]>
Commit | Line | Data |
---|---|---|
0793a61d | 1 | /* |
57c0c15b | 2 | * Performance events core code: |
0793a61d | 3 | * |
98144511 | 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
e7e7ee2e IM |
5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
7b732a75 | 8 | * |
57c0c15b | 9 | * For licensing details see kernel-base/COPYING |
0793a61d TG |
10 | */ |
11 | ||
12 | #include <linux/fs.h> | |
b9cacc7b | 13 | #include <linux/mm.h> |
0793a61d TG |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> | |
2e80a82a | 16 | #include <linux/idr.h> |
04289bb9 | 17 | #include <linux/file.h> |
0793a61d | 18 | #include <linux/poll.h> |
5a0e3ad6 | 19 | #include <linux/slab.h> |
76e1d904 | 20 | #include <linux/hash.h> |
12351ef8 | 21 | #include <linux/tick.h> |
0793a61d | 22 | #include <linux/sysfs.h> |
22a4f650 | 23 | #include <linux/dcache.h> |
0793a61d | 24 | #include <linux/percpu.h> |
22a4f650 | 25 | #include <linux/ptrace.h> |
c277443c | 26 | #include <linux/reboot.h> |
b9cacc7b | 27 | #include <linux/vmstat.h> |
abe43400 | 28 | #include <linux/device.h> |
6e5fdeed | 29 | #include <linux/export.h> |
906010b2 | 30 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
31 | #include <linux/hardirq.h> |
32 | #include <linux/rculist.h> | |
0793a61d TG |
33 | #include <linux/uaccess.h> |
34 | #include <linux/syscalls.h> | |
35 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 36 | #include <linux/kernel_stat.h> |
cdd6c482 | 37 | #include <linux/perf_event.h> |
6fb2915d | 38 | #include <linux/ftrace_event.h> |
3c502e7a | 39 | #include <linux/hw_breakpoint.h> |
c5ebcedb | 40 | #include <linux/mm_types.h> |
877c6856 | 41 | #include <linux/cgroup.h> |
0793a61d | 42 | |
76369139 FW |
43 | #include "internal.h" |
44 | ||
4e193bd4 TB |
45 | #include <asm/irq_regs.h> |
46 | ||
fe4b04fa | 47 | struct remote_function_call { |
e7e7ee2e IM |
48 | struct task_struct *p; |
49 | int (*func)(void *info); | |
50 | void *info; | |
51 | int ret; | |
fe4b04fa PZ |
52 | }; |
53 | ||
54 | static void remote_function(void *data) | |
55 | { | |
56 | struct remote_function_call *tfc = data; | |
57 | struct task_struct *p = tfc->p; | |
58 | ||
59 | if (p) { | |
60 | tfc->ret = -EAGAIN; | |
61 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
62 | return; | |
63 | } | |
64 | ||
65 | tfc->ret = tfc->func(tfc->info); | |
66 | } | |
67 | ||
68 | /** | |
69 | * task_function_call - call a function on the cpu on which a task runs | |
70 | * @p: the task to evaluate | |
71 | * @func: the function to be called | |
72 | * @info: the function call argument | |
73 | * | |
74 | * Calls the function @func when the task is currently running. This might | |
75 | * be on the current CPU, which just calls the function directly | |
76 | * | |
77 | * returns: @func return value, or | |
78 | * -ESRCH - when the process isn't running | |
79 | * -EAGAIN - when the process moved away | |
80 | */ | |
81 | static int | |
82 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
83 | { | |
84 | struct remote_function_call data = { | |
e7e7ee2e IM |
85 | .p = p, |
86 | .func = func, | |
87 | .info = info, | |
88 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
89 | }; |
90 | ||
91 | if (task_curr(p)) | |
92 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
93 | ||
94 | return data.ret; | |
95 | } | |
96 | ||
97 | /** | |
98 | * cpu_function_call - call a function on the cpu | |
99 | * @func: the function to be called | |
100 | * @info: the function call argument | |
101 | * | |
102 | * Calls the function @func on the remote cpu. | |
103 | * | |
104 | * returns: @func return value or -ENXIO when the cpu is offline | |
105 | */ | |
106 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
107 | { | |
108 | struct remote_function_call data = { | |
e7e7ee2e IM |
109 | .p = NULL, |
110 | .func = func, | |
111 | .info = info, | |
112 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
113 | }; |
114 | ||
115 | smp_call_function_single(cpu, remote_function, &data, 1); | |
116 | ||
117 | return data.ret; | |
118 | } | |
119 | ||
e5d1367f SE |
120 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
121 | PERF_FLAG_FD_OUTPUT |\ | |
122 | PERF_FLAG_PID_CGROUP) | |
123 | ||
bce38cd5 SE |
124 | /* |
125 | * branch priv levels that need permission checks | |
126 | */ | |
127 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
128 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
129 | PERF_SAMPLE_BRANCH_HV) | |
130 | ||
0b3fcf17 SE |
131 | enum event_type_t { |
132 | EVENT_FLEXIBLE = 0x1, | |
133 | EVENT_PINNED = 0x2, | |
134 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
135 | }; | |
136 | ||
e5d1367f SE |
137 | /* |
138 | * perf_sched_events : >0 events exist | |
139 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
140 | */ | |
c5905afb | 141 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 142 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 143 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 144 | |
cdd6c482 IM |
145 | static atomic_t nr_mmap_events __read_mostly; |
146 | static atomic_t nr_comm_events __read_mostly; | |
147 | static atomic_t nr_task_events __read_mostly; | |
9ee318a7 | 148 | |
108b02cf PZ |
149 | static LIST_HEAD(pmus); |
150 | static DEFINE_MUTEX(pmus_lock); | |
151 | static struct srcu_struct pmus_srcu; | |
152 | ||
0764771d | 153 | /* |
cdd6c482 | 154 | * perf event paranoia level: |
0fbdea19 IM |
155 | * -1 - not paranoid at all |
156 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 157 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 158 | * 2 - disallow kernel profiling for unpriv |
0764771d | 159 | */ |
cdd6c482 | 160 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 161 | |
20443384 FW |
162 | /* Minimum for 512 kiB + 1 user control page */ |
163 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
164 | |
165 | /* | |
cdd6c482 | 166 | * max perf event sample rate |
df58ab24 | 167 | */ |
163ec435 PZ |
168 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
169 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
170 | static int max_samples_per_tick __read_mostly = | |
171 | DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
172 | ||
9e630205 SE |
173 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
174 | ||
163ec435 PZ |
175 | int perf_proc_update_handler(struct ctl_table *table, int write, |
176 | void __user *buffer, size_t *lenp, | |
177 | loff_t *ppos) | |
178 | { | |
179 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
180 | ||
181 | if (ret || !write) | |
182 | return ret; | |
183 | ||
184 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
185 | ||
186 | return 0; | |
187 | } | |
1ccd1549 | 188 | |
cdd6c482 | 189 | static atomic64_t perf_event_id; |
a96bbc16 | 190 | |
0b3fcf17 SE |
191 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
192 | enum event_type_t event_type); | |
193 | ||
194 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
195 | enum event_type_t event_type, |
196 | struct task_struct *task); | |
197 | ||
198 | static void update_context_time(struct perf_event_context *ctx); | |
199 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 200 | |
cdd6c482 | 201 | void __weak perf_event_print_debug(void) { } |
0793a61d | 202 | |
84c79910 | 203 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 204 | { |
84c79910 | 205 | return "pmu"; |
0793a61d TG |
206 | } |
207 | ||
0b3fcf17 SE |
208 | static inline u64 perf_clock(void) |
209 | { | |
210 | return local_clock(); | |
211 | } | |
212 | ||
e5d1367f SE |
213 | static inline struct perf_cpu_context * |
214 | __get_cpu_context(struct perf_event_context *ctx) | |
215 | { | |
216 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
217 | } | |
218 | ||
facc4307 PZ |
219 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
220 | struct perf_event_context *ctx) | |
221 | { | |
222 | raw_spin_lock(&cpuctx->ctx.lock); | |
223 | if (ctx) | |
224 | raw_spin_lock(&ctx->lock); | |
225 | } | |
226 | ||
227 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
228 | struct perf_event_context *ctx) | |
229 | { | |
230 | if (ctx) | |
231 | raw_spin_unlock(&ctx->lock); | |
232 | raw_spin_unlock(&cpuctx->ctx.lock); | |
233 | } | |
234 | ||
e5d1367f SE |
235 | #ifdef CONFIG_CGROUP_PERF |
236 | ||
877c6856 LZ |
237 | /* |
238 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
239 | * This is a per-cpu dynamically allocated data structure. | |
240 | */ | |
241 | struct perf_cgroup_info { | |
242 | u64 time; | |
243 | u64 timestamp; | |
244 | }; | |
245 | ||
246 | struct perf_cgroup { | |
247 | struct cgroup_subsys_state css; | |
86e213e1 | 248 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
249 | }; |
250 | ||
3f7cce3c SE |
251 | /* |
252 | * Must ensure cgroup is pinned (css_get) before calling | |
253 | * this function. In other words, we cannot call this function | |
254 | * if there is no cgroup event for the current CPU context. | |
255 | */ | |
e5d1367f SE |
256 | static inline struct perf_cgroup * |
257 | perf_cgroup_from_task(struct task_struct *task) | |
258 | { | |
259 | return container_of(task_subsys_state(task, perf_subsys_id), | |
260 | struct perf_cgroup, css); | |
261 | } | |
262 | ||
263 | static inline bool | |
264 | perf_cgroup_match(struct perf_event *event) | |
265 | { | |
266 | struct perf_event_context *ctx = event->ctx; | |
267 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
268 | ||
ef824fa1 TH |
269 | /* @event doesn't care about cgroup */ |
270 | if (!event->cgrp) | |
271 | return true; | |
272 | ||
273 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
274 | if (!cpuctx->cgrp) | |
275 | return false; | |
276 | ||
277 | /* | |
278 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
279 | * also enabled for all its descendant cgroups. If @cpuctx's | |
280 | * cgroup is a descendant of @event's (the test covers identity | |
281 | * case), it's a match. | |
282 | */ | |
283 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
284 | event->cgrp->css.cgroup); | |
e5d1367f SE |
285 | } |
286 | ||
9c5da09d | 287 | static inline bool perf_tryget_cgroup(struct perf_event *event) |
e5d1367f | 288 | { |
9c5da09d | 289 | return css_tryget(&event->cgrp->css); |
e5d1367f SE |
290 | } |
291 | ||
292 | static inline void perf_put_cgroup(struct perf_event *event) | |
293 | { | |
294 | css_put(&event->cgrp->css); | |
295 | } | |
296 | ||
297 | static inline void perf_detach_cgroup(struct perf_event *event) | |
298 | { | |
299 | perf_put_cgroup(event); | |
300 | event->cgrp = NULL; | |
301 | } | |
302 | ||
303 | static inline int is_cgroup_event(struct perf_event *event) | |
304 | { | |
305 | return event->cgrp != NULL; | |
306 | } | |
307 | ||
308 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
309 | { | |
310 | struct perf_cgroup_info *t; | |
311 | ||
312 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
313 | return t->time; | |
314 | } | |
315 | ||
316 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
317 | { | |
318 | struct perf_cgroup_info *info; | |
319 | u64 now; | |
320 | ||
321 | now = perf_clock(); | |
322 | ||
323 | info = this_cpu_ptr(cgrp->info); | |
324 | ||
325 | info->time += now - info->timestamp; | |
326 | info->timestamp = now; | |
327 | } | |
328 | ||
329 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
330 | { | |
331 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
332 | if (cgrp_out) | |
333 | __update_cgrp_time(cgrp_out); | |
334 | } | |
335 | ||
336 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
337 | { | |
3f7cce3c SE |
338 | struct perf_cgroup *cgrp; |
339 | ||
e5d1367f | 340 | /* |
3f7cce3c SE |
341 | * ensure we access cgroup data only when needed and |
342 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 343 | */ |
3f7cce3c | 344 | if (!is_cgroup_event(event)) |
e5d1367f SE |
345 | return; |
346 | ||
3f7cce3c SE |
347 | cgrp = perf_cgroup_from_task(current); |
348 | /* | |
349 | * Do not update time when cgroup is not active | |
350 | */ | |
351 | if (cgrp == event->cgrp) | |
352 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
353 | } |
354 | ||
355 | static inline void | |
3f7cce3c SE |
356 | perf_cgroup_set_timestamp(struct task_struct *task, |
357 | struct perf_event_context *ctx) | |
e5d1367f SE |
358 | { |
359 | struct perf_cgroup *cgrp; | |
360 | struct perf_cgroup_info *info; | |
361 | ||
3f7cce3c SE |
362 | /* |
363 | * ctx->lock held by caller | |
364 | * ensure we do not access cgroup data | |
365 | * unless we have the cgroup pinned (css_get) | |
366 | */ | |
367 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
368 | return; |
369 | ||
370 | cgrp = perf_cgroup_from_task(task); | |
371 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 372 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
373 | } |
374 | ||
375 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
376 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
377 | ||
378 | /* | |
379 | * reschedule events based on the cgroup constraint of task. | |
380 | * | |
381 | * mode SWOUT : schedule out everything | |
382 | * mode SWIN : schedule in based on cgroup for next | |
383 | */ | |
384 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
385 | { | |
386 | struct perf_cpu_context *cpuctx; | |
387 | struct pmu *pmu; | |
388 | unsigned long flags; | |
389 | ||
390 | /* | |
391 | * disable interrupts to avoid geting nr_cgroup | |
392 | * changes via __perf_event_disable(). Also | |
393 | * avoids preemption. | |
394 | */ | |
395 | local_irq_save(flags); | |
396 | ||
397 | /* | |
398 | * we reschedule only in the presence of cgroup | |
399 | * constrained events. | |
400 | */ | |
401 | rcu_read_lock(); | |
402 | ||
403 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 404 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
405 | if (cpuctx->unique_pmu != pmu) |
406 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 407 | |
e5d1367f SE |
408 | /* |
409 | * perf_cgroup_events says at least one | |
410 | * context on this CPU has cgroup events. | |
411 | * | |
412 | * ctx->nr_cgroups reports the number of cgroup | |
413 | * events for a context. | |
414 | */ | |
415 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
416 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
417 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
418 | |
419 | if (mode & PERF_CGROUP_SWOUT) { | |
420 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
421 | /* | |
422 | * must not be done before ctxswout due | |
423 | * to event_filter_match() in event_sched_out() | |
424 | */ | |
425 | cpuctx->cgrp = NULL; | |
426 | } | |
427 | ||
428 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 429 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
430 | /* |
431 | * set cgrp before ctxsw in to allow | |
432 | * event_filter_match() to not have to pass | |
433 | * task around | |
e5d1367f SE |
434 | */ |
435 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
436 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
437 | } | |
facc4307 PZ |
438 | perf_pmu_enable(cpuctx->ctx.pmu); |
439 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 440 | } |
e5d1367f SE |
441 | } |
442 | ||
443 | rcu_read_unlock(); | |
444 | ||
445 | local_irq_restore(flags); | |
446 | } | |
447 | ||
a8d757ef SE |
448 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
449 | struct task_struct *next) | |
e5d1367f | 450 | { |
a8d757ef SE |
451 | struct perf_cgroup *cgrp1; |
452 | struct perf_cgroup *cgrp2 = NULL; | |
453 | ||
454 | /* | |
455 | * we come here when we know perf_cgroup_events > 0 | |
456 | */ | |
457 | cgrp1 = perf_cgroup_from_task(task); | |
458 | ||
459 | /* | |
460 | * next is NULL when called from perf_event_enable_on_exec() | |
461 | * that will systematically cause a cgroup_switch() | |
462 | */ | |
463 | if (next) | |
464 | cgrp2 = perf_cgroup_from_task(next); | |
465 | ||
466 | /* | |
467 | * only schedule out current cgroup events if we know | |
468 | * that we are switching to a different cgroup. Otherwise, | |
469 | * do no touch the cgroup events. | |
470 | */ | |
471 | if (cgrp1 != cgrp2) | |
472 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
473 | } |
474 | ||
a8d757ef SE |
475 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
476 | struct task_struct *task) | |
e5d1367f | 477 | { |
a8d757ef SE |
478 | struct perf_cgroup *cgrp1; |
479 | struct perf_cgroup *cgrp2 = NULL; | |
480 | ||
481 | /* | |
482 | * we come here when we know perf_cgroup_events > 0 | |
483 | */ | |
484 | cgrp1 = perf_cgroup_from_task(task); | |
485 | ||
486 | /* prev can never be NULL */ | |
487 | cgrp2 = perf_cgroup_from_task(prev); | |
488 | ||
489 | /* | |
490 | * only need to schedule in cgroup events if we are changing | |
491 | * cgroup during ctxsw. Cgroup events were not scheduled | |
492 | * out of ctxsw out if that was not the case. | |
493 | */ | |
494 | if (cgrp1 != cgrp2) | |
495 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
496 | } |
497 | ||
498 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
499 | struct perf_event_attr *attr, | |
500 | struct perf_event *group_leader) | |
501 | { | |
502 | struct perf_cgroup *cgrp; | |
503 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
504 | struct fd f = fdget(fd); |
505 | int ret = 0; | |
e5d1367f | 506 | |
2903ff01 | 507 | if (!f.file) |
e5d1367f SE |
508 | return -EBADF; |
509 | ||
2903ff01 | 510 | css = cgroup_css_from_dir(f.file, perf_subsys_id); |
3db272c0 LZ |
511 | if (IS_ERR(css)) { |
512 | ret = PTR_ERR(css); | |
513 | goto out; | |
514 | } | |
e5d1367f SE |
515 | |
516 | cgrp = container_of(css, struct perf_cgroup, css); | |
517 | event->cgrp = cgrp; | |
518 | ||
f75e18cb | 519 | /* must be done before we fput() the file */ |
9c5da09d SQ |
520 | if (!perf_tryget_cgroup(event)) { |
521 | event->cgrp = NULL; | |
522 | ret = -ENOENT; | |
523 | goto out; | |
524 | } | |
f75e18cb | 525 | |
e5d1367f SE |
526 | /* |
527 | * all events in a group must monitor | |
528 | * the same cgroup because a task belongs | |
529 | * to only one perf cgroup at a time | |
530 | */ | |
531 | if (group_leader && group_leader->cgrp != cgrp) { | |
532 | perf_detach_cgroup(event); | |
533 | ret = -EINVAL; | |
e5d1367f | 534 | } |
3db272c0 | 535 | out: |
2903ff01 | 536 | fdput(f); |
e5d1367f SE |
537 | return ret; |
538 | } | |
539 | ||
540 | static inline void | |
541 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
542 | { | |
543 | struct perf_cgroup_info *t; | |
544 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
545 | event->shadow_ctx_time = now - t->timestamp; | |
546 | } | |
547 | ||
548 | static inline void | |
549 | perf_cgroup_defer_enabled(struct perf_event *event) | |
550 | { | |
551 | /* | |
552 | * when the current task's perf cgroup does not match | |
553 | * the event's, we need to remember to call the | |
554 | * perf_mark_enable() function the first time a task with | |
555 | * a matching perf cgroup is scheduled in. | |
556 | */ | |
557 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
558 | event->cgrp_defer_enabled = 1; | |
559 | } | |
560 | ||
561 | static inline void | |
562 | perf_cgroup_mark_enabled(struct perf_event *event, | |
563 | struct perf_event_context *ctx) | |
564 | { | |
565 | struct perf_event *sub; | |
566 | u64 tstamp = perf_event_time(event); | |
567 | ||
568 | if (!event->cgrp_defer_enabled) | |
569 | return; | |
570 | ||
571 | event->cgrp_defer_enabled = 0; | |
572 | ||
573 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
574 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
575 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
576 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
577 | sub->cgrp_defer_enabled = 0; | |
578 | } | |
579 | } | |
580 | } | |
581 | #else /* !CONFIG_CGROUP_PERF */ | |
582 | ||
583 | static inline bool | |
584 | perf_cgroup_match(struct perf_event *event) | |
585 | { | |
586 | return true; | |
587 | } | |
588 | ||
589 | static inline void perf_detach_cgroup(struct perf_event *event) | |
590 | {} | |
591 | ||
592 | static inline int is_cgroup_event(struct perf_event *event) | |
593 | { | |
594 | return 0; | |
595 | } | |
596 | ||
597 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
598 | { | |
599 | return 0; | |
600 | } | |
601 | ||
602 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
603 | { | |
604 | } | |
605 | ||
606 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
607 | { | |
608 | } | |
609 | ||
a8d757ef SE |
610 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
611 | struct task_struct *next) | |
e5d1367f SE |
612 | { |
613 | } | |
614 | ||
a8d757ef SE |
615 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
616 | struct task_struct *task) | |
e5d1367f SE |
617 | { |
618 | } | |
619 | ||
620 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
621 | struct perf_event_attr *attr, | |
622 | struct perf_event *group_leader) | |
623 | { | |
624 | return -EINVAL; | |
625 | } | |
626 | ||
627 | static inline void | |
3f7cce3c SE |
628 | perf_cgroup_set_timestamp(struct task_struct *task, |
629 | struct perf_event_context *ctx) | |
e5d1367f SE |
630 | { |
631 | } | |
632 | ||
633 | void | |
634 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
635 | { | |
636 | } | |
637 | ||
638 | static inline void | |
639 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
640 | { | |
641 | } | |
642 | ||
643 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
644 | { | |
645 | return 0; | |
646 | } | |
647 | ||
648 | static inline void | |
649 | perf_cgroup_defer_enabled(struct perf_event *event) | |
650 | { | |
651 | } | |
652 | ||
653 | static inline void | |
654 | perf_cgroup_mark_enabled(struct perf_event *event, | |
655 | struct perf_event_context *ctx) | |
656 | { | |
657 | } | |
658 | #endif | |
659 | ||
9e630205 SE |
660 | /* |
661 | * set default to be dependent on timer tick just | |
662 | * like original code | |
663 | */ | |
664 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
665 | /* | |
666 | * function must be called with interrupts disbled | |
667 | */ | |
668 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
669 | { | |
670 | struct perf_cpu_context *cpuctx; | |
671 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
672 | int rotations = 0; | |
673 | ||
674 | WARN_ON(!irqs_disabled()); | |
675 | ||
676 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
677 | ||
678 | rotations = perf_rotate_context(cpuctx); | |
679 | ||
680 | /* | |
681 | * arm timer if needed | |
682 | */ | |
683 | if (rotations) { | |
684 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
685 | ret = HRTIMER_RESTART; | |
686 | } | |
687 | ||
688 | return ret; | |
689 | } | |
690 | ||
691 | /* CPU is going down */ | |
692 | void perf_cpu_hrtimer_cancel(int cpu) | |
693 | { | |
694 | struct perf_cpu_context *cpuctx; | |
695 | struct pmu *pmu; | |
696 | unsigned long flags; | |
697 | ||
698 | if (WARN_ON(cpu != smp_processor_id())) | |
699 | return; | |
700 | ||
701 | local_irq_save(flags); | |
702 | ||
703 | rcu_read_lock(); | |
704 | ||
705 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
706 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
707 | ||
708 | if (pmu->task_ctx_nr == perf_sw_context) | |
709 | continue; | |
710 | ||
711 | hrtimer_cancel(&cpuctx->hrtimer); | |
712 | } | |
713 | ||
714 | rcu_read_unlock(); | |
715 | ||
716 | local_irq_restore(flags); | |
717 | } | |
718 | ||
719 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
720 | { | |
721 | struct hrtimer *hr = &cpuctx->hrtimer; | |
722 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 723 | int timer; |
9e630205 SE |
724 | |
725 | /* no multiplexing needed for SW PMU */ | |
726 | if (pmu->task_ctx_nr == perf_sw_context) | |
727 | return; | |
728 | ||
62b85639 SE |
729 | /* |
730 | * check default is sane, if not set then force to | |
731 | * default interval (1/tick) | |
732 | */ | |
733 | timer = pmu->hrtimer_interval_ms; | |
734 | if (timer < 1) | |
735 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
736 | ||
737 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
738 | |
739 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
740 | hr->function = perf_cpu_hrtimer_handler; | |
741 | } | |
742 | ||
743 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
744 | { | |
745 | struct hrtimer *hr = &cpuctx->hrtimer; | |
746 | struct pmu *pmu = cpuctx->ctx.pmu; | |
747 | ||
748 | /* not for SW PMU */ | |
749 | if (pmu->task_ctx_nr == perf_sw_context) | |
750 | return; | |
751 | ||
752 | if (hrtimer_active(hr)) | |
753 | return; | |
754 | ||
755 | if (!hrtimer_callback_running(hr)) | |
756 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
757 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
758 | } | |
759 | ||
33696fc0 | 760 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 761 | { |
33696fc0 PZ |
762 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
763 | if (!(*count)++) | |
764 | pmu->pmu_disable(pmu); | |
9e35ad38 | 765 | } |
9e35ad38 | 766 | |
33696fc0 | 767 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 768 | { |
33696fc0 PZ |
769 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
770 | if (!--(*count)) | |
771 | pmu->pmu_enable(pmu); | |
9e35ad38 | 772 | } |
9e35ad38 | 773 | |
e9d2b064 PZ |
774 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
775 | ||
776 | /* | |
777 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
778 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
779 | * disabled, while rotate_context is called from IRQ context. | |
780 | */ | |
108b02cf | 781 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 782 | { |
108b02cf | 783 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 784 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 785 | |
e9d2b064 | 786 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 787 | |
12351ef8 FW |
788 | if (list_empty(&cpuctx->rotation_list)) { |
789 | int was_empty = list_empty(head); | |
e9d2b064 | 790 | list_add(&cpuctx->rotation_list, head); |
12351ef8 FW |
791 | if (was_empty) |
792 | tick_nohz_full_kick(); | |
793 | } | |
9e35ad38 | 794 | } |
9e35ad38 | 795 | |
cdd6c482 | 796 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 797 | { |
e5289d4a | 798 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
799 | } |
800 | ||
cdd6c482 | 801 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 802 | { |
564c2b21 PM |
803 | if (atomic_dec_and_test(&ctx->refcount)) { |
804 | if (ctx->parent_ctx) | |
805 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
806 | if (ctx->task) |
807 | put_task_struct(ctx->task); | |
cb796ff3 | 808 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 809 | } |
a63eaf34 PM |
810 | } |
811 | ||
cdd6c482 | 812 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
813 | { |
814 | if (ctx->parent_ctx) { | |
815 | put_ctx(ctx->parent_ctx); | |
816 | ctx->parent_ctx = NULL; | |
817 | } | |
818 | } | |
819 | ||
6844c09d ACM |
820 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
821 | { | |
822 | /* | |
823 | * only top level events have the pid namespace they were created in | |
824 | */ | |
825 | if (event->parent) | |
826 | event = event->parent; | |
827 | ||
828 | return task_tgid_nr_ns(p, event->ns); | |
829 | } | |
830 | ||
831 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
832 | { | |
833 | /* | |
834 | * only top level events have the pid namespace they were created in | |
835 | */ | |
836 | if (event->parent) | |
837 | event = event->parent; | |
838 | ||
839 | return task_pid_nr_ns(p, event->ns); | |
840 | } | |
841 | ||
7f453c24 | 842 | /* |
cdd6c482 | 843 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
844 | * to userspace. |
845 | */ | |
cdd6c482 | 846 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 847 | { |
cdd6c482 | 848 | u64 id = event->id; |
7f453c24 | 849 | |
cdd6c482 IM |
850 | if (event->parent) |
851 | id = event->parent->id; | |
7f453c24 PZ |
852 | |
853 | return id; | |
854 | } | |
855 | ||
25346b93 | 856 | /* |
cdd6c482 | 857 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
858 | * This has to cope with with the fact that until it is locked, |
859 | * the context could get moved to another task. | |
860 | */ | |
cdd6c482 | 861 | static struct perf_event_context * |
8dc85d54 | 862 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 863 | { |
cdd6c482 | 864 | struct perf_event_context *ctx; |
25346b93 PM |
865 | |
866 | rcu_read_lock(); | |
9ed6060d | 867 | retry: |
8dc85d54 | 868 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
869 | if (ctx) { |
870 | /* | |
871 | * If this context is a clone of another, it might | |
872 | * get swapped for another underneath us by | |
cdd6c482 | 873 | * perf_event_task_sched_out, though the |
25346b93 PM |
874 | * rcu_read_lock() protects us from any context |
875 | * getting freed. Lock the context and check if it | |
876 | * got swapped before we could get the lock, and retry | |
877 | * if so. If we locked the right context, then it | |
878 | * can't get swapped on us any more. | |
879 | */ | |
e625cce1 | 880 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 881 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 882 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
25346b93 PM |
883 | goto retry; |
884 | } | |
b49a9e7e PZ |
885 | |
886 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 887 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
888 | ctx = NULL; |
889 | } | |
25346b93 PM |
890 | } |
891 | rcu_read_unlock(); | |
892 | return ctx; | |
893 | } | |
894 | ||
895 | /* | |
896 | * Get the context for a task and increment its pin_count so it | |
897 | * can't get swapped to another task. This also increments its | |
898 | * reference count so that the context can't get freed. | |
899 | */ | |
8dc85d54 PZ |
900 | static struct perf_event_context * |
901 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 902 | { |
cdd6c482 | 903 | struct perf_event_context *ctx; |
25346b93 PM |
904 | unsigned long flags; |
905 | ||
8dc85d54 | 906 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
907 | if (ctx) { |
908 | ++ctx->pin_count; | |
e625cce1 | 909 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
910 | } |
911 | return ctx; | |
912 | } | |
913 | ||
cdd6c482 | 914 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
915 | { |
916 | unsigned long flags; | |
917 | ||
e625cce1 | 918 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 919 | --ctx->pin_count; |
e625cce1 | 920 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
921 | } |
922 | ||
f67218c3 PZ |
923 | /* |
924 | * Update the record of the current time in a context. | |
925 | */ | |
926 | static void update_context_time(struct perf_event_context *ctx) | |
927 | { | |
928 | u64 now = perf_clock(); | |
929 | ||
930 | ctx->time += now - ctx->timestamp; | |
931 | ctx->timestamp = now; | |
932 | } | |
933 | ||
4158755d SE |
934 | static u64 perf_event_time(struct perf_event *event) |
935 | { | |
936 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
937 | |
938 | if (is_cgroup_event(event)) | |
939 | return perf_cgroup_event_time(event); | |
940 | ||
4158755d SE |
941 | return ctx ? ctx->time : 0; |
942 | } | |
943 | ||
f67218c3 PZ |
944 | /* |
945 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 946 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
947 | */ |
948 | static void update_event_times(struct perf_event *event) | |
949 | { | |
950 | struct perf_event_context *ctx = event->ctx; | |
951 | u64 run_end; | |
952 | ||
953 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
954 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
955 | return; | |
e5d1367f SE |
956 | /* |
957 | * in cgroup mode, time_enabled represents | |
958 | * the time the event was enabled AND active | |
959 | * tasks were in the monitored cgroup. This is | |
960 | * independent of the activity of the context as | |
961 | * there may be a mix of cgroup and non-cgroup events. | |
962 | * | |
963 | * That is why we treat cgroup events differently | |
964 | * here. | |
965 | */ | |
966 | if (is_cgroup_event(event)) | |
46cd6a7f | 967 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
968 | else if (ctx->is_active) |
969 | run_end = ctx->time; | |
acd1d7c1 PZ |
970 | else |
971 | run_end = event->tstamp_stopped; | |
972 | ||
973 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
974 | |
975 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
976 | run_end = event->tstamp_stopped; | |
977 | else | |
4158755d | 978 | run_end = perf_event_time(event); |
f67218c3 PZ |
979 | |
980 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 981 | |
f67218c3 PZ |
982 | } |
983 | ||
96c21a46 PZ |
984 | /* |
985 | * Update total_time_enabled and total_time_running for all events in a group. | |
986 | */ | |
987 | static void update_group_times(struct perf_event *leader) | |
988 | { | |
989 | struct perf_event *event; | |
990 | ||
991 | update_event_times(leader); | |
992 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
993 | update_event_times(event); | |
994 | } | |
995 | ||
889ff015 FW |
996 | static struct list_head * |
997 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
998 | { | |
999 | if (event->attr.pinned) | |
1000 | return &ctx->pinned_groups; | |
1001 | else | |
1002 | return &ctx->flexible_groups; | |
1003 | } | |
1004 | ||
fccc714b | 1005 | /* |
cdd6c482 | 1006 | * Add a event from the lists for its context. |
fccc714b PZ |
1007 | * Must be called with ctx->mutex and ctx->lock held. |
1008 | */ | |
04289bb9 | 1009 | static void |
cdd6c482 | 1010 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1011 | { |
8a49542c PZ |
1012 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1013 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1014 | |
1015 | /* | |
8a49542c PZ |
1016 | * If we're a stand alone event or group leader, we go to the context |
1017 | * list, group events are kept attached to the group so that | |
1018 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1019 | */ |
8a49542c | 1020 | if (event->group_leader == event) { |
889ff015 FW |
1021 | struct list_head *list; |
1022 | ||
d6f962b5 FW |
1023 | if (is_software_event(event)) |
1024 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1025 | ||
889ff015 FW |
1026 | list = ctx_group_list(event, ctx); |
1027 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1028 | } |
592903cd | 1029 | |
08309379 | 1030 | if (is_cgroup_event(event)) |
e5d1367f | 1031 | ctx->nr_cgroups++; |
e5d1367f | 1032 | |
d010b332 SE |
1033 | if (has_branch_stack(event)) |
1034 | ctx->nr_branch_stack++; | |
1035 | ||
cdd6c482 | 1036 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1037 | if (!ctx->nr_events) |
108b02cf | 1038 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1039 | ctx->nr_events++; |
1040 | if (event->attr.inherit_stat) | |
bfbd3381 | 1041 | ctx->nr_stat++; |
04289bb9 IM |
1042 | } |
1043 | ||
0231bb53 JO |
1044 | /* |
1045 | * Initialize event state based on the perf_event_attr::disabled. | |
1046 | */ | |
1047 | static inline void perf_event__state_init(struct perf_event *event) | |
1048 | { | |
1049 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1050 | PERF_EVENT_STATE_INACTIVE; | |
1051 | } | |
1052 | ||
c320c7b7 ACM |
1053 | /* |
1054 | * Called at perf_event creation and when events are attached/detached from a | |
1055 | * group. | |
1056 | */ | |
1057 | static void perf_event__read_size(struct perf_event *event) | |
1058 | { | |
1059 | int entry = sizeof(u64); /* value */ | |
1060 | int size = 0; | |
1061 | int nr = 1; | |
1062 | ||
1063 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1064 | size += sizeof(u64); | |
1065 | ||
1066 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1067 | size += sizeof(u64); | |
1068 | ||
1069 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1070 | entry += sizeof(u64); | |
1071 | ||
1072 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1073 | nr += event->group_leader->nr_siblings; | |
1074 | size += sizeof(u64); | |
1075 | } | |
1076 | ||
1077 | size += entry * nr; | |
1078 | event->read_size = size; | |
1079 | } | |
1080 | ||
1081 | static void perf_event__header_size(struct perf_event *event) | |
1082 | { | |
1083 | struct perf_sample_data *data; | |
1084 | u64 sample_type = event->attr.sample_type; | |
1085 | u16 size = 0; | |
1086 | ||
1087 | perf_event__read_size(event); | |
1088 | ||
1089 | if (sample_type & PERF_SAMPLE_IP) | |
1090 | size += sizeof(data->ip); | |
1091 | ||
6844c09d ACM |
1092 | if (sample_type & PERF_SAMPLE_ADDR) |
1093 | size += sizeof(data->addr); | |
1094 | ||
1095 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1096 | size += sizeof(data->period); | |
1097 | ||
c3feedf2 AK |
1098 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1099 | size += sizeof(data->weight); | |
1100 | ||
6844c09d ACM |
1101 | if (sample_type & PERF_SAMPLE_READ) |
1102 | size += event->read_size; | |
1103 | ||
d6be9ad6 SE |
1104 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1105 | size += sizeof(data->data_src.val); | |
1106 | ||
6844c09d ACM |
1107 | event->header_size = size; |
1108 | } | |
1109 | ||
1110 | static void perf_event__id_header_size(struct perf_event *event) | |
1111 | { | |
1112 | struct perf_sample_data *data; | |
1113 | u64 sample_type = event->attr.sample_type; | |
1114 | u16 size = 0; | |
1115 | ||
c320c7b7 ACM |
1116 | if (sample_type & PERF_SAMPLE_TID) |
1117 | size += sizeof(data->tid_entry); | |
1118 | ||
1119 | if (sample_type & PERF_SAMPLE_TIME) | |
1120 | size += sizeof(data->time); | |
1121 | ||
c320c7b7 ACM |
1122 | if (sample_type & PERF_SAMPLE_ID) |
1123 | size += sizeof(data->id); | |
1124 | ||
1125 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1126 | size += sizeof(data->stream_id); | |
1127 | ||
1128 | if (sample_type & PERF_SAMPLE_CPU) | |
1129 | size += sizeof(data->cpu_entry); | |
1130 | ||
6844c09d | 1131 | event->id_header_size = size; |
c320c7b7 ACM |
1132 | } |
1133 | ||
8a49542c PZ |
1134 | static void perf_group_attach(struct perf_event *event) |
1135 | { | |
c320c7b7 | 1136 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1137 | |
74c3337c PZ |
1138 | /* |
1139 | * We can have double attach due to group movement in perf_event_open. | |
1140 | */ | |
1141 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1142 | return; | |
1143 | ||
8a49542c PZ |
1144 | event->attach_state |= PERF_ATTACH_GROUP; |
1145 | ||
1146 | if (group_leader == event) | |
1147 | return; | |
1148 | ||
1149 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1150 | !is_software_event(event)) | |
1151 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1152 | ||
1153 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1154 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1155 | |
1156 | perf_event__header_size(group_leader); | |
1157 | ||
1158 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1159 | perf_event__header_size(pos); | |
8a49542c PZ |
1160 | } |
1161 | ||
a63eaf34 | 1162 | /* |
cdd6c482 | 1163 | * Remove a event from the lists for its context. |
fccc714b | 1164 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1165 | */ |
04289bb9 | 1166 | static void |
cdd6c482 | 1167 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1168 | { |
68cacd29 | 1169 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1170 | /* |
1171 | * We can have double detach due to exit/hot-unplug + close. | |
1172 | */ | |
1173 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1174 | return; |
8a49542c PZ |
1175 | |
1176 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1177 | ||
68cacd29 | 1178 | if (is_cgroup_event(event)) { |
e5d1367f | 1179 | ctx->nr_cgroups--; |
68cacd29 SE |
1180 | cpuctx = __get_cpu_context(ctx); |
1181 | /* | |
1182 | * if there are no more cgroup events | |
1183 | * then cler cgrp to avoid stale pointer | |
1184 | * in update_cgrp_time_from_cpuctx() | |
1185 | */ | |
1186 | if (!ctx->nr_cgroups) | |
1187 | cpuctx->cgrp = NULL; | |
1188 | } | |
e5d1367f | 1189 | |
d010b332 SE |
1190 | if (has_branch_stack(event)) |
1191 | ctx->nr_branch_stack--; | |
1192 | ||
cdd6c482 IM |
1193 | ctx->nr_events--; |
1194 | if (event->attr.inherit_stat) | |
bfbd3381 | 1195 | ctx->nr_stat--; |
8bc20959 | 1196 | |
cdd6c482 | 1197 | list_del_rcu(&event->event_entry); |
04289bb9 | 1198 | |
8a49542c PZ |
1199 | if (event->group_leader == event) |
1200 | list_del_init(&event->group_entry); | |
5c148194 | 1201 | |
96c21a46 | 1202 | update_group_times(event); |
b2e74a26 SE |
1203 | |
1204 | /* | |
1205 | * If event was in error state, then keep it | |
1206 | * that way, otherwise bogus counts will be | |
1207 | * returned on read(). The only way to get out | |
1208 | * of error state is by explicit re-enabling | |
1209 | * of the event | |
1210 | */ | |
1211 | if (event->state > PERF_EVENT_STATE_OFF) | |
1212 | event->state = PERF_EVENT_STATE_OFF; | |
050735b0 PZ |
1213 | } |
1214 | ||
8a49542c | 1215 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1216 | { |
1217 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1218 | struct list_head *list = NULL; |
1219 | ||
1220 | /* | |
1221 | * We can have double detach due to exit/hot-unplug + close. | |
1222 | */ | |
1223 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1224 | return; | |
1225 | ||
1226 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1227 | ||
1228 | /* | |
1229 | * If this is a sibling, remove it from its group. | |
1230 | */ | |
1231 | if (event->group_leader != event) { | |
1232 | list_del_init(&event->group_entry); | |
1233 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1234 | goto out; |
8a49542c PZ |
1235 | } |
1236 | ||
1237 | if (!list_empty(&event->group_entry)) | |
1238 | list = &event->group_entry; | |
2e2af50b | 1239 | |
04289bb9 | 1240 | /* |
cdd6c482 IM |
1241 | * If this was a group event with sibling events then |
1242 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1243 | * to whatever list we are on. |
04289bb9 | 1244 | */ |
cdd6c482 | 1245 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1246 | if (list) |
1247 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1248 | sibling->group_leader = sibling; |
d6f962b5 FW |
1249 | |
1250 | /* Inherit group flags from the previous leader */ | |
1251 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1252 | } |
c320c7b7 ACM |
1253 | |
1254 | out: | |
1255 | perf_event__header_size(event->group_leader); | |
1256 | ||
1257 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1258 | perf_event__header_size(tmp); | |
04289bb9 IM |
1259 | } |
1260 | ||
fa66f07a SE |
1261 | static inline int |
1262 | event_filter_match(struct perf_event *event) | |
1263 | { | |
e5d1367f SE |
1264 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1265 | && perf_cgroup_match(event); | |
fa66f07a SE |
1266 | } |
1267 | ||
9ffcfa6f SE |
1268 | static void |
1269 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1270 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1271 | struct perf_event_context *ctx) |
3b6f9e5c | 1272 | { |
4158755d | 1273 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1274 | u64 delta; |
1275 | /* | |
1276 | * An event which could not be activated because of | |
1277 | * filter mismatch still needs to have its timings | |
1278 | * maintained, otherwise bogus information is return | |
1279 | * via read() for time_enabled, time_running: | |
1280 | */ | |
1281 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1282 | && !event_filter_match(event)) { | |
e5d1367f | 1283 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1284 | event->tstamp_running += delta; |
4158755d | 1285 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1286 | } |
1287 | ||
cdd6c482 | 1288 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1289 | return; |
3b6f9e5c | 1290 | |
cdd6c482 IM |
1291 | event->state = PERF_EVENT_STATE_INACTIVE; |
1292 | if (event->pending_disable) { | |
1293 | event->pending_disable = 0; | |
1294 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1295 | } |
4158755d | 1296 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1297 | event->pmu->del(event, 0); |
cdd6c482 | 1298 | event->oncpu = -1; |
3b6f9e5c | 1299 | |
cdd6c482 | 1300 | if (!is_software_event(event)) |
3b6f9e5c PM |
1301 | cpuctx->active_oncpu--; |
1302 | ctx->nr_active--; | |
0f5a2601 PZ |
1303 | if (event->attr.freq && event->attr.sample_freq) |
1304 | ctx->nr_freq--; | |
cdd6c482 | 1305 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c PM |
1306 | cpuctx->exclusive = 0; |
1307 | } | |
1308 | ||
d859e29f | 1309 | static void |
cdd6c482 | 1310 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1311 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1312 | struct perf_event_context *ctx) |
d859e29f | 1313 | { |
cdd6c482 | 1314 | struct perf_event *event; |
fa66f07a | 1315 | int state = group_event->state; |
d859e29f | 1316 | |
cdd6c482 | 1317 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1318 | |
1319 | /* | |
1320 | * Schedule out siblings (if any): | |
1321 | */ | |
cdd6c482 IM |
1322 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1323 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1324 | |
fa66f07a | 1325 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1326 | cpuctx->exclusive = 0; |
1327 | } | |
1328 | ||
0793a61d | 1329 | /* |
cdd6c482 | 1330 | * Cross CPU call to remove a performance event |
0793a61d | 1331 | * |
cdd6c482 | 1332 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1333 | * remove it from the context list. |
1334 | */ | |
fe4b04fa | 1335 | static int __perf_remove_from_context(void *info) |
0793a61d | 1336 | { |
cdd6c482 IM |
1337 | struct perf_event *event = info; |
1338 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1339 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1340 | |
e625cce1 | 1341 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1342 | event_sched_out(event, cpuctx, ctx); |
cdd6c482 | 1343 | list_del_event(event, ctx); |
64ce3126 PZ |
1344 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1345 | ctx->is_active = 0; | |
1346 | cpuctx->task_ctx = NULL; | |
1347 | } | |
e625cce1 | 1348 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1349 | |
1350 | return 0; | |
0793a61d TG |
1351 | } |
1352 | ||
1353 | ||
1354 | /* | |
cdd6c482 | 1355 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1356 | * |
cdd6c482 | 1357 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1358 | * call when the task is on a CPU. |
c93f7669 | 1359 | * |
cdd6c482 IM |
1360 | * If event->ctx is a cloned context, callers must make sure that |
1361 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1362 | * remains valid. This is OK when called from perf_release since |
1363 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1364 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1365 | * context has been detached from its task. |
0793a61d | 1366 | */ |
fe4b04fa | 1367 | static void perf_remove_from_context(struct perf_event *event) |
0793a61d | 1368 | { |
cdd6c482 | 1369 | struct perf_event_context *ctx = event->ctx; |
0793a61d TG |
1370 | struct task_struct *task = ctx->task; |
1371 | ||
fe4b04fa PZ |
1372 | lockdep_assert_held(&ctx->mutex); |
1373 | ||
0793a61d TG |
1374 | if (!task) { |
1375 | /* | |
cdd6c482 | 1376 | * Per cpu events are removed via an smp call and |
af901ca1 | 1377 | * the removal is always successful. |
0793a61d | 1378 | */ |
fe4b04fa | 1379 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
0793a61d TG |
1380 | return; |
1381 | } | |
1382 | ||
1383 | retry: | |
fe4b04fa PZ |
1384 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1385 | return; | |
0793a61d | 1386 | |
e625cce1 | 1387 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1388 | /* |
fe4b04fa PZ |
1389 | * If we failed to find a running task, but find the context active now |
1390 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1391 | */ |
fe4b04fa | 1392 | if (ctx->is_active) { |
e625cce1 | 1393 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1394 | goto retry; |
1395 | } | |
1396 | ||
1397 | /* | |
fe4b04fa PZ |
1398 | * Since the task isn't running, its safe to remove the event, us |
1399 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1400 | */ |
fe4b04fa | 1401 | list_del_event(event, ctx); |
e625cce1 | 1402 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1403 | } |
1404 | ||
d859e29f | 1405 | /* |
cdd6c482 | 1406 | * Cross CPU call to disable a performance event |
d859e29f | 1407 | */ |
500ad2d8 | 1408 | int __perf_event_disable(void *info) |
d859e29f | 1409 | { |
cdd6c482 | 1410 | struct perf_event *event = info; |
cdd6c482 | 1411 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1412 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1413 | |
1414 | /* | |
cdd6c482 IM |
1415 | * If this is a per-task event, need to check whether this |
1416 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1417 | * |
1418 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1419 | * flipping contexts around. | |
d859e29f | 1420 | */ |
665c2142 | 1421 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1422 | return -EINVAL; |
d859e29f | 1423 | |
e625cce1 | 1424 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1425 | |
1426 | /* | |
cdd6c482 | 1427 | * If the event is on, turn it off. |
d859e29f PM |
1428 | * If it is in error state, leave it in error state. |
1429 | */ | |
cdd6c482 | 1430 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1431 | update_context_time(ctx); |
e5d1367f | 1432 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1433 | update_group_times(event); |
1434 | if (event == event->group_leader) | |
1435 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1436 | else |
cdd6c482 IM |
1437 | event_sched_out(event, cpuctx, ctx); |
1438 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1439 | } |
1440 | ||
e625cce1 | 1441 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1442 | |
1443 | return 0; | |
d859e29f PM |
1444 | } |
1445 | ||
1446 | /* | |
cdd6c482 | 1447 | * Disable a event. |
c93f7669 | 1448 | * |
cdd6c482 IM |
1449 | * If event->ctx is a cloned context, callers must make sure that |
1450 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1451 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1452 | * perf_event_for_each_child or perf_event_for_each because they |
1453 | * hold the top-level event's child_mutex, so any descendant that | |
1454 | * goes to exit will block in sync_child_event. | |
1455 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1456 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1457 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1458 | */ |
44234adc | 1459 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1460 | { |
cdd6c482 | 1461 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1462 | struct task_struct *task = ctx->task; |
1463 | ||
1464 | if (!task) { | |
1465 | /* | |
cdd6c482 | 1466 | * Disable the event on the cpu that it's on |
d859e29f | 1467 | */ |
fe4b04fa | 1468 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1469 | return; |
1470 | } | |
1471 | ||
9ed6060d | 1472 | retry: |
fe4b04fa PZ |
1473 | if (!task_function_call(task, __perf_event_disable, event)) |
1474 | return; | |
d859e29f | 1475 | |
e625cce1 | 1476 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1477 | /* |
cdd6c482 | 1478 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1479 | */ |
cdd6c482 | 1480 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1481 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1482 | /* |
1483 | * Reload the task pointer, it might have been changed by | |
1484 | * a concurrent perf_event_context_sched_out(). | |
1485 | */ | |
1486 | task = ctx->task; | |
d859e29f PM |
1487 | goto retry; |
1488 | } | |
1489 | ||
1490 | /* | |
1491 | * Since we have the lock this context can't be scheduled | |
1492 | * in, so we can change the state safely. | |
1493 | */ | |
cdd6c482 IM |
1494 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1495 | update_group_times(event); | |
1496 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1497 | } |
e625cce1 | 1498 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1499 | } |
dcfce4a0 | 1500 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1501 | |
e5d1367f SE |
1502 | static void perf_set_shadow_time(struct perf_event *event, |
1503 | struct perf_event_context *ctx, | |
1504 | u64 tstamp) | |
1505 | { | |
1506 | /* | |
1507 | * use the correct time source for the time snapshot | |
1508 | * | |
1509 | * We could get by without this by leveraging the | |
1510 | * fact that to get to this function, the caller | |
1511 | * has most likely already called update_context_time() | |
1512 | * and update_cgrp_time_xx() and thus both timestamp | |
1513 | * are identical (or very close). Given that tstamp is, | |
1514 | * already adjusted for cgroup, we could say that: | |
1515 | * tstamp - ctx->timestamp | |
1516 | * is equivalent to | |
1517 | * tstamp - cgrp->timestamp. | |
1518 | * | |
1519 | * Then, in perf_output_read(), the calculation would | |
1520 | * work with no changes because: | |
1521 | * - event is guaranteed scheduled in | |
1522 | * - no scheduled out in between | |
1523 | * - thus the timestamp would be the same | |
1524 | * | |
1525 | * But this is a bit hairy. | |
1526 | * | |
1527 | * So instead, we have an explicit cgroup call to remain | |
1528 | * within the time time source all along. We believe it | |
1529 | * is cleaner and simpler to understand. | |
1530 | */ | |
1531 | if (is_cgroup_event(event)) | |
1532 | perf_cgroup_set_shadow_time(event, tstamp); | |
1533 | else | |
1534 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1535 | } | |
1536 | ||
4fe757dd PZ |
1537 | #define MAX_INTERRUPTS (~0ULL) |
1538 | ||
1539 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1540 | ||
235c7fc7 | 1541 | static int |
9ffcfa6f | 1542 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1543 | struct perf_cpu_context *cpuctx, |
6e37738a | 1544 | struct perf_event_context *ctx) |
235c7fc7 | 1545 | { |
4158755d SE |
1546 | u64 tstamp = perf_event_time(event); |
1547 | ||
cdd6c482 | 1548 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1549 | return 0; |
1550 | ||
cdd6c482 | 1551 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1552 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1553 | |
1554 | /* | |
1555 | * Unthrottle events, since we scheduled we might have missed several | |
1556 | * ticks already, also for a heavily scheduling task there is little | |
1557 | * guarantee it'll get a tick in a timely manner. | |
1558 | */ | |
1559 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1560 | perf_log_throttle(event, 1); | |
1561 | event->hw.interrupts = 0; | |
1562 | } | |
1563 | ||
235c7fc7 IM |
1564 | /* |
1565 | * The new state must be visible before we turn it on in the hardware: | |
1566 | */ | |
1567 | smp_wmb(); | |
1568 | ||
a4eaf7f1 | 1569 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1570 | event->state = PERF_EVENT_STATE_INACTIVE; |
1571 | event->oncpu = -1; | |
235c7fc7 IM |
1572 | return -EAGAIN; |
1573 | } | |
1574 | ||
4158755d | 1575 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1576 | |
e5d1367f | 1577 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1578 | |
cdd6c482 | 1579 | if (!is_software_event(event)) |
3b6f9e5c | 1580 | cpuctx->active_oncpu++; |
235c7fc7 | 1581 | ctx->nr_active++; |
0f5a2601 PZ |
1582 | if (event->attr.freq && event->attr.sample_freq) |
1583 | ctx->nr_freq++; | |
235c7fc7 | 1584 | |
cdd6c482 | 1585 | if (event->attr.exclusive) |
3b6f9e5c PM |
1586 | cpuctx->exclusive = 1; |
1587 | ||
235c7fc7 IM |
1588 | return 0; |
1589 | } | |
1590 | ||
6751b71e | 1591 | static int |
cdd6c482 | 1592 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1593 | struct perf_cpu_context *cpuctx, |
6e37738a | 1594 | struct perf_event_context *ctx) |
6751b71e | 1595 | { |
6bde9b6c | 1596 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1597 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1598 | u64 now = ctx->time; |
1599 | bool simulate = false; | |
6751b71e | 1600 | |
cdd6c482 | 1601 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1602 | return 0; |
1603 | ||
ad5133b7 | 1604 | pmu->start_txn(pmu); |
6bde9b6c | 1605 | |
9ffcfa6f | 1606 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1607 | pmu->cancel_txn(pmu); |
9e630205 | 1608 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1609 | return -EAGAIN; |
90151c35 | 1610 | } |
6751b71e PM |
1611 | |
1612 | /* | |
1613 | * Schedule in siblings as one group (if any): | |
1614 | */ | |
cdd6c482 | 1615 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1616 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1617 | partial_group = event; |
6751b71e PM |
1618 | goto group_error; |
1619 | } | |
1620 | } | |
1621 | ||
9ffcfa6f | 1622 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1623 | return 0; |
9ffcfa6f | 1624 | |
6751b71e PM |
1625 | group_error: |
1626 | /* | |
1627 | * Groups can be scheduled in as one unit only, so undo any | |
1628 | * partial group before returning: | |
d7842da4 SE |
1629 | * The events up to the failed event are scheduled out normally, |
1630 | * tstamp_stopped will be updated. | |
1631 | * | |
1632 | * The failed events and the remaining siblings need to have | |
1633 | * their timings updated as if they had gone thru event_sched_in() | |
1634 | * and event_sched_out(). This is required to get consistent timings | |
1635 | * across the group. This also takes care of the case where the group | |
1636 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1637 | * the time the event was actually stopped, such that time delta | |
1638 | * calculation in update_event_times() is correct. | |
6751b71e | 1639 | */ |
cdd6c482 IM |
1640 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1641 | if (event == partial_group) | |
d7842da4 SE |
1642 | simulate = true; |
1643 | ||
1644 | if (simulate) { | |
1645 | event->tstamp_running += now - event->tstamp_stopped; | |
1646 | event->tstamp_stopped = now; | |
1647 | } else { | |
1648 | event_sched_out(event, cpuctx, ctx); | |
1649 | } | |
6751b71e | 1650 | } |
9ffcfa6f | 1651 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1652 | |
ad5133b7 | 1653 | pmu->cancel_txn(pmu); |
90151c35 | 1654 | |
9e630205 SE |
1655 | perf_cpu_hrtimer_restart(cpuctx); |
1656 | ||
6751b71e PM |
1657 | return -EAGAIN; |
1658 | } | |
1659 | ||
3b6f9e5c | 1660 | /* |
cdd6c482 | 1661 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1662 | */ |
cdd6c482 | 1663 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1664 | struct perf_cpu_context *cpuctx, |
1665 | int can_add_hw) | |
1666 | { | |
1667 | /* | |
cdd6c482 | 1668 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1669 | */ |
d6f962b5 | 1670 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1671 | return 1; |
1672 | /* | |
1673 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1674 | * events can go on. |
3b6f9e5c PM |
1675 | */ |
1676 | if (cpuctx->exclusive) | |
1677 | return 0; | |
1678 | /* | |
1679 | * If this group is exclusive and there are already | |
cdd6c482 | 1680 | * events on the CPU, it can't go on. |
3b6f9e5c | 1681 | */ |
cdd6c482 | 1682 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1683 | return 0; |
1684 | /* | |
1685 | * Otherwise, try to add it if all previous groups were able | |
1686 | * to go on. | |
1687 | */ | |
1688 | return can_add_hw; | |
1689 | } | |
1690 | ||
cdd6c482 IM |
1691 | static void add_event_to_ctx(struct perf_event *event, |
1692 | struct perf_event_context *ctx) | |
53cfbf59 | 1693 | { |
4158755d SE |
1694 | u64 tstamp = perf_event_time(event); |
1695 | ||
cdd6c482 | 1696 | list_add_event(event, ctx); |
8a49542c | 1697 | perf_group_attach(event); |
4158755d SE |
1698 | event->tstamp_enabled = tstamp; |
1699 | event->tstamp_running = tstamp; | |
1700 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1701 | } |
1702 | ||
2c29ef0f PZ |
1703 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1704 | static void | |
1705 | ctx_sched_in(struct perf_event_context *ctx, | |
1706 | struct perf_cpu_context *cpuctx, | |
1707 | enum event_type_t event_type, | |
1708 | struct task_struct *task); | |
fe4b04fa | 1709 | |
dce5855b PZ |
1710 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1711 | struct perf_event_context *ctx, | |
1712 | struct task_struct *task) | |
1713 | { | |
1714 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1715 | if (ctx) | |
1716 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1717 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1718 | if (ctx) | |
1719 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1720 | } | |
1721 | ||
0793a61d | 1722 | /* |
cdd6c482 | 1723 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1724 | * |
1725 | * Must be called with ctx->mutex held | |
0793a61d | 1726 | */ |
fe4b04fa | 1727 | static int __perf_install_in_context(void *info) |
0793a61d | 1728 | { |
cdd6c482 IM |
1729 | struct perf_event *event = info; |
1730 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1731 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1732 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1733 | struct task_struct *task = current; | |
1734 | ||
b58f6b0d | 1735 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1736 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1737 | |
1738 | /* | |
2c29ef0f | 1739 | * If there was an active task_ctx schedule it out. |
0793a61d | 1740 | */ |
b58f6b0d | 1741 | if (task_ctx) |
2c29ef0f | 1742 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1743 | |
1744 | /* | |
1745 | * If the context we're installing events in is not the | |
1746 | * active task_ctx, flip them. | |
1747 | */ | |
1748 | if (ctx->task && task_ctx != ctx) { | |
1749 | if (task_ctx) | |
1750 | raw_spin_unlock(&task_ctx->lock); | |
1751 | raw_spin_lock(&ctx->lock); | |
1752 | task_ctx = ctx; | |
1753 | } | |
1754 | ||
1755 | if (task_ctx) { | |
1756 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1757 | task = task_ctx->task; |
1758 | } | |
b58f6b0d | 1759 | |
2c29ef0f | 1760 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1761 | |
4af4998b | 1762 | update_context_time(ctx); |
e5d1367f SE |
1763 | /* |
1764 | * update cgrp time only if current cgrp | |
1765 | * matches event->cgrp. Must be done before | |
1766 | * calling add_event_to_ctx() | |
1767 | */ | |
1768 | update_cgrp_time_from_event(event); | |
0793a61d | 1769 | |
cdd6c482 | 1770 | add_event_to_ctx(event, ctx); |
0793a61d | 1771 | |
d859e29f | 1772 | /* |
2c29ef0f | 1773 | * Schedule everything back in |
d859e29f | 1774 | */ |
dce5855b | 1775 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1776 | |
1777 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1778 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1779 | |
1780 | return 0; | |
0793a61d TG |
1781 | } |
1782 | ||
1783 | /* | |
cdd6c482 | 1784 | * Attach a performance event to a context |
0793a61d | 1785 | * |
cdd6c482 IM |
1786 | * First we add the event to the list with the hardware enable bit |
1787 | * in event->hw_config cleared. | |
0793a61d | 1788 | * |
cdd6c482 | 1789 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1790 | * call to enable it in the task context. The task might have been |
1791 | * scheduled away, but we check this in the smp call again. | |
1792 | */ | |
1793 | static void | |
cdd6c482 IM |
1794 | perf_install_in_context(struct perf_event_context *ctx, |
1795 | struct perf_event *event, | |
0793a61d TG |
1796 | int cpu) |
1797 | { | |
1798 | struct task_struct *task = ctx->task; | |
1799 | ||
fe4b04fa PZ |
1800 | lockdep_assert_held(&ctx->mutex); |
1801 | ||
c3f00c70 | 1802 | event->ctx = ctx; |
0cda4c02 YZ |
1803 | if (event->cpu != -1) |
1804 | event->cpu = cpu; | |
c3f00c70 | 1805 | |
0793a61d TG |
1806 | if (!task) { |
1807 | /* | |
cdd6c482 | 1808 | * Per cpu events are installed via an smp call and |
af901ca1 | 1809 | * the install is always successful. |
0793a61d | 1810 | */ |
fe4b04fa | 1811 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1812 | return; |
1813 | } | |
1814 | ||
0793a61d | 1815 | retry: |
fe4b04fa PZ |
1816 | if (!task_function_call(task, __perf_install_in_context, event)) |
1817 | return; | |
0793a61d | 1818 | |
e625cce1 | 1819 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1820 | /* |
fe4b04fa PZ |
1821 | * If we failed to find a running task, but find the context active now |
1822 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1823 | */ |
fe4b04fa | 1824 | if (ctx->is_active) { |
e625cce1 | 1825 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1826 | goto retry; |
1827 | } | |
1828 | ||
1829 | /* | |
fe4b04fa PZ |
1830 | * Since the task isn't running, its safe to add the event, us holding |
1831 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1832 | */ |
fe4b04fa | 1833 | add_event_to_ctx(event, ctx); |
e625cce1 | 1834 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1835 | } |
1836 | ||
fa289bec | 1837 | /* |
cdd6c482 | 1838 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1839 | * Enabling the leader of a group effectively enables all |
1840 | * the group members that aren't explicitly disabled, so we | |
1841 | * have to update their ->tstamp_enabled also. | |
1842 | * Note: this works for group members as well as group leaders | |
1843 | * since the non-leader members' sibling_lists will be empty. | |
1844 | */ | |
1d9b482e | 1845 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1846 | { |
cdd6c482 | 1847 | struct perf_event *sub; |
4158755d | 1848 | u64 tstamp = perf_event_time(event); |
fa289bec | 1849 | |
cdd6c482 | 1850 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1851 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1852 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1853 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1854 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1855 | } |
fa289bec PM |
1856 | } |
1857 | ||
d859e29f | 1858 | /* |
cdd6c482 | 1859 | * Cross CPU call to enable a performance event |
d859e29f | 1860 | */ |
fe4b04fa | 1861 | static int __perf_event_enable(void *info) |
04289bb9 | 1862 | { |
cdd6c482 | 1863 | struct perf_event *event = info; |
cdd6c482 IM |
1864 | struct perf_event_context *ctx = event->ctx; |
1865 | struct perf_event *leader = event->group_leader; | |
108b02cf | 1866 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 1867 | int err; |
04289bb9 | 1868 | |
fe4b04fa PZ |
1869 | if (WARN_ON_ONCE(!ctx->is_active)) |
1870 | return -EINVAL; | |
3cbed429 | 1871 | |
e625cce1 | 1872 | raw_spin_lock(&ctx->lock); |
4af4998b | 1873 | update_context_time(ctx); |
d859e29f | 1874 | |
cdd6c482 | 1875 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 1876 | goto unlock; |
e5d1367f SE |
1877 | |
1878 | /* | |
1879 | * set current task's cgroup time reference point | |
1880 | */ | |
3f7cce3c | 1881 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 1882 | |
1d9b482e | 1883 | __perf_event_mark_enabled(event); |
04289bb9 | 1884 | |
e5d1367f SE |
1885 | if (!event_filter_match(event)) { |
1886 | if (is_cgroup_event(event)) | |
1887 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 1888 | goto unlock; |
e5d1367f | 1889 | } |
f4c4176f | 1890 | |
04289bb9 | 1891 | /* |
cdd6c482 | 1892 | * If the event is in a group and isn't the group leader, |
d859e29f | 1893 | * then don't put it on unless the group is on. |
04289bb9 | 1894 | */ |
cdd6c482 | 1895 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 1896 | goto unlock; |
3b6f9e5c | 1897 | |
cdd6c482 | 1898 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 1899 | err = -EEXIST; |
e758a33d | 1900 | } else { |
cdd6c482 | 1901 | if (event == leader) |
6e37738a | 1902 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 1903 | else |
6e37738a | 1904 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 1905 | } |
d859e29f PM |
1906 | |
1907 | if (err) { | |
1908 | /* | |
cdd6c482 | 1909 | * If this event can't go on and it's part of a |
d859e29f PM |
1910 | * group, then the whole group has to come off. |
1911 | */ | |
9e630205 | 1912 | if (leader != event) { |
d859e29f | 1913 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
1914 | perf_cpu_hrtimer_restart(cpuctx); |
1915 | } | |
0d48696f | 1916 | if (leader->attr.pinned) { |
53cfbf59 | 1917 | update_group_times(leader); |
cdd6c482 | 1918 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 1919 | } |
d859e29f PM |
1920 | } |
1921 | ||
9ed6060d | 1922 | unlock: |
e625cce1 | 1923 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1924 | |
1925 | return 0; | |
d859e29f PM |
1926 | } |
1927 | ||
1928 | /* | |
cdd6c482 | 1929 | * Enable a event. |
c93f7669 | 1930 | * |
cdd6c482 IM |
1931 | * If event->ctx is a cloned context, callers must make sure that |
1932 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1933 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
1934 | * perf_event_for_each_child or perf_event_for_each as described |
1935 | * for perf_event_disable. | |
d859e29f | 1936 | */ |
44234adc | 1937 | void perf_event_enable(struct perf_event *event) |
d859e29f | 1938 | { |
cdd6c482 | 1939 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1940 | struct task_struct *task = ctx->task; |
1941 | ||
1942 | if (!task) { | |
1943 | /* | |
cdd6c482 | 1944 | * Enable the event on the cpu that it's on |
d859e29f | 1945 | */ |
fe4b04fa | 1946 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
1947 | return; |
1948 | } | |
1949 | ||
e625cce1 | 1950 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 1951 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
1952 | goto out; |
1953 | ||
1954 | /* | |
cdd6c482 IM |
1955 | * If the event is in error state, clear that first. |
1956 | * That way, if we see the event in error state below, we | |
d859e29f PM |
1957 | * know that it has gone back into error state, as distinct |
1958 | * from the task having been scheduled away before the | |
1959 | * cross-call arrived. | |
1960 | */ | |
cdd6c482 IM |
1961 | if (event->state == PERF_EVENT_STATE_ERROR) |
1962 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 1963 | |
9ed6060d | 1964 | retry: |
fe4b04fa | 1965 | if (!ctx->is_active) { |
1d9b482e | 1966 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
1967 | goto out; |
1968 | } | |
1969 | ||
e625cce1 | 1970 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1971 | |
1972 | if (!task_function_call(task, __perf_event_enable, event)) | |
1973 | return; | |
d859e29f | 1974 | |
e625cce1 | 1975 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
1976 | |
1977 | /* | |
cdd6c482 | 1978 | * If the context is active and the event is still off, |
d859e29f PM |
1979 | * we need to retry the cross-call. |
1980 | */ | |
fe4b04fa PZ |
1981 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
1982 | /* | |
1983 | * task could have been flipped by a concurrent | |
1984 | * perf_event_context_sched_out() | |
1985 | */ | |
1986 | task = ctx->task; | |
d859e29f | 1987 | goto retry; |
fe4b04fa | 1988 | } |
fa289bec | 1989 | |
9ed6060d | 1990 | out: |
e625cce1 | 1991 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1992 | } |
dcfce4a0 | 1993 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 1994 | |
26ca5c11 | 1995 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 1996 | { |
2023b359 | 1997 | /* |
cdd6c482 | 1998 | * not supported on inherited events |
2023b359 | 1999 | */ |
2e939d1d | 2000 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2001 | return -EINVAL; |
2002 | ||
cdd6c482 IM |
2003 | atomic_add(refresh, &event->event_limit); |
2004 | perf_event_enable(event); | |
2023b359 PZ |
2005 | |
2006 | return 0; | |
79f14641 | 2007 | } |
26ca5c11 | 2008 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2009 | |
5b0311e1 FW |
2010 | static void ctx_sched_out(struct perf_event_context *ctx, |
2011 | struct perf_cpu_context *cpuctx, | |
2012 | enum event_type_t event_type) | |
235c7fc7 | 2013 | { |
cdd6c482 | 2014 | struct perf_event *event; |
db24d33e | 2015 | int is_active = ctx->is_active; |
235c7fc7 | 2016 | |
db24d33e | 2017 | ctx->is_active &= ~event_type; |
cdd6c482 | 2018 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2019 | return; |
2020 | ||
4af4998b | 2021 | update_context_time(ctx); |
e5d1367f | 2022 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2023 | if (!ctx->nr_active) |
facc4307 | 2024 | return; |
5b0311e1 | 2025 | |
075e0b00 | 2026 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2027 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2028 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2029 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2030 | } |
889ff015 | 2031 | |
db24d33e | 2032 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2033 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2034 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2035 | } |
1b9a644f | 2036 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2037 | } |
2038 | ||
564c2b21 PM |
2039 | /* |
2040 | * Test whether two contexts are equivalent, i.e. whether they | |
2041 | * have both been cloned from the same version of the same context | |
cdd6c482 IM |
2042 | * and they both have the same number of enabled events. |
2043 | * If the number of enabled events is the same, then the set | |
2044 | * of enabled events should be the same, because these are both | |
2045 | * inherited contexts, therefore we can't access individual events | |
564c2b21 | 2046 | * in them directly with an fd; we can only enable/disable all |
cdd6c482 | 2047 | * events via prctl, or enable/disable all events in a family |
564c2b21 PM |
2048 | * via ioctl, which will have the same effect on both contexts. |
2049 | */ | |
cdd6c482 IM |
2050 | static int context_equiv(struct perf_event_context *ctx1, |
2051 | struct perf_event_context *ctx2) | |
564c2b21 PM |
2052 | { |
2053 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx | |
ad3a37de | 2054 | && ctx1->parent_gen == ctx2->parent_gen |
25346b93 | 2055 | && !ctx1->pin_count && !ctx2->pin_count; |
564c2b21 PM |
2056 | } |
2057 | ||
cdd6c482 IM |
2058 | static void __perf_event_sync_stat(struct perf_event *event, |
2059 | struct perf_event *next_event) | |
bfbd3381 PZ |
2060 | { |
2061 | u64 value; | |
2062 | ||
cdd6c482 | 2063 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2064 | return; |
2065 | ||
2066 | /* | |
cdd6c482 | 2067 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2068 | * because we're in the middle of a context switch and have IRQs |
2069 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2070 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2071 | * don't need to use it. |
2072 | */ | |
cdd6c482 IM |
2073 | switch (event->state) { |
2074 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2075 | event->pmu->read(event); |
2076 | /* fall-through */ | |
bfbd3381 | 2077 | |
cdd6c482 IM |
2078 | case PERF_EVENT_STATE_INACTIVE: |
2079 | update_event_times(event); | |
bfbd3381 PZ |
2080 | break; |
2081 | ||
2082 | default: | |
2083 | break; | |
2084 | } | |
2085 | ||
2086 | /* | |
cdd6c482 | 2087 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2088 | * values when we flip the contexts. |
2089 | */ | |
e7850595 PZ |
2090 | value = local64_read(&next_event->count); |
2091 | value = local64_xchg(&event->count, value); | |
2092 | local64_set(&next_event->count, value); | |
bfbd3381 | 2093 | |
cdd6c482 IM |
2094 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2095 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2096 | |
bfbd3381 | 2097 | /* |
19d2e755 | 2098 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2099 | */ |
cdd6c482 IM |
2100 | perf_event_update_userpage(event); |
2101 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2102 | } |
2103 | ||
2104 | #define list_next_entry(pos, member) \ | |
2105 | list_entry(pos->member.next, typeof(*pos), member) | |
2106 | ||
cdd6c482 IM |
2107 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2108 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2109 | { |
cdd6c482 | 2110 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2111 | |
2112 | if (!ctx->nr_stat) | |
2113 | return; | |
2114 | ||
02ffdbc8 PZ |
2115 | update_context_time(ctx); |
2116 | ||
cdd6c482 IM |
2117 | event = list_first_entry(&ctx->event_list, |
2118 | struct perf_event, event_entry); | |
bfbd3381 | 2119 | |
cdd6c482 IM |
2120 | next_event = list_first_entry(&next_ctx->event_list, |
2121 | struct perf_event, event_entry); | |
bfbd3381 | 2122 | |
cdd6c482 IM |
2123 | while (&event->event_entry != &ctx->event_list && |
2124 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2125 | |
cdd6c482 | 2126 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2127 | |
cdd6c482 IM |
2128 | event = list_next_entry(event, event_entry); |
2129 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2130 | } |
2131 | } | |
2132 | ||
fe4b04fa PZ |
2133 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2134 | struct task_struct *next) | |
0793a61d | 2135 | { |
8dc85d54 | 2136 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 IM |
2137 | struct perf_event_context *next_ctx; |
2138 | struct perf_event_context *parent; | |
108b02cf | 2139 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2140 | int do_switch = 1; |
0793a61d | 2141 | |
108b02cf PZ |
2142 | if (likely(!ctx)) |
2143 | return; | |
10989fb2 | 2144 | |
108b02cf PZ |
2145 | cpuctx = __get_cpu_context(ctx); |
2146 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2147 | return; |
2148 | ||
c93f7669 PM |
2149 | rcu_read_lock(); |
2150 | parent = rcu_dereference(ctx->parent_ctx); | |
8dc85d54 | 2151 | next_ctx = next->perf_event_ctxp[ctxn]; |
c93f7669 PM |
2152 | if (parent && next_ctx && |
2153 | rcu_dereference(next_ctx->parent_ctx) == parent) { | |
2154 | /* | |
2155 | * Looks like the two contexts are clones, so we might be | |
2156 | * able to optimize the context switch. We lock both | |
2157 | * contexts and check that they are clones under the | |
2158 | * lock (including re-checking that neither has been | |
2159 | * uncloned in the meantime). It doesn't matter which | |
2160 | * order we take the locks because no other cpu could | |
2161 | * be trying to lock both of these tasks. | |
2162 | */ | |
e625cce1 TG |
2163 | raw_spin_lock(&ctx->lock); |
2164 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2165 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2166 | /* |
2167 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2168 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2169 | */ |
8dc85d54 PZ |
2170 | task->perf_event_ctxp[ctxn] = next_ctx; |
2171 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2172 | ctx->task = next; |
2173 | next_ctx->task = task; | |
2174 | do_switch = 0; | |
bfbd3381 | 2175 | |
cdd6c482 | 2176 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2177 | } |
e625cce1 TG |
2178 | raw_spin_unlock(&next_ctx->lock); |
2179 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2180 | } |
c93f7669 | 2181 | rcu_read_unlock(); |
564c2b21 | 2182 | |
c93f7669 | 2183 | if (do_switch) { |
facc4307 | 2184 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2185 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2186 | cpuctx->task_ctx = NULL; |
facc4307 | 2187 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2188 | } |
0793a61d TG |
2189 | } |
2190 | ||
8dc85d54 PZ |
2191 | #define for_each_task_context_nr(ctxn) \ |
2192 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2193 | ||
2194 | /* | |
2195 | * Called from scheduler to remove the events of the current task, | |
2196 | * with interrupts disabled. | |
2197 | * | |
2198 | * We stop each event and update the event value in event->count. | |
2199 | * | |
2200 | * This does not protect us against NMI, but disable() | |
2201 | * sets the disabled bit in the control field of event _before_ | |
2202 | * accessing the event control register. If a NMI hits, then it will | |
2203 | * not restart the event. | |
2204 | */ | |
ab0cce56 JO |
2205 | void __perf_event_task_sched_out(struct task_struct *task, |
2206 | struct task_struct *next) | |
8dc85d54 PZ |
2207 | { |
2208 | int ctxn; | |
2209 | ||
8dc85d54 PZ |
2210 | for_each_task_context_nr(ctxn) |
2211 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2212 | |
2213 | /* | |
2214 | * if cgroup events exist on this CPU, then we need | |
2215 | * to check if we have to switch out PMU state. | |
2216 | * cgroup event are system-wide mode only | |
2217 | */ | |
2218 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2219 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2220 | } |
2221 | ||
04dc2dbb | 2222 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2223 | { |
108b02cf | 2224 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2225 | |
a63eaf34 PM |
2226 | if (!cpuctx->task_ctx) |
2227 | return; | |
012b84da IM |
2228 | |
2229 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2230 | return; | |
2231 | ||
04dc2dbb | 2232 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2233 | cpuctx->task_ctx = NULL; |
2234 | } | |
2235 | ||
5b0311e1 FW |
2236 | /* |
2237 | * Called with IRQs disabled | |
2238 | */ | |
2239 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2240 | enum event_type_t event_type) | |
2241 | { | |
2242 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2243 | } |
2244 | ||
235c7fc7 | 2245 | static void |
5b0311e1 | 2246 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2247 | struct perf_cpu_context *cpuctx) |
0793a61d | 2248 | { |
cdd6c482 | 2249 | struct perf_event *event; |
0793a61d | 2250 | |
889ff015 FW |
2251 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2252 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2253 | continue; |
5632ab12 | 2254 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2255 | continue; |
2256 | ||
e5d1367f SE |
2257 | /* may need to reset tstamp_enabled */ |
2258 | if (is_cgroup_event(event)) | |
2259 | perf_cgroup_mark_enabled(event, ctx); | |
2260 | ||
8c9ed8e1 | 2261 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2262 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2263 | |
2264 | /* | |
2265 | * If this pinned group hasn't been scheduled, | |
2266 | * put it in error state. | |
2267 | */ | |
cdd6c482 IM |
2268 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2269 | update_group_times(event); | |
2270 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2271 | } |
3b6f9e5c | 2272 | } |
5b0311e1 FW |
2273 | } |
2274 | ||
2275 | static void | |
2276 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2277 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2278 | { |
2279 | struct perf_event *event; | |
2280 | int can_add_hw = 1; | |
3b6f9e5c | 2281 | |
889ff015 FW |
2282 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2283 | /* Ignore events in OFF or ERROR state */ | |
2284 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2285 | continue; |
04289bb9 IM |
2286 | /* |
2287 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2288 | * of events: |
04289bb9 | 2289 | */ |
5632ab12 | 2290 | if (!event_filter_match(event)) |
0793a61d TG |
2291 | continue; |
2292 | ||
e5d1367f SE |
2293 | /* may need to reset tstamp_enabled */ |
2294 | if (is_cgroup_event(event)) | |
2295 | perf_cgroup_mark_enabled(event, ctx); | |
2296 | ||
9ed6060d | 2297 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2298 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2299 | can_add_hw = 0; |
9ed6060d | 2300 | } |
0793a61d | 2301 | } |
5b0311e1 FW |
2302 | } |
2303 | ||
2304 | static void | |
2305 | ctx_sched_in(struct perf_event_context *ctx, | |
2306 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2307 | enum event_type_t event_type, |
2308 | struct task_struct *task) | |
5b0311e1 | 2309 | { |
e5d1367f | 2310 | u64 now; |
db24d33e | 2311 | int is_active = ctx->is_active; |
e5d1367f | 2312 | |
db24d33e | 2313 | ctx->is_active |= event_type; |
5b0311e1 | 2314 | if (likely(!ctx->nr_events)) |
facc4307 | 2315 | return; |
5b0311e1 | 2316 | |
e5d1367f SE |
2317 | now = perf_clock(); |
2318 | ctx->timestamp = now; | |
3f7cce3c | 2319 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2320 | /* |
2321 | * First go through the list and put on any pinned groups | |
2322 | * in order to give them the best chance of going on. | |
2323 | */ | |
db24d33e | 2324 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2325 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2326 | |
2327 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2328 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2329 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2330 | } |
2331 | ||
329c0e01 | 2332 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2333 | enum event_type_t event_type, |
2334 | struct task_struct *task) | |
329c0e01 FW |
2335 | { |
2336 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2337 | ||
e5d1367f | 2338 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2339 | } |
2340 | ||
e5d1367f SE |
2341 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2342 | struct task_struct *task) | |
235c7fc7 | 2343 | { |
108b02cf | 2344 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2345 | |
108b02cf | 2346 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2347 | if (cpuctx->task_ctx == ctx) |
2348 | return; | |
2349 | ||
facc4307 | 2350 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2351 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2352 | /* |
2353 | * We want to keep the following priority order: | |
2354 | * cpu pinned (that don't need to move), task pinned, | |
2355 | * cpu flexible, task flexible. | |
2356 | */ | |
2357 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2358 | ||
1d5f003f GN |
2359 | if (ctx->nr_events) |
2360 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2361 | |
86b47c25 GN |
2362 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2363 | ||
facc4307 PZ |
2364 | perf_pmu_enable(ctx->pmu); |
2365 | perf_ctx_unlock(cpuctx, ctx); | |
2366 | ||
b5ab4cd5 PZ |
2367 | /* |
2368 | * Since these rotations are per-cpu, we need to ensure the | |
2369 | * cpu-context we got scheduled on is actually rotating. | |
2370 | */ | |
108b02cf | 2371 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2372 | } |
2373 | ||
d010b332 SE |
2374 | /* |
2375 | * When sampling the branck stack in system-wide, it may be necessary | |
2376 | * to flush the stack on context switch. This happens when the branch | |
2377 | * stack does not tag its entries with the pid of the current task. | |
2378 | * Otherwise it becomes impossible to associate a branch entry with a | |
2379 | * task. This ambiguity is more likely to appear when the branch stack | |
2380 | * supports priv level filtering and the user sets it to monitor only | |
2381 | * at the user level (which could be a useful measurement in system-wide | |
2382 | * mode). In that case, the risk is high of having a branch stack with | |
2383 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2384 | * entries or stashing them somewhere in the PMU specific code layer. | |
2385 | * | |
2386 | * This function provides the context switch callback to the lower code | |
2387 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2388 | * with at least one active event using taken branch sampling. | |
2389 | */ | |
2390 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2391 | struct task_struct *task) | |
2392 | { | |
2393 | struct perf_cpu_context *cpuctx; | |
2394 | struct pmu *pmu; | |
2395 | unsigned long flags; | |
2396 | ||
2397 | /* no need to flush branch stack if not changing task */ | |
2398 | if (prev == task) | |
2399 | return; | |
2400 | ||
2401 | local_irq_save(flags); | |
2402 | ||
2403 | rcu_read_lock(); | |
2404 | ||
2405 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2406 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2407 | ||
2408 | /* | |
2409 | * check if the context has at least one | |
2410 | * event using PERF_SAMPLE_BRANCH_STACK | |
2411 | */ | |
2412 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2413 | && pmu->flush_branch_stack) { | |
2414 | ||
2415 | pmu = cpuctx->ctx.pmu; | |
2416 | ||
2417 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2418 | ||
2419 | perf_pmu_disable(pmu); | |
2420 | ||
2421 | pmu->flush_branch_stack(); | |
2422 | ||
2423 | perf_pmu_enable(pmu); | |
2424 | ||
2425 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2426 | } | |
2427 | } | |
2428 | ||
2429 | rcu_read_unlock(); | |
2430 | ||
2431 | local_irq_restore(flags); | |
2432 | } | |
2433 | ||
8dc85d54 PZ |
2434 | /* |
2435 | * Called from scheduler to add the events of the current task | |
2436 | * with interrupts disabled. | |
2437 | * | |
2438 | * We restore the event value and then enable it. | |
2439 | * | |
2440 | * This does not protect us against NMI, but enable() | |
2441 | * sets the enabled bit in the control field of event _before_ | |
2442 | * accessing the event control register. If a NMI hits, then it will | |
2443 | * keep the event running. | |
2444 | */ | |
ab0cce56 JO |
2445 | void __perf_event_task_sched_in(struct task_struct *prev, |
2446 | struct task_struct *task) | |
8dc85d54 PZ |
2447 | { |
2448 | struct perf_event_context *ctx; | |
2449 | int ctxn; | |
2450 | ||
2451 | for_each_task_context_nr(ctxn) { | |
2452 | ctx = task->perf_event_ctxp[ctxn]; | |
2453 | if (likely(!ctx)) | |
2454 | continue; | |
2455 | ||
e5d1367f | 2456 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2457 | } |
e5d1367f SE |
2458 | /* |
2459 | * if cgroup events exist on this CPU, then we need | |
2460 | * to check if we have to switch in PMU state. | |
2461 | * cgroup event are system-wide mode only | |
2462 | */ | |
2463 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2464 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2465 | |
2466 | /* check for system-wide branch_stack events */ | |
2467 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2468 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2469 | } |
2470 | ||
abd50713 PZ |
2471 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2472 | { | |
2473 | u64 frequency = event->attr.sample_freq; | |
2474 | u64 sec = NSEC_PER_SEC; | |
2475 | u64 divisor, dividend; | |
2476 | ||
2477 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2478 | ||
2479 | count_fls = fls64(count); | |
2480 | nsec_fls = fls64(nsec); | |
2481 | frequency_fls = fls64(frequency); | |
2482 | sec_fls = 30; | |
2483 | ||
2484 | /* | |
2485 | * We got @count in @nsec, with a target of sample_freq HZ | |
2486 | * the target period becomes: | |
2487 | * | |
2488 | * @count * 10^9 | |
2489 | * period = ------------------- | |
2490 | * @nsec * sample_freq | |
2491 | * | |
2492 | */ | |
2493 | ||
2494 | /* | |
2495 | * Reduce accuracy by one bit such that @a and @b converge | |
2496 | * to a similar magnitude. | |
2497 | */ | |
fe4b04fa | 2498 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2499 | do { \ |
2500 | if (a##_fls > b##_fls) { \ | |
2501 | a >>= 1; \ | |
2502 | a##_fls--; \ | |
2503 | } else { \ | |
2504 | b >>= 1; \ | |
2505 | b##_fls--; \ | |
2506 | } \ | |
2507 | } while (0) | |
2508 | ||
2509 | /* | |
2510 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2511 | * the other, so that finally we can do a u64/u64 division. | |
2512 | */ | |
2513 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2514 | REDUCE_FLS(nsec, frequency); | |
2515 | REDUCE_FLS(sec, count); | |
2516 | } | |
2517 | ||
2518 | if (count_fls + sec_fls > 64) { | |
2519 | divisor = nsec * frequency; | |
2520 | ||
2521 | while (count_fls + sec_fls > 64) { | |
2522 | REDUCE_FLS(count, sec); | |
2523 | divisor >>= 1; | |
2524 | } | |
2525 | ||
2526 | dividend = count * sec; | |
2527 | } else { | |
2528 | dividend = count * sec; | |
2529 | ||
2530 | while (nsec_fls + frequency_fls > 64) { | |
2531 | REDUCE_FLS(nsec, frequency); | |
2532 | dividend >>= 1; | |
2533 | } | |
2534 | ||
2535 | divisor = nsec * frequency; | |
2536 | } | |
2537 | ||
f6ab91ad PZ |
2538 | if (!divisor) |
2539 | return dividend; | |
2540 | ||
abd50713 PZ |
2541 | return div64_u64(dividend, divisor); |
2542 | } | |
2543 | ||
e050e3f0 SE |
2544 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2545 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2546 | ||
f39d47ff | 2547 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2548 | { |
cdd6c482 | 2549 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2550 | s64 period, sample_period; |
bd2b5b12 PZ |
2551 | s64 delta; |
2552 | ||
abd50713 | 2553 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2554 | |
2555 | delta = (s64)(period - hwc->sample_period); | |
2556 | delta = (delta + 7) / 8; /* low pass filter */ | |
2557 | ||
2558 | sample_period = hwc->sample_period + delta; | |
2559 | ||
2560 | if (!sample_period) | |
2561 | sample_period = 1; | |
2562 | ||
bd2b5b12 | 2563 | hwc->sample_period = sample_period; |
abd50713 | 2564 | |
e7850595 | 2565 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2566 | if (disable) |
2567 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2568 | ||
e7850595 | 2569 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2570 | |
2571 | if (disable) | |
2572 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2573 | } |
bd2b5b12 PZ |
2574 | } |
2575 | ||
e050e3f0 SE |
2576 | /* |
2577 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2578 | * events. At the same time, make sure, having freq events does not change | |
2579 | * the rate of unthrottling as that would introduce bias. | |
2580 | */ | |
2581 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2582 | int needs_unthr) | |
60db5e09 | 2583 | { |
cdd6c482 IM |
2584 | struct perf_event *event; |
2585 | struct hw_perf_event *hwc; | |
e050e3f0 | 2586 | u64 now, period = TICK_NSEC; |
abd50713 | 2587 | s64 delta; |
60db5e09 | 2588 | |
e050e3f0 SE |
2589 | /* |
2590 | * only need to iterate over all events iff: | |
2591 | * - context have events in frequency mode (needs freq adjust) | |
2592 | * - there are events to unthrottle on this cpu | |
2593 | */ | |
2594 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2595 | return; |
2596 | ||
e050e3f0 | 2597 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2598 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2599 | |
03541f8b | 2600 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2601 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2602 | continue; |
2603 | ||
5632ab12 | 2604 | if (!event_filter_match(event)) |
5d27c23d PZ |
2605 | continue; |
2606 | ||
cdd6c482 | 2607 | hwc = &event->hw; |
6a24ed6c | 2608 | |
e050e3f0 SE |
2609 | if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) { |
2610 | hwc->interrupts = 0; | |
cdd6c482 | 2611 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2612 | event->pmu->start(event, 0); |
a78ac325 PZ |
2613 | } |
2614 | ||
cdd6c482 | 2615 | if (!event->attr.freq || !event->attr.sample_freq) |
60db5e09 PZ |
2616 | continue; |
2617 | ||
e050e3f0 SE |
2618 | /* |
2619 | * stop the event and update event->count | |
2620 | */ | |
2621 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2622 | ||
e7850595 | 2623 | now = local64_read(&event->count); |
abd50713 PZ |
2624 | delta = now - hwc->freq_count_stamp; |
2625 | hwc->freq_count_stamp = now; | |
60db5e09 | 2626 | |
e050e3f0 SE |
2627 | /* |
2628 | * restart the event | |
2629 | * reload only if value has changed | |
f39d47ff SE |
2630 | * we have stopped the event so tell that |
2631 | * to perf_adjust_period() to avoid stopping it | |
2632 | * twice. | |
e050e3f0 | 2633 | */ |
abd50713 | 2634 | if (delta > 0) |
f39d47ff | 2635 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2636 | |
2637 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
60db5e09 | 2638 | } |
e050e3f0 | 2639 | |
f39d47ff | 2640 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2641 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2642 | } |
2643 | ||
235c7fc7 | 2644 | /* |
cdd6c482 | 2645 | * Round-robin a context's events: |
235c7fc7 | 2646 | */ |
cdd6c482 | 2647 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2648 | { |
dddd3379 TG |
2649 | /* |
2650 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2651 | * disabled by the inheritance code. | |
2652 | */ | |
2653 | if (!ctx->rotate_disable) | |
2654 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2655 | } |
2656 | ||
b5ab4cd5 | 2657 | /* |
e9d2b064 PZ |
2658 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2659 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2660 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2661 | */ |
9e630205 | 2662 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2663 | { |
8dc85d54 | 2664 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2665 | int rotate = 0, remove = 1; |
7fc23a53 | 2666 | |
b5ab4cd5 | 2667 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2668 | remove = 0; |
b5ab4cd5 PZ |
2669 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2670 | rotate = 1; | |
2671 | } | |
235c7fc7 | 2672 | |
8dc85d54 | 2673 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2674 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2675 | remove = 0; |
b5ab4cd5 PZ |
2676 | if (ctx->nr_events != ctx->nr_active) |
2677 | rotate = 1; | |
2678 | } | |
9717e6cd | 2679 | |
e050e3f0 | 2680 | if (!rotate) |
0f5a2601 PZ |
2681 | goto done; |
2682 | ||
facc4307 | 2683 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2684 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2685 | |
e050e3f0 SE |
2686 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2687 | if (ctx) | |
2688 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2689 | |
e050e3f0 SE |
2690 | rotate_ctx(&cpuctx->ctx); |
2691 | if (ctx) | |
2692 | rotate_ctx(ctx); | |
235c7fc7 | 2693 | |
e050e3f0 | 2694 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2695 | |
0f5a2601 PZ |
2696 | perf_pmu_enable(cpuctx->ctx.pmu); |
2697 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2698 | done: |
e9d2b064 PZ |
2699 | if (remove) |
2700 | list_del_init(&cpuctx->rotation_list); | |
9e630205 SE |
2701 | |
2702 | return rotate; | |
e9d2b064 PZ |
2703 | } |
2704 | ||
026249ef FW |
2705 | #ifdef CONFIG_NO_HZ_FULL |
2706 | bool perf_event_can_stop_tick(void) | |
2707 | { | |
2708 | if (list_empty(&__get_cpu_var(rotation_list))) | |
2709 | return true; | |
2710 | else | |
2711 | return false; | |
2712 | } | |
2713 | #endif | |
2714 | ||
e9d2b064 PZ |
2715 | void perf_event_task_tick(void) |
2716 | { | |
2717 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2718 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2719 | struct perf_event_context *ctx; |
2720 | int throttled; | |
b5ab4cd5 | 2721 | |
e9d2b064 PZ |
2722 | WARN_ON(!irqs_disabled()); |
2723 | ||
e050e3f0 SE |
2724 | __this_cpu_inc(perf_throttled_seq); |
2725 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2726 | ||
e9d2b064 | 2727 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2728 | ctx = &cpuctx->ctx; |
2729 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2730 | ||
2731 | ctx = cpuctx->task_ctx; | |
2732 | if (ctx) | |
2733 | perf_adjust_freq_unthr_context(ctx, throttled); | |
e9d2b064 | 2734 | } |
0793a61d TG |
2735 | } |
2736 | ||
889ff015 FW |
2737 | static int event_enable_on_exec(struct perf_event *event, |
2738 | struct perf_event_context *ctx) | |
2739 | { | |
2740 | if (!event->attr.enable_on_exec) | |
2741 | return 0; | |
2742 | ||
2743 | event->attr.enable_on_exec = 0; | |
2744 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2745 | return 0; | |
2746 | ||
1d9b482e | 2747 | __perf_event_mark_enabled(event); |
889ff015 FW |
2748 | |
2749 | return 1; | |
2750 | } | |
2751 | ||
57e7986e | 2752 | /* |
cdd6c482 | 2753 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2754 | * This expects task == current. |
2755 | */ | |
8dc85d54 | 2756 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2757 | { |
cdd6c482 | 2758 | struct perf_event *event; |
57e7986e PM |
2759 | unsigned long flags; |
2760 | int enabled = 0; | |
889ff015 | 2761 | int ret; |
57e7986e PM |
2762 | |
2763 | local_irq_save(flags); | |
cdd6c482 | 2764 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2765 | goto out; |
2766 | ||
e566b76e SE |
2767 | /* |
2768 | * We must ctxsw out cgroup events to avoid conflict | |
2769 | * when invoking perf_task_event_sched_in() later on | |
2770 | * in this function. Otherwise we end up trying to | |
2771 | * ctxswin cgroup events which are already scheduled | |
2772 | * in. | |
2773 | */ | |
a8d757ef | 2774 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2775 | |
e625cce1 | 2776 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2777 | task_ctx_sched_out(ctx); |
57e7986e | 2778 | |
b79387ef | 2779 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2780 | ret = event_enable_on_exec(event, ctx); |
2781 | if (ret) | |
2782 | enabled = 1; | |
57e7986e PM |
2783 | } |
2784 | ||
2785 | /* | |
cdd6c482 | 2786 | * Unclone this context if we enabled any event. |
57e7986e | 2787 | */ |
71a851b4 PZ |
2788 | if (enabled) |
2789 | unclone_ctx(ctx); | |
57e7986e | 2790 | |
e625cce1 | 2791 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2792 | |
e566b76e SE |
2793 | /* |
2794 | * Also calls ctxswin for cgroup events, if any: | |
2795 | */ | |
e5d1367f | 2796 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2797 | out: |
57e7986e PM |
2798 | local_irq_restore(flags); |
2799 | } | |
2800 | ||
0793a61d | 2801 | /* |
cdd6c482 | 2802 | * Cross CPU call to read the hardware event |
0793a61d | 2803 | */ |
cdd6c482 | 2804 | static void __perf_event_read(void *info) |
0793a61d | 2805 | { |
cdd6c482 IM |
2806 | struct perf_event *event = info; |
2807 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2808 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2809 | |
e1ac3614 PM |
2810 | /* |
2811 | * If this is a task context, we need to check whether it is | |
2812 | * the current task context of this cpu. If not it has been | |
2813 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2814 | * event->count would have been updated to a recent sample |
2815 | * when the event was scheduled out. | |
e1ac3614 PM |
2816 | */ |
2817 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2818 | return; | |
2819 | ||
e625cce1 | 2820 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2821 | if (ctx->is_active) { |
542e72fc | 2822 | update_context_time(ctx); |
e5d1367f SE |
2823 | update_cgrp_time_from_event(event); |
2824 | } | |
cdd6c482 | 2825 | update_event_times(event); |
542e72fc PZ |
2826 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2827 | event->pmu->read(event); | |
e625cce1 | 2828 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2829 | } |
2830 | ||
b5e58793 PZ |
2831 | static inline u64 perf_event_count(struct perf_event *event) |
2832 | { | |
e7850595 | 2833 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
2834 | } |
2835 | ||
cdd6c482 | 2836 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
2837 | { |
2838 | /* | |
cdd6c482 IM |
2839 | * If event is enabled and currently active on a CPU, update the |
2840 | * value in the event structure: | |
0793a61d | 2841 | */ |
cdd6c482 IM |
2842 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2843 | smp_call_function_single(event->oncpu, | |
2844 | __perf_event_read, event, 1); | |
2845 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
2846 | struct perf_event_context *ctx = event->ctx; |
2847 | unsigned long flags; | |
2848 | ||
e625cce1 | 2849 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
2850 | /* |
2851 | * may read while context is not active | |
2852 | * (e.g., thread is blocked), in that case | |
2853 | * we cannot update context time | |
2854 | */ | |
e5d1367f | 2855 | if (ctx->is_active) { |
c530ccd9 | 2856 | update_context_time(ctx); |
e5d1367f SE |
2857 | update_cgrp_time_from_event(event); |
2858 | } | |
cdd6c482 | 2859 | update_event_times(event); |
e625cce1 | 2860 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
2861 | } |
2862 | ||
b5e58793 | 2863 | return perf_event_count(event); |
0793a61d TG |
2864 | } |
2865 | ||
a63eaf34 | 2866 | /* |
cdd6c482 | 2867 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 2868 | */ |
eb184479 | 2869 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 2870 | { |
e625cce1 | 2871 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 2872 | mutex_init(&ctx->mutex); |
889ff015 FW |
2873 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2874 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
2875 | INIT_LIST_HEAD(&ctx->event_list); |
2876 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
2877 | } |
2878 | ||
2879 | static struct perf_event_context * | |
2880 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
2881 | { | |
2882 | struct perf_event_context *ctx; | |
2883 | ||
2884 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
2885 | if (!ctx) | |
2886 | return NULL; | |
2887 | ||
2888 | __perf_event_init_context(ctx); | |
2889 | if (task) { | |
2890 | ctx->task = task; | |
2891 | get_task_struct(task); | |
0793a61d | 2892 | } |
eb184479 PZ |
2893 | ctx->pmu = pmu; |
2894 | ||
2895 | return ctx; | |
a63eaf34 PM |
2896 | } |
2897 | ||
2ebd4ffb MH |
2898 | static struct task_struct * |
2899 | find_lively_task_by_vpid(pid_t vpid) | |
2900 | { | |
2901 | struct task_struct *task; | |
2902 | int err; | |
0793a61d TG |
2903 | |
2904 | rcu_read_lock(); | |
2ebd4ffb | 2905 | if (!vpid) |
0793a61d TG |
2906 | task = current; |
2907 | else | |
2ebd4ffb | 2908 | task = find_task_by_vpid(vpid); |
0793a61d TG |
2909 | if (task) |
2910 | get_task_struct(task); | |
2911 | rcu_read_unlock(); | |
2912 | ||
2913 | if (!task) | |
2914 | return ERR_PTR(-ESRCH); | |
2915 | ||
0793a61d | 2916 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
2917 | err = -EACCES; |
2918 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
2919 | goto errout; | |
2920 | ||
2ebd4ffb MH |
2921 | return task; |
2922 | errout: | |
2923 | put_task_struct(task); | |
2924 | return ERR_PTR(err); | |
2925 | ||
2926 | } | |
2927 | ||
fe4b04fa PZ |
2928 | /* |
2929 | * Returns a matching context with refcount and pincount. | |
2930 | */ | |
108b02cf | 2931 | static struct perf_event_context * |
38a81da2 | 2932 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 2933 | { |
cdd6c482 | 2934 | struct perf_event_context *ctx; |
22a4f650 | 2935 | struct perf_cpu_context *cpuctx; |
25346b93 | 2936 | unsigned long flags; |
8dc85d54 | 2937 | int ctxn, err; |
0793a61d | 2938 | |
22a4ec72 | 2939 | if (!task) { |
cdd6c482 | 2940 | /* Must be root to operate on a CPU event: */ |
0764771d | 2941 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
2942 | return ERR_PTR(-EACCES); |
2943 | ||
0793a61d | 2944 | /* |
cdd6c482 | 2945 | * We could be clever and allow to attach a event to an |
0793a61d TG |
2946 | * offline CPU and activate it when the CPU comes up, but |
2947 | * that's for later. | |
2948 | */ | |
f6325e30 | 2949 | if (!cpu_online(cpu)) |
0793a61d TG |
2950 | return ERR_PTR(-ENODEV); |
2951 | ||
108b02cf | 2952 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 2953 | ctx = &cpuctx->ctx; |
c93f7669 | 2954 | get_ctx(ctx); |
fe4b04fa | 2955 | ++ctx->pin_count; |
0793a61d | 2956 | |
0793a61d TG |
2957 | return ctx; |
2958 | } | |
2959 | ||
8dc85d54 PZ |
2960 | err = -EINVAL; |
2961 | ctxn = pmu->task_ctx_nr; | |
2962 | if (ctxn < 0) | |
2963 | goto errout; | |
2964 | ||
9ed6060d | 2965 | retry: |
8dc85d54 | 2966 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 2967 | if (ctx) { |
71a851b4 | 2968 | unclone_ctx(ctx); |
fe4b04fa | 2969 | ++ctx->pin_count; |
e625cce1 | 2970 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 2971 | } else { |
eb184479 | 2972 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
2973 | err = -ENOMEM; |
2974 | if (!ctx) | |
2975 | goto errout; | |
eb184479 | 2976 | |
dbe08d82 ON |
2977 | err = 0; |
2978 | mutex_lock(&task->perf_event_mutex); | |
2979 | /* | |
2980 | * If it has already passed perf_event_exit_task(). | |
2981 | * we must see PF_EXITING, it takes this mutex too. | |
2982 | */ | |
2983 | if (task->flags & PF_EXITING) | |
2984 | err = -ESRCH; | |
2985 | else if (task->perf_event_ctxp[ctxn]) | |
2986 | err = -EAGAIN; | |
fe4b04fa | 2987 | else { |
9137fb28 | 2988 | get_ctx(ctx); |
fe4b04fa | 2989 | ++ctx->pin_count; |
dbe08d82 | 2990 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 2991 | } |
dbe08d82 ON |
2992 | mutex_unlock(&task->perf_event_mutex); |
2993 | ||
2994 | if (unlikely(err)) { | |
9137fb28 | 2995 | put_ctx(ctx); |
dbe08d82 ON |
2996 | |
2997 | if (err == -EAGAIN) | |
2998 | goto retry; | |
2999 | goto errout; | |
a63eaf34 PM |
3000 | } |
3001 | } | |
3002 | ||
0793a61d | 3003 | return ctx; |
c93f7669 | 3004 | |
9ed6060d | 3005 | errout: |
c93f7669 | 3006 | return ERR_PTR(err); |
0793a61d TG |
3007 | } |
3008 | ||
6fb2915d LZ |
3009 | static void perf_event_free_filter(struct perf_event *event); |
3010 | ||
cdd6c482 | 3011 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3012 | { |
cdd6c482 | 3013 | struct perf_event *event; |
592903cd | 3014 | |
cdd6c482 IM |
3015 | event = container_of(head, struct perf_event, rcu_head); |
3016 | if (event->ns) | |
3017 | put_pid_ns(event->ns); | |
6fb2915d | 3018 | perf_event_free_filter(event); |
cdd6c482 | 3019 | kfree(event); |
592903cd PZ |
3020 | } |
3021 | ||
76369139 | 3022 | static void ring_buffer_put(struct ring_buffer *rb); |
9bb5d40c | 3023 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb); |
925d519a | 3024 | |
cdd6c482 | 3025 | static void free_event(struct perf_event *event) |
f1600952 | 3026 | { |
e360adbe | 3027 | irq_work_sync(&event->pending); |
925d519a | 3028 | |
cdd6c482 | 3029 | if (!event->parent) { |
82cd6def | 3030 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 3031 | static_key_slow_dec_deferred(&perf_sched_events); |
3af9e859 | 3032 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
3033 | atomic_dec(&nr_mmap_events); |
3034 | if (event->attr.comm) | |
3035 | atomic_dec(&nr_comm_events); | |
3036 | if (event->attr.task) | |
3037 | atomic_dec(&nr_task_events); | |
927c7a9e FW |
3038 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3039 | put_callchain_buffers(); | |
08309379 PZ |
3040 | if (is_cgroup_event(event)) { |
3041 | atomic_dec(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 3042 | static_key_slow_dec_deferred(&perf_sched_events); |
08309379 | 3043 | } |
d010b332 SE |
3044 | |
3045 | if (has_branch_stack(event)) { | |
3046 | static_key_slow_dec_deferred(&perf_sched_events); | |
3047 | /* is system-wide event */ | |
9bb5d40c | 3048 | if (!(event->attach_state & PERF_ATTACH_TASK)) { |
d010b332 SE |
3049 | atomic_dec(&per_cpu(perf_branch_stack_events, |
3050 | event->cpu)); | |
9bb5d40c | 3051 | } |
d010b332 | 3052 | } |
f344011c | 3053 | } |
9ee318a7 | 3054 | |
76369139 | 3055 | if (event->rb) { |
9bb5d40c PZ |
3056 | struct ring_buffer *rb; |
3057 | ||
3058 | /* | |
3059 | * Can happen when we close an event with re-directed output. | |
3060 | * | |
3061 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3062 | * over us; possibly making our ring_buffer_put() the last. | |
3063 | */ | |
3064 | mutex_lock(&event->mmap_mutex); | |
3065 | rb = event->rb; | |
3066 | if (rb) { | |
3067 | rcu_assign_pointer(event->rb, NULL); | |
3068 | ring_buffer_detach(event, rb); | |
3069 | ring_buffer_put(rb); /* could be last */ | |
3070 | } | |
3071 | mutex_unlock(&event->mmap_mutex); | |
a4be7c27 PZ |
3072 | } |
3073 | ||
e5d1367f SE |
3074 | if (is_cgroup_event(event)) |
3075 | perf_detach_cgroup(event); | |
3076 | ||
cdd6c482 IM |
3077 | if (event->destroy) |
3078 | event->destroy(event); | |
e077df4f | 3079 | |
0c67b408 PZ |
3080 | if (event->ctx) |
3081 | put_ctx(event->ctx); | |
3082 | ||
cdd6c482 | 3083 | call_rcu(&event->rcu_head, free_event_rcu); |
f1600952 PZ |
3084 | } |
3085 | ||
a66a3052 | 3086 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 3087 | { |
cdd6c482 | 3088 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 3089 | |
ad3a37de | 3090 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
3091 | /* |
3092 | * There are two ways this annotation is useful: | |
3093 | * | |
3094 | * 1) there is a lock recursion from perf_event_exit_task | |
3095 | * see the comment there. | |
3096 | * | |
3097 | * 2) there is a lock-inversion with mmap_sem through | |
3098 | * perf_event_read_group(), which takes faults while | |
3099 | * holding ctx->mutex, however this is called after | |
3100 | * the last filedesc died, so there is no possibility | |
3101 | * to trigger the AB-BA case. | |
3102 | */ | |
3103 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
050735b0 | 3104 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 3105 | perf_group_detach(event); |
050735b0 | 3106 | raw_spin_unlock_irq(&ctx->lock); |
e03a9a55 | 3107 | perf_remove_from_context(event); |
d859e29f | 3108 | mutex_unlock(&ctx->mutex); |
0793a61d | 3109 | |
cdd6c482 | 3110 | free_event(event); |
0793a61d TG |
3111 | |
3112 | return 0; | |
3113 | } | |
a66a3052 | 3114 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 3115 | |
a66a3052 PZ |
3116 | /* |
3117 | * Called when the last reference to the file is gone. | |
3118 | */ | |
a6fa941d | 3119 | static void put_event(struct perf_event *event) |
fb0459d7 | 3120 | { |
8882135b | 3121 | struct task_struct *owner; |
fb0459d7 | 3122 | |
a6fa941d AV |
3123 | if (!atomic_long_dec_and_test(&event->refcount)) |
3124 | return; | |
fb0459d7 | 3125 | |
8882135b PZ |
3126 | rcu_read_lock(); |
3127 | owner = ACCESS_ONCE(event->owner); | |
3128 | /* | |
3129 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3130 | * !owner it means the list deletion is complete and we can indeed | |
3131 | * free this event, otherwise we need to serialize on | |
3132 | * owner->perf_event_mutex. | |
3133 | */ | |
3134 | smp_read_barrier_depends(); | |
3135 | if (owner) { | |
3136 | /* | |
3137 | * Since delayed_put_task_struct() also drops the last | |
3138 | * task reference we can safely take a new reference | |
3139 | * while holding the rcu_read_lock(). | |
3140 | */ | |
3141 | get_task_struct(owner); | |
3142 | } | |
3143 | rcu_read_unlock(); | |
3144 | ||
3145 | if (owner) { | |
3146 | mutex_lock(&owner->perf_event_mutex); | |
3147 | /* | |
3148 | * We have to re-check the event->owner field, if it is cleared | |
3149 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3150 | * ensured they're done, and we can proceed with freeing the | |
3151 | * event. | |
3152 | */ | |
3153 | if (event->owner) | |
3154 | list_del_init(&event->owner_entry); | |
3155 | mutex_unlock(&owner->perf_event_mutex); | |
3156 | put_task_struct(owner); | |
3157 | } | |
3158 | ||
a6fa941d AV |
3159 | perf_event_release_kernel(event); |
3160 | } | |
3161 | ||
3162 | static int perf_release(struct inode *inode, struct file *file) | |
3163 | { | |
3164 | put_event(file->private_data); | |
3165 | return 0; | |
fb0459d7 | 3166 | } |
fb0459d7 | 3167 | |
59ed446f | 3168 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3169 | { |
cdd6c482 | 3170 | struct perf_event *child; |
e53c0994 PZ |
3171 | u64 total = 0; |
3172 | ||
59ed446f PZ |
3173 | *enabled = 0; |
3174 | *running = 0; | |
3175 | ||
6f10581a | 3176 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3177 | total += perf_event_read(event); |
59ed446f PZ |
3178 | *enabled += event->total_time_enabled + |
3179 | atomic64_read(&event->child_total_time_enabled); | |
3180 | *running += event->total_time_running + | |
3181 | atomic64_read(&event->child_total_time_running); | |
3182 | ||
3183 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3184 | total += perf_event_read(child); |
59ed446f PZ |
3185 | *enabled += child->total_time_enabled; |
3186 | *running += child->total_time_running; | |
3187 | } | |
6f10581a | 3188 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3189 | |
3190 | return total; | |
3191 | } | |
fb0459d7 | 3192 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3193 | |
cdd6c482 | 3194 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3195 | u64 read_format, char __user *buf) |
3196 | { | |
cdd6c482 | 3197 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3198 | int n = 0, size = 0, ret = -EFAULT; |
3199 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3200 | u64 values[5]; |
59ed446f | 3201 | u64 count, enabled, running; |
abf4868b | 3202 | |
6f10581a | 3203 | mutex_lock(&ctx->mutex); |
59ed446f | 3204 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3205 | |
3206 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3207 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3208 | values[n++] = enabled; | |
3209 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3210 | values[n++] = running; | |
abf4868b PZ |
3211 | values[n++] = count; |
3212 | if (read_format & PERF_FORMAT_ID) | |
3213 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3214 | |
3215 | size = n * sizeof(u64); | |
3216 | ||
3217 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3218 | goto unlock; |
3dab77fb | 3219 | |
6f10581a | 3220 | ret = size; |
3dab77fb | 3221 | |
65abc865 | 3222 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3223 | n = 0; |
3dab77fb | 3224 | |
59ed446f | 3225 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3226 | if (read_format & PERF_FORMAT_ID) |
3227 | values[n++] = primary_event_id(sub); | |
3228 | ||
3229 | size = n * sizeof(u64); | |
3230 | ||
184d3da8 | 3231 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3232 | ret = -EFAULT; |
3233 | goto unlock; | |
3234 | } | |
abf4868b PZ |
3235 | |
3236 | ret += size; | |
3dab77fb | 3237 | } |
6f10581a PZ |
3238 | unlock: |
3239 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3240 | |
abf4868b | 3241 | return ret; |
3dab77fb PZ |
3242 | } |
3243 | ||
cdd6c482 | 3244 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3245 | u64 read_format, char __user *buf) |
3246 | { | |
59ed446f | 3247 | u64 enabled, running; |
3dab77fb PZ |
3248 | u64 values[4]; |
3249 | int n = 0; | |
3250 | ||
59ed446f PZ |
3251 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3252 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3253 | values[n++] = enabled; | |
3254 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3255 | values[n++] = running; | |
3dab77fb | 3256 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3257 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3258 | |
3259 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3260 | return -EFAULT; | |
3261 | ||
3262 | return n * sizeof(u64); | |
3263 | } | |
3264 | ||
0793a61d | 3265 | /* |
cdd6c482 | 3266 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3267 | */ |
3268 | static ssize_t | |
cdd6c482 | 3269 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3270 | { |
cdd6c482 | 3271 | u64 read_format = event->attr.read_format; |
3dab77fb | 3272 | int ret; |
0793a61d | 3273 | |
3b6f9e5c | 3274 | /* |
cdd6c482 | 3275 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3276 | * error state (i.e. because it was pinned but it couldn't be |
3277 | * scheduled on to the CPU at some point). | |
3278 | */ | |
cdd6c482 | 3279 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3280 | return 0; |
3281 | ||
c320c7b7 | 3282 | if (count < event->read_size) |
3dab77fb PZ |
3283 | return -ENOSPC; |
3284 | ||
cdd6c482 | 3285 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3286 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3287 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3288 | else |
cdd6c482 | 3289 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3290 | |
3dab77fb | 3291 | return ret; |
0793a61d TG |
3292 | } |
3293 | ||
0793a61d TG |
3294 | static ssize_t |
3295 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3296 | { | |
cdd6c482 | 3297 | struct perf_event *event = file->private_data; |
0793a61d | 3298 | |
cdd6c482 | 3299 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3300 | } |
3301 | ||
3302 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3303 | { | |
cdd6c482 | 3304 | struct perf_event *event = file->private_data; |
76369139 | 3305 | struct ring_buffer *rb; |
c33a0bc4 | 3306 | unsigned int events = POLL_HUP; |
c7138f37 | 3307 | |
10c6db11 | 3308 | /* |
9bb5d40c PZ |
3309 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3310 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3311 | */ |
3312 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3313 | rb = event->rb; |
3314 | if (rb) | |
76369139 | 3315 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 PZ |
3316 | mutex_unlock(&event->mmap_mutex); |
3317 | ||
cdd6c482 | 3318 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3319 | |
0793a61d TG |
3320 | return events; |
3321 | } | |
3322 | ||
cdd6c482 | 3323 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3324 | { |
cdd6c482 | 3325 | (void)perf_event_read(event); |
e7850595 | 3326 | local64_set(&event->count, 0); |
cdd6c482 | 3327 | perf_event_update_userpage(event); |
3df5edad PZ |
3328 | } |
3329 | ||
c93f7669 | 3330 | /* |
cdd6c482 IM |
3331 | * Holding the top-level event's child_mutex means that any |
3332 | * descendant process that has inherited this event will block | |
3333 | * in sync_child_event if it goes to exit, thus satisfying the | |
3334 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3335 | */ |
cdd6c482 IM |
3336 | static void perf_event_for_each_child(struct perf_event *event, |
3337 | void (*func)(struct perf_event *)) | |
3df5edad | 3338 | { |
cdd6c482 | 3339 | struct perf_event *child; |
3df5edad | 3340 | |
cdd6c482 IM |
3341 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3342 | mutex_lock(&event->child_mutex); | |
3343 | func(event); | |
3344 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3345 | func(child); |
cdd6c482 | 3346 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3347 | } |
3348 | ||
cdd6c482 IM |
3349 | static void perf_event_for_each(struct perf_event *event, |
3350 | void (*func)(struct perf_event *)) | |
3df5edad | 3351 | { |
cdd6c482 IM |
3352 | struct perf_event_context *ctx = event->ctx; |
3353 | struct perf_event *sibling; | |
3df5edad | 3354 | |
75f937f2 PZ |
3355 | WARN_ON_ONCE(ctx->parent_ctx); |
3356 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3357 | event = event->group_leader; |
75f937f2 | 3358 | |
cdd6c482 | 3359 | perf_event_for_each_child(event, func); |
cdd6c482 | 3360 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3361 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3362 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3363 | } |
3364 | ||
cdd6c482 | 3365 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3366 | { |
cdd6c482 | 3367 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3368 | int ret = 0; |
3369 | u64 value; | |
3370 | ||
6c7e550f | 3371 | if (!is_sampling_event(event)) |
08247e31 PZ |
3372 | return -EINVAL; |
3373 | ||
ad0cf347 | 3374 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3375 | return -EFAULT; |
3376 | ||
3377 | if (!value) | |
3378 | return -EINVAL; | |
3379 | ||
e625cce1 | 3380 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3381 | if (event->attr.freq) { |
3382 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3383 | ret = -EINVAL; |
3384 | goto unlock; | |
3385 | } | |
3386 | ||
cdd6c482 | 3387 | event->attr.sample_freq = value; |
08247e31 | 3388 | } else { |
cdd6c482 IM |
3389 | event->attr.sample_period = value; |
3390 | event->hw.sample_period = value; | |
08247e31 PZ |
3391 | } |
3392 | unlock: | |
e625cce1 | 3393 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3394 | |
3395 | return ret; | |
3396 | } | |
3397 | ||
ac9721f3 PZ |
3398 | static const struct file_operations perf_fops; |
3399 | ||
2903ff01 | 3400 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3401 | { |
2903ff01 AV |
3402 | struct fd f = fdget(fd); |
3403 | if (!f.file) | |
3404 | return -EBADF; | |
ac9721f3 | 3405 | |
2903ff01 AV |
3406 | if (f.file->f_op != &perf_fops) { |
3407 | fdput(f); | |
3408 | return -EBADF; | |
ac9721f3 | 3409 | } |
2903ff01 AV |
3410 | *p = f; |
3411 | return 0; | |
ac9721f3 PZ |
3412 | } |
3413 | ||
3414 | static int perf_event_set_output(struct perf_event *event, | |
3415 | struct perf_event *output_event); | |
6fb2915d | 3416 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3417 | |
d859e29f PM |
3418 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3419 | { | |
cdd6c482 IM |
3420 | struct perf_event *event = file->private_data; |
3421 | void (*func)(struct perf_event *); | |
3df5edad | 3422 | u32 flags = arg; |
d859e29f PM |
3423 | |
3424 | switch (cmd) { | |
cdd6c482 IM |
3425 | case PERF_EVENT_IOC_ENABLE: |
3426 | func = perf_event_enable; | |
d859e29f | 3427 | break; |
cdd6c482 IM |
3428 | case PERF_EVENT_IOC_DISABLE: |
3429 | func = perf_event_disable; | |
79f14641 | 3430 | break; |
cdd6c482 IM |
3431 | case PERF_EVENT_IOC_RESET: |
3432 | func = perf_event_reset; | |
6de6a7b9 | 3433 | break; |
3df5edad | 3434 | |
cdd6c482 IM |
3435 | case PERF_EVENT_IOC_REFRESH: |
3436 | return perf_event_refresh(event, arg); | |
08247e31 | 3437 | |
cdd6c482 IM |
3438 | case PERF_EVENT_IOC_PERIOD: |
3439 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3440 | |
cdd6c482 | 3441 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3442 | { |
ac9721f3 | 3443 | int ret; |
ac9721f3 | 3444 | if (arg != -1) { |
2903ff01 AV |
3445 | struct perf_event *output_event; |
3446 | struct fd output; | |
3447 | ret = perf_fget_light(arg, &output); | |
3448 | if (ret) | |
3449 | return ret; | |
3450 | output_event = output.file->private_data; | |
3451 | ret = perf_event_set_output(event, output_event); | |
3452 | fdput(output); | |
3453 | } else { | |
3454 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3455 | } |
ac9721f3 PZ |
3456 | return ret; |
3457 | } | |
a4be7c27 | 3458 | |
6fb2915d LZ |
3459 | case PERF_EVENT_IOC_SET_FILTER: |
3460 | return perf_event_set_filter(event, (void __user *)arg); | |
3461 | ||
d859e29f | 3462 | default: |
3df5edad | 3463 | return -ENOTTY; |
d859e29f | 3464 | } |
3df5edad PZ |
3465 | |
3466 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3467 | perf_event_for_each(event, func); |
3df5edad | 3468 | else |
cdd6c482 | 3469 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3470 | |
3471 | return 0; | |
d859e29f PM |
3472 | } |
3473 | ||
cdd6c482 | 3474 | int perf_event_task_enable(void) |
771d7cde | 3475 | { |
cdd6c482 | 3476 | struct perf_event *event; |
771d7cde | 3477 | |
cdd6c482 IM |
3478 | mutex_lock(¤t->perf_event_mutex); |
3479 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3480 | perf_event_for_each_child(event, perf_event_enable); | |
3481 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3482 | |
3483 | return 0; | |
3484 | } | |
3485 | ||
cdd6c482 | 3486 | int perf_event_task_disable(void) |
771d7cde | 3487 | { |
cdd6c482 | 3488 | struct perf_event *event; |
771d7cde | 3489 | |
cdd6c482 IM |
3490 | mutex_lock(¤t->perf_event_mutex); |
3491 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3492 | perf_event_for_each_child(event, perf_event_disable); | |
3493 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3494 | |
3495 | return 0; | |
3496 | } | |
3497 | ||
cdd6c482 | 3498 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3499 | { |
a4eaf7f1 PZ |
3500 | if (event->hw.state & PERF_HES_STOPPED) |
3501 | return 0; | |
3502 | ||
cdd6c482 | 3503 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3504 | return 0; |
3505 | ||
35edc2a5 | 3506 | return event->pmu->event_idx(event); |
194002b2 PZ |
3507 | } |
3508 | ||
c4794295 | 3509 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3510 | u64 *now, |
7f310a5d EM |
3511 | u64 *enabled, |
3512 | u64 *running) | |
c4794295 | 3513 | { |
e3f3541c | 3514 | u64 ctx_time; |
c4794295 | 3515 | |
e3f3541c PZ |
3516 | *now = perf_clock(); |
3517 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3518 | *enabled = ctx_time - event->tstamp_enabled; |
3519 | *running = ctx_time - event->tstamp_running; | |
3520 | } | |
3521 | ||
c7206205 | 3522 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3523 | { |
3524 | } | |
3525 | ||
38ff667b PZ |
3526 | /* |
3527 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3528 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3529 | * code calls this from NMI context. | |
3530 | */ | |
cdd6c482 | 3531 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3532 | { |
cdd6c482 | 3533 | struct perf_event_mmap_page *userpg; |
76369139 | 3534 | struct ring_buffer *rb; |
e3f3541c | 3535 | u64 enabled, running, now; |
38ff667b PZ |
3536 | |
3537 | rcu_read_lock(); | |
0d641208 EM |
3538 | /* |
3539 | * compute total_time_enabled, total_time_running | |
3540 | * based on snapshot values taken when the event | |
3541 | * was last scheduled in. | |
3542 | * | |
3543 | * we cannot simply called update_context_time() | |
3544 | * because of locking issue as we can be called in | |
3545 | * NMI context | |
3546 | */ | |
e3f3541c | 3547 | calc_timer_values(event, &now, &enabled, &running); |
76369139 FW |
3548 | rb = rcu_dereference(event->rb); |
3549 | if (!rb) | |
38ff667b PZ |
3550 | goto unlock; |
3551 | ||
76369139 | 3552 | userpg = rb->user_page; |
37d81828 | 3553 | |
7b732a75 PZ |
3554 | /* |
3555 | * Disable preemption so as to not let the corresponding user-space | |
3556 | * spin too long if we get preempted. | |
3557 | */ | |
3558 | preempt_disable(); | |
37d81828 | 3559 | ++userpg->lock; |
92f22a38 | 3560 | barrier(); |
cdd6c482 | 3561 | userpg->index = perf_event_index(event); |
b5e58793 | 3562 | userpg->offset = perf_event_count(event); |
365a4038 | 3563 | if (userpg->index) |
e7850595 | 3564 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3565 | |
0d641208 | 3566 | userpg->time_enabled = enabled + |
cdd6c482 | 3567 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3568 | |
0d641208 | 3569 | userpg->time_running = running + |
cdd6c482 | 3570 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3571 | |
c7206205 | 3572 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3573 | |
92f22a38 | 3574 | barrier(); |
37d81828 | 3575 | ++userpg->lock; |
7b732a75 | 3576 | preempt_enable(); |
38ff667b | 3577 | unlock: |
7b732a75 | 3578 | rcu_read_unlock(); |
37d81828 PM |
3579 | } |
3580 | ||
906010b2 PZ |
3581 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3582 | { | |
3583 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3584 | struct ring_buffer *rb; |
906010b2 PZ |
3585 | int ret = VM_FAULT_SIGBUS; |
3586 | ||
3587 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3588 | if (vmf->pgoff == 0) | |
3589 | ret = 0; | |
3590 | return ret; | |
3591 | } | |
3592 | ||
3593 | rcu_read_lock(); | |
76369139 FW |
3594 | rb = rcu_dereference(event->rb); |
3595 | if (!rb) | |
906010b2 PZ |
3596 | goto unlock; |
3597 | ||
3598 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3599 | goto unlock; | |
3600 | ||
76369139 | 3601 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3602 | if (!vmf->page) |
3603 | goto unlock; | |
3604 | ||
3605 | get_page(vmf->page); | |
3606 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3607 | vmf->page->index = vmf->pgoff; | |
3608 | ||
3609 | ret = 0; | |
3610 | unlock: | |
3611 | rcu_read_unlock(); | |
3612 | ||
3613 | return ret; | |
3614 | } | |
3615 | ||
10c6db11 PZ |
3616 | static void ring_buffer_attach(struct perf_event *event, |
3617 | struct ring_buffer *rb) | |
3618 | { | |
3619 | unsigned long flags; | |
3620 | ||
3621 | if (!list_empty(&event->rb_entry)) | |
3622 | return; | |
3623 | ||
3624 | spin_lock_irqsave(&rb->event_lock, flags); | |
9bb5d40c PZ |
3625 | if (list_empty(&event->rb_entry)) |
3626 | list_add(&event->rb_entry, &rb->event_list); | |
10c6db11 PZ |
3627 | spin_unlock_irqrestore(&rb->event_lock, flags); |
3628 | } | |
3629 | ||
9bb5d40c | 3630 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb) |
10c6db11 PZ |
3631 | { |
3632 | unsigned long flags; | |
3633 | ||
3634 | if (list_empty(&event->rb_entry)) | |
3635 | return; | |
3636 | ||
3637 | spin_lock_irqsave(&rb->event_lock, flags); | |
3638 | list_del_init(&event->rb_entry); | |
3639 | wake_up_all(&event->waitq); | |
3640 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3641 | } | |
3642 | ||
3643 | static void ring_buffer_wakeup(struct perf_event *event) | |
3644 | { | |
3645 | struct ring_buffer *rb; | |
3646 | ||
3647 | rcu_read_lock(); | |
3648 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
3649 | if (rb) { |
3650 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
3651 | wake_up_all(&event->waitq); | |
3652 | } | |
10c6db11 PZ |
3653 | rcu_read_unlock(); |
3654 | } | |
3655 | ||
76369139 | 3656 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3657 | { |
76369139 | 3658 | struct ring_buffer *rb; |
906010b2 | 3659 | |
76369139 FW |
3660 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3661 | rb_free(rb); | |
7b732a75 PZ |
3662 | } |
3663 | ||
76369139 | 3664 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3665 | { |
76369139 | 3666 | struct ring_buffer *rb; |
7b732a75 | 3667 | |
ac9721f3 | 3668 | rcu_read_lock(); |
76369139 FW |
3669 | rb = rcu_dereference(event->rb); |
3670 | if (rb) { | |
3671 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3672 | rb = NULL; | |
ac9721f3 PZ |
3673 | } |
3674 | rcu_read_unlock(); | |
3675 | ||
76369139 | 3676 | return rb; |
ac9721f3 PZ |
3677 | } |
3678 | ||
76369139 | 3679 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3680 | { |
76369139 | 3681 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3682 | return; |
7b732a75 | 3683 | |
9bb5d40c | 3684 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 3685 | |
76369139 | 3686 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3687 | } |
3688 | ||
3689 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3690 | { | |
cdd6c482 | 3691 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3692 | |
cdd6c482 | 3693 | atomic_inc(&event->mmap_count); |
9bb5d40c | 3694 | atomic_inc(&event->rb->mmap_count); |
7b732a75 PZ |
3695 | } |
3696 | ||
9bb5d40c PZ |
3697 | /* |
3698 | * A buffer can be mmap()ed multiple times; either directly through the same | |
3699 | * event, or through other events by use of perf_event_set_output(). | |
3700 | * | |
3701 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
3702 | * the buffer here, where we still have a VM context. This means we need | |
3703 | * to detach all events redirecting to us. | |
3704 | */ | |
7b732a75 PZ |
3705 | static void perf_mmap_close(struct vm_area_struct *vma) |
3706 | { | |
cdd6c482 | 3707 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3708 | |
9bb5d40c PZ |
3709 | struct ring_buffer *rb = event->rb; |
3710 | struct user_struct *mmap_user = rb->mmap_user; | |
3711 | int mmap_locked = rb->mmap_locked; | |
3712 | unsigned long size = perf_data_size(rb); | |
789f90fc | 3713 | |
9bb5d40c PZ |
3714 | atomic_dec(&rb->mmap_count); |
3715 | ||
3716 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
3717 | return; | |
3718 | ||
3719 | /* Detach current event from the buffer. */ | |
3720 | rcu_assign_pointer(event->rb, NULL); | |
3721 | ring_buffer_detach(event, rb); | |
3722 | mutex_unlock(&event->mmap_mutex); | |
3723 | ||
3724 | /* If there's still other mmap()s of this buffer, we're done. */ | |
3725 | if (atomic_read(&rb->mmap_count)) { | |
3726 | ring_buffer_put(rb); /* can't be last */ | |
3727 | return; | |
3728 | } | |
ac9721f3 | 3729 | |
9bb5d40c PZ |
3730 | /* |
3731 | * No other mmap()s, detach from all other events that might redirect | |
3732 | * into the now unreachable buffer. Somewhat complicated by the | |
3733 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
3734 | */ | |
3735 | again: | |
3736 | rcu_read_lock(); | |
3737 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
3738 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
3739 | /* | |
3740 | * This event is en-route to free_event() which will | |
3741 | * detach it and remove it from the list. | |
3742 | */ | |
3743 | continue; | |
3744 | } | |
3745 | rcu_read_unlock(); | |
789f90fc | 3746 | |
9bb5d40c PZ |
3747 | mutex_lock(&event->mmap_mutex); |
3748 | /* | |
3749 | * Check we didn't race with perf_event_set_output() which can | |
3750 | * swizzle the rb from under us while we were waiting to | |
3751 | * acquire mmap_mutex. | |
3752 | * | |
3753 | * If we find a different rb; ignore this event, a next | |
3754 | * iteration will no longer find it on the list. We have to | |
3755 | * still restart the iteration to make sure we're not now | |
3756 | * iterating the wrong list. | |
3757 | */ | |
3758 | if (event->rb == rb) { | |
3759 | rcu_assign_pointer(event->rb, NULL); | |
3760 | ring_buffer_detach(event, rb); | |
3761 | ring_buffer_put(rb); /* can't be last, we still have one */ | |
26cb63ad | 3762 | } |
cdd6c482 | 3763 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 3764 | put_event(event); |
ac9721f3 | 3765 | |
9bb5d40c PZ |
3766 | /* |
3767 | * Restart the iteration; either we're on the wrong list or | |
3768 | * destroyed its integrity by doing a deletion. | |
3769 | */ | |
3770 | goto again; | |
7b732a75 | 3771 | } |
9bb5d40c PZ |
3772 | rcu_read_unlock(); |
3773 | ||
3774 | /* | |
3775 | * It could be there's still a few 0-ref events on the list; they'll | |
3776 | * get cleaned up by free_event() -- they'll also still have their | |
3777 | * ref on the rb and will free it whenever they are done with it. | |
3778 | * | |
3779 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
3780 | * undo the VM accounting. | |
3781 | */ | |
3782 | ||
3783 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
3784 | vma->vm_mm->pinned_vm -= mmap_locked; | |
3785 | free_uid(mmap_user); | |
3786 | ||
3787 | ring_buffer_put(rb); /* could be last */ | |
37d81828 PM |
3788 | } |
3789 | ||
f0f37e2f | 3790 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3791 | .open = perf_mmap_open, |
3792 | .close = perf_mmap_close, | |
3793 | .fault = perf_mmap_fault, | |
3794 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3795 | }; |
3796 | ||
3797 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3798 | { | |
cdd6c482 | 3799 | struct perf_event *event = file->private_data; |
22a4f650 | 3800 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3801 | struct user_struct *user = current_user(); |
22a4f650 | 3802 | unsigned long locked, lock_limit; |
76369139 | 3803 | struct ring_buffer *rb; |
7b732a75 PZ |
3804 | unsigned long vma_size; |
3805 | unsigned long nr_pages; | |
789f90fc | 3806 | long user_extra, extra; |
d57e34fd | 3807 | int ret = 0, flags = 0; |
37d81828 | 3808 | |
c7920614 PZ |
3809 | /* |
3810 | * Don't allow mmap() of inherited per-task counters. This would | |
3811 | * create a performance issue due to all children writing to the | |
76369139 | 3812 | * same rb. |
c7920614 PZ |
3813 | */ |
3814 | if (event->cpu == -1 && event->attr.inherit) | |
3815 | return -EINVAL; | |
3816 | ||
43a21ea8 | 3817 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3818 | return -EINVAL; |
7b732a75 PZ |
3819 | |
3820 | vma_size = vma->vm_end - vma->vm_start; | |
3821 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3822 | ||
7730d865 | 3823 | /* |
76369139 | 3824 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3825 | * can do bitmasks instead of modulo. |
3826 | */ | |
3827 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3828 | return -EINVAL; |
3829 | ||
7b732a75 | 3830 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3831 | return -EINVAL; |
3832 | ||
7b732a75 PZ |
3833 | if (vma->vm_pgoff != 0) |
3834 | return -EINVAL; | |
37d81828 | 3835 | |
cdd6c482 | 3836 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 3837 | again: |
cdd6c482 | 3838 | mutex_lock(&event->mmap_mutex); |
76369139 | 3839 | if (event->rb) { |
9bb5d40c | 3840 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 3841 | ret = -EINVAL; |
9bb5d40c PZ |
3842 | goto unlock; |
3843 | } | |
3844 | ||
3845 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
3846 | /* | |
3847 | * Raced against perf_mmap_close() through | |
3848 | * perf_event_set_output(). Try again, hope for better | |
3849 | * luck. | |
3850 | */ | |
3851 | mutex_unlock(&event->mmap_mutex); | |
3852 | goto again; | |
3853 | } | |
3854 | ||
ebb3c4c4 PZ |
3855 | goto unlock; |
3856 | } | |
3857 | ||
789f90fc | 3858 | user_extra = nr_pages + 1; |
cdd6c482 | 3859 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
3860 | |
3861 | /* | |
3862 | * Increase the limit linearly with more CPUs: | |
3863 | */ | |
3864 | user_lock_limit *= num_online_cpus(); | |
3865 | ||
789f90fc | 3866 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 3867 | |
789f90fc PZ |
3868 | extra = 0; |
3869 | if (user_locked > user_lock_limit) | |
3870 | extra = user_locked - user_lock_limit; | |
7b732a75 | 3871 | |
78d7d407 | 3872 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 3873 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 3874 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 3875 | |
459ec28a IM |
3876 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
3877 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
3878 | ret = -EPERM; |
3879 | goto unlock; | |
3880 | } | |
7b732a75 | 3881 | |
76369139 | 3882 | WARN_ON(event->rb); |
906010b2 | 3883 | |
d57e34fd | 3884 | if (vma->vm_flags & VM_WRITE) |
76369139 | 3885 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 3886 | |
4ec8363d VW |
3887 | rb = rb_alloc(nr_pages, |
3888 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
3889 | event->cpu, flags); | |
3890 | ||
76369139 | 3891 | if (!rb) { |
ac9721f3 | 3892 | ret = -ENOMEM; |
ebb3c4c4 | 3893 | goto unlock; |
ac9721f3 | 3894 | } |
26cb63ad | 3895 | |
9bb5d40c | 3896 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
3897 | rb->mmap_locked = extra; |
3898 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 3899 | |
ac9721f3 | 3900 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
3901 | vma->vm_mm->pinned_vm += extra; |
3902 | ||
9bb5d40c | 3903 | ring_buffer_attach(event, rb); |
26cb63ad | 3904 | rcu_assign_pointer(event->rb, rb); |
ac9721f3 | 3905 | |
9a0f05cb PZ |
3906 | perf_event_update_userpage(event); |
3907 | ||
ebb3c4c4 | 3908 | unlock: |
ac9721f3 PZ |
3909 | if (!ret) |
3910 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 3911 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 3912 | |
9bb5d40c PZ |
3913 | /* |
3914 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
3915 | * vma. | |
3916 | */ | |
26cb63ad | 3917 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 3918 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
3919 | |
3920 | return ret; | |
37d81828 PM |
3921 | } |
3922 | ||
3c446b3d PZ |
3923 | static int perf_fasync(int fd, struct file *filp, int on) |
3924 | { | |
496ad9aa | 3925 | struct inode *inode = file_inode(filp); |
cdd6c482 | 3926 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
3927 | int retval; |
3928 | ||
3929 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 3930 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
3931 | mutex_unlock(&inode->i_mutex); |
3932 | ||
3933 | if (retval < 0) | |
3934 | return retval; | |
3935 | ||
3936 | return 0; | |
3937 | } | |
3938 | ||
0793a61d | 3939 | static const struct file_operations perf_fops = { |
3326c1ce | 3940 | .llseek = no_llseek, |
0793a61d TG |
3941 | .release = perf_release, |
3942 | .read = perf_read, | |
3943 | .poll = perf_poll, | |
d859e29f PM |
3944 | .unlocked_ioctl = perf_ioctl, |
3945 | .compat_ioctl = perf_ioctl, | |
37d81828 | 3946 | .mmap = perf_mmap, |
3c446b3d | 3947 | .fasync = perf_fasync, |
0793a61d TG |
3948 | }; |
3949 | ||
925d519a | 3950 | /* |
cdd6c482 | 3951 | * Perf event wakeup |
925d519a PZ |
3952 | * |
3953 | * If there's data, ensure we set the poll() state and publish everything | |
3954 | * to user-space before waking everybody up. | |
3955 | */ | |
3956 | ||
cdd6c482 | 3957 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 3958 | { |
10c6db11 | 3959 | ring_buffer_wakeup(event); |
4c9e2542 | 3960 | |
cdd6c482 IM |
3961 | if (event->pending_kill) { |
3962 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
3963 | event->pending_kill = 0; | |
4c9e2542 | 3964 | } |
925d519a PZ |
3965 | } |
3966 | ||
e360adbe | 3967 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 3968 | { |
cdd6c482 IM |
3969 | struct perf_event *event = container_of(entry, |
3970 | struct perf_event, pending); | |
79f14641 | 3971 | |
cdd6c482 IM |
3972 | if (event->pending_disable) { |
3973 | event->pending_disable = 0; | |
3974 | __perf_event_disable(event); | |
79f14641 PZ |
3975 | } |
3976 | ||
cdd6c482 IM |
3977 | if (event->pending_wakeup) { |
3978 | event->pending_wakeup = 0; | |
3979 | perf_event_wakeup(event); | |
79f14641 PZ |
3980 | } |
3981 | } | |
3982 | ||
39447b38 ZY |
3983 | /* |
3984 | * We assume there is only KVM supporting the callbacks. | |
3985 | * Later on, we might change it to a list if there is | |
3986 | * another virtualization implementation supporting the callbacks. | |
3987 | */ | |
3988 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
3989 | ||
3990 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3991 | { | |
3992 | perf_guest_cbs = cbs; | |
3993 | return 0; | |
3994 | } | |
3995 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
3996 | ||
3997 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
3998 | { | |
3999 | perf_guest_cbs = NULL; | |
4000 | return 0; | |
4001 | } | |
4002 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4003 | ||
4018994f JO |
4004 | static void |
4005 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4006 | struct pt_regs *regs, u64 mask) | |
4007 | { | |
4008 | int bit; | |
4009 | ||
4010 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4011 | sizeof(mask) * BITS_PER_BYTE) { | |
4012 | u64 val; | |
4013 | ||
4014 | val = perf_reg_value(regs, bit); | |
4015 | perf_output_put(handle, val); | |
4016 | } | |
4017 | } | |
4018 | ||
4019 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
4020 | struct pt_regs *regs) | |
4021 | { | |
4022 | if (!user_mode(regs)) { | |
4023 | if (current->mm) | |
4024 | regs = task_pt_regs(current); | |
4025 | else | |
4026 | regs = NULL; | |
4027 | } | |
4028 | ||
4029 | if (regs) { | |
4030 | regs_user->regs = regs; | |
4031 | regs_user->abi = perf_reg_abi(current); | |
4032 | } | |
4033 | } | |
4034 | ||
c5ebcedb JO |
4035 | /* |
4036 | * Get remaining task size from user stack pointer. | |
4037 | * | |
4038 | * It'd be better to take stack vma map and limit this more | |
4039 | * precisly, but there's no way to get it safely under interrupt, | |
4040 | * so using TASK_SIZE as limit. | |
4041 | */ | |
4042 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4043 | { | |
4044 | unsigned long addr = perf_user_stack_pointer(regs); | |
4045 | ||
4046 | if (!addr || addr >= TASK_SIZE) | |
4047 | return 0; | |
4048 | ||
4049 | return TASK_SIZE - addr; | |
4050 | } | |
4051 | ||
4052 | static u16 | |
4053 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4054 | struct pt_regs *regs) | |
4055 | { | |
4056 | u64 task_size; | |
4057 | ||
4058 | /* No regs, no stack pointer, no dump. */ | |
4059 | if (!regs) | |
4060 | return 0; | |
4061 | ||
4062 | /* | |
4063 | * Check if we fit in with the requested stack size into the: | |
4064 | * - TASK_SIZE | |
4065 | * If we don't, we limit the size to the TASK_SIZE. | |
4066 | * | |
4067 | * - remaining sample size | |
4068 | * If we don't, we customize the stack size to | |
4069 | * fit in to the remaining sample size. | |
4070 | */ | |
4071 | ||
4072 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4073 | stack_size = min(stack_size, (u16) task_size); | |
4074 | ||
4075 | /* Current header size plus static size and dynamic size. */ | |
4076 | header_size += 2 * sizeof(u64); | |
4077 | ||
4078 | /* Do we fit in with the current stack dump size? */ | |
4079 | if ((u16) (header_size + stack_size) < header_size) { | |
4080 | /* | |
4081 | * If we overflow the maximum size for the sample, | |
4082 | * we customize the stack dump size to fit in. | |
4083 | */ | |
4084 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4085 | stack_size = round_up(stack_size, sizeof(u64)); | |
4086 | } | |
4087 | ||
4088 | return stack_size; | |
4089 | } | |
4090 | ||
4091 | static void | |
4092 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4093 | struct pt_regs *regs) | |
4094 | { | |
4095 | /* Case of a kernel thread, nothing to dump */ | |
4096 | if (!regs) { | |
4097 | u64 size = 0; | |
4098 | perf_output_put(handle, size); | |
4099 | } else { | |
4100 | unsigned long sp; | |
4101 | unsigned int rem; | |
4102 | u64 dyn_size; | |
4103 | ||
4104 | /* | |
4105 | * We dump: | |
4106 | * static size | |
4107 | * - the size requested by user or the best one we can fit | |
4108 | * in to the sample max size | |
4109 | * data | |
4110 | * - user stack dump data | |
4111 | * dynamic size | |
4112 | * - the actual dumped size | |
4113 | */ | |
4114 | ||
4115 | /* Static size. */ | |
4116 | perf_output_put(handle, dump_size); | |
4117 | ||
4118 | /* Data. */ | |
4119 | sp = perf_user_stack_pointer(regs); | |
4120 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4121 | dyn_size = dump_size - rem; | |
4122 | ||
4123 | perf_output_skip(handle, rem); | |
4124 | ||
4125 | /* Dynamic size. */ | |
4126 | perf_output_put(handle, dyn_size); | |
4127 | } | |
4128 | } | |
4129 | ||
c980d109 ACM |
4130 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4131 | struct perf_sample_data *data, | |
4132 | struct perf_event *event) | |
6844c09d ACM |
4133 | { |
4134 | u64 sample_type = event->attr.sample_type; | |
4135 | ||
4136 | data->type = sample_type; | |
4137 | header->size += event->id_header_size; | |
4138 | ||
4139 | if (sample_type & PERF_SAMPLE_TID) { | |
4140 | /* namespace issues */ | |
4141 | data->tid_entry.pid = perf_event_pid(event, current); | |
4142 | data->tid_entry.tid = perf_event_tid(event, current); | |
4143 | } | |
4144 | ||
4145 | if (sample_type & PERF_SAMPLE_TIME) | |
4146 | data->time = perf_clock(); | |
4147 | ||
4148 | if (sample_type & PERF_SAMPLE_ID) | |
4149 | data->id = primary_event_id(event); | |
4150 | ||
4151 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4152 | data->stream_id = event->id; | |
4153 | ||
4154 | if (sample_type & PERF_SAMPLE_CPU) { | |
4155 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4156 | data->cpu_entry.reserved = 0; | |
4157 | } | |
4158 | } | |
4159 | ||
76369139 FW |
4160 | void perf_event_header__init_id(struct perf_event_header *header, |
4161 | struct perf_sample_data *data, | |
4162 | struct perf_event *event) | |
c980d109 ACM |
4163 | { |
4164 | if (event->attr.sample_id_all) | |
4165 | __perf_event_header__init_id(header, data, event); | |
4166 | } | |
4167 | ||
4168 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4169 | struct perf_sample_data *data) | |
4170 | { | |
4171 | u64 sample_type = data->type; | |
4172 | ||
4173 | if (sample_type & PERF_SAMPLE_TID) | |
4174 | perf_output_put(handle, data->tid_entry); | |
4175 | ||
4176 | if (sample_type & PERF_SAMPLE_TIME) | |
4177 | perf_output_put(handle, data->time); | |
4178 | ||
4179 | if (sample_type & PERF_SAMPLE_ID) | |
4180 | perf_output_put(handle, data->id); | |
4181 | ||
4182 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4183 | perf_output_put(handle, data->stream_id); | |
4184 | ||
4185 | if (sample_type & PERF_SAMPLE_CPU) | |
4186 | perf_output_put(handle, data->cpu_entry); | |
4187 | } | |
4188 | ||
76369139 FW |
4189 | void perf_event__output_id_sample(struct perf_event *event, |
4190 | struct perf_output_handle *handle, | |
4191 | struct perf_sample_data *sample) | |
c980d109 ACM |
4192 | { |
4193 | if (event->attr.sample_id_all) | |
4194 | __perf_event__output_id_sample(handle, sample); | |
4195 | } | |
4196 | ||
3dab77fb | 4197 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4198 | struct perf_event *event, |
4199 | u64 enabled, u64 running) | |
3dab77fb | 4200 | { |
cdd6c482 | 4201 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4202 | u64 values[4]; |
4203 | int n = 0; | |
4204 | ||
b5e58793 | 4205 | values[n++] = perf_event_count(event); |
3dab77fb | 4206 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4207 | values[n++] = enabled + |
cdd6c482 | 4208 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4209 | } |
4210 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4211 | values[n++] = running + |
cdd6c482 | 4212 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4213 | } |
4214 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4215 | values[n++] = primary_event_id(event); |
3dab77fb | 4216 | |
76369139 | 4217 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4218 | } |
4219 | ||
4220 | /* | |
cdd6c482 | 4221 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4222 | */ |
4223 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4224 | struct perf_event *event, |
4225 | u64 enabled, u64 running) | |
3dab77fb | 4226 | { |
cdd6c482 IM |
4227 | struct perf_event *leader = event->group_leader, *sub; |
4228 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4229 | u64 values[5]; |
4230 | int n = 0; | |
4231 | ||
4232 | values[n++] = 1 + leader->nr_siblings; | |
4233 | ||
4234 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4235 | values[n++] = enabled; |
3dab77fb PZ |
4236 | |
4237 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4238 | values[n++] = running; |
3dab77fb | 4239 | |
cdd6c482 | 4240 | if (leader != event) |
3dab77fb PZ |
4241 | leader->pmu->read(leader); |
4242 | ||
b5e58793 | 4243 | values[n++] = perf_event_count(leader); |
3dab77fb | 4244 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4245 | values[n++] = primary_event_id(leader); |
3dab77fb | 4246 | |
76369139 | 4247 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4248 | |
65abc865 | 4249 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4250 | n = 0; |
4251 | ||
cdd6c482 | 4252 | if (sub != event) |
3dab77fb PZ |
4253 | sub->pmu->read(sub); |
4254 | ||
b5e58793 | 4255 | values[n++] = perf_event_count(sub); |
3dab77fb | 4256 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4257 | values[n++] = primary_event_id(sub); |
3dab77fb | 4258 | |
76369139 | 4259 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4260 | } |
4261 | } | |
4262 | ||
eed01528 SE |
4263 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4264 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4265 | ||
3dab77fb | 4266 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4267 | struct perf_event *event) |
3dab77fb | 4268 | { |
e3f3541c | 4269 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4270 | u64 read_format = event->attr.read_format; |
4271 | ||
4272 | /* | |
4273 | * compute total_time_enabled, total_time_running | |
4274 | * based on snapshot values taken when the event | |
4275 | * was last scheduled in. | |
4276 | * | |
4277 | * we cannot simply called update_context_time() | |
4278 | * because of locking issue as we are called in | |
4279 | * NMI context | |
4280 | */ | |
c4794295 | 4281 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4282 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4283 | |
cdd6c482 | 4284 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4285 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4286 | else |
eed01528 | 4287 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4288 | } |
4289 | ||
5622f295 MM |
4290 | void perf_output_sample(struct perf_output_handle *handle, |
4291 | struct perf_event_header *header, | |
4292 | struct perf_sample_data *data, | |
cdd6c482 | 4293 | struct perf_event *event) |
5622f295 MM |
4294 | { |
4295 | u64 sample_type = data->type; | |
4296 | ||
4297 | perf_output_put(handle, *header); | |
4298 | ||
4299 | if (sample_type & PERF_SAMPLE_IP) | |
4300 | perf_output_put(handle, data->ip); | |
4301 | ||
4302 | if (sample_type & PERF_SAMPLE_TID) | |
4303 | perf_output_put(handle, data->tid_entry); | |
4304 | ||
4305 | if (sample_type & PERF_SAMPLE_TIME) | |
4306 | perf_output_put(handle, data->time); | |
4307 | ||
4308 | if (sample_type & PERF_SAMPLE_ADDR) | |
4309 | perf_output_put(handle, data->addr); | |
4310 | ||
4311 | if (sample_type & PERF_SAMPLE_ID) | |
4312 | perf_output_put(handle, data->id); | |
4313 | ||
4314 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4315 | perf_output_put(handle, data->stream_id); | |
4316 | ||
4317 | if (sample_type & PERF_SAMPLE_CPU) | |
4318 | perf_output_put(handle, data->cpu_entry); | |
4319 | ||
4320 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4321 | perf_output_put(handle, data->period); | |
4322 | ||
4323 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4324 | perf_output_read(handle, event); |
5622f295 MM |
4325 | |
4326 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4327 | if (data->callchain) { | |
4328 | int size = 1; | |
4329 | ||
4330 | if (data->callchain) | |
4331 | size += data->callchain->nr; | |
4332 | ||
4333 | size *= sizeof(u64); | |
4334 | ||
76369139 | 4335 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4336 | } else { |
4337 | u64 nr = 0; | |
4338 | perf_output_put(handle, nr); | |
4339 | } | |
4340 | } | |
4341 | ||
4342 | if (sample_type & PERF_SAMPLE_RAW) { | |
4343 | if (data->raw) { | |
4344 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4345 | __output_copy(handle, data->raw->data, |
4346 | data->raw->size); | |
5622f295 MM |
4347 | } else { |
4348 | struct { | |
4349 | u32 size; | |
4350 | u32 data; | |
4351 | } raw = { | |
4352 | .size = sizeof(u32), | |
4353 | .data = 0, | |
4354 | }; | |
4355 | perf_output_put(handle, raw); | |
4356 | } | |
4357 | } | |
a7ac67ea PZ |
4358 | |
4359 | if (!event->attr.watermark) { | |
4360 | int wakeup_events = event->attr.wakeup_events; | |
4361 | ||
4362 | if (wakeup_events) { | |
4363 | struct ring_buffer *rb = handle->rb; | |
4364 | int events = local_inc_return(&rb->events); | |
4365 | ||
4366 | if (events >= wakeup_events) { | |
4367 | local_sub(wakeup_events, &rb->events); | |
4368 | local_inc(&rb->wakeup); | |
4369 | } | |
4370 | } | |
4371 | } | |
bce38cd5 SE |
4372 | |
4373 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4374 | if (data->br_stack) { | |
4375 | size_t size; | |
4376 | ||
4377 | size = data->br_stack->nr | |
4378 | * sizeof(struct perf_branch_entry); | |
4379 | ||
4380 | perf_output_put(handle, data->br_stack->nr); | |
4381 | perf_output_copy(handle, data->br_stack->entries, size); | |
4382 | } else { | |
4383 | /* | |
4384 | * we always store at least the value of nr | |
4385 | */ | |
4386 | u64 nr = 0; | |
4387 | perf_output_put(handle, nr); | |
4388 | } | |
4389 | } | |
4018994f JO |
4390 | |
4391 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4392 | u64 abi = data->regs_user.abi; | |
4393 | ||
4394 | /* | |
4395 | * If there are no regs to dump, notice it through | |
4396 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4397 | */ | |
4398 | perf_output_put(handle, abi); | |
4399 | ||
4400 | if (abi) { | |
4401 | u64 mask = event->attr.sample_regs_user; | |
4402 | perf_output_sample_regs(handle, | |
4403 | data->regs_user.regs, | |
4404 | mask); | |
4405 | } | |
4406 | } | |
c5ebcedb JO |
4407 | |
4408 | if (sample_type & PERF_SAMPLE_STACK_USER) | |
4409 | perf_output_sample_ustack(handle, | |
4410 | data->stack_user_size, | |
4411 | data->regs_user.regs); | |
c3feedf2 AK |
4412 | |
4413 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
4414 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
4415 | |
4416 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
4417 | perf_output_put(handle, data->data_src.val); | |
5622f295 MM |
4418 | } |
4419 | ||
4420 | void perf_prepare_sample(struct perf_event_header *header, | |
4421 | struct perf_sample_data *data, | |
cdd6c482 | 4422 | struct perf_event *event, |
5622f295 | 4423 | struct pt_regs *regs) |
7b732a75 | 4424 | { |
cdd6c482 | 4425 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4426 | |
cdd6c482 | 4427 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4428 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4429 | |
4430 | header->misc = 0; | |
4431 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4432 | |
c980d109 | 4433 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4434 | |
c320c7b7 | 4435 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4436 | data->ip = perf_instruction_pointer(regs); |
4437 | ||
b23f3325 | 4438 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4439 | int size = 1; |
394ee076 | 4440 | |
e6dab5ff | 4441 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4442 | |
4443 | if (data->callchain) | |
4444 | size += data->callchain->nr; | |
4445 | ||
4446 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4447 | } |
4448 | ||
3a43ce68 | 4449 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4450 | int size = sizeof(u32); |
4451 | ||
4452 | if (data->raw) | |
4453 | size += data->raw->size; | |
4454 | else | |
4455 | size += sizeof(u32); | |
4456 | ||
4457 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4458 | header->size += size; |
7f453c24 | 4459 | } |
bce38cd5 SE |
4460 | |
4461 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4462 | int size = sizeof(u64); /* nr */ | |
4463 | if (data->br_stack) { | |
4464 | size += data->br_stack->nr | |
4465 | * sizeof(struct perf_branch_entry); | |
4466 | } | |
4467 | header->size += size; | |
4468 | } | |
4018994f JO |
4469 | |
4470 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4471 | /* regs dump ABI info */ | |
4472 | int size = sizeof(u64); | |
4473 | ||
4474 | perf_sample_regs_user(&data->regs_user, regs); | |
4475 | ||
4476 | if (data->regs_user.regs) { | |
4477 | u64 mask = event->attr.sample_regs_user; | |
4478 | size += hweight64(mask) * sizeof(u64); | |
4479 | } | |
4480 | ||
4481 | header->size += size; | |
4482 | } | |
c5ebcedb JO |
4483 | |
4484 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4485 | /* | |
4486 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4487 | * processed as the last one or have additional check added | |
4488 | * in case new sample type is added, because we could eat | |
4489 | * up the rest of the sample size. | |
4490 | */ | |
4491 | struct perf_regs_user *uregs = &data->regs_user; | |
4492 | u16 stack_size = event->attr.sample_stack_user; | |
4493 | u16 size = sizeof(u64); | |
4494 | ||
4495 | if (!uregs->abi) | |
4496 | perf_sample_regs_user(uregs, regs); | |
4497 | ||
4498 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4499 | uregs->regs); | |
4500 | ||
4501 | /* | |
4502 | * If there is something to dump, add space for the dump | |
4503 | * itself and for the field that tells the dynamic size, | |
4504 | * which is how many have been actually dumped. | |
4505 | */ | |
4506 | if (stack_size) | |
4507 | size += sizeof(u64) + stack_size; | |
4508 | ||
4509 | data->stack_user_size = stack_size; | |
4510 | header->size += size; | |
4511 | } | |
5622f295 | 4512 | } |
7f453c24 | 4513 | |
a8b0ca17 | 4514 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4515 | struct perf_sample_data *data, |
4516 | struct pt_regs *regs) | |
4517 | { | |
4518 | struct perf_output_handle handle; | |
4519 | struct perf_event_header header; | |
689802b2 | 4520 | |
927c7a9e FW |
4521 | /* protect the callchain buffers */ |
4522 | rcu_read_lock(); | |
4523 | ||
cdd6c482 | 4524 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4525 | |
a7ac67ea | 4526 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4527 | goto exit; |
0322cd6e | 4528 | |
cdd6c482 | 4529 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4530 | |
8a057d84 | 4531 | perf_output_end(&handle); |
927c7a9e FW |
4532 | |
4533 | exit: | |
4534 | rcu_read_unlock(); | |
0322cd6e PZ |
4535 | } |
4536 | ||
38b200d6 | 4537 | /* |
cdd6c482 | 4538 | * read event_id |
38b200d6 PZ |
4539 | */ |
4540 | ||
4541 | struct perf_read_event { | |
4542 | struct perf_event_header header; | |
4543 | ||
4544 | u32 pid; | |
4545 | u32 tid; | |
38b200d6 PZ |
4546 | }; |
4547 | ||
4548 | static void | |
cdd6c482 | 4549 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4550 | struct task_struct *task) |
4551 | { | |
4552 | struct perf_output_handle handle; | |
c980d109 | 4553 | struct perf_sample_data sample; |
dfc65094 | 4554 | struct perf_read_event read_event = { |
38b200d6 | 4555 | .header = { |
cdd6c482 | 4556 | .type = PERF_RECORD_READ, |
38b200d6 | 4557 | .misc = 0, |
c320c7b7 | 4558 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4559 | }, |
cdd6c482 IM |
4560 | .pid = perf_event_pid(event, task), |
4561 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4562 | }; |
3dab77fb | 4563 | int ret; |
38b200d6 | 4564 | |
c980d109 | 4565 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4566 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4567 | if (ret) |
4568 | return; | |
4569 | ||
dfc65094 | 4570 | perf_output_put(&handle, read_event); |
cdd6c482 | 4571 | perf_output_read(&handle, event); |
c980d109 | 4572 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4573 | |
38b200d6 PZ |
4574 | perf_output_end(&handle); |
4575 | } | |
4576 | ||
52d857a8 JO |
4577 | typedef int (perf_event_aux_match_cb)(struct perf_event *event, void *data); |
4578 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); | |
4579 | ||
4580 | static void | |
4581 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
4582 | perf_event_aux_match_cb match, | |
4583 | perf_event_aux_output_cb output, | |
4584 | void *data) | |
4585 | { | |
4586 | struct perf_event *event; | |
4587 | ||
4588 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
4589 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
4590 | continue; | |
4591 | if (!event_filter_match(event)) | |
4592 | continue; | |
4593 | if (match(event, data)) | |
4594 | output(event, data); | |
4595 | } | |
4596 | } | |
4597 | ||
4598 | static void | |
4599 | perf_event_aux(perf_event_aux_match_cb match, | |
4600 | perf_event_aux_output_cb output, | |
4601 | void *data, | |
4602 | struct perf_event_context *task_ctx) | |
4603 | { | |
4604 | struct perf_cpu_context *cpuctx; | |
4605 | struct perf_event_context *ctx; | |
4606 | struct pmu *pmu; | |
4607 | int ctxn; | |
4608 | ||
4609 | rcu_read_lock(); | |
4610 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
4611 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
4612 | if (cpuctx->unique_pmu != pmu) | |
4613 | goto next; | |
4614 | perf_event_aux_ctx(&cpuctx->ctx, match, output, data); | |
4615 | if (task_ctx) | |
4616 | goto next; | |
4617 | ctxn = pmu->task_ctx_nr; | |
4618 | if (ctxn < 0) | |
4619 | goto next; | |
4620 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4621 | if (ctx) | |
4622 | perf_event_aux_ctx(ctx, match, output, data); | |
4623 | next: | |
4624 | put_cpu_ptr(pmu->pmu_cpu_context); | |
4625 | } | |
4626 | ||
4627 | if (task_ctx) { | |
4628 | preempt_disable(); | |
4629 | perf_event_aux_ctx(task_ctx, match, output, data); | |
4630 | preempt_enable(); | |
4631 | } | |
4632 | rcu_read_unlock(); | |
4633 | } | |
4634 | ||
60313ebe | 4635 | /* |
9f498cc5 PZ |
4636 | * task tracking -- fork/exit |
4637 | * | |
3af9e859 | 4638 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
60313ebe PZ |
4639 | */ |
4640 | ||
9f498cc5 | 4641 | struct perf_task_event { |
3a80b4a3 | 4642 | struct task_struct *task; |
cdd6c482 | 4643 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4644 | |
4645 | struct { | |
4646 | struct perf_event_header header; | |
4647 | ||
4648 | u32 pid; | |
4649 | u32 ppid; | |
9f498cc5 PZ |
4650 | u32 tid; |
4651 | u32 ptid; | |
393b2ad8 | 4652 | u64 time; |
cdd6c482 | 4653 | } event_id; |
60313ebe PZ |
4654 | }; |
4655 | ||
cdd6c482 | 4656 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 4657 | void *data) |
60313ebe | 4658 | { |
52d857a8 | 4659 | struct perf_task_event *task_event = data; |
60313ebe | 4660 | struct perf_output_handle handle; |
c980d109 | 4661 | struct perf_sample_data sample; |
9f498cc5 | 4662 | struct task_struct *task = task_event->task; |
c980d109 | 4663 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4664 | |
c980d109 | 4665 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4666 | |
c980d109 | 4667 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4668 | task_event->event_id.header.size); |
ef60777c | 4669 | if (ret) |
c980d109 | 4670 | goto out; |
60313ebe | 4671 | |
cdd6c482 IM |
4672 | task_event->event_id.pid = perf_event_pid(event, task); |
4673 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4674 | |
cdd6c482 IM |
4675 | task_event->event_id.tid = perf_event_tid(event, task); |
4676 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4677 | |
cdd6c482 | 4678 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4679 | |
c980d109 ACM |
4680 | perf_event__output_id_sample(event, &handle, &sample); |
4681 | ||
60313ebe | 4682 | perf_output_end(&handle); |
c980d109 ACM |
4683 | out: |
4684 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4685 | } |
4686 | ||
52d857a8 JO |
4687 | static int perf_event_task_match(struct perf_event *event, |
4688 | void *data __maybe_unused) | |
60313ebe | 4689 | { |
52d857a8 JO |
4690 | return event->attr.comm || event->attr.mmap || |
4691 | event->attr.mmap_data || event->attr.task; | |
60313ebe PZ |
4692 | } |
4693 | ||
cdd6c482 IM |
4694 | static void perf_event_task(struct task_struct *task, |
4695 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4696 | int new) |
60313ebe | 4697 | { |
9f498cc5 | 4698 | struct perf_task_event task_event; |
60313ebe | 4699 | |
cdd6c482 IM |
4700 | if (!atomic_read(&nr_comm_events) && |
4701 | !atomic_read(&nr_mmap_events) && | |
4702 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4703 | return; |
4704 | ||
9f498cc5 | 4705 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4706 | .task = task, |
4707 | .task_ctx = task_ctx, | |
cdd6c482 | 4708 | .event_id = { |
60313ebe | 4709 | .header = { |
cdd6c482 | 4710 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4711 | .misc = 0, |
cdd6c482 | 4712 | .size = sizeof(task_event.event_id), |
60313ebe | 4713 | }, |
573402db PZ |
4714 | /* .pid */ |
4715 | /* .ppid */ | |
9f498cc5 PZ |
4716 | /* .tid */ |
4717 | /* .ptid */ | |
6f93d0a7 | 4718 | .time = perf_clock(), |
60313ebe PZ |
4719 | }, |
4720 | }; | |
4721 | ||
52d857a8 JO |
4722 | perf_event_aux(perf_event_task_match, |
4723 | perf_event_task_output, | |
4724 | &task_event, | |
4725 | task_ctx); | |
9f498cc5 PZ |
4726 | } |
4727 | ||
cdd6c482 | 4728 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4729 | { |
cdd6c482 | 4730 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4731 | } |
4732 | ||
8d1b2d93 PZ |
4733 | /* |
4734 | * comm tracking | |
4735 | */ | |
4736 | ||
4737 | struct perf_comm_event { | |
22a4f650 IM |
4738 | struct task_struct *task; |
4739 | char *comm; | |
8d1b2d93 PZ |
4740 | int comm_size; |
4741 | ||
4742 | struct { | |
4743 | struct perf_event_header header; | |
4744 | ||
4745 | u32 pid; | |
4746 | u32 tid; | |
cdd6c482 | 4747 | } event_id; |
8d1b2d93 PZ |
4748 | }; |
4749 | ||
cdd6c482 | 4750 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 4751 | void *data) |
8d1b2d93 | 4752 | { |
52d857a8 | 4753 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 4754 | struct perf_output_handle handle; |
c980d109 | 4755 | struct perf_sample_data sample; |
cdd6c482 | 4756 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4757 | int ret; |
4758 | ||
4759 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); | |
4760 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4761 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4762 | |
4763 | if (ret) | |
c980d109 | 4764 | goto out; |
8d1b2d93 | 4765 | |
cdd6c482 IM |
4766 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4767 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4768 | |
cdd6c482 | 4769 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4770 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4771 | comm_event->comm_size); |
c980d109 ACM |
4772 | |
4773 | perf_event__output_id_sample(event, &handle, &sample); | |
4774 | ||
8d1b2d93 | 4775 | perf_output_end(&handle); |
c980d109 ACM |
4776 | out: |
4777 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4778 | } |
4779 | ||
52d857a8 JO |
4780 | static int perf_event_comm_match(struct perf_event *event, |
4781 | void *data __maybe_unused) | |
8d1b2d93 | 4782 | { |
52d857a8 | 4783 | return event->attr.comm; |
8d1b2d93 PZ |
4784 | } |
4785 | ||
cdd6c482 | 4786 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 4787 | { |
413ee3b4 | 4788 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4789 | unsigned int size; |
8d1b2d93 | 4790 | |
413ee3b4 | 4791 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4792 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4793 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4794 | |
4795 | comm_event->comm = comm; | |
4796 | comm_event->comm_size = size; | |
4797 | ||
cdd6c482 | 4798 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 4799 | |
52d857a8 JO |
4800 | perf_event_aux(perf_event_comm_match, |
4801 | perf_event_comm_output, | |
4802 | comm_event, | |
4803 | NULL); | |
8d1b2d93 PZ |
4804 | } |
4805 | ||
cdd6c482 | 4806 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4807 | { |
9ee318a7 | 4808 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4809 | struct perf_event_context *ctx; |
4810 | int ctxn; | |
9ee318a7 | 4811 | |
c79aa0d9 | 4812 | rcu_read_lock(); |
8dc85d54 PZ |
4813 | for_each_task_context_nr(ctxn) { |
4814 | ctx = task->perf_event_ctxp[ctxn]; | |
4815 | if (!ctx) | |
4816 | continue; | |
9ee318a7 | 4817 | |
8dc85d54 PZ |
4818 | perf_event_enable_on_exec(ctx); |
4819 | } | |
c79aa0d9 | 4820 | rcu_read_unlock(); |
9ee318a7 | 4821 | |
cdd6c482 | 4822 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4823 | return; |
a63eaf34 | 4824 | |
9ee318a7 | 4825 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4826 | .task = task, |
573402db PZ |
4827 | /* .comm */ |
4828 | /* .comm_size */ | |
cdd6c482 | 4829 | .event_id = { |
573402db | 4830 | .header = { |
cdd6c482 | 4831 | .type = PERF_RECORD_COMM, |
573402db PZ |
4832 | .misc = 0, |
4833 | /* .size */ | |
4834 | }, | |
4835 | /* .pid */ | |
4836 | /* .tid */ | |
8d1b2d93 PZ |
4837 | }, |
4838 | }; | |
4839 | ||
cdd6c482 | 4840 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
4841 | } |
4842 | ||
0a4a9391 PZ |
4843 | /* |
4844 | * mmap tracking | |
4845 | */ | |
4846 | ||
4847 | struct perf_mmap_event { | |
089dd79d PZ |
4848 | struct vm_area_struct *vma; |
4849 | ||
4850 | const char *file_name; | |
4851 | int file_size; | |
0a4a9391 PZ |
4852 | |
4853 | struct { | |
4854 | struct perf_event_header header; | |
4855 | ||
4856 | u32 pid; | |
4857 | u32 tid; | |
4858 | u64 start; | |
4859 | u64 len; | |
4860 | u64 pgoff; | |
cdd6c482 | 4861 | } event_id; |
0a4a9391 PZ |
4862 | }; |
4863 | ||
cdd6c482 | 4864 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 4865 | void *data) |
0a4a9391 | 4866 | { |
52d857a8 | 4867 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 4868 | struct perf_output_handle handle; |
c980d109 | 4869 | struct perf_sample_data sample; |
cdd6c482 | 4870 | int size = mmap_event->event_id.header.size; |
c980d109 | 4871 | int ret; |
0a4a9391 | 4872 | |
c980d109 ACM |
4873 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
4874 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4875 | mmap_event->event_id.header.size); |
0a4a9391 | 4876 | if (ret) |
c980d109 | 4877 | goto out; |
0a4a9391 | 4878 | |
cdd6c482 IM |
4879 | mmap_event->event_id.pid = perf_event_pid(event, current); |
4880 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 4881 | |
cdd6c482 | 4882 | perf_output_put(&handle, mmap_event->event_id); |
76369139 | 4883 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 4884 | mmap_event->file_size); |
c980d109 ACM |
4885 | |
4886 | perf_event__output_id_sample(event, &handle, &sample); | |
4887 | ||
78d613eb | 4888 | perf_output_end(&handle); |
c980d109 ACM |
4889 | out: |
4890 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
4891 | } |
4892 | ||
cdd6c482 | 4893 | static int perf_event_mmap_match(struct perf_event *event, |
52d857a8 | 4894 | void *data) |
0a4a9391 | 4895 | { |
52d857a8 JO |
4896 | struct perf_mmap_event *mmap_event = data; |
4897 | struct vm_area_struct *vma = mmap_event->vma; | |
4898 | int executable = vma->vm_flags & VM_EXEC; | |
0a4a9391 | 4899 | |
52d857a8 JO |
4900 | return (!executable && event->attr.mmap_data) || |
4901 | (executable && event->attr.mmap); | |
0a4a9391 PZ |
4902 | } |
4903 | ||
cdd6c482 | 4904 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 4905 | { |
089dd79d PZ |
4906 | struct vm_area_struct *vma = mmap_event->vma; |
4907 | struct file *file = vma->vm_file; | |
0a4a9391 PZ |
4908 | unsigned int size; |
4909 | char tmp[16]; | |
4910 | char *buf = NULL; | |
089dd79d | 4911 | const char *name; |
0a4a9391 | 4912 | |
413ee3b4 AB |
4913 | memset(tmp, 0, sizeof(tmp)); |
4914 | ||
0a4a9391 | 4915 | if (file) { |
413ee3b4 | 4916 | /* |
76369139 | 4917 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
4918 | * need to add enough zero bytes after the string to handle |
4919 | * the 64bit alignment we do later. | |
4920 | */ | |
4921 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
4922 | if (!buf) { |
4923 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
4924 | goto got_name; | |
4925 | } | |
d3d21c41 | 4926 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
4927 | if (IS_ERR(name)) { |
4928 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
4929 | goto got_name; | |
4930 | } | |
4931 | } else { | |
413ee3b4 AB |
4932 | if (arch_vma_name(mmap_event->vma)) { |
4933 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
c97847d2 CG |
4934 | sizeof(tmp) - 1); |
4935 | tmp[sizeof(tmp) - 1] = '\0'; | |
089dd79d | 4936 | goto got_name; |
413ee3b4 | 4937 | } |
089dd79d PZ |
4938 | |
4939 | if (!vma->vm_mm) { | |
4940 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
4941 | goto got_name; | |
3af9e859 EM |
4942 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
4943 | vma->vm_end >= vma->vm_mm->brk) { | |
4944 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
4945 | goto got_name; | |
4946 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
4947 | vma->vm_end >= vma->vm_mm->start_stack) { | |
4948 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
4949 | goto got_name; | |
089dd79d PZ |
4950 | } |
4951 | ||
0a4a9391 PZ |
4952 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
4953 | goto got_name; | |
4954 | } | |
4955 | ||
4956 | got_name: | |
888fcee0 | 4957 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
4958 | |
4959 | mmap_event->file_name = name; | |
4960 | mmap_event->file_size = size; | |
4961 | ||
2fe85427 SE |
4962 | if (!(vma->vm_flags & VM_EXEC)) |
4963 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
4964 | ||
cdd6c482 | 4965 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 4966 | |
52d857a8 JO |
4967 | perf_event_aux(perf_event_mmap_match, |
4968 | perf_event_mmap_output, | |
4969 | mmap_event, | |
4970 | NULL); | |
665c2142 | 4971 | |
0a4a9391 PZ |
4972 | kfree(buf); |
4973 | } | |
4974 | ||
3af9e859 | 4975 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 4976 | { |
9ee318a7 PZ |
4977 | struct perf_mmap_event mmap_event; |
4978 | ||
cdd6c482 | 4979 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
4980 | return; |
4981 | ||
4982 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 4983 | .vma = vma, |
573402db PZ |
4984 | /* .file_name */ |
4985 | /* .file_size */ | |
cdd6c482 | 4986 | .event_id = { |
573402db | 4987 | .header = { |
cdd6c482 | 4988 | .type = PERF_RECORD_MMAP, |
39447b38 | 4989 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
4990 | /* .size */ |
4991 | }, | |
4992 | /* .pid */ | |
4993 | /* .tid */ | |
089dd79d PZ |
4994 | .start = vma->vm_start, |
4995 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 4996 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 PZ |
4997 | }, |
4998 | }; | |
4999 | ||
cdd6c482 | 5000 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5001 | } |
5002 | ||
a78ac325 PZ |
5003 | /* |
5004 | * IRQ throttle logging | |
5005 | */ | |
5006 | ||
cdd6c482 | 5007 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5008 | { |
5009 | struct perf_output_handle handle; | |
c980d109 | 5010 | struct perf_sample_data sample; |
a78ac325 PZ |
5011 | int ret; |
5012 | ||
5013 | struct { | |
5014 | struct perf_event_header header; | |
5015 | u64 time; | |
cca3f454 | 5016 | u64 id; |
7f453c24 | 5017 | u64 stream_id; |
a78ac325 PZ |
5018 | } throttle_event = { |
5019 | .header = { | |
cdd6c482 | 5020 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5021 | .misc = 0, |
5022 | .size = sizeof(throttle_event), | |
5023 | }, | |
def0a9b2 | 5024 | .time = perf_clock(), |
cdd6c482 IM |
5025 | .id = primary_event_id(event), |
5026 | .stream_id = event->id, | |
a78ac325 PZ |
5027 | }; |
5028 | ||
966ee4d6 | 5029 | if (enable) |
cdd6c482 | 5030 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5031 | |
c980d109 ACM |
5032 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5033 | ||
5034 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5035 | throttle_event.header.size); |
a78ac325 PZ |
5036 | if (ret) |
5037 | return; | |
5038 | ||
5039 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5040 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5041 | perf_output_end(&handle); |
5042 | } | |
5043 | ||
f6c7d5fe | 5044 | /* |
cdd6c482 | 5045 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5046 | */ |
5047 | ||
a8b0ca17 | 5048 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5049 | int throttle, struct perf_sample_data *data, |
5050 | struct pt_regs *regs) | |
f6c7d5fe | 5051 | { |
cdd6c482 IM |
5052 | int events = atomic_read(&event->event_limit); |
5053 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5054 | u64 seq; |
79f14641 PZ |
5055 | int ret = 0; |
5056 | ||
96398826 PZ |
5057 | /* |
5058 | * Non-sampling counters might still use the PMI to fold short | |
5059 | * hardware counters, ignore those. | |
5060 | */ | |
5061 | if (unlikely(!is_sampling_event(event))) | |
5062 | return 0; | |
5063 | ||
e050e3f0 SE |
5064 | seq = __this_cpu_read(perf_throttled_seq); |
5065 | if (seq != hwc->interrupts_seq) { | |
5066 | hwc->interrupts_seq = seq; | |
5067 | hwc->interrupts = 1; | |
5068 | } else { | |
5069 | hwc->interrupts++; | |
5070 | if (unlikely(throttle | |
5071 | && hwc->interrupts >= max_samples_per_tick)) { | |
5072 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5073 | hwc->interrupts = MAX_INTERRUPTS; |
5074 | perf_log_throttle(event, 0); | |
a78ac325 PZ |
5075 | ret = 1; |
5076 | } | |
e050e3f0 | 5077 | } |
60db5e09 | 5078 | |
cdd6c482 | 5079 | if (event->attr.freq) { |
def0a9b2 | 5080 | u64 now = perf_clock(); |
abd50713 | 5081 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5082 | |
abd50713 | 5083 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5084 | |
abd50713 | 5085 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5086 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5087 | } |
5088 | ||
2023b359 PZ |
5089 | /* |
5090 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5091 | * events |
2023b359 PZ |
5092 | */ |
5093 | ||
cdd6c482 IM |
5094 | event->pending_kill = POLL_IN; |
5095 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5096 | ret = 1; |
cdd6c482 | 5097 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5098 | event->pending_disable = 1; |
5099 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5100 | } |
5101 | ||
453f19ee | 5102 | if (event->overflow_handler) |
a8b0ca17 | 5103 | event->overflow_handler(event, data, regs); |
453f19ee | 5104 | else |
a8b0ca17 | 5105 | perf_event_output(event, data, regs); |
453f19ee | 5106 | |
f506b3dc | 5107 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5108 | event->pending_wakeup = 1; |
5109 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5110 | } |
5111 | ||
79f14641 | 5112 | return ret; |
f6c7d5fe PZ |
5113 | } |
5114 | ||
a8b0ca17 | 5115 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5116 | struct perf_sample_data *data, |
5117 | struct pt_regs *regs) | |
850bc73f | 5118 | { |
a8b0ca17 | 5119 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5120 | } |
5121 | ||
15dbf27c | 5122 | /* |
cdd6c482 | 5123 | * Generic software event infrastructure |
15dbf27c PZ |
5124 | */ |
5125 | ||
b28ab83c PZ |
5126 | struct swevent_htable { |
5127 | struct swevent_hlist *swevent_hlist; | |
5128 | struct mutex hlist_mutex; | |
5129 | int hlist_refcount; | |
5130 | ||
5131 | /* Recursion avoidance in each contexts */ | |
5132 | int recursion[PERF_NR_CONTEXTS]; | |
5133 | }; | |
5134 | ||
5135 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5136 | ||
7b4b6658 | 5137 | /* |
cdd6c482 IM |
5138 | * We directly increment event->count and keep a second value in |
5139 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5140 | * is kept in the range [-sample_period, 0] so that we can use the |
5141 | * sign as trigger. | |
5142 | */ | |
5143 | ||
ab573844 | 5144 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5145 | { |
cdd6c482 | 5146 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5147 | u64 period = hwc->last_period; |
5148 | u64 nr, offset; | |
5149 | s64 old, val; | |
5150 | ||
5151 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5152 | |
5153 | again: | |
e7850595 | 5154 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5155 | if (val < 0) |
5156 | return 0; | |
15dbf27c | 5157 | |
7b4b6658 PZ |
5158 | nr = div64_u64(period + val, period); |
5159 | offset = nr * period; | |
5160 | val -= offset; | |
e7850595 | 5161 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5162 | goto again; |
15dbf27c | 5163 | |
7b4b6658 | 5164 | return nr; |
15dbf27c PZ |
5165 | } |
5166 | ||
0cff784a | 5167 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5168 | struct perf_sample_data *data, |
5622f295 | 5169 | struct pt_regs *regs) |
15dbf27c | 5170 | { |
cdd6c482 | 5171 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5172 | int throttle = 0; |
15dbf27c | 5173 | |
0cff784a PZ |
5174 | if (!overflow) |
5175 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5176 | |
7b4b6658 PZ |
5177 | if (hwc->interrupts == MAX_INTERRUPTS) |
5178 | return; | |
15dbf27c | 5179 | |
7b4b6658 | 5180 | for (; overflow; overflow--) { |
a8b0ca17 | 5181 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5182 | data, regs)) { |
7b4b6658 PZ |
5183 | /* |
5184 | * We inhibit the overflow from happening when | |
5185 | * hwc->interrupts == MAX_INTERRUPTS. | |
5186 | */ | |
5187 | break; | |
5188 | } | |
cf450a73 | 5189 | throttle = 1; |
7b4b6658 | 5190 | } |
15dbf27c PZ |
5191 | } |
5192 | ||
a4eaf7f1 | 5193 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5194 | struct perf_sample_data *data, |
5622f295 | 5195 | struct pt_regs *regs) |
7b4b6658 | 5196 | { |
cdd6c482 | 5197 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5198 | |
e7850595 | 5199 | local64_add(nr, &event->count); |
d6d020e9 | 5200 | |
0cff784a PZ |
5201 | if (!regs) |
5202 | return; | |
5203 | ||
6c7e550f | 5204 | if (!is_sampling_event(event)) |
7b4b6658 | 5205 | return; |
d6d020e9 | 5206 | |
5d81e5cf AV |
5207 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5208 | data->period = nr; | |
5209 | return perf_swevent_overflow(event, 1, data, regs); | |
5210 | } else | |
5211 | data->period = event->hw.last_period; | |
5212 | ||
0cff784a | 5213 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5214 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5215 | |
e7850595 | 5216 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5217 | return; |
df1a132b | 5218 | |
a8b0ca17 | 5219 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5220 | } |
5221 | ||
f5ffe02e FW |
5222 | static int perf_exclude_event(struct perf_event *event, |
5223 | struct pt_regs *regs) | |
5224 | { | |
a4eaf7f1 | 5225 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5226 | return 1; |
a4eaf7f1 | 5227 | |
f5ffe02e FW |
5228 | if (regs) { |
5229 | if (event->attr.exclude_user && user_mode(regs)) | |
5230 | return 1; | |
5231 | ||
5232 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5233 | return 1; | |
5234 | } | |
5235 | ||
5236 | return 0; | |
5237 | } | |
5238 | ||
cdd6c482 | 5239 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5240 | enum perf_type_id type, |
6fb2915d LZ |
5241 | u32 event_id, |
5242 | struct perf_sample_data *data, | |
5243 | struct pt_regs *regs) | |
15dbf27c | 5244 | { |
cdd6c482 | 5245 | if (event->attr.type != type) |
a21ca2ca | 5246 | return 0; |
f5ffe02e | 5247 | |
cdd6c482 | 5248 | if (event->attr.config != event_id) |
15dbf27c PZ |
5249 | return 0; |
5250 | ||
f5ffe02e FW |
5251 | if (perf_exclude_event(event, regs)) |
5252 | return 0; | |
15dbf27c PZ |
5253 | |
5254 | return 1; | |
5255 | } | |
5256 | ||
76e1d904 FW |
5257 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5258 | { | |
5259 | u64 val = event_id | (type << 32); | |
5260 | ||
5261 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5262 | } | |
5263 | ||
49f135ed FW |
5264 | static inline struct hlist_head * |
5265 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5266 | { |
49f135ed FW |
5267 | u64 hash = swevent_hash(type, event_id); |
5268 | ||
5269 | return &hlist->heads[hash]; | |
5270 | } | |
76e1d904 | 5271 | |
49f135ed FW |
5272 | /* For the read side: events when they trigger */ |
5273 | static inline struct hlist_head * | |
b28ab83c | 5274 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5275 | { |
5276 | struct swevent_hlist *hlist; | |
76e1d904 | 5277 | |
b28ab83c | 5278 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5279 | if (!hlist) |
5280 | return NULL; | |
5281 | ||
49f135ed FW |
5282 | return __find_swevent_head(hlist, type, event_id); |
5283 | } | |
5284 | ||
5285 | /* For the event head insertion and removal in the hlist */ | |
5286 | static inline struct hlist_head * | |
b28ab83c | 5287 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5288 | { |
5289 | struct swevent_hlist *hlist; | |
5290 | u32 event_id = event->attr.config; | |
5291 | u64 type = event->attr.type; | |
5292 | ||
5293 | /* | |
5294 | * Event scheduling is always serialized against hlist allocation | |
5295 | * and release. Which makes the protected version suitable here. | |
5296 | * The context lock guarantees that. | |
5297 | */ | |
b28ab83c | 5298 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5299 | lockdep_is_held(&event->ctx->lock)); |
5300 | if (!hlist) | |
5301 | return NULL; | |
5302 | ||
5303 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5304 | } |
5305 | ||
5306 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5307 | u64 nr, |
76e1d904 FW |
5308 | struct perf_sample_data *data, |
5309 | struct pt_regs *regs) | |
15dbf27c | 5310 | { |
b28ab83c | 5311 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5312 | struct perf_event *event; |
76e1d904 | 5313 | struct hlist_head *head; |
15dbf27c | 5314 | |
76e1d904 | 5315 | rcu_read_lock(); |
b28ab83c | 5316 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5317 | if (!head) |
5318 | goto end; | |
5319 | ||
b67bfe0d | 5320 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5321 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5322 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5323 | } |
76e1d904 FW |
5324 | end: |
5325 | rcu_read_unlock(); | |
15dbf27c PZ |
5326 | } |
5327 | ||
4ed7c92d | 5328 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5329 | { |
b28ab83c | 5330 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5331 | |
b28ab83c | 5332 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5333 | } |
645e8cc0 | 5334 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5335 | |
fa9f90be | 5336 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5337 | { |
b28ab83c | 5338 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5339 | |
b28ab83c | 5340 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5341 | } |
15dbf27c | 5342 | |
a8b0ca17 | 5343 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5344 | { |
a4234bfc | 5345 | struct perf_sample_data data; |
4ed7c92d PZ |
5346 | int rctx; |
5347 | ||
1c024eca | 5348 | preempt_disable_notrace(); |
4ed7c92d PZ |
5349 | rctx = perf_swevent_get_recursion_context(); |
5350 | if (rctx < 0) | |
5351 | return; | |
a4234bfc | 5352 | |
fd0d000b | 5353 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5354 | |
a8b0ca17 | 5355 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5356 | |
5357 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5358 | preempt_enable_notrace(); |
b8e83514 PZ |
5359 | } |
5360 | ||
cdd6c482 | 5361 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5362 | { |
15dbf27c PZ |
5363 | } |
5364 | ||
a4eaf7f1 | 5365 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5366 | { |
b28ab83c | 5367 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5368 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5369 | struct hlist_head *head; |
5370 | ||
6c7e550f | 5371 | if (is_sampling_event(event)) { |
7b4b6658 | 5372 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5373 | perf_swevent_set_period(event); |
7b4b6658 | 5374 | } |
76e1d904 | 5375 | |
a4eaf7f1 PZ |
5376 | hwc->state = !(flags & PERF_EF_START); |
5377 | ||
b28ab83c | 5378 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
5379 | if (WARN_ON_ONCE(!head)) |
5380 | return -EINVAL; | |
5381 | ||
5382 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5383 | ||
15dbf27c PZ |
5384 | return 0; |
5385 | } | |
5386 | ||
a4eaf7f1 | 5387 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5388 | { |
76e1d904 | 5389 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5390 | } |
5391 | ||
a4eaf7f1 | 5392 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5393 | { |
a4eaf7f1 | 5394 | event->hw.state = 0; |
d6d020e9 | 5395 | } |
aa9c4c0f | 5396 | |
a4eaf7f1 | 5397 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5398 | { |
a4eaf7f1 | 5399 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5400 | } |
5401 | ||
49f135ed FW |
5402 | /* Deref the hlist from the update side */ |
5403 | static inline struct swevent_hlist * | |
b28ab83c | 5404 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5405 | { |
b28ab83c PZ |
5406 | return rcu_dereference_protected(swhash->swevent_hlist, |
5407 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5408 | } |
5409 | ||
b28ab83c | 5410 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5411 | { |
b28ab83c | 5412 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5413 | |
49f135ed | 5414 | if (!hlist) |
76e1d904 FW |
5415 | return; |
5416 | ||
b28ab83c | 5417 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5418 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5419 | } |
5420 | ||
5421 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5422 | { | |
b28ab83c | 5423 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5424 | |
b28ab83c | 5425 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5426 | |
b28ab83c PZ |
5427 | if (!--swhash->hlist_refcount) |
5428 | swevent_hlist_release(swhash); | |
76e1d904 | 5429 | |
b28ab83c | 5430 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5431 | } |
5432 | ||
5433 | static void swevent_hlist_put(struct perf_event *event) | |
5434 | { | |
5435 | int cpu; | |
5436 | ||
5437 | if (event->cpu != -1) { | |
5438 | swevent_hlist_put_cpu(event, event->cpu); | |
5439 | return; | |
5440 | } | |
5441 | ||
5442 | for_each_possible_cpu(cpu) | |
5443 | swevent_hlist_put_cpu(event, cpu); | |
5444 | } | |
5445 | ||
5446 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5447 | { | |
b28ab83c | 5448 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5449 | int err = 0; |
5450 | ||
b28ab83c | 5451 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5452 | |
b28ab83c | 5453 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5454 | struct swevent_hlist *hlist; |
5455 | ||
5456 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5457 | if (!hlist) { | |
5458 | err = -ENOMEM; | |
5459 | goto exit; | |
5460 | } | |
b28ab83c | 5461 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5462 | } |
b28ab83c | 5463 | swhash->hlist_refcount++; |
9ed6060d | 5464 | exit: |
b28ab83c | 5465 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5466 | |
5467 | return err; | |
5468 | } | |
5469 | ||
5470 | static int swevent_hlist_get(struct perf_event *event) | |
5471 | { | |
5472 | int err; | |
5473 | int cpu, failed_cpu; | |
5474 | ||
5475 | if (event->cpu != -1) | |
5476 | return swevent_hlist_get_cpu(event, event->cpu); | |
5477 | ||
5478 | get_online_cpus(); | |
5479 | for_each_possible_cpu(cpu) { | |
5480 | err = swevent_hlist_get_cpu(event, cpu); | |
5481 | if (err) { | |
5482 | failed_cpu = cpu; | |
5483 | goto fail; | |
5484 | } | |
5485 | } | |
5486 | put_online_cpus(); | |
5487 | ||
5488 | return 0; | |
9ed6060d | 5489 | fail: |
76e1d904 FW |
5490 | for_each_possible_cpu(cpu) { |
5491 | if (cpu == failed_cpu) | |
5492 | break; | |
5493 | swevent_hlist_put_cpu(event, cpu); | |
5494 | } | |
5495 | ||
5496 | put_online_cpus(); | |
5497 | return err; | |
5498 | } | |
5499 | ||
c5905afb | 5500 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5501 | |
b0a873eb PZ |
5502 | static void sw_perf_event_destroy(struct perf_event *event) |
5503 | { | |
5504 | u64 event_id = event->attr.config; | |
95476b64 | 5505 | |
b0a873eb PZ |
5506 | WARN_ON(event->parent); |
5507 | ||
c5905afb | 5508 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5509 | swevent_hlist_put(event); |
5510 | } | |
5511 | ||
5512 | static int perf_swevent_init(struct perf_event *event) | |
5513 | { | |
8176cced | 5514 | u64 event_id = event->attr.config; |
b0a873eb PZ |
5515 | |
5516 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5517 | return -ENOENT; | |
5518 | ||
2481c5fa SE |
5519 | /* |
5520 | * no branch sampling for software events | |
5521 | */ | |
5522 | if (has_branch_stack(event)) | |
5523 | return -EOPNOTSUPP; | |
5524 | ||
b0a873eb PZ |
5525 | switch (event_id) { |
5526 | case PERF_COUNT_SW_CPU_CLOCK: | |
5527 | case PERF_COUNT_SW_TASK_CLOCK: | |
5528 | return -ENOENT; | |
5529 | ||
5530 | default: | |
5531 | break; | |
5532 | } | |
5533 | ||
ce677831 | 5534 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5535 | return -ENOENT; |
5536 | ||
5537 | if (!event->parent) { | |
5538 | int err; | |
5539 | ||
5540 | err = swevent_hlist_get(event); | |
5541 | if (err) | |
5542 | return err; | |
5543 | ||
c5905afb | 5544 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5545 | event->destroy = sw_perf_event_destroy; |
5546 | } | |
5547 | ||
5548 | return 0; | |
5549 | } | |
5550 | ||
35edc2a5 PZ |
5551 | static int perf_swevent_event_idx(struct perf_event *event) |
5552 | { | |
5553 | return 0; | |
5554 | } | |
5555 | ||
b0a873eb | 5556 | static struct pmu perf_swevent = { |
89a1e187 | 5557 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5558 | |
b0a873eb | 5559 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5560 | .add = perf_swevent_add, |
5561 | .del = perf_swevent_del, | |
5562 | .start = perf_swevent_start, | |
5563 | .stop = perf_swevent_stop, | |
1c024eca | 5564 | .read = perf_swevent_read, |
35edc2a5 PZ |
5565 | |
5566 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5567 | }; |
5568 | ||
b0a873eb PZ |
5569 | #ifdef CONFIG_EVENT_TRACING |
5570 | ||
1c024eca PZ |
5571 | static int perf_tp_filter_match(struct perf_event *event, |
5572 | struct perf_sample_data *data) | |
5573 | { | |
5574 | void *record = data->raw->data; | |
5575 | ||
5576 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5577 | return 1; | |
5578 | return 0; | |
5579 | } | |
5580 | ||
5581 | static int perf_tp_event_match(struct perf_event *event, | |
5582 | struct perf_sample_data *data, | |
5583 | struct pt_regs *regs) | |
5584 | { | |
a0f7d0f7 FW |
5585 | if (event->hw.state & PERF_HES_STOPPED) |
5586 | return 0; | |
580d607c PZ |
5587 | /* |
5588 | * All tracepoints are from kernel-space. | |
5589 | */ | |
5590 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5591 | return 0; |
5592 | ||
5593 | if (!perf_tp_filter_match(event, data)) | |
5594 | return 0; | |
5595 | ||
5596 | return 1; | |
5597 | } | |
5598 | ||
5599 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5600 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5601 | struct task_struct *task) | |
95476b64 FW |
5602 | { |
5603 | struct perf_sample_data data; | |
1c024eca | 5604 | struct perf_event *event; |
1c024eca | 5605 | |
95476b64 FW |
5606 | struct perf_raw_record raw = { |
5607 | .size = entry_size, | |
5608 | .data = record, | |
5609 | }; | |
5610 | ||
fd0d000b | 5611 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5612 | data.raw = &raw; |
5613 | ||
b67bfe0d | 5614 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 5615 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 5616 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5617 | } |
ecc55f84 | 5618 | |
e6dab5ff AV |
5619 | /* |
5620 | * If we got specified a target task, also iterate its context and | |
5621 | * deliver this event there too. | |
5622 | */ | |
5623 | if (task && task != current) { | |
5624 | struct perf_event_context *ctx; | |
5625 | struct trace_entry *entry = record; | |
5626 | ||
5627 | rcu_read_lock(); | |
5628 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5629 | if (!ctx) | |
5630 | goto unlock; | |
5631 | ||
5632 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5633 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5634 | continue; | |
5635 | if (event->attr.config != entry->type) | |
5636 | continue; | |
5637 | if (perf_tp_event_match(event, &data, regs)) | |
5638 | perf_swevent_event(event, count, &data, regs); | |
5639 | } | |
5640 | unlock: | |
5641 | rcu_read_unlock(); | |
5642 | } | |
5643 | ||
ecc55f84 | 5644 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5645 | } |
5646 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5647 | ||
cdd6c482 | 5648 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5649 | { |
1c024eca | 5650 | perf_trace_destroy(event); |
e077df4f PZ |
5651 | } |
5652 | ||
b0a873eb | 5653 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5654 | { |
76e1d904 FW |
5655 | int err; |
5656 | ||
b0a873eb PZ |
5657 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5658 | return -ENOENT; | |
5659 | ||
2481c5fa SE |
5660 | /* |
5661 | * no branch sampling for tracepoint events | |
5662 | */ | |
5663 | if (has_branch_stack(event)) | |
5664 | return -EOPNOTSUPP; | |
5665 | ||
1c024eca PZ |
5666 | err = perf_trace_init(event); |
5667 | if (err) | |
b0a873eb | 5668 | return err; |
e077df4f | 5669 | |
cdd6c482 | 5670 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5671 | |
b0a873eb PZ |
5672 | return 0; |
5673 | } | |
5674 | ||
5675 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5676 | .task_ctx_nr = perf_sw_context, |
5677 | ||
b0a873eb | 5678 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5679 | .add = perf_trace_add, |
5680 | .del = perf_trace_del, | |
5681 | .start = perf_swevent_start, | |
5682 | .stop = perf_swevent_stop, | |
b0a873eb | 5683 | .read = perf_swevent_read, |
35edc2a5 PZ |
5684 | |
5685 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5686 | }; |
5687 | ||
5688 | static inline void perf_tp_register(void) | |
5689 | { | |
2e80a82a | 5690 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5691 | } |
6fb2915d LZ |
5692 | |
5693 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5694 | { | |
5695 | char *filter_str; | |
5696 | int ret; | |
5697 | ||
5698 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5699 | return -EINVAL; | |
5700 | ||
5701 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5702 | if (IS_ERR(filter_str)) | |
5703 | return PTR_ERR(filter_str); | |
5704 | ||
5705 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5706 | ||
5707 | kfree(filter_str); | |
5708 | return ret; | |
5709 | } | |
5710 | ||
5711 | static void perf_event_free_filter(struct perf_event *event) | |
5712 | { | |
5713 | ftrace_profile_free_filter(event); | |
5714 | } | |
5715 | ||
e077df4f | 5716 | #else |
6fb2915d | 5717 | |
b0a873eb | 5718 | static inline void perf_tp_register(void) |
e077df4f | 5719 | { |
e077df4f | 5720 | } |
6fb2915d LZ |
5721 | |
5722 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5723 | { | |
5724 | return -ENOENT; | |
5725 | } | |
5726 | ||
5727 | static void perf_event_free_filter(struct perf_event *event) | |
5728 | { | |
5729 | } | |
5730 | ||
07b139c8 | 5731 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5732 | |
24f1e32c | 5733 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5734 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5735 | { |
f5ffe02e FW |
5736 | struct perf_sample_data sample; |
5737 | struct pt_regs *regs = data; | |
5738 | ||
fd0d000b | 5739 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 5740 | |
a4eaf7f1 | 5741 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5742 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5743 | } |
5744 | #endif | |
5745 | ||
b0a873eb PZ |
5746 | /* |
5747 | * hrtimer based swevent callback | |
5748 | */ | |
f29ac756 | 5749 | |
b0a873eb | 5750 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5751 | { |
b0a873eb PZ |
5752 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5753 | struct perf_sample_data data; | |
5754 | struct pt_regs *regs; | |
5755 | struct perf_event *event; | |
5756 | u64 period; | |
f29ac756 | 5757 | |
b0a873eb | 5758 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5759 | |
5760 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5761 | return HRTIMER_NORESTART; | |
5762 | ||
b0a873eb | 5763 | event->pmu->read(event); |
f344011c | 5764 | |
fd0d000b | 5765 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
5766 | regs = get_irq_regs(); |
5767 | ||
5768 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5769 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 5770 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
5771 | ret = HRTIMER_NORESTART; |
5772 | } | |
24f1e32c | 5773 | |
b0a873eb PZ |
5774 | period = max_t(u64, 10000, event->hw.sample_period); |
5775 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5776 | |
b0a873eb | 5777 | return ret; |
f29ac756 PZ |
5778 | } |
5779 | ||
b0a873eb | 5780 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5781 | { |
b0a873eb | 5782 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5783 | s64 period; |
5784 | ||
5785 | if (!is_sampling_event(event)) | |
5786 | return; | |
f5ffe02e | 5787 | |
5d508e82 FBH |
5788 | period = local64_read(&hwc->period_left); |
5789 | if (period) { | |
5790 | if (period < 0) | |
5791 | period = 10000; | |
fa407f35 | 5792 | |
5d508e82 FBH |
5793 | local64_set(&hwc->period_left, 0); |
5794 | } else { | |
5795 | period = max_t(u64, 10000, hwc->sample_period); | |
5796 | } | |
5797 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 5798 | ns_to_ktime(period), 0, |
b5ab4cd5 | 5799 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 5800 | } |
b0a873eb PZ |
5801 | |
5802 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 5803 | { |
b0a873eb PZ |
5804 | struct hw_perf_event *hwc = &event->hw; |
5805 | ||
6c7e550f | 5806 | if (is_sampling_event(event)) { |
b0a873eb | 5807 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 5808 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
5809 | |
5810 | hrtimer_cancel(&hwc->hrtimer); | |
5811 | } | |
24f1e32c FW |
5812 | } |
5813 | ||
ba3dd36c PZ |
5814 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5815 | { | |
5816 | struct hw_perf_event *hwc = &event->hw; | |
5817 | ||
5818 | if (!is_sampling_event(event)) | |
5819 | return; | |
5820 | ||
5821 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5822 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
5823 | ||
5824 | /* | |
5825 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
5826 | * mapping and avoid the whole period adjust feedback stuff. | |
5827 | */ | |
5828 | if (event->attr.freq) { | |
5829 | long freq = event->attr.sample_freq; | |
5830 | ||
5831 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
5832 | hwc->sample_period = event->attr.sample_period; | |
5833 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 5834 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
5835 | event->attr.freq = 0; |
5836 | } | |
5837 | } | |
5838 | ||
b0a873eb PZ |
5839 | /* |
5840 | * Software event: cpu wall time clock | |
5841 | */ | |
5842 | ||
5843 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 5844 | { |
b0a873eb PZ |
5845 | s64 prev; |
5846 | u64 now; | |
5847 | ||
a4eaf7f1 | 5848 | now = local_clock(); |
b0a873eb PZ |
5849 | prev = local64_xchg(&event->hw.prev_count, now); |
5850 | local64_add(now - prev, &event->count); | |
24f1e32c | 5851 | } |
24f1e32c | 5852 | |
a4eaf7f1 | 5853 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5854 | { |
a4eaf7f1 | 5855 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 5856 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5857 | } |
5858 | ||
a4eaf7f1 | 5859 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 5860 | { |
b0a873eb PZ |
5861 | perf_swevent_cancel_hrtimer(event); |
5862 | cpu_clock_event_update(event); | |
5863 | } | |
f29ac756 | 5864 | |
a4eaf7f1 PZ |
5865 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
5866 | { | |
5867 | if (flags & PERF_EF_START) | |
5868 | cpu_clock_event_start(event, flags); | |
5869 | ||
5870 | return 0; | |
5871 | } | |
5872 | ||
5873 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
5874 | { | |
5875 | cpu_clock_event_stop(event, flags); | |
5876 | } | |
5877 | ||
b0a873eb PZ |
5878 | static void cpu_clock_event_read(struct perf_event *event) |
5879 | { | |
5880 | cpu_clock_event_update(event); | |
5881 | } | |
f344011c | 5882 | |
b0a873eb PZ |
5883 | static int cpu_clock_event_init(struct perf_event *event) |
5884 | { | |
5885 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5886 | return -ENOENT; | |
5887 | ||
5888 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
5889 | return -ENOENT; | |
5890 | ||
2481c5fa SE |
5891 | /* |
5892 | * no branch sampling for software events | |
5893 | */ | |
5894 | if (has_branch_stack(event)) | |
5895 | return -EOPNOTSUPP; | |
5896 | ||
ba3dd36c PZ |
5897 | perf_swevent_init_hrtimer(event); |
5898 | ||
b0a873eb | 5899 | return 0; |
f29ac756 PZ |
5900 | } |
5901 | ||
b0a873eb | 5902 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
5903 | .task_ctx_nr = perf_sw_context, |
5904 | ||
b0a873eb | 5905 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
5906 | .add = cpu_clock_event_add, |
5907 | .del = cpu_clock_event_del, | |
5908 | .start = cpu_clock_event_start, | |
5909 | .stop = cpu_clock_event_stop, | |
b0a873eb | 5910 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
5911 | |
5912 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5913 | }; |
5914 | ||
5915 | /* | |
5916 | * Software event: task time clock | |
5917 | */ | |
5918 | ||
5919 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 5920 | { |
b0a873eb PZ |
5921 | u64 prev; |
5922 | s64 delta; | |
5c92d124 | 5923 | |
b0a873eb PZ |
5924 | prev = local64_xchg(&event->hw.prev_count, now); |
5925 | delta = now - prev; | |
5926 | local64_add(delta, &event->count); | |
5927 | } | |
5c92d124 | 5928 | |
a4eaf7f1 | 5929 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5930 | { |
a4eaf7f1 | 5931 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 5932 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5933 | } |
5934 | ||
a4eaf7f1 | 5935 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
5936 | { |
5937 | perf_swevent_cancel_hrtimer(event); | |
5938 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
5939 | } |
5940 | ||
5941 | static int task_clock_event_add(struct perf_event *event, int flags) | |
5942 | { | |
5943 | if (flags & PERF_EF_START) | |
5944 | task_clock_event_start(event, flags); | |
b0a873eb | 5945 | |
a4eaf7f1 PZ |
5946 | return 0; |
5947 | } | |
5948 | ||
5949 | static void task_clock_event_del(struct perf_event *event, int flags) | |
5950 | { | |
5951 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
5952 | } |
5953 | ||
5954 | static void task_clock_event_read(struct perf_event *event) | |
5955 | { | |
768a06e2 PZ |
5956 | u64 now = perf_clock(); |
5957 | u64 delta = now - event->ctx->timestamp; | |
5958 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
5959 | |
5960 | task_clock_event_update(event, time); | |
5961 | } | |
5962 | ||
5963 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 5964 | { |
b0a873eb PZ |
5965 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
5966 | return -ENOENT; | |
5967 | ||
5968 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
5969 | return -ENOENT; | |
5970 | ||
2481c5fa SE |
5971 | /* |
5972 | * no branch sampling for software events | |
5973 | */ | |
5974 | if (has_branch_stack(event)) | |
5975 | return -EOPNOTSUPP; | |
5976 | ||
ba3dd36c PZ |
5977 | perf_swevent_init_hrtimer(event); |
5978 | ||
b0a873eb | 5979 | return 0; |
6fb2915d LZ |
5980 | } |
5981 | ||
b0a873eb | 5982 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
5983 | .task_ctx_nr = perf_sw_context, |
5984 | ||
b0a873eb | 5985 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
5986 | .add = task_clock_event_add, |
5987 | .del = task_clock_event_del, | |
5988 | .start = task_clock_event_start, | |
5989 | .stop = task_clock_event_stop, | |
b0a873eb | 5990 | .read = task_clock_event_read, |
35edc2a5 PZ |
5991 | |
5992 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 5993 | }; |
6fb2915d | 5994 | |
ad5133b7 | 5995 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 5996 | { |
e077df4f | 5997 | } |
6fb2915d | 5998 | |
ad5133b7 | 5999 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6000 | { |
ad5133b7 | 6001 | return 0; |
6fb2915d LZ |
6002 | } |
6003 | ||
ad5133b7 | 6004 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6005 | { |
ad5133b7 | 6006 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6007 | } |
6008 | ||
ad5133b7 PZ |
6009 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6010 | { | |
6011 | perf_pmu_enable(pmu); | |
6012 | return 0; | |
6013 | } | |
e077df4f | 6014 | |
ad5133b7 | 6015 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6016 | { |
ad5133b7 | 6017 | perf_pmu_enable(pmu); |
24f1e32c FW |
6018 | } |
6019 | ||
35edc2a5 PZ |
6020 | static int perf_event_idx_default(struct perf_event *event) |
6021 | { | |
6022 | return event->hw.idx + 1; | |
6023 | } | |
6024 | ||
8dc85d54 PZ |
6025 | /* |
6026 | * Ensures all contexts with the same task_ctx_nr have the same | |
6027 | * pmu_cpu_context too. | |
6028 | */ | |
6029 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 6030 | { |
8dc85d54 | 6031 | struct pmu *pmu; |
b326e956 | 6032 | |
8dc85d54 PZ |
6033 | if (ctxn < 0) |
6034 | return NULL; | |
24f1e32c | 6035 | |
8dc85d54 PZ |
6036 | list_for_each_entry(pmu, &pmus, entry) { |
6037 | if (pmu->task_ctx_nr == ctxn) | |
6038 | return pmu->pmu_cpu_context; | |
6039 | } | |
24f1e32c | 6040 | |
8dc85d54 | 6041 | return NULL; |
24f1e32c FW |
6042 | } |
6043 | ||
51676957 | 6044 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6045 | { |
51676957 PZ |
6046 | int cpu; |
6047 | ||
6048 | for_each_possible_cpu(cpu) { | |
6049 | struct perf_cpu_context *cpuctx; | |
6050 | ||
6051 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6052 | ||
3f1f3320 PZ |
6053 | if (cpuctx->unique_pmu == old_pmu) |
6054 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6055 | } |
6056 | } | |
6057 | ||
6058 | static void free_pmu_context(struct pmu *pmu) | |
6059 | { | |
6060 | struct pmu *i; | |
f5ffe02e | 6061 | |
8dc85d54 | 6062 | mutex_lock(&pmus_lock); |
0475f9ea | 6063 | /* |
8dc85d54 | 6064 | * Like a real lame refcount. |
0475f9ea | 6065 | */ |
51676957 PZ |
6066 | list_for_each_entry(i, &pmus, entry) { |
6067 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6068 | update_pmu_context(i, pmu); | |
8dc85d54 | 6069 | goto out; |
51676957 | 6070 | } |
8dc85d54 | 6071 | } |
d6d020e9 | 6072 | |
51676957 | 6073 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6074 | out: |
6075 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6076 | } |
2e80a82a | 6077 | static struct idr pmu_idr; |
d6d020e9 | 6078 | |
abe43400 PZ |
6079 | static ssize_t |
6080 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6081 | { | |
6082 | struct pmu *pmu = dev_get_drvdata(dev); | |
6083 | ||
6084 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6085 | } | |
6086 | ||
62b85639 SE |
6087 | static ssize_t |
6088 | perf_event_mux_interval_ms_show(struct device *dev, | |
6089 | struct device_attribute *attr, | |
6090 | char *page) | |
6091 | { | |
6092 | struct pmu *pmu = dev_get_drvdata(dev); | |
6093 | ||
6094 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6095 | } | |
6096 | ||
6097 | static ssize_t | |
6098 | perf_event_mux_interval_ms_store(struct device *dev, | |
6099 | struct device_attribute *attr, | |
6100 | const char *buf, size_t count) | |
6101 | { | |
6102 | struct pmu *pmu = dev_get_drvdata(dev); | |
6103 | int timer, cpu, ret; | |
6104 | ||
6105 | ret = kstrtoint(buf, 0, &timer); | |
6106 | if (ret) | |
6107 | return ret; | |
6108 | ||
6109 | if (timer < 1) | |
6110 | return -EINVAL; | |
6111 | ||
6112 | /* same value, noting to do */ | |
6113 | if (timer == pmu->hrtimer_interval_ms) | |
6114 | return count; | |
6115 | ||
6116 | pmu->hrtimer_interval_ms = timer; | |
6117 | ||
6118 | /* update all cpuctx for this PMU */ | |
6119 | for_each_possible_cpu(cpu) { | |
6120 | struct perf_cpu_context *cpuctx; | |
6121 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6122 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6123 | ||
6124 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6125 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6126 | } | |
6127 | ||
6128 | return count; | |
6129 | } | |
6130 | ||
6131 | #define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store) | |
6132 | ||
abe43400 | 6133 | static struct device_attribute pmu_dev_attrs[] = { |
62b85639 SE |
6134 | __ATTR_RO(type), |
6135 | __ATTR_RW(perf_event_mux_interval_ms), | |
6136 | __ATTR_NULL, | |
abe43400 PZ |
6137 | }; |
6138 | ||
6139 | static int pmu_bus_running; | |
6140 | static struct bus_type pmu_bus = { | |
6141 | .name = "event_source", | |
6142 | .dev_attrs = pmu_dev_attrs, | |
6143 | }; | |
6144 | ||
6145 | static void pmu_dev_release(struct device *dev) | |
6146 | { | |
6147 | kfree(dev); | |
6148 | } | |
6149 | ||
6150 | static int pmu_dev_alloc(struct pmu *pmu) | |
6151 | { | |
6152 | int ret = -ENOMEM; | |
6153 | ||
6154 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6155 | if (!pmu->dev) | |
6156 | goto out; | |
6157 | ||
0c9d42ed | 6158 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6159 | device_initialize(pmu->dev); |
6160 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6161 | if (ret) | |
6162 | goto free_dev; | |
6163 | ||
6164 | dev_set_drvdata(pmu->dev, pmu); | |
6165 | pmu->dev->bus = &pmu_bus; | |
6166 | pmu->dev->release = pmu_dev_release; | |
6167 | ret = device_add(pmu->dev); | |
6168 | if (ret) | |
6169 | goto free_dev; | |
6170 | ||
6171 | out: | |
6172 | return ret; | |
6173 | ||
6174 | free_dev: | |
6175 | put_device(pmu->dev); | |
6176 | goto out; | |
6177 | } | |
6178 | ||
547e9fd7 | 6179 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6180 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6181 | |
03d8e80b | 6182 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6183 | { |
108b02cf | 6184 | int cpu, ret; |
24f1e32c | 6185 | |
b0a873eb | 6186 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6187 | ret = -ENOMEM; |
6188 | pmu->pmu_disable_count = alloc_percpu(int); | |
6189 | if (!pmu->pmu_disable_count) | |
6190 | goto unlock; | |
f29ac756 | 6191 | |
2e80a82a PZ |
6192 | pmu->type = -1; |
6193 | if (!name) | |
6194 | goto skip_type; | |
6195 | pmu->name = name; | |
6196 | ||
6197 | if (type < 0) { | |
0e9c3be2 TH |
6198 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6199 | if (type < 0) { | |
6200 | ret = type; | |
2e80a82a PZ |
6201 | goto free_pdc; |
6202 | } | |
6203 | } | |
6204 | pmu->type = type; | |
6205 | ||
abe43400 PZ |
6206 | if (pmu_bus_running) { |
6207 | ret = pmu_dev_alloc(pmu); | |
6208 | if (ret) | |
6209 | goto free_idr; | |
6210 | } | |
6211 | ||
2e80a82a | 6212 | skip_type: |
8dc85d54 PZ |
6213 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6214 | if (pmu->pmu_cpu_context) | |
6215 | goto got_cpu_context; | |
f29ac756 | 6216 | |
c4814202 | 6217 | ret = -ENOMEM; |
108b02cf PZ |
6218 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6219 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6220 | goto free_dev; |
f344011c | 6221 | |
108b02cf PZ |
6222 | for_each_possible_cpu(cpu) { |
6223 | struct perf_cpu_context *cpuctx; | |
6224 | ||
6225 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6226 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6227 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6228 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6229 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6230 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6231 | |
6232 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6233 | ||
e9d2b064 | 6234 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
3f1f3320 | 6235 | cpuctx->unique_pmu = pmu; |
108b02cf | 6236 | } |
76e1d904 | 6237 | |
8dc85d54 | 6238 | got_cpu_context: |
ad5133b7 PZ |
6239 | if (!pmu->start_txn) { |
6240 | if (pmu->pmu_enable) { | |
6241 | /* | |
6242 | * If we have pmu_enable/pmu_disable calls, install | |
6243 | * transaction stubs that use that to try and batch | |
6244 | * hardware accesses. | |
6245 | */ | |
6246 | pmu->start_txn = perf_pmu_start_txn; | |
6247 | pmu->commit_txn = perf_pmu_commit_txn; | |
6248 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6249 | } else { | |
6250 | pmu->start_txn = perf_pmu_nop_void; | |
6251 | pmu->commit_txn = perf_pmu_nop_int; | |
6252 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6253 | } |
5c92d124 | 6254 | } |
15dbf27c | 6255 | |
ad5133b7 PZ |
6256 | if (!pmu->pmu_enable) { |
6257 | pmu->pmu_enable = perf_pmu_nop_void; | |
6258 | pmu->pmu_disable = perf_pmu_nop_void; | |
6259 | } | |
6260 | ||
35edc2a5 PZ |
6261 | if (!pmu->event_idx) |
6262 | pmu->event_idx = perf_event_idx_default; | |
6263 | ||
b0a873eb | 6264 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6265 | ret = 0; |
6266 | unlock: | |
b0a873eb PZ |
6267 | mutex_unlock(&pmus_lock); |
6268 | ||
33696fc0 | 6269 | return ret; |
108b02cf | 6270 | |
abe43400 PZ |
6271 | free_dev: |
6272 | device_del(pmu->dev); | |
6273 | put_device(pmu->dev); | |
6274 | ||
2e80a82a PZ |
6275 | free_idr: |
6276 | if (pmu->type >= PERF_TYPE_MAX) | |
6277 | idr_remove(&pmu_idr, pmu->type); | |
6278 | ||
108b02cf PZ |
6279 | free_pdc: |
6280 | free_percpu(pmu->pmu_disable_count); | |
6281 | goto unlock; | |
f29ac756 PZ |
6282 | } |
6283 | ||
b0a873eb | 6284 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6285 | { |
b0a873eb PZ |
6286 | mutex_lock(&pmus_lock); |
6287 | list_del_rcu(&pmu->entry); | |
6288 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6289 | |
0475f9ea | 6290 | /* |
cde8e884 PZ |
6291 | * We dereference the pmu list under both SRCU and regular RCU, so |
6292 | * synchronize against both of those. | |
0475f9ea | 6293 | */ |
b0a873eb | 6294 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6295 | synchronize_rcu(); |
d6d020e9 | 6296 | |
33696fc0 | 6297 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6298 | if (pmu->type >= PERF_TYPE_MAX) |
6299 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6300 | device_del(pmu->dev); |
6301 | put_device(pmu->dev); | |
51676957 | 6302 | free_pmu_context(pmu); |
b0a873eb | 6303 | } |
d6d020e9 | 6304 | |
b0a873eb PZ |
6305 | struct pmu *perf_init_event(struct perf_event *event) |
6306 | { | |
6307 | struct pmu *pmu = NULL; | |
6308 | int idx; | |
940c5b29 | 6309 | int ret; |
b0a873eb PZ |
6310 | |
6311 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6312 | |
6313 | rcu_read_lock(); | |
6314 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6315 | rcu_read_unlock(); | |
940c5b29 | 6316 | if (pmu) { |
7e5b2a01 | 6317 | event->pmu = pmu; |
940c5b29 LM |
6318 | ret = pmu->event_init(event); |
6319 | if (ret) | |
6320 | pmu = ERR_PTR(ret); | |
2e80a82a | 6321 | goto unlock; |
940c5b29 | 6322 | } |
2e80a82a | 6323 | |
b0a873eb | 6324 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 6325 | event->pmu = pmu; |
940c5b29 | 6326 | ret = pmu->event_init(event); |
b0a873eb | 6327 | if (!ret) |
e5f4d339 | 6328 | goto unlock; |
76e1d904 | 6329 | |
b0a873eb PZ |
6330 | if (ret != -ENOENT) { |
6331 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6332 | goto unlock; |
f344011c | 6333 | } |
5c92d124 | 6334 | } |
e5f4d339 PZ |
6335 | pmu = ERR_PTR(-ENOENT); |
6336 | unlock: | |
b0a873eb | 6337 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6338 | |
4aeb0b42 | 6339 | return pmu; |
5c92d124 IM |
6340 | } |
6341 | ||
0793a61d | 6342 | /* |
cdd6c482 | 6343 | * Allocate and initialize a event structure |
0793a61d | 6344 | */ |
cdd6c482 | 6345 | static struct perf_event * |
c3f00c70 | 6346 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6347 | struct task_struct *task, |
6348 | struct perf_event *group_leader, | |
6349 | struct perf_event *parent_event, | |
4dc0da86 AK |
6350 | perf_overflow_handler_t overflow_handler, |
6351 | void *context) | |
0793a61d | 6352 | { |
51b0fe39 | 6353 | struct pmu *pmu; |
cdd6c482 IM |
6354 | struct perf_event *event; |
6355 | struct hw_perf_event *hwc; | |
d5d2bc0d | 6356 | long err; |
0793a61d | 6357 | |
66832eb4 ON |
6358 | if ((unsigned)cpu >= nr_cpu_ids) { |
6359 | if (!task || cpu != -1) | |
6360 | return ERR_PTR(-EINVAL); | |
6361 | } | |
6362 | ||
c3f00c70 | 6363 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6364 | if (!event) |
d5d2bc0d | 6365 | return ERR_PTR(-ENOMEM); |
0793a61d | 6366 | |
04289bb9 | 6367 | /* |
cdd6c482 | 6368 | * Single events are their own group leaders, with an |
04289bb9 IM |
6369 | * empty sibling list: |
6370 | */ | |
6371 | if (!group_leader) | |
cdd6c482 | 6372 | group_leader = event; |
04289bb9 | 6373 | |
cdd6c482 IM |
6374 | mutex_init(&event->child_mutex); |
6375 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6376 | |
cdd6c482 IM |
6377 | INIT_LIST_HEAD(&event->group_entry); |
6378 | INIT_LIST_HEAD(&event->event_entry); | |
6379 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
6380 | INIT_LIST_HEAD(&event->rb_entry); |
6381 | ||
cdd6c482 | 6382 | init_waitqueue_head(&event->waitq); |
e360adbe | 6383 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6384 | |
cdd6c482 | 6385 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6386 | |
a6fa941d | 6387 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6388 | event->cpu = cpu; |
6389 | event->attr = *attr; | |
6390 | event->group_leader = group_leader; | |
6391 | event->pmu = NULL; | |
cdd6c482 | 6392 | event->oncpu = -1; |
a96bbc16 | 6393 | |
cdd6c482 | 6394 | event->parent = parent_event; |
b84fbc9f | 6395 | |
17cf22c3 | 6396 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6397 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6398 | |
cdd6c482 | 6399 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6400 | |
d580ff86 PZ |
6401 | if (task) { |
6402 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6403 | |
6404 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6405 | event->hw.tp_target = task; | |
d580ff86 PZ |
6406 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6407 | /* | |
6408 | * hw_breakpoint is a bit difficult here.. | |
6409 | */ | |
f22c1bb6 | 6410 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6411 | event->hw.bp_target = task; |
6412 | #endif | |
6413 | } | |
6414 | ||
4dc0da86 | 6415 | if (!overflow_handler && parent_event) { |
b326e956 | 6416 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6417 | context = parent_event->overflow_handler_context; |
6418 | } | |
66832eb4 | 6419 | |
b326e956 | 6420 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6421 | event->overflow_handler_context = context; |
97eaf530 | 6422 | |
0231bb53 | 6423 | perf_event__state_init(event); |
a86ed508 | 6424 | |
4aeb0b42 | 6425 | pmu = NULL; |
b8e83514 | 6426 | |
cdd6c482 | 6427 | hwc = &event->hw; |
bd2b5b12 | 6428 | hwc->sample_period = attr->sample_period; |
0d48696f | 6429 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6430 | hwc->sample_period = 1; |
eced1dfc | 6431 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6432 | |
e7850595 | 6433 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6434 | |
2023b359 | 6435 | /* |
cdd6c482 | 6436 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6437 | */ |
3dab77fb | 6438 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
2023b359 PZ |
6439 | goto done; |
6440 | ||
b0a873eb | 6441 | pmu = perf_init_event(event); |
974802ea | 6442 | |
d5d2bc0d PM |
6443 | done: |
6444 | err = 0; | |
4aeb0b42 | 6445 | if (!pmu) |
d5d2bc0d | 6446 | err = -EINVAL; |
4aeb0b42 RR |
6447 | else if (IS_ERR(pmu)) |
6448 | err = PTR_ERR(pmu); | |
5c92d124 | 6449 | |
d5d2bc0d | 6450 | if (err) { |
cdd6c482 IM |
6451 | if (event->ns) |
6452 | put_pid_ns(event->ns); | |
6453 | kfree(event); | |
d5d2bc0d | 6454 | return ERR_PTR(err); |
621a01ea | 6455 | } |
d5d2bc0d | 6456 | |
cdd6c482 | 6457 | if (!event->parent) { |
82cd6def | 6458 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 6459 | static_key_slow_inc(&perf_sched_events.key); |
3af9e859 | 6460 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
6461 | atomic_inc(&nr_mmap_events); |
6462 | if (event->attr.comm) | |
6463 | atomic_inc(&nr_comm_events); | |
6464 | if (event->attr.task) | |
6465 | atomic_inc(&nr_task_events); | |
927c7a9e FW |
6466 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6467 | err = get_callchain_buffers(); | |
6468 | if (err) { | |
6469 | free_event(event); | |
6470 | return ERR_PTR(err); | |
6471 | } | |
6472 | } | |
d010b332 SE |
6473 | if (has_branch_stack(event)) { |
6474 | static_key_slow_inc(&perf_sched_events.key); | |
6475 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6476 | atomic_inc(&per_cpu(perf_branch_stack_events, | |
6477 | event->cpu)); | |
6478 | } | |
f344011c | 6479 | } |
9ee318a7 | 6480 | |
cdd6c482 | 6481 | return event; |
0793a61d TG |
6482 | } |
6483 | ||
cdd6c482 IM |
6484 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6485 | struct perf_event_attr *attr) | |
974802ea | 6486 | { |
974802ea | 6487 | u32 size; |
cdf8073d | 6488 | int ret; |
974802ea PZ |
6489 | |
6490 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6491 | return -EFAULT; | |
6492 | ||
6493 | /* | |
6494 | * zero the full structure, so that a short copy will be nice. | |
6495 | */ | |
6496 | memset(attr, 0, sizeof(*attr)); | |
6497 | ||
6498 | ret = get_user(size, &uattr->size); | |
6499 | if (ret) | |
6500 | return ret; | |
6501 | ||
6502 | if (size > PAGE_SIZE) /* silly large */ | |
6503 | goto err_size; | |
6504 | ||
6505 | if (!size) /* abi compat */ | |
6506 | size = PERF_ATTR_SIZE_VER0; | |
6507 | ||
6508 | if (size < PERF_ATTR_SIZE_VER0) | |
6509 | goto err_size; | |
6510 | ||
6511 | /* | |
6512 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6513 | * ensure all the unknown bits are 0 - i.e. new |
6514 | * user-space does not rely on any kernel feature | |
6515 | * extensions we dont know about yet. | |
974802ea PZ |
6516 | */ |
6517 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6518 | unsigned char __user *addr; |
6519 | unsigned char __user *end; | |
6520 | unsigned char val; | |
974802ea | 6521 | |
cdf8073d IS |
6522 | addr = (void __user *)uattr + sizeof(*attr); |
6523 | end = (void __user *)uattr + size; | |
974802ea | 6524 | |
cdf8073d | 6525 | for (; addr < end; addr++) { |
974802ea PZ |
6526 | ret = get_user(val, addr); |
6527 | if (ret) | |
6528 | return ret; | |
6529 | if (val) | |
6530 | goto err_size; | |
6531 | } | |
b3e62e35 | 6532 | size = sizeof(*attr); |
974802ea PZ |
6533 | } |
6534 | ||
6535 | ret = copy_from_user(attr, uattr, size); | |
6536 | if (ret) | |
6537 | return -EFAULT; | |
6538 | ||
cd757645 | 6539 | if (attr->__reserved_1) |
974802ea PZ |
6540 | return -EINVAL; |
6541 | ||
6542 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6543 | return -EINVAL; | |
6544 | ||
6545 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6546 | return -EINVAL; | |
6547 | ||
bce38cd5 SE |
6548 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6549 | u64 mask = attr->branch_sample_type; | |
6550 | ||
6551 | /* only using defined bits */ | |
6552 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6553 | return -EINVAL; | |
6554 | ||
6555 | /* at least one branch bit must be set */ | |
6556 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6557 | return -EINVAL; | |
6558 | ||
bce38cd5 SE |
6559 | /* propagate priv level, when not set for branch */ |
6560 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6561 | ||
6562 | /* exclude_kernel checked on syscall entry */ | |
6563 | if (!attr->exclude_kernel) | |
6564 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6565 | ||
6566 | if (!attr->exclude_user) | |
6567 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6568 | ||
6569 | if (!attr->exclude_hv) | |
6570 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6571 | /* | |
6572 | * adjust user setting (for HW filter setup) | |
6573 | */ | |
6574 | attr->branch_sample_type = mask; | |
6575 | } | |
e712209a SE |
6576 | /* privileged levels capture (kernel, hv): check permissions */ |
6577 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
6578 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
6579 | return -EACCES; | |
bce38cd5 | 6580 | } |
4018994f | 6581 | |
c5ebcedb | 6582 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 6583 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
6584 | if (ret) |
6585 | return ret; | |
6586 | } | |
6587 | ||
6588 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
6589 | if (!arch_perf_have_user_stack_dump()) | |
6590 | return -ENOSYS; | |
6591 | ||
6592 | /* | |
6593 | * We have __u32 type for the size, but so far | |
6594 | * we can only use __u16 as maximum due to the | |
6595 | * __u16 sample size limit. | |
6596 | */ | |
6597 | if (attr->sample_stack_user >= USHRT_MAX) | |
6598 | ret = -EINVAL; | |
6599 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
6600 | ret = -EINVAL; | |
6601 | } | |
4018994f | 6602 | |
974802ea PZ |
6603 | out: |
6604 | return ret; | |
6605 | ||
6606 | err_size: | |
6607 | put_user(sizeof(*attr), &uattr->size); | |
6608 | ret = -E2BIG; | |
6609 | goto out; | |
6610 | } | |
6611 | ||
ac9721f3 PZ |
6612 | static int |
6613 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6614 | { |
76369139 | 6615 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6616 | int ret = -EINVAL; |
6617 | ||
ac9721f3 | 6618 | if (!output_event) |
a4be7c27 PZ |
6619 | goto set; |
6620 | ||
ac9721f3 PZ |
6621 | /* don't allow circular references */ |
6622 | if (event == output_event) | |
a4be7c27 PZ |
6623 | goto out; |
6624 | ||
0f139300 PZ |
6625 | /* |
6626 | * Don't allow cross-cpu buffers | |
6627 | */ | |
6628 | if (output_event->cpu != event->cpu) | |
6629 | goto out; | |
6630 | ||
6631 | /* | |
76369139 | 6632 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6633 | */ |
6634 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6635 | goto out; | |
6636 | ||
a4be7c27 | 6637 | set: |
cdd6c482 | 6638 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6639 | /* Can't redirect output if we've got an active mmap() */ |
6640 | if (atomic_read(&event->mmap_count)) | |
6641 | goto unlock; | |
a4be7c27 | 6642 | |
9bb5d40c PZ |
6643 | old_rb = event->rb; |
6644 | ||
ac9721f3 | 6645 | if (output_event) { |
76369139 FW |
6646 | /* get the rb we want to redirect to */ |
6647 | rb = ring_buffer_get(output_event); | |
6648 | if (!rb) | |
ac9721f3 | 6649 | goto unlock; |
a4be7c27 PZ |
6650 | } |
6651 | ||
10c6db11 PZ |
6652 | if (old_rb) |
6653 | ring_buffer_detach(event, old_rb); | |
9bb5d40c PZ |
6654 | |
6655 | if (rb) | |
6656 | ring_buffer_attach(event, rb); | |
6657 | ||
6658 | rcu_assign_pointer(event->rb, rb); | |
6659 | ||
6660 | if (old_rb) { | |
6661 | ring_buffer_put(old_rb); | |
6662 | /* | |
6663 | * Since we detached before setting the new rb, so that we | |
6664 | * could attach the new rb, we could have missed a wakeup. | |
6665 | * Provide it now. | |
6666 | */ | |
6667 | wake_up_all(&event->waitq); | |
6668 | } | |
6669 | ||
a4be7c27 | 6670 | ret = 0; |
ac9721f3 PZ |
6671 | unlock: |
6672 | mutex_unlock(&event->mmap_mutex); | |
6673 | ||
a4be7c27 | 6674 | out: |
a4be7c27 PZ |
6675 | return ret; |
6676 | } | |
6677 | ||
0793a61d | 6678 | /** |
cdd6c482 | 6679 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6680 | * |
cdd6c482 | 6681 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6682 | * @pid: target pid |
9f66a381 | 6683 | * @cpu: target cpu |
cdd6c482 | 6684 | * @group_fd: group leader event fd |
0793a61d | 6685 | */ |
cdd6c482 IM |
6686 | SYSCALL_DEFINE5(perf_event_open, |
6687 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 6688 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 6689 | { |
b04243ef PZ |
6690 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6691 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
6692 | struct perf_event_attr attr; |
6693 | struct perf_event_context *ctx; | |
6694 | struct file *event_file = NULL; | |
2903ff01 | 6695 | struct fd group = {NULL, 0}; |
38a81da2 | 6696 | struct task_struct *task = NULL; |
89a1e187 | 6697 | struct pmu *pmu; |
ea635c64 | 6698 | int event_fd; |
b04243ef | 6699 | int move_group = 0; |
dc86cabe | 6700 | int err; |
0793a61d | 6701 | |
2743a5b0 | 6702 | /* for future expandability... */ |
e5d1367f | 6703 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
6704 | return -EINVAL; |
6705 | ||
dc86cabe IM |
6706 | err = perf_copy_attr(attr_uptr, &attr); |
6707 | if (err) | |
6708 | return err; | |
eab656ae | 6709 | |
0764771d PZ |
6710 | if (!attr.exclude_kernel) { |
6711 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6712 | return -EACCES; | |
6713 | } | |
6714 | ||
df58ab24 | 6715 | if (attr.freq) { |
cdd6c482 | 6716 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
6717 | return -EINVAL; |
6718 | } | |
6719 | ||
e5d1367f SE |
6720 | /* |
6721 | * In cgroup mode, the pid argument is used to pass the fd | |
6722 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6723 | * designates the cpu on which to monitor threads from that | |
6724 | * cgroup. | |
6725 | */ | |
6726 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6727 | return -EINVAL; | |
6728 | ||
ab72a702 | 6729 | event_fd = get_unused_fd(); |
ea635c64 AV |
6730 | if (event_fd < 0) |
6731 | return event_fd; | |
6732 | ||
ac9721f3 | 6733 | if (group_fd != -1) { |
2903ff01 AV |
6734 | err = perf_fget_light(group_fd, &group); |
6735 | if (err) | |
d14b12d7 | 6736 | goto err_fd; |
2903ff01 | 6737 | group_leader = group.file->private_data; |
ac9721f3 PZ |
6738 | if (flags & PERF_FLAG_FD_OUTPUT) |
6739 | output_event = group_leader; | |
6740 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6741 | group_leader = NULL; | |
6742 | } | |
6743 | ||
e5d1367f | 6744 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6745 | task = find_lively_task_by_vpid(pid); |
6746 | if (IS_ERR(task)) { | |
6747 | err = PTR_ERR(task); | |
6748 | goto err_group_fd; | |
6749 | } | |
6750 | } | |
6751 | ||
fbfc623f YZ |
6752 | get_online_cpus(); |
6753 | ||
4dc0da86 AK |
6754 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6755 | NULL, NULL); | |
d14b12d7 SE |
6756 | if (IS_ERR(event)) { |
6757 | err = PTR_ERR(event); | |
c6be5a5c | 6758 | goto err_task; |
d14b12d7 SE |
6759 | } |
6760 | ||
e5d1367f SE |
6761 | if (flags & PERF_FLAG_PID_CGROUP) { |
6762 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
6763 | if (err) | |
6764 | goto err_alloc; | |
08309379 PZ |
6765 | /* |
6766 | * one more event: | |
6767 | * - that has cgroup constraint on event->cpu | |
6768 | * - that may need work on context switch | |
6769 | */ | |
6770 | atomic_inc(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 6771 | static_key_slow_inc(&perf_sched_events.key); |
e5d1367f SE |
6772 | } |
6773 | ||
89a1e187 PZ |
6774 | /* |
6775 | * Special case software events and allow them to be part of | |
6776 | * any hardware group. | |
6777 | */ | |
6778 | pmu = event->pmu; | |
b04243ef PZ |
6779 | |
6780 | if (group_leader && | |
6781 | (is_software_event(event) != is_software_event(group_leader))) { | |
6782 | if (is_software_event(event)) { | |
6783 | /* | |
6784 | * If event and group_leader are not both a software | |
6785 | * event, and event is, then group leader is not. | |
6786 | * | |
6787 | * Allow the addition of software events to !software | |
6788 | * groups, this is safe because software events never | |
6789 | * fail to schedule. | |
6790 | */ | |
6791 | pmu = group_leader->pmu; | |
6792 | } else if (is_software_event(group_leader) && | |
6793 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
6794 | /* | |
6795 | * In case the group is a pure software group, and we | |
6796 | * try to add a hardware event, move the whole group to | |
6797 | * the hardware context. | |
6798 | */ | |
6799 | move_group = 1; | |
6800 | } | |
6801 | } | |
89a1e187 PZ |
6802 | |
6803 | /* | |
6804 | * Get the target context (task or percpu): | |
6805 | */ | |
e2d37cd2 | 6806 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
6807 | if (IS_ERR(ctx)) { |
6808 | err = PTR_ERR(ctx); | |
c6be5a5c | 6809 | goto err_alloc; |
89a1e187 PZ |
6810 | } |
6811 | ||
fd1edb3a PZ |
6812 | if (task) { |
6813 | put_task_struct(task); | |
6814 | task = NULL; | |
6815 | } | |
6816 | ||
ccff286d | 6817 | /* |
cdd6c482 | 6818 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 6819 | */ |
ac9721f3 | 6820 | if (group_leader) { |
dc86cabe | 6821 | err = -EINVAL; |
04289bb9 | 6822 | |
04289bb9 | 6823 | /* |
ccff286d IM |
6824 | * Do not allow a recursive hierarchy (this new sibling |
6825 | * becoming part of another group-sibling): | |
6826 | */ | |
6827 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 6828 | goto err_context; |
ccff286d IM |
6829 | /* |
6830 | * Do not allow to attach to a group in a different | |
6831 | * task or CPU context: | |
04289bb9 | 6832 | */ |
b04243ef PZ |
6833 | if (move_group) { |
6834 | if (group_leader->ctx->type != ctx->type) | |
6835 | goto err_context; | |
6836 | } else { | |
6837 | if (group_leader->ctx != ctx) | |
6838 | goto err_context; | |
6839 | } | |
6840 | ||
3b6f9e5c PM |
6841 | /* |
6842 | * Only a group leader can be exclusive or pinned | |
6843 | */ | |
0d48696f | 6844 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 6845 | goto err_context; |
ac9721f3 PZ |
6846 | } |
6847 | ||
6848 | if (output_event) { | |
6849 | err = perf_event_set_output(event, output_event); | |
6850 | if (err) | |
c3f00c70 | 6851 | goto err_context; |
ac9721f3 | 6852 | } |
0793a61d | 6853 | |
ea635c64 AV |
6854 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
6855 | if (IS_ERR(event_file)) { | |
6856 | err = PTR_ERR(event_file); | |
c3f00c70 | 6857 | goto err_context; |
ea635c64 | 6858 | } |
9b51f66d | 6859 | |
b04243ef PZ |
6860 | if (move_group) { |
6861 | struct perf_event_context *gctx = group_leader->ctx; | |
6862 | ||
6863 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 6864 | perf_remove_from_context(group_leader); |
0231bb53 JO |
6865 | |
6866 | /* | |
6867 | * Removing from the context ends up with disabled | |
6868 | * event. What we want here is event in the initial | |
6869 | * startup state, ready to be add into new context. | |
6870 | */ | |
6871 | perf_event__state_init(group_leader); | |
b04243ef PZ |
6872 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6873 | group_entry) { | |
fe4b04fa | 6874 | perf_remove_from_context(sibling); |
0231bb53 | 6875 | perf_event__state_init(sibling); |
b04243ef PZ |
6876 | put_ctx(gctx); |
6877 | } | |
6878 | mutex_unlock(&gctx->mutex); | |
6879 | put_ctx(gctx); | |
ea635c64 | 6880 | } |
9b51f66d | 6881 | |
ad3a37de | 6882 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 6883 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
6884 | |
6885 | if (move_group) { | |
0cda4c02 | 6886 | synchronize_rcu(); |
e2d37cd2 | 6887 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
6888 | get_ctx(ctx); |
6889 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
6890 | group_entry) { | |
e2d37cd2 | 6891 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
6892 | get_ctx(ctx); |
6893 | } | |
6894 | } | |
6895 | ||
e2d37cd2 | 6896 | perf_install_in_context(ctx, event, event->cpu); |
ad3a37de | 6897 | ++ctx->generation; |
fe4b04fa | 6898 | perf_unpin_context(ctx); |
d859e29f | 6899 | mutex_unlock(&ctx->mutex); |
9b51f66d | 6900 | |
fbfc623f YZ |
6901 | put_online_cpus(); |
6902 | ||
cdd6c482 | 6903 | event->owner = current; |
8882135b | 6904 | |
cdd6c482 IM |
6905 | mutex_lock(¤t->perf_event_mutex); |
6906 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
6907 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 6908 | |
c320c7b7 ACM |
6909 | /* |
6910 | * Precalculate sample_data sizes | |
6911 | */ | |
6912 | perf_event__header_size(event); | |
6844c09d | 6913 | perf_event__id_header_size(event); |
c320c7b7 | 6914 | |
8a49542c PZ |
6915 | /* |
6916 | * Drop the reference on the group_event after placing the | |
6917 | * new event on the sibling_list. This ensures destruction | |
6918 | * of the group leader will find the pointer to itself in | |
6919 | * perf_group_detach(). | |
6920 | */ | |
2903ff01 | 6921 | fdput(group); |
ea635c64 AV |
6922 | fd_install(event_fd, event_file); |
6923 | return event_fd; | |
0793a61d | 6924 | |
c3f00c70 | 6925 | err_context: |
fe4b04fa | 6926 | perf_unpin_context(ctx); |
ea635c64 | 6927 | put_ctx(ctx); |
c6be5a5c | 6928 | err_alloc: |
ea635c64 | 6929 | free_event(event); |
e7d0bc04 | 6930 | err_task: |
fbfc623f | 6931 | put_online_cpus(); |
e7d0bc04 PZ |
6932 | if (task) |
6933 | put_task_struct(task); | |
89a1e187 | 6934 | err_group_fd: |
2903ff01 | 6935 | fdput(group); |
ea635c64 AV |
6936 | err_fd: |
6937 | put_unused_fd(event_fd); | |
dc86cabe | 6938 | return err; |
0793a61d TG |
6939 | } |
6940 | ||
fb0459d7 AV |
6941 | /** |
6942 | * perf_event_create_kernel_counter | |
6943 | * | |
6944 | * @attr: attributes of the counter to create | |
6945 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 6946 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
6947 | */ |
6948 | struct perf_event * | |
6949 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 6950 | struct task_struct *task, |
4dc0da86 AK |
6951 | perf_overflow_handler_t overflow_handler, |
6952 | void *context) | |
fb0459d7 | 6953 | { |
fb0459d7 | 6954 | struct perf_event_context *ctx; |
c3f00c70 | 6955 | struct perf_event *event; |
fb0459d7 | 6956 | int err; |
d859e29f | 6957 | |
fb0459d7 AV |
6958 | /* |
6959 | * Get the target context (task or percpu): | |
6960 | */ | |
d859e29f | 6961 | |
4dc0da86 AK |
6962 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
6963 | overflow_handler, context); | |
c3f00c70 PZ |
6964 | if (IS_ERR(event)) { |
6965 | err = PTR_ERR(event); | |
6966 | goto err; | |
6967 | } | |
d859e29f | 6968 | |
38a81da2 | 6969 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
6970 | if (IS_ERR(ctx)) { |
6971 | err = PTR_ERR(ctx); | |
c3f00c70 | 6972 | goto err_free; |
d859e29f | 6973 | } |
fb0459d7 | 6974 | |
fb0459d7 AV |
6975 | WARN_ON_ONCE(ctx->parent_ctx); |
6976 | mutex_lock(&ctx->mutex); | |
6977 | perf_install_in_context(ctx, event, cpu); | |
6978 | ++ctx->generation; | |
fe4b04fa | 6979 | perf_unpin_context(ctx); |
fb0459d7 AV |
6980 | mutex_unlock(&ctx->mutex); |
6981 | ||
fb0459d7 AV |
6982 | return event; |
6983 | ||
c3f00c70 PZ |
6984 | err_free: |
6985 | free_event(event); | |
6986 | err: | |
c6567f64 | 6987 | return ERR_PTR(err); |
9b51f66d | 6988 | } |
fb0459d7 | 6989 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 6990 | |
0cda4c02 YZ |
6991 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
6992 | { | |
6993 | struct perf_event_context *src_ctx; | |
6994 | struct perf_event_context *dst_ctx; | |
6995 | struct perf_event *event, *tmp; | |
6996 | LIST_HEAD(events); | |
6997 | ||
6998 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
6999 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7000 | ||
7001 | mutex_lock(&src_ctx->mutex); | |
7002 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
7003 | event_entry) { | |
7004 | perf_remove_from_context(event); | |
7005 | put_ctx(src_ctx); | |
7006 | list_add(&event->event_entry, &events); | |
7007 | } | |
7008 | mutex_unlock(&src_ctx->mutex); | |
7009 | ||
7010 | synchronize_rcu(); | |
7011 | ||
7012 | mutex_lock(&dst_ctx->mutex); | |
7013 | list_for_each_entry_safe(event, tmp, &events, event_entry) { | |
7014 | list_del(&event->event_entry); | |
7015 | if (event->state >= PERF_EVENT_STATE_OFF) | |
7016 | event->state = PERF_EVENT_STATE_INACTIVE; | |
7017 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
7018 | get_ctx(dst_ctx); | |
7019 | } | |
7020 | mutex_unlock(&dst_ctx->mutex); | |
7021 | } | |
7022 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7023 | ||
cdd6c482 | 7024 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7025 | struct task_struct *child) |
d859e29f | 7026 | { |
cdd6c482 | 7027 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7028 | u64 child_val; |
d859e29f | 7029 | |
cdd6c482 IM |
7030 | if (child_event->attr.inherit_stat) |
7031 | perf_event_read_event(child_event, child); | |
38b200d6 | 7032 | |
b5e58793 | 7033 | child_val = perf_event_count(child_event); |
d859e29f PM |
7034 | |
7035 | /* | |
7036 | * Add back the child's count to the parent's count: | |
7037 | */ | |
a6e6dea6 | 7038 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7039 | atomic64_add(child_event->total_time_enabled, |
7040 | &parent_event->child_total_time_enabled); | |
7041 | atomic64_add(child_event->total_time_running, | |
7042 | &parent_event->child_total_time_running); | |
d859e29f PM |
7043 | |
7044 | /* | |
cdd6c482 | 7045 | * Remove this event from the parent's list |
d859e29f | 7046 | */ |
cdd6c482 IM |
7047 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7048 | mutex_lock(&parent_event->child_mutex); | |
7049 | list_del_init(&child_event->child_list); | |
7050 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
7051 | |
7052 | /* | |
cdd6c482 | 7053 | * Release the parent event, if this was the last |
d859e29f PM |
7054 | * reference to it. |
7055 | */ | |
a6fa941d | 7056 | put_event(parent_event); |
d859e29f PM |
7057 | } |
7058 | ||
9b51f66d | 7059 | static void |
cdd6c482 IM |
7060 | __perf_event_exit_task(struct perf_event *child_event, |
7061 | struct perf_event_context *child_ctx, | |
38b200d6 | 7062 | struct task_struct *child) |
9b51f66d | 7063 | { |
38b435b1 PZ |
7064 | if (child_event->parent) { |
7065 | raw_spin_lock_irq(&child_ctx->lock); | |
7066 | perf_group_detach(child_event); | |
7067 | raw_spin_unlock_irq(&child_ctx->lock); | |
7068 | } | |
9b51f66d | 7069 | |
fe4b04fa | 7070 | perf_remove_from_context(child_event); |
0cc0c027 | 7071 | |
9b51f66d | 7072 | /* |
38b435b1 | 7073 | * It can happen that the parent exits first, and has events |
9b51f66d | 7074 | * that are still around due to the child reference. These |
38b435b1 | 7075 | * events need to be zapped. |
9b51f66d | 7076 | */ |
38b435b1 | 7077 | if (child_event->parent) { |
cdd6c482 IM |
7078 | sync_child_event(child_event, child); |
7079 | free_event(child_event); | |
4bcf349a | 7080 | } |
9b51f66d IM |
7081 | } |
7082 | ||
8dc85d54 | 7083 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7084 | { |
cdd6c482 IM |
7085 | struct perf_event *child_event, *tmp; |
7086 | struct perf_event_context *child_ctx; | |
a63eaf34 | 7087 | unsigned long flags; |
9b51f66d | 7088 | |
8dc85d54 | 7089 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7090 | perf_event_task(child, NULL, 0); |
9b51f66d | 7091 | return; |
9f498cc5 | 7092 | } |
9b51f66d | 7093 | |
a63eaf34 | 7094 | local_irq_save(flags); |
ad3a37de PM |
7095 | /* |
7096 | * We can't reschedule here because interrupts are disabled, | |
7097 | * and either child is current or it is a task that can't be | |
7098 | * scheduled, so we are now safe from rescheduling changing | |
7099 | * our context. | |
7100 | */ | |
806839b2 | 7101 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7102 | |
7103 | /* | |
7104 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7105 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7106 | * incremented the context's refcount before we do put_ctx below. |
7107 | */ | |
e625cce1 | 7108 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7109 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7110 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
7111 | /* |
7112 | * If this context is a clone; unclone it so it can't get | |
7113 | * swapped to another process while we're removing all | |
cdd6c482 | 7114 | * the events from it. |
71a851b4 PZ |
7115 | */ |
7116 | unclone_ctx(child_ctx); | |
5e942bb3 | 7117 | update_context_time(child_ctx); |
e625cce1 | 7118 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
7119 | |
7120 | /* | |
cdd6c482 IM |
7121 | * Report the task dead after unscheduling the events so that we |
7122 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7123 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7124 | */ |
cdd6c482 | 7125 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 7126 | |
66fff224 PZ |
7127 | /* |
7128 | * We can recurse on the same lock type through: | |
7129 | * | |
cdd6c482 IM |
7130 | * __perf_event_exit_task() |
7131 | * sync_child_event() | |
a6fa941d AV |
7132 | * put_event() |
7133 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
7134 | * |
7135 | * But since its the parent context it won't be the same instance. | |
7136 | */ | |
a0507c84 | 7137 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 7138 | |
8bc20959 | 7139 | again: |
889ff015 FW |
7140 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
7141 | group_entry) | |
7142 | __perf_event_exit_task(child_event, child_ctx, child); | |
7143 | ||
7144 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 7145 | group_entry) |
cdd6c482 | 7146 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
7147 | |
7148 | /* | |
cdd6c482 | 7149 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
7150 | * its siblings to the list, but we obtained 'tmp' before that which |
7151 | * will still point to the list head terminating the iteration. | |
7152 | */ | |
889ff015 FW |
7153 | if (!list_empty(&child_ctx->pinned_groups) || |
7154 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 7155 | goto again; |
a63eaf34 PM |
7156 | |
7157 | mutex_unlock(&child_ctx->mutex); | |
7158 | ||
7159 | put_ctx(child_ctx); | |
9b51f66d IM |
7160 | } |
7161 | ||
8dc85d54 PZ |
7162 | /* |
7163 | * When a child task exits, feed back event values to parent events. | |
7164 | */ | |
7165 | void perf_event_exit_task(struct task_struct *child) | |
7166 | { | |
8882135b | 7167 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
7168 | int ctxn; |
7169 | ||
8882135b PZ |
7170 | mutex_lock(&child->perf_event_mutex); |
7171 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
7172 | owner_entry) { | |
7173 | list_del_init(&event->owner_entry); | |
7174 | ||
7175 | /* | |
7176 | * Ensure the list deletion is visible before we clear | |
7177 | * the owner, closes a race against perf_release() where | |
7178 | * we need to serialize on the owner->perf_event_mutex. | |
7179 | */ | |
7180 | smp_wmb(); | |
7181 | event->owner = NULL; | |
7182 | } | |
7183 | mutex_unlock(&child->perf_event_mutex); | |
7184 | ||
8dc85d54 PZ |
7185 | for_each_task_context_nr(ctxn) |
7186 | perf_event_exit_task_context(child, ctxn); | |
7187 | } | |
7188 | ||
889ff015 FW |
7189 | static void perf_free_event(struct perf_event *event, |
7190 | struct perf_event_context *ctx) | |
7191 | { | |
7192 | struct perf_event *parent = event->parent; | |
7193 | ||
7194 | if (WARN_ON_ONCE(!parent)) | |
7195 | return; | |
7196 | ||
7197 | mutex_lock(&parent->child_mutex); | |
7198 | list_del_init(&event->child_list); | |
7199 | mutex_unlock(&parent->child_mutex); | |
7200 | ||
a6fa941d | 7201 | put_event(parent); |
889ff015 | 7202 | |
8a49542c | 7203 | perf_group_detach(event); |
889ff015 FW |
7204 | list_del_event(event, ctx); |
7205 | free_event(event); | |
7206 | } | |
7207 | ||
bbbee908 PZ |
7208 | /* |
7209 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 7210 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 7211 | */ |
cdd6c482 | 7212 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 7213 | { |
8dc85d54 | 7214 | struct perf_event_context *ctx; |
cdd6c482 | 7215 | struct perf_event *event, *tmp; |
8dc85d54 | 7216 | int ctxn; |
bbbee908 | 7217 | |
8dc85d54 PZ |
7218 | for_each_task_context_nr(ctxn) { |
7219 | ctx = task->perf_event_ctxp[ctxn]; | |
7220 | if (!ctx) | |
7221 | continue; | |
bbbee908 | 7222 | |
8dc85d54 | 7223 | mutex_lock(&ctx->mutex); |
bbbee908 | 7224 | again: |
8dc85d54 PZ |
7225 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
7226 | group_entry) | |
7227 | perf_free_event(event, ctx); | |
bbbee908 | 7228 | |
8dc85d54 PZ |
7229 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
7230 | group_entry) | |
7231 | perf_free_event(event, ctx); | |
bbbee908 | 7232 | |
8dc85d54 PZ |
7233 | if (!list_empty(&ctx->pinned_groups) || |
7234 | !list_empty(&ctx->flexible_groups)) | |
7235 | goto again; | |
bbbee908 | 7236 | |
8dc85d54 | 7237 | mutex_unlock(&ctx->mutex); |
bbbee908 | 7238 | |
8dc85d54 PZ |
7239 | put_ctx(ctx); |
7240 | } | |
889ff015 FW |
7241 | } |
7242 | ||
4e231c79 PZ |
7243 | void perf_event_delayed_put(struct task_struct *task) |
7244 | { | |
7245 | int ctxn; | |
7246 | ||
7247 | for_each_task_context_nr(ctxn) | |
7248 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7249 | } | |
7250 | ||
97dee4f3 PZ |
7251 | /* |
7252 | * inherit a event from parent task to child task: | |
7253 | */ | |
7254 | static struct perf_event * | |
7255 | inherit_event(struct perf_event *parent_event, | |
7256 | struct task_struct *parent, | |
7257 | struct perf_event_context *parent_ctx, | |
7258 | struct task_struct *child, | |
7259 | struct perf_event *group_leader, | |
7260 | struct perf_event_context *child_ctx) | |
7261 | { | |
7262 | struct perf_event *child_event; | |
cee010ec | 7263 | unsigned long flags; |
97dee4f3 PZ |
7264 | |
7265 | /* | |
7266 | * Instead of creating recursive hierarchies of events, | |
7267 | * we link inherited events back to the original parent, | |
7268 | * which has a filp for sure, which we use as the reference | |
7269 | * count: | |
7270 | */ | |
7271 | if (parent_event->parent) | |
7272 | parent_event = parent_event->parent; | |
7273 | ||
7274 | child_event = perf_event_alloc(&parent_event->attr, | |
7275 | parent_event->cpu, | |
d580ff86 | 7276 | child, |
97dee4f3 | 7277 | group_leader, parent_event, |
4dc0da86 | 7278 | NULL, NULL); |
97dee4f3 PZ |
7279 | if (IS_ERR(child_event)) |
7280 | return child_event; | |
a6fa941d AV |
7281 | |
7282 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7283 | free_event(child_event); | |
7284 | return NULL; | |
7285 | } | |
7286 | ||
97dee4f3 PZ |
7287 | get_ctx(child_ctx); |
7288 | ||
7289 | /* | |
7290 | * Make the child state follow the state of the parent event, | |
7291 | * not its attr.disabled bit. We hold the parent's mutex, | |
7292 | * so we won't race with perf_event_{en, dis}able_family. | |
7293 | */ | |
7294 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7295 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7296 | else | |
7297 | child_event->state = PERF_EVENT_STATE_OFF; | |
7298 | ||
7299 | if (parent_event->attr.freq) { | |
7300 | u64 sample_period = parent_event->hw.sample_period; | |
7301 | struct hw_perf_event *hwc = &child_event->hw; | |
7302 | ||
7303 | hwc->sample_period = sample_period; | |
7304 | hwc->last_period = sample_period; | |
7305 | ||
7306 | local64_set(&hwc->period_left, sample_period); | |
7307 | } | |
7308 | ||
7309 | child_event->ctx = child_ctx; | |
7310 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7311 | child_event->overflow_handler_context |
7312 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7313 | |
614b6780 TG |
7314 | /* |
7315 | * Precalculate sample_data sizes | |
7316 | */ | |
7317 | perf_event__header_size(child_event); | |
6844c09d | 7318 | perf_event__id_header_size(child_event); |
614b6780 | 7319 | |
97dee4f3 PZ |
7320 | /* |
7321 | * Link it up in the child's context: | |
7322 | */ | |
cee010ec | 7323 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7324 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7325 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7326 | |
97dee4f3 PZ |
7327 | /* |
7328 | * Link this into the parent event's child list | |
7329 | */ | |
7330 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7331 | mutex_lock(&parent_event->child_mutex); | |
7332 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7333 | mutex_unlock(&parent_event->child_mutex); | |
7334 | ||
7335 | return child_event; | |
7336 | } | |
7337 | ||
7338 | static int inherit_group(struct perf_event *parent_event, | |
7339 | struct task_struct *parent, | |
7340 | struct perf_event_context *parent_ctx, | |
7341 | struct task_struct *child, | |
7342 | struct perf_event_context *child_ctx) | |
7343 | { | |
7344 | struct perf_event *leader; | |
7345 | struct perf_event *sub; | |
7346 | struct perf_event *child_ctr; | |
7347 | ||
7348 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7349 | child, NULL, child_ctx); | |
7350 | if (IS_ERR(leader)) | |
7351 | return PTR_ERR(leader); | |
7352 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7353 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7354 | child, leader, child_ctx); | |
7355 | if (IS_ERR(child_ctr)) | |
7356 | return PTR_ERR(child_ctr); | |
7357 | } | |
7358 | return 0; | |
889ff015 FW |
7359 | } |
7360 | ||
7361 | static int | |
7362 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7363 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7364 | struct task_struct *child, int ctxn, |
889ff015 FW |
7365 | int *inherited_all) |
7366 | { | |
7367 | int ret; | |
8dc85d54 | 7368 | struct perf_event_context *child_ctx; |
889ff015 FW |
7369 | |
7370 | if (!event->attr.inherit) { | |
7371 | *inherited_all = 0; | |
7372 | return 0; | |
bbbee908 PZ |
7373 | } |
7374 | ||
fe4b04fa | 7375 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7376 | if (!child_ctx) { |
7377 | /* | |
7378 | * This is executed from the parent task context, so | |
7379 | * inherit events that have been marked for cloning. | |
7380 | * First allocate and initialize a context for the | |
7381 | * child. | |
7382 | */ | |
bbbee908 | 7383 | |
eb184479 | 7384 | child_ctx = alloc_perf_context(event->pmu, child); |
889ff015 FW |
7385 | if (!child_ctx) |
7386 | return -ENOMEM; | |
bbbee908 | 7387 | |
8dc85d54 | 7388 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7389 | } |
7390 | ||
7391 | ret = inherit_group(event, parent, parent_ctx, | |
7392 | child, child_ctx); | |
7393 | ||
7394 | if (ret) | |
7395 | *inherited_all = 0; | |
7396 | ||
7397 | return ret; | |
bbbee908 PZ |
7398 | } |
7399 | ||
9b51f66d | 7400 | /* |
cdd6c482 | 7401 | * Initialize the perf_event context in task_struct |
9b51f66d | 7402 | */ |
8dc85d54 | 7403 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7404 | { |
889ff015 | 7405 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7406 | struct perf_event_context *cloned_ctx; |
7407 | struct perf_event *event; | |
9b51f66d | 7408 | struct task_struct *parent = current; |
564c2b21 | 7409 | int inherited_all = 1; |
dddd3379 | 7410 | unsigned long flags; |
6ab423e0 | 7411 | int ret = 0; |
9b51f66d | 7412 | |
8dc85d54 | 7413 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7414 | return 0; |
7415 | ||
ad3a37de | 7416 | /* |
25346b93 PM |
7417 | * If the parent's context is a clone, pin it so it won't get |
7418 | * swapped under us. | |
ad3a37de | 7419 | */ |
8dc85d54 | 7420 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 7421 | |
ad3a37de PM |
7422 | /* |
7423 | * No need to check if parent_ctx != NULL here; since we saw | |
7424 | * it non-NULL earlier, the only reason for it to become NULL | |
7425 | * is if we exit, and since we're currently in the middle of | |
7426 | * a fork we can't be exiting at the same time. | |
7427 | */ | |
ad3a37de | 7428 | |
9b51f66d IM |
7429 | /* |
7430 | * Lock the parent list. No need to lock the child - not PID | |
7431 | * hashed yet and not running, so nobody can access it. | |
7432 | */ | |
d859e29f | 7433 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7434 | |
7435 | /* | |
7436 | * We dont have to disable NMIs - we are only looking at | |
7437 | * the list, not manipulating it: | |
7438 | */ | |
889ff015 | 7439 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7440 | ret = inherit_task_group(event, parent, parent_ctx, |
7441 | child, ctxn, &inherited_all); | |
889ff015 FW |
7442 | if (ret) |
7443 | break; | |
7444 | } | |
b93f7978 | 7445 | |
dddd3379 TG |
7446 | /* |
7447 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7448 | * to allocations, but we need to prevent rotation because | |
7449 | * rotate_ctx() will change the list from interrupt context. | |
7450 | */ | |
7451 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7452 | parent_ctx->rotate_disable = 1; | |
7453 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7454 | ||
889ff015 | 7455 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7456 | ret = inherit_task_group(event, parent, parent_ctx, |
7457 | child, ctxn, &inherited_all); | |
889ff015 | 7458 | if (ret) |
9b51f66d | 7459 | break; |
564c2b21 PM |
7460 | } |
7461 | ||
dddd3379 TG |
7462 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7463 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7464 | |
8dc85d54 | 7465 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7466 | |
05cbaa28 | 7467 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7468 | /* |
7469 | * Mark the child context as a clone of the parent | |
7470 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7471 | * |
7472 | * Note that if the parent is a clone, the holding of | |
7473 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7474 | */ |
c5ed5145 | 7475 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7476 | if (cloned_ctx) { |
7477 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7478 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7479 | } else { |
7480 | child_ctx->parent_ctx = parent_ctx; | |
7481 | child_ctx->parent_gen = parent_ctx->generation; | |
7482 | } | |
7483 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7484 | } |
7485 | ||
c5ed5145 | 7486 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7487 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7488 | |
25346b93 | 7489 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7490 | put_ctx(parent_ctx); |
ad3a37de | 7491 | |
6ab423e0 | 7492 | return ret; |
9b51f66d IM |
7493 | } |
7494 | ||
8dc85d54 PZ |
7495 | /* |
7496 | * Initialize the perf_event context in task_struct | |
7497 | */ | |
7498 | int perf_event_init_task(struct task_struct *child) | |
7499 | { | |
7500 | int ctxn, ret; | |
7501 | ||
8550d7cb ON |
7502 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7503 | mutex_init(&child->perf_event_mutex); | |
7504 | INIT_LIST_HEAD(&child->perf_event_list); | |
7505 | ||
8dc85d54 PZ |
7506 | for_each_task_context_nr(ctxn) { |
7507 | ret = perf_event_init_context(child, ctxn); | |
7508 | if (ret) | |
7509 | return ret; | |
7510 | } | |
7511 | ||
7512 | return 0; | |
7513 | } | |
7514 | ||
220b140b PM |
7515 | static void __init perf_event_init_all_cpus(void) |
7516 | { | |
b28ab83c | 7517 | struct swevent_htable *swhash; |
220b140b | 7518 | int cpu; |
220b140b PM |
7519 | |
7520 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7521 | swhash = &per_cpu(swevent_htable, cpu); |
7522 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7523 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7524 | } |
7525 | } | |
7526 | ||
cdd6c482 | 7527 | static void __cpuinit perf_event_init_cpu(int cpu) |
0793a61d | 7528 | { |
108b02cf | 7529 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7530 | |
b28ab83c | 7531 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7532 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7533 | struct swevent_hlist *hlist; |
7534 | ||
b28ab83c PZ |
7535 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7536 | WARN_ON(!hlist); | |
7537 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7538 | } |
b28ab83c | 7539 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7540 | } |
7541 | ||
c277443c | 7542 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7543 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7544 | { |
e9d2b064 PZ |
7545 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7546 | ||
7547 | WARN_ON(!irqs_disabled()); | |
7548 | ||
7549 | list_del_init(&cpuctx->rotation_list); | |
7550 | } | |
7551 | ||
108b02cf | 7552 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7553 | { |
108b02cf | 7554 | struct perf_event_context *ctx = __info; |
cdd6c482 | 7555 | struct perf_event *event, *tmp; |
0793a61d | 7556 | |
108b02cf | 7557 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7558 | |
889ff015 | 7559 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 7560 | __perf_remove_from_context(event); |
889ff015 | 7561 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 7562 | __perf_remove_from_context(event); |
0793a61d | 7563 | } |
108b02cf PZ |
7564 | |
7565 | static void perf_event_exit_cpu_context(int cpu) | |
7566 | { | |
7567 | struct perf_event_context *ctx; | |
7568 | struct pmu *pmu; | |
7569 | int idx; | |
7570 | ||
7571 | idx = srcu_read_lock(&pmus_srcu); | |
7572 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7573 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7574 | |
7575 | mutex_lock(&ctx->mutex); | |
7576 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7577 | mutex_unlock(&ctx->mutex); | |
7578 | } | |
7579 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7580 | } |
7581 | ||
cdd6c482 | 7582 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7583 | { |
b28ab83c | 7584 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7585 | |
b28ab83c PZ |
7586 | mutex_lock(&swhash->hlist_mutex); |
7587 | swevent_hlist_release(swhash); | |
7588 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 7589 | |
108b02cf | 7590 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7591 | } |
7592 | #else | |
cdd6c482 | 7593 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7594 | #endif |
7595 | ||
c277443c PZ |
7596 | static int |
7597 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7598 | { | |
7599 | int cpu; | |
7600 | ||
7601 | for_each_online_cpu(cpu) | |
7602 | perf_event_exit_cpu(cpu); | |
7603 | ||
7604 | return NOTIFY_OK; | |
7605 | } | |
7606 | ||
7607 | /* | |
7608 | * Run the perf reboot notifier at the very last possible moment so that | |
7609 | * the generic watchdog code runs as long as possible. | |
7610 | */ | |
7611 | static struct notifier_block perf_reboot_notifier = { | |
7612 | .notifier_call = perf_reboot, | |
7613 | .priority = INT_MIN, | |
7614 | }; | |
7615 | ||
0793a61d TG |
7616 | static int __cpuinit |
7617 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
7618 | { | |
7619 | unsigned int cpu = (long)hcpu; | |
7620 | ||
4536e4d1 | 7621 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7622 | |
7623 | case CPU_UP_PREPARE: | |
5e11637e | 7624 | case CPU_DOWN_FAILED: |
cdd6c482 | 7625 | perf_event_init_cpu(cpu); |
0793a61d TG |
7626 | break; |
7627 | ||
5e11637e | 7628 | case CPU_UP_CANCELED: |
0793a61d | 7629 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7630 | perf_event_exit_cpu(cpu); |
0793a61d | 7631 | break; |
0793a61d TG |
7632 | default: |
7633 | break; | |
7634 | } | |
7635 | ||
7636 | return NOTIFY_OK; | |
7637 | } | |
7638 | ||
cdd6c482 | 7639 | void __init perf_event_init(void) |
0793a61d | 7640 | { |
3c502e7a JW |
7641 | int ret; |
7642 | ||
2e80a82a PZ |
7643 | idr_init(&pmu_idr); |
7644 | ||
220b140b | 7645 | perf_event_init_all_cpus(); |
b0a873eb | 7646 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7647 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7648 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7649 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7650 | perf_tp_register(); |
7651 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7652 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7653 | |
7654 | ret = init_hw_breakpoint(); | |
7655 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7656 | |
7657 | /* do not patch jump label more than once per second */ | |
7658 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7659 | |
7660 | /* | |
7661 | * Build time assertion that we keep the data_head at the intended | |
7662 | * location. IOW, validation we got the __reserved[] size right. | |
7663 | */ | |
7664 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
7665 | != 1024); | |
0793a61d | 7666 | } |
abe43400 PZ |
7667 | |
7668 | static int __init perf_event_sysfs_init(void) | |
7669 | { | |
7670 | struct pmu *pmu; | |
7671 | int ret; | |
7672 | ||
7673 | mutex_lock(&pmus_lock); | |
7674 | ||
7675 | ret = bus_register(&pmu_bus); | |
7676 | if (ret) | |
7677 | goto unlock; | |
7678 | ||
7679 | list_for_each_entry(pmu, &pmus, entry) { | |
7680 | if (!pmu->name || pmu->type < 0) | |
7681 | continue; | |
7682 | ||
7683 | ret = pmu_dev_alloc(pmu); | |
7684 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
7685 | } | |
7686 | pmu_bus_running = 1; | |
7687 | ret = 0; | |
7688 | ||
7689 | unlock: | |
7690 | mutex_unlock(&pmus_lock); | |
7691 | ||
7692 | return ret; | |
7693 | } | |
7694 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
7695 | |
7696 | #ifdef CONFIG_CGROUP_PERF | |
92fb9748 | 7697 | static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont) |
e5d1367f SE |
7698 | { |
7699 | struct perf_cgroup *jc; | |
e5d1367f | 7700 | |
1b15d055 | 7701 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
7702 | if (!jc) |
7703 | return ERR_PTR(-ENOMEM); | |
7704 | ||
e5d1367f SE |
7705 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7706 | if (!jc->info) { | |
7707 | kfree(jc); | |
7708 | return ERR_PTR(-ENOMEM); | |
7709 | } | |
7710 | ||
e5d1367f SE |
7711 | return &jc->css; |
7712 | } | |
7713 | ||
92fb9748 | 7714 | static void perf_cgroup_css_free(struct cgroup *cont) |
e5d1367f SE |
7715 | { |
7716 | struct perf_cgroup *jc; | |
7717 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), | |
7718 | struct perf_cgroup, css); | |
7719 | free_percpu(jc->info); | |
7720 | kfree(jc); | |
7721 | } | |
7722 | ||
7723 | static int __perf_cgroup_move(void *info) | |
7724 | { | |
7725 | struct task_struct *task = info; | |
7726 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
7727 | return 0; | |
7728 | } | |
7729 | ||
761b3ef5 | 7730 | static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
e5d1367f | 7731 | { |
bb9d97b6 TH |
7732 | struct task_struct *task; |
7733 | ||
7734 | cgroup_taskset_for_each(task, cgrp, tset) | |
7735 | task_function_call(task, __perf_cgroup_move, task); | |
e5d1367f SE |
7736 | } |
7737 | ||
761b3ef5 LZ |
7738 | static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, |
7739 | struct task_struct *task) | |
e5d1367f SE |
7740 | { |
7741 | /* | |
7742 | * cgroup_exit() is called in the copy_process() failure path. | |
7743 | * Ignore this case since the task hasn't ran yet, this avoids | |
7744 | * trying to poke a half freed task state from generic code. | |
7745 | */ | |
7746 | if (!(task->flags & PF_EXITING)) | |
7747 | return; | |
7748 | ||
bb9d97b6 | 7749 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7750 | } |
7751 | ||
7752 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
7753 | .name = "perf_event", |
7754 | .subsys_id = perf_subsys_id, | |
92fb9748 TH |
7755 | .css_alloc = perf_cgroup_css_alloc, |
7756 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 7757 | .exit = perf_cgroup_exit, |
bb9d97b6 | 7758 | .attach = perf_cgroup_attach, |
e5d1367f SE |
7759 | }; |
7760 | #endif /* CONFIG_CGROUP_PERF */ |