Commit | Line | Data |
---|---|---|
0793a61d | 1 | /* |
57c0c15b | 2 | * Performance events core code: |
0793a61d | 3 | * |
98144511 | 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
e7e7ee2e IM |
5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
7b732a75 | 8 | * |
57c0c15b | 9 | * For licensing details see kernel-base/COPYING |
0793a61d TG |
10 | */ |
11 | ||
12 | #include <linux/fs.h> | |
b9cacc7b | 13 | #include <linux/mm.h> |
0793a61d TG |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> | |
2e80a82a | 16 | #include <linux/idr.h> |
04289bb9 | 17 | #include <linux/file.h> |
0793a61d | 18 | #include <linux/poll.h> |
5a0e3ad6 | 19 | #include <linux/slab.h> |
76e1d904 | 20 | #include <linux/hash.h> |
0793a61d | 21 | #include <linux/sysfs.h> |
22a4f650 | 22 | #include <linux/dcache.h> |
0793a61d | 23 | #include <linux/percpu.h> |
22a4f650 | 24 | #include <linux/ptrace.h> |
c277443c | 25 | #include <linux/reboot.h> |
b9cacc7b | 26 | #include <linux/vmstat.h> |
abe43400 | 27 | #include <linux/device.h> |
6e5fdeed | 28 | #include <linux/export.h> |
906010b2 | 29 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
30 | #include <linux/hardirq.h> |
31 | #include <linux/rculist.h> | |
0793a61d TG |
32 | #include <linux/uaccess.h> |
33 | #include <linux/syscalls.h> | |
34 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 35 | #include <linux/kernel_stat.h> |
cdd6c482 | 36 | #include <linux/perf_event.h> |
6fb2915d | 37 | #include <linux/ftrace_event.h> |
3c502e7a | 38 | #include <linux/hw_breakpoint.h> |
c5ebcedb | 39 | #include <linux/mm_types.h> |
0793a61d | 40 | |
76369139 FW |
41 | #include "internal.h" |
42 | ||
4e193bd4 TB |
43 | #include <asm/irq_regs.h> |
44 | ||
fe4b04fa | 45 | struct remote_function_call { |
e7e7ee2e IM |
46 | struct task_struct *p; |
47 | int (*func)(void *info); | |
48 | void *info; | |
49 | int ret; | |
fe4b04fa PZ |
50 | }; |
51 | ||
52 | static void remote_function(void *data) | |
53 | { | |
54 | struct remote_function_call *tfc = data; | |
55 | struct task_struct *p = tfc->p; | |
56 | ||
57 | if (p) { | |
58 | tfc->ret = -EAGAIN; | |
59 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
60 | return; | |
61 | } | |
62 | ||
63 | tfc->ret = tfc->func(tfc->info); | |
64 | } | |
65 | ||
66 | /** | |
67 | * task_function_call - call a function on the cpu on which a task runs | |
68 | * @p: the task to evaluate | |
69 | * @func: the function to be called | |
70 | * @info: the function call argument | |
71 | * | |
72 | * Calls the function @func when the task is currently running. This might | |
73 | * be on the current CPU, which just calls the function directly | |
74 | * | |
75 | * returns: @func return value, or | |
76 | * -ESRCH - when the process isn't running | |
77 | * -EAGAIN - when the process moved away | |
78 | */ | |
79 | static int | |
80 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
81 | { | |
82 | struct remote_function_call data = { | |
e7e7ee2e IM |
83 | .p = p, |
84 | .func = func, | |
85 | .info = info, | |
86 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
87 | }; |
88 | ||
89 | if (task_curr(p)) | |
90 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
91 | ||
92 | return data.ret; | |
93 | } | |
94 | ||
95 | /** | |
96 | * cpu_function_call - call a function on the cpu | |
97 | * @func: the function to be called | |
98 | * @info: the function call argument | |
99 | * | |
100 | * Calls the function @func on the remote cpu. | |
101 | * | |
102 | * returns: @func return value or -ENXIO when the cpu is offline | |
103 | */ | |
104 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
105 | { | |
106 | struct remote_function_call data = { | |
e7e7ee2e IM |
107 | .p = NULL, |
108 | .func = func, | |
109 | .info = info, | |
110 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
111 | }; |
112 | ||
113 | smp_call_function_single(cpu, remote_function, &data, 1); | |
114 | ||
115 | return data.ret; | |
116 | } | |
117 | ||
e5d1367f SE |
118 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
119 | PERF_FLAG_FD_OUTPUT |\ | |
120 | PERF_FLAG_PID_CGROUP) | |
121 | ||
bce38cd5 SE |
122 | /* |
123 | * branch priv levels that need permission checks | |
124 | */ | |
125 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
126 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
127 | PERF_SAMPLE_BRANCH_HV) | |
128 | ||
0b3fcf17 SE |
129 | enum event_type_t { |
130 | EVENT_FLEXIBLE = 0x1, | |
131 | EVENT_PINNED = 0x2, | |
132 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
133 | }; | |
134 | ||
e5d1367f SE |
135 | /* |
136 | * perf_sched_events : >0 events exist | |
137 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
138 | */ | |
c5905afb | 139 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 140 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 141 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 142 | |
cdd6c482 IM |
143 | static atomic_t nr_mmap_events __read_mostly; |
144 | static atomic_t nr_comm_events __read_mostly; | |
145 | static atomic_t nr_task_events __read_mostly; | |
9ee318a7 | 146 | |
108b02cf PZ |
147 | static LIST_HEAD(pmus); |
148 | static DEFINE_MUTEX(pmus_lock); | |
149 | static struct srcu_struct pmus_srcu; | |
150 | ||
0764771d | 151 | /* |
cdd6c482 | 152 | * perf event paranoia level: |
0fbdea19 IM |
153 | * -1 - not paranoid at all |
154 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 155 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 156 | * 2 - disallow kernel profiling for unpriv |
0764771d | 157 | */ |
cdd6c482 | 158 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 159 | |
20443384 FW |
160 | /* Minimum for 512 kiB + 1 user control page */ |
161 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
162 | |
163 | /* | |
cdd6c482 | 164 | * max perf event sample rate |
df58ab24 | 165 | */ |
163ec435 PZ |
166 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
167 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
168 | static int max_samples_per_tick __read_mostly = | |
169 | DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
170 | ||
171 | int perf_proc_update_handler(struct ctl_table *table, int write, | |
172 | void __user *buffer, size_t *lenp, | |
173 | loff_t *ppos) | |
174 | { | |
175 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
176 | ||
177 | if (ret || !write) | |
178 | return ret; | |
179 | ||
180 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
181 | ||
182 | return 0; | |
183 | } | |
1ccd1549 | 184 | |
cdd6c482 | 185 | static atomic64_t perf_event_id; |
a96bbc16 | 186 | |
0b3fcf17 SE |
187 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
188 | enum event_type_t event_type); | |
189 | ||
190 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
191 | enum event_type_t event_type, |
192 | struct task_struct *task); | |
193 | ||
194 | static void update_context_time(struct perf_event_context *ctx); | |
195 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 196 | |
10c6db11 PZ |
197 | static void ring_buffer_attach(struct perf_event *event, |
198 | struct ring_buffer *rb); | |
199 | ||
cdd6c482 | 200 | void __weak perf_event_print_debug(void) { } |
0793a61d | 201 | |
84c79910 | 202 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 203 | { |
84c79910 | 204 | return "pmu"; |
0793a61d TG |
205 | } |
206 | ||
0b3fcf17 SE |
207 | static inline u64 perf_clock(void) |
208 | { | |
209 | return local_clock(); | |
210 | } | |
211 | ||
e5d1367f SE |
212 | static inline struct perf_cpu_context * |
213 | __get_cpu_context(struct perf_event_context *ctx) | |
214 | { | |
215 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
216 | } | |
217 | ||
facc4307 PZ |
218 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
219 | struct perf_event_context *ctx) | |
220 | { | |
221 | raw_spin_lock(&cpuctx->ctx.lock); | |
222 | if (ctx) | |
223 | raw_spin_lock(&ctx->lock); | |
224 | } | |
225 | ||
226 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
227 | struct perf_event_context *ctx) | |
228 | { | |
229 | if (ctx) | |
230 | raw_spin_unlock(&ctx->lock); | |
231 | raw_spin_unlock(&cpuctx->ctx.lock); | |
232 | } | |
233 | ||
e5d1367f SE |
234 | #ifdef CONFIG_CGROUP_PERF |
235 | ||
3f7cce3c SE |
236 | /* |
237 | * Must ensure cgroup is pinned (css_get) before calling | |
238 | * this function. In other words, we cannot call this function | |
239 | * if there is no cgroup event for the current CPU context. | |
240 | */ | |
e5d1367f SE |
241 | static inline struct perf_cgroup * |
242 | perf_cgroup_from_task(struct task_struct *task) | |
243 | { | |
244 | return container_of(task_subsys_state(task, perf_subsys_id), | |
245 | struct perf_cgroup, css); | |
246 | } | |
247 | ||
248 | static inline bool | |
249 | perf_cgroup_match(struct perf_event *event) | |
250 | { | |
251 | struct perf_event_context *ctx = event->ctx; | |
252 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
253 | ||
254 | return !event->cgrp || event->cgrp == cpuctx->cgrp; | |
255 | } | |
256 | ||
9c5da09d | 257 | static inline bool perf_tryget_cgroup(struct perf_event *event) |
e5d1367f | 258 | { |
9c5da09d | 259 | return css_tryget(&event->cgrp->css); |
e5d1367f SE |
260 | } |
261 | ||
262 | static inline void perf_put_cgroup(struct perf_event *event) | |
263 | { | |
264 | css_put(&event->cgrp->css); | |
265 | } | |
266 | ||
267 | static inline void perf_detach_cgroup(struct perf_event *event) | |
268 | { | |
269 | perf_put_cgroup(event); | |
270 | event->cgrp = NULL; | |
271 | } | |
272 | ||
273 | static inline int is_cgroup_event(struct perf_event *event) | |
274 | { | |
275 | return event->cgrp != NULL; | |
276 | } | |
277 | ||
278 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
279 | { | |
280 | struct perf_cgroup_info *t; | |
281 | ||
282 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
283 | return t->time; | |
284 | } | |
285 | ||
286 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
287 | { | |
288 | struct perf_cgroup_info *info; | |
289 | u64 now; | |
290 | ||
291 | now = perf_clock(); | |
292 | ||
293 | info = this_cpu_ptr(cgrp->info); | |
294 | ||
295 | info->time += now - info->timestamp; | |
296 | info->timestamp = now; | |
297 | } | |
298 | ||
299 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
300 | { | |
301 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
302 | if (cgrp_out) | |
303 | __update_cgrp_time(cgrp_out); | |
304 | } | |
305 | ||
306 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
307 | { | |
3f7cce3c SE |
308 | struct perf_cgroup *cgrp; |
309 | ||
e5d1367f | 310 | /* |
3f7cce3c SE |
311 | * ensure we access cgroup data only when needed and |
312 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 313 | */ |
3f7cce3c | 314 | if (!is_cgroup_event(event)) |
e5d1367f SE |
315 | return; |
316 | ||
3f7cce3c SE |
317 | cgrp = perf_cgroup_from_task(current); |
318 | /* | |
319 | * Do not update time when cgroup is not active | |
320 | */ | |
321 | if (cgrp == event->cgrp) | |
322 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
323 | } |
324 | ||
325 | static inline void | |
3f7cce3c SE |
326 | perf_cgroup_set_timestamp(struct task_struct *task, |
327 | struct perf_event_context *ctx) | |
e5d1367f SE |
328 | { |
329 | struct perf_cgroup *cgrp; | |
330 | struct perf_cgroup_info *info; | |
331 | ||
3f7cce3c SE |
332 | /* |
333 | * ctx->lock held by caller | |
334 | * ensure we do not access cgroup data | |
335 | * unless we have the cgroup pinned (css_get) | |
336 | */ | |
337 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
338 | return; |
339 | ||
340 | cgrp = perf_cgroup_from_task(task); | |
341 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 342 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
343 | } |
344 | ||
345 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
346 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
347 | ||
348 | /* | |
349 | * reschedule events based on the cgroup constraint of task. | |
350 | * | |
351 | * mode SWOUT : schedule out everything | |
352 | * mode SWIN : schedule in based on cgroup for next | |
353 | */ | |
354 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
355 | { | |
356 | struct perf_cpu_context *cpuctx; | |
357 | struct pmu *pmu; | |
358 | unsigned long flags; | |
359 | ||
360 | /* | |
361 | * disable interrupts to avoid geting nr_cgroup | |
362 | * changes via __perf_event_disable(). Also | |
363 | * avoids preemption. | |
364 | */ | |
365 | local_irq_save(flags); | |
366 | ||
367 | /* | |
368 | * we reschedule only in the presence of cgroup | |
369 | * constrained events. | |
370 | */ | |
371 | rcu_read_lock(); | |
372 | ||
373 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 374 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
375 | if (cpuctx->unique_pmu != pmu) |
376 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 377 | |
e5d1367f SE |
378 | /* |
379 | * perf_cgroup_events says at least one | |
380 | * context on this CPU has cgroup events. | |
381 | * | |
382 | * ctx->nr_cgroups reports the number of cgroup | |
383 | * events for a context. | |
384 | */ | |
385 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
386 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
387 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
388 | |
389 | if (mode & PERF_CGROUP_SWOUT) { | |
390 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
391 | /* | |
392 | * must not be done before ctxswout due | |
393 | * to event_filter_match() in event_sched_out() | |
394 | */ | |
395 | cpuctx->cgrp = NULL; | |
396 | } | |
397 | ||
398 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 399 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
400 | /* |
401 | * set cgrp before ctxsw in to allow | |
402 | * event_filter_match() to not have to pass | |
403 | * task around | |
e5d1367f SE |
404 | */ |
405 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
406 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
407 | } | |
facc4307 PZ |
408 | perf_pmu_enable(cpuctx->ctx.pmu); |
409 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 410 | } |
e5d1367f SE |
411 | } |
412 | ||
413 | rcu_read_unlock(); | |
414 | ||
415 | local_irq_restore(flags); | |
416 | } | |
417 | ||
a8d757ef SE |
418 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
419 | struct task_struct *next) | |
e5d1367f | 420 | { |
a8d757ef SE |
421 | struct perf_cgroup *cgrp1; |
422 | struct perf_cgroup *cgrp2 = NULL; | |
423 | ||
424 | /* | |
425 | * we come here when we know perf_cgroup_events > 0 | |
426 | */ | |
427 | cgrp1 = perf_cgroup_from_task(task); | |
428 | ||
429 | /* | |
430 | * next is NULL when called from perf_event_enable_on_exec() | |
431 | * that will systematically cause a cgroup_switch() | |
432 | */ | |
433 | if (next) | |
434 | cgrp2 = perf_cgroup_from_task(next); | |
435 | ||
436 | /* | |
437 | * only schedule out current cgroup events if we know | |
438 | * that we are switching to a different cgroup. Otherwise, | |
439 | * do no touch the cgroup events. | |
440 | */ | |
441 | if (cgrp1 != cgrp2) | |
442 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
443 | } |
444 | ||
a8d757ef SE |
445 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
446 | struct task_struct *task) | |
e5d1367f | 447 | { |
a8d757ef SE |
448 | struct perf_cgroup *cgrp1; |
449 | struct perf_cgroup *cgrp2 = NULL; | |
450 | ||
451 | /* | |
452 | * we come here when we know perf_cgroup_events > 0 | |
453 | */ | |
454 | cgrp1 = perf_cgroup_from_task(task); | |
455 | ||
456 | /* prev can never be NULL */ | |
457 | cgrp2 = perf_cgroup_from_task(prev); | |
458 | ||
459 | /* | |
460 | * only need to schedule in cgroup events if we are changing | |
461 | * cgroup during ctxsw. Cgroup events were not scheduled | |
462 | * out of ctxsw out if that was not the case. | |
463 | */ | |
464 | if (cgrp1 != cgrp2) | |
465 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
466 | } |
467 | ||
468 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
469 | struct perf_event_attr *attr, | |
470 | struct perf_event *group_leader) | |
471 | { | |
472 | struct perf_cgroup *cgrp; | |
473 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
474 | struct fd f = fdget(fd); |
475 | int ret = 0; | |
e5d1367f | 476 | |
2903ff01 | 477 | if (!f.file) |
e5d1367f SE |
478 | return -EBADF; |
479 | ||
2903ff01 | 480 | css = cgroup_css_from_dir(f.file, perf_subsys_id); |
3db272c0 LZ |
481 | if (IS_ERR(css)) { |
482 | ret = PTR_ERR(css); | |
483 | goto out; | |
484 | } | |
e5d1367f SE |
485 | |
486 | cgrp = container_of(css, struct perf_cgroup, css); | |
487 | event->cgrp = cgrp; | |
488 | ||
f75e18cb | 489 | /* must be done before we fput() the file */ |
9c5da09d SQ |
490 | if (!perf_tryget_cgroup(event)) { |
491 | event->cgrp = NULL; | |
492 | ret = -ENOENT; | |
493 | goto out; | |
494 | } | |
f75e18cb | 495 | |
e5d1367f SE |
496 | /* |
497 | * all events in a group must monitor | |
498 | * the same cgroup because a task belongs | |
499 | * to only one perf cgroup at a time | |
500 | */ | |
501 | if (group_leader && group_leader->cgrp != cgrp) { | |
502 | perf_detach_cgroup(event); | |
503 | ret = -EINVAL; | |
e5d1367f | 504 | } |
3db272c0 | 505 | out: |
2903ff01 | 506 | fdput(f); |
e5d1367f SE |
507 | return ret; |
508 | } | |
509 | ||
510 | static inline void | |
511 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
512 | { | |
513 | struct perf_cgroup_info *t; | |
514 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
515 | event->shadow_ctx_time = now - t->timestamp; | |
516 | } | |
517 | ||
518 | static inline void | |
519 | perf_cgroup_defer_enabled(struct perf_event *event) | |
520 | { | |
521 | /* | |
522 | * when the current task's perf cgroup does not match | |
523 | * the event's, we need to remember to call the | |
524 | * perf_mark_enable() function the first time a task with | |
525 | * a matching perf cgroup is scheduled in. | |
526 | */ | |
527 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
528 | event->cgrp_defer_enabled = 1; | |
529 | } | |
530 | ||
531 | static inline void | |
532 | perf_cgroup_mark_enabled(struct perf_event *event, | |
533 | struct perf_event_context *ctx) | |
534 | { | |
535 | struct perf_event *sub; | |
536 | u64 tstamp = perf_event_time(event); | |
537 | ||
538 | if (!event->cgrp_defer_enabled) | |
539 | return; | |
540 | ||
541 | event->cgrp_defer_enabled = 0; | |
542 | ||
543 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
544 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
545 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
546 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
547 | sub->cgrp_defer_enabled = 0; | |
548 | } | |
549 | } | |
550 | } | |
551 | #else /* !CONFIG_CGROUP_PERF */ | |
552 | ||
553 | static inline bool | |
554 | perf_cgroup_match(struct perf_event *event) | |
555 | { | |
556 | return true; | |
557 | } | |
558 | ||
559 | static inline void perf_detach_cgroup(struct perf_event *event) | |
560 | {} | |
561 | ||
562 | static inline int is_cgroup_event(struct perf_event *event) | |
563 | { | |
564 | return 0; | |
565 | } | |
566 | ||
567 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
568 | { | |
569 | return 0; | |
570 | } | |
571 | ||
572 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
573 | { | |
574 | } | |
575 | ||
576 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
577 | { | |
578 | } | |
579 | ||
a8d757ef SE |
580 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
581 | struct task_struct *next) | |
e5d1367f SE |
582 | { |
583 | } | |
584 | ||
a8d757ef SE |
585 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
586 | struct task_struct *task) | |
e5d1367f SE |
587 | { |
588 | } | |
589 | ||
590 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
591 | struct perf_event_attr *attr, | |
592 | struct perf_event *group_leader) | |
593 | { | |
594 | return -EINVAL; | |
595 | } | |
596 | ||
597 | static inline void | |
3f7cce3c SE |
598 | perf_cgroup_set_timestamp(struct task_struct *task, |
599 | struct perf_event_context *ctx) | |
e5d1367f SE |
600 | { |
601 | } | |
602 | ||
603 | void | |
604 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
605 | { | |
606 | } | |
607 | ||
608 | static inline void | |
609 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
610 | { | |
611 | } | |
612 | ||
613 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
614 | { | |
615 | return 0; | |
616 | } | |
617 | ||
618 | static inline void | |
619 | perf_cgroup_defer_enabled(struct perf_event *event) | |
620 | { | |
621 | } | |
622 | ||
623 | static inline void | |
624 | perf_cgroup_mark_enabled(struct perf_event *event, | |
625 | struct perf_event_context *ctx) | |
626 | { | |
627 | } | |
628 | #endif | |
629 | ||
33696fc0 | 630 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 631 | { |
33696fc0 PZ |
632 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
633 | if (!(*count)++) | |
634 | pmu->pmu_disable(pmu); | |
9e35ad38 | 635 | } |
9e35ad38 | 636 | |
33696fc0 | 637 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 638 | { |
33696fc0 PZ |
639 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
640 | if (!--(*count)) | |
641 | pmu->pmu_enable(pmu); | |
9e35ad38 | 642 | } |
9e35ad38 | 643 | |
e9d2b064 PZ |
644 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
645 | ||
646 | /* | |
647 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
648 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
649 | * disabled, while rotate_context is called from IRQ context. | |
650 | */ | |
108b02cf | 651 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 652 | { |
108b02cf | 653 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 654 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 655 | |
e9d2b064 | 656 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 657 | |
e9d2b064 PZ |
658 | if (list_empty(&cpuctx->rotation_list)) |
659 | list_add(&cpuctx->rotation_list, head); | |
9e35ad38 | 660 | } |
9e35ad38 | 661 | |
cdd6c482 | 662 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 663 | { |
e5289d4a | 664 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
665 | } |
666 | ||
cdd6c482 | 667 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 668 | { |
564c2b21 PM |
669 | if (atomic_dec_and_test(&ctx->refcount)) { |
670 | if (ctx->parent_ctx) | |
671 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
672 | if (ctx->task) |
673 | put_task_struct(ctx->task); | |
cb796ff3 | 674 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 675 | } |
a63eaf34 PM |
676 | } |
677 | ||
cdd6c482 | 678 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
679 | { |
680 | if (ctx->parent_ctx) { | |
681 | put_ctx(ctx->parent_ctx); | |
682 | ctx->parent_ctx = NULL; | |
683 | } | |
684 | } | |
685 | ||
6844c09d ACM |
686 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
687 | { | |
688 | /* | |
689 | * only top level events have the pid namespace they were created in | |
690 | */ | |
691 | if (event->parent) | |
692 | event = event->parent; | |
693 | ||
694 | return task_tgid_nr_ns(p, event->ns); | |
695 | } | |
696 | ||
697 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
698 | { | |
699 | /* | |
700 | * only top level events have the pid namespace they were created in | |
701 | */ | |
702 | if (event->parent) | |
703 | event = event->parent; | |
704 | ||
705 | return task_pid_nr_ns(p, event->ns); | |
706 | } | |
707 | ||
7f453c24 | 708 | /* |
cdd6c482 | 709 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
710 | * to userspace. |
711 | */ | |
cdd6c482 | 712 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 713 | { |
cdd6c482 | 714 | u64 id = event->id; |
7f453c24 | 715 | |
cdd6c482 IM |
716 | if (event->parent) |
717 | id = event->parent->id; | |
7f453c24 PZ |
718 | |
719 | return id; | |
720 | } | |
721 | ||
25346b93 | 722 | /* |
cdd6c482 | 723 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
724 | * This has to cope with with the fact that until it is locked, |
725 | * the context could get moved to another task. | |
726 | */ | |
cdd6c482 | 727 | static struct perf_event_context * |
8dc85d54 | 728 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 729 | { |
cdd6c482 | 730 | struct perf_event_context *ctx; |
25346b93 PM |
731 | |
732 | rcu_read_lock(); | |
9ed6060d | 733 | retry: |
8dc85d54 | 734 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
735 | if (ctx) { |
736 | /* | |
737 | * If this context is a clone of another, it might | |
738 | * get swapped for another underneath us by | |
cdd6c482 | 739 | * perf_event_task_sched_out, though the |
25346b93 PM |
740 | * rcu_read_lock() protects us from any context |
741 | * getting freed. Lock the context and check if it | |
742 | * got swapped before we could get the lock, and retry | |
743 | * if so. If we locked the right context, then it | |
744 | * can't get swapped on us any more. | |
745 | */ | |
e625cce1 | 746 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 747 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 748 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
25346b93 PM |
749 | goto retry; |
750 | } | |
b49a9e7e PZ |
751 | |
752 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 753 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
754 | ctx = NULL; |
755 | } | |
25346b93 PM |
756 | } |
757 | rcu_read_unlock(); | |
758 | return ctx; | |
759 | } | |
760 | ||
761 | /* | |
762 | * Get the context for a task and increment its pin_count so it | |
763 | * can't get swapped to another task. This also increments its | |
764 | * reference count so that the context can't get freed. | |
765 | */ | |
8dc85d54 PZ |
766 | static struct perf_event_context * |
767 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 768 | { |
cdd6c482 | 769 | struct perf_event_context *ctx; |
25346b93 PM |
770 | unsigned long flags; |
771 | ||
8dc85d54 | 772 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
773 | if (ctx) { |
774 | ++ctx->pin_count; | |
e625cce1 | 775 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
776 | } |
777 | return ctx; | |
778 | } | |
779 | ||
cdd6c482 | 780 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
781 | { |
782 | unsigned long flags; | |
783 | ||
e625cce1 | 784 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 785 | --ctx->pin_count; |
e625cce1 | 786 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
787 | } |
788 | ||
f67218c3 PZ |
789 | /* |
790 | * Update the record of the current time in a context. | |
791 | */ | |
792 | static void update_context_time(struct perf_event_context *ctx) | |
793 | { | |
794 | u64 now = perf_clock(); | |
795 | ||
796 | ctx->time += now - ctx->timestamp; | |
797 | ctx->timestamp = now; | |
798 | } | |
799 | ||
4158755d SE |
800 | static u64 perf_event_time(struct perf_event *event) |
801 | { | |
802 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
803 | |
804 | if (is_cgroup_event(event)) | |
805 | return perf_cgroup_event_time(event); | |
806 | ||
4158755d SE |
807 | return ctx ? ctx->time : 0; |
808 | } | |
809 | ||
f67218c3 PZ |
810 | /* |
811 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 812 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
813 | */ |
814 | static void update_event_times(struct perf_event *event) | |
815 | { | |
816 | struct perf_event_context *ctx = event->ctx; | |
817 | u64 run_end; | |
818 | ||
819 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
820 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
821 | return; | |
e5d1367f SE |
822 | /* |
823 | * in cgroup mode, time_enabled represents | |
824 | * the time the event was enabled AND active | |
825 | * tasks were in the monitored cgroup. This is | |
826 | * independent of the activity of the context as | |
827 | * there may be a mix of cgroup and non-cgroup events. | |
828 | * | |
829 | * That is why we treat cgroup events differently | |
830 | * here. | |
831 | */ | |
832 | if (is_cgroup_event(event)) | |
46cd6a7f | 833 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
834 | else if (ctx->is_active) |
835 | run_end = ctx->time; | |
acd1d7c1 PZ |
836 | else |
837 | run_end = event->tstamp_stopped; | |
838 | ||
839 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
840 | |
841 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
842 | run_end = event->tstamp_stopped; | |
843 | else | |
4158755d | 844 | run_end = perf_event_time(event); |
f67218c3 PZ |
845 | |
846 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 847 | |
f67218c3 PZ |
848 | } |
849 | ||
96c21a46 PZ |
850 | /* |
851 | * Update total_time_enabled and total_time_running for all events in a group. | |
852 | */ | |
853 | static void update_group_times(struct perf_event *leader) | |
854 | { | |
855 | struct perf_event *event; | |
856 | ||
857 | update_event_times(leader); | |
858 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
859 | update_event_times(event); | |
860 | } | |
861 | ||
889ff015 FW |
862 | static struct list_head * |
863 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
864 | { | |
865 | if (event->attr.pinned) | |
866 | return &ctx->pinned_groups; | |
867 | else | |
868 | return &ctx->flexible_groups; | |
869 | } | |
870 | ||
fccc714b | 871 | /* |
cdd6c482 | 872 | * Add a event from the lists for its context. |
fccc714b PZ |
873 | * Must be called with ctx->mutex and ctx->lock held. |
874 | */ | |
04289bb9 | 875 | static void |
cdd6c482 | 876 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 877 | { |
8a49542c PZ |
878 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
879 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
880 | |
881 | /* | |
8a49542c PZ |
882 | * If we're a stand alone event or group leader, we go to the context |
883 | * list, group events are kept attached to the group so that | |
884 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 885 | */ |
8a49542c | 886 | if (event->group_leader == event) { |
889ff015 FW |
887 | struct list_head *list; |
888 | ||
d6f962b5 FW |
889 | if (is_software_event(event)) |
890 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
891 | ||
889ff015 FW |
892 | list = ctx_group_list(event, ctx); |
893 | list_add_tail(&event->group_entry, list); | |
5c148194 | 894 | } |
592903cd | 895 | |
08309379 | 896 | if (is_cgroup_event(event)) |
e5d1367f | 897 | ctx->nr_cgroups++; |
e5d1367f | 898 | |
d010b332 SE |
899 | if (has_branch_stack(event)) |
900 | ctx->nr_branch_stack++; | |
901 | ||
cdd6c482 | 902 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 903 | if (!ctx->nr_events) |
108b02cf | 904 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
905 | ctx->nr_events++; |
906 | if (event->attr.inherit_stat) | |
bfbd3381 | 907 | ctx->nr_stat++; |
04289bb9 IM |
908 | } |
909 | ||
0231bb53 JO |
910 | /* |
911 | * Initialize event state based on the perf_event_attr::disabled. | |
912 | */ | |
913 | static inline void perf_event__state_init(struct perf_event *event) | |
914 | { | |
915 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
916 | PERF_EVENT_STATE_INACTIVE; | |
917 | } | |
918 | ||
c320c7b7 ACM |
919 | /* |
920 | * Called at perf_event creation and when events are attached/detached from a | |
921 | * group. | |
922 | */ | |
923 | static void perf_event__read_size(struct perf_event *event) | |
924 | { | |
925 | int entry = sizeof(u64); /* value */ | |
926 | int size = 0; | |
927 | int nr = 1; | |
928 | ||
929 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
930 | size += sizeof(u64); | |
931 | ||
932 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
933 | size += sizeof(u64); | |
934 | ||
935 | if (event->attr.read_format & PERF_FORMAT_ID) | |
936 | entry += sizeof(u64); | |
937 | ||
938 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
939 | nr += event->group_leader->nr_siblings; | |
940 | size += sizeof(u64); | |
941 | } | |
942 | ||
943 | size += entry * nr; | |
944 | event->read_size = size; | |
945 | } | |
946 | ||
947 | static void perf_event__header_size(struct perf_event *event) | |
948 | { | |
949 | struct perf_sample_data *data; | |
950 | u64 sample_type = event->attr.sample_type; | |
951 | u16 size = 0; | |
952 | ||
953 | perf_event__read_size(event); | |
954 | ||
955 | if (sample_type & PERF_SAMPLE_IP) | |
956 | size += sizeof(data->ip); | |
957 | ||
6844c09d ACM |
958 | if (sample_type & PERF_SAMPLE_ADDR) |
959 | size += sizeof(data->addr); | |
960 | ||
961 | if (sample_type & PERF_SAMPLE_PERIOD) | |
962 | size += sizeof(data->period); | |
963 | ||
964 | if (sample_type & PERF_SAMPLE_READ) | |
965 | size += event->read_size; | |
966 | ||
967 | event->header_size = size; | |
968 | } | |
969 | ||
970 | static void perf_event__id_header_size(struct perf_event *event) | |
971 | { | |
972 | struct perf_sample_data *data; | |
973 | u64 sample_type = event->attr.sample_type; | |
974 | u16 size = 0; | |
975 | ||
c320c7b7 ACM |
976 | if (sample_type & PERF_SAMPLE_TID) |
977 | size += sizeof(data->tid_entry); | |
978 | ||
979 | if (sample_type & PERF_SAMPLE_TIME) | |
980 | size += sizeof(data->time); | |
981 | ||
c320c7b7 ACM |
982 | if (sample_type & PERF_SAMPLE_ID) |
983 | size += sizeof(data->id); | |
984 | ||
985 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
986 | size += sizeof(data->stream_id); | |
987 | ||
988 | if (sample_type & PERF_SAMPLE_CPU) | |
989 | size += sizeof(data->cpu_entry); | |
990 | ||
6844c09d | 991 | event->id_header_size = size; |
c320c7b7 ACM |
992 | } |
993 | ||
8a49542c PZ |
994 | static void perf_group_attach(struct perf_event *event) |
995 | { | |
c320c7b7 | 996 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 997 | |
74c3337c PZ |
998 | /* |
999 | * We can have double attach due to group movement in perf_event_open. | |
1000 | */ | |
1001 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1002 | return; | |
1003 | ||
8a49542c PZ |
1004 | event->attach_state |= PERF_ATTACH_GROUP; |
1005 | ||
1006 | if (group_leader == event) | |
1007 | return; | |
1008 | ||
1009 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1010 | !is_software_event(event)) | |
1011 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1012 | ||
1013 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1014 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1015 | |
1016 | perf_event__header_size(group_leader); | |
1017 | ||
1018 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1019 | perf_event__header_size(pos); | |
8a49542c PZ |
1020 | } |
1021 | ||
a63eaf34 | 1022 | /* |
cdd6c482 | 1023 | * Remove a event from the lists for its context. |
fccc714b | 1024 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1025 | */ |
04289bb9 | 1026 | static void |
cdd6c482 | 1027 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1028 | { |
68cacd29 | 1029 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1030 | /* |
1031 | * We can have double detach due to exit/hot-unplug + close. | |
1032 | */ | |
1033 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1034 | return; |
8a49542c PZ |
1035 | |
1036 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1037 | ||
68cacd29 | 1038 | if (is_cgroup_event(event)) { |
e5d1367f | 1039 | ctx->nr_cgroups--; |
68cacd29 SE |
1040 | cpuctx = __get_cpu_context(ctx); |
1041 | /* | |
1042 | * if there are no more cgroup events | |
1043 | * then cler cgrp to avoid stale pointer | |
1044 | * in update_cgrp_time_from_cpuctx() | |
1045 | */ | |
1046 | if (!ctx->nr_cgroups) | |
1047 | cpuctx->cgrp = NULL; | |
1048 | } | |
e5d1367f | 1049 | |
d010b332 SE |
1050 | if (has_branch_stack(event)) |
1051 | ctx->nr_branch_stack--; | |
1052 | ||
cdd6c482 IM |
1053 | ctx->nr_events--; |
1054 | if (event->attr.inherit_stat) | |
bfbd3381 | 1055 | ctx->nr_stat--; |
8bc20959 | 1056 | |
cdd6c482 | 1057 | list_del_rcu(&event->event_entry); |
04289bb9 | 1058 | |
8a49542c PZ |
1059 | if (event->group_leader == event) |
1060 | list_del_init(&event->group_entry); | |
5c148194 | 1061 | |
96c21a46 | 1062 | update_group_times(event); |
b2e74a26 SE |
1063 | |
1064 | /* | |
1065 | * If event was in error state, then keep it | |
1066 | * that way, otherwise bogus counts will be | |
1067 | * returned on read(). The only way to get out | |
1068 | * of error state is by explicit re-enabling | |
1069 | * of the event | |
1070 | */ | |
1071 | if (event->state > PERF_EVENT_STATE_OFF) | |
1072 | event->state = PERF_EVENT_STATE_OFF; | |
050735b0 PZ |
1073 | } |
1074 | ||
8a49542c | 1075 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1076 | { |
1077 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1078 | struct list_head *list = NULL; |
1079 | ||
1080 | /* | |
1081 | * We can have double detach due to exit/hot-unplug + close. | |
1082 | */ | |
1083 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1084 | return; | |
1085 | ||
1086 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1087 | ||
1088 | /* | |
1089 | * If this is a sibling, remove it from its group. | |
1090 | */ | |
1091 | if (event->group_leader != event) { | |
1092 | list_del_init(&event->group_entry); | |
1093 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1094 | goto out; |
8a49542c PZ |
1095 | } |
1096 | ||
1097 | if (!list_empty(&event->group_entry)) | |
1098 | list = &event->group_entry; | |
2e2af50b | 1099 | |
04289bb9 | 1100 | /* |
cdd6c482 IM |
1101 | * If this was a group event with sibling events then |
1102 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1103 | * to whatever list we are on. |
04289bb9 | 1104 | */ |
cdd6c482 | 1105 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1106 | if (list) |
1107 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1108 | sibling->group_leader = sibling; |
d6f962b5 FW |
1109 | |
1110 | /* Inherit group flags from the previous leader */ | |
1111 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1112 | } |
c320c7b7 ACM |
1113 | |
1114 | out: | |
1115 | perf_event__header_size(event->group_leader); | |
1116 | ||
1117 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1118 | perf_event__header_size(tmp); | |
04289bb9 IM |
1119 | } |
1120 | ||
fa66f07a SE |
1121 | static inline int |
1122 | event_filter_match(struct perf_event *event) | |
1123 | { | |
e5d1367f SE |
1124 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1125 | && perf_cgroup_match(event); | |
fa66f07a SE |
1126 | } |
1127 | ||
9ffcfa6f SE |
1128 | static void |
1129 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1130 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1131 | struct perf_event_context *ctx) |
3b6f9e5c | 1132 | { |
4158755d | 1133 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1134 | u64 delta; |
1135 | /* | |
1136 | * An event which could not be activated because of | |
1137 | * filter mismatch still needs to have its timings | |
1138 | * maintained, otherwise bogus information is return | |
1139 | * via read() for time_enabled, time_running: | |
1140 | */ | |
1141 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1142 | && !event_filter_match(event)) { | |
e5d1367f | 1143 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1144 | event->tstamp_running += delta; |
4158755d | 1145 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1146 | } |
1147 | ||
cdd6c482 | 1148 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1149 | return; |
3b6f9e5c | 1150 | |
cdd6c482 IM |
1151 | event->state = PERF_EVENT_STATE_INACTIVE; |
1152 | if (event->pending_disable) { | |
1153 | event->pending_disable = 0; | |
1154 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1155 | } |
4158755d | 1156 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1157 | event->pmu->del(event, 0); |
cdd6c482 | 1158 | event->oncpu = -1; |
3b6f9e5c | 1159 | |
cdd6c482 | 1160 | if (!is_software_event(event)) |
3b6f9e5c PM |
1161 | cpuctx->active_oncpu--; |
1162 | ctx->nr_active--; | |
0f5a2601 PZ |
1163 | if (event->attr.freq && event->attr.sample_freq) |
1164 | ctx->nr_freq--; | |
cdd6c482 | 1165 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c PM |
1166 | cpuctx->exclusive = 0; |
1167 | } | |
1168 | ||
d859e29f | 1169 | static void |
cdd6c482 | 1170 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1171 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1172 | struct perf_event_context *ctx) |
d859e29f | 1173 | { |
cdd6c482 | 1174 | struct perf_event *event; |
fa66f07a | 1175 | int state = group_event->state; |
d859e29f | 1176 | |
cdd6c482 | 1177 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1178 | |
1179 | /* | |
1180 | * Schedule out siblings (if any): | |
1181 | */ | |
cdd6c482 IM |
1182 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1183 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1184 | |
fa66f07a | 1185 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1186 | cpuctx->exclusive = 0; |
1187 | } | |
1188 | ||
0793a61d | 1189 | /* |
cdd6c482 | 1190 | * Cross CPU call to remove a performance event |
0793a61d | 1191 | * |
cdd6c482 | 1192 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1193 | * remove it from the context list. |
1194 | */ | |
fe4b04fa | 1195 | static int __perf_remove_from_context(void *info) |
0793a61d | 1196 | { |
cdd6c482 IM |
1197 | struct perf_event *event = info; |
1198 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1199 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1200 | |
e625cce1 | 1201 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1202 | event_sched_out(event, cpuctx, ctx); |
cdd6c482 | 1203 | list_del_event(event, ctx); |
64ce3126 PZ |
1204 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1205 | ctx->is_active = 0; | |
1206 | cpuctx->task_ctx = NULL; | |
1207 | } | |
e625cce1 | 1208 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1209 | |
1210 | return 0; | |
0793a61d TG |
1211 | } |
1212 | ||
1213 | ||
1214 | /* | |
cdd6c482 | 1215 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1216 | * |
cdd6c482 | 1217 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1218 | * call when the task is on a CPU. |
c93f7669 | 1219 | * |
cdd6c482 IM |
1220 | * If event->ctx is a cloned context, callers must make sure that |
1221 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1222 | * remains valid. This is OK when called from perf_release since |
1223 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1224 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1225 | * context has been detached from its task. |
0793a61d | 1226 | */ |
fe4b04fa | 1227 | static void perf_remove_from_context(struct perf_event *event) |
0793a61d | 1228 | { |
cdd6c482 | 1229 | struct perf_event_context *ctx = event->ctx; |
0793a61d TG |
1230 | struct task_struct *task = ctx->task; |
1231 | ||
fe4b04fa PZ |
1232 | lockdep_assert_held(&ctx->mutex); |
1233 | ||
0793a61d TG |
1234 | if (!task) { |
1235 | /* | |
cdd6c482 | 1236 | * Per cpu events are removed via an smp call and |
af901ca1 | 1237 | * the removal is always successful. |
0793a61d | 1238 | */ |
fe4b04fa | 1239 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
0793a61d TG |
1240 | return; |
1241 | } | |
1242 | ||
1243 | retry: | |
fe4b04fa PZ |
1244 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1245 | return; | |
0793a61d | 1246 | |
e625cce1 | 1247 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1248 | /* |
fe4b04fa PZ |
1249 | * If we failed to find a running task, but find the context active now |
1250 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1251 | */ |
fe4b04fa | 1252 | if (ctx->is_active) { |
e625cce1 | 1253 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1254 | goto retry; |
1255 | } | |
1256 | ||
1257 | /* | |
fe4b04fa PZ |
1258 | * Since the task isn't running, its safe to remove the event, us |
1259 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1260 | */ |
fe4b04fa | 1261 | list_del_event(event, ctx); |
e625cce1 | 1262 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1263 | } |
1264 | ||
d859e29f | 1265 | /* |
cdd6c482 | 1266 | * Cross CPU call to disable a performance event |
d859e29f | 1267 | */ |
500ad2d8 | 1268 | int __perf_event_disable(void *info) |
d859e29f | 1269 | { |
cdd6c482 | 1270 | struct perf_event *event = info; |
cdd6c482 | 1271 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1272 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1273 | |
1274 | /* | |
cdd6c482 IM |
1275 | * If this is a per-task event, need to check whether this |
1276 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1277 | * |
1278 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1279 | * flipping contexts around. | |
d859e29f | 1280 | */ |
665c2142 | 1281 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1282 | return -EINVAL; |
d859e29f | 1283 | |
e625cce1 | 1284 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1285 | |
1286 | /* | |
cdd6c482 | 1287 | * If the event is on, turn it off. |
d859e29f PM |
1288 | * If it is in error state, leave it in error state. |
1289 | */ | |
cdd6c482 | 1290 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1291 | update_context_time(ctx); |
e5d1367f | 1292 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1293 | update_group_times(event); |
1294 | if (event == event->group_leader) | |
1295 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1296 | else |
cdd6c482 IM |
1297 | event_sched_out(event, cpuctx, ctx); |
1298 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1299 | } |
1300 | ||
e625cce1 | 1301 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1302 | |
1303 | return 0; | |
d859e29f PM |
1304 | } |
1305 | ||
1306 | /* | |
cdd6c482 | 1307 | * Disable a event. |
c93f7669 | 1308 | * |
cdd6c482 IM |
1309 | * If event->ctx is a cloned context, callers must make sure that |
1310 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1311 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1312 | * perf_event_for_each_child or perf_event_for_each because they |
1313 | * hold the top-level event's child_mutex, so any descendant that | |
1314 | * goes to exit will block in sync_child_event. | |
1315 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1316 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1317 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1318 | */ |
44234adc | 1319 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1320 | { |
cdd6c482 | 1321 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1322 | struct task_struct *task = ctx->task; |
1323 | ||
1324 | if (!task) { | |
1325 | /* | |
cdd6c482 | 1326 | * Disable the event on the cpu that it's on |
d859e29f | 1327 | */ |
fe4b04fa | 1328 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1329 | return; |
1330 | } | |
1331 | ||
9ed6060d | 1332 | retry: |
fe4b04fa PZ |
1333 | if (!task_function_call(task, __perf_event_disable, event)) |
1334 | return; | |
d859e29f | 1335 | |
e625cce1 | 1336 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1337 | /* |
cdd6c482 | 1338 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1339 | */ |
cdd6c482 | 1340 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1341 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1342 | /* |
1343 | * Reload the task pointer, it might have been changed by | |
1344 | * a concurrent perf_event_context_sched_out(). | |
1345 | */ | |
1346 | task = ctx->task; | |
d859e29f PM |
1347 | goto retry; |
1348 | } | |
1349 | ||
1350 | /* | |
1351 | * Since we have the lock this context can't be scheduled | |
1352 | * in, so we can change the state safely. | |
1353 | */ | |
cdd6c482 IM |
1354 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1355 | update_group_times(event); | |
1356 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1357 | } |
e625cce1 | 1358 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1359 | } |
dcfce4a0 | 1360 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1361 | |
e5d1367f SE |
1362 | static void perf_set_shadow_time(struct perf_event *event, |
1363 | struct perf_event_context *ctx, | |
1364 | u64 tstamp) | |
1365 | { | |
1366 | /* | |
1367 | * use the correct time source for the time snapshot | |
1368 | * | |
1369 | * We could get by without this by leveraging the | |
1370 | * fact that to get to this function, the caller | |
1371 | * has most likely already called update_context_time() | |
1372 | * and update_cgrp_time_xx() and thus both timestamp | |
1373 | * are identical (or very close). Given that tstamp is, | |
1374 | * already adjusted for cgroup, we could say that: | |
1375 | * tstamp - ctx->timestamp | |
1376 | * is equivalent to | |
1377 | * tstamp - cgrp->timestamp. | |
1378 | * | |
1379 | * Then, in perf_output_read(), the calculation would | |
1380 | * work with no changes because: | |
1381 | * - event is guaranteed scheduled in | |
1382 | * - no scheduled out in between | |
1383 | * - thus the timestamp would be the same | |
1384 | * | |
1385 | * But this is a bit hairy. | |
1386 | * | |
1387 | * So instead, we have an explicit cgroup call to remain | |
1388 | * within the time time source all along. We believe it | |
1389 | * is cleaner and simpler to understand. | |
1390 | */ | |
1391 | if (is_cgroup_event(event)) | |
1392 | perf_cgroup_set_shadow_time(event, tstamp); | |
1393 | else | |
1394 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1395 | } | |
1396 | ||
4fe757dd PZ |
1397 | #define MAX_INTERRUPTS (~0ULL) |
1398 | ||
1399 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1400 | ||
235c7fc7 | 1401 | static int |
9ffcfa6f | 1402 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1403 | struct perf_cpu_context *cpuctx, |
6e37738a | 1404 | struct perf_event_context *ctx) |
235c7fc7 | 1405 | { |
4158755d SE |
1406 | u64 tstamp = perf_event_time(event); |
1407 | ||
cdd6c482 | 1408 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1409 | return 0; |
1410 | ||
cdd6c482 | 1411 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1412 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1413 | |
1414 | /* | |
1415 | * Unthrottle events, since we scheduled we might have missed several | |
1416 | * ticks already, also for a heavily scheduling task there is little | |
1417 | * guarantee it'll get a tick in a timely manner. | |
1418 | */ | |
1419 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1420 | perf_log_throttle(event, 1); | |
1421 | event->hw.interrupts = 0; | |
1422 | } | |
1423 | ||
235c7fc7 IM |
1424 | /* |
1425 | * The new state must be visible before we turn it on in the hardware: | |
1426 | */ | |
1427 | smp_wmb(); | |
1428 | ||
a4eaf7f1 | 1429 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1430 | event->state = PERF_EVENT_STATE_INACTIVE; |
1431 | event->oncpu = -1; | |
235c7fc7 IM |
1432 | return -EAGAIN; |
1433 | } | |
1434 | ||
4158755d | 1435 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1436 | |
e5d1367f | 1437 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1438 | |
cdd6c482 | 1439 | if (!is_software_event(event)) |
3b6f9e5c | 1440 | cpuctx->active_oncpu++; |
235c7fc7 | 1441 | ctx->nr_active++; |
0f5a2601 PZ |
1442 | if (event->attr.freq && event->attr.sample_freq) |
1443 | ctx->nr_freq++; | |
235c7fc7 | 1444 | |
cdd6c482 | 1445 | if (event->attr.exclusive) |
3b6f9e5c PM |
1446 | cpuctx->exclusive = 1; |
1447 | ||
235c7fc7 IM |
1448 | return 0; |
1449 | } | |
1450 | ||
6751b71e | 1451 | static int |
cdd6c482 | 1452 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1453 | struct perf_cpu_context *cpuctx, |
6e37738a | 1454 | struct perf_event_context *ctx) |
6751b71e | 1455 | { |
6bde9b6c | 1456 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1457 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1458 | u64 now = ctx->time; |
1459 | bool simulate = false; | |
6751b71e | 1460 | |
cdd6c482 | 1461 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1462 | return 0; |
1463 | ||
ad5133b7 | 1464 | pmu->start_txn(pmu); |
6bde9b6c | 1465 | |
9ffcfa6f | 1466 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1467 | pmu->cancel_txn(pmu); |
6751b71e | 1468 | return -EAGAIN; |
90151c35 | 1469 | } |
6751b71e PM |
1470 | |
1471 | /* | |
1472 | * Schedule in siblings as one group (if any): | |
1473 | */ | |
cdd6c482 | 1474 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1475 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1476 | partial_group = event; |
6751b71e PM |
1477 | goto group_error; |
1478 | } | |
1479 | } | |
1480 | ||
9ffcfa6f | 1481 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1482 | return 0; |
9ffcfa6f | 1483 | |
6751b71e PM |
1484 | group_error: |
1485 | /* | |
1486 | * Groups can be scheduled in as one unit only, so undo any | |
1487 | * partial group before returning: | |
d7842da4 SE |
1488 | * The events up to the failed event are scheduled out normally, |
1489 | * tstamp_stopped will be updated. | |
1490 | * | |
1491 | * The failed events and the remaining siblings need to have | |
1492 | * their timings updated as if they had gone thru event_sched_in() | |
1493 | * and event_sched_out(). This is required to get consistent timings | |
1494 | * across the group. This also takes care of the case where the group | |
1495 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1496 | * the time the event was actually stopped, such that time delta | |
1497 | * calculation in update_event_times() is correct. | |
6751b71e | 1498 | */ |
cdd6c482 IM |
1499 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1500 | if (event == partial_group) | |
d7842da4 SE |
1501 | simulate = true; |
1502 | ||
1503 | if (simulate) { | |
1504 | event->tstamp_running += now - event->tstamp_stopped; | |
1505 | event->tstamp_stopped = now; | |
1506 | } else { | |
1507 | event_sched_out(event, cpuctx, ctx); | |
1508 | } | |
6751b71e | 1509 | } |
9ffcfa6f | 1510 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1511 | |
ad5133b7 | 1512 | pmu->cancel_txn(pmu); |
90151c35 | 1513 | |
6751b71e PM |
1514 | return -EAGAIN; |
1515 | } | |
1516 | ||
3b6f9e5c | 1517 | /* |
cdd6c482 | 1518 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1519 | */ |
cdd6c482 | 1520 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1521 | struct perf_cpu_context *cpuctx, |
1522 | int can_add_hw) | |
1523 | { | |
1524 | /* | |
cdd6c482 | 1525 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1526 | */ |
d6f962b5 | 1527 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1528 | return 1; |
1529 | /* | |
1530 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1531 | * events can go on. |
3b6f9e5c PM |
1532 | */ |
1533 | if (cpuctx->exclusive) | |
1534 | return 0; | |
1535 | /* | |
1536 | * If this group is exclusive and there are already | |
cdd6c482 | 1537 | * events on the CPU, it can't go on. |
3b6f9e5c | 1538 | */ |
cdd6c482 | 1539 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1540 | return 0; |
1541 | /* | |
1542 | * Otherwise, try to add it if all previous groups were able | |
1543 | * to go on. | |
1544 | */ | |
1545 | return can_add_hw; | |
1546 | } | |
1547 | ||
cdd6c482 IM |
1548 | static void add_event_to_ctx(struct perf_event *event, |
1549 | struct perf_event_context *ctx) | |
53cfbf59 | 1550 | { |
4158755d SE |
1551 | u64 tstamp = perf_event_time(event); |
1552 | ||
cdd6c482 | 1553 | list_add_event(event, ctx); |
8a49542c | 1554 | perf_group_attach(event); |
4158755d SE |
1555 | event->tstamp_enabled = tstamp; |
1556 | event->tstamp_running = tstamp; | |
1557 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1558 | } |
1559 | ||
2c29ef0f PZ |
1560 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1561 | static void | |
1562 | ctx_sched_in(struct perf_event_context *ctx, | |
1563 | struct perf_cpu_context *cpuctx, | |
1564 | enum event_type_t event_type, | |
1565 | struct task_struct *task); | |
fe4b04fa | 1566 | |
dce5855b PZ |
1567 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1568 | struct perf_event_context *ctx, | |
1569 | struct task_struct *task) | |
1570 | { | |
1571 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1572 | if (ctx) | |
1573 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1574 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1575 | if (ctx) | |
1576 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1577 | } | |
1578 | ||
0793a61d | 1579 | /* |
cdd6c482 | 1580 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1581 | * |
1582 | * Must be called with ctx->mutex held | |
0793a61d | 1583 | */ |
fe4b04fa | 1584 | static int __perf_install_in_context(void *info) |
0793a61d | 1585 | { |
cdd6c482 IM |
1586 | struct perf_event *event = info; |
1587 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1588 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1589 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1590 | struct task_struct *task = current; | |
1591 | ||
b58f6b0d | 1592 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1593 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1594 | |
1595 | /* | |
2c29ef0f | 1596 | * If there was an active task_ctx schedule it out. |
0793a61d | 1597 | */ |
b58f6b0d | 1598 | if (task_ctx) |
2c29ef0f | 1599 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1600 | |
1601 | /* | |
1602 | * If the context we're installing events in is not the | |
1603 | * active task_ctx, flip them. | |
1604 | */ | |
1605 | if (ctx->task && task_ctx != ctx) { | |
1606 | if (task_ctx) | |
1607 | raw_spin_unlock(&task_ctx->lock); | |
1608 | raw_spin_lock(&ctx->lock); | |
1609 | task_ctx = ctx; | |
1610 | } | |
1611 | ||
1612 | if (task_ctx) { | |
1613 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1614 | task = task_ctx->task; |
1615 | } | |
b58f6b0d | 1616 | |
2c29ef0f | 1617 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1618 | |
4af4998b | 1619 | update_context_time(ctx); |
e5d1367f SE |
1620 | /* |
1621 | * update cgrp time only if current cgrp | |
1622 | * matches event->cgrp. Must be done before | |
1623 | * calling add_event_to_ctx() | |
1624 | */ | |
1625 | update_cgrp_time_from_event(event); | |
0793a61d | 1626 | |
cdd6c482 | 1627 | add_event_to_ctx(event, ctx); |
0793a61d | 1628 | |
d859e29f | 1629 | /* |
2c29ef0f | 1630 | * Schedule everything back in |
d859e29f | 1631 | */ |
dce5855b | 1632 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1633 | |
1634 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1635 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1636 | |
1637 | return 0; | |
0793a61d TG |
1638 | } |
1639 | ||
1640 | /* | |
cdd6c482 | 1641 | * Attach a performance event to a context |
0793a61d | 1642 | * |
cdd6c482 IM |
1643 | * First we add the event to the list with the hardware enable bit |
1644 | * in event->hw_config cleared. | |
0793a61d | 1645 | * |
cdd6c482 | 1646 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1647 | * call to enable it in the task context. The task might have been |
1648 | * scheduled away, but we check this in the smp call again. | |
1649 | */ | |
1650 | static void | |
cdd6c482 IM |
1651 | perf_install_in_context(struct perf_event_context *ctx, |
1652 | struct perf_event *event, | |
0793a61d TG |
1653 | int cpu) |
1654 | { | |
1655 | struct task_struct *task = ctx->task; | |
1656 | ||
fe4b04fa PZ |
1657 | lockdep_assert_held(&ctx->mutex); |
1658 | ||
c3f00c70 | 1659 | event->ctx = ctx; |
0cda4c02 YZ |
1660 | if (event->cpu != -1) |
1661 | event->cpu = cpu; | |
c3f00c70 | 1662 | |
0793a61d TG |
1663 | if (!task) { |
1664 | /* | |
cdd6c482 | 1665 | * Per cpu events are installed via an smp call and |
af901ca1 | 1666 | * the install is always successful. |
0793a61d | 1667 | */ |
fe4b04fa | 1668 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1669 | return; |
1670 | } | |
1671 | ||
0793a61d | 1672 | retry: |
fe4b04fa PZ |
1673 | if (!task_function_call(task, __perf_install_in_context, event)) |
1674 | return; | |
0793a61d | 1675 | |
e625cce1 | 1676 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1677 | /* |
fe4b04fa PZ |
1678 | * If we failed to find a running task, but find the context active now |
1679 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1680 | */ |
fe4b04fa | 1681 | if (ctx->is_active) { |
e625cce1 | 1682 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1683 | goto retry; |
1684 | } | |
1685 | ||
1686 | /* | |
fe4b04fa PZ |
1687 | * Since the task isn't running, its safe to add the event, us holding |
1688 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1689 | */ |
fe4b04fa | 1690 | add_event_to_ctx(event, ctx); |
e625cce1 | 1691 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1692 | } |
1693 | ||
fa289bec | 1694 | /* |
cdd6c482 | 1695 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1696 | * Enabling the leader of a group effectively enables all |
1697 | * the group members that aren't explicitly disabled, so we | |
1698 | * have to update their ->tstamp_enabled also. | |
1699 | * Note: this works for group members as well as group leaders | |
1700 | * since the non-leader members' sibling_lists will be empty. | |
1701 | */ | |
1d9b482e | 1702 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1703 | { |
cdd6c482 | 1704 | struct perf_event *sub; |
4158755d | 1705 | u64 tstamp = perf_event_time(event); |
fa289bec | 1706 | |
cdd6c482 | 1707 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1708 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1709 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1710 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1711 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1712 | } |
fa289bec PM |
1713 | } |
1714 | ||
d859e29f | 1715 | /* |
cdd6c482 | 1716 | * Cross CPU call to enable a performance event |
d859e29f | 1717 | */ |
fe4b04fa | 1718 | static int __perf_event_enable(void *info) |
04289bb9 | 1719 | { |
cdd6c482 | 1720 | struct perf_event *event = info; |
cdd6c482 IM |
1721 | struct perf_event_context *ctx = event->ctx; |
1722 | struct perf_event *leader = event->group_leader; | |
108b02cf | 1723 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 1724 | int err; |
04289bb9 | 1725 | |
fe4b04fa PZ |
1726 | if (WARN_ON_ONCE(!ctx->is_active)) |
1727 | return -EINVAL; | |
3cbed429 | 1728 | |
e625cce1 | 1729 | raw_spin_lock(&ctx->lock); |
4af4998b | 1730 | update_context_time(ctx); |
d859e29f | 1731 | |
cdd6c482 | 1732 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 1733 | goto unlock; |
e5d1367f SE |
1734 | |
1735 | /* | |
1736 | * set current task's cgroup time reference point | |
1737 | */ | |
3f7cce3c | 1738 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 1739 | |
1d9b482e | 1740 | __perf_event_mark_enabled(event); |
04289bb9 | 1741 | |
e5d1367f SE |
1742 | if (!event_filter_match(event)) { |
1743 | if (is_cgroup_event(event)) | |
1744 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 1745 | goto unlock; |
e5d1367f | 1746 | } |
f4c4176f | 1747 | |
04289bb9 | 1748 | /* |
cdd6c482 | 1749 | * If the event is in a group and isn't the group leader, |
d859e29f | 1750 | * then don't put it on unless the group is on. |
04289bb9 | 1751 | */ |
cdd6c482 | 1752 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 1753 | goto unlock; |
3b6f9e5c | 1754 | |
cdd6c482 | 1755 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 1756 | err = -EEXIST; |
e758a33d | 1757 | } else { |
cdd6c482 | 1758 | if (event == leader) |
6e37738a | 1759 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 1760 | else |
6e37738a | 1761 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 1762 | } |
d859e29f PM |
1763 | |
1764 | if (err) { | |
1765 | /* | |
cdd6c482 | 1766 | * If this event can't go on and it's part of a |
d859e29f PM |
1767 | * group, then the whole group has to come off. |
1768 | */ | |
cdd6c482 | 1769 | if (leader != event) |
d859e29f | 1770 | group_sched_out(leader, cpuctx, ctx); |
0d48696f | 1771 | if (leader->attr.pinned) { |
53cfbf59 | 1772 | update_group_times(leader); |
cdd6c482 | 1773 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 1774 | } |
d859e29f PM |
1775 | } |
1776 | ||
9ed6060d | 1777 | unlock: |
e625cce1 | 1778 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1779 | |
1780 | return 0; | |
d859e29f PM |
1781 | } |
1782 | ||
1783 | /* | |
cdd6c482 | 1784 | * Enable a event. |
c93f7669 | 1785 | * |
cdd6c482 IM |
1786 | * If event->ctx is a cloned context, callers must make sure that |
1787 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1788 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
1789 | * perf_event_for_each_child or perf_event_for_each as described |
1790 | * for perf_event_disable. | |
d859e29f | 1791 | */ |
44234adc | 1792 | void perf_event_enable(struct perf_event *event) |
d859e29f | 1793 | { |
cdd6c482 | 1794 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1795 | struct task_struct *task = ctx->task; |
1796 | ||
1797 | if (!task) { | |
1798 | /* | |
cdd6c482 | 1799 | * Enable the event on the cpu that it's on |
d859e29f | 1800 | */ |
fe4b04fa | 1801 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
1802 | return; |
1803 | } | |
1804 | ||
e625cce1 | 1805 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 1806 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
1807 | goto out; |
1808 | ||
1809 | /* | |
cdd6c482 IM |
1810 | * If the event is in error state, clear that first. |
1811 | * That way, if we see the event in error state below, we | |
d859e29f PM |
1812 | * know that it has gone back into error state, as distinct |
1813 | * from the task having been scheduled away before the | |
1814 | * cross-call arrived. | |
1815 | */ | |
cdd6c482 IM |
1816 | if (event->state == PERF_EVENT_STATE_ERROR) |
1817 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 1818 | |
9ed6060d | 1819 | retry: |
fe4b04fa | 1820 | if (!ctx->is_active) { |
1d9b482e | 1821 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
1822 | goto out; |
1823 | } | |
1824 | ||
e625cce1 | 1825 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1826 | |
1827 | if (!task_function_call(task, __perf_event_enable, event)) | |
1828 | return; | |
d859e29f | 1829 | |
e625cce1 | 1830 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
1831 | |
1832 | /* | |
cdd6c482 | 1833 | * If the context is active and the event is still off, |
d859e29f PM |
1834 | * we need to retry the cross-call. |
1835 | */ | |
fe4b04fa PZ |
1836 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
1837 | /* | |
1838 | * task could have been flipped by a concurrent | |
1839 | * perf_event_context_sched_out() | |
1840 | */ | |
1841 | task = ctx->task; | |
d859e29f | 1842 | goto retry; |
fe4b04fa | 1843 | } |
fa289bec | 1844 | |
9ed6060d | 1845 | out: |
e625cce1 | 1846 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1847 | } |
dcfce4a0 | 1848 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 1849 | |
26ca5c11 | 1850 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 1851 | { |
2023b359 | 1852 | /* |
cdd6c482 | 1853 | * not supported on inherited events |
2023b359 | 1854 | */ |
2e939d1d | 1855 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
1856 | return -EINVAL; |
1857 | ||
cdd6c482 IM |
1858 | atomic_add(refresh, &event->event_limit); |
1859 | perf_event_enable(event); | |
2023b359 PZ |
1860 | |
1861 | return 0; | |
79f14641 | 1862 | } |
26ca5c11 | 1863 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 1864 | |
5b0311e1 FW |
1865 | static void ctx_sched_out(struct perf_event_context *ctx, |
1866 | struct perf_cpu_context *cpuctx, | |
1867 | enum event_type_t event_type) | |
235c7fc7 | 1868 | { |
cdd6c482 | 1869 | struct perf_event *event; |
db24d33e | 1870 | int is_active = ctx->is_active; |
235c7fc7 | 1871 | |
db24d33e | 1872 | ctx->is_active &= ~event_type; |
cdd6c482 | 1873 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
1874 | return; |
1875 | ||
4af4998b | 1876 | update_context_time(ctx); |
e5d1367f | 1877 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 1878 | if (!ctx->nr_active) |
facc4307 | 1879 | return; |
5b0311e1 | 1880 | |
075e0b00 | 1881 | perf_pmu_disable(ctx->pmu); |
db24d33e | 1882 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
1883 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
1884 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 1885 | } |
889ff015 | 1886 | |
db24d33e | 1887 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 1888 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 1889 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 1890 | } |
1b9a644f | 1891 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
1892 | } |
1893 | ||
564c2b21 PM |
1894 | /* |
1895 | * Test whether two contexts are equivalent, i.e. whether they | |
1896 | * have both been cloned from the same version of the same context | |
cdd6c482 IM |
1897 | * and they both have the same number of enabled events. |
1898 | * If the number of enabled events is the same, then the set | |
1899 | * of enabled events should be the same, because these are both | |
1900 | * inherited contexts, therefore we can't access individual events | |
564c2b21 | 1901 | * in them directly with an fd; we can only enable/disable all |
cdd6c482 | 1902 | * events via prctl, or enable/disable all events in a family |
564c2b21 PM |
1903 | * via ioctl, which will have the same effect on both contexts. |
1904 | */ | |
cdd6c482 IM |
1905 | static int context_equiv(struct perf_event_context *ctx1, |
1906 | struct perf_event_context *ctx2) | |
564c2b21 PM |
1907 | { |
1908 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx | |
ad3a37de | 1909 | && ctx1->parent_gen == ctx2->parent_gen |
25346b93 | 1910 | && !ctx1->pin_count && !ctx2->pin_count; |
564c2b21 PM |
1911 | } |
1912 | ||
cdd6c482 IM |
1913 | static void __perf_event_sync_stat(struct perf_event *event, |
1914 | struct perf_event *next_event) | |
bfbd3381 PZ |
1915 | { |
1916 | u64 value; | |
1917 | ||
cdd6c482 | 1918 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
1919 | return; |
1920 | ||
1921 | /* | |
cdd6c482 | 1922 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
1923 | * because we're in the middle of a context switch and have IRQs |
1924 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 1925 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
1926 | * don't need to use it. |
1927 | */ | |
cdd6c482 IM |
1928 | switch (event->state) { |
1929 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
1930 | event->pmu->read(event); |
1931 | /* fall-through */ | |
bfbd3381 | 1932 | |
cdd6c482 IM |
1933 | case PERF_EVENT_STATE_INACTIVE: |
1934 | update_event_times(event); | |
bfbd3381 PZ |
1935 | break; |
1936 | ||
1937 | default: | |
1938 | break; | |
1939 | } | |
1940 | ||
1941 | /* | |
cdd6c482 | 1942 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
1943 | * values when we flip the contexts. |
1944 | */ | |
e7850595 PZ |
1945 | value = local64_read(&next_event->count); |
1946 | value = local64_xchg(&event->count, value); | |
1947 | local64_set(&next_event->count, value); | |
bfbd3381 | 1948 | |
cdd6c482 IM |
1949 | swap(event->total_time_enabled, next_event->total_time_enabled); |
1950 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 1951 | |
bfbd3381 | 1952 | /* |
19d2e755 | 1953 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 1954 | */ |
cdd6c482 IM |
1955 | perf_event_update_userpage(event); |
1956 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
1957 | } |
1958 | ||
1959 | #define list_next_entry(pos, member) \ | |
1960 | list_entry(pos->member.next, typeof(*pos), member) | |
1961 | ||
cdd6c482 IM |
1962 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
1963 | struct perf_event_context *next_ctx) | |
bfbd3381 | 1964 | { |
cdd6c482 | 1965 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
1966 | |
1967 | if (!ctx->nr_stat) | |
1968 | return; | |
1969 | ||
02ffdbc8 PZ |
1970 | update_context_time(ctx); |
1971 | ||
cdd6c482 IM |
1972 | event = list_first_entry(&ctx->event_list, |
1973 | struct perf_event, event_entry); | |
bfbd3381 | 1974 | |
cdd6c482 IM |
1975 | next_event = list_first_entry(&next_ctx->event_list, |
1976 | struct perf_event, event_entry); | |
bfbd3381 | 1977 | |
cdd6c482 IM |
1978 | while (&event->event_entry != &ctx->event_list && |
1979 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 1980 | |
cdd6c482 | 1981 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 1982 | |
cdd6c482 IM |
1983 | event = list_next_entry(event, event_entry); |
1984 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
1985 | } |
1986 | } | |
1987 | ||
fe4b04fa PZ |
1988 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
1989 | struct task_struct *next) | |
0793a61d | 1990 | { |
8dc85d54 | 1991 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 IM |
1992 | struct perf_event_context *next_ctx; |
1993 | struct perf_event_context *parent; | |
108b02cf | 1994 | struct perf_cpu_context *cpuctx; |
c93f7669 | 1995 | int do_switch = 1; |
0793a61d | 1996 | |
108b02cf PZ |
1997 | if (likely(!ctx)) |
1998 | return; | |
10989fb2 | 1999 | |
108b02cf PZ |
2000 | cpuctx = __get_cpu_context(ctx); |
2001 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2002 | return; |
2003 | ||
c93f7669 PM |
2004 | rcu_read_lock(); |
2005 | parent = rcu_dereference(ctx->parent_ctx); | |
8dc85d54 | 2006 | next_ctx = next->perf_event_ctxp[ctxn]; |
c93f7669 PM |
2007 | if (parent && next_ctx && |
2008 | rcu_dereference(next_ctx->parent_ctx) == parent) { | |
2009 | /* | |
2010 | * Looks like the two contexts are clones, so we might be | |
2011 | * able to optimize the context switch. We lock both | |
2012 | * contexts and check that they are clones under the | |
2013 | * lock (including re-checking that neither has been | |
2014 | * uncloned in the meantime). It doesn't matter which | |
2015 | * order we take the locks because no other cpu could | |
2016 | * be trying to lock both of these tasks. | |
2017 | */ | |
e625cce1 TG |
2018 | raw_spin_lock(&ctx->lock); |
2019 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2020 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2021 | /* |
2022 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2023 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2024 | */ |
8dc85d54 PZ |
2025 | task->perf_event_ctxp[ctxn] = next_ctx; |
2026 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2027 | ctx->task = next; |
2028 | next_ctx->task = task; | |
2029 | do_switch = 0; | |
bfbd3381 | 2030 | |
cdd6c482 | 2031 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2032 | } |
e625cce1 TG |
2033 | raw_spin_unlock(&next_ctx->lock); |
2034 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2035 | } |
c93f7669 | 2036 | rcu_read_unlock(); |
564c2b21 | 2037 | |
c93f7669 | 2038 | if (do_switch) { |
facc4307 | 2039 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2040 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2041 | cpuctx->task_ctx = NULL; |
facc4307 | 2042 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2043 | } |
0793a61d TG |
2044 | } |
2045 | ||
8dc85d54 PZ |
2046 | #define for_each_task_context_nr(ctxn) \ |
2047 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2048 | ||
2049 | /* | |
2050 | * Called from scheduler to remove the events of the current task, | |
2051 | * with interrupts disabled. | |
2052 | * | |
2053 | * We stop each event and update the event value in event->count. | |
2054 | * | |
2055 | * This does not protect us against NMI, but disable() | |
2056 | * sets the disabled bit in the control field of event _before_ | |
2057 | * accessing the event control register. If a NMI hits, then it will | |
2058 | * not restart the event. | |
2059 | */ | |
ab0cce56 JO |
2060 | void __perf_event_task_sched_out(struct task_struct *task, |
2061 | struct task_struct *next) | |
8dc85d54 PZ |
2062 | { |
2063 | int ctxn; | |
2064 | ||
8dc85d54 PZ |
2065 | for_each_task_context_nr(ctxn) |
2066 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2067 | |
2068 | /* | |
2069 | * if cgroup events exist on this CPU, then we need | |
2070 | * to check if we have to switch out PMU state. | |
2071 | * cgroup event are system-wide mode only | |
2072 | */ | |
2073 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2074 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2075 | } |
2076 | ||
04dc2dbb | 2077 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2078 | { |
108b02cf | 2079 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2080 | |
a63eaf34 PM |
2081 | if (!cpuctx->task_ctx) |
2082 | return; | |
012b84da IM |
2083 | |
2084 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2085 | return; | |
2086 | ||
04dc2dbb | 2087 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2088 | cpuctx->task_ctx = NULL; |
2089 | } | |
2090 | ||
5b0311e1 FW |
2091 | /* |
2092 | * Called with IRQs disabled | |
2093 | */ | |
2094 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2095 | enum event_type_t event_type) | |
2096 | { | |
2097 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2098 | } |
2099 | ||
235c7fc7 | 2100 | static void |
5b0311e1 | 2101 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2102 | struct perf_cpu_context *cpuctx) |
0793a61d | 2103 | { |
cdd6c482 | 2104 | struct perf_event *event; |
0793a61d | 2105 | |
889ff015 FW |
2106 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2107 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2108 | continue; |
5632ab12 | 2109 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2110 | continue; |
2111 | ||
e5d1367f SE |
2112 | /* may need to reset tstamp_enabled */ |
2113 | if (is_cgroup_event(event)) | |
2114 | perf_cgroup_mark_enabled(event, ctx); | |
2115 | ||
8c9ed8e1 | 2116 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2117 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2118 | |
2119 | /* | |
2120 | * If this pinned group hasn't been scheduled, | |
2121 | * put it in error state. | |
2122 | */ | |
cdd6c482 IM |
2123 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2124 | update_group_times(event); | |
2125 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2126 | } |
3b6f9e5c | 2127 | } |
5b0311e1 FW |
2128 | } |
2129 | ||
2130 | static void | |
2131 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2132 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2133 | { |
2134 | struct perf_event *event; | |
2135 | int can_add_hw = 1; | |
3b6f9e5c | 2136 | |
889ff015 FW |
2137 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2138 | /* Ignore events in OFF or ERROR state */ | |
2139 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2140 | continue; |
04289bb9 IM |
2141 | /* |
2142 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2143 | * of events: |
04289bb9 | 2144 | */ |
5632ab12 | 2145 | if (!event_filter_match(event)) |
0793a61d TG |
2146 | continue; |
2147 | ||
e5d1367f SE |
2148 | /* may need to reset tstamp_enabled */ |
2149 | if (is_cgroup_event(event)) | |
2150 | perf_cgroup_mark_enabled(event, ctx); | |
2151 | ||
9ed6060d | 2152 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2153 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2154 | can_add_hw = 0; |
9ed6060d | 2155 | } |
0793a61d | 2156 | } |
5b0311e1 FW |
2157 | } |
2158 | ||
2159 | static void | |
2160 | ctx_sched_in(struct perf_event_context *ctx, | |
2161 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2162 | enum event_type_t event_type, |
2163 | struct task_struct *task) | |
5b0311e1 | 2164 | { |
e5d1367f | 2165 | u64 now; |
db24d33e | 2166 | int is_active = ctx->is_active; |
e5d1367f | 2167 | |
db24d33e | 2168 | ctx->is_active |= event_type; |
5b0311e1 | 2169 | if (likely(!ctx->nr_events)) |
facc4307 | 2170 | return; |
5b0311e1 | 2171 | |
e5d1367f SE |
2172 | now = perf_clock(); |
2173 | ctx->timestamp = now; | |
3f7cce3c | 2174 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2175 | /* |
2176 | * First go through the list and put on any pinned groups | |
2177 | * in order to give them the best chance of going on. | |
2178 | */ | |
db24d33e | 2179 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2180 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2181 | |
2182 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2183 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2184 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2185 | } |
2186 | ||
329c0e01 | 2187 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2188 | enum event_type_t event_type, |
2189 | struct task_struct *task) | |
329c0e01 FW |
2190 | { |
2191 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2192 | ||
e5d1367f | 2193 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2194 | } |
2195 | ||
e5d1367f SE |
2196 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2197 | struct task_struct *task) | |
235c7fc7 | 2198 | { |
108b02cf | 2199 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2200 | |
108b02cf | 2201 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2202 | if (cpuctx->task_ctx == ctx) |
2203 | return; | |
2204 | ||
facc4307 | 2205 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2206 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2207 | /* |
2208 | * We want to keep the following priority order: | |
2209 | * cpu pinned (that don't need to move), task pinned, | |
2210 | * cpu flexible, task flexible. | |
2211 | */ | |
2212 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2213 | ||
1d5f003f GN |
2214 | if (ctx->nr_events) |
2215 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2216 | |
86b47c25 GN |
2217 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2218 | ||
facc4307 PZ |
2219 | perf_pmu_enable(ctx->pmu); |
2220 | perf_ctx_unlock(cpuctx, ctx); | |
2221 | ||
b5ab4cd5 PZ |
2222 | /* |
2223 | * Since these rotations are per-cpu, we need to ensure the | |
2224 | * cpu-context we got scheduled on is actually rotating. | |
2225 | */ | |
108b02cf | 2226 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2227 | } |
2228 | ||
d010b332 SE |
2229 | /* |
2230 | * When sampling the branck stack in system-wide, it may be necessary | |
2231 | * to flush the stack on context switch. This happens when the branch | |
2232 | * stack does not tag its entries with the pid of the current task. | |
2233 | * Otherwise it becomes impossible to associate a branch entry with a | |
2234 | * task. This ambiguity is more likely to appear when the branch stack | |
2235 | * supports priv level filtering and the user sets it to monitor only | |
2236 | * at the user level (which could be a useful measurement in system-wide | |
2237 | * mode). In that case, the risk is high of having a branch stack with | |
2238 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2239 | * entries or stashing them somewhere in the PMU specific code layer. | |
2240 | * | |
2241 | * This function provides the context switch callback to the lower code | |
2242 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2243 | * with at least one active event using taken branch sampling. | |
2244 | */ | |
2245 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2246 | struct task_struct *task) | |
2247 | { | |
2248 | struct perf_cpu_context *cpuctx; | |
2249 | struct pmu *pmu; | |
2250 | unsigned long flags; | |
2251 | ||
2252 | /* no need to flush branch stack if not changing task */ | |
2253 | if (prev == task) | |
2254 | return; | |
2255 | ||
2256 | local_irq_save(flags); | |
2257 | ||
2258 | rcu_read_lock(); | |
2259 | ||
2260 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2261 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2262 | ||
2263 | /* | |
2264 | * check if the context has at least one | |
2265 | * event using PERF_SAMPLE_BRANCH_STACK | |
2266 | */ | |
2267 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2268 | && pmu->flush_branch_stack) { | |
2269 | ||
2270 | pmu = cpuctx->ctx.pmu; | |
2271 | ||
2272 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2273 | ||
2274 | perf_pmu_disable(pmu); | |
2275 | ||
2276 | pmu->flush_branch_stack(); | |
2277 | ||
2278 | perf_pmu_enable(pmu); | |
2279 | ||
2280 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2281 | } | |
2282 | } | |
2283 | ||
2284 | rcu_read_unlock(); | |
2285 | ||
2286 | local_irq_restore(flags); | |
2287 | } | |
2288 | ||
8dc85d54 PZ |
2289 | /* |
2290 | * Called from scheduler to add the events of the current task | |
2291 | * with interrupts disabled. | |
2292 | * | |
2293 | * We restore the event value and then enable it. | |
2294 | * | |
2295 | * This does not protect us against NMI, but enable() | |
2296 | * sets the enabled bit in the control field of event _before_ | |
2297 | * accessing the event control register. If a NMI hits, then it will | |
2298 | * keep the event running. | |
2299 | */ | |
ab0cce56 JO |
2300 | void __perf_event_task_sched_in(struct task_struct *prev, |
2301 | struct task_struct *task) | |
8dc85d54 PZ |
2302 | { |
2303 | struct perf_event_context *ctx; | |
2304 | int ctxn; | |
2305 | ||
2306 | for_each_task_context_nr(ctxn) { | |
2307 | ctx = task->perf_event_ctxp[ctxn]; | |
2308 | if (likely(!ctx)) | |
2309 | continue; | |
2310 | ||
e5d1367f | 2311 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2312 | } |
e5d1367f SE |
2313 | /* |
2314 | * if cgroup events exist on this CPU, then we need | |
2315 | * to check if we have to switch in PMU state. | |
2316 | * cgroup event are system-wide mode only | |
2317 | */ | |
2318 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2319 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2320 | |
2321 | /* check for system-wide branch_stack events */ | |
2322 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2323 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2324 | } |
2325 | ||
abd50713 PZ |
2326 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2327 | { | |
2328 | u64 frequency = event->attr.sample_freq; | |
2329 | u64 sec = NSEC_PER_SEC; | |
2330 | u64 divisor, dividend; | |
2331 | ||
2332 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2333 | ||
2334 | count_fls = fls64(count); | |
2335 | nsec_fls = fls64(nsec); | |
2336 | frequency_fls = fls64(frequency); | |
2337 | sec_fls = 30; | |
2338 | ||
2339 | /* | |
2340 | * We got @count in @nsec, with a target of sample_freq HZ | |
2341 | * the target period becomes: | |
2342 | * | |
2343 | * @count * 10^9 | |
2344 | * period = ------------------- | |
2345 | * @nsec * sample_freq | |
2346 | * | |
2347 | */ | |
2348 | ||
2349 | /* | |
2350 | * Reduce accuracy by one bit such that @a and @b converge | |
2351 | * to a similar magnitude. | |
2352 | */ | |
fe4b04fa | 2353 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2354 | do { \ |
2355 | if (a##_fls > b##_fls) { \ | |
2356 | a >>= 1; \ | |
2357 | a##_fls--; \ | |
2358 | } else { \ | |
2359 | b >>= 1; \ | |
2360 | b##_fls--; \ | |
2361 | } \ | |
2362 | } while (0) | |
2363 | ||
2364 | /* | |
2365 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2366 | * the other, so that finally we can do a u64/u64 division. | |
2367 | */ | |
2368 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2369 | REDUCE_FLS(nsec, frequency); | |
2370 | REDUCE_FLS(sec, count); | |
2371 | } | |
2372 | ||
2373 | if (count_fls + sec_fls > 64) { | |
2374 | divisor = nsec * frequency; | |
2375 | ||
2376 | while (count_fls + sec_fls > 64) { | |
2377 | REDUCE_FLS(count, sec); | |
2378 | divisor >>= 1; | |
2379 | } | |
2380 | ||
2381 | dividend = count * sec; | |
2382 | } else { | |
2383 | dividend = count * sec; | |
2384 | ||
2385 | while (nsec_fls + frequency_fls > 64) { | |
2386 | REDUCE_FLS(nsec, frequency); | |
2387 | dividend >>= 1; | |
2388 | } | |
2389 | ||
2390 | divisor = nsec * frequency; | |
2391 | } | |
2392 | ||
f6ab91ad PZ |
2393 | if (!divisor) |
2394 | return dividend; | |
2395 | ||
abd50713 PZ |
2396 | return div64_u64(dividend, divisor); |
2397 | } | |
2398 | ||
e050e3f0 SE |
2399 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2400 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2401 | ||
f39d47ff | 2402 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2403 | { |
cdd6c482 | 2404 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2405 | s64 period, sample_period; |
bd2b5b12 PZ |
2406 | s64 delta; |
2407 | ||
abd50713 | 2408 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2409 | |
2410 | delta = (s64)(period - hwc->sample_period); | |
2411 | delta = (delta + 7) / 8; /* low pass filter */ | |
2412 | ||
2413 | sample_period = hwc->sample_period + delta; | |
2414 | ||
2415 | if (!sample_period) | |
2416 | sample_period = 1; | |
2417 | ||
bd2b5b12 | 2418 | hwc->sample_period = sample_period; |
abd50713 | 2419 | |
e7850595 | 2420 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2421 | if (disable) |
2422 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2423 | ||
e7850595 | 2424 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2425 | |
2426 | if (disable) | |
2427 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2428 | } |
bd2b5b12 PZ |
2429 | } |
2430 | ||
e050e3f0 SE |
2431 | /* |
2432 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2433 | * events. At the same time, make sure, having freq events does not change | |
2434 | * the rate of unthrottling as that would introduce bias. | |
2435 | */ | |
2436 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2437 | int needs_unthr) | |
60db5e09 | 2438 | { |
cdd6c482 IM |
2439 | struct perf_event *event; |
2440 | struct hw_perf_event *hwc; | |
e050e3f0 | 2441 | u64 now, period = TICK_NSEC; |
abd50713 | 2442 | s64 delta; |
60db5e09 | 2443 | |
e050e3f0 SE |
2444 | /* |
2445 | * only need to iterate over all events iff: | |
2446 | * - context have events in frequency mode (needs freq adjust) | |
2447 | * - there are events to unthrottle on this cpu | |
2448 | */ | |
2449 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2450 | return; |
2451 | ||
e050e3f0 | 2452 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2453 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2454 | |
03541f8b | 2455 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2456 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2457 | continue; |
2458 | ||
5632ab12 | 2459 | if (!event_filter_match(event)) |
5d27c23d PZ |
2460 | continue; |
2461 | ||
cdd6c482 | 2462 | hwc = &event->hw; |
6a24ed6c | 2463 | |
e050e3f0 SE |
2464 | if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) { |
2465 | hwc->interrupts = 0; | |
cdd6c482 | 2466 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2467 | event->pmu->start(event, 0); |
a78ac325 PZ |
2468 | } |
2469 | ||
cdd6c482 | 2470 | if (!event->attr.freq || !event->attr.sample_freq) |
60db5e09 PZ |
2471 | continue; |
2472 | ||
e050e3f0 SE |
2473 | /* |
2474 | * stop the event and update event->count | |
2475 | */ | |
2476 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2477 | ||
e7850595 | 2478 | now = local64_read(&event->count); |
abd50713 PZ |
2479 | delta = now - hwc->freq_count_stamp; |
2480 | hwc->freq_count_stamp = now; | |
60db5e09 | 2481 | |
e050e3f0 SE |
2482 | /* |
2483 | * restart the event | |
2484 | * reload only if value has changed | |
f39d47ff SE |
2485 | * we have stopped the event so tell that |
2486 | * to perf_adjust_period() to avoid stopping it | |
2487 | * twice. | |
e050e3f0 | 2488 | */ |
abd50713 | 2489 | if (delta > 0) |
f39d47ff | 2490 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2491 | |
2492 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
60db5e09 | 2493 | } |
e050e3f0 | 2494 | |
f39d47ff | 2495 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2496 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2497 | } |
2498 | ||
235c7fc7 | 2499 | /* |
cdd6c482 | 2500 | * Round-robin a context's events: |
235c7fc7 | 2501 | */ |
cdd6c482 | 2502 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2503 | { |
dddd3379 TG |
2504 | /* |
2505 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2506 | * disabled by the inheritance code. | |
2507 | */ | |
2508 | if (!ctx->rotate_disable) | |
2509 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2510 | } |
2511 | ||
b5ab4cd5 | 2512 | /* |
e9d2b064 PZ |
2513 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2514 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2515 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2516 | */ |
e9d2b064 | 2517 | static void perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2518 | { |
8dc85d54 | 2519 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2520 | int rotate = 0, remove = 1; |
7fc23a53 | 2521 | |
b5ab4cd5 | 2522 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2523 | remove = 0; |
b5ab4cd5 PZ |
2524 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2525 | rotate = 1; | |
2526 | } | |
235c7fc7 | 2527 | |
8dc85d54 | 2528 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2529 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2530 | remove = 0; |
b5ab4cd5 PZ |
2531 | if (ctx->nr_events != ctx->nr_active) |
2532 | rotate = 1; | |
2533 | } | |
9717e6cd | 2534 | |
e050e3f0 | 2535 | if (!rotate) |
0f5a2601 PZ |
2536 | goto done; |
2537 | ||
facc4307 | 2538 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2539 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2540 | |
e050e3f0 SE |
2541 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2542 | if (ctx) | |
2543 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2544 | |
e050e3f0 SE |
2545 | rotate_ctx(&cpuctx->ctx); |
2546 | if (ctx) | |
2547 | rotate_ctx(ctx); | |
235c7fc7 | 2548 | |
e050e3f0 | 2549 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2550 | |
0f5a2601 PZ |
2551 | perf_pmu_enable(cpuctx->ctx.pmu); |
2552 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2553 | done: |
e9d2b064 PZ |
2554 | if (remove) |
2555 | list_del_init(&cpuctx->rotation_list); | |
e9d2b064 PZ |
2556 | } |
2557 | ||
2558 | void perf_event_task_tick(void) | |
2559 | { | |
2560 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2561 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2562 | struct perf_event_context *ctx; |
2563 | int throttled; | |
b5ab4cd5 | 2564 | |
e9d2b064 PZ |
2565 | WARN_ON(!irqs_disabled()); |
2566 | ||
e050e3f0 SE |
2567 | __this_cpu_inc(perf_throttled_seq); |
2568 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2569 | ||
e9d2b064 | 2570 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2571 | ctx = &cpuctx->ctx; |
2572 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2573 | ||
2574 | ctx = cpuctx->task_ctx; | |
2575 | if (ctx) | |
2576 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2577 | ||
e9d2b064 PZ |
2578 | if (cpuctx->jiffies_interval == 1 || |
2579 | !(jiffies % cpuctx->jiffies_interval)) | |
2580 | perf_rotate_context(cpuctx); | |
2581 | } | |
0793a61d TG |
2582 | } |
2583 | ||
889ff015 FW |
2584 | static int event_enable_on_exec(struct perf_event *event, |
2585 | struct perf_event_context *ctx) | |
2586 | { | |
2587 | if (!event->attr.enable_on_exec) | |
2588 | return 0; | |
2589 | ||
2590 | event->attr.enable_on_exec = 0; | |
2591 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2592 | return 0; | |
2593 | ||
1d9b482e | 2594 | __perf_event_mark_enabled(event); |
889ff015 FW |
2595 | |
2596 | return 1; | |
2597 | } | |
2598 | ||
57e7986e | 2599 | /* |
cdd6c482 | 2600 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2601 | * This expects task == current. |
2602 | */ | |
8dc85d54 | 2603 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2604 | { |
cdd6c482 | 2605 | struct perf_event *event; |
57e7986e PM |
2606 | unsigned long flags; |
2607 | int enabled = 0; | |
889ff015 | 2608 | int ret; |
57e7986e PM |
2609 | |
2610 | local_irq_save(flags); | |
cdd6c482 | 2611 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2612 | goto out; |
2613 | ||
e566b76e SE |
2614 | /* |
2615 | * We must ctxsw out cgroup events to avoid conflict | |
2616 | * when invoking perf_task_event_sched_in() later on | |
2617 | * in this function. Otherwise we end up trying to | |
2618 | * ctxswin cgroup events which are already scheduled | |
2619 | * in. | |
2620 | */ | |
a8d757ef | 2621 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2622 | |
e625cce1 | 2623 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2624 | task_ctx_sched_out(ctx); |
57e7986e | 2625 | |
b79387ef | 2626 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2627 | ret = event_enable_on_exec(event, ctx); |
2628 | if (ret) | |
2629 | enabled = 1; | |
57e7986e PM |
2630 | } |
2631 | ||
2632 | /* | |
cdd6c482 | 2633 | * Unclone this context if we enabled any event. |
57e7986e | 2634 | */ |
71a851b4 PZ |
2635 | if (enabled) |
2636 | unclone_ctx(ctx); | |
57e7986e | 2637 | |
e625cce1 | 2638 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2639 | |
e566b76e SE |
2640 | /* |
2641 | * Also calls ctxswin for cgroup events, if any: | |
2642 | */ | |
e5d1367f | 2643 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2644 | out: |
57e7986e PM |
2645 | local_irq_restore(flags); |
2646 | } | |
2647 | ||
0793a61d | 2648 | /* |
cdd6c482 | 2649 | * Cross CPU call to read the hardware event |
0793a61d | 2650 | */ |
cdd6c482 | 2651 | static void __perf_event_read(void *info) |
0793a61d | 2652 | { |
cdd6c482 IM |
2653 | struct perf_event *event = info; |
2654 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2655 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2656 | |
e1ac3614 PM |
2657 | /* |
2658 | * If this is a task context, we need to check whether it is | |
2659 | * the current task context of this cpu. If not it has been | |
2660 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2661 | * event->count would have been updated to a recent sample |
2662 | * when the event was scheduled out. | |
e1ac3614 PM |
2663 | */ |
2664 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2665 | return; | |
2666 | ||
e625cce1 | 2667 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2668 | if (ctx->is_active) { |
542e72fc | 2669 | update_context_time(ctx); |
e5d1367f SE |
2670 | update_cgrp_time_from_event(event); |
2671 | } | |
cdd6c482 | 2672 | update_event_times(event); |
542e72fc PZ |
2673 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2674 | event->pmu->read(event); | |
e625cce1 | 2675 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2676 | } |
2677 | ||
b5e58793 PZ |
2678 | static inline u64 perf_event_count(struct perf_event *event) |
2679 | { | |
e7850595 | 2680 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
2681 | } |
2682 | ||
cdd6c482 | 2683 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
2684 | { |
2685 | /* | |
cdd6c482 IM |
2686 | * If event is enabled and currently active on a CPU, update the |
2687 | * value in the event structure: | |
0793a61d | 2688 | */ |
cdd6c482 IM |
2689 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2690 | smp_call_function_single(event->oncpu, | |
2691 | __perf_event_read, event, 1); | |
2692 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
2693 | struct perf_event_context *ctx = event->ctx; |
2694 | unsigned long flags; | |
2695 | ||
e625cce1 | 2696 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
2697 | /* |
2698 | * may read while context is not active | |
2699 | * (e.g., thread is blocked), in that case | |
2700 | * we cannot update context time | |
2701 | */ | |
e5d1367f | 2702 | if (ctx->is_active) { |
c530ccd9 | 2703 | update_context_time(ctx); |
e5d1367f SE |
2704 | update_cgrp_time_from_event(event); |
2705 | } | |
cdd6c482 | 2706 | update_event_times(event); |
e625cce1 | 2707 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
2708 | } |
2709 | ||
b5e58793 | 2710 | return perf_event_count(event); |
0793a61d TG |
2711 | } |
2712 | ||
a63eaf34 | 2713 | /* |
cdd6c482 | 2714 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 2715 | */ |
eb184479 | 2716 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 2717 | { |
e625cce1 | 2718 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 2719 | mutex_init(&ctx->mutex); |
889ff015 FW |
2720 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2721 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
2722 | INIT_LIST_HEAD(&ctx->event_list); |
2723 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
2724 | } |
2725 | ||
2726 | static struct perf_event_context * | |
2727 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
2728 | { | |
2729 | struct perf_event_context *ctx; | |
2730 | ||
2731 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
2732 | if (!ctx) | |
2733 | return NULL; | |
2734 | ||
2735 | __perf_event_init_context(ctx); | |
2736 | if (task) { | |
2737 | ctx->task = task; | |
2738 | get_task_struct(task); | |
0793a61d | 2739 | } |
eb184479 PZ |
2740 | ctx->pmu = pmu; |
2741 | ||
2742 | return ctx; | |
a63eaf34 PM |
2743 | } |
2744 | ||
2ebd4ffb MH |
2745 | static struct task_struct * |
2746 | find_lively_task_by_vpid(pid_t vpid) | |
2747 | { | |
2748 | struct task_struct *task; | |
2749 | int err; | |
0793a61d TG |
2750 | |
2751 | rcu_read_lock(); | |
2ebd4ffb | 2752 | if (!vpid) |
0793a61d TG |
2753 | task = current; |
2754 | else | |
2ebd4ffb | 2755 | task = find_task_by_vpid(vpid); |
0793a61d TG |
2756 | if (task) |
2757 | get_task_struct(task); | |
2758 | rcu_read_unlock(); | |
2759 | ||
2760 | if (!task) | |
2761 | return ERR_PTR(-ESRCH); | |
2762 | ||
0793a61d | 2763 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
2764 | err = -EACCES; |
2765 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
2766 | goto errout; | |
2767 | ||
2ebd4ffb MH |
2768 | return task; |
2769 | errout: | |
2770 | put_task_struct(task); | |
2771 | return ERR_PTR(err); | |
2772 | ||
2773 | } | |
2774 | ||
fe4b04fa PZ |
2775 | /* |
2776 | * Returns a matching context with refcount and pincount. | |
2777 | */ | |
108b02cf | 2778 | static struct perf_event_context * |
38a81da2 | 2779 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 2780 | { |
cdd6c482 | 2781 | struct perf_event_context *ctx; |
22a4f650 | 2782 | struct perf_cpu_context *cpuctx; |
25346b93 | 2783 | unsigned long flags; |
8dc85d54 | 2784 | int ctxn, err; |
0793a61d | 2785 | |
22a4ec72 | 2786 | if (!task) { |
cdd6c482 | 2787 | /* Must be root to operate on a CPU event: */ |
0764771d | 2788 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
2789 | return ERR_PTR(-EACCES); |
2790 | ||
0793a61d | 2791 | /* |
cdd6c482 | 2792 | * We could be clever and allow to attach a event to an |
0793a61d TG |
2793 | * offline CPU and activate it when the CPU comes up, but |
2794 | * that's for later. | |
2795 | */ | |
f6325e30 | 2796 | if (!cpu_online(cpu)) |
0793a61d TG |
2797 | return ERR_PTR(-ENODEV); |
2798 | ||
108b02cf | 2799 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 2800 | ctx = &cpuctx->ctx; |
c93f7669 | 2801 | get_ctx(ctx); |
fe4b04fa | 2802 | ++ctx->pin_count; |
0793a61d | 2803 | |
0793a61d TG |
2804 | return ctx; |
2805 | } | |
2806 | ||
8dc85d54 PZ |
2807 | err = -EINVAL; |
2808 | ctxn = pmu->task_ctx_nr; | |
2809 | if (ctxn < 0) | |
2810 | goto errout; | |
2811 | ||
9ed6060d | 2812 | retry: |
8dc85d54 | 2813 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 2814 | if (ctx) { |
71a851b4 | 2815 | unclone_ctx(ctx); |
fe4b04fa | 2816 | ++ctx->pin_count; |
e625cce1 | 2817 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 2818 | } else { |
eb184479 | 2819 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
2820 | err = -ENOMEM; |
2821 | if (!ctx) | |
2822 | goto errout; | |
eb184479 | 2823 | |
dbe08d82 ON |
2824 | err = 0; |
2825 | mutex_lock(&task->perf_event_mutex); | |
2826 | /* | |
2827 | * If it has already passed perf_event_exit_task(). | |
2828 | * we must see PF_EXITING, it takes this mutex too. | |
2829 | */ | |
2830 | if (task->flags & PF_EXITING) | |
2831 | err = -ESRCH; | |
2832 | else if (task->perf_event_ctxp[ctxn]) | |
2833 | err = -EAGAIN; | |
fe4b04fa | 2834 | else { |
9137fb28 | 2835 | get_ctx(ctx); |
fe4b04fa | 2836 | ++ctx->pin_count; |
dbe08d82 | 2837 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 2838 | } |
dbe08d82 ON |
2839 | mutex_unlock(&task->perf_event_mutex); |
2840 | ||
2841 | if (unlikely(err)) { | |
9137fb28 | 2842 | put_ctx(ctx); |
dbe08d82 ON |
2843 | |
2844 | if (err == -EAGAIN) | |
2845 | goto retry; | |
2846 | goto errout; | |
a63eaf34 PM |
2847 | } |
2848 | } | |
2849 | ||
0793a61d | 2850 | return ctx; |
c93f7669 | 2851 | |
9ed6060d | 2852 | errout: |
c93f7669 | 2853 | return ERR_PTR(err); |
0793a61d TG |
2854 | } |
2855 | ||
6fb2915d LZ |
2856 | static void perf_event_free_filter(struct perf_event *event); |
2857 | ||
cdd6c482 | 2858 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 2859 | { |
cdd6c482 | 2860 | struct perf_event *event; |
592903cd | 2861 | |
cdd6c482 IM |
2862 | event = container_of(head, struct perf_event, rcu_head); |
2863 | if (event->ns) | |
2864 | put_pid_ns(event->ns); | |
6fb2915d | 2865 | perf_event_free_filter(event); |
cdd6c482 | 2866 | kfree(event); |
592903cd PZ |
2867 | } |
2868 | ||
76369139 | 2869 | static void ring_buffer_put(struct ring_buffer *rb); |
925d519a | 2870 | |
cdd6c482 | 2871 | static void free_event(struct perf_event *event) |
f1600952 | 2872 | { |
e360adbe | 2873 | irq_work_sync(&event->pending); |
925d519a | 2874 | |
cdd6c482 | 2875 | if (!event->parent) { |
82cd6def | 2876 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 2877 | static_key_slow_dec_deferred(&perf_sched_events); |
3af9e859 | 2878 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
2879 | atomic_dec(&nr_mmap_events); |
2880 | if (event->attr.comm) | |
2881 | atomic_dec(&nr_comm_events); | |
2882 | if (event->attr.task) | |
2883 | atomic_dec(&nr_task_events); | |
927c7a9e FW |
2884 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
2885 | put_callchain_buffers(); | |
08309379 PZ |
2886 | if (is_cgroup_event(event)) { |
2887 | atomic_dec(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 2888 | static_key_slow_dec_deferred(&perf_sched_events); |
08309379 | 2889 | } |
d010b332 SE |
2890 | |
2891 | if (has_branch_stack(event)) { | |
2892 | static_key_slow_dec_deferred(&perf_sched_events); | |
2893 | /* is system-wide event */ | |
2894 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
2895 | atomic_dec(&per_cpu(perf_branch_stack_events, | |
2896 | event->cpu)); | |
2897 | } | |
f344011c | 2898 | } |
9ee318a7 | 2899 | |
76369139 FW |
2900 | if (event->rb) { |
2901 | ring_buffer_put(event->rb); | |
2902 | event->rb = NULL; | |
a4be7c27 PZ |
2903 | } |
2904 | ||
e5d1367f SE |
2905 | if (is_cgroup_event(event)) |
2906 | perf_detach_cgroup(event); | |
2907 | ||
cdd6c482 IM |
2908 | if (event->destroy) |
2909 | event->destroy(event); | |
e077df4f | 2910 | |
0c67b408 PZ |
2911 | if (event->ctx) |
2912 | put_ctx(event->ctx); | |
2913 | ||
cdd6c482 | 2914 | call_rcu(&event->rcu_head, free_event_rcu); |
f1600952 PZ |
2915 | } |
2916 | ||
a66a3052 | 2917 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 2918 | { |
cdd6c482 | 2919 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 2920 | |
ad3a37de | 2921 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
2922 | /* |
2923 | * There are two ways this annotation is useful: | |
2924 | * | |
2925 | * 1) there is a lock recursion from perf_event_exit_task | |
2926 | * see the comment there. | |
2927 | * | |
2928 | * 2) there is a lock-inversion with mmap_sem through | |
2929 | * perf_event_read_group(), which takes faults while | |
2930 | * holding ctx->mutex, however this is called after | |
2931 | * the last filedesc died, so there is no possibility | |
2932 | * to trigger the AB-BA case. | |
2933 | */ | |
2934 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
050735b0 | 2935 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 2936 | perf_group_detach(event); |
050735b0 | 2937 | raw_spin_unlock_irq(&ctx->lock); |
e03a9a55 | 2938 | perf_remove_from_context(event); |
d859e29f | 2939 | mutex_unlock(&ctx->mutex); |
0793a61d | 2940 | |
cdd6c482 | 2941 | free_event(event); |
0793a61d TG |
2942 | |
2943 | return 0; | |
2944 | } | |
a66a3052 | 2945 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 2946 | |
a66a3052 PZ |
2947 | /* |
2948 | * Called when the last reference to the file is gone. | |
2949 | */ | |
a6fa941d | 2950 | static void put_event(struct perf_event *event) |
fb0459d7 | 2951 | { |
8882135b | 2952 | struct task_struct *owner; |
fb0459d7 | 2953 | |
a6fa941d AV |
2954 | if (!atomic_long_dec_and_test(&event->refcount)) |
2955 | return; | |
fb0459d7 | 2956 | |
8882135b PZ |
2957 | rcu_read_lock(); |
2958 | owner = ACCESS_ONCE(event->owner); | |
2959 | /* | |
2960 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
2961 | * !owner it means the list deletion is complete and we can indeed | |
2962 | * free this event, otherwise we need to serialize on | |
2963 | * owner->perf_event_mutex. | |
2964 | */ | |
2965 | smp_read_barrier_depends(); | |
2966 | if (owner) { | |
2967 | /* | |
2968 | * Since delayed_put_task_struct() also drops the last | |
2969 | * task reference we can safely take a new reference | |
2970 | * while holding the rcu_read_lock(). | |
2971 | */ | |
2972 | get_task_struct(owner); | |
2973 | } | |
2974 | rcu_read_unlock(); | |
2975 | ||
2976 | if (owner) { | |
2977 | mutex_lock(&owner->perf_event_mutex); | |
2978 | /* | |
2979 | * We have to re-check the event->owner field, if it is cleared | |
2980 | * we raced with perf_event_exit_task(), acquiring the mutex | |
2981 | * ensured they're done, and we can proceed with freeing the | |
2982 | * event. | |
2983 | */ | |
2984 | if (event->owner) | |
2985 | list_del_init(&event->owner_entry); | |
2986 | mutex_unlock(&owner->perf_event_mutex); | |
2987 | put_task_struct(owner); | |
2988 | } | |
2989 | ||
a6fa941d AV |
2990 | perf_event_release_kernel(event); |
2991 | } | |
2992 | ||
2993 | static int perf_release(struct inode *inode, struct file *file) | |
2994 | { | |
2995 | put_event(file->private_data); | |
2996 | return 0; | |
fb0459d7 | 2997 | } |
fb0459d7 | 2998 | |
59ed446f | 2999 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3000 | { |
cdd6c482 | 3001 | struct perf_event *child; |
e53c0994 PZ |
3002 | u64 total = 0; |
3003 | ||
59ed446f PZ |
3004 | *enabled = 0; |
3005 | *running = 0; | |
3006 | ||
6f10581a | 3007 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3008 | total += perf_event_read(event); |
59ed446f PZ |
3009 | *enabled += event->total_time_enabled + |
3010 | atomic64_read(&event->child_total_time_enabled); | |
3011 | *running += event->total_time_running + | |
3012 | atomic64_read(&event->child_total_time_running); | |
3013 | ||
3014 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3015 | total += perf_event_read(child); |
59ed446f PZ |
3016 | *enabled += child->total_time_enabled; |
3017 | *running += child->total_time_running; | |
3018 | } | |
6f10581a | 3019 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3020 | |
3021 | return total; | |
3022 | } | |
fb0459d7 | 3023 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3024 | |
cdd6c482 | 3025 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3026 | u64 read_format, char __user *buf) |
3027 | { | |
cdd6c482 | 3028 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3029 | int n = 0, size = 0, ret = -EFAULT; |
3030 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3031 | u64 values[5]; |
59ed446f | 3032 | u64 count, enabled, running; |
abf4868b | 3033 | |
6f10581a | 3034 | mutex_lock(&ctx->mutex); |
59ed446f | 3035 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3036 | |
3037 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3038 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3039 | values[n++] = enabled; | |
3040 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3041 | values[n++] = running; | |
abf4868b PZ |
3042 | values[n++] = count; |
3043 | if (read_format & PERF_FORMAT_ID) | |
3044 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3045 | |
3046 | size = n * sizeof(u64); | |
3047 | ||
3048 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3049 | goto unlock; |
3dab77fb | 3050 | |
6f10581a | 3051 | ret = size; |
3dab77fb | 3052 | |
65abc865 | 3053 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3054 | n = 0; |
3dab77fb | 3055 | |
59ed446f | 3056 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3057 | if (read_format & PERF_FORMAT_ID) |
3058 | values[n++] = primary_event_id(sub); | |
3059 | ||
3060 | size = n * sizeof(u64); | |
3061 | ||
184d3da8 | 3062 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3063 | ret = -EFAULT; |
3064 | goto unlock; | |
3065 | } | |
abf4868b PZ |
3066 | |
3067 | ret += size; | |
3dab77fb | 3068 | } |
6f10581a PZ |
3069 | unlock: |
3070 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3071 | |
abf4868b | 3072 | return ret; |
3dab77fb PZ |
3073 | } |
3074 | ||
cdd6c482 | 3075 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3076 | u64 read_format, char __user *buf) |
3077 | { | |
59ed446f | 3078 | u64 enabled, running; |
3dab77fb PZ |
3079 | u64 values[4]; |
3080 | int n = 0; | |
3081 | ||
59ed446f PZ |
3082 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3083 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3084 | values[n++] = enabled; | |
3085 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3086 | values[n++] = running; | |
3dab77fb | 3087 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3088 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3089 | |
3090 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3091 | return -EFAULT; | |
3092 | ||
3093 | return n * sizeof(u64); | |
3094 | } | |
3095 | ||
0793a61d | 3096 | /* |
cdd6c482 | 3097 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3098 | */ |
3099 | static ssize_t | |
cdd6c482 | 3100 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3101 | { |
cdd6c482 | 3102 | u64 read_format = event->attr.read_format; |
3dab77fb | 3103 | int ret; |
0793a61d | 3104 | |
3b6f9e5c | 3105 | /* |
cdd6c482 | 3106 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3107 | * error state (i.e. because it was pinned but it couldn't be |
3108 | * scheduled on to the CPU at some point). | |
3109 | */ | |
cdd6c482 | 3110 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3111 | return 0; |
3112 | ||
c320c7b7 | 3113 | if (count < event->read_size) |
3dab77fb PZ |
3114 | return -ENOSPC; |
3115 | ||
cdd6c482 | 3116 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3117 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3118 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3119 | else |
cdd6c482 | 3120 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3121 | |
3dab77fb | 3122 | return ret; |
0793a61d TG |
3123 | } |
3124 | ||
0793a61d TG |
3125 | static ssize_t |
3126 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3127 | { | |
cdd6c482 | 3128 | struct perf_event *event = file->private_data; |
0793a61d | 3129 | |
cdd6c482 | 3130 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3131 | } |
3132 | ||
3133 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3134 | { | |
cdd6c482 | 3135 | struct perf_event *event = file->private_data; |
76369139 | 3136 | struct ring_buffer *rb; |
c33a0bc4 | 3137 | unsigned int events = POLL_HUP; |
c7138f37 | 3138 | |
10c6db11 PZ |
3139 | /* |
3140 | * Race between perf_event_set_output() and perf_poll(): perf_poll() | |
3141 | * grabs the rb reference but perf_event_set_output() overrides it. | |
3142 | * Here is the timeline for two threads T1, T2: | |
3143 | * t0: T1, rb = rcu_dereference(event->rb) | |
3144 | * t1: T2, old_rb = event->rb | |
3145 | * t2: T2, event->rb = new rb | |
3146 | * t3: T2, ring_buffer_detach(old_rb) | |
3147 | * t4: T1, ring_buffer_attach(rb1) | |
3148 | * t5: T1, poll_wait(event->waitq) | |
3149 | * | |
3150 | * To avoid this problem, we grab mmap_mutex in perf_poll() | |
3151 | * thereby ensuring that the assignment of the new ring buffer | |
3152 | * and the detachment of the old buffer appear atomic to perf_poll() | |
3153 | */ | |
3154 | mutex_lock(&event->mmap_mutex); | |
3155 | ||
c7138f37 | 3156 | rcu_read_lock(); |
76369139 | 3157 | rb = rcu_dereference(event->rb); |
10c6db11 PZ |
3158 | if (rb) { |
3159 | ring_buffer_attach(event, rb); | |
76369139 | 3160 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3161 | } |
c7138f37 | 3162 | rcu_read_unlock(); |
0793a61d | 3163 | |
10c6db11 PZ |
3164 | mutex_unlock(&event->mmap_mutex); |
3165 | ||
cdd6c482 | 3166 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3167 | |
0793a61d TG |
3168 | return events; |
3169 | } | |
3170 | ||
cdd6c482 | 3171 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3172 | { |
cdd6c482 | 3173 | (void)perf_event_read(event); |
e7850595 | 3174 | local64_set(&event->count, 0); |
cdd6c482 | 3175 | perf_event_update_userpage(event); |
3df5edad PZ |
3176 | } |
3177 | ||
c93f7669 | 3178 | /* |
cdd6c482 IM |
3179 | * Holding the top-level event's child_mutex means that any |
3180 | * descendant process that has inherited this event will block | |
3181 | * in sync_child_event if it goes to exit, thus satisfying the | |
3182 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3183 | */ |
cdd6c482 IM |
3184 | static void perf_event_for_each_child(struct perf_event *event, |
3185 | void (*func)(struct perf_event *)) | |
3df5edad | 3186 | { |
cdd6c482 | 3187 | struct perf_event *child; |
3df5edad | 3188 | |
cdd6c482 IM |
3189 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3190 | mutex_lock(&event->child_mutex); | |
3191 | func(event); | |
3192 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3193 | func(child); |
cdd6c482 | 3194 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3195 | } |
3196 | ||
cdd6c482 IM |
3197 | static void perf_event_for_each(struct perf_event *event, |
3198 | void (*func)(struct perf_event *)) | |
3df5edad | 3199 | { |
cdd6c482 IM |
3200 | struct perf_event_context *ctx = event->ctx; |
3201 | struct perf_event *sibling; | |
3df5edad | 3202 | |
75f937f2 PZ |
3203 | WARN_ON_ONCE(ctx->parent_ctx); |
3204 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3205 | event = event->group_leader; |
75f937f2 | 3206 | |
cdd6c482 | 3207 | perf_event_for_each_child(event, func); |
cdd6c482 | 3208 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3209 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3210 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3211 | } |
3212 | ||
cdd6c482 | 3213 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3214 | { |
cdd6c482 | 3215 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3216 | int ret = 0; |
3217 | u64 value; | |
3218 | ||
6c7e550f | 3219 | if (!is_sampling_event(event)) |
08247e31 PZ |
3220 | return -EINVAL; |
3221 | ||
ad0cf347 | 3222 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3223 | return -EFAULT; |
3224 | ||
3225 | if (!value) | |
3226 | return -EINVAL; | |
3227 | ||
e625cce1 | 3228 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3229 | if (event->attr.freq) { |
3230 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3231 | ret = -EINVAL; |
3232 | goto unlock; | |
3233 | } | |
3234 | ||
cdd6c482 | 3235 | event->attr.sample_freq = value; |
08247e31 | 3236 | } else { |
cdd6c482 IM |
3237 | event->attr.sample_period = value; |
3238 | event->hw.sample_period = value; | |
08247e31 PZ |
3239 | } |
3240 | unlock: | |
e625cce1 | 3241 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3242 | |
3243 | return ret; | |
3244 | } | |
3245 | ||
ac9721f3 PZ |
3246 | static const struct file_operations perf_fops; |
3247 | ||
2903ff01 | 3248 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3249 | { |
2903ff01 AV |
3250 | struct fd f = fdget(fd); |
3251 | if (!f.file) | |
3252 | return -EBADF; | |
ac9721f3 | 3253 | |
2903ff01 AV |
3254 | if (f.file->f_op != &perf_fops) { |
3255 | fdput(f); | |
3256 | return -EBADF; | |
ac9721f3 | 3257 | } |
2903ff01 AV |
3258 | *p = f; |
3259 | return 0; | |
ac9721f3 PZ |
3260 | } |
3261 | ||
3262 | static int perf_event_set_output(struct perf_event *event, | |
3263 | struct perf_event *output_event); | |
6fb2915d | 3264 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3265 | |
d859e29f PM |
3266 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3267 | { | |
cdd6c482 IM |
3268 | struct perf_event *event = file->private_data; |
3269 | void (*func)(struct perf_event *); | |
3df5edad | 3270 | u32 flags = arg; |
d859e29f PM |
3271 | |
3272 | switch (cmd) { | |
cdd6c482 IM |
3273 | case PERF_EVENT_IOC_ENABLE: |
3274 | func = perf_event_enable; | |
d859e29f | 3275 | break; |
cdd6c482 IM |
3276 | case PERF_EVENT_IOC_DISABLE: |
3277 | func = perf_event_disable; | |
79f14641 | 3278 | break; |
cdd6c482 IM |
3279 | case PERF_EVENT_IOC_RESET: |
3280 | func = perf_event_reset; | |
6de6a7b9 | 3281 | break; |
3df5edad | 3282 | |
cdd6c482 IM |
3283 | case PERF_EVENT_IOC_REFRESH: |
3284 | return perf_event_refresh(event, arg); | |
08247e31 | 3285 | |
cdd6c482 IM |
3286 | case PERF_EVENT_IOC_PERIOD: |
3287 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3288 | |
cdd6c482 | 3289 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3290 | { |
ac9721f3 | 3291 | int ret; |
ac9721f3 | 3292 | if (arg != -1) { |
2903ff01 AV |
3293 | struct perf_event *output_event; |
3294 | struct fd output; | |
3295 | ret = perf_fget_light(arg, &output); | |
3296 | if (ret) | |
3297 | return ret; | |
3298 | output_event = output.file->private_data; | |
3299 | ret = perf_event_set_output(event, output_event); | |
3300 | fdput(output); | |
3301 | } else { | |
3302 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3303 | } |
ac9721f3 PZ |
3304 | return ret; |
3305 | } | |
a4be7c27 | 3306 | |
6fb2915d LZ |
3307 | case PERF_EVENT_IOC_SET_FILTER: |
3308 | return perf_event_set_filter(event, (void __user *)arg); | |
3309 | ||
d859e29f | 3310 | default: |
3df5edad | 3311 | return -ENOTTY; |
d859e29f | 3312 | } |
3df5edad PZ |
3313 | |
3314 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3315 | perf_event_for_each(event, func); |
3df5edad | 3316 | else |
cdd6c482 | 3317 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3318 | |
3319 | return 0; | |
d859e29f PM |
3320 | } |
3321 | ||
cdd6c482 | 3322 | int perf_event_task_enable(void) |
771d7cde | 3323 | { |
cdd6c482 | 3324 | struct perf_event *event; |
771d7cde | 3325 | |
cdd6c482 IM |
3326 | mutex_lock(¤t->perf_event_mutex); |
3327 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3328 | perf_event_for_each_child(event, perf_event_enable); | |
3329 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3330 | |
3331 | return 0; | |
3332 | } | |
3333 | ||
cdd6c482 | 3334 | int perf_event_task_disable(void) |
771d7cde | 3335 | { |
cdd6c482 | 3336 | struct perf_event *event; |
771d7cde | 3337 | |
cdd6c482 IM |
3338 | mutex_lock(¤t->perf_event_mutex); |
3339 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3340 | perf_event_for_each_child(event, perf_event_disable); | |
3341 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3342 | |
3343 | return 0; | |
3344 | } | |
3345 | ||
cdd6c482 | 3346 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3347 | { |
a4eaf7f1 PZ |
3348 | if (event->hw.state & PERF_HES_STOPPED) |
3349 | return 0; | |
3350 | ||
cdd6c482 | 3351 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3352 | return 0; |
3353 | ||
35edc2a5 | 3354 | return event->pmu->event_idx(event); |
194002b2 PZ |
3355 | } |
3356 | ||
c4794295 | 3357 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3358 | u64 *now, |
7f310a5d EM |
3359 | u64 *enabled, |
3360 | u64 *running) | |
c4794295 | 3361 | { |
e3f3541c | 3362 | u64 ctx_time; |
c4794295 | 3363 | |
e3f3541c PZ |
3364 | *now = perf_clock(); |
3365 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3366 | *enabled = ctx_time - event->tstamp_enabled; |
3367 | *running = ctx_time - event->tstamp_running; | |
3368 | } | |
3369 | ||
c7206205 | 3370 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3371 | { |
3372 | } | |
3373 | ||
38ff667b PZ |
3374 | /* |
3375 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3376 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3377 | * code calls this from NMI context. | |
3378 | */ | |
cdd6c482 | 3379 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3380 | { |
cdd6c482 | 3381 | struct perf_event_mmap_page *userpg; |
76369139 | 3382 | struct ring_buffer *rb; |
e3f3541c | 3383 | u64 enabled, running, now; |
38ff667b PZ |
3384 | |
3385 | rcu_read_lock(); | |
0d641208 EM |
3386 | /* |
3387 | * compute total_time_enabled, total_time_running | |
3388 | * based on snapshot values taken when the event | |
3389 | * was last scheduled in. | |
3390 | * | |
3391 | * we cannot simply called update_context_time() | |
3392 | * because of locking issue as we can be called in | |
3393 | * NMI context | |
3394 | */ | |
e3f3541c | 3395 | calc_timer_values(event, &now, &enabled, &running); |
76369139 FW |
3396 | rb = rcu_dereference(event->rb); |
3397 | if (!rb) | |
38ff667b PZ |
3398 | goto unlock; |
3399 | ||
76369139 | 3400 | userpg = rb->user_page; |
37d81828 | 3401 | |
7b732a75 PZ |
3402 | /* |
3403 | * Disable preemption so as to not let the corresponding user-space | |
3404 | * spin too long if we get preempted. | |
3405 | */ | |
3406 | preempt_disable(); | |
37d81828 | 3407 | ++userpg->lock; |
92f22a38 | 3408 | barrier(); |
cdd6c482 | 3409 | userpg->index = perf_event_index(event); |
b5e58793 | 3410 | userpg->offset = perf_event_count(event); |
365a4038 | 3411 | if (userpg->index) |
e7850595 | 3412 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3413 | |
0d641208 | 3414 | userpg->time_enabled = enabled + |
cdd6c482 | 3415 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3416 | |
0d641208 | 3417 | userpg->time_running = running + |
cdd6c482 | 3418 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3419 | |
c7206205 | 3420 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3421 | |
92f22a38 | 3422 | barrier(); |
37d81828 | 3423 | ++userpg->lock; |
7b732a75 | 3424 | preempt_enable(); |
38ff667b | 3425 | unlock: |
7b732a75 | 3426 | rcu_read_unlock(); |
37d81828 PM |
3427 | } |
3428 | ||
906010b2 PZ |
3429 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3430 | { | |
3431 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3432 | struct ring_buffer *rb; |
906010b2 PZ |
3433 | int ret = VM_FAULT_SIGBUS; |
3434 | ||
3435 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3436 | if (vmf->pgoff == 0) | |
3437 | ret = 0; | |
3438 | return ret; | |
3439 | } | |
3440 | ||
3441 | rcu_read_lock(); | |
76369139 FW |
3442 | rb = rcu_dereference(event->rb); |
3443 | if (!rb) | |
906010b2 PZ |
3444 | goto unlock; |
3445 | ||
3446 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3447 | goto unlock; | |
3448 | ||
76369139 | 3449 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3450 | if (!vmf->page) |
3451 | goto unlock; | |
3452 | ||
3453 | get_page(vmf->page); | |
3454 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3455 | vmf->page->index = vmf->pgoff; | |
3456 | ||
3457 | ret = 0; | |
3458 | unlock: | |
3459 | rcu_read_unlock(); | |
3460 | ||
3461 | return ret; | |
3462 | } | |
3463 | ||
10c6db11 PZ |
3464 | static void ring_buffer_attach(struct perf_event *event, |
3465 | struct ring_buffer *rb) | |
3466 | { | |
3467 | unsigned long flags; | |
3468 | ||
3469 | if (!list_empty(&event->rb_entry)) | |
3470 | return; | |
3471 | ||
3472 | spin_lock_irqsave(&rb->event_lock, flags); | |
3473 | if (!list_empty(&event->rb_entry)) | |
3474 | goto unlock; | |
3475 | ||
3476 | list_add(&event->rb_entry, &rb->event_list); | |
3477 | unlock: | |
3478 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3479 | } | |
3480 | ||
3481 | static void ring_buffer_detach(struct perf_event *event, | |
3482 | struct ring_buffer *rb) | |
3483 | { | |
3484 | unsigned long flags; | |
3485 | ||
3486 | if (list_empty(&event->rb_entry)) | |
3487 | return; | |
3488 | ||
3489 | spin_lock_irqsave(&rb->event_lock, flags); | |
3490 | list_del_init(&event->rb_entry); | |
3491 | wake_up_all(&event->waitq); | |
3492 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3493 | } | |
3494 | ||
3495 | static void ring_buffer_wakeup(struct perf_event *event) | |
3496 | { | |
3497 | struct ring_buffer *rb; | |
3498 | ||
3499 | rcu_read_lock(); | |
3500 | rb = rcu_dereference(event->rb); | |
44b7f4b9 WD |
3501 | if (!rb) |
3502 | goto unlock; | |
3503 | ||
3504 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
10c6db11 | 3505 | wake_up_all(&event->waitq); |
44b7f4b9 WD |
3506 | |
3507 | unlock: | |
10c6db11 PZ |
3508 | rcu_read_unlock(); |
3509 | } | |
3510 | ||
76369139 | 3511 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3512 | { |
76369139 | 3513 | struct ring_buffer *rb; |
906010b2 | 3514 | |
76369139 FW |
3515 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3516 | rb_free(rb); | |
7b732a75 PZ |
3517 | } |
3518 | ||
76369139 | 3519 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3520 | { |
76369139 | 3521 | struct ring_buffer *rb; |
7b732a75 | 3522 | |
ac9721f3 | 3523 | rcu_read_lock(); |
76369139 FW |
3524 | rb = rcu_dereference(event->rb); |
3525 | if (rb) { | |
3526 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3527 | rb = NULL; | |
ac9721f3 PZ |
3528 | } |
3529 | rcu_read_unlock(); | |
3530 | ||
76369139 | 3531 | return rb; |
ac9721f3 PZ |
3532 | } |
3533 | ||
76369139 | 3534 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3535 | { |
10c6db11 PZ |
3536 | struct perf_event *event, *n; |
3537 | unsigned long flags; | |
3538 | ||
76369139 | 3539 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3540 | return; |
7b732a75 | 3541 | |
10c6db11 PZ |
3542 | spin_lock_irqsave(&rb->event_lock, flags); |
3543 | list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) { | |
3544 | list_del_init(&event->rb_entry); | |
3545 | wake_up_all(&event->waitq); | |
3546 | } | |
3547 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3548 | ||
76369139 | 3549 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3550 | } |
3551 | ||
3552 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3553 | { | |
cdd6c482 | 3554 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3555 | |
cdd6c482 | 3556 | atomic_inc(&event->mmap_count); |
7b732a75 PZ |
3557 | } |
3558 | ||
3559 | static void perf_mmap_close(struct vm_area_struct *vma) | |
3560 | { | |
cdd6c482 | 3561 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3562 | |
cdd6c482 | 3563 | if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) { |
76369139 | 3564 | unsigned long size = perf_data_size(event->rb); |
ac9721f3 | 3565 | struct user_struct *user = event->mmap_user; |
76369139 | 3566 | struct ring_buffer *rb = event->rb; |
789f90fc | 3567 | |
906010b2 | 3568 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm); |
bc3e53f6 | 3569 | vma->vm_mm->pinned_vm -= event->mmap_locked; |
76369139 | 3570 | rcu_assign_pointer(event->rb, NULL); |
10c6db11 | 3571 | ring_buffer_detach(event, rb); |
cdd6c482 | 3572 | mutex_unlock(&event->mmap_mutex); |
ac9721f3 | 3573 | |
76369139 | 3574 | ring_buffer_put(rb); |
ac9721f3 | 3575 | free_uid(user); |
7b732a75 | 3576 | } |
37d81828 PM |
3577 | } |
3578 | ||
f0f37e2f | 3579 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3580 | .open = perf_mmap_open, |
3581 | .close = perf_mmap_close, | |
3582 | .fault = perf_mmap_fault, | |
3583 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3584 | }; |
3585 | ||
3586 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3587 | { | |
cdd6c482 | 3588 | struct perf_event *event = file->private_data; |
22a4f650 | 3589 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3590 | struct user_struct *user = current_user(); |
22a4f650 | 3591 | unsigned long locked, lock_limit; |
76369139 | 3592 | struct ring_buffer *rb; |
7b732a75 PZ |
3593 | unsigned long vma_size; |
3594 | unsigned long nr_pages; | |
789f90fc | 3595 | long user_extra, extra; |
d57e34fd | 3596 | int ret = 0, flags = 0; |
37d81828 | 3597 | |
c7920614 PZ |
3598 | /* |
3599 | * Don't allow mmap() of inherited per-task counters. This would | |
3600 | * create a performance issue due to all children writing to the | |
76369139 | 3601 | * same rb. |
c7920614 PZ |
3602 | */ |
3603 | if (event->cpu == -1 && event->attr.inherit) | |
3604 | return -EINVAL; | |
3605 | ||
43a21ea8 | 3606 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3607 | return -EINVAL; |
7b732a75 PZ |
3608 | |
3609 | vma_size = vma->vm_end - vma->vm_start; | |
3610 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3611 | ||
7730d865 | 3612 | /* |
76369139 | 3613 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3614 | * can do bitmasks instead of modulo. |
3615 | */ | |
3616 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3617 | return -EINVAL; |
3618 | ||
7b732a75 | 3619 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3620 | return -EINVAL; |
3621 | ||
7b732a75 PZ |
3622 | if (vma->vm_pgoff != 0) |
3623 | return -EINVAL; | |
37d81828 | 3624 | |
cdd6c482 IM |
3625 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3626 | mutex_lock(&event->mmap_mutex); | |
76369139 FW |
3627 | if (event->rb) { |
3628 | if (event->rb->nr_pages == nr_pages) | |
3629 | atomic_inc(&event->rb->refcount); | |
ac9721f3 | 3630 | else |
ebb3c4c4 PZ |
3631 | ret = -EINVAL; |
3632 | goto unlock; | |
3633 | } | |
3634 | ||
789f90fc | 3635 | user_extra = nr_pages + 1; |
cdd6c482 | 3636 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
3637 | |
3638 | /* | |
3639 | * Increase the limit linearly with more CPUs: | |
3640 | */ | |
3641 | user_lock_limit *= num_online_cpus(); | |
3642 | ||
789f90fc | 3643 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 3644 | |
789f90fc PZ |
3645 | extra = 0; |
3646 | if (user_locked > user_lock_limit) | |
3647 | extra = user_locked - user_lock_limit; | |
7b732a75 | 3648 | |
78d7d407 | 3649 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 3650 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 3651 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 3652 | |
459ec28a IM |
3653 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
3654 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
3655 | ret = -EPERM; |
3656 | goto unlock; | |
3657 | } | |
7b732a75 | 3658 | |
76369139 | 3659 | WARN_ON(event->rb); |
906010b2 | 3660 | |
d57e34fd | 3661 | if (vma->vm_flags & VM_WRITE) |
76369139 | 3662 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 3663 | |
4ec8363d VW |
3664 | rb = rb_alloc(nr_pages, |
3665 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
3666 | event->cpu, flags); | |
3667 | ||
76369139 | 3668 | if (!rb) { |
ac9721f3 | 3669 | ret = -ENOMEM; |
ebb3c4c4 | 3670 | goto unlock; |
ac9721f3 | 3671 | } |
76369139 | 3672 | rcu_assign_pointer(event->rb, rb); |
43a21ea8 | 3673 | |
ac9721f3 PZ |
3674 | atomic_long_add(user_extra, &user->locked_vm); |
3675 | event->mmap_locked = extra; | |
3676 | event->mmap_user = get_current_user(); | |
bc3e53f6 | 3677 | vma->vm_mm->pinned_vm += event->mmap_locked; |
ac9721f3 | 3678 | |
9a0f05cb PZ |
3679 | perf_event_update_userpage(event); |
3680 | ||
ebb3c4c4 | 3681 | unlock: |
ac9721f3 PZ |
3682 | if (!ret) |
3683 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 3684 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 3685 | |
314e51b9 | 3686 | vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 3687 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
3688 | |
3689 | return ret; | |
37d81828 PM |
3690 | } |
3691 | ||
3c446b3d PZ |
3692 | static int perf_fasync(int fd, struct file *filp, int on) |
3693 | { | |
496ad9aa | 3694 | struct inode *inode = file_inode(filp); |
cdd6c482 | 3695 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
3696 | int retval; |
3697 | ||
3698 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 3699 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
3700 | mutex_unlock(&inode->i_mutex); |
3701 | ||
3702 | if (retval < 0) | |
3703 | return retval; | |
3704 | ||
3705 | return 0; | |
3706 | } | |
3707 | ||
0793a61d | 3708 | static const struct file_operations perf_fops = { |
3326c1ce | 3709 | .llseek = no_llseek, |
0793a61d TG |
3710 | .release = perf_release, |
3711 | .read = perf_read, | |
3712 | .poll = perf_poll, | |
d859e29f PM |
3713 | .unlocked_ioctl = perf_ioctl, |
3714 | .compat_ioctl = perf_ioctl, | |
37d81828 | 3715 | .mmap = perf_mmap, |
3c446b3d | 3716 | .fasync = perf_fasync, |
0793a61d TG |
3717 | }; |
3718 | ||
925d519a | 3719 | /* |
cdd6c482 | 3720 | * Perf event wakeup |
925d519a PZ |
3721 | * |
3722 | * If there's data, ensure we set the poll() state and publish everything | |
3723 | * to user-space before waking everybody up. | |
3724 | */ | |
3725 | ||
cdd6c482 | 3726 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 3727 | { |
10c6db11 | 3728 | ring_buffer_wakeup(event); |
4c9e2542 | 3729 | |
cdd6c482 IM |
3730 | if (event->pending_kill) { |
3731 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
3732 | event->pending_kill = 0; | |
4c9e2542 | 3733 | } |
925d519a PZ |
3734 | } |
3735 | ||
e360adbe | 3736 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 3737 | { |
cdd6c482 IM |
3738 | struct perf_event *event = container_of(entry, |
3739 | struct perf_event, pending); | |
79f14641 | 3740 | |
cdd6c482 IM |
3741 | if (event->pending_disable) { |
3742 | event->pending_disable = 0; | |
3743 | __perf_event_disable(event); | |
79f14641 PZ |
3744 | } |
3745 | ||
cdd6c482 IM |
3746 | if (event->pending_wakeup) { |
3747 | event->pending_wakeup = 0; | |
3748 | perf_event_wakeup(event); | |
79f14641 PZ |
3749 | } |
3750 | } | |
3751 | ||
39447b38 ZY |
3752 | /* |
3753 | * We assume there is only KVM supporting the callbacks. | |
3754 | * Later on, we might change it to a list if there is | |
3755 | * another virtualization implementation supporting the callbacks. | |
3756 | */ | |
3757 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
3758 | ||
3759 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3760 | { | |
3761 | perf_guest_cbs = cbs; | |
3762 | return 0; | |
3763 | } | |
3764 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
3765 | ||
3766 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3767 | { | |
3768 | perf_guest_cbs = NULL; | |
3769 | return 0; | |
3770 | } | |
3771 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
3772 | ||
4018994f JO |
3773 | static void |
3774 | perf_output_sample_regs(struct perf_output_handle *handle, | |
3775 | struct pt_regs *regs, u64 mask) | |
3776 | { | |
3777 | int bit; | |
3778 | ||
3779 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
3780 | sizeof(mask) * BITS_PER_BYTE) { | |
3781 | u64 val; | |
3782 | ||
3783 | val = perf_reg_value(regs, bit); | |
3784 | perf_output_put(handle, val); | |
3785 | } | |
3786 | } | |
3787 | ||
3788 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
3789 | struct pt_regs *regs) | |
3790 | { | |
3791 | if (!user_mode(regs)) { | |
3792 | if (current->mm) | |
3793 | regs = task_pt_regs(current); | |
3794 | else | |
3795 | regs = NULL; | |
3796 | } | |
3797 | ||
3798 | if (regs) { | |
3799 | regs_user->regs = regs; | |
3800 | regs_user->abi = perf_reg_abi(current); | |
3801 | } | |
3802 | } | |
3803 | ||
c5ebcedb JO |
3804 | /* |
3805 | * Get remaining task size from user stack pointer. | |
3806 | * | |
3807 | * It'd be better to take stack vma map and limit this more | |
3808 | * precisly, but there's no way to get it safely under interrupt, | |
3809 | * so using TASK_SIZE as limit. | |
3810 | */ | |
3811 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
3812 | { | |
3813 | unsigned long addr = perf_user_stack_pointer(regs); | |
3814 | ||
3815 | if (!addr || addr >= TASK_SIZE) | |
3816 | return 0; | |
3817 | ||
3818 | return TASK_SIZE - addr; | |
3819 | } | |
3820 | ||
3821 | static u16 | |
3822 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
3823 | struct pt_regs *regs) | |
3824 | { | |
3825 | u64 task_size; | |
3826 | ||
3827 | /* No regs, no stack pointer, no dump. */ | |
3828 | if (!regs) | |
3829 | return 0; | |
3830 | ||
3831 | /* | |
3832 | * Check if we fit in with the requested stack size into the: | |
3833 | * - TASK_SIZE | |
3834 | * If we don't, we limit the size to the TASK_SIZE. | |
3835 | * | |
3836 | * - remaining sample size | |
3837 | * If we don't, we customize the stack size to | |
3838 | * fit in to the remaining sample size. | |
3839 | */ | |
3840 | ||
3841 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
3842 | stack_size = min(stack_size, (u16) task_size); | |
3843 | ||
3844 | /* Current header size plus static size and dynamic size. */ | |
3845 | header_size += 2 * sizeof(u64); | |
3846 | ||
3847 | /* Do we fit in with the current stack dump size? */ | |
3848 | if ((u16) (header_size + stack_size) < header_size) { | |
3849 | /* | |
3850 | * If we overflow the maximum size for the sample, | |
3851 | * we customize the stack dump size to fit in. | |
3852 | */ | |
3853 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
3854 | stack_size = round_up(stack_size, sizeof(u64)); | |
3855 | } | |
3856 | ||
3857 | return stack_size; | |
3858 | } | |
3859 | ||
3860 | static void | |
3861 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
3862 | struct pt_regs *regs) | |
3863 | { | |
3864 | /* Case of a kernel thread, nothing to dump */ | |
3865 | if (!regs) { | |
3866 | u64 size = 0; | |
3867 | perf_output_put(handle, size); | |
3868 | } else { | |
3869 | unsigned long sp; | |
3870 | unsigned int rem; | |
3871 | u64 dyn_size; | |
3872 | ||
3873 | /* | |
3874 | * We dump: | |
3875 | * static size | |
3876 | * - the size requested by user or the best one we can fit | |
3877 | * in to the sample max size | |
3878 | * data | |
3879 | * - user stack dump data | |
3880 | * dynamic size | |
3881 | * - the actual dumped size | |
3882 | */ | |
3883 | ||
3884 | /* Static size. */ | |
3885 | perf_output_put(handle, dump_size); | |
3886 | ||
3887 | /* Data. */ | |
3888 | sp = perf_user_stack_pointer(regs); | |
3889 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
3890 | dyn_size = dump_size - rem; | |
3891 | ||
3892 | perf_output_skip(handle, rem); | |
3893 | ||
3894 | /* Dynamic size. */ | |
3895 | perf_output_put(handle, dyn_size); | |
3896 | } | |
3897 | } | |
3898 | ||
c980d109 ACM |
3899 | static void __perf_event_header__init_id(struct perf_event_header *header, |
3900 | struct perf_sample_data *data, | |
3901 | struct perf_event *event) | |
6844c09d ACM |
3902 | { |
3903 | u64 sample_type = event->attr.sample_type; | |
3904 | ||
3905 | data->type = sample_type; | |
3906 | header->size += event->id_header_size; | |
3907 | ||
3908 | if (sample_type & PERF_SAMPLE_TID) { | |
3909 | /* namespace issues */ | |
3910 | data->tid_entry.pid = perf_event_pid(event, current); | |
3911 | data->tid_entry.tid = perf_event_tid(event, current); | |
3912 | } | |
3913 | ||
3914 | if (sample_type & PERF_SAMPLE_TIME) | |
3915 | data->time = perf_clock(); | |
3916 | ||
3917 | if (sample_type & PERF_SAMPLE_ID) | |
3918 | data->id = primary_event_id(event); | |
3919 | ||
3920 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3921 | data->stream_id = event->id; | |
3922 | ||
3923 | if (sample_type & PERF_SAMPLE_CPU) { | |
3924 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
3925 | data->cpu_entry.reserved = 0; | |
3926 | } | |
3927 | } | |
3928 | ||
76369139 FW |
3929 | void perf_event_header__init_id(struct perf_event_header *header, |
3930 | struct perf_sample_data *data, | |
3931 | struct perf_event *event) | |
c980d109 ACM |
3932 | { |
3933 | if (event->attr.sample_id_all) | |
3934 | __perf_event_header__init_id(header, data, event); | |
3935 | } | |
3936 | ||
3937 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
3938 | struct perf_sample_data *data) | |
3939 | { | |
3940 | u64 sample_type = data->type; | |
3941 | ||
3942 | if (sample_type & PERF_SAMPLE_TID) | |
3943 | perf_output_put(handle, data->tid_entry); | |
3944 | ||
3945 | if (sample_type & PERF_SAMPLE_TIME) | |
3946 | perf_output_put(handle, data->time); | |
3947 | ||
3948 | if (sample_type & PERF_SAMPLE_ID) | |
3949 | perf_output_put(handle, data->id); | |
3950 | ||
3951 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
3952 | perf_output_put(handle, data->stream_id); | |
3953 | ||
3954 | if (sample_type & PERF_SAMPLE_CPU) | |
3955 | perf_output_put(handle, data->cpu_entry); | |
3956 | } | |
3957 | ||
76369139 FW |
3958 | void perf_event__output_id_sample(struct perf_event *event, |
3959 | struct perf_output_handle *handle, | |
3960 | struct perf_sample_data *sample) | |
c980d109 ACM |
3961 | { |
3962 | if (event->attr.sample_id_all) | |
3963 | __perf_event__output_id_sample(handle, sample); | |
3964 | } | |
3965 | ||
3dab77fb | 3966 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
3967 | struct perf_event *event, |
3968 | u64 enabled, u64 running) | |
3dab77fb | 3969 | { |
cdd6c482 | 3970 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
3971 | u64 values[4]; |
3972 | int n = 0; | |
3973 | ||
b5e58793 | 3974 | values[n++] = perf_event_count(event); |
3dab77fb | 3975 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 3976 | values[n++] = enabled + |
cdd6c482 | 3977 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
3978 | } |
3979 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 3980 | values[n++] = running + |
cdd6c482 | 3981 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
3982 | } |
3983 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 3984 | values[n++] = primary_event_id(event); |
3dab77fb | 3985 | |
76369139 | 3986 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
3987 | } |
3988 | ||
3989 | /* | |
cdd6c482 | 3990 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
3991 | */ |
3992 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
3993 | struct perf_event *event, |
3994 | u64 enabled, u64 running) | |
3dab77fb | 3995 | { |
cdd6c482 IM |
3996 | struct perf_event *leader = event->group_leader, *sub; |
3997 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
3998 | u64 values[5]; |
3999 | int n = 0; | |
4000 | ||
4001 | values[n++] = 1 + leader->nr_siblings; | |
4002 | ||
4003 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4004 | values[n++] = enabled; |
3dab77fb PZ |
4005 | |
4006 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4007 | values[n++] = running; |
3dab77fb | 4008 | |
cdd6c482 | 4009 | if (leader != event) |
3dab77fb PZ |
4010 | leader->pmu->read(leader); |
4011 | ||
b5e58793 | 4012 | values[n++] = perf_event_count(leader); |
3dab77fb | 4013 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4014 | values[n++] = primary_event_id(leader); |
3dab77fb | 4015 | |
76369139 | 4016 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4017 | |
65abc865 | 4018 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4019 | n = 0; |
4020 | ||
cdd6c482 | 4021 | if (sub != event) |
3dab77fb PZ |
4022 | sub->pmu->read(sub); |
4023 | ||
b5e58793 | 4024 | values[n++] = perf_event_count(sub); |
3dab77fb | 4025 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4026 | values[n++] = primary_event_id(sub); |
3dab77fb | 4027 | |
76369139 | 4028 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4029 | } |
4030 | } | |
4031 | ||
eed01528 SE |
4032 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4033 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4034 | ||
3dab77fb | 4035 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4036 | struct perf_event *event) |
3dab77fb | 4037 | { |
e3f3541c | 4038 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4039 | u64 read_format = event->attr.read_format; |
4040 | ||
4041 | /* | |
4042 | * compute total_time_enabled, total_time_running | |
4043 | * based on snapshot values taken when the event | |
4044 | * was last scheduled in. | |
4045 | * | |
4046 | * we cannot simply called update_context_time() | |
4047 | * because of locking issue as we are called in | |
4048 | * NMI context | |
4049 | */ | |
c4794295 | 4050 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4051 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4052 | |
cdd6c482 | 4053 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4054 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4055 | else |
eed01528 | 4056 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4057 | } |
4058 | ||
5622f295 MM |
4059 | void perf_output_sample(struct perf_output_handle *handle, |
4060 | struct perf_event_header *header, | |
4061 | struct perf_sample_data *data, | |
cdd6c482 | 4062 | struct perf_event *event) |
5622f295 MM |
4063 | { |
4064 | u64 sample_type = data->type; | |
4065 | ||
4066 | perf_output_put(handle, *header); | |
4067 | ||
4068 | if (sample_type & PERF_SAMPLE_IP) | |
4069 | perf_output_put(handle, data->ip); | |
4070 | ||
4071 | if (sample_type & PERF_SAMPLE_TID) | |
4072 | perf_output_put(handle, data->tid_entry); | |
4073 | ||
4074 | if (sample_type & PERF_SAMPLE_TIME) | |
4075 | perf_output_put(handle, data->time); | |
4076 | ||
4077 | if (sample_type & PERF_SAMPLE_ADDR) | |
4078 | perf_output_put(handle, data->addr); | |
4079 | ||
4080 | if (sample_type & PERF_SAMPLE_ID) | |
4081 | perf_output_put(handle, data->id); | |
4082 | ||
4083 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4084 | perf_output_put(handle, data->stream_id); | |
4085 | ||
4086 | if (sample_type & PERF_SAMPLE_CPU) | |
4087 | perf_output_put(handle, data->cpu_entry); | |
4088 | ||
4089 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4090 | perf_output_put(handle, data->period); | |
4091 | ||
4092 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4093 | perf_output_read(handle, event); |
5622f295 MM |
4094 | |
4095 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4096 | if (data->callchain) { | |
4097 | int size = 1; | |
4098 | ||
4099 | if (data->callchain) | |
4100 | size += data->callchain->nr; | |
4101 | ||
4102 | size *= sizeof(u64); | |
4103 | ||
76369139 | 4104 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4105 | } else { |
4106 | u64 nr = 0; | |
4107 | perf_output_put(handle, nr); | |
4108 | } | |
4109 | } | |
4110 | ||
4111 | if (sample_type & PERF_SAMPLE_RAW) { | |
4112 | if (data->raw) { | |
4113 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4114 | __output_copy(handle, data->raw->data, |
4115 | data->raw->size); | |
5622f295 MM |
4116 | } else { |
4117 | struct { | |
4118 | u32 size; | |
4119 | u32 data; | |
4120 | } raw = { | |
4121 | .size = sizeof(u32), | |
4122 | .data = 0, | |
4123 | }; | |
4124 | perf_output_put(handle, raw); | |
4125 | } | |
4126 | } | |
a7ac67ea PZ |
4127 | |
4128 | if (!event->attr.watermark) { | |
4129 | int wakeup_events = event->attr.wakeup_events; | |
4130 | ||
4131 | if (wakeup_events) { | |
4132 | struct ring_buffer *rb = handle->rb; | |
4133 | int events = local_inc_return(&rb->events); | |
4134 | ||
4135 | if (events >= wakeup_events) { | |
4136 | local_sub(wakeup_events, &rb->events); | |
4137 | local_inc(&rb->wakeup); | |
4138 | } | |
4139 | } | |
4140 | } | |
bce38cd5 SE |
4141 | |
4142 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4143 | if (data->br_stack) { | |
4144 | size_t size; | |
4145 | ||
4146 | size = data->br_stack->nr | |
4147 | * sizeof(struct perf_branch_entry); | |
4148 | ||
4149 | perf_output_put(handle, data->br_stack->nr); | |
4150 | perf_output_copy(handle, data->br_stack->entries, size); | |
4151 | } else { | |
4152 | /* | |
4153 | * we always store at least the value of nr | |
4154 | */ | |
4155 | u64 nr = 0; | |
4156 | perf_output_put(handle, nr); | |
4157 | } | |
4158 | } | |
4018994f JO |
4159 | |
4160 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4161 | u64 abi = data->regs_user.abi; | |
4162 | ||
4163 | /* | |
4164 | * If there are no regs to dump, notice it through | |
4165 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4166 | */ | |
4167 | perf_output_put(handle, abi); | |
4168 | ||
4169 | if (abi) { | |
4170 | u64 mask = event->attr.sample_regs_user; | |
4171 | perf_output_sample_regs(handle, | |
4172 | data->regs_user.regs, | |
4173 | mask); | |
4174 | } | |
4175 | } | |
c5ebcedb JO |
4176 | |
4177 | if (sample_type & PERF_SAMPLE_STACK_USER) | |
4178 | perf_output_sample_ustack(handle, | |
4179 | data->stack_user_size, | |
4180 | data->regs_user.regs); | |
5622f295 MM |
4181 | } |
4182 | ||
4183 | void perf_prepare_sample(struct perf_event_header *header, | |
4184 | struct perf_sample_data *data, | |
cdd6c482 | 4185 | struct perf_event *event, |
5622f295 | 4186 | struct pt_regs *regs) |
7b732a75 | 4187 | { |
cdd6c482 | 4188 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4189 | |
cdd6c482 | 4190 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4191 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4192 | |
4193 | header->misc = 0; | |
4194 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4195 | |
c980d109 | 4196 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4197 | |
c320c7b7 | 4198 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4199 | data->ip = perf_instruction_pointer(regs); |
4200 | ||
b23f3325 | 4201 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4202 | int size = 1; |
394ee076 | 4203 | |
e6dab5ff | 4204 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4205 | |
4206 | if (data->callchain) | |
4207 | size += data->callchain->nr; | |
4208 | ||
4209 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4210 | } |
4211 | ||
3a43ce68 | 4212 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4213 | int size = sizeof(u32); |
4214 | ||
4215 | if (data->raw) | |
4216 | size += data->raw->size; | |
4217 | else | |
4218 | size += sizeof(u32); | |
4219 | ||
4220 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4221 | header->size += size; |
7f453c24 | 4222 | } |
bce38cd5 SE |
4223 | |
4224 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4225 | int size = sizeof(u64); /* nr */ | |
4226 | if (data->br_stack) { | |
4227 | size += data->br_stack->nr | |
4228 | * sizeof(struct perf_branch_entry); | |
4229 | } | |
4230 | header->size += size; | |
4231 | } | |
4018994f JO |
4232 | |
4233 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4234 | /* regs dump ABI info */ | |
4235 | int size = sizeof(u64); | |
4236 | ||
4237 | perf_sample_regs_user(&data->regs_user, regs); | |
4238 | ||
4239 | if (data->regs_user.regs) { | |
4240 | u64 mask = event->attr.sample_regs_user; | |
4241 | size += hweight64(mask) * sizeof(u64); | |
4242 | } | |
4243 | ||
4244 | header->size += size; | |
4245 | } | |
c5ebcedb JO |
4246 | |
4247 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4248 | /* | |
4249 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4250 | * processed as the last one or have additional check added | |
4251 | * in case new sample type is added, because we could eat | |
4252 | * up the rest of the sample size. | |
4253 | */ | |
4254 | struct perf_regs_user *uregs = &data->regs_user; | |
4255 | u16 stack_size = event->attr.sample_stack_user; | |
4256 | u16 size = sizeof(u64); | |
4257 | ||
4258 | if (!uregs->abi) | |
4259 | perf_sample_regs_user(uregs, regs); | |
4260 | ||
4261 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4262 | uregs->regs); | |
4263 | ||
4264 | /* | |
4265 | * If there is something to dump, add space for the dump | |
4266 | * itself and for the field that tells the dynamic size, | |
4267 | * which is how many have been actually dumped. | |
4268 | */ | |
4269 | if (stack_size) | |
4270 | size += sizeof(u64) + stack_size; | |
4271 | ||
4272 | data->stack_user_size = stack_size; | |
4273 | header->size += size; | |
4274 | } | |
5622f295 | 4275 | } |
7f453c24 | 4276 | |
a8b0ca17 | 4277 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4278 | struct perf_sample_data *data, |
4279 | struct pt_regs *regs) | |
4280 | { | |
4281 | struct perf_output_handle handle; | |
4282 | struct perf_event_header header; | |
689802b2 | 4283 | |
927c7a9e FW |
4284 | /* protect the callchain buffers */ |
4285 | rcu_read_lock(); | |
4286 | ||
cdd6c482 | 4287 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4288 | |
a7ac67ea | 4289 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4290 | goto exit; |
0322cd6e | 4291 | |
cdd6c482 | 4292 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4293 | |
8a057d84 | 4294 | perf_output_end(&handle); |
927c7a9e FW |
4295 | |
4296 | exit: | |
4297 | rcu_read_unlock(); | |
0322cd6e PZ |
4298 | } |
4299 | ||
38b200d6 | 4300 | /* |
cdd6c482 | 4301 | * read event_id |
38b200d6 PZ |
4302 | */ |
4303 | ||
4304 | struct perf_read_event { | |
4305 | struct perf_event_header header; | |
4306 | ||
4307 | u32 pid; | |
4308 | u32 tid; | |
38b200d6 PZ |
4309 | }; |
4310 | ||
4311 | static void | |
cdd6c482 | 4312 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4313 | struct task_struct *task) |
4314 | { | |
4315 | struct perf_output_handle handle; | |
c980d109 | 4316 | struct perf_sample_data sample; |
dfc65094 | 4317 | struct perf_read_event read_event = { |
38b200d6 | 4318 | .header = { |
cdd6c482 | 4319 | .type = PERF_RECORD_READ, |
38b200d6 | 4320 | .misc = 0, |
c320c7b7 | 4321 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4322 | }, |
cdd6c482 IM |
4323 | .pid = perf_event_pid(event, task), |
4324 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4325 | }; |
3dab77fb | 4326 | int ret; |
38b200d6 | 4327 | |
c980d109 | 4328 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4329 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4330 | if (ret) |
4331 | return; | |
4332 | ||
dfc65094 | 4333 | perf_output_put(&handle, read_event); |
cdd6c482 | 4334 | perf_output_read(&handle, event); |
c980d109 | 4335 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4336 | |
38b200d6 PZ |
4337 | perf_output_end(&handle); |
4338 | } | |
4339 | ||
60313ebe | 4340 | /* |
9f498cc5 PZ |
4341 | * task tracking -- fork/exit |
4342 | * | |
3af9e859 | 4343 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
60313ebe PZ |
4344 | */ |
4345 | ||
9f498cc5 | 4346 | struct perf_task_event { |
3a80b4a3 | 4347 | struct task_struct *task; |
cdd6c482 | 4348 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4349 | |
4350 | struct { | |
4351 | struct perf_event_header header; | |
4352 | ||
4353 | u32 pid; | |
4354 | u32 ppid; | |
9f498cc5 PZ |
4355 | u32 tid; |
4356 | u32 ptid; | |
393b2ad8 | 4357 | u64 time; |
cdd6c482 | 4358 | } event_id; |
60313ebe PZ |
4359 | }; |
4360 | ||
cdd6c482 | 4361 | static void perf_event_task_output(struct perf_event *event, |
9f498cc5 | 4362 | struct perf_task_event *task_event) |
60313ebe PZ |
4363 | { |
4364 | struct perf_output_handle handle; | |
c980d109 | 4365 | struct perf_sample_data sample; |
9f498cc5 | 4366 | struct task_struct *task = task_event->task; |
c980d109 | 4367 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4368 | |
c980d109 | 4369 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4370 | |
c980d109 | 4371 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4372 | task_event->event_id.header.size); |
ef60777c | 4373 | if (ret) |
c980d109 | 4374 | goto out; |
60313ebe | 4375 | |
cdd6c482 IM |
4376 | task_event->event_id.pid = perf_event_pid(event, task); |
4377 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4378 | |
cdd6c482 IM |
4379 | task_event->event_id.tid = perf_event_tid(event, task); |
4380 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4381 | |
cdd6c482 | 4382 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4383 | |
c980d109 ACM |
4384 | perf_event__output_id_sample(event, &handle, &sample); |
4385 | ||
60313ebe | 4386 | perf_output_end(&handle); |
c980d109 ACM |
4387 | out: |
4388 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4389 | } |
4390 | ||
cdd6c482 | 4391 | static int perf_event_task_match(struct perf_event *event) |
60313ebe | 4392 | { |
6f93d0a7 | 4393 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4394 | return 0; |
4395 | ||
5632ab12 | 4396 | if (!event_filter_match(event)) |
5d27c23d PZ |
4397 | return 0; |
4398 | ||
3af9e859 EM |
4399 | if (event->attr.comm || event->attr.mmap || |
4400 | event->attr.mmap_data || event->attr.task) | |
60313ebe PZ |
4401 | return 1; |
4402 | ||
4403 | return 0; | |
4404 | } | |
4405 | ||
cdd6c482 | 4406 | static void perf_event_task_ctx(struct perf_event_context *ctx, |
9f498cc5 | 4407 | struct perf_task_event *task_event) |
60313ebe | 4408 | { |
cdd6c482 | 4409 | struct perf_event *event; |
60313ebe | 4410 | |
cdd6c482 IM |
4411 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4412 | if (perf_event_task_match(event)) | |
4413 | perf_event_task_output(event, task_event); | |
60313ebe | 4414 | } |
60313ebe PZ |
4415 | } |
4416 | ||
cdd6c482 | 4417 | static void perf_event_task_event(struct perf_task_event *task_event) |
60313ebe PZ |
4418 | { |
4419 | struct perf_cpu_context *cpuctx; | |
8dc85d54 | 4420 | struct perf_event_context *ctx; |
108b02cf | 4421 | struct pmu *pmu; |
8dc85d54 | 4422 | int ctxn; |
60313ebe | 4423 | |
d6ff86cf | 4424 | rcu_read_lock(); |
108b02cf | 4425 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4426 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
3f1f3320 | 4427 | if (cpuctx->unique_pmu != pmu) |
51676957 | 4428 | goto next; |
108b02cf | 4429 | perf_event_task_ctx(&cpuctx->ctx, task_event); |
8dc85d54 PZ |
4430 | |
4431 | ctx = task_event->task_ctx; | |
4432 | if (!ctx) { | |
4433 | ctxn = pmu->task_ctx_nr; | |
4434 | if (ctxn < 0) | |
41945f6c | 4435 | goto next; |
8dc85d54 | 4436 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); |
d610d98b NK |
4437 | if (ctx) |
4438 | perf_event_task_ctx(ctx, task_event); | |
8dc85d54 | 4439 | } |
41945f6c PZ |
4440 | next: |
4441 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4442 | } |
d610d98b NK |
4443 | if (task_event->task_ctx) |
4444 | perf_event_task_ctx(task_event->task_ctx, task_event); | |
4445 | ||
60313ebe PZ |
4446 | rcu_read_unlock(); |
4447 | } | |
4448 | ||
cdd6c482 IM |
4449 | static void perf_event_task(struct task_struct *task, |
4450 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4451 | int new) |
60313ebe | 4452 | { |
9f498cc5 | 4453 | struct perf_task_event task_event; |
60313ebe | 4454 | |
cdd6c482 IM |
4455 | if (!atomic_read(&nr_comm_events) && |
4456 | !atomic_read(&nr_mmap_events) && | |
4457 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4458 | return; |
4459 | ||
9f498cc5 | 4460 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4461 | .task = task, |
4462 | .task_ctx = task_ctx, | |
cdd6c482 | 4463 | .event_id = { |
60313ebe | 4464 | .header = { |
cdd6c482 | 4465 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4466 | .misc = 0, |
cdd6c482 | 4467 | .size = sizeof(task_event.event_id), |
60313ebe | 4468 | }, |
573402db PZ |
4469 | /* .pid */ |
4470 | /* .ppid */ | |
9f498cc5 PZ |
4471 | /* .tid */ |
4472 | /* .ptid */ | |
6f93d0a7 | 4473 | .time = perf_clock(), |
60313ebe PZ |
4474 | }, |
4475 | }; | |
4476 | ||
cdd6c482 | 4477 | perf_event_task_event(&task_event); |
9f498cc5 PZ |
4478 | } |
4479 | ||
cdd6c482 | 4480 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4481 | { |
cdd6c482 | 4482 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4483 | } |
4484 | ||
8d1b2d93 PZ |
4485 | /* |
4486 | * comm tracking | |
4487 | */ | |
4488 | ||
4489 | struct perf_comm_event { | |
22a4f650 IM |
4490 | struct task_struct *task; |
4491 | char *comm; | |
8d1b2d93 PZ |
4492 | int comm_size; |
4493 | ||
4494 | struct { | |
4495 | struct perf_event_header header; | |
4496 | ||
4497 | u32 pid; | |
4498 | u32 tid; | |
cdd6c482 | 4499 | } event_id; |
8d1b2d93 PZ |
4500 | }; |
4501 | ||
cdd6c482 | 4502 | static void perf_event_comm_output(struct perf_event *event, |
8d1b2d93 PZ |
4503 | struct perf_comm_event *comm_event) |
4504 | { | |
4505 | struct perf_output_handle handle; | |
c980d109 | 4506 | struct perf_sample_data sample; |
cdd6c482 | 4507 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4508 | int ret; |
4509 | ||
4510 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); | |
4511 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4512 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4513 | |
4514 | if (ret) | |
c980d109 | 4515 | goto out; |
8d1b2d93 | 4516 | |
cdd6c482 IM |
4517 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4518 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4519 | |
cdd6c482 | 4520 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4521 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4522 | comm_event->comm_size); |
c980d109 ACM |
4523 | |
4524 | perf_event__output_id_sample(event, &handle, &sample); | |
4525 | ||
8d1b2d93 | 4526 | perf_output_end(&handle); |
c980d109 ACM |
4527 | out: |
4528 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4529 | } |
4530 | ||
cdd6c482 | 4531 | static int perf_event_comm_match(struct perf_event *event) |
8d1b2d93 | 4532 | { |
6f93d0a7 | 4533 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4534 | return 0; |
4535 | ||
5632ab12 | 4536 | if (!event_filter_match(event)) |
5d27c23d PZ |
4537 | return 0; |
4538 | ||
cdd6c482 | 4539 | if (event->attr.comm) |
8d1b2d93 PZ |
4540 | return 1; |
4541 | ||
4542 | return 0; | |
4543 | } | |
4544 | ||
cdd6c482 | 4545 | static void perf_event_comm_ctx(struct perf_event_context *ctx, |
8d1b2d93 PZ |
4546 | struct perf_comm_event *comm_event) |
4547 | { | |
cdd6c482 | 4548 | struct perf_event *event; |
8d1b2d93 | 4549 | |
cdd6c482 IM |
4550 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
4551 | if (perf_event_comm_match(event)) | |
4552 | perf_event_comm_output(event, comm_event); | |
8d1b2d93 | 4553 | } |
8d1b2d93 PZ |
4554 | } |
4555 | ||
cdd6c482 | 4556 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 PZ |
4557 | { |
4558 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4559 | struct perf_event_context *ctx; |
413ee3b4 | 4560 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4561 | unsigned int size; |
108b02cf | 4562 | struct pmu *pmu; |
8dc85d54 | 4563 | int ctxn; |
8d1b2d93 | 4564 | |
413ee3b4 | 4565 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4566 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4567 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4568 | |
4569 | comm_event->comm = comm; | |
4570 | comm_event->comm_size = size; | |
4571 | ||
cdd6c482 | 4572 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
f6595f3a | 4573 | rcu_read_lock(); |
108b02cf | 4574 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4575 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
3f1f3320 | 4576 | if (cpuctx->unique_pmu != pmu) |
51676957 | 4577 | goto next; |
108b02cf | 4578 | perf_event_comm_ctx(&cpuctx->ctx, comm_event); |
8dc85d54 PZ |
4579 | |
4580 | ctxn = pmu->task_ctx_nr; | |
4581 | if (ctxn < 0) | |
41945f6c | 4582 | goto next; |
8dc85d54 PZ |
4583 | |
4584 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4585 | if (ctx) | |
4586 | perf_event_comm_ctx(ctx, comm_event); | |
41945f6c PZ |
4587 | next: |
4588 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4589 | } |
665c2142 | 4590 | rcu_read_unlock(); |
8d1b2d93 PZ |
4591 | } |
4592 | ||
cdd6c482 | 4593 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4594 | { |
9ee318a7 | 4595 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4596 | struct perf_event_context *ctx; |
4597 | int ctxn; | |
9ee318a7 | 4598 | |
8dc85d54 PZ |
4599 | for_each_task_context_nr(ctxn) { |
4600 | ctx = task->perf_event_ctxp[ctxn]; | |
4601 | if (!ctx) | |
4602 | continue; | |
9ee318a7 | 4603 | |
8dc85d54 PZ |
4604 | perf_event_enable_on_exec(ctx); |
4605 | } | |
9ee318a7 | 4606 | |
cdd6c482 | 4607 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4608 | return; |
a63eaf34 | 4609 | |
9ee318a7 | 4610 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4611 | .task = task, |
573402db PZ |
4612 | /* .comm */ |
4613 | /* .comm_size */ | |
cdd6c482 | 4614 | .event_id = { |
573402db | 4615 | .header = { |
cdd6c482 | 4616 | .type = PERF_RECORD_COMM, |
573402db PZ |
4617 | .misc = 0, |
4618 | /* .size */ | |
4619 | }, | |
4620 | /* .pid */ | |
4621 | /* .tid */ | |
8d1b2d93 PZ |
4622 | }, |
4623 | }; | |
4624 | ||
cdd6c482 | 4625 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
4626 | } |
4627 | ||
0a4a9391 PZ |
4628 | /* |
4629 | * mmap tracking | |
4630 | */ | |
4631 | ||
4632 | struct perf_mmap_event { | |
089dd79d PZ |
4633 | struct vm_area_struct *vma; |
4634 | ||
4635 | const char *file_name; | |
4636 | int file_size; | |
0a4a9391 PZ |
4637 | |
4638 | struct { | |
4639 | struct perf_event_header header; | |
4640 | ||
4641 | u32 pid; | |
4642 | u32 tid; | |
4643 | u64 start; | |
4644 | u64 len; | |
4645 | u64 pgoff; | |
cdd6c482 | 4646 | } event_id; |
0a4a9391 PZ |
4647 | }; |
4648 | ||
cdd6c482 | 4649 | static void perf_event_mmap_output(struct perf_event *event, |
0a4a9391 PZ |
4650 | struct perf_mmap_event *mmap_event) |
4651 | { | |
4652 | struct perf_output_handle handle; | |
c980d109 | 4653 | struct perf_sample_data sample; |
cdd6c482 | 4654 | int size = mmap_event->event_id.header.size; |
c980d109 | 4655 | int ret; |
0a4a9391 | 4656 | |
c980d109 ACM |
4657 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
4658 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4659 | mmap_event->event_id.header.size); |
0a4a9391 | 4660 | if (ret) |
c980d109 | 4661 | goto out; |
0a4a9391 | 4662 | |
cdd6c482 IM |
4663 | mmap_event->event_id.pid = perf_event_pid(event, current); |
4664 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 4665 | |
cdd6c482 | 4666 | perf_output_put(&handle, mmap_event->event_id); |
76369139 | 4667 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 4668 | mmap_event->file_size); |
c980d109 ACM |
4669 | |
4670 | perf_event__output_id_sample(event, &handle, &sample); | |
4671 | ||
78d613eb | 4672 | perf_output_end(&handle); |
c980d109 ACM |
4673 | out: |
4674 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
4675 | } |
4676 | ||
cdd6c482 | 4677 | static int perf_event_mmap_match(struct perf_event *event, |
3af9e859 EM |
4678 | struct perf_mmap_event *mmap_event, |
4679 | int executable) | |
0a4a9391 | 4680 | { |
6f93d0a7 | 4681 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
22e19085 PZ |
4682 | return 0; |
4683 | ||
5632ab12 | 4684 | if (!event_filter_match(event)) |
5d27c23d PZ |
4685 | return 0; |
4686 | ||
3af9e859 EM |
4687 | if ((!executable && event->attr.mmap_data) || |
4688 | (executable && event->attr.mmap)) | |
0a4a9391 PZ |
4689 | return 1; |
4690 | ||
4691 | return 0; | |
4692 | } | |
4693 | ||
cdd6c482 | 4694 | static void perf_event_mmap_ctx(struct perf_event_context *ctx, |
3af9e859 EM |
4695 | struct perf_mmap_event *mmap_event, |
4696 | int executable) | |
0a4a9391 | 4697 | { |
cdd6c482 | 4698 | struct perf_event *event; |
0a4a9391 | 4699 | |
cdd6c482 | 4700 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
3af9e859 | 4701 | if (perf_event_mmap_match(event, mmap_event, executable)) |
cdd6c482 | 4702 | perf_event_mmap_output(event, mmap_event); |
0a4a9391 | 4703 | } |
0a4a9391 PZ |
4704 | } |
4705 | ||
cdd6c482 | 4706 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 PZ |
4707 | { |
4708 | struct perf_cpu_context *cpuctx; | |
cdd6c482 | 4709 | struct perf_event_context *ctx; |
089dd79d PZ |
4710 | struct vm_area_struct *vma = mmap_event->vma; |
4711 | struct file *file = vma->vm_file; | |
0a4a9391 PZ |
4712 | unsigned int size; |
4713 | char tmp[16]; | |
4714 | char *buf = NULL; | |
089dd79d | 4715 | const char *name; |
108b02cf | 4716 | struct pmu *pmu; |
8dc85d54 | 4717 | int ctxn; |
0a4a9391 | 4718 | |
413ee3b4 AB |
4719 | memset(tmp, 0, sizeof(tmp)); |
4720 | ||
0a4a9391 | 4721 | if (file) { |
413ee3b4 | 4722 | /* |
76369139 | 4723 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
4724 | * need to add enough zero bytes after the string to handle |
4725 | * the 64bit alignment we do later. | |
4726 | */ | |
4727 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
4728 | if (!buf) { |
4729 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
4730 | goto got_name; | |
4731 | } | |
d3d21c41 | 4732 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
4733 | if (IS_ERR(name)) { |
4734 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
4735 | goto got_name; | |
4736 | } | |
4737 | } else { | |
413ee3b4 AB |
4738 | if (arch_vma_name(mmap_event->vma)) { |
4739 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
4740 | sizeof(tmp)); | |
089dd79d | 4741 | goto got_name; |
413ee3b4 | 4742 | } |
089dd79d PZ |
4743 | |
4744 | if (!vma->vm_mm) { | |
4745 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
4746 | goto got_name; | |
3af9e859 EM |
4747 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
4748 | vma->vm_end >= vma->vm_mm->brk) { | |
4749 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
4750 | goto got_name; | |
4751 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
4752 | vma->vm_end >= vma->vm_mm->start_stack) { | |
4753 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
4754 | goto got_name; | |
089dd79d PZ |
4755 | } |
4756 | ||
0a4a9391 PZ |
4757 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
4758 | goto got_name; | |
4759 | } | |
4760 | ||
4761 | got_name: | |
888fcee0 | 4762 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
4763 | |
4764 | mmap_event->file_name = name; | |
4765 | mmap_event->file_size = size; | |
4766 | ||
cdd6c482 | 4767 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 4768 | |
f6d9dd23 | 4769 | rcu_read_lock(); |
108b02cf | 4770 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
41945f6c | 4771 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
3f1f3320 | 4772 | if (cpuctx->unique_pmu != pmu) |
51676957 | 4773 | goto next; |
108b02cf PZ |
4774 | perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, |
4775 | vma->vm_flags & VM_EXEC); | |
8dc85d54 PZ |
4776 | |
4777 | ctxn = pmu->task_ctx_nr; | |
4778 | if (ctxn < 0) | |
41945f6c | 4779 | goto next; |
8dc85d54 PZ |
4780 | |
4781 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4782 | if (ctx) { | |
4783 | perf_event_mmap_ctx(ctx, mmap_event, | |
4784 | vma->vm_flags & VM_EXEC); | |
4785 | } | |
41945f6c PZ |
4786 | next: |
4787 | put_cpu_ptr(pmu->pmu_cpu_context); | |
108b02cf | 4788 | } |
665c2142 PZ |
4789 | rcu_read_unlock(); |
4790 | ||
0a4a9391 PZ |
4791 | kfree(buf); |
4792 | } | |
4793 | ||
3af9e859 | 4794 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 4795 | { |
9ee318a7 PZ |
4796 | struct perf_mmap_event mmap_event; |
4797 | ||
cdd6c482 | 4798 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
4799 | return; |
4800 | ||
4801 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 4802 | .vma = vma, |
573402db PZ |
4803 | /* .file_name */ |
4804 | /* .file_size */ | |
cdd6c482 | 4805 | .event_id = { |
573402db | 4806 | .header = { |
cdd6c482 | 4807 | .type = PERF_RECORD_MMAP, |
39447b38 | 4808 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
4809 | /* .size */ |
4810 | }, | |
4811 | /* .pid */ | |
4812 | /* .tid */ | |
089dd79d PZ |
4813 | .start = vma->vm_start, |
4814 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 4815 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 PZ |
4816 | }, |
4817 | }; | |
4818 | ||
cdd6c482 | 4819 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
4820 | } |
4821 | ||
a78ac325 PZ |
4822 | /* |
4823 | * IRQ throttle logging | |
4824 | */ | |
4825 | ||
cdd6c482 | 4826 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
4827 | { |
4828 | struct perf_output_handle handle; | |
c980d109 | 4829 | struct perf_sample_data sample; |
a78ac325 PZ |
4830 | int ret; |
4831 | ||
4832 | struct { | |
4833 | struct perf_event_header header; | |
4834 | u64 time; | |
cca3f454 | 4835 | u64 id; |
7f453c24 | 4836 | u64 stream_id; |
a78ac325 PZ |
4837 | } throttle_event = { |
4838 | .header = { | |
cdd6c482 | 4839 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
4840 | .misc = 0, |
4841 | .size = sizeof(throttle_event), | |
4842 | }, | |
def0a9b2 | 4843 | .time = perf_clock(), |
cdd6c482 IM |
4844 | .id = primary_event_id(event), |
4845 | .stream_id = event->id, | |
a78ac325 PZ |
4846 | }; |
4847 | ||
966ee4d6 | 4848 | if (enable) |
cdd6c482 | 4849 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 4850 | |
c980d109 ACM |
4851 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
4852 | ||
4853 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4854 | throttle_event.header.size); |
a78ac325 PZ |
4855 | if (ret) |
4856 | return; | |
4857 | ||
4858 | perf_output_put(&handle, throttle_event); | |
c980d109 | 4859 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
4860 | perf_output_end(&handle); |
4861 | } | |
4862 | ||
f6c7d5fe | 4863 | /* |
cdd6c482 | 4864 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
4865 | */ |
4866 | ||
a8b0ca17 | 4867 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4868 | int throttle, struct perf_sample_data *data, |
4869 | struct pt_regs *regs) | |
f6c7d5fe | 4870 | { |
cdd6c482 IM |
4871 | int events = atomic_read(&event->event_limit); |
4872 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 4873 | u64 seq; |
79f14641 PZ |
4874 | int ret = 0; |
4875 | ||
96398826 PZ |
4876 | /* |
4877 | * Non-sampling counters might still use the PMI to fold short | |
4878 | * hardware counters, ignore those. | |
4879 | */ | |
4880 | if (unlikely(!is_sampling_event(event))) | |
4881 | return 0; | |
4882 | ||
e050e3f0 SE |
4883 | seq = __this_cpu_read(perf_throttled_seq); |
4884 | if (seq != hwc->interrupts_seq) { | |
4885 | hwc->interrupts_seq = seq; | |
4886 | hwc->interrupts = 1; | |
4887 | } else { | |
4888 | hwc->interrupts++; | |
4889 | if (unlikely(throttle | |
4890 | && hwc->interrupts >= max_samples_per_tick)) { | |
4891 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
4892 | hwc->interrupts = MAX_INTERRUPTS; |
4893 | perf_log_throttle(event, 0); | |
a78ac325 PZ |
4894 | ret = 1; |
4895 | } | |
e050e3f0 | 4896 | } |
60db5e09 | 4897 | |
cdd6c482 | 4898 | if (event->attr.freq) { |
def0a9b2 | 4899 | u64 now = perf_clock(); |
abd50713 | 4900 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 4901 | |
abd50713 | 4902 | hwc->freq_time_stamp = now; |
bd2b5b12 | 4903 | |
abd50713 | 4904 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 4905 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
4906 | } |
4907 | ||
2023b359 PZ |
4908 | /* |
4909 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 4910 | * events |
2023b359 PZ |
4911 | */ |
4912 | ||
cdd6c482 IM |
4913 | event->pending_kill = POLL_IN; |
4914 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 4915 | ret = 1; |
cdd6c482 | 4916 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
4917 | event->pending_disable = 1; |
4918 | irq_work_queue(&event->pending); | |
79f14641 PZ |
4919 | } |
4920 | ||
453f19ee | 4921 | if (event->overflow_handler) |
a8b0ca17 | 4922 | event->overflow_handler(event, data, regs); |
453f19ee | 4923 | else |
a8b0ca17 | 4924 | perf_event_output(event, data, regs); |
453f19ee | 4925 | |
f506b3dc | 4926 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
4927 | event->pending_wakeup = 1; |
4928 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
4929 | } |
4930 | ||
79f14641 | 4931 | return ret; |
f6c7d5fe PZ |
4932 | } |
4933 | ||
a8b0ca17 | 4934 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
4935 | struct perf_sample_data *data, |
4936 | struct pt_regs *regs) | |
850bc73f | 4937 | { |
a8b0ca17 | 4938 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
4939 | } |
4940 | ||
15dbf27c | 4941 | /* |
cdd6c482 | 4942 | * Generic software event infrastructure |
15dbf27c PZ |
4943 | */ |
4944 | ||
b28ab83c PZ |
4945 | struct swevent_htable { |
4946 | struct swevent_hlist *swevent_hlist; | |
4947 | struct mutex hlist_mutex; | |
4948 | int hlist_refcount; | |
4949 | ||
4950 | /* Recursion avoidance in each contexts */ | |
4951 | int recursion[PERF_NR_CONTEXTS]; | |
4952 | }; | |
4953 | ||
4954 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
4955 | ||
7b4b6658 | 4956 | /* |
cdd6c482 IM |
4957 | * We directly increment event->count and keep a second value in |
4958 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
4959 | * is kept in the range [-sample_period, 0] so that we can use the |
4960 | * sign as trigger. | |
4961 | */ | |
4962 | ||
cdd6c482 | 4963 | static u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 4964 | { |
cdd6c482 | 4965 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
4966 | u64 period = hwc->last_period; |
4967 | u64 nr, offset; | |
4968 | s64 old, val; | |
4969 | ||
4970 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
4971 | |
4972 | again: | |
e7850595 | 4973 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
4974 | if (val < 0) |
4975 | return 0; | |
15dbf27c | 4976 | |
7b4b6658 PZ |
4977 | nr = div64_u64(period + val, period); |
4978 | offset = nr * period; | |
4979 | val -= offset; | |
e7850595 | 4980 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 4981 | goto again; |
15dbf27c | 4982 | |
7b4b6658 | 4983 | return nr; |
15dbf27c PZ |
4984 | } |
4985 | ||
0cff784a | 4986 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 4987 | struct perf_sample_data *data, |
5622f295 | 4988 | struct pt_regs *regs) |
15dbf27c | 4989 | { |
cdd6c482 | 4990 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 4991 | int throttle = 0; |
15dbf27c | 4992 | |
0cff784a PZ |
4993 | if (!overflow) |
4994 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 4995 | |
7b4b6658 PZ |
4996 | if (hwc->interrupts == MAX_INTERRUPTS) |
4997 | return; | |
15dbf27c | 4998 | |
7b4b6658 | 4999 | for (; overflow; overflow--) { |
a8b0ca17 | 5000 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5001 | data, regs)) { |
7b4b6658 PZ |
5002 | /* |
5003 | * We inhibit the overflow from happening when | |
5004 | * hwc->interrupts == MAX_INTERRUPTS. | |
5005 | */ | |
5006 | break; | |
5007 | } | |
cf450a73 | 5008 | throttle = 1; |
7b4b6658 | 5009 | } |
15dbf27c PZ |
5010 | } |
5011 | ||
a4eaf7f1 | 5012 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5013 | struct perf_sample_data *data, |
5622f295 | 5014 | struct pt_regs *regs) |
7b4b6658 | 5015 | { |
cdd6c482 | 5016 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5017 | |
e7850595 | 5018 | local64_add(nr, &event->count); |
d6d020e9 | 5019 | |
0cff784a PZ |
5020 | if (!regs) |
5021 | return; | |
5022 | ||
6c7e550f | 5023 | if (!is_sampling_event(event)) |
7b4b6658 | 5024 | return; |
d6d020e9 | 5025 | |
5d81e5cf AV |
5026 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5027 | data->period = nr; | |
5028 | return perf_swevent_overflow(event, 1, data, regs); | |
5029 | } else | |
5030 | data->period = event->hw.last_period; | |
5031 | ||
0cff784a | 5032 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5033 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5034 | |
e7850595 | 5035 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5036 | return; |
df1a132b | 5037 | |
a8b0ca17 | 5038 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5039 | } |
5040 | ||
f5ffe02e FW |
5041 | static int perf_exclude_event(struct perf_event *event, |
5042 | struct pt_regs *regs) | |
5043 | { | |
a4eaf7f1 | 5044 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5045 | return 1; |
a4eaf7f1 | 5046 | |
f5ffe02e FW |
5047 | if (regs) { |
5048 | if (event->attr.exclude_user && user_mode(regs)) | |
5049 | return 1; | |
5050 | ||
5051 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5052 | return 1; | |
5053 | } | |
5054 | ||
5055 | return 0; | |
5056 | } | |
5057 | ||
cdd6c482 | 5058 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5059 | enum perf_type_id type, |
6fb2915d LZ |
5060 | u32 event_id, |
5061 | struct perf_sample_data *data, | |
5062 | struct pt_regs *regs) | |
15dbf27c | 5063 | { |
cdd6c482 | 5064 | if (event->attr.type != type) |
a21ca2ca | 5065 | return 0; |
f5ffe02e | 5066 | |
cdd6c482 | 5067 | if (event->attr.config != event_id) |
15dbf27c PZ |
5068 | return 0; |
5069 | ||
f5ffe02e FW |
5070 | if (perf_exclude_event(event, regs)) |
5071 | return 0; | |
15dbf27c PZ |
5072 | |
5073 | return 1; | |
5074 | } | |
5075 | ||
76e1d904 FW |
5076 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5077 | { | |
5078 | u64 val = event_id | (type << 32); | |
5079 | ||
5080 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5081 | } | |
5082 | ||
49f135ed FW |
5083 | static inline struct hlist_head * |
5084 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5085 | { |
49f135ed FW |
5086 | u64 hash = swevent_hash(type, event_id); |
5087 | ||
5088 | return &hlist->heads[hash]; | |
5089 | } | |
76e1d904 | 5090 | |
49f135ed FW |
5091 | /* For the read side: events when they trigger */ |
5092 | static inline struct hlist_head * | |
b28ab83c | 5093 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5094 | { |
5095 | struct swevent_hlist *hlist; | |
76e1d904 | 5096 | |
b28ab83c | 5097 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5098 | if (!hlist) |
5099 | return NULL; | |
5100 | ||
49f135ed FW |
5101 | return __find_swevent_head(hlist, type, event_id); |
5102 | } | |
5103 | ||
5104 | /* For the event head insertion and removal in the hlist */ | |
5105 | static inline struct hlist_head * | |
b28ab83c | 5106 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5107 | { |
5108 | struct swevent_hlist *hlist; | |
5109 | u32 event_id = event->attr.config; | |
5110 | u64 type = event->attr.type; | |
5111 | ||
5112 | /* | |
5113 | * Event scheduling is always serialized against hlist allocation | |
5114 | * and release. Which makes the protected version suitable here. | |
5115 | * The context lock guarantees that. | |
5116 | */ | |
b28ab83c | 5117 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5118 | lockdep_is_held(&event->ctx->lock)); |
5119 | if (!hlist) | |
5120 | return NULL; | |
5121 | ||
5122 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5123 | } |
5124 | ||
5125 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5126 | u64 nr, |
76e1d904 FW |
5127 | struct perf_sample_data *data, |
5128 | struct pt_regs *regs) | |
15dbf27c | 5129 | { |
b28ab83c | 5130 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5131 | struct perf_event *event; |
76e1d904 | 5132 | struct hlist_head *head; |
15dbf27c | 5133 | |
76e1d904 | 5134 | rcu_read_lock(); |
b28ab83c | 5135 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5136 | if (!head) |
5137 | goto end; | |
5138 | ||
b67bfe0d | 5139 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5140 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5141 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5142 | } |
76e1d904 FW |
5143 | end: |
5144 | rcu_read_unlock(); | |
15dbf27c PZ |
5145 | } |
5146 | ||
4ed7c92d | 5147 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5148 | { |
b28ab83c | 5149 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5150 | |
b28ab83c | 5151 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5152 | } |
645e8cc0 | 5153 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5154 | |
fa9f90be | 5155 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5156 | { |
b28ab83c | 5157 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5158 | |
b28ab83c | 5159 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5160 | } |
15dbf27c | 5161 | |
a8b0ca17 | 5162 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5163 | { |
a4234bfc | 5164 | struct perf_sample_data data; |
4ed7c92d PZ |
5165 | int rctx; |
5166 | ||
1c024eca | 5167 | preempt_disable_notrace(); |
4ed7c92d PZ |
5168 | rctx = perf_swevent_get_recursion_context(); |
5169 | if (rctx < 0) | |
5170 | return; | |
a4234bfc | 5171 | |
fd0d000b | 5172 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5173 | |
a8b0ca17 | 5174 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5175 | |
5176 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5177 | preempt_enable_notrace(); |
b8e83514 PZ |
5178 | } |
5179 | ||
cdd6c482 | 5180 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5181 | { |
15dbf27c PZ |
5182 | } |
5183 | ||
a4eaf7f1 | 5184 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5185 | { |
b28ab83c | 5186 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5187 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5188 | struct hlist_head *head; |
5189 | ||
6c7e550f | 5190 | if (is_sampling_event(event)) { |
7b4b6658 | 5191 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5192 | perf_swevent_set_period(event); |
7b4b6658 | 5193 | } |
76e1d904 | 5194 | |
a4eaf7f1 PZ |
5195 | hwc->state = !(flags & PERF_EF_START); |
5196 | ||
b28ab83c | 5197 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
5198 | if (WARN_ON_ONCE(!head)) |
5199 | return -EINVAL; | |
5200 | ||
5201 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5202 | ||
15dbf27c PZ |
5203 | return 0; |
5204 | } | |
5205 | ||
a4eaf7f1 | 5206 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5207 | { |
76e1d904 | 5208 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5209 | } |
5210 | ||
a4eaf7f1 | 5211 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5212 | { |
a4eaf7f1 | 5213 | event->hw.state = 0; |
d6d020e9 | 5214 | } |
aa9c4c0f | 5215 | |
a4eaf7f1 | 5216 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5217 | { |
a4eaf7f1 | 5218 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5219 | } |
5220 | ||
49f135ed FW |
5221 | /* Deref the hlist from the update side */ |
5222 | static inline struct swevent_hlist * | |
b28ab83c | 5223 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5224 | { |
b28ab83c PZ |
5225 | return rcu_dereference_protected(swhash->swevent_hlist, |
5226 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5227 | } |
5228 | ||
b28ab83c | 5229 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5230 | { |
b28ab83c | 5231 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5232 | |
49f135ed | 5233 | if (!hlist) |
76e1d904 FW |
5234 | return; |
5235 | ||
b28ab83c | 5236 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5237 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5238 | } |
5239 | ||
5240 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5241 | { | |
b28ab83c | 5242 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5243 | |
b28ab83c | 5244 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5245 | |
b28ab83c PZ |
5246 | if (!--swhash->hlist_refcount) |
5247 | swevent_hlist_release(swhash); | |
76e1d904 | 5248 | |
b28ab83c | 5249 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5250 | } |
5251 | ||
5252 | static void swevent_hlist_put(struct perf_event *event) | |
5253 | { | |
5254 | int cpu; | |
5255 | ||
5256 | if (event->cpu != -1) { | |
5257 | swevent_hlist_put_cpu(event, event->cpu); | |
5258 | return; | |
5259 | } | |
5260 | ||
5261 | for_each_possible_cpu(cpu) | |
5262 | swevent_hlist_put_cpu(event, cpu); | |
5263 | } | |
5264 | ||
5265 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5266 | { | |
b28ab83c | 5267 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5268 | int err = 0; |
5269 | ||
b28ab83c | 5270 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5271 | |
b28ab83c | 5272 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5273 | struct swevent_hlist *hlist; |
5274 | ||
5275 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5276 | if (!hlist) { | |
5277 | err = -ENOMEM; | |
5278 | goto exit; | |
5279 | } | |
b28ab83c | 5280 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5281 | } |
b28ab83c | 5282 | swhash->hlist_refcount++; |
9ed6060d | 5283 | exit: |
b28ab83c | 5284 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5285 | |
5286 | return err; | |
5287 | } | |
5288 | ||
5289 | static int swevent_hlist_get(struct perf_event *event) | |
5290 | { | |
5291 | int err; | |
5292 | int cpu, failed_cpu; | |
5293 | ||
5294 | if (event->cpu != -1) | |
5295 | return swevent_hlist_get_cpu(event, event->cpu); | |
5296 | ||
5297 | get_online_cpus(); | |
5298 | for_each_possible_cpu(cpu) { | |
5299 | err = swevent_hlist_get_cpu(event, cpu); | |
5300 | if (err) { | |
5301 | failed_cpu = cpu; | |
5302 | goto fail; | |
5303 | } | |
5304 | } | |
5305 | put_online_cpus(); | |
5306 | ||
5307 | return 0; | |
9ed6060d | 5308 | fail: |
76e1d904 FW |
5309 | for_each_possible_cpu(cpu) { |
5310 | if (cpu == failed_cpu) | |
5311 | break; | |
5312 | swevent_hlist_put_cpu(event, cpu); | |
5313 | } | |
5314 | ||
5315 | put_online_cpus(); | |
5316 | return err; | |
5317 | } | |
5318 | ||
c5905afb | 5319 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5320 | |
b0a873eb PZ |
5321 | static void sw_perf_event_destroy(struct perf_event *event) |
5322 | { | |
5323 | u64 event_id = event->attr.config; | |
95476b64 | 5324 | |
b0a873eb PZ |
5325 | WARN_ON(event->parent); |
5326 | ||
c5905afb | 5327 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5328 | swevent_hlist_put(event); |
5329 | } | |
5330 | ||
5331 | static int perf_swevent_init(struct perf_event *event) | |
5332 | { | |
5333 | int event_id = event->attr.config; | |
5334 | ||
5335 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5336 | return -ENOENT; | |
5337 | ||
2481c5fa SE |
5338 | /* |
5339 | * no branch sampling for software events | |
5340 | */ | |
5341 | if (has_branch_stack(event)) | |
5342 | return -EOPNOTSUPP; | |
5343 | ||
b0a873eb PZ |
5344 | switch (event_id) { |
5345 | case PERF_COUNT_SW_CPU_CLOCK: | |
5346 | case PERF_COUNT_SW_TASK_CLOCK: | |
5347 | return -ENOENT; | |
5348 | ||
5349 | default: | |
5350 | break; | |
5351 | } | |
5352 | ||
ce677831 | 5353 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5354 | return -ENOENT; |
5355 | ||
5356 | if (!event->parent) { | |
5357 | int err; | |
5358 | ||
5359 | err = swevent_hlist_get(event); | |
5360 | if (err) | |
5361 | return err; | |
5362 | ||
c5905afb | 5363 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5364 | event->destroy = sw_perf_event_destroy; |
5365 | } | |
5366 | ||
5367 | return 0; | |
5368 | } | |
5369 | ||
35edc2a5 PZ |
5370 | static int perf_swevent_event_idx(struct perf_event *event) |
5371 | { | |
5372 | return 0; | |
5373 | } | |
5374 | ||
b0a873eb | 5375 | static struct pmu perf_swevent = { |
89a1e187 | 5376 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5377 | |
b0a873eb | 5378 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5379 | .add = perf_swevent_add, |
5380 | .del = perf_swevent_del, | |
5381 | .start = perf_swevent_start, | |
5382 | .stop = perf_swevent_stop, | |
1c024eca | 5383 | .read = perf_swevent_read, |
35edc2a5 PZ |
5384 | |
5385 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5386 | }; |
5387 | ||
b0a873eb PZ |
5388 | #ifdef CONFIG_EVENT_TRACING |
5389 | ||
1c024eca PZ |
5390 | static int perf_tp_filter_match(struct perf_event *event, |
5391 | struct perf_sample_data *data) | |
5392 | { | |
5393 | void *record = data->raw->data; | |
5394 | ||
5395 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5396 | return 1; | |
5397 | return 0; | |
5398 | } | |
5399 | ||
5400 | static int perf_tp_event_match(struct perf_event *event, | |
5401 | struct perf_sample_data *data, | |
5402 | struct pt_regs *regs) | |
5403 | { | |
a0f7d0f7 FW |
5404 | if (event->hw.state & PERF_HES_STOPPED) |
5405 | return 0; | |
580d607c PZ |
5406 | /* |
5407 | * All tracepoints are from kernel-space. | |
5408 | */ | |
5409 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5410 | return 0; |
5411 | ||
5412 | if (!perf_tp_filter_match(event, data)) | |
5413 | return 0; | |
5414 | ||
5415 | return 1; | |
5416 | } | |
5417 | ||
5418 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5419 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5420 | struct task_struct *task) | |
95476b64 FW |
5421 | { |
5422 | struct perf_sample_data data; | |
1c024eca | 5423 | struct perf_event *event; |
1c024eca | 5424 | |
95476b64 FW |
5425 | struct perf_raw_record raw = { |
5426 | .size = entry_size, | |
5427 | .data = record, | |
5428 | }; | |
5429 | ||
fd0d000b | 5430 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5431 | data.raw = &raw; |
5432 | ||
b67bfe0d | 5433 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 5434 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 5435 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5436 | } |
ecc55f84 | 5437 | |
e6dab5ff AV |
5438 | /* |
5439 | * If we got specified a target task, also iterate its context and | |
5440 | * deliver this event there too. | |
5441 | */ | |
5442 | if (task && task != current) { | |
5443 | struct perf_event_context *ctx; | |
5444 | struct trace_entry *entry = record; | |
5445 | ||
5446 | rcu_read_lock(); | |
5447 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5448 | if (!ctx) | |
5449 | goto unlock; | |
5450 | ||
5451 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5452 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5453 | continue; | |
5454 | if (event->attr.config != entry->type) | |
5455 | continue; | |
5456 | if (perf_tp_event_match(event, &data, regs)) | |
5457 | perf_swevent_event(event, count, &data, regs); | |
5458 | } | |
5459 | unlock: | |
5460 | rcu_read_unlock(); | |
5461 | } | |
5462 | ||
ecc55f84 | 5463 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5464 | } |
5465 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5466 | ||
cdd6c482 | 5467 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5468 | { |
1c024eca | 5469 | perf_trace_destroy(event); |
e077df4f PZ |
5470 | } |
5471 | ||
b0a873eb | 5472 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5473 | { |
76e1d904 FW |
5474 | int err; |
5475 | ||
b0a873eb PZ |
5476 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5477 | return -ENOENT; | |
5478 | ||
2481c5fa SE |
5479 | /* |
5480 | * no branch sampling for tracepoint events | |
5481 | */ | |
5482 | if (has_branch_stack(event)) | |
5483 | return -EOPNOTSUPP; | |
5484 | ||
1c024eca PZ |
5485 | err = perf_trace_init(event); |
5486 | if (err) | |
b0a873eb | 5487 | return err; |
e077df4f | 5488 | |
cdd6c482 | 5489 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5490 | |
b0a873eb PZ |
5491 | return 0; |
5492 | } | |
5493 | ||
5494 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5495 | .task_ctx_nr = perf_sw_context, |
5496 | ||
b0a873eb | 5497 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5498 | .add = perf_trace_add, |
5499 | .del = perf_trace_del, | |
5500 | .start = perf_swevent_start, | |
5501 | .stop = perf_swevent_stop, | |
b0a873eb | 5502 | .read = perf_swevent_read, |
35edc2a5 PZ |
5503 | |
5504 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5505 | }; |
5506 | ||
5507 | static inline void perf_tp_register(void) | |
5508 | { | |
2e80a82a | 5509 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5510 | } |
6fb2915d LZ |
5511 | |
5512 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5513 | { | |
5514 | char *filter_str; | |
5515 | int ret; | |
5516 | ||
5517 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5518 | return -EINVAL; | |
5519 | ||
5520 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5521 | if (IS_ERR(filter_str)) | |
5522 | return PTR_ERR(filter_str); | |
5523 | ||
5524 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5525 | ||
5526 | kfree(filter_str); | |
5527 | return ret; | |
5528 | } | |
5529 | ||
5530 | static void perf_event_free_filter(struct perf_event *event) | |
5531 | { | |
5532 | ftrace_profile_free_filter(event); | |
5533 | } | |
5534 | ||
e077df4f | 5535 | #else |
6fb2915d | 5536 | |
b0a873eb | 5537 | static inline void perf_tp_register(void) |
e077df4f | 5538 | { |
e077df4f | 5539 | } |
6fb2915d LZ |
5540 | |
5541 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5542 | { | |
5543 | return -ENOENT; | |
5544 | } | |
5545 | ||
5546 | static void perf_event_free_filter(struct perf_event *event) | |
5547 | { | |
5548 | } | |
5549 | ||
07b139c8 | 5550 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5551 | |
24f1e32c | 5552 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5553 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5554 | { |
f5ffe02e FW |
5555 | struct perf_sample_data sample; |
5556 | struct pt_regs *regs = data; | |
5557 | ||
fd0d000b | 5558 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 5559 | |
a4eaf7f1 | 5560 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5561 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5562 | } |
5563 | #endif | |
5564 | ||
b0a873eb PZ |
5565 | /* |
5566 | * hrtimer based swevent callback | |
5567 | */ | |
f29ac756 | 5568 | |
b0a873eb | 5569 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5570 | { |
b0a873eb PZ |
5571 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5572 | struct perf_sample_data data; | |
5573 | struct pt_regs *regs; | |
5574 | struct perf_event *event; | |
5575 | u64 period; | |
f29ac756 | 5576 | |
b0a873eb | 5577 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5578 | |
5579 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5580 | return HRTIMER_NORESTART; | |
5581 | ||
b0a873eb | 5582 | event->pmu->read(event); |
f344011c | 5583 | |
fd0d000b | 5584 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
5585 | regs = get_irq_regs(); |
5586 | ||
5587 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5588 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 5589 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
5590 | ret = HRTIMER_NORESTART; |
5591 | } | |
24f1e32c | 5592 | |
b0a873eb PZ |
5593 | period = max_t(u64, 10000, event->hw.sample_period); |
5594 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5595 | |
b0a873eb | 5596 | return ret; |
f29ac756 PZ |
5597 | } |
5598 | ||
b0a873eb | 5599 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5600 | { |
b0a873eb | 5601 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5602 | s64 period; |
5603 | ||
5604 | if (!is_sampling_event(event)) | |
5605 | return; | |
f5ffe02e | 5606 | |
5d508e82 FBH |
5607 | period = local64_read(&hwc->period_left); |
5608 | if (period) { | |
5609 | if (period < 0) | |
5610 | period = 10000; | |
fa407f35 | 5611 | |
5d508e82 FBH |
5612 | local64_set(&hwc->period_left, 0); |
5613 | } else { | |
5614 | period = max_t(u64, 10000, hwc->sample_period); | |
5615 | } | |
5616 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 5617 | ns_to_ktime(period), 0, |
b5ab4cd5 | 5618 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 5619 | } |
b0a873eb PZ |
5620 | |
5621 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 5622 | { |
b0a873eb PZ |
5623 | struct hw_perf_event *hwc = &event->hw; |
5624 | ||
6c7e550f | 5625 | if (is_sampling_event(event)) { |
b0a873eb | 5626 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 5627 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
5628 | |
5629 | hrtimer_cancel(&hwc->hrtimer); | |
5630 | } | |
24f1e32c FW |
5631 | } |
5632 | ||
ba3dd36c PZ |
5633 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5634 | { | |
5635 | struct hw_perf_event *hwc = &event->hw; | |
5636 | ||
5637 | if (!is_sampling_event(event)) | |
5638 | return; | |
5639 | ||
5640 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5641 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
5642 | ||
5643 | /* | |
5644 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
5645 | * mapping and avoid the whole period adjust feedback stuff. | |
5646 | */ | |
5647 | if (event->attr.freq) { | |
5648 | long freq = event->attr.sample_freq; | |
5649 | ||
5650 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
5651 | hwc->sample_period = event->attr.sample_period; | |
5652 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 5653 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
5654 | event->attr.freq = 0; |
5655 | } | |
5656 | } | |
5657 | ||
b0a873eb PZ |
5658 | /* |
5659 | * Software event: cpu wall time clock | |
5660 | */ | |
5661 | ||
5662 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 5663 | { |
b0a873eb PZ |
5664 | s64 prev; |
5665 | u64 now; | |
5666 | ||
a4eaf7f1 | 5667 | now = local_clock(); |
b0a873eb PZ |
5668 | prev = local64_xchg(&event->hw.prev_count, now); |
5669 | local64_add(now - prev, &event->count); | |
24f1e32c | 5670 | } |
24f1e32c | 5671 | |
a4eaf7f1 | 5672 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5673 | { |
a4eaf7f1 | 5674 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 5675 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5676 | } |
5677 | ||
a4eaf7f1 | 5678 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 5679 | { |
b0a873eb PZ |
5680 | perf_swevent_cancel_hrtimer(event); |
5681 | cpu_clock_event_update(event); | |
5682 | } | |
f29ac756 | 5683 | |
a4eaf7f1 PZ |
5684 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
5685 | { | |
5686 | if (flags & PERF_EF_START) | |
5687 | cpu_clock_event_start(event, flags); | |
5688 | ||
5689 | return 0; | |
5690 | } | |
5691 | ||
5692 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
5693 | { | |
5694 | cpu_clock_event_stop(event, flags); | |
5695 | } | |
5696 | ||
b0a873eb PZ |
5697 | static void cpu_clock_event_read(struct perf_event *event) |
5698 | { | |
5699 | cpu_clock_event_update(event); | |
5700 | } | |
f344011c | 5701 | |
b0a873eb PZ |
5702 | static int cpu_clock_event_init(struct perf_event *event) |
5703 | { | |
5704 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5705 | return -ENOENT; | |
5706 | ||
5707 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
5708 | return -ENOENT; | |
5709 | ||
2481c5fa SE |
5710 | /* |
5711 | * no branch sampling for software events | |
5712 | */ | |
5713 | if (has_branch_stack(event)) | |
5714 | return -EOPNOTSUPP; | |
5715 | ||
ba3dd36c PZ |
5716 | perf_swevent_init_hrtimer(event); |
5717 | ||
b0a873eb | 5718 | return 0; |
f29ac756 PZ |
5719 | } |
5720 | ||
b0a873eb | 5721 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
5722 | .task_ctx_nr = perf_sw_context, |
5723 | ||
b0a873eb | 5724 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
5725 | .add = cpu_clock_event_add, |
5726 | .del = cpu_clock_event_del, | |
5727 | .start = cpu_clock_event_start, | |
5728 | .stop = cpu_clock_event_stop, | |
b0a873eb | 5729 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
5730 | |
5731 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5732 | }; |
5733 | ||
5734 | /* | |
5735 | * Software event: task time clock | |
5736 | */ | |
5737 | ||
5738 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 5739 | { |
b0a873eb PZ |
5740 | u64 prev; |
5741 | s64 delta; | |
5c92d124 | 5742 | |
b0a873eb PZ |
5743 | prev = local64_xchg(&event->hw.prev_count, now); |
5744 | delta = now - prev; | |
5745 | local64_add(delta, &event->count); | |
5746 | } | |
5c92d124 | 5747 | |
a4eaf7f1 | 5748 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5749 | { |
a4eaf7f1 | 5750 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 5751 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5752 | } |
5753 | ||
a4eaf7f1 | 5754 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
5755 | { |
5756 | perf_swevent_cancel_hrtimer(event); | |
5757 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
5758 | } |
5759 | ||
5760 | static int task_clock_event_add(struct perf_event *event, int flags) | |
5761 | { | |
5762 | if (flags & PERF_EF_START) | |
5763 | task_clock_event_start(event, flags); | |
b0a873eb | 5764 | |
a4eaf7f1 PZ |
5765 | return 0; |
5766 | } | |
5767 | ||
5768 | static void task_clock_event_del(struct perf_event *event, int flags) | |
5769 | { | |
5770 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
5771 | } |
5772 | ||
5773 | static void task_clock_event_read(struct perf_event *event) | |
5774 | { | |
768a06e2 PZ |
5775 | u64 now = perf_clock(); |
5776 | u64 delta = now - event->ctx->timestamp; | |
5777 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
5778 | |
5779 | task_clock_event_update(event, time); | |
5780 | } | |
5781 | ||
5782 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 5783 | { |
b0a873eb PZ |
5784 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5785 | return -ENOENT; | |
5786 | ||
5787 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
5788 | return -ENOENT; | |
5789 | ||
2481c5fa SE |
5790 | /* |
5791 | * no branch sampling for software events | |
5792 | */ | |
5793 | if (has_branch_stack(event)) | |
5794 | return -EOPNOTSUPP; | |
5795 | ||
ba3dd36c PZ |
5796 | perf_swevent_init_hrtimer(event); |
5797 | ||
b0a873eb | 5798 | return 0; |
6fb2915d LZ |
5799 | } |
5800 | ||
b0a873eb | 5801 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
5802 | .task_ctx_nr = perf_sw_context, |
5803 | ||
b0a873eb | 5804 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
5805 | .add = task_clock_event_add, |
5806 | .del = task_clock_event_del, | |
5807 | .start = task_clock_event_start, | |
5808 | .stop = task_clock_event_stop, | |
b0a873eb | 5809 | .read = task_clock_event_read, |
35edc2a5 PZ |
5810 | |
5811 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 5812 | }; |
6fb2915d | 5813 | |
ad5133b7 | 5814 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 5815 | { |
e077df4f | 5816 | } |
6fb2915d | 5817 | |
ad5133b7 | 5818 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 5819 | { |
ad5133b7 | 5820 | return 0; |
6fb2915d LZ |
5821 | } |
5822 | ||
ad5133b7 | 5823 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 5824 | { |
ad5133b7 | 5825 | perf_pmu_disable(pmu); |
6fb2915d LZ |
5826 | } |
5827 | ||
ad5133b7 PZ |
5828 | static int perf_pmu_commit_txn(struct pmu *pmu) |
5829 | { | |
5830 | perf_pmu_enable(pmu); | |
5831 | return 0; | |
5832 | } | |
e077df4f | 5833 | |
ad5133b7 | 5834 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 5835 | { |
ad5133b7 | 5836 | perf_pmu_enable(pmu); |
24f1e32c FW |
5837 | } |
5838 | ||
35edc2a5 PZ |
5839 | static int perf_event_idx_default(struct perf_event *event) |
5840 | { | |
5841 | return event->hw.idx + 1; | |
5842 | } | |
5843 | ||
8dc85d54 PZ |
5844 | /* |
5845 | * Ensures all contexts with the same task_ctx_nr have the same | |
5846 | * pmu_cpu_context too. | |
5847 | */ | |
5848 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 5849 | { |
8dc85d54 | 5850 | struct pmu *pmu; |
b326e956 | 5851 | |
8dc85d54 PZ |
5852 | if (ctxn < 0) |
5853 | return NULL; | |
24f1e32c | 5854 | |
8dc85d54 PZ |
5855 | list_for_each_entry(pmu, &pmus, entry) { |
5856 | if (pmu->task_ctx_nr == ctxn) | |
5857 | return pmu->pmu_cpu_context; | |
5858 | } | |
24f1e32c | 5859 | |
8dc85d54 | 5860 | return NULL; |
24f1e32c FW |
5861 | } |
5862 | ||
51676957 | 5863 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 5864 | { |
51676957 PZ |
5865 | int cpu; |
5866 | ||
5867 | for_each_possible_cpu(cpu) { | |
5868 | struct perf_cpu_context *cpuctx; | |
5869 | ||
5870 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
5871 | ||
3f1f3320 PZ |
5872 | if (cpuctx->unique_pmu == old_pmu) |
5873 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
5874 | } |
5875 | } | |
5876 | ||
5877 | static void free_pmu_context(struct pmu *pmu) | |
5878 | { | |
5879 | struct pmu *i; | |
f5ffe02e | 5880 | |
8dc85d54 | 5881 | mutex_lock(&pmus_lock); |
0475f9ea | 5882 | /* |
8dc85d54 | 5883 | * Like a real lame refcount. |
0475f9ea | 5884 | */ |
51676957 PZ |
5885 | list_for_each_entry(i, &pmus, entry) { |
5886 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
5887 | update_pmu_context(i, pmu); | |
8dc85d54 | 5888 | goto out; |
51676957 | 5889 | } |
8dc85d54 | 5890 | } |
d6d020e9 | 5891 | |
51676957 | 5892 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
5893 | out: |
5894 | mutex_unlock(&pmus_lock); | |
24f1e32c | 5895 | } |
2e80a82a | 5896 | static struct idr pmu_idr; |
d6d020e9 | 5897 | |
abe43400 PZ |
5898 | static ssize_t |
5899 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
5900 | { | |
5901 | struct pmu *pmu = dev_get_drvdata(dev); | |
5902 | ||
5903 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
5904 | } | |
5905 | ||
5906 | static struct device_attribute pmu_dev_attrs[] = { | |
5907 | __ATTR_RO(type), | |
5908 | __ATTR_NULL, | |
5909 | }; | |
5910 | ||
5911 | static int pmu_bus_running; | |
5912 | static struct bus_type pmu_bus = { | |
5913 | .name = "event_source", | |
5914 | .dev_attrs = pmu_dev_attrs, | |
5915 | }; | |
5916 | ||
5917 | static void pmu_dev_release(struct device *dev) | |
5918 | { | |
5919 | kfree(dev); | |
5920 | } | |
5921 | ||
5922 | static int pmu_dev_alloc(struct pmu *pmu) | |
5923 | { | |
5924 | int ret = -ENOMEM; | |
5925 | ||
5926 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
5927 | if (!pmu->dev) | |
5928 | goto out; | |
5929 | ||
0c9d42ed | 5930 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
5931 | device_initialize(pmu->dev); |
5932 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
5933 | if (ret) | |
5934 | goto free_dev; | |
5935 | ||
5936 | dev_set_drvdata(pmu->dev, pmu); | |
5937 | pmu->dev->bus = &pmu_bus; | |
5938 | pmu->dev->release = pmu_dev_release; | |
5939 | ret = device_add(pmu->dev); | |
5940 | if (ret) | |
5941 | goto free_dev; | |
5942 | ||
5943 | out: | |
5944 | return ret; | |
5945 | ||
5946 | free_dev: | |
5947 | put_device(pmu->dev); | |
5948 | goto out; | |
5949 | } | |
5950 | ||
547e9fd7 | 5951 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 5952 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 5953 | |
2e80a82a | 5954 | int perf_pmu_register(struct pmu *pmu, char *name, int type) |
24f1e32c | 5955 | { |
108b02cf | 5956 | int cpu, ret; |
24f1e32c | 5957 | |
b0a873eb | 5958 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
5959 | ret = -ENOMEM; |
5960 | pmu->pmu_disable_count = alloc_percpu(int); | |
5961 | if (!pmu->pmu_disable_count) | |
5962 | goto unlock; | |
f29ac756 | 5963 | |
2e80a82a PZ |
5964 | pmu->type = -1; |
5965 | if (!name) | |
5966 | goto skip_type; | |
5967 | pmu->name = name; | |
5968 | ||
5969 | if (type < 0) { | |
0e9c3be2 TH |
5970 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
5971 | if (type < 0) { | |
5972 | ret = type; | |
2e80a82a PZ |
5973 | goto free_pdc; |
5974 | } | |
5975 | } | |
5976 | pmu->type = type; | |
5977 | ||
abe43400 PZ |
5978 | if (pmu_bus_running) { |
5979 | ret = pmu_dev_alloc(pmu); | |
5980 | if (ret) | |
5981 | goto free_idr; | |
5982 | } | |
5983 | ||
2e80a82a | 5984 | skip_type: |
8dc85d54 PZ |
5985 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
5986 | if (pmu->pmu_cpu_context) | |
5987 | goto got_cpu_context; | |
f29ac756 | 5988 | |
108b02cf PZ |
5989 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
5990 | if (!pmu->pmu_cpu_context) | |
abe43400 | 5991 | goto free_dev; |
f344011c | 5992 | |
108b02cf PZ |
5993 | for_each_possible_cpu(cpu) { |
5994 | struct perf_cpu_context *cpuctx; | |
5995 | ||
5996 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 5997 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 5998 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 5999 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6000 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6001 | cpuctx->ctx.pmu = pmu; |
e9d2b064 PZ |
6002 | cpuctx->jiffies_interval = 1; |
6003 | INIT_LIST_HEAD(&cpuctx->rotation_list); | |
3f1f3320 | 6004 | cpuctx->unique_pmu = pmu; |
108b02cf | 6005 | } |
76e1d904 | 6006 | |
8dc85d54 | 6007 | got_cpu_context: |
ad5133b7 PZ |
6008 | if (!pmu->start_txn) { |
6009 | if (pmu->pmu_enable) { | |
6010 | /* | |
6011 | * If we have pmu_enable/pmu_disable calls, install | |
6012 | * transaction stubs that use that to try and batch | |
6013 | * hardware accesses. | |
6014 | */ | |
6015 | pmu->start_txn = perf_pmu_start_txn; | |
6016 | pmu->commit_txn = perf_pmu_commit_txn; | |
6017 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6018 | } else { | |
6019 | pmu->start_txn = perf_pmu_nop_void; | |
6020 | pmu->commit_txn = perf_pmu_nop_int; | |
6021 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6022 | } |
5c92d124 | 6023 | } |
15dbf27c | 6024 | |
ad5133b7 PZ |
6025 | if (!pmu->pmu_enable) { |
6026 | pmu->pmu_enable = perf_pmu_nop_void; | |
6027 | pmu->pmu_disable = perf_pmu_nop_void; | |
6028 | } | |
6029 | ||
35edc2a5 PZ |
6030 | if (!pmu->event_idx) |
6031 | pmu->event_idx = perf_event_idx_default; | |
6032 | ||
b0a873eb | 6033 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6034 | ret = 0; |
6035 | unlock: | |
b0a873eb PZ |
6036 | mutex_unlock(&pmus_lock); |
6037 | ||
33696fc0 | 6038 | return ret; |
108b02cf | 6039 | |
abe43400 PZ |
6040 | free_dev: |
6041 | device_del(pmu->dev); | |
6042 | put_device(pmu->dev); | |
6043 | ||
2e80a82a PZ |
6044 | free_idr: |
6045 | if (pmu->type >= PERF_TYPE_MAX) | |
6046 | idr_remove(&pmu_idr, pmu->type); | |
6047 | ||
108b02cf PZ |
6048 | free_pdc: |
6049 | free_percpu(pmu->pmu_disable_count); | |
6050 | goto unlock; | |
f29ac756 PZ |
6051 | } |
6052 | ||
b0a873eb | 6053 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6054 | { |
b0a873eb PZ |
6055 | mutex_lock(&pmus_lock); |
6056 | list_del_rcu(&pmu->entry); | |
6057 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6058 | |
0475f9ea | 6059 | /* |
cde8e884 PZ |
6060 | * We dereference the pmu list under both SRCU and regular RCU, so |
6061 | * synchronize against both of those. | |
0475f9ea | 6062 | */ |
b0a873eb | 6063 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6064 | synchronize_rcu(); |
d6d020e9 | 6065 | |
33696fc0 | 6066 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6067 | if (pmu->type >= PERF_TYPE_MAX) |
6068 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6069 | device_del(pmu->dev); |
6070 | put_device(pmu->dev); | |
51676957 | 6071 | free_pmu_context(pmu); |
b0a873eb | 6072 | } |
d6d020e9 | 6073 | |
b0a873eb PZ |
6074 | struct pmu *perf_init_event(struct perf_event *event) |
6075 | { | |
6076 | struct pmu *pmu = NULL; | |
6077 | int idx; | |
940c5b29 | 6078 | int ret; |
b0a873eb PZ |
6079 | |
6080 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6081 | |
6082 | rcu_read_lock(); | |
6083 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6084 | rcu_read_unlock(); | |
940c5b29 | 6085 | if (pmu) { |
7e5b2a01 | 6086 | event->pmu = pmu; |
940c5b29 LM |
6087 | ret = pmu->event_init(event); |
6088 | if (ret) | |
6089 | pmu = ERR_PTR(ret); | |
2e80a82a | 6090 | goto unlock; |
940c5b29 | 6091 | } |
2e80a82a | 6092 | |
b0a873eb | 6093 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 6094 | event->pmu = pmu; |
940c5b29 | 6095 | ret = pmu->event_init(event); |
b0a873eb | 6096 | if (!ret) |
e5f4d339 | 6097 | goto unlock; |
76e1d904 | 6098 | |
b0a873eb PZ |
6099 | if (ret != -ENOENT) { |
6100 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6101 | goto unlock; |
f344011c | 6102 | } |
5c92d124 | 6103 | } |
e5f4d339 PZ |
6104 | pmu = ERR_PTR(-ENOENT); |
6105 | unlock: | |
b0a873eb | 6106 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6107 | |
4aeb0b42 | 6108 | return pmu; |
5c92d124 IM |
6109 | } |
6110 | ||
0793a61d | 6111 | /* |
cdd6c482 | 6112 | * Allocate and initialize a event structure |
0793a61d | 6113 | */ |
cdd6c482 | 6114 | static struct perf_event * |
c3f00c70 | 6115 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6116 | struct task_struct *task, |
6117 | struct perf_event *group_leader, | |
6118 | struct perf_event *parent_event, | |
4dc0da86 AK |
6119 | perf_overflow_handler_t overflow_handler, |
6120 | void *context) | |
0793a61d | 6121 | { |
51b0fe39 | 6122 | struct pmu *pmu; |
cdd6c482 IM |
6123 | struct perf_event *event; |
6124 | struct hw_perf_event *hwc; | |
d5d2bc0d | 6125 | long err; |
0793a61d | 6126 | |
66832eb4 ON |
6127 | if ((unsigned)cpu >= nr_cpu_ids) { |
6128 | if (!task || cpu != -1) | |
6129 | return ERR_PTR(-EINVAL); | |
6130 | } | |
6131 | ||
c3f00c70 | 6132 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6133 | if (!event) |
d5d2bc0d | 6134 | return ERR_PTR(-ENOMEM); |
0793a61d | 6135 | |
04289bb9 | 6136 | /* |
cdd6c482 | 6137 | * Single events are their own group leaders, with an |
04289bb9 IM |
6138 | * empty sibling list: |
6139 | */ | |
6140 | if (!group_leader) | |
cdd6c482 | 6141 | group_leader = event; |
04289bb9 | 6142 | |
cdd6c482 IM |
6143 | mutex_init(&event->child_mutex); |
6144 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6145 | |
cdd6c482 IM |
6146 | INIT_LIST_HEAD(&event->group_entry); |
6147 | INIT_LIST_HEAD(&event->event_entry); | |
6148 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
6149 | INIT_LIST_HEAD(&event->rb_entry); |
6150 | ||
cdd6c482 | 6151 | init_waitqueue_head(&event->waitq); |
e360adbe | 6152 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6153 | |
cdd6c482 | 6154 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6155 | |
a6fa941d | 6156 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6157 | event->cpu = cpu; |
6158 | event->attr = *attr; | |
6159 | event->group_leader = group_leader; | |
6160 | event->pmu = NULL; | |
cdd6c482 | 6161 | event->oncpu = -1; |
a96bbc16 | 6162 | |
cdd6c482 | 6163 | event->parent = parent_event; |
b84fbc9f | 6164 | |
17cf22c3 | 6165 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6166 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6167 | |
cdd6c482 | 6168 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6169 | |
d580ff86 PZ |
6170 | if (task) { |
6171 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6172 | |
6173 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6174 | event->hw.tp_target = task; | |
d580ff86 PZ |
6175 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6176 | /* | |
6177 | * hw_breakpoint is a bit difficult here.. | |
6178 | */ | |
f22c1bb6 | 6179 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6180 | event->hw.bp_target = task; |
6181 | #endif | |
6182 | } | |
6183 | ||
4dc0da86 | 6184 | if (!overflow_handler && parent_event) { |
b326e956 | 6185 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6186 | context = parent_event->overflow_handler_context; |
6187 | } | |
66832eb4 | 6188 | |
b326e956 | 6189 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6190 | event->overflow_handler_context = context; |
97eaf530 | 6191 | |
0231bb53 | 6192 | perf_event__state_init(event); |
a86ed508 | 6193 | |
4aeb0b42 | 6194 | pmu = NULL; |
b8e83514 | 6195 | |
cdd6c482 | 6196 | hwc = &event->hw; |
bd2b5b12 | 6197 | hwc->sample_period = attr->sample_period; |
0d48696f | 6198 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6199 | hwc->sample_period = 1; |
eced1dfc | 6200 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6201 | |
e7850595 | 6202 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6203 | |
2023b359 | 6204 | /* |
cdd6c482 | 6205 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6206 | */ |
3dab77fb | 6207 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
2023b359 PZ |
6208 | goto done; |
6209 | ||
b0a873eb | 6210 | pmu = perf_init_event(event); |
974802ea | 6211 | |
d5d2bc0d PM |
6212 | done: |
6213 | err = 0; | |
4aeb0b42 | 6214 | if (!pmu) |
d5d2bc0d | 6215 | err = -EINVAL; |
4aeb0b42 RR |
6216 | else if (IS_ERR(pmu)) |
6217 | err = PTR_ERR(pmu); | |
5c92d124 | 6218 | |
d5d2bc0d | 6219 | if (err) { |
cdd6c482 IM |
6220 | if (event->ns) |
6221 | put_pid_ns(event->ns); | |
6222 | kfree(event); | |
d5d2bc0d | 6223 | return ERR_PTR(err); |
621a01ea | 6224 | } |
d5d2bc0d | 6225 | |
cdd6c482 | 6226 | if (!event->parent) { |
82cd6def | 6227 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 6228 | static_key_slow_inc(&perf_sched_events.key); |
3af9e859 | 6229 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
6230 | atomic_inc(&nr_mmap_events); |
6231 | if (event->attr.comm) | |
6232 | atomic_inc(&nr_comm_events); | |
6233 | if (event->attr.task) | |
6234 | atomic_inc(&nr_task_events); | |
927c7a9e FW |
6235 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6236 | err = get_callchain_buffers(); | |
6237 | if (err) { | |
6238 | free_event(event); | |
6239 | return ERR_PTR(err); | |
6240 | } | |
6241 | } | |
d010b332 SE |
6242 | if (has_branch_stack(event)) { |
6243 | static_key_slow_inc(&perf_sched_events.key); | |
6244 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6245 | atomic_inc(&per_cpu(perf_branch_stack_events, | |
6246 | event->cpu)); | |
6247 | } | |
f344011c | 6248 | } |
9ee318a7 | 6249 | |
cdd6c482 | 6250 | return event; |
0793a61d TG |
6251 | } |
6252 | ||
cdd6c482 IM |
6253 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6254 | struct perf_event_attr *attr) | |
974802ea | 6255 | { |
974802ea | 6256 | u32 size; |
cdf8073d | 6257 | int ret; |
974802ea PZ |
6258 | |
6259 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6260 | return -EFAULT; | |
6261 | ||
6262 | /* | |
6263 | * zero the full structure, so that a short copy will be nice. | |
6264 | */ | |
6265 | memset(attr, 0, sizeof(*attr)); | |
6266 | ||
6267 | ret = get_user(size, &uattr->size); | |
6268 | if (ret) | |
6269 | return ret; | |
6270 | ||
6271 | if (size > PAGE_SIZE) /* silly large */ | |
6272 | goto err_size; | |
6273 | ||
6274 | if (!size) /* abi compat */ | |
6275 | size = PERF_ATTR_SIZE_VER0; | |
6276 | ||
6277 | if (size < PERF_ATTR_SIZE_VER0) | |
6278 | goto err_size; | |
6279 | ||
6280 | /* | |
6281 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6282 | * ensure all the unknown bits are 0 - i.e. new |
6283 | * user-space does not rely on any kernel feature | |
6284 | * extensions we dont know about yet. | |
974802ea PZ |
6285 | */ |
6286 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6287 | unsigned char __user *addr; |
6288 | unsigned char __user *end; | |
6289 | unsigned char val; | |
974802ea | 6290 | |
cdf8073d IS |
6291 | addr = (void __user *)uattr + sizeof(*attr); |
6292 | end = (void __user *)uattr + size; | |
974802ea | 6293 | |
cdf8073d | 6294 | for (; addr < end; addr++) { |
974802ea PZ |
6295 | ret = get_user(val, addr); |
6296 | if (ret) | |
6297 | return ret; | |
6298 | if (val) | |
6299 | goto err_size; | |
6300 | } | |
b3e62e35 | 6301 | size = sizeof(*attr); |
974802ea PZ |
6302 | } |
6303 | ||
6304 | ret = copy_from_user(attr, uattr, size); | |
6305 | if (ret) | |
6306 | return -EFAULT; | |
6307 | ||
cd757645 | 6308 | if (attr->__reserved_1) |
974802ea PZ |
6309 | return -EINVAL; |
6310 | ||
6311 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6312 | return -EINVAL; | |
6313 | ||
6314 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6315 | return -EINVAL; | |
6316 | ||
bce38cd5 SE |
6317 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6318 | u64 mask = attr->branch_sample_type; | |
6319 | ||
6320 | /* only using defined bits */ | |
6321 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6322 | return -EINVAL; | |
6323 | ||
6324 | /* at least one branch bit must be set */ | |
6325 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6326 | return -EINVAL; | |
6327 | ||
6328 | /* kernel level capture: check permissions */ | |
6329 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
6330 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6331 | return -EACCES; | |
6332 | ||
6333 | /* propagate priv level, when not set for branch */ | |
6334 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6335 | ||
6336 | /* exclude_kernel checked on syscall entry */ | |
6337 | if (!attr->exclude_kernel) | |
6338 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6339 | ||
6340 | if (!attr->exclude_user) | |
6341 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6342 | ||
6343 | if (!attr->exclude_hv) | |
6344 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6345 | /* | |
6346 | * adjust user setting (for HW filter setup) | |
6347 | */ | |
6348 | attr->branch_sample_type = mask; | |
6349 | } | |
6350 | } | |
4018994f | 6351 | |
c5ebcedb | 6352 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 6353 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
6354 | if (ret) |
6355 | return ret; | |
6356 | } | |
6357 | ||
6358 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
6359 | if (!arch_perf_have_user_stack_dump()) | |
6360 | return -ENOSYS; | |
6361 | ||
6362 | /* | |
6363 | * We have __u32 type for the size, but so far | |
6364 | * we can only use __u16 as maximum due to the | |
6365 | * __u16 sample size limit. | |
6366 | */ | |
6367 | if (attr->sample_stack_user >= USHRT_MAX) | |
6368 | ret = -EINVAL; | |
6369 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
6370 | ret = -EINVAL; | |
6371 | } | |
4018994f | 6372 | |
974802ea PZ |
6373 | out: |
6374 | return ret; | |
6375 | ||
6376 | err_size: | |
6377 | put_user(sizeof(*attr), &uattr->size); | |
6378 | ret = -E2BIG; | |
6379 | goto out; | |
6380 | } | |
6381 | ||
ac9721f3 PZ |
6382 | static int |
6383 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6384 | { |
76369139 | 6385 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6386 | int ret = -EINVAL; |
6387 | ||
ac9721f3 | 6388 | if (!output_event) |
a4be7c27 PZ |
6389 | goto set; |
6390 | ||
ac9721f3 PZ |
6391 | /* don't allow circular references */ |
6392 | if (event == output_event) | |
a4be7c27 PZ |
6393 | goto out; |
6394 | ||
0f139300 PZ |
6395 | /* |
6396 | * Don't allow cross-cpu buffers | |
6397 | */ | |
6398 | if (output_event->cpu != event->cpu) | |
6399 | goto out; | |
6400 | ||
6401 | /* | |
76369139 | 6402 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6403 | */ |
6404 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6405 | goto out; | |
6406 | ||
a4be7c27 | 6407 | set: |
cdd6c482 | 6408 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6409 | /* Can't redirect output if we've got an active mmap() */ |
6410 | if (atomic_read(&event->mmap_count)) | |
6411 | goto unlock; | |
a4be7c27 | 6412 | |
ac9721f3 | 6413 | if (output_event) { |
76369139 FW |
6414 | /* get the rb we want to redirect to */ |
6415 | rb = ring_buffer_get(output_event); | |
6416 | if (!rb) | |
ac9721f3 | 6417 | goto unlock; |
a4be7c27 PZ |
6418 | } |
6419 | ||
76369139 FW |
6420 | old_rb = event->rb; |
6421 | rcu_assign_pointer(event->rb, rb); | |
10c6db11 PZ |
6422 | if (old_rb) |
6423 | ring_buffer_detach(event, old_rb); | |
a4be7c27 | 6424 | ret = 0; |
ac9721f3 PZ |
6425 | unlock: |
6426 | mutex_unlock(&event->mmap_mutex); | |
6427 | ||
76369139 FW |
6428 | if (old_rb) |
6429 | ring_buffer_put(old_rb); | |
a4be7c27 | 6430 | out: |
a4be7c27 PZ |
6431 | return ret; |
6432 | } | |
6433 | ||
0793a61d | 6434 | /** |
cdd6c482 | 6435 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6436 | * |
cdd6c482 | 6437 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6438 | * @pid: target pid |
9f66a381 | 6439 | * @cpu: target cpu |
cdd6c482 | 6440 | * @group_fd: group leader event fd |
0793a61d | 6441 | */ |
cdd6c482 IM |
6442 | SYSCALL_DEFINE5(perf_event_open, |
6443 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 6444 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 6445 | { |
b04243ef PZ |
6446 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6447 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
6448 | struct perf_event_attr attr; |
6449 | struct perf_event_context *ctx; | |
6450 | struct file *event_file = NULL; | |
2903ff01 | 6451 | struct fd group = {NULL, 0}; |
38a81da2 | 6452 | struct task_struct *task = NULL; |
89a1e187 | 6453 | struct pmu *pmu; |
ea635c64 | 6454 | int event_fd; |
b04243ef | 6455 | int move_group = 0; |
dc86cabe | 6456 | int err; |
0793a61d | 6457 | |
2743a5b0 | 6458 | /* for future expandability... */ |
e5d1367f | 6459 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
6460 | return -EINVAL; |
6461 | ||
dc86cabe IM |
6462 | err = perf_copy_attr(attr_uptr, &attr); |
6463 | if (err) | |
6464 | return err; | |
eab656ae | 6465 | |
0764771d PZ |
6466 | if (!attr.exclude_kernel) { |
6467 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6468 | return -EACCES; | |
6469 | } | |
6470 | ||
df58ab24 | 6471 | if (attr.freq) { |
cdd6c482 | 6472 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
6473 | return -EINVAL; |
6474 | } | |
6475 | ||
e5d1367f SE |
6476 | /* |
6477 | * In cgroup mode, the pid argument is used to pass the fd | |
6478 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6479 | * designates the cpu on which to monitor threads from that | |
6480 | * cgroup. | |
6481 | */ | |
6482 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6483 | return -EINVAL; | |
6484 | ||
ab72a702 | 6485 | event_fd = get_unused_fd(); |
ea635c64 AV |
6486 | if (event_fd < 0) |
6487 | return event_fd; | |
6488 | ||
ac9721f3 | 6489 | if (group_fd != -1) { |
2903ff01 AV |
6490 | err = perf_fget_light(group_fd, &group); |
6491 | if (err) | |
d14b12d7 | 6492 | goto err_fd; |
2903ff01 | 6493 | group_leader = group.file->private_data; |
ac9721f3 PZ |
6494 | if (flags & PERF_FLAG_FD_OUTPUT) |
6495 | output_event = group_leader; | |
6496 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6497 | group_leader = NULL; | |
6498 | } | |
6499 | ||
e5d1367f | 6500 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6501 | task = find_lively_task_by_vpid(pid); |
6502 | if (IS_ERR(task)) { | |
6503 | err = PTR_ERR(task); | |
6504 | goto err_group_fd; | |
6505 | } | |
6506 | } | |
6507 | ||
fbfc623f YZ |
6508 | get_online_cpus(); |
6509 | ||
4dc0da86 AK |
6510 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6511 | NULL, NULL); | |
d14b12d7 SE |
6512 | if (IS_ERR(event)) { |
6513 | err = PTR_ERR(event); | |
c6be5a5c | 6514 | goto err_task; |
d14b12d7 SE |
6515 | } |
6516 | ||
e5d1367f SE |
6517 | if (flags & PERF_FLAG_PID_CGROUP) { |
6518 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
6519 | if (err) | |
6520 | goto err_alloc; | |
08309379 PZ |
6521 | /* |
6522 | * one more event: | |
6523 | * - that has cgroup constraint on event->cpu | |
6524 | * - that may need work on context switch | |
6525 | */ | |
6526 | atomic_inc(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 6527 | static_key_slow_inc(&perf_sched_events.key); |
e5d1367f SE |
6528 | } |
6529 | ||
89a1e187 PZ |
6530 | /* |
6531 | * Special case software events and allow them to be part of | |
6532 | * any hardware group. | |
6533 | */ | |
6534 | pmu = event->pmu; | |
b04243ef PZ |
6535 | |
6536 | if (group_leader && | |
6537 | (is_software_event(event) != is_software_event(group_leader))) { | |
6538 | if (is_software_event(event)) { | |
6539 | /* | |
6540 | * If event and group_leader are not both a software | |
6541 | * event, and event is, then group leader is not. | |
6542 | * | |
6543 | * Allow the addition of software events to !software | |
6544 | * groups, this is safe because software events never | |
6545 | * fail to schedule. | |
6546 | */ | |
6547 | pmu = group_leader->pmu; | |
6548 | } else if (is_software_event(group_leader) && | |
6549 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
6550 | /* | |
6551 | * In case the group is a pure software group, and we | |
6552 | * try to add a hardware event, move the whole group to | |
6553 | * the hardware context. | |
6554 | */ | |
6555 | move_group = 1; | |
6556 | } | |
6557 | } | |
89a1e187 PZ |
6558 | |
6559 | /* | |
6560 | * Get the target context (task or percpu): | |
6561 | */ | |
e2d37cd2 | 6562 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
6563 | if (IS_ERR(ctx)) { |
6564 | err = PTR_ERR(ctx); | |
c6be5a5c | 6565 | goto err_alloc; |
89a1e187 PZ |
6566 | } |
6567 | ||
fd1edb3a PZ |
6568 | if (task) { |
6569 | put_task_struct(task); | |
6570 | task = NULL; | |
6571 | } | |
6572 | ||
ccff286d | 6573 | /* |
cdd6c482 | 6574 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 6575 | */ |
ac9721f3 | 6576 | if (group_leader) { |
dc86cabe | 6577 | err = -EINVAL; |
04289bb9 | 6578 | |
04289bb9 | 6579 | /* |
ccff286d IM |
6580 | * Do not allow a recursive hierarchy (this new sibling |
6581 | * becoming part of another group-sibling): | |
6582 | */ | |
6583 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 6584 | goto err_context; |
ccff286d IM |
6585 | /* |
6586 | * Do not allow to attach to a group in a different | |
6587 | * task or CPU context: | |
04289bb9 | 6588 | */ |
b04243ef PZ |
6589 | if (move_group) { |
6590 | if (group_leader->ctx->type != ctx->type) | |
6591 | goto err_context; | |
6592 | } else { | |
6593 | if (group_leader->ctx != ctx) | |
6594 | goto err_context; | |
6595 | } | |
6596 | ||
3b6f9e5c PM |
6597 | /* |
6598 | * Only a group leader can be exclusive or pinned | |
6599 | */ | |
0d48696f | 6600 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 6601 | goto err_context; |
ac9721f3 PZ |
6602 | } |
6603 | ||
6604 | if (output_event) { | |
6605 | err = perf_event_set_output(event, output_event); | |
6606 | if (err) | |
c3f00c70 | 6607 | goto err_context; |
ac9721f3 | 6608 | } |
0793a61d | 6609 | |
ea635c64 AV |
6610 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
6611 | if (IS_ERR(event_file)) { | |
6612 | err = PTR_ERR(event_file); | |
c3f00c70 | 6613 | goto err_context; |
ea635c64 | 6614 | } |
9b51f66d | 6615 | |
b04243ef PZ |
6616 | if (move_group) { |
6617 | struct perf_event_context *gctx = group_leader->ctx; | |
6618 | ||
6619 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 6620 | perf_remove_from_context(group_leader); |
0231bb53 JO |
6621 | |
6622 | /* | |
6623 | * Removing from the context ends up with disabled | |
6624 | * event. What we want here is event in the initial | |
6625 | * startup state, ready to be add into new context. | |
6626 | */ | |
6627 | perf_event__state_init(group_leader); | |
b04243ef PZ |
6628 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6629 | group_entry) { | |
fe4b04fa | 6630 | perf_remove_from_context(sibling); |
0231bb53 | 6631 | perf_event__state_init(sibling); |
b04243ef PZ |
6632 | put_ctx(gctx); |
6633 | } | |
6634 | mutex_unlock(&gctx->mutex); | |
6635 | put_ctx(gctx); | |
ea635c64 | 6636 | } |
9b51f66d | 6637 | |
ad3a37de | 6638 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 6639 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
6640 | |
6641 | if (move_group) { | |
0cda4c02 | 6642 | synchronize_rcu(); |
e2d37cd2 | 6643 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
6644 | get_ctx(ctx); |
6645 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
6646 | group_entry) { | |
e2d37cd2 | 6647 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
6648 | get_ctx(ctx); |
6649 | } | |
6650 | } | |
6651 | ||
e2d37cd2 | 6652 | perf_install_in_context(ctx, event, event->cpu); |
ad3a37de | 6653 | ++ctx->generation; |
fe4b04fa | 6654 | perf_unpin_context(ctx); |
d859e29f | 6655 | mutex_unlock(&ctx->mutex); |
9b51f66d | 6656 | |
fbfc623f YZ |
6657 | put_online_cpus(); |
6658 | ||
cdd6c482 | 6659 | event->owner = current; |
8882135b | 6660 | |
cdd6c482 IM |
6661 | mutex_lock(¤t->perf_event_mutex); |
6662 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
6663 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 6664 | |
c320c7b7 ACM |
6665 | /* |
6666 | * Precalculate sample_data sizes | |
6667 | */ | |
6668 | perf_event__header_size(event); | |
6844c09d | 6669 | perf_event__id_header_size(event); |
c320c7b7 | 6670 | |
8a49542c PZ |
6671 | /* |
6672 | * Drop the reference on the group_event after placing the | |
6673 | * new event on the sibling_list. This ensures destruction | |
6674 | * of the group leader will find the pointer to itself in | |
6675 | * perf_group_detach(). | |
6676 | */ | |
2903ff01 | 6677 | fdput(group); |
ea635c64 AV |
6678 | fd_install(event_fd, event_file); |
6679 | return event_fd; | |
0793a61d | 6680 | |
c3f00c70 | 6681 | err_context: |
fe4b04fa | 6682 | perf_unpin_context(ctx); |
ea635c64 | 6683 | put_ctx(ctx); |
c6be5a5c | 6684 | err_alloc: |
ea635c64 | 6685 | free_event(event); |
e7d0bc04 | 6686 | err_task: |
fbfc623f | 6687 | put_online_cpus(); |
e7d0bc04 PZ |
6688 | if (task) |
6689 | put_task_struct(task); | |
89a1e187 | 6690 | err_group_fd: |
2903ff01 | 6691 | fdput(group); |
ea635c64 AV |
6692 | err_fd: |
6693 | put_unused_fd(event_fd); | |
dc86cabe | 6694 | return err; |
0793a61d TG |
6695 | } |
6696 | ||
fb0459d7 AV |
6697 | /** |
6698 | * perf_event_create_kernel_counter | |
6699 | * | |
6700 | * @attr: attributes of the counter to create | |
6701 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 6702 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
6703 | */ |
6704 | struct perf_event * | |
6705 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 6706 | struct task_struct *task, |
4dc0da86 AK |
6707 | perf_overflow_handler_t overflow_handler, |
6708 | void *context) | |
fb0459d7 | 6709 | { |
fb0459d7 | 6710 | struct perf_event_context *ctx; |
c3f00c70 | 6711 | struct perf_event *event; |
fb0459d7 | 6712 | int err; |
d859e29f | 6713 | |
fb0459d7 AV |
6714 | /* |
6715 | * Get the target context (task or percpu): | |
6716 | */ | |
d859e29f | 6717 | |
4dc0da86 AK |
6718 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
6719 | overflow_handler, context); | |
c3f00c70 PZ |
6720 | if (IS_ERR(event)) { |
6721 | err = PTR_ERR(event); | |
6722 | goto err; | |
6723 | } | |
d859e29f | 6724 | |
38a81da2 | 6725 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
6726 | if (IS_ERR(ctx)) { |
6727 | err = PTR_ERR(ctx); | |
c3f00c70 | 6728 | goto err_free; |
d859e29f | 6729 | } |
fb0459d7 | 6730 | |
fb0459d7 AV |
6731 | WARN_ON_ONCE(ctx->parent_ctx); |
6732 | mutex_lock(&ctx->mutex); | |
6733 | perf_install_in_context(ctx, event, cpu); | |
6734 | ++ctx->generation; | |
fe4b04fa | 6735 | perf_unpin_context(ctx); |
fb0459d7 AV |
6736 | mutex_unlock(&ctx->mutex); |
6737 | ||
fb0459d7 AV |
6738 | return event; |
6739 | ||
c3f00c70 PZ |
6740 | err_free: |
6741 | free_event(event); | |
6742 | err: | |
c6567f64 | 6743 | return ERR_PTR(err); |
9b51f66d | 6744 | } |
fb0459d7 | 6745 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 6746 | |
0cda4c02 YZ |
6747 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
6748 | { | |
6749 | struct perf_event_context *src_ctx; | |
6750 | struct perf_event_context *dst_ctx; | |
6751 | struct perf_event *event, *tmp; | |
6752 | LIST_HEAD(events); | |
6753 | ||
6754 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
6755 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
6756 | ||
6757 | mutex_lock(&src_ctx->mutex); | |
6758 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
6759 | event_entry) { | |
6760 | perf_remove_from_context(event); | |
6761 | put_ctx(src_ctx); | |
6762 | list_add(&event->event_entry, &events); | |
6763 | } | |
6764 | mutex_unlock(&src_ctx->mutex); | |
6765 | ||
6766 | synchronize_rcu(); | |
6767 | ||
6768 | mutex_lock(&dst_ctx->mutex); | |
6769 | list_for_each_entry_safe(event, tmp, &events, event_entry) { | |
6770 | list_del(&event->event_entry); | |
6771 | if (event->state >= PERF_EVENT_STATE_OFF) | |
6772 | event->state = PERF_EVENT_STATE_INACTIVE; | |
6773 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
6774 | get_ctx(dst_ctx); | |
6775 | } | |
6776 | mutex_unlock(&dst_ctx->mutex); | |
6777 | } | |
6778 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
6779 | ||
cdd6c482 | 6780 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 6781 | struct task_struct *child) |
d859e29f | 6782 | { |
cdd6c482 | 6783 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 6784 | u64 child_val; |
d859e29f | 6785 | |
cdd6c482 IM |
6786 | if (child_event->attr.inherit_stat) |
6787 | perf_event_read_event(child_event, child); | |
38b200d6 | 6788 | |
b5e58793 | 6789 | child_val = perf_event_count(child_event); |
d859e29f PM |
6790 | |
6791 | /* | |
6792 | * Add back the child's count to the parent's count: | |
6793 | */ | |
a6e6dea6 | 6794 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
6795 | atomic64_add(child_event->total_time_enabled, |
6796 | &parent_event->child_total_time_enabled); | |
6797 | atomic64_add(child_event->total_time_running, | |
6798 | &parent_event->child_total_time_running); | |
d859e29f PM |
6799 | |
6800 | /* | |
cdd6c482 | 6801 | * Remove this event from the parent's list |
d859e29f | 6802 | */ |
cdd6c482 IM |
6803 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
6804 | mutex_lock(&parent_event->child_mutex); | |
6805 | list_del_init(&child_event->child_list); | |
6806 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
6807 | |
6808 | /* | |
cdd6c482 | 6809 | * Release the parent event, if this was the last |
d859e29f PM |
6810 | * reference to it. |
6811 | */ | |
a6fa941d | 6812 | put_event(parent_event); |
d859e29f PM |
6813 | } |
6814 | ||
9b51f66d | 6815 | static void |
cdd6c482 IM |
6816 | __perf_event_exit_task(struct perf_event *child_event, |
6817 | struct perf_event_context *child_ctx, | |
38b200d6 | 6818 | struct task_struct *child) |
9b51f66d | 6819 | { |
38b435b1 PZ |
6820 | if (child_event->parent) { |
6821 | raw_spin_lock_irq(&child_ctx->lock); | |
6822 | perf_group_detach(child_event); | |
6823 | raw_spin_unlock_irq(&child_ctx->lock); | |
6824 | } | |
9b51f66d | 6825 | |
fe4b04fa | 6826 | perf_remove_from_context(child_event); |
0cc0c027 | 6827 | |
9b51f66d | 6828 | /* |
38b435b1 | 6829 | * It can happen that the parent exits first, and has events |
9b51f66d | 6830 | * that are still around due to the child reference. These |
38b435b1 | 6831 | * events need to be zapped. |
9b51f66d | 6832 | */ |
38b435b1 | 6833 | if (child_event->parent) { |
cdd6c482 IM |
6834 | sync_child_event(child_event, child); |
6835 | free_event(child_event); | |
4bcf349a | 6836 | } |
9b51f66d IM |
6837 | } |
6838 | ||
8dc85d54 | 6839 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 6840 | { |
cdd6c482 IM |
6841 | struct perf_event *child_event, *tmp; |
6842 | struct perf_event_context *child_ctx; | |
a63eaf34 | 6843 | unsigned long flags; |
9b51f66d | 6844 | |
8dc85d54 | 6845 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 6846 | perf_event_task(child, NULL, 0); |
9b51f66d | 6847 | return; |
9f498cc5 | 6848 | } |
9b51f66d | 6849 | |
a63eaf34 | 6850 | local_irq_save(flags); |
ad3a37de PM |
6851 | /* |
6852 | * We can't reschedule here because interrupts are disabled, | |
6853 | * and either child is current or it is a task that can't be | |
6854 | * scheduled, so we are now safe from rescheduling changing | |
6855 | * our context. | |
6856 | */ | |
806839b2 | 6857 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
6858 | |
6859 | /* | |
6860 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 6861 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
6862 | * incremented the context's refcount before we do put_ctx below. |
6863 | */ | |
e625cce1 | 6864 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 6865 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 6866 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
6867 | /* |
6868 | * If this context is a clone; unclone it so it can't get | |
6869 | * swapped to another process while we're removing all | |
cdd6c482 | 6870 | * the events from it. |
71a851b4 PZ |
6871 | */ |
6872 | unclone_ctx(child_ctx); | |
5e942bb3 | 6873 | update_context_time(child_ctx); |
e625cce1 | 6874 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
6875 | |
6876 | /* | |
cdd6c482 IM |
6877 | * Report the task dead after unscheduling the events so that we |
6878 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
6879 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 6880 | */ |
cdd6c482 | 6881 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 6882 | |
66fff224 PZ |
6883 | /* |
6884 | * We can recurse on the same lock type through: | |
6885 | * | |
cdd6c482 IM |
6886 | * __perf_event_exit_task() |
6887 | * sync_child_event() | |
a6fa941d AV |
6888 | * put_event() |
6889 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
6890 | * |
6891 | * But since its the parent context it won't be the same instance. | |
6892 | */ | |
a0507c84 | 6893 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 6894 | |
8bc20959 | 6895 | again: |
889ff015 FW |
6896 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
6897 | group_entry) | |
6898 | __perf_event_exit_task(child_event, child_ctx, child); | |
6899 | ||
6900 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 6901 | group_entry) |
cdd6c482 | 6902 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
6903 | |
6904 | /* | |
cdd6c482 | 6905 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
6906 | * its siblings to the list, but we obtained 'tmp' before that which |
6907 | * will still point to the list head terminating the iteration. | |
6908 | */ | |
889ff015 FW |
6909 | if (!list_empty(&child_ctx->pinned_groups) || |
6910 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 6911 | goto again; |
a63eaf34 PM |
6912 | |
6913 | mutex_unlock(&child_ctx->mutex); | |
6914 | ||
6915 | put_ctx(child_ctx); | |
9b51f66d IM |
6916 | } |
6917 | ||
8dc85d54 PZ |
6918 | /* |
6919 | * When a child task exits, feed back event values to parent events. | |
6920 | */ | |
6921 | void perf_event_exit_task(struct task_struct *child) | |
6922 | { | |
8882135b | 6923 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
6924 | int ctxn; |
6925 | ||
8882135b PZ |
6926 | mutex_lock(&child->perf_event_mutex); |
6927 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
6928 | owner_entry) { | |
6929 | list_del_init(&event->owner_entry); | |
6930 | ||
6931 | /* | |
6932 | * Ensure the list deletion is visible before we clear | |
6933 | * the owner, closes a race against perf_release() where | |
6934 | * we need to serialize on the owner->perf_event_mutex. | |
6935 | */ | |
6936 | smp_wmb(); | |
6937 | event->owner = NULL; | |
6938 | } | |
6939 | mutex_unlock(&child->perf_event_mutex); | |
6940 | ||
8dc85d54 PZ |
6941 | for_each_task_context_nr(ctxn) |
6942 | perf_event_exit_task_context(child, ctxn); | |
6943 | } | |
6944 | ||
889ff015 FW |
6945 | static void perf_free_event(struct perf_event *event, |
6946 | struct perf_event_context *ctx) | |
6947 | { | |
6948 | struct perf_event *parent = event->parent; | |
6949 | ||
6950 | if (WARN_ON_ONCE(!parent)) | |
6951 | return; | |
6952 | ||
6953 | mutex_lock(&parent->child_mutex); | |
6954 | list_del_init(&event->child_list); | |
6955 | mutex_unlock(&parent->child_mutex); | |
6956 | ||
a6fa941d | 6957 | put_event(parent); |
889ff015 | 6958 | |
8a49542c | 6959 | perf_group_detach(event); |
889ff015 FW |
6960 | list_del_event(event, ctx); |
6961 | free_event(event); | |
6962 | } | |
6963 | ||
bbbee908 PZ |
6964 | /* |
6965 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 6966 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 6967 | */ |
cdd6c482 | 6968 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 6969 | { |
8dc85d54 | 6970 | struct perf_event_context *ctx; |
cdd6c482 | 6971 | struct perf_event *event, *tmp; |
8dc85d54 | 6972 | int ctxn; |
bbbee908 | 6973 | |
8dc85d54 PZ |
6974 | for_each_task_context_nr(ctxn) { |
6975 | ctx = task->perf_event_ctxp[ctxn]; | |
6976 | if (!ctx) | |
6977 | continue; | |
bbbee908 | 6978 | |
8dc85d54 | 6979 | mutex_lock(&ctx->mutex); |
bbbee908 | 6980 | again: |
8dc85d54 PZ |
6981 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
6982 | group_entry) | |
6983 | perf_free_event(event, ctx); | |
bbbee908 | 6984 | |
8dc85d54 PZ |
6985 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
6986 | group_entry) | |
6987 | perf_free_event(event, ctx); | |
bbbee908 | 6988 | |
8dc85d54 PZ |
6989 | if (!list_empty(&ctx->pinned_groups) || |
6990 | !list_empty(&ctx->flexible_groups)) | |
6991 | goto again; | |
bbbee908 | 6992 | |
8dc85d54 | 6993 | mutex_unlock(&ctx->mutex); |
bbbee908 | 6994 | |
8dc85d54 PZ |
6995 | put_ctx(ctx); |
6996 | } | |
889ff015 FW |
6997 | } |
6998 | ||
4e231c79 PZ |
6999 | void perf_event_delayed_put(struct task_struct *task) |
7000 | { | |
7001 | int ctxn; | |
7002 | ||
7003 | for_each_task_context_nr(ctxn) | |
7004 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7005 | } | |
7006 | ||
97dee4f3 PZ |
7007 | /* |
7008 | * inherit a event from parent task to child task: | |
7009 | */ | |
7010 | static struct perf_event * | |
7011 | inherit_event(struct perf_event *parent_event, | |
7012 | struct task_struct *parent, | |
7013 | struct perf_event_context *parent_ctx, | |
7014 | struct task_struct *child, | |
7015 | struct perf_event *group_leader, | |
7016 | struct perf_event_context *child_ctx) | |
7017 | { | |
7018 | struct perf_event *child_event; | |
cee010ec | 7019 | unsigned long flags; |
97dee4f3 PZ |
7020 | |
7021 | /* | |
7022 | * Instead of creating recursive hierarchies of events, | |
7023 | * we link inherited events back to the original parent, | |
7024 | * which has a filp for sure, which we use as the reference | |
7025 | * count: | |
7026 | */ | |
7027 | if (parent_event->parent) | |
7028 | parent_event = parent_event->parent; | |
7029 | ||
7030 | child_event = perf_event_alloc(&parent_event->attr, | |
7031 | parent_event->cpu, | |
d580ff86 | 7032 | child, |
97dee4f3 | 7033 | group_leader, parent_event, |
4dc0da86 | 7034 | NULL, NULL); |
97dee4f3 PZ |
7035 | if (IS_ERR(child_event)) |
7036 | return child_event; | |
a6fa941d AV |
7037 | |
7038 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7039 | free_event(child_event); | |
7040 | return NULL; | |
7041 | } | |
7042 | ||
97dee4f3 PZ |
7043 | get_ctx(child_ctx); |
7044 | ||
7045 | /* | |
7046 | * Make the child state follow the state of the parent event, | |
7047 | * not its attr.disabled bit. We hold the parent's mutex, | |
7048 | * so we won't race with perf_event_{en, dis}able_family. | |
7049 | */ | |
7050 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7051 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7052 | else | |
7053 | child_event->state = PERF_EVENT_STATE_OFF; | |
7054 | ||
7055 | if (parent_event->attr.freq) { | |
7056 | u64 sample_period = parent_event->hw.sample_period; | |
7057 | struct hw_perf_event *hwc = &child_event->hw; | |
7058 | ||
7059 | hwc->sample_period = sample_period; | |
7060 | hwc->last_period = sample_period; | |
7061 | ||
7062 | local64_set(&hwc->period_left, sample_period); | |
7063 | } | |
7064 | ||
7065 | child_event->ctx = child_ctx; | |
7066 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7067 | child_event->overflow_handler_context |
7068 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7069 | |
614b6780 TG |
7070 | /* |
7071 | * Precalculate sample_data sizes | |
7072 | */ | |
7073 | perf_event__header_size(child_event); | |
6844c09d | 7074 | perf_event__id_header_size(child_event); |
614b6780 | 7075 | |
97dee4f3 PZ |
7076 | /* |
7077 | * Link it up in the child's context: | |
7078 | */ | |
cee010ec | 7079 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7080 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7081 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7082 | |
97dee4f3 PZ |
7083 | /* |
7084 | * Link this into the parent event's child list | |
7085 | */ | |
7086 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7087 | mutex_lock(&parent_event->child_mutex); | |
7088 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7089 | mutex_unlock(&parent_event->child_mutex); | |
7090 | ||
7091 | return child_event; | |
7092 | } | |
7093 | ||
7094 | static int inherit_group(struct perf_event *parent_event, | |
7095 | struct task_struct *parent, | |
7096 | struct perf_event_context *parent_ctx, | |
7097 | struct task_struct *child, | |
7098 | struct perf_event_context *child_ctx) | |
7099 | { | |
7100 | struct perf_event *leader; | |
7101 | struct perf_event *sub; | |
7102 | struct perf_event *child_ctr; | |
7103 | ||
7104 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7105 | child, NULL, child_ctx); | |
7106 | if (IS_ERR(leader)) | |
7107 | return PTR_ERR(leader); | |
7108 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7109 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7110 | child, leader, child_ctx); | |
7111 | if (IS_ERR(child_ctr)) | |
7112 | return PTR_ERR(child_ctr); | |
7113 | } | |
7114 | return 0; | |
889ff015 FW |
7115 | } |
7116 | ||
7117 | static int | |
7118 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7119 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7120 | struct task_struct *child, int ctxn, |
889ff015 FW |
7121 | int *inherited_all) |
7122 | { | |
7123 | int ret; | |
8dc85d54 | 7124 | struct perf_event_context *child_ctx; |
889ff015 FW |
7125 | |
7126 | if (!event->attr.inherit) { | |
7127 | *inherited_all = 0; | |
7128 | return 0; | |
bbbee908 PZ |
7129 | } |
7130 | ||
fe4b04fa | 7131 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7132 | if (!child_ctx) { |
7133 | /* | |
7134 | * This is executed from the parent task context, so | |
7135 | * inherit events that have been marked for cloning. | |
7136 | * First allocate and initialize a context for the | |
7137 | * child. | |
7138 | */ | |
bbbee908 | 7139 | |
eb184479 | 7140 | child_ctx = alloc_perf_context(event->pmu, child); |
889ff015 FW |
7141 | if (!child_ctx) |
7142 | return -ENOMEM; | |
bbbee908 | 7143 | |
8dc85d54 | 7144 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7145 | } |
7146 | ||
7147 | ret = inherit_group(event, parent, parent_ctx, | |
7148 | child, child_ctx); | |
7149 | ||
7150 | if (ret) | |
7151 | *inherited_all = 0; | |
7152 | ||
7153 | return ret; | |
bbbee908 PZ |
7154 | } |
7155 | ||
9b51f66d | 7156 | /* |
cdd6c482 | 7157 | * Initialize the perf_event context in task_struct |
9b51f66d | 7158 | */ |
8dc85d54 | 7159 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7160 | { |
889ff015 | 7161 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7162 | struct perf_event_context *cloned_ctx; |
7163 | struct perf_event *event; | |
9b51f66d | 7164 | struct task_struct *parent = current; |
564c2b21 | 7165 | int inherited_all = 1; |
dddd3379 | 7166 | unsigned long flags; |
6ab423e0 | 7167 | int ret = 0; |
9b51f66d | 7168 | |
8dc85d54 | 7169 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7170 | return 0; |
7171 | ||
ad3a37de | 7172 | /* |
25346b93 PM |
7173 | * If the parent's context is a clone, pin it so it won't get |
7174 | * swapped under us. | |
ad3a37de | 7175 | */ |
8dc85d54 | 7176 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 7177 | |
ad3a37de PM |
7178 | /* |
7179 | * No need to check if parent_ctx != NULL here; since we saw | |
7180 | * it non-NULL earlier, the only reason for it to become NULL | |
7181 | * is if we exit, and since we're currently in the middle of | |
7182 | * a fork we can't be exiting at the same time. | |
7183 | */ | |
ad3a37de | 7184 | |
9b51f66d IM |
7185 | /* |
7186 | * Lock the parent list. No need to lock the child - not PID | |
7187 | * hashed yet and not running, so nobody can access it. | |
7188 | */ | |
d859e29f | 7189 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7190 | |
7191 | /* | |
7192 | * We dont have to disable NMIs - we are only looking at | |
7193 | * the list, not manipulating it: | |
7194 | */ | |
889ff015 | 7195 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7196 | ret = inherit_task_group(event, parent, parent_ctx, |
7197 | child, ctxn, &inherited_all); | |
889ff015 FW |
7198 | if (ret) |
7199 | break; | |
7200 | } | |
b93f7978 | 7201 | |
dddd3379 TG |
7202 | /* |
7203 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7204 | * to allocations, but we need to prevent rotation because | |
7205 | * rotate_ctx() will change the list from interrupt context. | |
7206 | */ | |
7207 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7208 | parent_ctx->rotate_disable = 1; | |
7209 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7210 | ||
889ff015 | 7211 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7212 | ret = inherit_task_group(event, parent, parent_ctx, |
7213 | child, ctxn, &inherited_all); | |
889ff015 | 7214 | if (ret) |
9b51f66d | 7215 | break; |
564c2b21 PM |
7216 | } |
7217 | ||
dddd3379 TG |
7218 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7219 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7220 | |
8dc85d54 | 7221 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7222 | |
05cbaa28 | 7223 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7224 | /* |
7225 | * Mark the child context as a clone of the parent | |
7226 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7227 | * |
7228 | * Note that if the parent is a clone, the holding of | |
7229 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7230 | */ |
c5ed5145 | 7231 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7232 | if (cloned_ctx) { |
7233 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7234 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7235 | } else { |
7236 | child_ctx->parent_ctx = parent_ctx; | |
7237 | child_ctx->parent_gen = parent_ctx->generation; | |
7238 | } | |
7239 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7240 | } |
7241 | ||
c5ed5145 | 7242 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7243 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7244 | |
25346b93 | 7245 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7246 | put_ctx(parent_ctx); |
ad3a37de | 7247 | |
6ab423e0 | 7248 | return ret; |
9b51f66d IM |
7249 | } |
7250 | ||
8dc85d54 PZ |
7251 | /* |
7252 | * Initialize the perf_event context in task_struct | |
7253 | */ | |
7254 | int perf_event_init_task(struct task_struct *child) | |
7255 | { | |
7256 | int ctxn, ret; | |
7257 | ||
8550d7cb ON |
7258 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7259 | mutex_init(&child->perf_event_mutex); | |
7260 | INIT_LIST_HEAD(&child->perf_event_list); | |
7261 | ||
8dc85d54 PZ |
7262 | for_each_task_context_nr(ctxn) { |
7263 | ret = perf_event_init_context(child, ctxn); | |
7264 | if (ret) | |
7265 | return ret; | |
7266 | } | |
7267 | ||
7268 | return 0; | |
7269 | } | |
7270 | ||
220b140b PM |
7271 | static void __init perf_event_init_all_cpus(void) |
7272 | { | |
b28ab83c | 7273 | struct swevent_htable *swhash; |
220b140b | 7274 | int cpu; |
220b140b PM |
7275 | |
7276 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7277 | swhash = &per_cpu(swevent_htable, cpu); |
7278 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7279 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7280 | } |
7281 | } | |
7282 | ||
cdd6c482 | 7283 | static void __cpuinit perf_event_init_cpu(int cpu) |
0793a61d | 7284 | { |
108b02cf | 7285 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7286 | |
b28ab83c | 7287 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7288 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7289 | struct swevent_hlist *hlist; |
7290 | ||
b28ab83c PZ |
7291 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7292 | WARN_ON(!hlist); | |
7293 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7294 | } |
b28ab83c | 7295 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7296 | } |
7297 | ||
c277443c | 7298 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7299 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7300 | { |
e9d2b064 PZ |
7301 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7302 | ||
7303 | WARN_ON(!irqs_disabled()); | |
7304 | ||
7305 | list_del_init(&cpuctx->rotation_list); | |
7306 | } | |
7307 | ||
108b02cf | 7308 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7309 | { |
108b02cf | 7310 | struct perf_event_context *ctx = __info; |
cdd6c482 | 7311 | struct perf_event *event, *tmp; |
0793a61d | 7312 | |
108b02cf | 7313 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7314 | |
889ff015 | 7315 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 7316 | __perf_remove_from_context(event); |
889ff015 | 7317 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 7318 | __perf_remove_from_context(event); |
0793a61d | 7319 | } |
108b02cf PZ |
7320 | |
7321 | static void perf_event_exit_cpu_context(int cpu) | |
7322 | { | |
7323 | struct perf_event_context *ctx; | |
7324 | struct pmu *pmu; | |
7325 | int idx; | |
7326 | ||
7327 | idx = srcu_read_lock(&pmus_srcu); | |
7328 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7329 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7330 | |
7331 | mutex_lock(&ctx->mutex); | |
7332 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7333 | mutex_unlock(&ctx->mutex); | |
7334 | } | |
7335 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7336 | } |
7337 | ||
cdd6c482 | 7338 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7339 | { |
b28ab83c | 7340 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7341 | |
b28ab83c PZ |
7342 | mutex_lock(&swhash->hlist_mutex); |
7343 | swevent_hlist_release(swhash); | |
7344 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 7345 | |
108b02cf | 7346 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7347 | } |
7348 | #else | |
cdd6c482 | 7349 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7350 | #endif |
7351 | ||
c277443c PZ |
7352 | static int |
7353 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7354 | { | |
7355 | int cpu; | |
7356 | ||
7357 | for_each_online_cpu(cpu) | |
7358 | perf_event_exit_cpu(cpu); | |
7359 | ||
7360 | return NOTIFY_OK; | |
7361 | } | |
7362 | ||
7363 | /* | |
7364 | * Run the perf reboot notifier at the very last possible moment so that | |
7365 | * the generic watchdog code runs as long as possible. | |
7366 | */ | |
7367 | static struct notifier_block perf_reboot_notifier = { | |
7368 | .notifier_call = perf_reboot, | |
7369 | .priority = INT_MIN, | |
7370 | }; | |
7371 | ||
0793a61d TG |
7372 | static int __cpuinit |
7373 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
7374 | { | |
7375 | unsigned int cpu = (long)hcpu; | |
7376 | ||
4536e4d1 | 7377 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7378 | |
7379 | case CPU_UP_PREPARE: | |
5e11637e | 7380 | case CPU_DOWN_FAILED: |
cdd6c482 | 7381 | perf_event_init_cpu(cpu); |
0793a61d TG |
7382 | break; |
7383 | ||
5e11637e | 7384 | case CPU_UP_CANCELED: |
0793a61d | 7385 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7386 | perf_event_exit_cpu(cpu); |
0793a61d TG |
7387 | break; |
7388 | ||
7389 | default: | |
7390 | break; | |
7391 | } | |
7392 | ||
7393 | return NOTIFY_OK; | |
7394 | } | |
7395 | ||
cdd6c482 | 7396 | void __init perf_event_init(void) |
0793a61d | 7397 | { |
3c502e7a JW |
7398 | int ret; |
7399 | ||
2e80a82a PZ |
7400 | idr_init(&pmu_idr); |
7401 | ||
220b140b | 7402 | perf_event_init_all_cpus(); |
b0a873eb | 7403 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7404 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7405 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7406 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7407 | perf_tp_register(); |
7408 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7409 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7410 | |
7411 | ret = init_hw_breakpoint(); | |
7412 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7413 | |
7414 | /* do not patch jump label more than once per second */ | |
7415 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7416 | |
7417 | /* | |
7418 | * Build time assertion that we keep the data_head at the intended | |
7419 | * location. IOW, validation we got the __reserved[] size right. | |
7420 | */ | |
7421 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
7422 | != 1024); | |
0793a61d | 7423 | } |
abe43400 PZ |
7424 | |
7425 | static int __init perf_event_sysfs_init(void) | |
7426 | { | |
7427 | struct pmu *pmu; | |
7428 | int ret; | |
7429 | ||
7430 | mutex_lock(&pmus_lock); | |
7431 | ||
7432 | ret = bus_register(&pmu_bus); | |
7433 | if (ret) | |
7434 | goto unlock; | |
7435 | ||
7436 | list_for_each_entry(pmu, &pmus, entry) { | |
7437 | if (!pmu->name || pmu->type < 0) | |
7438 | continue; | |
7439 | ||
7440 | ret = pmu_dev_alloc(pmu); | |
7441 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
7442 | } | |
7443 | pmu_bus_running = 1; | |
7444 | ret = 0; | |
7445 | ||
7446 | unlock: | |
7447 | mutex_unlock(&pmus_lock); | |
7448 | ||
7449 | return ret; | |
7450 | } | |
7451 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
7452 | |
7453 | #ifdef CONFIG_CGROUP_PERF | |
92fb9748 | 7454 | static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont) |
e5d1367f SE |
7455 | { |
7456 | struct perf_cgroup *jc; | |
e5d1367f | 7457 | |
1b15d055 | 7458 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
7459 | if (!jc) |
7460 | return ERR_PTR(-ENOMEM); | |
7461 | ||
e5d1367f SE |
7462 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7463 | if (!jc->info) { | |
7464 | kfree(jc); | |
7465 | return ERR_PTR(-ENOMEM); | |
7466 | } | |
7467 | ||
e5d1367f SE |
7468 | return &jc->css; |
7469 | } | |
7470 | ||
92fb9748 | 7471 | static void perf_cgroup_css_free(struct cgroup *cont) |
e5d1367f SE |
7472 | { |
7473 | struct perf_cgroup *jc; | |
7474 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), | |
7475 | struct perf_cgroup, css); | |
7476 | free_percpu(jc->info); | |
7477 | kfree(jc); | |
7478 | } | |
7479 | ||
7480 | static int __perf_cgroup_move(void *info) | |
7481 | { | |
7482 | struct task_struct *task = info; | |
7483 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
7484 | return 0; | |
7485 | } | |
7486 | ||
761b3ef5 | 7487 | static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
e5d1367f | 7488 | { |
bb9d97b6 TH |
7489 | struct task_struct *task; |
7490 | ||
7491 | cgroup_taskset_for_each(task, cgrp, tset) | |
7492 | task_function_call(task, __perf_cgroup_move, task); | |
e5d1367f SE |
7493 | } |
7494 | ||
761b3ef5 LZ |
7495 | static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, |
7496 | struct task_struct *task) | |
e5d1367f SE |
7497 | { |
7498 | /* | |
7499 | * cgroup_exit() is called in the copy_process() failure path. | |
7500 | * Ignore this case since the task hasn't ran yet, this avoids | |
7501 | * trying to poke a half freed task state from generic code. | |
7502 | */ | |
7503 | if (!(task->flags & PF_EXITING)) | |
7504 | return; | |
7505 | ||
bb9d97b6 | 7506 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7507 | } |
7508 | ||
7509 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
7510 | .name = "perf_event", |
7511 | .subsys_id = perf_subsys_id, | |
92fb9748 TH |
7512 | .css_alloc = perf_cgroup_css_alloc, |
7513 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 7514 | .exit = perf_cgroup_exit, |
bb9d97b6 | 7515 | .attach = perf_cgroup_attach, |
8c7f6edb TH |
7516 | |
7517 | /* | |
7518 | * perf_event cgroup doesn't handle nesting correctly. | |
7519 | * ctx->nr_cgroups adjustments should be propagated through the | |
7520 | * cgroup hierarchy. Fix it and remove the following. | |
7521 | */ | |
7522 | .broken_hierarchy = true, | |
e5d1367f SE |
7523 | }; |
7524 | #endif /* CONFIG_CGROUP_PERF */ |