645 | * groups, this is safe because software events never | |
6646 | * fail to schedule. | |
6647 | */ | |
6648 | pmu = group_leader->pmu; | |
6649 | } else if (is_software_event(group_leader) && | |
6650 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
6651 | /* | |
6652 | * In case the group is a pure software group, and we | |
6653 | * try to add a hardware event, move the whole group to | |
6654 | * the hardware context. | |
6655 | */ | |
6656 | move_group = 1; | |
6657 | } | |
6658 | } | |
89a1e187 PZ |
6659 | |
6660 | /* | |
6661 | * Get the target context (task or percpu): | |
6662 | */ | |
e2d37cd2 | 6663 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
6664 | if (IS_ERR(ctx)) { |
6665 | err = PTR_ERR(ctx); | |
c6be5a5c | 6666 | goto err_alloc; |
89a1e187 PZ |
6667 | } |
6668 | ||
fd1edb3a PZ |
6669 | if (task) { |
6670 | put_task_struct(task); | |
6671 | task = NULL; | |
6672 | } | |
6673 | ||
ccff286d | 6674 | /* |
cdd6c482 | 6675 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 6676 | */ |
ac9721f3 | 6677 | if (group_leader) { |
dc86cabe | 6678 | err = -EINVAL; |
04289bb9 | 6679 | |
04289bb9 | 6680 | /* |
ccff286d IM |
6681 | * Do not allow a recursive hierarchy (this new sibling |
6682 | * becoming part of another group-sibling): | |
6683 | */ | |
6684 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 6685 | goto err_context; |
ccff286d IM |
6686 | /* |
6687 | * Do not allow to attach to a group in a different | |
6688 | * task or CPU context: | |
04289bb9 | 6689 | */ |
b04243ef PZ |
6690 | if (move_group) { |
6691 | if (group_leader->ctx->type != ctx->type) | |
6692 | goto err_context; | |
6693 | } else { | |
6694 | if (group_leader->ctx != ctx) | |
6695 | goto err_context; | |
6696 | } | |
6697 | ||
3b6f9e5c PM |
6698 | /* |
6699 | * Only a group leader can be exclusive or pinned | |
6700 | */ | |
0d48696f | 6701 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 6702 | goto err_context; |
ac9721f3 PZ |
6703 | } |
6704 | ||
6705 | if (output_event) { | |
6706 | err = perf_event_set_output(event, output_event); | |
6707 | if (err) | |
c3f00c70 | 6708 | goto err_context; |
ac9721f3 | 6709 | } |
0793a61d | 6710 | |
ea635c64 AV |
6711 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
6712 | if (IS_ERR(event_file)) { | |
6713 | err = PTR_ERR(event_file); | |
c3f00c70 | 6714 | goto err_context; |
ea635c64 | 6715 | } |
9b51f66d | 6716 | |
b04243ef PZ |
6717 | if (move_group) { |
6718 | struct perf_event_context *gctx = group_leader->ctx; | |
6719 | ||
6720 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 6721 | perf_remove_from_context(group_leader); |
0231bb53 JO |
6722 | |
6723 | /* | |
6724 | * Removing from the context ends up with disabled | |
6725 | * event. What we want here is event in the initial | |
6726 | * startup state, ready to be add into new context. | |
6727 | */ | |
6728 | perf_event__state_init(group_leader); | |
b04243ef PZ |
6729 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6730 | group_entry) { | |
fe4b04fa | 6731 | perf_remove_from_context(sibling); |
0231bb53 | 6732 | perf_event__state_init(sibling); |
b04243ef PZ |
6733 | put_ctx(gctx); |
6734 | } | |
6735 | mutex_unlock(&gctx->mutex); | |
6736 | put_ctx(gctx); | |
ea635c64 | 6737 | } |
9b51f66d | 6738 | |
ad3a37de | 6739 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 6740 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
6741 | |
6742 | if (move_group) { | |
0cda4c02 | 6743 | synchronize_rcu(); |
e2d37cd2 | 6744 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
6745 | get_ctx(ctx); |
6746 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
6747 | group_entry) { | |
e2d37cd2 | 6748 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
6749 | get_ctx(ctx); |
6750 | } | |
6751 | } | |
6752 | ||
e2d37cd2 | 6753 | perf_install_in_context(ctx, event, event->cpu); |
ad3a37de | 6754 | ++ctx->generation; |
fe4b04fa | 6755 | perf_unpin_context(ctx); |
d859e29f | 6756 | mutex_unlock(&ctx->mutex); |
9b51f66d | 6757 | |
fbfc623f YZ |
6758 | put_online_cpus(); |
6759 | ||
cdd6c482 | 6760 | event->owner = current; |
8882135b | 6761 | |
cdd6c482 IM |
6762 | mutex_lock(¤t->perf_event_mutex); |
6763 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
6764 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 6765 | |
c320c7b7 ACM |
6766 | /* |
6767 | * Precalculate sample_data sizes | |
6768 | */ | |
6769 | perf_event__header_size(event); | |
6844c09d | 6770 | perf_event__id_header_size(event); |
c320c7b7 | 6771 | |
8a49542c PZ |
6772 | /* |
6773 | * Drop the reference on the group_event after placing the | |
6774 | * new event on the sibling_list. This ensures destruction | |
6775 | * of the group leader will find the pointer to itself in | |
6776 | * perf_group_detach(). | |
6777 | */ | |
2903ff01 | 6778 | fdput(group); |
ea635c64 AV |
6779 | fd_install(event_fd, event_file); |
6780 | return event_fd; | |
0793a61d | 6781 | |
c3f00c70 | 6782 | err_context: |
fe4b04fa | 6783 | perf_unpin_context(ctx); |
ea635c64 | 6784 | put_ctx(ctx); |
c6be5a5c | 6785 | err_alloc: |
ea635c64 | 6786 | free_event(event); |
e7d0bc04 | 6787 | err_task: |
fbfc623f | 6788 | put_online_cpus(); |
e7d0bc04 PZ |
6789 | if (task) |
6790 | put_task_struct(task); | |
89a1e187 | 6791 | err_group_fd: |
2903ff01 | 6792 | fdput(group); |
ea635c64 AV |
6793 | err_fd: |
6794 | put_unused_fd(event_fd); | |
dc86cabe | 6795 | return err; |
0793a61d TG |
6796 | } |
6797 | ||
fb0459d7 AV |
6798 | /** |
6799 | * perf_event_create_kernel_counter | |
6800 | * | |
6801 | * @attr: attributes of the counter to create | |
6802 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 6803 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
6804 | */ |
6805 | struct perf_event * | |
6806 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 6807 | struct task_struct *task, |
4dc0da86 AK |
6808 | perf_overflow_handler_t overflow_handler, |
6809 | void *context) | |
fb0459d7 | 6810 | { |
fb0459d7 | 6811 | struct perf_event_context *ctx; |
c3f00c70 | 6812 | struct perf_event *event; |
fb0459d7 | 6813 | int err; |
d859e29f | 6814 | |
fb0459d7 AV |
6815 | /* |
6816 | * Get the target context (task or percpu): | |
6817 | */ | |
d859e29f | 6818 | |
4dc0da86 AK |
6819 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
6820 | overflow_handler, context); | |
c3f00c70 PZ |
6821 | if (IS_ERR(event)) { |
6822 | err = PTR_ERR(event); | |
6823 | goto err; | |
6824 | } | |
d859e29f | 6825 | |
38a81da2 | 6826 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
6827 | if (IS_ERR(ctx)) { |
6828 | err = PTR_ERR(ctx); | |
c3f00c70 | 6829 | goto err_free; |
d859e29f | 6830 | } |
fb0459d7 | 6831 | |
fb0459d7 AV |
6832 | WARN_ON_ONCE(ctx->parent_ctx); |
6833 | mutex_lock(&ctx->mutex); | |
6834 | perf_install_in_context(ctx, event, cpu); | |
6835 | ++ctx->generation; | |
fe4b04fa | 6836 | perf_unpin_context(ctx); |
fb0459d7 AV |
6837 | mutex_unlock(&ctx->mutex); |
6838 | ||
fb0459d7 AV |
6839 | return event; |
6840 | ||
c3f00c70 PZ |
6841 | err_free: |
6842 | free_event(event); | |
6843 | err: | |
c6567f64 | 6844 | return ERR_PTR(err); |
9b51f66d | 6845 | } |
fb0459d7 | 6846 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 6847 | |
0cda4c02 YZ |
6848 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
6849 | { | |
6850 | struct perf_event_context *src_ctx; | |
6851 | struct perf_event_context *dst_ctx; | |
6852 | struct perf_event *event, *tmp; | |
6853 | LIST_HEAD(events); | |
6854 | ||
6855 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
6856 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
6857 | ||
6858 | mutex_lock(&src_ctx->mutex); | |
6859 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
6860 | event_entry) { | |
6861 | perf_remove_from_context(event); | |
6862 | put_ctx(src_ctx); | |
6863 | list_add(&event->event_entry, &events); | |
6864 | } | |
6865 | mutex_unlock(&src_ctx->mutex); | |
6866 | ||
6867 | synchronize_rcu(); | |
6868 | ||
6869 | mutex_lock(&dst_ctx->mutex); | |
6870 | list_for_each_entry_safe(event, tmp, &events, event_entry) { | |
6871 | list_del(&event->event_entry); | |
6872 | if (event->state >= PERF_EVENT_STATE_OFF) | |
6873 | event->state = PERF_EVENT_STATE_INACTIVE; | |
6874 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
6875 | get_ctx(dst_ctx); | |
6876 | } | |
6877 | mutex_unlock(&dst_ctx->mutex); | |
6878 | } | |
6879 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
6880 | ||
cdd6c482 | 6881 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 6882 | struct task_struct *child) |
d859e29f | 6883 | { |
cdd6c482 | 6884 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 6885 | u64 child_val; |
d859e29f | 6886 | |
cdd6c482 IM |
6887 | if (child_event->attr.inherit_stat) |
6888 | perf_event_read_event(child_event, child); | |
38b200d6 | 6889 | |
b5e58793 | 6890 | child_val = perf_event_count(child_event); |
d859e29f PM |
6891 | |
6892 | /* | |
6893 | * Add back the child's count to the parent's count: | |
6894 | */ | |
a6e6dea6 | 6895 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
6896 | atomic64_add(child_event->total_time_enabled, |
6897 | &parent_event->child_total_time_enabled); | |
6898 | atomic64_add(child_event->total_time_running, | |
6899 | &parent_event->child_total_time_running); | |
d859e29f PM |
6900 | |
6901 | /* | |
cdd6c482 | 6902 | * Remove this event from the parent's list |
d859e29f | 6903 | */ |
cdd6c482 IM |
6904 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
6905 | mutex_lock(&parent_event->child_mutex); | |
6906 | list_del_init(&child_event->child_list); | |
6907 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
6908 | |
6909 | /* | |
cdd6c482 | 6910 | * Release the parent event, if this was the last |
d859e29f PM |
6911 | * reference to it. |
6912 | */ | |
a6fa941d | 6913 | put_event(parent_event); |
d859e29f PM |
6914 | } |
6915 | ||
9b51f66d | 6916 | static void |
cdd6c482 IM |
6917 | __perf_event_exit_task(struct perf_event *child_event, |
6918 | struct perf_event_context *child_ctx, | |
38b200d6 | 6919 | struct task_struct *child) |
9b51f66d | 6920 | { |
38b435b1 PZ |
6921 | if (child_event->parent) { |
6922 | raw_spin_lock_irq(&child_ctx->lock); | |
6923 | perf_group_detach(child_event); | |
6924 | raw_spin_unlock_irq(&child_ctx->lock); | |
6925 | } | |
9b51f66d | 6926 | |
fe4b04fa | 6927 | perf_remove_from_context(child_event); |
0cc0c027 | 6928 | |
9b51f66d | 6929 | /* |
38b435b1 | 6930 | * It can happen that the parent exits first, and has events |
9b51f66d | 6931 | * that are still around due to the child reference. These |
38b435b1 | 6932 | * events need to be zapped. |
9b51f66d | 6933 | */ |
38b435b1 | 6934 | if (child_event->parent) { |
cdd6c482 IM |
6935 | sync_child_event(child_event, child); |
6936 | free_event(child_event); | |
4bcf349a | 6937 | } |
9b51f66d IM |
6938 | } |
6939 | ||
8dc85d54 | 6940 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 6941 | { |
cdd6c482 IM |
6942 | struct perf_event *child_event, *tmp; |
6943 | struct perf_event_context *child_ctx; | |
a63eaf34 | 6944 | unsigned long flags; |
9b51f66d | 6945 | |
8dc85d54 | 6946 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 6947 | perf_event_task(child, NULL, 0); |
9b51f66d | 6948 | return; |
9f498cc5 | 6949 | } |
9b51f66d | 6950 | |
a63eaf34 | 6951 | local_irq_save(flags); |
ad3a37de PM |
6952 | /* |
6953 | * We can't reschedule here because interrupts are disabled, | |
6954 | * and either child is current or it is a task that can't be | |
6955 | * scheduled, so we are now safe from rescheduling changing | |
6956 | * our context. | |
6957 | */ | |
806839b2 | 6958 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
6959 | |
6960 | /* | |
6961 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 6962 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
6963 | * incremented the context's refcount before we do put_ctx below. |
6964 | */ | |
e625cce1 | 6965 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 6966 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 6967 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
6968 | /* |
6969 | * If this context is a clone; unclone it so it can't get | |
6970 | * swapped to another process while we're removing all | |
cdd6c482 | 6971 | * the events from it. |
71a851b4 PZ |
6972 | */ |
6973 | unclone_ctx(child_ctx); | |
5e942bb3 | 6974 | update_context_time(child_ctx); |
e625cce1 | 6975 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
6976 | |
6977 | /* | |
cdd6c482 IM |
6978 | * Report the task dead after unscheduling the events so that we |
6979 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
6980 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 6981 | */ |
cdd6c482 | 6982 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 6983 | |
66fff224 PZ |
6984 | /* |
6985 | * We can recurse on the same lock type through: | |
6986 | * | |
cdd6c482 IM |
6987 | * __perf_event_exit_task() |
6988 | * sync_child_event() | |
a6fa941d AV |
6989 | * put_event() |
6990 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
6991 | * |
6992 | * But since its the parent context it won't be the same instance. | |
6993 | */ | |
a0507c84 | 6994 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 6995 | |
8bc20959 | 6996 | again: |
889ff015 FW |
6997 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
6998 | group_entry) | |
6999 | __perf_event_exit_task(child_event, child_ctx, child); | |
7000 | ||
7001 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 7002 | group_entry) |
cdd6c482 | 7003 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
7004 | |
7005 | /* | |
cdd6c482 | 7006 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
7007 | * its siblings to the list, but we obtained 'tmp' before that which |
7008 | * will still point to the list head terminating the iteration. | |
7009 | */ | |
889ff015 FW |
7010 | if (!list_empty(&child_ctx->pinned_groups) || |
7011 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 7012 | goto again; |
a63eaf34 PM |
7013 | |
7014 | mutex_unlock(&child_ctx->mutex); | |
7015 | ||
7016 | put_ctx(child_ctx); | |
9b51f66d IM |
7017 | } |
7018 | ||
8dc85d54 PZ |
7019 | /* |
7020 | * When a child task exits, feed back event values to parent events. | |
7021 | */ | |
7022 | void perf_event_exit_task(struct task_struct *child) | |
7023 | { | |
8882135b | 7024 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
7025 | int ctxn; |
7026 | ||
8882135b PZ |
7027 | mutex_lock(&child->perf_event_mutex); |
7028 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
7029 | owner_entry) { | |
7030 | list_del_init(&event->owner_entry); | |
7031 | ||
7032 | /* | |
7033 | * Ensure the list deletion is visible before we clear | |
7034 | * the owner, closes a race against perf_release() where | |
7035 | * we need to serialize on the owner->perf_event_mutex. | |
7036 | */ | |
7037 | smp_wmb(); | |
7038 | event->owner = NULL; | |
7039 | } | |
7040 | mutex_unlock(&child->perf_event_mutex); | |
7041 | ||
8dc85d54 PZ |
7042 | for_each_task_context_nr(ctxn) |
7043 | perf_event_exit_task_context(child, ctxn); | |
7044 | } | |
7045 | ||
889ff015 FW |
7046 | static void perf_free_event(struct perf_event *event, |
7047 | struct perf_event_context *ctx) | |
7048 | { | |
7049 | struct perf_event *parent = event->parent; | |
7050 | ||
7051 | if (WARN_ON_ONCE(!parent)) | |
7052 | return; | |
7053 | ||
7054 | mutex_lock(&parent->child_mutex); | |
7055 | list_del_init(&event->child_list); | |
7056 | mutex_unlock(&parent->child_mutex); | |
7057 | ||
a6fa941d | 7058 | put_event(parent); |
889ff015 | 7059 | |
8a49542c | 7060 | perf_group_detach(event); |
889ff015 FW |
7061 | list_del_event(event, ctx); |
7062 | free_event(event); | |
7063 | } | |
7064 | ||
bbbee908 PZ |
7065 | /* |
7066 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 7067 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 7068 | */ |
cdd6c482 | 7069 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 7070 | { |
8dc85d54 | 7071 | struct perf_event_context *ctx; |
cdd6c482 | 7072 | struct perf_event *event, *tmp; |
8dc85d54 | 7073 | int ctxn; |
bbbee908 | 7074 | |
8dc85d54 PZ |
7075 | for_each_task_context_nr(ctxn) { |
7076 | ctx = task->perf_event_ctxp[ctxn]; | |
7077 | if (!ctx) | |
7078 | continue; | |
bbbee908 | 7079 | |
8dc85d54 | 7080 | mutex_lock(&ctx->mutex); |
bbbee908 | 7081 | again: |
8dc85d54 PZ |
7082 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
7083 | group_entry) | |
7084 | perf_free_event(event, ctx); | |
bbbee908 | 7085 | |
8dc85d54 PZ |
7086 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
7087 | group_entry) | |
7088 | perf_free_event(event, ctx); | |
bbbee908 | 7089 | |
8dc85d54 PZ |
7090 | if (!list_empty(&ctx->pinned_groups) || |
7091 | !list_empty(&ctx->flexible_groups)) | |
7092 | goto again; | |
bbbee908 | 7093 | |
8dc85d54 | 7094 | mutex_unlock(&ctx->mutex); |
bbbee908 | 7095 | |
8dc85d54 PZ |
7096 | put_ctx(ctx); |
7097 | } | |
889ff015 FW |
7098 | } |
7099 | ||
4e231c79 PZ |
7100 | void perf_event_delayed_put(struct task_struct *task) |
7101 | { | |
7102 | int ctxn; | |
7103 | ||
7104 | for_each_task_context_nr(ctxn) | |
7105 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7106 | } | |
7107 | ||
97dee4f3 PZ |
7108 | /* |
7109 | * inherit a event from parent task to child task: | |
7110 | */ | |
7111 | static struct perf_event * | |
7112 | inherit_event(struct perf_event *parent_event, | |
7113 | struct task_struct *parent, | |
7114 | struct perf_event_context *parent_ctx, | |
7115 | struct task_struct *child, | |
7116 | struct perf_event *group_leader, | |
7117 | struct perf_event_context *child_ctx) | |
7118 | { | |
7119 | struct perf_event *child_event; | |
cee010ec | 7120 | unsigned long flags; |
97dee4f3 PZ |
7121 | |
7122 | /* | |
7123 | * Instead of creating recursive hierarchies of events, | |
7124 | * we link inherited events back to the original parent, | |
7125 | * which has a filp for sure, which we use as the reference | |
7126 | * count: | |
7127 | */ | |
7128 | if (parent_event->parent) | |
7129 | parent_event = parent_event->parent; | |
7130 | ||
7131 | child_event = perf_event_alloc(&parent_event->attr, | |
7132 | parent_event->cpu, | |
d580ff86 | 7133 | child, |
97dee4f3 | 7134 | group_leader, parent_event, |
4dc0da86 | 7135 | NULL, NULL); |
97dee4f3 PZ |
7136 | if (IS_ERR(child_event)) |
7137 | return child_event; | |
a6fa941d AV |
7138 | |
7139 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7140 | free_event(child_event); | |
7141 | return NULL; | |
7142 | } | |
7143 | ||
97dee4f3 PZ |
7144 | get_ctx(child_ctx); |
7145 | ||
7146 | /* | |
7147 | * Make the child state follow the state of the parent event, | |
7148 | * not its attr.disabled bit. We hold the parent's mutex, | |
7149 | * so we won't race with perf_event_{en, dis}able_family. | |
7150 | */ | |
7151 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7152 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7153 | else | |
7154 | child_event->state = PERF_EVENT_STATE_OFF; | |
7155 | ||
7156 | if (parent_event->attr.freq) { | |
7157 | u64 sample_period = parent_event->hw.sample_period; | |
7158 | struct hw_perf_event *hwc = &child_event->hw; | |
7159 | ||
7160 | hwc->sample_period = sample_period; | |
7161 | hwc->last_period = sample_period; | |
7162 | ||
7163 | local64_set(&hwc->period_left, sample_period); | |
7164 | } | |
7165 | ||
7166 | child_event->ctx = child_ctx; | |
7167 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7168 | child_event->overflow_handler_context |
7169 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7170 | |
614b6780 TG |
7171 | /* |
7172 | * Precalculate sample_data sizes | |
7173 | */ | |
7174 | perf_event__header_size(child_event); | |
6844c09d | 7175 | perf_event__id_header_size(child_event); |
614b6780 | 7176 | |
97dee4f3 PZ |
7177 | /* |
7178 | * Link it up in the child's context: | |
7179 | */ | |
cee010ec | 7180 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7181 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7182 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7183 | |
97dee4f3 PZ |
7184 | /* |
7185 | * Link this into the parent event's child list | |
7186 | */ | |
7187 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7188 | mutex_lock(&parent_event->child_mutex); | |
7189 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7190 | mutex_unlock(&parent_event->child_mutex); | |
7191 | ||
7192 | return child_event; | |
7193 | } | |
7194 | ||
7195 | static int inherit_group(struct perf_event *parent_event, | |
7196 | struct task_struct *parent, | |
7197 | struct perf_event_context *parent_ctx, | |
7198 | struct task_struct *child, | |
7199 | struct perf_event_context *child_ctx) | |
7200 | { | |
7201 | struct perf_event *leader; | |
7202 | struct perf_event *sub; | |
7203 | struct perf_event *child_ctr; | |
7204 | ||
7205 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7206 | child, NULL, child_ctx); | |
7207 | if (IS_ERR(leader)) | |
7208 | return PTR_ERR(leader); | |
7209 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7210 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7211 | child, leader, child_ctx); | |
7212 | if (IS_ERR(child_ctr)) | |
7213 | return PTR_ERR(child_ctr); | |
7214 | } | |
7215 | return 0; | |
889ff015 FW |
7216 | } |
7217 | ||
7218 | static int | |
7219 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7220 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7221 | struct task_struct *child, int ctxn, |
889ff015 FW |
7222 | int *inherited_all) |
7223 | { | |
7224 | int ret; | |
8dc85d54 | 7225 | struct perf_event_context *child_ctx; |
889ff015 FW |
7226 | |
7227 | if (!event->attr.inherit) { | |
7228 | *inherited_all = 0; | |
7229 | return 0; | |
bbbee908 PZ |
7230 | } |
7231 | ||
fe4b04fa | 7232 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7233 | if (!child_ctx) { |
7234 | /* | |
7235 | * This is executed from the parent task context, so | |
7236 | * inherit events that have been marked for cloning. | |
7237 | * First allocate and initialize a context for the | |
7238 | * child. | |
7239 | */ | |
bbbee908 | 7240 | |
eb184479 | 7241 | child_ctx = alloc_perf_context(event->pmu, child); |
889ff015 FW |
7242 | if (!child_ctx) |
7243 | return -ENOMEM; | |
bbbee908 | 7244 | |
8dc85d54 | 7245 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7246 | } |
7247 | ||
7248 | ret = inherit_group(event, parent, parent_ctx, | |
7249 | child, child_ctx); | |
7250 | ||
7251 | if (ret) | |
7252 | *inherited_all = 0; | |
7253 | ||
7254 | return ret; | |
bbbee908 PZ |
7255 | } |
7256 | ||
9b51f66d | 7257 | /* |
cdd6c482 | 7258 | * Initialize the perf_event context in task_struct |
9b51f66d | 7259 | */ |
8dc85d54 | 7260 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7261 | { |
889ff015 | 7262 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7263 | struct perf_event_context *cloned_ctx; |
7264 | struct perf_event *event; | |
9b51f66d | 7265 | struct task_struct *parent = current; |
564c2b21 | 7266 | int inherited_all = 1; |
dddd3379 | 7267 | unsigned long flags; |
6ab423e0 | 7268 | int ret = 0; |
9b51f66d | 7269 | |
8dc85d54 | 7270 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7271 | return 0; |
7272 | ||
ad3a37de | 7273 | /* |
25346b93 PM |
7274 | * If the parent's context is a clone, pin it so it won't get |
7275 | * swapped under us. | |
ad3a37de | 7276 | */ |
8dc85d54 | 7277 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 7278 | |
ad3a37de PM |
7279 | /* |
7280 | * No need to check if parent_ctx != NULL here; since we saw | |
7281 | * it non-NULL earlier, the only reason for it to become NULL | |
7282 | * is if we exit, and since we're currently in the middle of | |
7283 | * a fork we can't be exiting at the same time. | |
7284 | */ | |
ad3a37de | 7285 | |
9b51f66d IM |
7286 | /* |
7287 | * Lock the parent list. No need to lock the child - not PID | |
7288 | * hashed yet and not running, so nobody can access it. | |
7289 | */ | |
d859e29f | 7290 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7291 | |
7292 | /* | |
7293 | * We dont have to disable NMIs - we are only looking at | |
7294 | * the list, not manipulating it: | |
7295 | */ | |
889ff015 | 7296 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7297 | ret = inherit_task_group(event, parent, parent_ctx, |
7298 | child, ctxn, &inherited_all); | |
889ff015 FW |
7299 | if (ret) |
7300 | break; | |
7301 | } | |
b93f7978 | 7302 | |
dddd3379 TG |
7303 | /* |
7304 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7305 | * to allocations, but we need to prevent rotation because | |
7306 | * rotate_ctx() will change the list from interrupt context. | |
7307 | */ | |
7308 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7309 | parent_ctx->rotate_disable = 1; | |
7310 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7311 | ||
889ff015 | 7312 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7313 | ret = inherit_task_group(event, parent, parent_ctx, |
7314 | child, ctxn, &inherited_all); | |
889ff015 | 7315 | if (ret) |
9b51f66d | 7316 | break; |
564c2b21 PM |
7317 | } |
7318 | ||
dddd3379 TG |
7319 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7320 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7321 | |
8dc85d54 | 7322 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7323 | |
05cbaa28 | 7324 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7325 | /* |
7326 | * Mark the child context as a clone of the parent | |
7327 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7328 | * |
7329 | * Note that if the parent is a clone, the holding of | |
7330 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7331 | */ |
c5ed5145 | 7332 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7333 | if (cloned_ctx) { |
7334 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7335 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7336 | } else { |
7337 | child_ctx->parent_ctx = parent_ctx; | |
7338 | child_ctx->parent_gen = parent_ctx->generation; | |
7339 | } | |
7340 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7341 | } |
7342 | ||
c5ed5145 | 7343 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7344 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7345 | |
25346b93 | 7346 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7347 | put_ctx(parent_ctx); |
ad3a37de | 7348 | |
6ab423e0 | 7349 | return ret; |
9b51f66d IM |
7350 | } |
7351 | ||
8dc85d54 PZ |
7352 | /* |
7353 | * Initialize the perf_event context in task_struct | |
7354 | */ | |
7355 | int perf_event_init_task(struct task_struct *child) | |
7356 | { | |
7357 | int ctxn, ret; | |
7358 | ||
8550d7cb ON |
7359 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7360 | mutex_init(&child->perf_event_mutex); | |
7361 | INIT_LIST_HEAD(&child->perf_event_list); | |
7362 | ||
8dc85d54 PZ |
7363 | for_each_task_context_nr(ctxn) { |
7364 | ret = perf_event_init_context(child, ctxn); | |
7365 | if (ret) | |
7366 | return ret; | |
7367 | } | |
7368 | ||
7369 | return 0; | |
7370 | } | |
7371 | ||
220b140b PM |
7372 | static void __init perf_event_init_all_cpus(void) |
7373 | { | |
b28ab83c | 7374 | struct swevent_htable *swhash; |
220b140b | 7375 | int cpu; |
220b140b PM |
7376 | |
7377 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7378 | swhash = &per_cpu(swevent_htable, cpu); |
7379 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7380 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7381 | } |
7382 | } | |
7383 | ||
cdd6c482 | 7384 | static void __cpuinit perf_event_init_cpu(int cpu) |
0793a61d | 7385 | { |
108b02cf | 7386 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7387 | |
b28ab83c | 7388 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7389 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7390 | struct swevent_hlist *hlist; |
7391 | ||
b28ab83c PZ |
7392 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7393 | WARN_ON(!hlist); | |
7394 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7395 | } |
b28ab83c | 7396 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7397 | } |
7398 | ||
c277443c | 7399 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7400 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7401 | { |
e9d2b064 PZ |
7402 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7403 | ||
7404 | WARN_ON(!irqs_disabled()); | |
7405 | ||
7406 | list_del_init(&cpuctx->rotation_list); | |
7407 | } | |
7408 | ||
108b02cf | 7409 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7410 | { |
108b02cf | 7411 | struct perf_event_context *ctx = __info; |
cdd6c482 | 7412 | struct perf_event *event, *tmp; |
0793a61d | 7413 | |
108b02cf | 7414 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7415 | |
889ff015 | 7416 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 7417 | __perf_remove_from_context(event); |
889ff015 | 7418 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 7419 | __perf_remove_from_context(event); |
0793a61d | 7420 | } |
108b02cf PZ |
7421 | |
7422 | static void perf_event_exit_cpu_context(int cpu) | |
7423 | { | |
7424 | struct perf_event_context *ctx; | |
7425 | struct pmu *pmu; | |
7426 | int idx; | |
7427 | ||
7428 | idx = srcu_read_lock(&pmus_srcu); | |
7429 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7430 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7431 | |
7432 | mutex_lock(&ctx->mutex); | |
7433 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7434 | mutex_unlock(&ctx->mutex); | |
7435 | } | |
7436 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7437 | } |
7438 | ||
cdd6c482 | 7439 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7440 | { |
b28ab83c | 7441 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7442 | |
b28ab83c PZ |
7443 | mutex_lock(&swhash->hlist_mutex); |
7444 | swevent_hlist_release(swhash); | |
7445 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 7446 | |
108b02cf | 7447 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7448 | } |
7449 | #else | |
cdd6c482 | 7450 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7451 | #endif |
7452 | ||
c277443c PZ |
7453 | static int |
7454 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7455 | { | |
7456 | int cpu; | |
7457 | ||
7458 | for_each_online_cpu(cpu) | |
7459 | perf_event_exit_cpu(cpu); | |
7460 | ||
7461 | return NOTIFY_OK; | |
7462 | } | |
7463 | ||
7464 | /* | |
7465 | * Run the perf reboot notifier at the very last possible moment so that | |
7466 | * the generic watchdog code runs as long as possible. | |
7467 | */ | |
7468 | static struct notifier_block perf_reboot_notifier = { | |
7469 | .notifier_call = perf_reboot, | |
7470 | .priority = INT_MIN, | |
7471 | }; | |
7472 | ||
0793a61d TG |
7473 | static int __cpuinit |
7474 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
7475 | { | |
7476 | unsigned int cpu = (long)hcpu; | |
7477 | ||
4536e4d1 | 7478 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7479 | |
7480 | case CPU_UP_PREPARE: | |
5e11637e | 7481 | case CPU_DOWN_FAILED: |
cdd6c482 | 7482 | perf_event_init_cpu(cpu); |
0793a61d TG |
7483 | break; |
7484 | ||
5e11637e | 7485 | case CPU_UP_CANCELED: |
0793a61d | 7486 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7487 | perf_event_exit_cpu(cpu); |
0793a61d | 7488 | break; |
0793a61d TG |
7489 | default: |
7490 | break; | |
7491 | } | |
7492 | ||
7493 | return NOTIFY_OK; | |
7494 | } | |
7495 | ||
cdd6c482 | 7496 | void __init perf_event_init(void) |
0793a61d | 7497 | { |
3c502e7a JW |
7498 | int ret; |
7499 | ||
2e80a82a PZ |
7500 | idr_init(&pmu_idr); |
7501 | ||
220b140b | 7502 | perf_event_init_all_cpus(); |
b0a873eb | 7503 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7504 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7505 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7506 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7507 | perf_tp_register(); |
7508 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7509 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7510 | |
7511 | ret = init_hw_breakpoint(); | |
7512 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7513 | |
7514 | /* do not patch jump label more than once per second */ | |
7515 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7516 | |
7517 | /* | |
7518 | * Build time assertion that we keep the data_head at the intended | |
7519 | * location. IOW, validation we got the __reserved[] size right. | |
7520 | */ | |
7521 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
7522 | != 1024); | |
0793a61d | 7523 | } |
abe43400 PZ |
7524 | |
7525 | static int __init perf_event_sysfs_init(void) | |
7526 | { | |
7527 | struct pmu *pmu; | |
7528 | int ret; | |
7529 | ||
7530 | mutex_lock(&pmus_lock); | |
7531 | ||
7532 | ret = bus_register(&pmu_bus); | |
7533 | if (ret) | |
7534 | goto unlock; | |
7535 | ||
7536 | list_for_each_entry(pmu, &pmus, entry) { | |
7537 | if (!pmu->name || pmu->type < 0) | |
7538 | continue; | |
7539 | ||
7540 | ret = pmu_dev_alloc(pmu); | |
7541 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
7542 | } | |
7543 | pmu_bus_running = 1; | |
7544 | ret = 0; | |
7545 | ||
7546 | unlock: | |
7547 | mutex_unlock(&pmus_lock); | |
7548 | ||
7549 | return ret; | |
7550 | } | |
7551 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
7552 | |
7553 | #ifdef CONFIG_CGROUP_PERF | |
92fb9748 | 7554 | static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont) |
e5d1367f SE |
7555 | { |
7556 | struct perf_cgroup *jc; | |
e5d1367f | 7557 | |
1b15d055 | 7558 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
7559 | if (!jc) |
7560 | return ERR_PTR(-ENOMEM); | |
7561 | ||
e5d1367f SE |
7562 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7563 | if (!jc->info) { | |
7564 | kfree(jc); | |
7565 | return ERR_PTR(-ENOMEM); | |
7566 | } | |
7567 | ||
e5d1367f SE |
7568 | return &jc->css; |
7569 | } | |
7570 | ||
92fb9748 | 7571 | static void perf_cgroup_css_free(struct cgroup *cont) |
e5d1367f SE |
7572 | { |
7573 | struct perf_cgroup *jc; | |
7574 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), | |
7575 | struct perf_cgroup, css); | |
7576 | free_percpu(jc->info); | |
7577 | kfree(jc); | |
7578 | } | |
7579 | ||
7580 | static int __perf_cgroup_move(void *info) | |
7581 | { | |
7582 | struct task_struct *task = info; | |
7583 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
7584 | return 0; | |
7585 | } | |
7586 | ||
761b3ef5 | 7587 | static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
e5d1367f | 7588 | { |
bb9d97b6 TH |
7589 | struct task_struct *task; |
7590 | ||
7591 | cgroup_taskset_for_each(task, cgrp, tset) | |
7592 | task_function_call(task, __perf_cgroup_move, task); | |
e5d1367f SE |
7593 | } |
7594 | ||
761b3ef5 LZ |
7595 | static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, |
7596 | struct task_struct *task) | |
e5d1367f SE |
7597 | { |
7598 | /* | |
7599 | * cgroup_exit() is called in the copy_process() failure path. | |
7600 | * Ignore this case since the task hasn't ran yet, this avoids | |
7601 | * trying to poke a half freed task state from generic code. | |
7602 | */ | |
7603 | if (!(task->flags & PF_EXITING)) | |
7604 | return; | |
7605 | ||
bb9d97b6 | 7606 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7607 | } |
7608 | ||
7609 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
7610 | .name = "perf_event", |
7611 | .subsys_id = perf_subsys_id, | |
92fb9748 TH |
7612 | .css_alloc = perf_cgroup_css_alloc, |
7613 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 7614 | .exit = perf_cgroup_exit, |
bb9d97b6 | 7615 | .attach = perf_cgroup_attach, |
e5d1367f SE |
7616 | }; |
7617 | #endif /* CONFIG_CGROUP_PERF */ |