]>
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 |\ | |
a21b0b35 YD |
122 | PERF_FLAG_PID_CGROUP |\ |
123 | PERF_FLAG_FD_CLOEXEC) | |
e5d1367f | 124 | |
bce38cd5 SE |
125 | /* |
126 | * branch priv levels that need permission checks | |
127 | */ | |
128 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
129 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
130 | PERF_SAMPLE_BRANCH_HV) | |
131 | ||
0b3fcf17 SE |
132 | enum event_type_t { |
133 | EVENT_FLEXIBLE = 0x1, | |
134 | EVENT_PINNED = 0x2, | |
135 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
136 | }; | |
137 | ||
e5d1367f SE |
138 | /* |
139 | * perf_sched_events : >0 events exist | |
140 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
141 | */ | |
c5905afb | 142 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 143 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 144 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 145 | |
cdd6c482 IM |
146 | static atomic_t nr_mmap_events __read_mostly; |
147 | static atomic_t nr_comm_events __read_mostly; | |
148 | static atomic_t nr_task_events __read_mostly; | |
948b26b6 | 149 | static atomic_t nr_freq_events __read_mostly; |
9ee318a7 | 150 | |
108b02cf PZ |
151 | static LIST_HEAD(pmus); |
152 | static DEFINE_MUTEX(pmus_lock); | |
153 | static struct srcu_struct pmus_srcu; | |
154 | ||
0764771d | 155 | /* |
cdd6c482 | 156 | * perf event paranoia level: |
0fbdea19 IM |
157 | * -1 - not paranoid at all |
158 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 159 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 160 | * 2 - disallow kernel profiling for unpriv |
0764771d | 161 | */ |
cdd6c482 | 162 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 163 | |
20443384 FW |
164 | /* Minimum for 512 kiB + 1 user control page */ |
165 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
166 | |
167 | /* | |
cdd6c482 | 168 | * max perf event sample rate |
df58ab24 | 169 | */ |
14c63f17 DH |
170 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
171 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
172 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
173 | ||
174 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
175 | ||
176 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
177 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
178 | ||
d9494cb4 PZ |
179 | static int perf_sample_allowed_ns __read_mostly = |
180 | DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100; | |
14c63f17 DH |
181 | |
182 | void update_perf_cpu_limits(void) | |
183 | { | |
184 | u64 tmp = perf_sample_period_ns; | |
185 | ||
186 | tmp *= sysctl_perf_cpu_time_max_percent; | |
e5302920 | 187 | do_div(tmp, 100); |
d9494cb4 | 188 | ACCESS_ONCE(perf_sample_allowed_ns) = tmp; |
14c63f17 | 189 | } |
163ec435 | 190 | |
9e630205 SE |
191 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
192 | ||
163ec435 PZ |
193 | int perf_proc_update_handler(struct ctl_table *table, int write, |
194 | void __user *buffer, size_t *lenp, | |
195 | loff_t *ppos) | |
196 | { | |
723478c8 | 197 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
163ec435 PZ |
198 | |
199 | if (ret || !write) | |
200 | return ret; | |
201 | ||
202 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
203 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
204 | update_perf_cpu_limits(); | |
205 | ||
206 | return 0; | |
207 | } | |
208 | ||
209 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
210 | ||
211 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
212 | void __user *buffer, size_t *lenp, | |
213 | loff_t *ppos) | |
214 | { | |
215 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
216 | ||
217 | if (ret || !write) | |
218 | return ret; | |
219 | ||
220 | update_perf_cpu_limits(); | |
163ec435 PZ |
221 | |
222 | return 0; | |
223 | } | |
1ccd1549 | 224 | |
14c63f17 DH |
225 | /* |
226 | * perf samples are done in some very critical code paths (NMIs). | |
227 | * If they take too much CPU time, the system can lock up and not | |
228 | * get any real work done. This will drop the sample rate when | |
229 | * we detect that events are taking too long. | |
230 | */ | |
231 | #define NR_ACCUMULATED_SAMPLES 128 | |
d9494cb4 | 232 | static DEFINE_PER_CPU(u64, running_sample_length); |
14c63f17 | 233 | |
6a02ad66 | 234 | static void perf_duration_warn(struct irq_work *w) |
14c63f17 | 235 | { |
6a02ad66 | 236 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
14c63f17 | 237 | u64 avg_local_sample_len; |
e5302920 | 238 | u64 local_samples_len; |
6a02ad66 PZ |
239 | |
240 | local_samples_len = __get_cpu_var(running_sample_length); | |
241 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
242 | ||
243 | printk_ratelimited(KERN_WARNING | |
244 | "perf interrupt took too long (%lld > %lld), lowering " | |
245 | "kernel.perf_event_max_sample_rate to %d\n", | |
246 | avg_local_sample_len, allowed_ns, | |
247 | sysctl_perf_event_sample_rate); | |
248 | } | |
249 | ||
250 | static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn); | |
251 | ||
252 | void perf_sample_event_took(u64 sample_len_ns) | |
253 | { | |
d9494cb4 | 254 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
6a02ad66 PZ |
255 | u64 avg_local_sample_len; |
256 | u64 local_samples_len; | |
14c63f17 | 257 | |
d9494cb4 | 258 | if (allowed_ns == 0) |
14c63f17 DH |
259 | return; |
260 | ||
261 | /* decay the counter by 1 average sample */ | |
262 | local_samples_len = __get_cpu_var(running_sample_length); | |
263 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; | |
264 | local_samples_len += sample_len_ns; | |
265 | __get_cpu_var(running_sample_length) = local_samples_len; | |
266 | ||
267 | /* | |
268 | * note: this will be biased artifically low until we have | |
269 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
270 | * from having to maintain a count. | |
271 | */ | |
272 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
273 | ||
d9494cb4 | 274 | if (avg_local_sample_len <= allowed_ns) |
14c63f17 DH |
275 | return; |
276 | ||
277 | if (max_samples_per_tick <= 1) | |
278 | return; | |
279 | ||
280 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
281 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
282 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
283 | ||
14c63f17 | 284 | update_perf_cpu_limits(); |
6a02ad66 PZ |
285 | |
286 | irq_work_queue(&perf_duration_work); | |
14c63f17 DH |
287 | } |
288 | ||
cdd6c482 | 289 | static atomic64_t perf_event_id; |
a96bbc16 | 290 | |
0b3fcf17 SE |
291 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
292 | enum event_type_t event_type); | |
293 | ||
294 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
295 | enum event_type_t event_type, |
296 | struct task_struct *task); | |
297 | ||
298 | static void update_context_time(struct perf_event_context *ctx); | |
299 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 300 | |
cdd6c482 | 301 | void __weak perf_event_print_debug(void) { } |
0793a61d | 302 | |
84c79910 | 303 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 304 | { |
84c79910 | 305 | return "pmu"; |
0793a61d TG |
306 | } |
307 | ||
0b3fcf17 SE |
308 | static inline u64 perf_clock(void) |
309 | { | |
310 | return local_clock(); | |
311 | } | |
312 | ||
e5d1367f SE |
313 | static inline struct perf_cpu_context * |
314 | __get_cpu_context(struct perf_event_context *ctx) | |
315 | { | |
316 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
317 | } | |
318 | ||
facc4307 PZ |
319 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
320 | struct perf_event_context *ctx) | |
321 | { | |
322 | raw_spin_lock(&cpuctx->ctx.lock); | |
323 | if (ctx) | |
324 | raw_spin_lock(&ctx->lock); | |
325 | } | |
326 | ||
327 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
328 | struct perf_event_context *ctx) | |
329 | { | |
330 | if (ctx) | |
331 | raw_spin_unlock(&ctx->lock); | |
332 | raw_spin_unlock(&cpuctx->ctx.lock); | |
333 | } | |
334 | ||
e5d1367f SE |
335 | #ifdef CONFIG_CGROUP_PERF |
336 | ||
877c6856 LZ |
337 | /* |
338 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
339 | * This is a per-cpu dynamically allocated data structure. | |
340 | */ | |
341 | struct perf_cgroup_info { | |
342 | u64 time; | |
343 | u64 timestamp; | |
344 | }; | |
345 | ||
346 | struct perf_cgroup { | |
347 | struct cgroup_subsys_state css; | |
86e213e1 | 348 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
349 | }; |
350 | ||
3f7cce3c SE |
351 | /* |
352 | * Must ensure cgroup is pinned (css_get) before calling | |
353 | * this function. In other words, we cannot call this function | |
354 | * if there is no cgroup event for the current CPU context. | |
355 | */ | |
e5d1367f SE |
356 | static inline struct perf_cgroup * |
357 | perf_cgroup_from_task(struct task_struct *task) | |
358 | { | |
8af01f56 TH |
359 | return container_of(task_css(task, perf_subsys_id), |
360 | struct perf_cgroup, css); | |
e5d1367f SE |
361 | } |
362 | ||
363 | static inline bool | |
364 | perf_cgroup_match(struct perf_event *event) | |
365 | { | |
366 | struct perf_event_context *ctx = event->ctx; | |
367 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
368 | ||
ef824fa1 TH |
369 | /* @event doesn't care about cgroup */ |
370 | if (!event->cgrp) | |
371 | return true; | |
372 | ||
373 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
374 | if (!cpuctx->cgrp) | |
375 | return false; | |
376 | ||
377 | /* | |
378 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
379 | * also enabled for all its descendant cgroups. If @cpuctx's | |
380 | * cgroup is a descendant of @event's (the test covers identity | |
381 | * case), it's a match. | |
382 | */ | |
383 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
384 | event->cgrp->css.cgroup); | |
e5d1367f SE |
385 | } |
386 | ||
9c5da09d | 387 | static inline bool perf_tryget_cgroup(struct perf_event *event) |
e5d1367f | 388 | { |
9c5da09d | 389 | return css_tryget(&event->cgrp->css); |
e5d1367f SE |
390 | } |
391 | ||
392 | static inline void perf_put_cgroup(struct perf_event *event) | |
393 | { | |
394 | css_put(&event->cgrp->css); | |
395 | } | |
396 | ||
397 | static inline void perf_detach_cgroup(struct perf_event *event) | |
398 | { | |
399 | perf_put_cgroup(event); | |
400 | event->cgrp = NULL; | |
401 | } | |
402 | ||
403 | static inline int is_cgroup_event(struct perf_event *event) | |
404 | { | |
405 | return event->cgrp != NULL; | |
406 | } | |
407 | ||
408 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
409 | { | |
410 | struct perf_cgroup_info *t; | |
411 | ||
412 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
413 | return t->time; | |
414 | } | |
415 | ||
416 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
417 | { | |
418 | struct perf_cgroup_info *info; | |
419 | u64 now; | |
420 | ||
421 | now = perf_clock(); | |
422 | ||
423 | info = this_cpu_ptr(cgrp->info); | |
424 | ||
425 | info->time += now - info->timestamp; | |
426 | info->timestamp = now; | |
427 | } | |
428 | ||
429 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
430 | { | |
431 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
432 | if (cgrp_out) | |
433 | __update_cgrp_time(cgrp_out); | |
434 | } | |
435 | ||
436 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
437 | { | |
3f7cce3c SE |
438 | struct perf_cgroup *cgrp; |
439 | ||
e5d1367f | 440 | /* |
3f7cce3c SE |
441 | * ensure we access cgroup data only when needed and |
442 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 443 | */ |
3f7cce3c | 444 | if (!is_cgroup_event(event)) |
e5d1367f SE |
445 | return; |
446 | ||
3f7cce3c SE |
447 | cgrp = perf_cgroup_from_task(current); |
448 | /* | |
449 | * Do not update time when cgroup is not active | |
450 | */ | |
451 | if (cgrp == event->cgrp) | |
452 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
453 | } |
454 | ||
455 | static inline void | |
3f7cce3c SE |
456 | perf_cgroup_set_timestamp(struct task_struct *task, |
457 | struct perf_event_context *ctx) | |
e5d1367f SE |
458 | { |
459 | struct perf_cgroup *cgrp; | |
460 | struct perf_cgroup_info *info; | |
461 | ||
3f7cce3c SE |
462 | /* |
463 | * ctx->lock held by caller | |
464 | * ensure we do not access cgroup data | |
465 | * unless we have the cgroup pinned (css_get) | |
466 | */ | |
467 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
468 | return; |
469 | ||
470 | cgrp = perf_cgroup_from_task(task); | |
471 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 472 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
473 | } |
474 | ||
475 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
476 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
477 | ||
478 | /* | |
479 | * reschedule events based on the cgroup constraint of task. | |
480 | * | |
481 | * mode SWOUT : schedule out everything | |
482 | * mode SWIN : schedule in based on cgroup for next | |
483 | */ | |
484 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
485 | { | |
486 | struct perf_cpu_context *cpuctx; | |
487 | struct pmu *pmu; | |
488 | unsigned long flags; | |
489 | ||
490 | /* | |
491 | * disable interrupts to avoid geting nr_cgroup | |
492 | * changes via __perf_event_disable(). Also | |
493 | * avoids preemption. | |
494 | */ | |
495 | local_irq_save(flags); | |
496 | ||
497 | /* | |
498 | * we reschedule only in the presence of cgroup | |
499 | * constrained events. | |
500 | */ | |
501 | rcu_read_lock(); | |
502 | ||
503 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 504 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
505 | if (cpuctx->unique_pmu != pmu) |
506 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 507 | |
e5d1367f SE |
508 | /* |
509 | * perf_cgroup_events says at least one | |
510 | * context on this CPU has cgroup events. | |
511 | * | |
512 | * ctx->nr_cgroups reports the number of cgroup | |
513 | * events for a context. | |
514 | */ | |
515 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
516 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
517 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
518 | |
519 | if (mode & PERF_CGROUP_SWOUT) { | |
520 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
521 | /* | |
522 | * must not be done before ctxswout due | |
523 | * to event_filter_match() in event_sched_out() | |
524 | */ | |
525 | cpuctx->cgrp = NULL; | |
526 | } | |
527 | ||
528 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 529 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
530 | /* |
531 | * set cgrp before ctxsw in to allow | |
532 | * event_filter_match() to not have to pass | |
533 | * task around | |
e5d1367f SE |
534 | */ |
535 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
536 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
537 | } | |
facc4307 PZ |
538 | perf_pmu_enable(cpuctx->ctx.pmu); |
539 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 540 | } |
e5d1367f SE |
541 | } |
542 | ||
543 | rcu_read_unlock(); | |
544 | ||
545 | local_irq_restore(flags); | |
546 | } | |
547 | ||
a8d757ef SE |
548 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
549 | struct task_struct *next) | |
e5d1367f | 550 | { |
a8d757ef SE |
551 | struct perf_cgroup *cgrp1; |
552 | struct perf_cgroup *cgrp2 = NULL; | |
553 | ||
554 | /* | |
555 | * we come here when we know perf_cgroup_events > 0 | |
556 | */ | |
557 | cgrp1 = perf_cgroup_from_task(task); | |
558 | ||
559 | /* | |
560 | * next is NULL when called from perf_event_enable_on_exec() | |
561 | * that will systematically cause a cgroup_switch() | |
562 | */ | |
563 | if (next) | |
564 | cgrp2 = perf_cgroup_from_task(next); | |
565 | ||
566 | /* | |
567 | * only schedule out current cgroup events if we know | |
568 | * that we are switching to a different cgroup. Otherwise, | |
569 | * do no touch the cgroup events. | |
570 | */ | |
571 | if (cgrp1 != cgrp2) | |
572 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
573 | } |
574 | ||
a8d757ef SE |
575 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
576 | struct task_struct *task) | |
e5d1367f | 577 | { |
a8d757ef SE |
578 | struct perf_cgroup *cgrp1; |
579 | struct perf_cgroup *cgrp2 = NULL; | |
580 | ||
581 | /* | |
582 | * we come here when we know perf_cgroup_events > 0 | |
583 | */ | |
584 | cgrp1 = perf_cgroup_from_task(task); | |
585 | ||
586 | /* prev can never be NULL */ | |
587 | cgrp2 = perf_cgroup_from_task(prev); | |
588 | ||
589 | /* | |
590 | * only need to schedule in cgroup events if we are changing | |
591 | * cgroup during ctxsw. Cgroup events were not scheduled | |
592 | * out of ctxsw out if that was not the case. | |
593 | */ | |
594 | if (cgrp1 != cgrp2) | |
595 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
596 | } |
597 | ||
598 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
599 | struct perf_event_attr *attr, | |
600 | struct perf_event *group_leader) | |
601 | { | |
602 | struct perf_cgroup *cgrp; | |
603 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
604 | struct fd f = fdget(fd); |
605 | int ret = 0; | |
e5d1367f | 606 | |
2903ff01 | 607 | if (!f.file) |
e5d1367f SE |
608 | return -EBADF; |
609 | ||
b77d7b60 TH |
610 | rcu_read_lock(); |
611 | ||
35cf0836 | 612 | css = css_from_dir(f.file->f_dentry, &perf_subsys); |
3db272c0 LZ |
613 | if (IS_ERR(css)) { |
614 | ret = PTR_ERR(css); | |
615 | goto out; | |
616 | } | |
e5d1367f SE |
617 | |
618 | cgrp = container_of(css, struct perf_cgroup, css); | |
619 | event->cgrp = cgrp; | |
620 | ||
f75e18cb | 621 | /* must be done before we fput() the file */ |
9c5da09d SQ |
622 | if (!perf_tryget_cgroup(event)) { |
623 | event->cgrp = NULL; | |
624 | ret = -ENOENT; | |
625 | goto out; | |
626 | } | |
f75e18cb | 627 | |
e5d1367f SE |
628 | /* |
629 | * all events in a group must monitor | |
630 | * the same cgroup because a task belongs | |
631 | * to only one perf cgroup at a time | |
632 | */ | |
633 | if (group_leader && group_leader->cgrp != cgrp) { | |
634 | perf_detach_cgroup(event); | |
635 | ret = -EINVAL; | |
e5d1367f | 636 | } |
3db272c0 | 637 | out: |
b77d7b60 | 638 | rcu_read_unlock(); |
2903ff01 | 639 | fdput(f); |
e5d1367f SE |
640 | return ret; |
641 | } | |
642 | ||
643 | static inline void | |
644 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
645 | { | |
646 | struct perf_cgroup_info *t; | |
647 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
648 | event->shadow_ctx_time = now - t->timestamp; | |
649 | } | |
650 | ||
651 | static inline void | |
652 | perf_cgroup_defer_enabled(struct perf_event *event) | |
653 | { | |
654 | /* | |
655 | * when the current task's perf cgroup does not match | |
656 | * the event's, we need to remember to call the | |
657 | * perf_mark_enable() function the first time a task with | |
658 | * a matching perf cgroup is scheduled in. | |
659 | */ | |
660 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
661 | event->cgrp_defer_enabled = 1; | |
662 | } | |
663 | ||
664 | static inline void | |
665 | perf_cgroup_mark_enabled(struct perf_event *event, | |
666 | struct perf_event_context *ctx) | |
667 | { | |
668 | struct perf_event *sub; | |
669 | u64 tstamp = perf_event_time(event); | |
670 | ||
671 | if (!event->cgrp_defer_enabled) | |
672 | return; | |
673 | ||
674 | event->cgrp_defer_enabled = 0; | |
675 | ||
676 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
677 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
678 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
679 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
680 | sub->cgrp_defer_enabled = 0; | |
681 | } | |
682 | } | |
683 | } | |
684 | #else /* !CONFIG_CGROUP_PERF */ | |
685 | ||
686 | static inline bool | |
687 | perf_cgroup_match(struct perf_event *event) | |
688 | { | |
689 | return true; | |
690 | } | |
691 | ||
692 | static inline void perf_detach_cgroup(struct perf_event *event) | |
693 | {} | |
694 | ||
695 | static inline int is_cgroup_event(struct perf_event *event) | |
696 | { | |
697 | return 0; | |
698 | } | |
699 | ||
700 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
701 | { | |
702 | return 0; | |
703 | } | |
704 | ||
705 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
706 | { | |
707 | } | |
708 | ||
709 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
710 | { | |
711 | } | |
712 | ||
a8d757ef SE |
713 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
714 | struct task_struct *next) | |
e5d1367f SE |
715 | { |
716 | } | |
717 | ||
a8d757ef SE |
718 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
719 | struct task_struct *task) | |
e5d1367f SE |
720 | { |
721 | } | |
722 | ||
723 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
724 | struct perf_event_attr *attr, | |
725 | struct perf_event *group_leader) | |
726 | { | |
727 | return -EINVAL; | |
728 | } | |
729 | ||
730 | static inline void | |
3f7cce3c SE |
731 | perf_cgroup_set_timestamp(struct task_struct *task, |
732 | struct perf_event_context *ctx) | |
e5d1367f SE |
733 | { |
734 | } | |
735 | ||
736 | void | |
737 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
738 | { | |
739 | } | |
740 | ||
741 | static inline void | |
742 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
743 | { | |
744 | } | |
745 | ||
746 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
747 | { | |
748 | return 0; | |
749 | } | |
750 | ||
751 | static inline void | |
752 | perf_cgroup_defer_enabled(struct perf_event *event) | |
753 | { | |
754 | } | |
755 | ||
756 | static inline void | |
757 | perf_cgroup_mark_enabled(struct perf_event *event, | |
758 | struct perf_event_context *ctx) | |
759 | { | |
760 | } | |
761 | #endif | |
762 | ||
9e630205 SE |
763 | /* |
764 | * set default to be dependent on timer tick just | |
765 | * like original code | |
766 | */ | |
767 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
768 | /* | |
769 | * function must be called with interrupts disbled | |
770 | */ | |
771 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
772 | { | |
773 | struct perf_cpu_context *cpuctx; | |
774 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
775 | int rotations = 0; | |
776 | ||
777 | WARN_ON(!irqs_disabled()); | |
778 | ||
779 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
780 | ||
781 | rotations = perf_rotate_context(cpuctx); | |
782 | ||
783 | /* | |
784 | * arm timer if needed | |
785 | */ | |
786 | if (rotations) { | |
787 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
788 | ret = HRTIMER_RESTART; | |
789 | } | |
790 | ||
791 | return ret; | |
792 | } | |
793 | ||
794 | /* CPU is going down */ | |
795 | void perf_cpu_hrtimer_cancel(int cpu) | |
796 | { | |
797 | struct perf_cpu_context *cpuctx; | |
798 | struct pmu *pmu; | |
799 | unsigned long flags; | |
800 | ||
801 | if (WARN_ON(cpu != smp_processor_id())) | |
802 | return; | |
803 | ||
804 | local_irq_save(flags); | |
805 | ||
806 | rcu_read_lock(); | |
807 | ||
808 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
809 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
810 | ||
811 | if (pmu->task_ctx_nr == perf_sw_context) | |
812 | continue; | |
813 | ||
814 | hrtimer_cancel(&cpuctx->hrtimer); | |
815 | } | |
816 | ||
817 | rcu_read_unlock(); | |
818 | ||
819 | local_irq_restore(flags); | |
820 | } | |
821 | ||
822 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
823 | { | |
824 | struct hrtimer *hr = &cpuctx->hrtimer; | |
825 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 826 | int timer; |
9e630205 SE |
827 | |
828 | /* no multiplexing needed for SW PMU */ | |
829 | if (pmu->task_ctx_nr == perf_sw_context) | |
830 | return; | |
831 | ||
62b85639 SE |
832 | /* |
833 | * check default is sane, if not set then force to | |
834 | * default interval (1/tick) | |
835 | */ | |
836 | timer = pmu->hrtimer_interval_ms; | |
837 | if (timer < 1) | |
838 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
839 | ||
840 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
841 | |
842 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
843 | hr->function = perf_cpu_hrtimer_handler; | |
844 | } | |
845 | ||
846 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
847 | { | |
848 | struct hrtimer *hr = &cpuctx->hrtimer; | |
849 | struct pmu *pmu = cpuctx->ctx.pmu; | |
850 | ||
851 | /* not for SW PMU */ | |
852 | if (pmu->task_ctx_nr == perf_sw_context) | |
853 | return; | |
854 | ||
855 | if (hrtimer_active(hr)) | |
856 | return; | |
857 | ||
858 | if (!hrtimer_callback_running(hr)) | |
859 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
860 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
861 | } | |
862 | ||
33696fc0 | 863 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 864 | { |
33696fc0 PZ |
865 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
866 | if (!(*count)++) | |
867 | pmu->pmu_disable(pmu); | |
9e35ad38 | 868 | } |
9e35ad38 | 869 | |
33696fc0 | 870 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 871 | { |
33696fc0 PZ |
872 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
873 | if (!--(*count)) | |
874 | pmu->pmu_enable(pmu); | |
9e35ad38 | 875 | } |
9e35ad38 | 876 | |
e9d2b064 PZ |
877 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
878 | ||
879 | /* | |
880 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
881 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
882 | * disabled, while rotate_context is called from IRQ context. | |
883 | */ | |
108b02cf | 884 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 885 | { |
108b02cf | 886 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 887 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 888 | |
e9d2b064 | 889 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 890 | |
d84153d6 | 891 | if (list_empty(&cpuctx->rotation_list)) |
e9d2b064 | 892 | list_add(&cpuctx->rotation_list, head); |
9e35ad38 | 893 | } |
9e35ad38 | 894 | |
cdd6c482 | 895 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 896 | { |
e5289d4a | 897 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
898 | } |
899 | ||
cdd6c482 | 900 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 901 | { |
564c2b21 PM |
902 | if (atomic_dec_and_test(&ctx->refcount)) { |
903 | if (ctx->parent_ctx) | |
904 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
905 | if (ctx->task) |
906 | put_task_struct(ctx->task); | |
cb796ff3 | 907 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 908 | } |
a63eaf34 PM |
909 | } |
910 | ||
cdd6c482 | 911 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
912 | { |
913 | if (ctx->parent_ctx) { | |
914 | put_ctx(ctx->parent_ctx); | |
915 | ctx->parent_ctx = NULL; | |
916 | } | |
5a3126d4 | 917 | ctx->generation++; |
71a851b4 PZ |
918 | } |
919 | ||
6844c09d ACM |
920 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
921 | { | |
922 | /* | |
923 | * only top level events have the pid namespace they were created in | |
924 | */ | |
925 | if (event->parent) | |
926 | event = event->parent; | |
927 | ||
928 | return task_tgid_nr_ns(p, event->ns); | |
929 | } | |
930 | ||
931 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
932 | { | |
933 | /* | |
934 | * only top level events have the pid namespace they were created in | |
935 | */ | |
936 | if (event->parent) | |
937 | event = event->parent; | |
938 | ||
939 | return task_pid_nr_ns(p, event->ns); | |
940 | } | |
941 | ||
7f453c24 | 942 | /* |
cdd6c482 | 943 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
944 | * to userspace. |
945 | */ | |
cdd6c482 | 946 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 947 | { |
cdd6c482 | 948 | u64 id = event->id; |
7f453c24 | 949 | |
cdd6c482 IM |
950 | if (event->parent) |
951 | id = event->parent->id; | |
7f453c24 PZ |
952 | |
953 | return id; | |
954 | } | |
955 | ||
25346b93 | 956 | /* |
cdd6c482 | 957 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
958 | * This has to cope with with the fact that until it is locked, |
959 | * the context could get moved to another task. | |
960 | */ | |
cdd6c482 | 961 | static struct perf_event_context * |
8dc85d54 | 962 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 963 | { |
cdd6c482 | 964 | struct perf_event_context *ctx; |
25346b93 | 965 | |
9ed6060d | 966 | retry: |
058ebd0e PZ |
967 | /* |
968 | * One of the few rules of preemptible RCU is that one cannot do | |
969 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
970 | * part of the read side critical section was preemptible -- see | |
971 | * rcu_read_unlock_special(). | |
972 | * | |
973 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
974 | * side critical section is non-preemptible. | |
975 | */ | |
976 | preempt_disable(); | |
977 | rcu_read_lock(); | |
8dc85d54 | 978 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
979 | if (ctx) { |
980 | /* | |
981 | * If this context is a clone of another, it might | |
982 | * get swapped for another underneath us by | |
cdd6c482 | 983 | * perf_event_task_sched_out, though the |
25346b93 PM |
984 | * rcu_read_lock() protects us from any context |
985 | * getting freed. Lock the context and check if it | |
986 | * got swapped before we could get the lock, and retry | |
987 | * if so. If we locked the right context, then it | |
988 | * can't get swapped on us any more. | |
989 | */ | |
e625cce1 | 990 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 991 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 992 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
993 | rcu_read_unlock(); |
994 | preempt_enable(); | |
25346b93 PM |
995 | goto retry; |
996 | } | |
b49a9e7e PZ |
997 | |
998 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 999 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
1000 | ctx = NULL; |
1001 | } | |
25346b93 PM |
1002 | } |
1003 | rcu_read_unlock(); | |
058ebd0e | 1004 | preempt_enable(); |
25346b93 PM |
1005 | return ctx; |
1006 | } | |
1007 | ||
1008 | /* | |
1009 | * Get the context for a task and increment its pin_count so it | |
1010 | * can't get swapped to another task. This also increments its | |
1011 | * reference count so that the context can't get freed. | |
1012 | */ | |
8dc85d54 PZ |
1013 | static struct perf_event_context * |
1014 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 1015 | { |
cdd6c482 | 1016 | struct perf_event_context *ctx; |
25346b93 PM |
1017 | unsigned long flags; |
1018 | ||
8dc85d54 | 1019 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1020 | if (ctx) { |
1021 | ++ctx->pin_count; | |
e625cce1 | 1022 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1023 | } |
1024 | return ctx; | |
1025 | } | |
1026 | ||
cdd6c482 | 1027 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1028 | { |
1029 | unsigned long flags; | |
1030 | ||
e625cce1 | 1031 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1032 | --ctx->pin_count; |
e625cce1 | 1033 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1034 | } |
1035 | ||
f67218c3 PZ |
1036 | /* |
1037 | * Update the record of the current time in a context. | |
1038 | */ | |
1039 | static void update_context_time(struct perf_event_context *ctx) | |
1040 | { | |
1041 | u64 now = perf_clock(); | |
1042 | ||
1043 | ctx->time += now - ctx->timestamp; | |
1044 | ctx->timestamp = now; | |
1045 | } | |
1046 | ||
4158755d SE |
1047 | static u64 perf_event_time(struct perf_event *event) |
1048 | { | |
1049 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1050 | |
1051 | if (is_cgroup_event(event)) | |
1052 | return perf_cgroup_event_time(event); | |
1053 | ||
4158755d SE |
1054 | return ctx ? ctx->time : 0; |
1055 | } | |
1056 | ||
f67218c3 PZ |
1057 | /* |
1058 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1059 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1060 | */ |
1061 | static void update_event_times(struct perf_event *event) | |
1062 | { | |
1063 | struct perf_event_context *ctx = event->ctx; | |
1064 | u64 run_end; | |
1065 | ||
1066 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1067 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1068 | return; | |
e5d1367f SE |
1069 | /* |
1070 | * in cgroup mode, time_enabled represents | |
1071 | * the time the event was enabled AND active | |
1072 | * tasks were in the monitored cgroup. This is | |
1073 | * independent of the activity of the context as | |
1074 | * there may be a mix of cgroup and non-cgroup events. | |
1075 | * | |
1076 | * That is why we treat cgroup events differently | |
1077 | * here. | |
1078 | */ | |
1079 | if (is_cgroup_event(event)) | |
46cd6a7f | 1080 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1081 | else if (ctx->is_active) |
1082 | run_end = ctx->time; | |
acd1d7c1 PZ |
1083 | else |
1084 | run_end = event->tstamp_stopped; | |
1085 | ||
1086 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1087 | |
1088 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1089 | run_end = event->tstamp_stopped; | |
1090 | else | |
4158755d | 1091 | run_end = perf_event_time(event); |
f67218c3 PZ |
1092 | |
1093 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1094 | |
f67218c3 PZ |
1095 | } |
1096 | ||
96c21a46 PZ |
1097 | /* |
1098 | * Update total_time_enabled and total_time_running for all events in a group. | |
1099 | */ | |
1100 | static void update_group_times(struct perf_event *leader) | |
1101 | { | |
1102 | struct perf_event *event; | |
1103 | ||
1104 | update_event_times(leader); | |
1105 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1106 | update_event_times(event); | |
1107 | } | |
1108 | ||
889ff015 FW |
1109 | static struct list_head * |
1110 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1111 | { | |
1112 | if (event->attr.pinned) | |
1113 | return &ctx->pinned_groups; | |
1114 | else | |
1115 | return &ctx->flexible_groups; | |
1116 | } | |
1117 | ||
fccc714b | 1118 | /* |
cdd6c482 | 1119 | * Add a event from the lists for its context. |
fccc714b PZ |
1120 | * Must be called with ctx->mutex and ctx->lock held. |
1121 | */ | |
04289bb9 | 1122 | static void |
cdd6c482 | 1123 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1124 | { |
8a49542c PZ |
1125 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1126 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1127 | |
1128 | /* | |
8a49542c PZ |
1129 | * If we're a stand alone event or group leader, we go to the context |
1130 | * list, group events are kept attached to the group so that | |
1131 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1132 | */ |
8a49542c | 1133 | if (event->group_leader == event) { |
889ff015 FW |
1134 | struct list_head *list; |
1135 | ||
d6f962b5 FW |
1136 | if (is_software_event(event)) |
1137 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1138 | ||
889ff015 FW |
1139 | list = ctx_group_list(event, ctx); |
1140 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1141 | } |
592903cd | 1142 | |
08309379 | 1143 | if (is_cgroup_event(event)) |
e5d1367f | 1144 | ctx->nr_cgroups++; |
e5d1367f | 1145 | |
d010b332 SE |
1146 | if (has_branch_stack(event)) |
1147 | ctx->nr_branch_stack++; | |
1148 | ||
cdd6c482 | 1149 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1150 | if (!ctx->nr_events) |
108b02cf | 1151 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1152 | ctx->nr_events++; |
1153 | if (event->attr.inherit_stat) | |
bfbd3381 | 1154 | ctx->nr_stat++; |
5a3126d4 PZ |
1155 | |
1156 | ctx->generation++; | |
04289bb9 IM |
1157 | } |
1158 | ||
0231bb53 JO |
1159 | /* |
1160 | * Initialize event state based on the perf_event_attr::disabled. | |
1161 | */ | |
1162 | static inline void perf_event__state_init(struct perf_event *event) | |
1163 | { | |
1164 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1165 | PERF_EVENT_STATE_INACTIVE; | |
1166 | } | |
1167 | ||
c320c7b7 ACM |
1168 | /* |
1169 | * Called at perf_event creation and when events are attached/detached from a | |
1170 | * group. | |
1171 | */ | |
1172 | static void perf_event__read_size(struct perf_event *event) | |
1173 | { | |
1174 | int entry = sizeof(u64); /* value */ | |
1175 | int size = 0; | |
1176 | int nr = 1; | |
1177 | ||
1178 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1179 | size += sizeof(u64); | |
1180 | ||
1181 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1182 | size += sizeof(u64); | |
1183 | ||
1184 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1185 | entry += sizeof(u64); | |
1186 | ||
1187 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1188 | nr += event->group_leader->nr_siblings; | |
1189 | size += sizeof(u64); | |
1190 | } | |
1191 | ||
1192 | size += entry * nr; | |
1193 | event->read_size = size; | |
1194 | } | |
1195 | ||
1196 | static void perf_event__header_size(struct perf_event *event) | |
1197 | { | |
1198 | struct perf_sample_data *data; | |
1199 | u64 sample_type = event->attr.sample_type; | |
1200 | u16 size = 0; | |
1201 | ||
1202 | perf_event__read_size(event); | |
1203 | ||
1204 | if (sample_type & PERF_SAMPLE_IP) | |
1205 | size += sizeof(data->ip); | |
1206 | ||
6844c09d ACM |
1207 | if (sample_type & PERF_SAMPLE_ADDR) |
1208 | size += sizeof(data->addr); | |
1209 | ||
1210 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1211 | size += sizeof(data->period); | |
1212 | ||
c3feedf2 AK |
1213 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1214 | size += sizeof(data->weight); | |
1215 | ||
6844c09d ACM |
1216 | if (sample_type & PERF_SAMPLE_READ) |
1217 | size += event->read_size; | |
1218 | ||
d6be9ad6 SE |
1219 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1220 | size += sizeof(data->data_src.val); | |
1221 | ||
fdfbbd07 AK |
1222 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
1223 | size += sizeof(data->txn); | |
1224 | ||
6844c09d ACM |
1225 | event->header_size = size; |
1226 | } | |
1227 | ||
1228 | static void perf_event__id_header_size(struct perf_event *event) | |
1229 | { | |
1230 | struct perf_sample_data *data; | |
1231 | u64 sample_type = event->attr.sample_type; | |
1232 | u16 size = 0; | |
1233 | ||
c320c7b7 ACM |
1234 | if (sample_type & PERF_SAMPLE_TID) |
1235 | size += sizeof(data->tid_entry); | |
1236 | ||
1237 | if (sample_type & PERF_SAMPLE_TIME) | |
1238 | size += sizeof(data->time); | |
1239 | ||
ff3d527c AH |
1240 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1241 | size += sizeof(data->id); | |
1242 | ||
c320c7b7 ACM |
1243 | if (sample_type & PERF_SAMPLE_ID) |
1244 | size += sizeof(data->id); | |
1245 | ||
1246 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1247 | size += sizeof(data->stream_id); | |
1248 | ||
1249 | if (sample_type & PERF_SAMPLE_CPU) | |
1250 | size += sizeof(data->cpu_entry); | |
1251 | ||
6844c09d | 1252 | event->id_header_size = size; |
c320c7b7 ACM |
1253 | } |
1254 | ||
8a49542c PZ |
1255 | static void perf_group_attach(struct perf_event *event) |
1256 | { | |
c320c7b7 | 1257 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1258 | |
74c3337c PZ |
1259 | /* |
1260 | * We can have double attach due to group movement in perf_event_open. | |
1261 | */ | |
1262 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1263 | return; | |
1264 | ||
8a49542c PZ |
1265 | event->attach_state |= PERF_ATTACH_GROUP; |
1266 | ||
1267 | if (group_leader == event) | |
1268 | return; | |
1269 | ||
1270 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1271 | !is_software_event(event)) | |
1272 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1273 | ||
1274 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1275 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1276 | |
1277 | perf_event__header_size(group_leader); | |
1278 | ||
1279 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1280 | perf_event__header_size(pos); | |
8a49542c PZ |
1281 | } |
1282 | ||
a63eaf34 | 1283 | /* |
cdd6c482 | 1284 | * Remove a event from the lists for its context. |
fccc714b | 1285 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1286 | */ |
04289bb9 | 1287 | static void |
cdd6c482 | 1288 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1289 | { |
68cacd29 | 1290 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1291 | /* |
1292 | * We can have double detach due to exit/hot-unplug + close. | |
1293 | */ | |
1294 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1295 | return; |
8a49542c PZ |
1296 | |
1297 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1298 | ||
68cacd29 | 1299 | if (is_cgroup_event(event)) { |
e5d1367f | 1300 | ctx->nr_cgroups--; |
68cacd29 SE |
1301 | cpuctx = __get_cpu_context(ctx); |
1302 | /* | |
1303 | * if there are no more cgroup events | |
1304 | * then cler cgrp to avoid stale pointer | |
1305 | * in update_cgrp_time_from_cpuctx() | |
1306 | */ | |
1307 | if (!ctx->nr_cgroups) | |
1308 | cpuctx->cgrp = NULL; | |
1309 | } | |
e5d1367f | 1310 | |
d010b332 SE |
1311 | if (has_branch_stack(event)) |
1312 | ctx->nr_branch_stack--; | |
1313 | ||
cdd6c482 IM |
1314 | ctx->nr_events--; |
1315 | if (event->attr.inherit_stat) | |
bfbd3381 | 1316 | ctx->nr_stat--; |
8bc20959 | 1317 | |
cdd6c482 | 1318 | list_del_rcu(&event->event_entry); |
04289bb9 | 1319 | |
8a49542c PZ |
1320 | if (event->group_leader == event) |
1321 | list_del_init(&event->group_entry); | |
5c148194 | 1322 | |
96c21a46 | 1323 | update_group_times(event); |
b2e74a26 SE |
1324 | |
1325 | /* | |
1326 | * If event was in error state, then keep it | |
1327 | * that way, otherwise bogus counts will be | |
1328 | * returned on read(). The only way to get out | |
1329 | * of error state is by explicit re-enabling | |
1330 | * of the event | |
1331 | */ | |
1332 | if (event->state > PERF_EVENT_STATE_OFF) | |
1333 | event->state = PERF_EVENT_STATE_OFF; | |
5a3126d4 PZ |
1334 | |
1335 | ctx->generation++; | |
050735b0 PZ |
1336 | } |
1337 | ||
8a49542c | 1338 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1339 | { |
1340 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1341 | struct list_head *list = NULL; |
1342 | ||
1343 | /* | |
1344 | * We can have double detach due to exit/hot-unplug + close. | |
1345 | */ | |
1346 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1347 | return; | |
1348 | ||
1349 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1350 | ||
1351 | /* | |
1352 | * If this is a sibling, remove it from its group. | |
1353 | */ | |
1354 | if (event->group_leader != event) { | |
1355 | list_del_init(&event->group_entry); | |
1356 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1357 | goto out; |
8a49542c PZ |
1358 | } |
1359 | ||
1360 | if (!list_empty(&event->group_entry)) | |
1361 | list = &event->group_entry; | |
2e2af50b | 1362 | |
04289bb9 | 1363 | /* |
cdd6c482 IM |
1364 | * If this was a group event with sibling events then |
1365 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1366 | * to whatever list we are on. |
04289bb9 | 1367 | */ |
cdd6c482 | 1368 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1369 | if (list) |
1370 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1371 | sibling->group_leader = sibling; |
d6f962b5 FW |
1372 | |
1373 | /* Inherit group flags from the previous leader */ | |
1374 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1375 | } |
c320c7b7 ACM |
1376 | |
1377 | out: | |
1378 | perf_event__header_size(event->group_leader); | |
1379 | ||
1380 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1381 | perf_event__header_size(tmp); | |
04289bb9 IM |
1382 | } |
1383 | ||
fa66f07a SE |
1384 | static inline int |
1385 | event_filter_match(struct perf_event *event) | |
1386 | { | |
e5d1367f SE |
1387 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1388 | && perf_cgroup_match(event); | |
fa66f07a SE |
1389 | } |
1390 | ||
9ffcfa6f SE |
1391 | static void |
1392 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1393 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1394 | struct perf_event_context *ctx) |
3b6f9e5c | 1395 | { |
4158755d | 1396 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1397 | u64 delta; |
1398 | /* | |
1399 | * An event which could not be activated because of | |
1400 | * filter mismatch still needs to have its timings | |
1401 | * maintained, otherwise bogus information is return | |
1402 | * via read() for time_enabled, time_running: | |
1403 | */ | |
1404 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1405 | && !event_filter_match(event)) { | |
e5d1367f | 1406 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1407 | event->tstamp_running += delta; |
4158755d | 1408 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1409 | } |
1410 | ||
cdd6c482 | 1411 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1412 | return; |
3b6f9e5c | 1413 | |
44377277 AS |
1414 | perf_pmu_disable(event->pmu); |
1415 | ||
cdd6c482 IM |
1416 | event->state = PERF_EVENT_STATE_INACTIVE; |
1417 | if (event->pending_disable) { | |
1418 | event->pending_disable = 0; | |
1419 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1420 | } |
4158755d | 1421 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1422 | event->pmu->del(event, 0); |
cdd6c482 | 1423 | event->oncpu = -1; |
3b6f9e5c | 1424 | |
cdd6c482 | 1425 | if (!is_software_event(event)) |
3b6f9e5c PM |
1426 | cpuctx->active_oncpu--; |
1427 | ctx->nr_active--; | |
0f5a2601 PZ |
1428 | if (event->attr.freq && event->attr.sample_freq) |
1429 | ctx->nr_freq--; | |
cdd6c482 | 1430 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c | 1431 | cpuctx->exclusive = 0; |
44377277 AS |
1432 | |
1433 | perf_pmu_enable(event->pmu); | |
3b6f9e5c PM |
1434 | } |
1435 | ||
d859e29f | 1436 | static void |
cdd6c482 | 1437 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1438 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1439 | struct perf_event_context *ctx) |
d859e29f | 1440 | { |
cdd6c482 | 1441 | struct perf_event *event; |
fa66f07a | 1442 | int state = group_event->state; |
d859e29f | 1443 | |
cdd6c482 | 1444 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1445 | |
1446 | /* | |
1447 | * Schedule out siblings (if any): | |
1448 | */ | |
cdd6c482 IM |
1449 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1450 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1451 | |
fa66f07a | 1452 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1453 | cpuctx->exclusive = 0; |
1454 | } | |
1455 | ||
0793a61d | 1456 | /* |
cdd6c482 | 1457 | * Cross CPU call to remove a performance event |
0793a61d | 1458 | * |
cdd6c482 | 1459 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1460 | * remove it from the context list. |
1461 | */ | |
fe4b04fa | 1462 | static int __perf_remove_from_context(void *info) |
0793a61d | 1463 | { |
cdd6c482 IM |
1464 | struct perf_event *event = info; |
1465 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1466 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1467 | |
e625cce1 | 1468 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1469 | event_sched_out(event, cpuctx, ctx); |
cdd6c482 | 1470 | list_del_event(event, ctx); |
64ce3126 PZ |
1471 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1472 | ctx->is_active = 0; | |
1473 | cpuctx->task_ctx = NULL; | |
1474 | } | |
e625cce1 | 1475 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1476 | |
1477 | return 0; | |
0793a61d TG |
1478 | } |
1479 | ||
1480 | ||
1481 | /* | |
cdd6c482 | 1482 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1483 | * |
cdd6c482 | 1484 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1485 | * call when the task is on a CPU. |
c93f7669 | 1486 | * |
cdd6c482 IM |
1487 | * If event->ctx is a cloned context, callers must make sure that |
1488 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1489 | * remains valid. This is OK when called from perf_release since |
1490 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1491 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1492 | * context has been detached from its task. |
0793a61d | 1493 | */ |
fe4b04fa | 1494 | static void perf_remove_from_context(struct perf_event *event) |
0793a61d | 1495 | { |
cdd6c482 | 1496 | struct perf_event_context *ctx = event->ctx; |
0793a61d TG |
1497 | struct task_struct *task = ctx->task; |
1498 | ||
fe4b04fa PZ |
1499 | lockdep_assert_held(&ctx->mutex); |
1500 | ||
0793a61d TG |
1501 | if (!task) { |
1502 | /* | |
cdd6c482 | 1503 | * Per cpu events are removed via an smp call and |
af901ca1 | 1504 | * the removal is always successful. |
0793a61d | 1505 | */ |
fe4b04fa | 1506 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
0793a61d TG |
1507 | return; |
1508 | } | |
1509 | ||
1510 | retry: | |
fe4b04fa PZ |
1511 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1512 | return; | |
0793a61d | 1513 | |
e625cce1 | 1514 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1515 | /* |
fe4b04fa PZ |
1516 | * If we failed to find a running task, but find the context active now |
1517 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1518 | */ |
fe4b04fa | 1519 | if (ctx->is_active) { |
e625cce1 | 1520 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1521 | goto retry; |
1522 | } | |
1523 | ||
1524 | /* | |
fe4b04fa PZ |
1525 | * Since the task isn't running, its safe to remove the event, us |
1526 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1527 | */ |
fe4b04fa | 1528 | list_del_event(event, ctx); |
e625cce1 | 1529 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1530 | } |
1531 | ||
d859e29f | 1532 | /* |
cdd6c482 | 1533 | * Cross CPU call to disable a performance event |
d859e29f | 1534 | */ |
500ad2d8 | 1535 | int __perf_event_disable(void *info) |
d859e29f | 1536 | { |
cdd6c482 | 1537 | struct perf_event *event = info; |
cdd6c482 | 1538 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1539 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1540 | |
1541 | /* | |
cdd6c482 IM |
1542 | * If this is a per-task event, need to check whether this |
1543 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1544 | * |
1545 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1546 | * flipping contexts around. | |
d859e29f | 1547 | */ |
665c2142 | 1548 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1549 | return -EINVAL; |
d859e29f | 1550 | |
e625cce1 | 1551 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1552 | |
1553 | /* | |
cdd6c482 | 1554 | * If the event is on, turn it off. |
d859e29f PM |
1555 | * If it is in error state, leave it in error state. |
1556 | */ | |
cdd6c482 | 1557 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1558 | update_context_time(ctx); |
e5d1367f | 1559 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1560 | update_group_times(event); |
1561 | if (event == event->group_leader) | |
1562 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1563 | else |
cdd6c482 IM |
1564 | event_sched_out(event, cpuctx, ctx); |
1565 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1566 | } |
1567 | ||
e625cce1 | 1568 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1569 | |
1570 | return 0; | |
d859e29f PM |
1571 | } |
1572 | ||
1573 | /* | |
cdd6c482 | 1574 | * Disable a event. |
c93f7669 | 1575 | * |
cdd6c482 IM |
1576 | * If event->ctx is a cloned context, callers must make sure that |
1577 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1578 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1579 | * perf_event_for_each_child or perf_event_for_each because they |
1580 | * hold the top-level event's child_mutex, so any descendant that | |
1581 | * goes to exit will block in sync_child_event. | |
1582 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1583 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1584 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1585 | */ |
44234adc | 1586 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1587 | { |
cdd6c482 | 1588 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1589 | struct task_struct *task = ctx->task; |
1590 | ||
1591 | if (!task) { | |
1592 | /* | |
cdd6c482 | 1593 | * Disable the event on the cpu that it's on |
d859e29f | 1594 | */ |
fe4b04fa | 1595 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1596 | return; |
1597 | } | |
1598 | ||
9ed6060d | 1599 | retry: |
fe4b04fa PZ |
1600 | if (!task_function_call(task, __perf_event_disable, event)) |
1601 | return; | |
d859e29f | 1602 | |
e625cce1 | 1603 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1604 | /* |
cdd6c482 | 1605 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1606 | */ |
cdd6c482 | 1607 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1608 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1609 | /* |
1610 | * Reload the task pointer, it might have been changed by | |
1611 | * a concurrent perf_event_context_sched_out(). | |
1612 | */ | |
1613 | task = ctx->task; | |
d859e29f PM |
1614 | goto retry; |
1615 | } | |
1616 | ||
1617 | /* | |
1618 | * Since we have the lock this context can't be scheduled | |
1619 | * in, so we can change the state safely. | |
1620 | */ | |
cdd6c482 IM |
1621 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1622 | update_group_times(event); | |
1623 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1624 | } |
e625cce1 | 1625 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1626 | } |
dcfce4a0 | 1627 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1628 | |
e5d1367f SE |
1629 | static void perf_set_shadow_time(struct perf_event *event, |
1630 | struct perf_event_context *ctx, | |
1631 | u64 tstamp) | |
1632 | { | |
1633 | /* | |
1634 | * use the correct time source for the time snapshot | |
1635 | * | |
1636 | * We could get by without this by leveraging the | |
1637 | * fact that to get to this function, the caller | |
1638 | * has most likely already called update_context_time() | |
1639 | * and update_cgrp_time_xx() and thus both timestamp | |
1640 | * are identical (or very close). Given that tstamp is, | |
1641 | * already adjusted for cgroup, we could say that: | |
1642 | * tstamp - ctx->timestamp | |
1643 | * is equivalent to | |
1644 | * tstamp - cgrp->timestamp. | |
1645 | * | |
1646 | * Then, in perf_output_read(), the calculation would | |
1647 | * work with no changes because: | |
1648 | * - event is guaranteed scheduled in | |
1649 | * - no scheduled out in between | |
1650 | * - thus the timestamp would be the same | |
1651 | * | |
1652 | * But this is a bit hairy. | |
1653 | * | |
1654 | * So instead, we have an explicit cgroup call to remain | |
1655 | * within the time time source all along. We believe it | |
1656 | * is cleaner and simpler to understand. | |
1657 | */ | |
1658 | if (is_cgroup_event(event)) | |
1659 | perf_cgroup_set_shadow_time(event, tstamp); | |
1660 | else | |
1661 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1662 | } | |
1663 | ||
4fe757dd PZ |
1664 | #define MAX_INTERRUPTS (~0ULL) |
1665 | ||
1666 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1667 | ||
235c7fc7 | 1668 | static int |
9ffcfa6f | 1669 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1670 | struct perf_cpu_context *cpuctx, |
6e37738a | 1671 | struct perf_event_context *ctx) |
235c7fc7 | 1672 | { |
4158755d | 1673 | u64 tstamp = perf_event_time(event); |
44377277 | 1674 | int ret = 0; |
4158755d | 1675 | |
cdd6c482 | 1676 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1677 | return 0; |
1678 | ||
cdd6c482 | 1679 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1680 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1681 | |
1682 | /* | |
1683 | * Unthrottle events, since we scheduled we might have missed several | |
1684 | * ticks already, also for a heavily scheduling task there is little | |
1685 | * guarantee it'll get a tick in a timely manner. | |
1686 | */ | |
1687 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1688 | perf_log_throttle(event, 1); | |
1689 | event->hw.interrupts = 0; | |
1690 | } | |
1691 | ||
235c7fc7 IM |
1692 | /* |
1693 | * The new state must be visible before we turn it on in the hardware: | |
1694 | */ | |
1695 | smp_wmb(); | |
1696 | ||
44377277 AS |
1697 | perf_pmu_disable(event->pmu); |
1698 | ||
a4eaf7f1 | 1699 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1700 | event->state = PERF_EVENT_STATE_INACTIVE; |
1701 | event->oncpu = -1; | |
44377277 AS |
1702 | ret = -EAGAIN; |
1703 | goto out; | |
235c7fc7 IM |
1704 | } |
1705 | ||
4158755d | 1706 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1707 | |
e5d1367f | 1708 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1709 | |
cdd6c482 | 1710 | if (!is_software_event(event)) |
3b6f9e5c | 1711 | cpuctx->active_oncpu++; |
235c7fc7 | 1712 | ctx->nr_active++; |
0f5a2601 PZ |
1713 | if (event->attr.freq && event->attr.sample_freq) |
1714 | ctx->nr_freq++; | |
235c7fc7 | 1715 | |
cdd6c482 | 1716 | if (event->attr.exclusive) |
3b6f9e5c PM |
1717 | cpuctx->exclusive = 1; |
1718 | ||
44377277 AS |
1719 | out: |
1720 | perf_pmu_enable(event->pmu); | |
1721 | ||
1722 | return ret; | |
235c7fc7 IM |
1723 | } |
1724 | ||
6751b71e | 1725 | static int |
cdd6c482 | 1726 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1727 | struct perf_cpu_context *cpuctx, |
6e37738a | 1728 | struct perf_event_context *ctx) |
6751b71e | 1729 | { |
6bde9b6c | 1730 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1731 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1732 | u64 now = ctx->time; |
1733 | bool simulate = false; | |
6751b71e | 1734 | |
cdd6c482 | 1735 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1736 | return 0; |
1737 | ||
ad5133b7 | 1738 | pmu->start_txn(pmu); |
6bde9b6c | 1739 | |
9ffcfa6f | 1740 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1741 | pmu->cancel_txn(pmu); |
9e630205 | 1742 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1743 | return -EAGAIN; |
90151c35 | 1744 | } |
6751b71e PM |
1745 | |
1746 | /* | |
1747 | * Schedule in siblings as one group (if any): | |
1748 | */ | |
cdd6c482 | 1749 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1750 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1751 | partial_group = event; |
6751b71e PM |
1752 | goto group_error; |
1753 | } | |
1754 | } | |
1755 | ||
9ffcfa6f | 1756 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1757 | return 0; |
9ffcfa6f | 1758 | |
6751b71e PM |
1759 | group_error: |
1760 | /* | |
1761 | * Groups can be scheduled in as one unit only, so undo any | |
1762 | * partial group before returning: | |
d7842da4 SE |
1763 | * The events up to the failed event are scheduled out normally, |
1764 | * tstamp_stopped will be updated. | |
1765 | * | |
1766 | * The failed events and the remaining siblings need to have | |
1767 | * their timings updated as if they had gone thru event_sched_in() | |
1768 | * and event_sched_out(). This is required to get consistent timings | |
1769 | * across the group. This also takes care of the case where the group | |
1770 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1771 | * the time the event was actually stopped, such that time delta | |
1772 | * calculation in update_event_times() is correct. | |
6751b71e | 1773 | */ |
cdd6c482 IM |
1774 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1775 | if (event == partial_group) | |
d7842da4 SE |
1776 | simulate = true; |
1777 | ||
1778 | if (simulate) { | |
1779 | event->tstamp_running += now - event->tstamp_stopped; | |
1780 | event->tstamp_stopped = now; | |
1781 | } else { | |
1782 | event_sched_out(event, cpuctx, ctx); | |
1783 | } | |
6751b71e | 1784 | } |
9ffcfa6f | 1785 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1786 | |
ad5133b7 | 1787 | pmu->cancel_txn(pmu); |
90151c35 | 1788 | |
9e630205 SE |
1789 | perf_cpu_hrtimer_restart(cpuctx); |
1790 | ||
6751b71e PM |
1791 | return -EAGAIN; |
1792 | } | |
1793 | ||
3b6f9e5c | 1794 | /* |
cdd6c482 | 1795 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1796 | */ |
cdd6c482 | 1797 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1798 | struct perf_cpu_context *cpuctx, |
1799 | int can_add_hw) | |
1800 | { | |
1801 | /* | |
cdd6c482 | 1802 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1803 | */ |
d6f962b5 | 1804 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1805 | return 1; |
1806 | /* | |
1807 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1808 | * events can go on. |
3b6f9e5c PM |
1809 | */ |
1810 | if (cpuctx->exclusive) | |
1811 | return 0; | |
1812 | /* | |
1813 | * If this group is exclusive and there are already | |
cdd6c482 | 1814 | * events on the CPU, it can't go on. |
3b6f9e5c | 1815 | */ |
cdd6c482 | 1816 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1817 | return 0; |
1818 | /* | |
1819 | * Otherwise, try to add it if all previous groups were able | |
1820 | * to go on. | |
1821 | */ | |
1822 | return can_add_hw; | |
1823 | } | |
1824 | ||
cdd6c482 IM |
1825 | static void add_event_to_ctx(struct perf_event *event, |
1826 | struct perf_event_context *ctx) | |
53cfbf59 | 1827 | { |
4158755d SE |
1828 | u64 tstamp = perf_event_time(event); |
1829 | ||
cdd6c482 | 1830 | list_add_event(event, ctx); |
8a49542c | 1831 | perf_group_attach(event); |
4158755d SE |
1832 | event->tstamp_enabled = tstamp; |
1833 | event->tstamp_running = tstamp; | |
1834 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1835 | } |
1836 | ||
2c29ef0f PZ |
1837 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1838 | static void | |
1839 | ctx_sched_in(struct perf_event_context *ctx, | |
1840 | struct perf_cpu_context *cpuctx, | |
1841 | enum event_type_t event_type, | |
1842 | struct task_struct *task); | |
fe4b04fa | 1843 | |
dce5855b PZ |
1844 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1845 | struct perf_event_context *ctx, | |
1846 | struct task_struct *task) | |
1847 | { | |
1848 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1849 | if (ctx) | |
1850 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1851 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1852 | if (ctx) | |
1853 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1854 | } | |
1855 | ||
0793a61d | 1856 | /* |
cdd6c482 | 1857 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1858 | * |
1859 | * Must be called with ctx->mutex held | |
0793a61d | 1860 | */ |
fe4b04fa | 1861 | static int __perf_install_in_context(void *info) |
0793a61d | 1862 | { |
cdd6c482 IM |
1863 | struct perf_event *event = info; |
1864 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1865 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1866 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1867 | struct task_struct *task = current; | |
1868 | ||
b58f6b0d | 1869 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1870 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1871 | |
1872 | /* | |
2c29ef0f | 1873 | * If there was an active task_ctx schedule it out. |
0793a61d | 1874 | */ |
b58f6b0d | 1875 | if (task_ctx) |
2c29ef0f | 1876 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1877 | |
1878 | /* | |
1879 | * If the context we're installing events in is not the | |
1880 | * active task_ctx, flip them. | |
1881 | */ | |
1882 | if (ctx->task && task_ctx != ctx) { | |
1883 | if (task_ctx) | |
1884 | raw_spin_unlock(&task_ctx->lock); | |
1885 | raw_spin_lock(&ctx->lock); | |
1886 | task_ctx = ctx; | |
1887 | } | |
1888 | ||
1889 | if (task_ctx) { | |
1890 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1891 | task = task_ctx->task; |
1892 | } | |
b58f6b0d | 1893 | |
2c29ef0f | 1894 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1895 | |
4af4998b | 1896 | update_context_time(ctx); |
e5d1367f SE |
1897 | /* |
1898 | * update cgrp time only if current cgrp | |
1899 | * matches event->cgrp. Must be done before | |
1900 | * calling add_event_to_ctx() | |
1901 | */ | |
1902 | update_cgrp_time_from_event(event); | |
0793a61d | 1903 | |
cdd6c482 | 1904 | add_event_to_ctx(event, ctx); |
0793a61d | 1905 | |
d859e29f | 1906 | /* |
2c29ef0f | 1907 | * Schedule everything back in |
d859e29f | 1908 | */ |
dce5855b | 1909 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1910 | |
1911 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1912 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1913 | |
1914 | return 0; | |
0793a61d TG |
1915 | } |
1916 | ||
1917 | /* | |
cdd6c482 | 1918 | * Attach a performance event to a context |
0793a61d | 1919 | * |
cdd6c482 IM |
1920 | * First we add the event to the list with the hardware enable bit |
1921 | * in event->hw_config cleared. | |
0793a61d | 1922 | * |
cdd6c482 | 1923 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1924 | * call to enable it in the task context. The task might have been |
1925 | * scheduled away, but we check this in the smp call again. | |
1926 | */ | |
1927 | static void | |
cdd6c482 IM |
1928 | perf_install_in_context(struct perf_event_context *ctx, |
1929 | struct perf_event *event, | |
0793a61d TG |
1930 | int cpu) |
1931 | { | |
1932 | struct task_struct *task = ctx->task; | |
1933 | ||
fe4b04fa PZ |
1934 | lockdep_assert_held(&ctx->mutex); |
1935 | ||
c3f00c70 | 1936 | event->ctx = ctx; |
0cda4c02 YZ |
1937 | if (event->cpu != -1) |
1938 | event->cpu = cpu; | |
c3f00c70 | 1939 | |
0793a61d TG |
1940 | if (!task) { |
1941 | /* | |
cdd6c482 | 1942 | * Per cpu events are installed via an smp call and |
af901ca1 | 1943 | * the install is always successful. |
0793a61d | 1944 | */ |
fe4b04fa | 1945 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1946 | return; |
1947 | } | |
1948 | ||
0793a61d | 1949 | retry: |
fe4b04fa PZ |
1950 | if (!task_function_call(task, __perf_install_in_context, event)) |
1951 | return; | |
0793a61d | 1952 | |
e625cce1 | 1953 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1954 | /* |
fe4b04fa PZ |
1955 | * If we failed to find a running task, but find the context active now |
1956 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1957 | */ |
fe4b04fa | 1958 | if (ctx->is_active) { |
e625cce1 | 1959 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1960 | goto retry; |
1961 | } | |
1962 | ||
1963 | /* | |
fe4b04fa PZ |
1964 | * Since the task isn't running, its safe to add the event, us holding |
1965 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1966 | */ |
fe4b04fa | 1967 | add_event_to_ctx(event, ctx); |
e625cce1 | 1968 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1969 | } |
1970 | ||
fa289bec | 1971 | /* |
cdd6c482 | 1972 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1973 | * Enabling the leader of a group effectively enables all |
1974 | * the group members that aren't explicitly disabled, so we | |
1975 | * have to update their ->tstamp_enabled also. | |
1976 | * Note: this works for group members as well as group leaders | |
1977 | * since the non-leader members' sibling_lists will be empty. | |
1978 | */ | |
1d9b482e | 1979 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1980 | { |
cdd6c482 | 1981 | struct perf_event *sub; |
4158755d | 1982 | u64 tstamp = perf_event_time(event); |
fa289bec | 1983 | |
cdd6c482 | 1984 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1985 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1986 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1987 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1988 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1989 | } |
fa289bec PM |
1990 | } |
1991 | ||
d859e29f | 1992 | /* |
cdd6c482 | 1993 | * Cross CPU call to enable a performance event |
d859e29f | 1994 | */ |
fe4b04fa | 1995 | static int __perf_event_enable(void *info) |
04289bb9 | 1996 | { |
cdd6c482 | 1997 | struct perf_event *event = info; |
cdd6c482 IM |
1998 | struct perf_event_context *ctx = event->ctx; |
1999 | struct perf_event *leader = event->group_leader; | |
108b02cf | 2000 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 2001 | int err; |
04289bb9 | 2002 | |
06f41796 JO |
2003 | /* |
2004 | * There's a time window between 'ctx->is_active' check | |
2005 | * in perf_event_enable function and this place having: | |
2006 | * - IRQs on | |
2007 | * - ctx->lock unlocked | |
2008 | * | |
2009 | * where the task could be killed and 'ctx' deactivated | |
2010 | * by perf_event_exit_task. | |
2011 | */ | |
2012 | if (!ctx->is_active) | |
fe4b04fa | 2013 | return -EINVAL; |
3cbed429 | 2014 | |
e625cce1 | 2015 | raw_spin_lock(&ctx->lock); |
4af4998b | 2016 | update_context_time(ctx); |
d859e29f | 2017 | |
cdd6c482 | 2018 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 2019 | goto unlock; |
e5d1367f SE |
2020 | |
2021 | /* | |
2022 | * set current task's cgroup time reference point | |
2023 | */ | |
3f7cce3c | 2024 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 2025 | |
1d9b482e | 2026 | __perf_event_mark_enabled(event); |
04289bb9 | 2027 | |
e5d1367f SE |
2028 | if (!event_filter_match(event)) { |
2029 | if (is_cgroup_event(event)) | |
2030 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 2031 | goto unlock; |
e5d1367f | 2032 | } |
f4c4176f | 2033 | |
04289bb9 | 2034 | /* |
cdd6c482 | 2035 | * If the event is in a group and isn't the group leader, |
d859e29f | 2036 | * then don't put it on unless the group is on. |
04289bb9 | 2037 | */ |
cdd6c482 | 2038 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2039 | goto unlock; |
3b6f9e5c | 2040 | |
cdd6c482 | 2041 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2042 | err = -EEXIST; |
e758a33d | 2043 | } else { |
cdd6c482 | 2044 | if (event == leader) |
6e37738a | 2045 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2046 | else |
6e37738a | 2047 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2048 | } |
d859e29f PM |
2049 | |
2050 | if (err) { | |
2051 | /* | |
cdd6c482 | 2052 | * If this event can't go on and it's part of a |
d859e29f PM |
2053 | * group, then the whole group has to come off. |
2054 | */ | |
9e630205 | 2055 | if (leader != event) { |
d859e29f | 2056 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
2057 | perf_cpu_hrtimer_restart(cpuctx); |
2058 | } | |
0d48696f | 2059 | if (leader->attr.pinned) { |
53cfbf59 | 2060 | update_group_times(leader); |
cdd6c482 | 2061 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2062 | } |
d859e29f PM |
2063 | } |
2064 | ||
9ed6060d | 2065 | unlock: |
e625cce1 | 2066 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2067 | |
2068 | return 0; | |
d859e29f PM |
2069 | } |
2070 | ||
2071 | /* | |
cdd6c482 | 2072 | * Enable a event. |
c93f7669 | 2073 | * |
cdd6c482 IM |
2074 | * If event->ctx is a cloned context, callers must make sure that |
2075 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2076 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2077 | * perf_event_for_each_child or perf_event_for_each as described |
2078 | * for perf_event_disable. | |
d859e29f | 2079 | */ |
44234adc | 2080 | void perf_event_enable(struct perf_event *event) |
d859e29f | 2081 | { |
cdd6c482 | 2082 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2083 | struct task_struct *task = ctx->task; |
2084 | ||
2085 | if (!task) { | |
2086 | /* | |
cdd6c482 | 2087 | * Enable the event on the cpu that it's on |
d859e29f | 2088 | */ |
fe4b04fa | 2089 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2090 | return; |
2091 | } | |
2092 | ||
e625cce1 | 2093 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2094 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2095 | goto out; |
2096 | ||
2097 | /* | |
cdd6c482 IM |
2098 | * If the event is in error state, clear that first. |
2099 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2100 | * know that it has gone back into error state, as distinct |
2101 | * from the task having been scheduled away before the | |
2102 | * cross-call arrived. | |
2103 | */ | |
cdd6c482 IM |
2104 | if (event->state == PERF_EVENT_STATE_ERROR) |
2105 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2106 | |
9ed6060d | 2107 | retry: |
fe4b04fa | 2108 | if (!ctx->is_active) { |
1d9b482e | 2109 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2110 | goto out; |
2111 | } | |
2112 | ||
e625cce1 | 2113 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2114 | |
2115 | if (!task_function_call(task, __perf_event_enable, event)) | |
2116 | return; | |
d859e29f | 2117 | |
e625cce1 | 2118 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2119 | |
2120 | /* | |
cdd6c482 | 2121 | * If the context is active and the event is still off, |
d859e29f PM |
2122 | * we need to retry the cross-call. |
2123 | */ | |
fe4b04fa PZ |
2124 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2125 | /* | |
2126 | * task could have been flipped by a concurrent | |
2127 | * perf_event_context_sched_out() | |
2128 | */ | |
2129 | task = ctx->task; | |
d859e29f | 2130 | goto retry; |
fe4b04fa | 2131 | } |
fa289bec | 2132 | |
9ed6060d | 2133 | out: |
e625cce1 | 2134 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2135 | } |
dcfce4a0 | 2136 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2137 | |
26ca5c11 | 2138 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2139 | { |
2023b359 | 2140 | /* |
cdd6c482 | 2141 | * not supported on inherited events |
2023b359 | 2142 | */ |
2e939d1d | 2143 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2144 | return -EINVAL; |
2145 | ||
cdd6c482 IM |
2146 | atomic_add(refresh, &event->event_limit); |
2147 | perf_event_enable(event); | |
2023b359 PZ |
2148 | |
2149 | return 0; | |
79f14641 | 2150 | } |
26ca5c11 | 2151 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2152 | |
5b0311e1 FW |
2153 | static void ctx_sched_out(struct perf_event_context *ctx, |
2154 | struct perf_cpu_context *cpuctx, | |
2155 | enum event_type_t event_type) | |
235c7fc7 | 2156 | { |
cdd6c482 | 2157 | struct perf_event *event; |
db24d33e | 2158 | int is_active = ctx->is_active; |
235c7fc7 | 2159 | |
db24d33e | 2160 | ctx->is_active &= ~event_type; |
cdd6c482 | 2161 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2162 | return; |
2163 | ||
4af4998b | 2164 | update_context_time(ctx); |
e5d1367f | 2165 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2166 | if (!ctx->nr_active) |
facc4307 | 2167 | return; |
5b0311e1 | 2168 | |
075e0b00 | 2169 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2170 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2171 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2172 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2173 | } |
889ff015 | 2174 | |
db24d33e | 2175 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2176 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2177 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2178 | } |
1b9a644f | 2179 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2180 | } |
2181 | ||
564c2b21 | 2182 | /* |
5a3126d4 PZ |
2183 | * Test whether two contexts are equivalent, i.e. whether they have both been |
2184 | * cloned from the same version of the same context. | |
2185 | * | |
2186 | * Equivalence is measured using a generation number in the context that is | |
2187 | * incremented on each modification to it; see unclone_ctx(), list_add_event() | |
2188 | * and list_del_event(). | |
564c2b21 | 2189 | */ |
cdd6c482 IM |
2190 | static int context_equiv(struct perf_event_context *ctx1, |
2191 | struct perf_event_context *ctx2) | |
564c2b21 | 2192 | { |
5a3126d4 PZ |
2193 | /* Pinning disables the swap optimization */ |
2194 | if (ctx1->pin_count || ctx2->pin_count) | |
2195 | return 0; | |
2196 | ||
2197 | /* If ctx1 is the parent of ctx2 */ | |
2198 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) | |
2199 | return 1; | |
2200 | ||
2201 | /* If ctx2 is the parent of ctx1 */ | |
2202 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) | |
2203 | return 1; | |
2204 | ||
2205 | /* | |
2206 | * If ctx1 and ctx2 have the same parent; we flatten the parent | |
2207 | * hierarchy, see perf_event_init_context(). | |
2208 | */ | |
2209 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && | |
2210 | ctx1->parent_gen == ctx2->parent_gen) | |
2211 | return 1; | |
2212 | ||
2213 | /* Unmatched */ | |
2214 | return 0; | |
564c2b21 PM |
2215 | } |
2216 | ||
cdd6c482 IM |
2217 | static void __perf_event_sync_stat(struct perf_event *event, |
2218 | struct perf_event *next_event) | |
bfbd3381 PZ |
2219 | { |
2220 | u64 value; | |
2221 | ||
cdd6c482 | 2222 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2223 | return; |
2224 | ||
2225 | /* | |
cdd6c482 | 2226 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2227 | * because we're in the middle of a context switch and have IRQs |
2228 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2229 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2230 | * don't need to use it. |
2231 | */ | |
cdd6c482 IM |
2232 | switch (event->state) { |
2233 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2234 | event->pmu->read(event); |
2235 | /* fall-through */ | |
bfbd3381 | 2236 | |
cdd6c482 IM |
2237 | case PERF_EVENT_STATE_INACTIVE: |
2238 | update_event_times(event); | |
bfbd3381 PZ |
2239 | break; |
2240 | ||
2241 | default: | |
2242 | break; | |
2243 | } | |
2244 | ||
2245 | /* | |
cdd6c482 | 2246 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2247 | * values when we flip the contexts. |
2248 | */ | |
e7850595 PZ |
2249 | value = local64_read(&next_event->count); |
2250 | value = local64_xchg(&event->count, value); | |
2251 | local64_set(&next_event->count, value); | |
bfbd3381 | 2252 | |
cdd6c482 IM |
2253 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2254 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2255 | |
bfbd3381 | 2256 | /* |
19d2e755 | 2257 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2258 | */ |
cdd6c482 IM |
2259 | perf_event_update_userpage(event); |
2260 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2261 | } |
2262 | ||
cdd6c482 IM |
2263 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2264 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2265 | { |
cdd6c482 | 2266 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2267 | |
2268 | if (!ctx->nr_stat) | |
2269 | return; | |
2270 | ||
02ffdbc8 PZ |
2271 | update_context_time(ctx); |
2272 | ||
cdd6c482 IM |
2273 | event = list_first_entry(&ctx->event_list, |
2274 | struct perf_event, event_entry); | |
bfbd3381 | 2275 | |
cdd6c482 IM |
2276 | next_event = list_first_entry(&next_ctx->event_list, |
2277 | struct perf_event, event_entry); | |
bfbd3381 | 2278 | |
cdd6c482 IM |
2279 | while (&event->event_entry != &ctx->event_list && |
2280 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2281 | |
cdd6c482 | 2282 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2283 | |
cdd6c482 IM |
2284 | event = list_next_entry(event, event_entry); |
2285 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2286 | } |
2287 | } | |
2288 | ||
fe4b04fa PZ |
2289 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2290 | struct task_struct *next) | |
0793a61d | 2291 | { |
8dc85d54 | 2292 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 | 2293 | struct perf_event_context *next_ctx; |
5a3126d4 | 2294 | struct perf_event_context *parent, *next_parent; |
108b02cf | 2295 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2296 | int do_switch = 1; |
0793a61d | 2297 | |
108b02cf PZ |
2298 | if (likely(!ctx)) |
2299 | return; | |
10989fb2 | 2300 | |
108b02cf PZ |
2301 | cpuctx = __get_cpu_context(ctx); |
2302 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2303 | return; |
2304 | ||
c93f7669 | 2305 | rcu_read_lock(); |
8dc85d54 | 2306 | next_ctx = next->perf_event_ctxp[ctxn]; |
5a3126d4 PZ |
2307 | if (!next_ctx) |
2308 | goto unlock; | |
2309 | ||
2310 | parent = rcu_dereference(ctx->parent_ctx); | |
2311 | next_parent = rcu_dereference(next_ctx->parent_ctx); | |
2312 | ||
2313 | /* If neither context have a parent context; they cannot be clones. */ | |
2314 | if (!parent && !next_parent) | |
2315 | goto unlock; | |
2316 | ||
2317 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { | |
c93f7669 PM |
2318 | /* |
2319 | * Looks like the two contexts are clones, so we might be | |
2320 | * able to optimize the context switch. We lock both | |
2321 | * contexts and check that they are clones under the | |
2322 | * lock (including re-checking that neither has been | |
2323 | * uncloned in the meantime). It doesn't matter which | |
2324 | * order we take the locks because no other cpu could | |
2325 | * be trying to lock both of these tasks. | |
2326 | */ | |
e625cce1 TG |
2327 | raw_spin_lock(&ctx->lock); |
2328 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2329 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2330 | /* |
2331 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2332 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2333 | */ |
8dc85d54 PZ |
2334 | task->perf_event_ctxp[ctxn] = next_ctx; |
2335 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2336 | ctx->task = next; |
2337 | next_ctx->task = task; | |
2338 | do_switch = 0; | |
bfbd3381 | 2339 | |
cdd6c482 | 2340 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2341 | } |
e625cce1 TG |
2342 | raw_spin_unlock(&next_ctx->lock); |
2343 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2344 | } |
5a3126d4 | 2345 | unlock: |
c93f7669 | 2346 | rcu_read_unlock(); |
564c2b21 | 2347 | |
c93f7669 | 2348 | if (do_switch) { |
facc4307 | 2349 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2350 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2351 | cpuctx->task_ctx = NULL; |
facc4307 | 2352 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2353 | } |
0793a61d TG |
2354 | } |
2355 | ||
8dc85d54 PZ |
2356 | #define for_each_task_context_nr(ctxn) \ |
2357 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2358 | ||
2359 | /* | |
2360 | * Called from scheduler to remove the events of the current task, | |
2361 | * with interrupts disabled. | |
2362 | * | |
2363 | * We stop each event and update the event value in event->count. | |
2364 | * | |
2365 | * This does not protect us against NMI, but disable() | |
2366 | * sets the disabled bit in the control field of event _before_ | |
2367 | * accessing the event control register. If a NMI hits, then it will | |
2368 | * not restart the event. | |
2369 | */ | |
ab0cce56 JO |
2370 | void __perf_event_task_sched_out(struct task_struct *task, |
2371 | struct task_struct *next) | |
8dc85d54 PZ |
2372 | { |
2373 | int ctxn; | |
2374 | ||
8dc85d54 PZ |
2375 | for_each_task_context_nr(ctxn) |
2376 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2377 | |
2378 | /* | |
2379 | * if cgroup events exist on this CPU, then we need | |
2380 | * to check if we have to switch out PMU state. | |
2381 | * cgroup event are system-wide mode only | |
2382 | */ | |
2383 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2384 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2385 | } |
2386 | ||
04dc2dbb | 2387 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2388 | { |
108b02cf | 2389 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2390 | |
a63eaf34 PM |
2391 | if (!cpuctx->task_ctx) |
2392 | return; | |
012b84da IM |
2393 | |
2394 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2395 | return; | |
2396 | ||
04dc2dbb | 2397 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2398 | cpuctx->task_ctx = NULL; |
2399 | } | |
2400 | ||
5b0311e1 FW |
2401 | /* |
2402 | * Called with IRQs disabled | |
2403 | */ | |
2404 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2405 | enum event_type_t event_type) | |
2406 | { | |
2407 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2408 | } |
2409 | ||
235c7fc7 | 2410 | static void |
5b0311e1 | 2411 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2412 | struct perf_cpu_context *cpuctx) |
0793a61d | 2413 | { |
cdd6c482 | 2414 | struct perf_event *event; |
0793a61d | 2415 | |
889ff015 FW |
2416 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2417 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2418 | continue; |
5632ab12 | 2419 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2420 | continue; |
2421 | ||
e5d1367f SE |
2422 | /* may need to reset tstamp_enabled */ |
2423 | if (is_cgroup_event(event)) | |
2424 | perf_cgroup_mark_enabled(event, ctx); | |
2425 | ||
8c9ed8e1 | 2426 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2427 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2428 | |
2429 | /* | |
2430 | * If this pinned group hasn't been scheduled, | |
2431 | * put it in error state. | |
2432 | */ | |
cdd6c482 IM |
2433 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2434 | update_group_times(event); | |
2435 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2436 | } |
3b6f9e5c | 2437 | } |
5b0311e1 FW |
2438 | } |
2439 | ||
2440 | static void | |
2441 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2442 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2443 | { |
2444 | struct perf_event *event; | |
2445 | int can_add_hw = 1; | |
3b6f9e5c | 2446 | |
889ff015 FW |
2447 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2448 | /* Ignore events in OFF or ERROR state */ | |
2449 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2450 | continue; |
04289bb9 IM |
2451 | /* |
2452 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2453 | * of events: |
04289bb9 | 2454 | */ |
5632ab12 | 2455 | if (!event_filter_match(event)) |
0793a61d TG |
2456 | continue; |
2457 | ||
e5d1367f SE |
2458 | /* may need to reset tstamp_enabled */ |
2459 | if (is_cgroup_event(event)) | |
2460 | perf_cgroup_mark_enabled(event, ctx); | |
2461 | ||
9ed6060d | 2462 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2463 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2464 | can_add_hw = 0; |
9ed6060d | 2465 | } |
0793a61d | 2466 | } |
5b0311e1 FW |
2467 | } |
2468 | ||
2469 | static void | |
2470 | ctx_sched_in(struct perf_event_context *ctx, | |
2471 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2472 | enum event_type_t event_type, |
2473 | struct task_struct *task) | |
5b0311e1 | 2474 | { |
e5d1367f | 2475 | u64 now; |
db24d33e | 2476 | int is_active = ctx->is_active; |
e5d1367f | 2477 | |
db24d33e | 2478 | ctx->is_active |= event_type; |
5b0311e1 | 2479 | if (likely(!ctx->nr_events)) |
facc4307 | 2480 | return; |
5b0311e1 | 2481 | |
e5d1367f SE |
2482 | now = perf_clock(); |
2483 | ctx->timestamp = now; | |
3f7cce3c | 2484 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2485 | /* |
2486 | * First go through the list and put on any pinned groups | |
2487 | * in order to give them the best chance of going on. | |
2488 | */ | |
db24d33e | 2489 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2490 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2491 | |
2492 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2493 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2494 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2495 | } |
2496 | ||
329c0e01 | 2497 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2498 | enum event_type_t event_type, |
2499 | struct task_struct *task) | |
329c0e01 FW |
2500 | { |
2501 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2502 | ||
e5d1367f | 2503 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2504 | } |
2505 | ||
e5d1367f SE |
2506 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2507 | struct task_struct *task) | |
235c7fc7 | 2508 | { |
108b02cf | 2509 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2510 | |
108b02cf | 2511 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2512 | if (cpuctx->task_ctx == ctx) |
2513 | return; | |
2514 | ||
facc4307 | 2515 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2516 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2517 | /* |
2518 | * We want to keep the following priority order: | |
2519 | * cpu pinned (that don't need to move), task pinned, | |
2520 | * cpu flexible, task flexible. | |
2521 | */ | |
2522 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2523 | ||
1d5f003f GN |
2524 | if (ctx->nr_events) |
2525 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2526 | |
86b47c25 GN |
2527 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2528 | ||
facc4307 PZ |
2529 | perf_pmu_enable(ctx->pmu); |
2530 | perf_ctx_unlock(cpuctx, ctx); | |
2531 | ||
b5ab4cd5 PZ |
2532 | /* |
2533 | * Since these rotations are per-cpu, we need to ensure the | |
2534 | * cpu-context we got scheduled on is actually rotating. | |
2535 | */ | |
108b02cf | 2536 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2537 | } |
2538 | ||
d010b332 SE |
2539 | /* |
2540 | * When sampling the branck stack in system-wide, it may be necessary | |
2541 | * to flush the stack on context switch. This happens when the branch | |
2542 | * stack does not tag its entries with the pid of the current task. | |
2543 | * Otherwise it becomes impossible to associate a branch entry with a | |
2544 | * task. This ambiguity is more likely to appear when the branch stack | |
2545 | * supports priv level filtering and the user sets it to monitor only | |
2546 | * at the user level (which could be a useful measurement in system-wide | |
2547 | * mode). In that case, the risk is high of having a branch stack with | |
2548 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2549 | * entries or stashing them somewhere in the PMU specific code layer. | |
2550 | * | |
2551 | * This function provides the context switch callback to the lower code | |
2552 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2553 | * with at least one active event using taken branch sampling. | |
2554 | */ | |
2555 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2556 | struct task_struct *task) | |
2557 | { | |
2558 | struct perf_cpu_context *cpuctx; | |
2559 | struct pmu *pmu; | |
2560 | unsigned long flags; | |
2561 | ||
2562 | /* no need to flush branch stack if not changing task */ | |
2563 | if (prev == task) | |
2564 | return; | |
2565 | ||
2566 | local_irq_save(flags); | |
2567 | ||
2568 | rcu_read_lock(); | |
2569 | ||
2570 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2571 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2572 | ||
2573 | /* | |
2574 | * check if the context has at least one | |
2575 | * event using PERF_SAMPLE_BRANCH_STACK | |
2576 | */ | |
2577 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2578 | && pmu->flush_branch_stack) { | |
2579 | ||
2580 | pmu = cpuctx->ctx.pmu; | |
2581 | ||
2582 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2583 | ||
2584 | perf_pmu_disable(pmu); | |
2585 | ||
2586 | pmu->flush_branch_stack(); | |
2587 | ||
2588 | perf_pmu_enable(pmu); | |
2589 | ||
2590 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2591 | } | |
2592 | } | |
2593 | ||
2594 | rcu_read_unlock(); | |
2595 | ||
2596 | local_irq_restore(flags); | |
2597 | } | |
2598 | ||
8dc85d54 PZ |
2599 | /* |
2600 | * Called from scheduler to add the events of the current task | |
2601 | * with interrupts disabled. | |
2602 | * | |
2603 | * We restore the event value and then enable it. | |
2604 | * | |
2605 | * This does not protect us against NMI, but enable() | |
2606 | * sets the enabled bit in the control field of event _before_ | |
2607 | * accessing the event control register. If a NMI hits, then it will | |
2608 | * keep the event running. | |
2609 | */ | |
ab0cce56 JO |
2610 | void __perf_event_task_sched_in(struct task_struct *prev, |
2611 | struct task_struct *task) | |
8dc85d54 PZ |
2612 | { |
2613 | struct perf_event_context *ctx; | |
2614 | int ctxn; | |
2615 | ||
2616 | for_each_task_context_nr(ctxn) { | |
2617 | ctx = task->perf_event_ctxp[ctxn]; | |
2618 | if (likely(!ctx)) | |
2619 | continue; | |
2620 | ||
e5d1367f | 2621 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2622 | } |
e5d1367f SE |
2623 | /* |
2624 | * if cgroup events exist on this CPU, then we need | |
2625 | * to check if we have to switch in PMU state. | |
2626 | * cgroup event are system-wide mode only | |
2627 | */ | |
2628 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2629 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2630 | |
2631 | /* check for system-wide branch_stack events */ | |
2632 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2633 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2634 | } |
2635 | ||
abd50713 PZ |
2636 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2637 | { | |
2638 | u64 frequency = event->attr.sample_freq; | |
2639 | u64 sec = NSEC_PER_SEC; | |
2640 | u64 divisor, dividend; | |
2641 | ||
2642 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2643 | ||
2644 | count_fls = fls64(count); | |
2645 | nsec_fls = fls64(nsec); | |
2646 | frequency_fls = fls64(frequency); | |
2647 | sec_fls = 30; | |
2648 | ||
2649 | /* | |
2650 | * We got @count in @nsec, with a target of sample_freq HZ | |
2651 | * the target period becomes: | |
2652 | * | |
2653 | * @count * 10^9 | |
2654 | * period = ------------------- | |
2655 | * @nsec * sample_freq | |
2656 | * | |
2657 | */ | |
2658 | ||
2659 | /* | |
2660 | * Reduce accuracy by one bit such that @a and @b converge | |
2661 | * to a similar magnitude. | |
2662 | */ | |
fe4b04fa | 2663 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2664 | do { \ |
2665 | if (a##_fls > b##_fls) { \ | |
2666 | a >>= 1; \ | |
2667 | a##_fls--; \ | |
2668 | } else { \ | |
2669 | b >>= 1; \ | |
2670 | b##_fls--; \ | |
2671 | } \ | |
2672 | } while (0) | |
2673 | ||
2674 | /* | |
2675 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2676 | * the other, so that finally we can do a u64/u64 division. | |
2677 | */ | |
2678 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2679 | REDUCE_FLS(nsec, frequency); | |
2680 | REDUCE_FLS(sec, count); | |
2681 | } | |
2682 | ||
2683 | if (count_fls + sec_fls > 64) { | |
2684 | divisor = nsec * frequency; | |
2685 | ||
2686 | while (count_fls + sec_fls > 64) { | |
2687 | REDUCE_FLS(count, sec); | |
2688 | divisor >>= 1; | |
2689 | } | |
2690 | ||
2691 | dividend = count * sec; | |
2692 | } else { | |
2693 | dividend = count * sec; | |
2694 | ||
2695 | while (nsec_fls + frequency_fls > 64) { | |
2696 | REDUCE_FLS(nsec, frequency); | |
2697 | dividend >>= 1; | |
2698 | } | |
2699 | ||
2700 | divisor = nsec * frequency; | |
2701 | } | |
2702 | ||
f6ab91ad PZ |
2703 | if (!divisor) |
2704 | return dividend; | |
2705 | ||
abd50713 PZ |
2706 | return div64_u64(dividend, divisor); |
2707 | } | |
2708 | ||
e050e3f0 SE |
2709 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2710 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2711 | ||
f39d47ff | 2712 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2713 | { |
cdd6c482 | 2714 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2715 | s64 period, sample_period; |
bd2b5b12 PZ |
2716 | s64 delta; |
2717 | ||
abd50713 | 2718 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2719 | |
2720 | delta = (s64)(period - hwc->sample_period); | |
2721 | delta = (delta + 7) / 8; /* low pass filter */ | |
2722 | ||
2723 | sample_period = hwc->sample_period + delta; | |
2724 | ||
2725 | if (!sample_period) | |
2726 | sample_period = 1; | |
2727 | ||
bd2b5b12 | 2728 | hwc->sample_period = sample_period; |
abd50713 | 2729 | |
e7850595 | 2730 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2731 | if (disable) |
2732 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2733 | ||
e7850595 | 2734 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2735 | |
2736 | if (disable) | |
2737 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2738 | } |
bd2b5b12 PZ |
2739 | } |
2740 | ||
e050e3f0 SE |
2741 | /* |
2742 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2743 | * events. At the same time, make sure, having freq events does not change | |
2744 | * the rate of unthrottling as that would introduce bias. | |
2745 | */ | |
2746 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2747 | int needs_unthr) | |
60db5e09 | 2748 | { |
cdd6c482 IM |
2749 | struct perf_event *event; |
2750 | struct hw_perf_event *hwc; | |
e050e3f0 | 2751 | u64 now, period = TICK_NSEC; |
abd50713 | 2752 | s64 delta; |
60db5e09 | 2753 | |
e050e3f0 SE |
2754 | /* |
2755 | * only need to iterate over all events iff: | |
2756 | * - context have events in frequency mode (needs freq adjust) | |
2757 | * - there are events to unthrottle on this cpu | |
2758 | */ | |
2759 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2760 | return; |
2761 | ||
e050e3f0 | 2762 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2763 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2764 | |
03541f8b | 2765 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2766 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2767 | continue; |
2768 | ||
5632ab12 | 2769 | if (!event_filter_match(event)) |
5d27c23d PZ |
2770 | continue; |
2771 | ||
44377277 AS |
2772 | perf_pmu_disable(event->pmu); |
2773 | ||
cdd6c482 | 2774 | hwc = &event->hw; |
6a24ed6c | 2775 | |
ae23bff1 | 2776 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2777 | hwc->interrupts = 0; |
cdd6c482 | 2778 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2779 | event->pmu->start(event, 0); |
a78ac325 PZ |
2780 | } |
2781 | ||
cdd6c482 | 2782 | if (!event->attr.freq || !event->attr.sample_freq) |
44377277 | 2783 | goto next; |
60db5e09 | 2784 | |
e050e3f0 SE |
2785 | /* |
2786 | * stop the event and update event->count | |
2787 | */ | |
2788 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2789 | ||
e7850595 | 2790 | now = local64_read(&event->count); |
abd50713 PZ |
2791 | delta = now - hwc->freq_count_stamp; |
2792 | hwc->freq_count_stamp = now; | |
60db5e09 | 2793 | |
e050e3f0 SE |
2794 | /* |
2795 | * restart the event | |
2796 | * reload only if value has changed | |
f39d47ff SE |
2797 | * we have stopped the event so tell that |
2798 | * to perf_adjust_period() to avoid stopping it | |
2799 | * twice. | |
e050e3f0 | 2800 | */ |
abd50713 | 2801 | if (delta > 0) |
f39d47ff | 2802 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2803 | |
2804 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
44377277 AS |
2805 | next: |
2806 | perf_pmu_enable(event->pmu); | |
60db5e09 | 2807 | } |
e050e3f0 | 2808 | |
f39d47ff | 2809 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2810 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2811 | } |
2812 | ||
235c7fc7 | 2813 | /* |
cdd6c482 | 2814 | * Round-robin a context's events: |
235c7fc7 | 2815 | */ |
cdd6c482 | 2816 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2817 | { |
dddd3379 TG |
2818 | /* |
2819 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2820 | * disabled by the inheritance code. | |
2821 | */ | |
2822 | if (!ctx->rotate_disable) | |
2823 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2824 | } |
2825 | ||
b5ab4cd5 | 2826 | /* |
e9d2b064 PZ |
2827 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2828 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2829 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2830 | */ |
9e630205 | 2831 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2832 | { |
8dc85d54 | 2833 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2834 | int rotate = 0, remove = 1; |
7fc23a53 | 2835 | |
b5ab4cd5 | 2836 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2837 | remove = 0; |
b5ab4cd5 PZ |
2838 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2839 | rotate = 1; | |
2840 | } | |
235c7fc7 | 2841 | |
8dc85d54 | 2842 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2843 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2844 | remove = 0; |
b5ab4cd5 PZ |
2845 | if (ctx->nr_events != ctx->nr_active) |
2846 | rotate = 1; | |
2847 | } | |
9717e6cd | 2848 | |
e050e3f0 | 2849 | if (!rotate) |
0f5a2601 PZ |
2850 | goto done; |
2851 | ||
facc4307 | 2852 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2853 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2854 | |
e050e3f0 SE |
2855 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2856 | if (ctx) | |
2857 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2858 | |
e050e3f0 SE |
2859 | rotate_ctx(&cpuctx->ctx); |
2860 | if (ctx) | |
2861 | rotate_ctx(ctx); | |
235c7fc7 | 2862 | |
e050e3f0 | 2863 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2864 | |
0f5a2601 PZ |
2865 | perf_pmu_enable(cpuctx->ctx.pmu); |
2866 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2867 | done: |
e9d2b064 PZ |
2868 | if (remove) |
2869 | list_del_init(&cpuctx->rotation_list); | |
9e630205 SE |
2870 | |
2871 | return rotate; | |
e9d2b064 PZ |
2872 | } |
2873 | ||
026249ef FW |
2874 | #ifdef CONFIG_NO_HZ_FULL |
2875 | bool perf_event_can_stop_tick(void) | |
2876 | { | |
948b26b6 | 2877 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 2878 | __this_cpu_read(perf_throttled_count)) |
026249ef | 2879 | return false; |
d84153d6 FW |
2880 | else |
2881 | return true; | |
026249ef FW |
2882 | } |
2883 | #endif | |
2884 | ||
e9d2b064 PZ |
2885 | void perf_event_task_tick(void) |
2886 | { | |
2887 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2888 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2889 | struct perf_event_context *ctx; |
2890 | int throttled; | |
b5ab4cd5 | 2891 | |
e9d2b064 PZ |
2892 | WARN_ON(!irqs_disabled()); |
2893 | ||
e050e3f0 SE |
2894 | __this_cpu_inc(perf_throttled_seq); |
2895 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2896 | ||
e9d2b064 | 2897 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2898 | ctx = &cpuctx->ctx; |
2899 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2900 | ||
2901 | ctx = cpuctx->task_ctx; | |
2902 | if (ctx) | |
2903 | perf_adjust_freq_unthr_context(ctx, throttled); | |
e9d2b064 | 2904 | } |
0793a61d TG |
2905 | } |
2906 | ||
889ff015 FW |
2907 | static int event_enable_on_exec(struct perf_event *event, |
2908 | struct perf_event_context *ctx) | |
2909 | { | |
2910 | if (!event->attr.enable_on_exec) | |
2911 | return 0; | |
2912 | ||
2913 | event->attr.enable_on_exec = 0; | |
2914 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2915 | return 0; | |
2916 | ||
1d9b482e | 2917 | __perf_event_mark_enabled(event); |
889ff015 FW |
2918 | |
2919 | return 1; | |
2920 | } | |
2921 | ||
57e7986e | 2922 | /* |
cdd6c482 | 2923 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2924 | * This expects task == current. |
2925 | */ | |
8dc85d54 | 2926 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2927 | { |
cdd6c482 | 2928 | struct perf_event *event; |
57e7986e PM |
2929 | unsigned long flags; |
2930 | int enabled = 0; | |
889ff015 | 2931 | int ret; |
57e7986e PM |
2932 | |
2933 | local_irq_save(flags); | |
cdd6c482 | 2934 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2935 | goto out; |
2936 | ||
e566b76e SE |
2937 | /* |
2938 | * We must ctxsw out cgroup events to avoid conflict | |
2939 | * when invoking perf_task_event_sched_in() later on | |
2940 | * in this function. Otherwise we end up trying to | |
2941 | * ctxswin cgroup events which are already scheduled | |
2942 | * in. | |
2943 | */ | |
a8d757ef | 2944 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2945 | |
e625cce1 | 2946 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2947 | task_ctx_sched_out(ctx); |
57e7986e | 2948 | |
b79387ef | 2949 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2950 | ret = event_enable_on_exec(event, ctx); |
2951 | if (ret) | |
2952 | enabled = 1; | |
57e7986e PM |
2953 | } |
2954 | ||
2955 | /* | |
cdd6c482 | 2956 | * Unclone this context if we enabled any event. |
57e7986e | 2957 | */ |
71a851b4 PZ |
2958 | if (enabled) |
2959 | unclone_ctx(ctx); | |
57e7986e | 2960 | |
e625cce1 | 2961 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2962 | |
e566b76e SE |
2963 | /* |
2964 | * Also calls ctxswin for cgroup events, if any: | |
2965 | */ | |
e5d1367f | 2966 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2967 | out: |
57e7986e PM |
2968 | local_irq_restore(flags); |
2969 | } | |
2970 | ||
0793a61d | 2971 | /* |
cdd6c482 | 2972 | * Cross CPU call to read the hardware event |
0793a61d | 2973 | */ |
cdd6c482 | 2974 | static void __perf_event_read(void *info) |
0793a61d | 2975 | { |
cdd6c482 IM |
2976 | struct perf_event *event = info; |
2977 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2978 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2979 | |
e1ac3614 PM |
2980 | /* |
2981 | * If this is a task context, we need to check whether it is | |
2982 | * the current task context of this cpu. If not it has been | |
2983 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2984 | * event->count would have been updated to a recent sample |
2985 | * when the event was scheduled out. | |
e1ac3614 PM |
2986 | */ |
2987 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2988 | return; | |
2989 | ||
e625cce1 | 2990 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2991 | if (ctx->is_active) { |
542e72fc | 2992 | update_context_time(ctx); |
e5d1367f SE |
2993 | update_cgrp_time_from_event(event); |
2994 | } | |
cdd6c482 | 2995 | update_event_times(event); |
542e72fc PZ |
2996 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2997 | event->pmu->read(event); | |
e625cce1 | 2998 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2999 | } |
3000 | ||
b5e58793 PZ |
3001 | static inline u64 perf_event_count(struct perf_event *event) |
3002 | { | |
e7850595 | 3003 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
3004 | } |
3005 | ||
cdd6c482 | 3006 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
3007 | { |
3008 | /* | |
cdd6c482 IM |
3009 | * If event is enabled and currently active on a CPU, update the |
3010 | * value in the event structure: | |
0793a61d | 3011 | */ |
cdd6c482 IM |
3012 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
3013 | smp_call_function_single(event->oncpu, | |
3014 | __perf_event_read, event, 1); | |
3015 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
3016 | struct perf_event_context *ctx = event->ctx; |
3017 | unsigned long flags; | |
3018 | ||
e625cce1 | 3019 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
3020 | /* |
3021 | * may read while context is not active | |
3022 | * (e.g., thread is blocked), in that case | |
3023 | * we cannot update context time | |
3024 | */ | |
e5d1367f | 3025 | if (ctx->is_active) { |
c530ccd9 | 3026 | update_context_time(ctx); |
e5d1367f SE |
3027 | update_cgrp_time_from_event(event); |
3028 | } | |
cdd6c482 | 3029 | update_event_times(event); |
e625cce1 | 3030 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
3031 | } |
3032 | ||
b5e58793 | 3033 | return perf_event_count(event); |
0793a61d TG |
3034 | } |
3035 | ||
a63eaf34 | 3036 | /* |
cdd6c482 | 3037 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 3038 | */ |
eb184479 | 3039 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 3040 | { |
e625cce1 | 3041 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 3042 | mutex_init(&ctx->mutex); |
889ff015 FW |
3043 | INIT_LIST_HEAD(&ctx->pinned_groups); |
3044 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
3045 | INIT_LIST_HEAD(&ctx->event_list); |
3046 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
3047 | } |
3048 | ||
3049 | static struct perf_event_context * | |
3050 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
3051 | { | |
3052 | struct perf_event_context *ctx; | |
3053 | ||
3054 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
3055 | if (!ctx) | |
3056 | return NULL; | |
3057 | ||
3058 | __perf_event_init_context(ctx); | |
3059 | if (task) { | |
3060 | ctx->task = task; | |
3061 | get_task_struct(task); | |
0793a61d | 3062 | } |
eb184479 PZ |
3063 | ctx->pmu = pmu; |
3064 | ||
3065 | return ctx; | |
a63eaf34 PM |
3066 | } |
3067 | ||
2ebd4ffb MH |
3068 | static struct task_struct * |
3069 | find_lively_task_by_vpid(pid_t vpid) | |
3070 | { | |
3071 | struct task_struct *task; | |
3072 | int err; | |
0793a61d TG |
3073 | |
3074 | rcu_read_lock(); | |
2ebd4ffb | 3075 | if (!vpid) |
0793a61d TG |
3076 | task = current; |
3077 | else | |
2ebd4ffb | 3078 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3079 | if (task) |
3080 | get_task_struct(task); | |
3081 | rcu_read_unlock(); | |
3082 | ||
3083 | if (!task) | |
3084 | return ERR_PTR(-ESRCH); | |
3085 | ||
0793a61d | 3086 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3087 | err = -EACCES; |
3088 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3089 | goto errout; | |
3090 | ||
2ebd4ffb MH |
3091 | return task; |
3092 | errout: | |
3093 | put_task_struct(task); | |
3094 | return ERR_PTR(err); | |
3095 | ||
3096 | } | |
3097 | ||
fe4b04fa PZ |
3098 | /* |
3099 | * Returns a matching context with refcount and pincount. | |
3100 | */ | |
108b02cf | 3101 | static struct perf_event_context * |
38a81da2 | 3102 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 3103 | { |
cdd6c482 | 3104 | struct perf_event_context *ctx; |
22a4f650 | 3105 | struct perf_cpu_context *cpuctx; |
25346b93 | 3106 | unsigned long flags; |
8dc85d54 | 3107 | int ctxn, err; |
0793a61d | 3108 | |
22a4ec72 | 3109 | if (!task) { |
cdd6c482 | 3110 | /* Must be root to operate on a CPU event: */ |
0764771d | 3111 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3112 | return ERR_PTR(-EACCES); |
3113 | ||
0793a61d | 3114 | /* |
cdd6c482 | 3115 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3116 | * offline CPU and activate it when the CPU comes up, but |
3117 | * that's for later. | |
3118 | */ | |
f6325e30 | 3119 | if (!cpu_online(cpu)) |
0793a61d TG |
3120 | return ERR_PTR(-ENODEV); |
3121 | ||
108b02cf | 3122 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3123 | ctx = &cpuctx->ctx; |
c93f7669 | 3124 | get_ctx(ctx); |
fe4b04fa | 3125 | ++ctx->pin_count; |
0793a61d | 3126 | |
0793a61d TG |
3127 | return ctx; |
3128 | } | |
3129 | ||
8dc85d54 PZ |
3130 | err = -EINVAL; |
3131 | ctxn = pmu->task_ctx_nr; | |
3132 | if (ctxn < 0) | |
3133 | goto errout; | |
3134 | ||
9ed6060d | 3135 | retry: |
8dc85d54 | 3136 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3137 | if (ctx) { |
71a851b4 | 3138 | unclone_ctx(ctx); |
fe4b04fa | 3139 | ++ctx->pin_count; |
e625cce1 | 3140 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 3141 | } else { |
eb184479 | 3142 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3143 | err = -ENOMEM; |
3144 | if (!ctx) | |
3145 | goto errout; | |
eb184479 | 3146 | |
dbe08d82 ON |
3147 | err = 0; |
3148 | mutex_lock(&task->perf_event_mutex); | |
3149 | /* | |
3150 | * If it has already passed perf_event_exit_task(). | |
3151 | * we must see PF_EXITING, it takes this mutex too. | |
3152 | */ | |
3153 | if (task->flags & PF_EXITING) | |
3154 | err = -ESRCH; | |
3155 | else if (task->perf_event_ctxp[ctxn]) | |
3156 | err = -EAGAIN; | |
fe4b04fa | 3157 | else { |
9137fb28 | 3158 | get_ctx(ctx); |
fe4b04fa | 3159 | ++ctx->pin_count; |
dbe08d82 | 3160 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3161 | } |
dbe08d82 ON |
3162 | mutex_unlock(&task->perf_event_mutex); |
3163 | ||
3164 | if (unlikely(err)) { | |
9137fb28 | 3165 | put_ctx(ctx); |
dbe08d82 ON |
3166 | |
3167 | if (err == -EAGAIN) | |
3168 | goto retry; | |
3169 | goto errout; | |
a63eaf34 PM |
3170 | } |
3171 | } | |
3172 | ||
0793a61d | 3173 | return ctx; |
c93f7669 | 3174 | |
9ed6060d | 3175 | errout: |
c93f7669 | 3176 | return ERR_PTR(err); |
0793a61d TG |
3177 | } |
3178 | ||
6fb2915d LZ |
3179 | static void perf_event_free_filter(struct perf_event *event); |
3180 | ||
cdd6c482 | 3181 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3182 | { |
cdd6c482 | 3183 | struct perf_event *event; |
592903cd | 3184 | |
cdd6c482 IM |
3185 | event = container_of(head, struct perf_event, rcu_head); |
3186 | if (event->ns) | |
3187 | put_pid_ns(event->ns); | |
6fb2915d | 3188 | perf_event_free_filter(event); |
cdd6c482 | 3189 | kfree(event); |
592903cd PZ |
3190 | } |
3191 | ||
76369139 | 3192 | static void ring_buffer_put(struct ring_buffer *rb); |
9bb5d40c | 3193 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb); |
925d519a | 3194 | |
4beb31f3 | 3195 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3196 | { |
4beb31f3 FW |
3197 | if (event->parent) |
3198 | return; | |
3199 | ||
3200 | if (has_branch_stack(event)) { | |
3201 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
3202 | atomic_dec(&per_cpu(perf_branch_stack_events, cpu)); | |
3203 | } | |
3204 | if (is_cgroup_event(event)) | |
3205 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3206 | } | |
925d519a | 3207 | |
4beb31f3 FW |
3208 | static void unaccount_event(struct perf_event *event) |
3209 | { | |
3210 | if (event->parent) | |
3211 | return; | |
3212 | ||
3213 | if (event->attach_state & PERF_ATTACH_TASK) | |
3214 | static_key_slow_dec_deferred(&perf_sched_events); | |
3215 | if (event->attr.mmap || event->attr.mmap_data) | |
3216 | atomic_dec(&nr_mmap_events); | |
3217 | if (event->attr.comm) | |
3218 | atomic_dec(&nr_comm_events); | |
3219 | if (event->attr.task) | |
3220 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3221 | if (event->attr.freq) |
3222 | atomic_dec(&nr_freq_events); | |
4beb31f3 FW |
3223 | if (is_cgroup_event(event)) |
3224 | static_key_slow_dec_deferred(&perf_sched_events); | |
3225 | if (has_branch_stack(event)) | |
3226 | static_key_slow_dec_deferred(&perf_sched_events); | |
3227 | ||
3228 | unaccount_event_cpu(event, event->cpu); | |
3229 | } | |
925d519a | 3230 | |
766d6c07 FW |
3231 | static void __free_event(struct perf_event *event) |
3232 | { | |
cdd6c482 | 3233 | if (!event->parent) { |
927c7a9e FW |
3234 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3235 | put_callchain_buffers(); | |
f344011c | 3236 | } |
9ee318a7 | 3237 | |
766d6c07 FW |
3238 | if (event->destroy) |
3239 | event->destroy(event); | |
3240 | ||
3241 | if (event->ctx) | |
3242 | put_ctx(event->ctx); | |
3243 | ||
3244 | call_rcu(&event->rcu_head, free_event_rcu); | |
3245 | } | |
cdd6c482 | 3246 | static void free_event(struct perf_event *event) |
f1600952 | 3247 | { |
e360adbe | 3248 | irq_work_sync(&event->pending); |
925d519a | 3249 | |
4beb31f3 | 3250 | unaccount_event(event); |
9ee318a7 | 3251 | |
76369139 | 3252 | if (event->rb) { |
9bb5d40c PZ |
3253 | struct ring_buffer *rb; |
3254 | ||
3255 | /* | |
3256 | * Can happen when we close an event with re-directed output. | |
3257 | * | |
3258 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3259 | * over us; possibly making our ring_buffer_put() the last. | |
3260 | */ | |
3261 | mutex_lock(&event->mmap_mutex); | |
3262 | rb = event->rb; | |
3263 | if (rb) { | |
3264 | rcu_assign_pointer(event->rb, NULL); | |
3265 | ring_buffer_detach(event, rb); | |
3266 | ring_buffer_put(rb); /* could be last */ | |
3267 | } | |
3268 | mutex_unlock(&event->mmap_mutex); | |
a4be7c27 PZ |
3269 | } |
3270 | ||
e5d1367f SE |
3271 | if (is_cgroup_event(event)) |
3272 | perf_detach_cgroup(event); | |
3273 | ||
0c67b408 | 3274 | |
766d6c07 | 3275 | __free_event(event); |
f1600952 PZ |
3276 | } |
3277 | ||
a66a3052 | 3278 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 3279 | { |
cdd6c482 | 3280 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 3281 | |
ad3a37de | 3282 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
3283 | /* |
3284 | * There are two ways this annotation is useful: | |
3285 | * | |
3286 | * 1) there is a lock recursion from perf_event_exit_task | |
3287 | * see the comment there. | |
3288 | * | |
3289 | * 2) there is a lock-inversion with mmap_sem through | |
3290 | * perf_event_read_group(), which takes faults while | |
3291 | * holding ctx->mutex, however this is called after | |
3292 | * the last filedesc died, so there is no possibility | |
3293 | * to trigger the AB-BA case. | |
3294 | */ | |
3295 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
050735b0 | 3296 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 3297 | perf_group_detach(event); |
050735b0 | 3298 | raw_spin_unlock_irq(&ctx->lock); |
e03a9a55 | 3299 | perf_remove_from_context(event); |
d859e29f | 3300 | mutex_unlock(&ctx->mutex); |
0793a61d | 3301 | |
cdd6c482 | 3302 | free_event(event); |
0793a61d TG |
3303 | |
3304 | return 0; | |
3305 | } | |
a66a3052 | 3306 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 3307 | |
a66a3052 PZ |
3308 | /* |
3309 | * Called when the last reference to the file is gone. | |
3310 | */ | |
a6fa941d | 3311 | static void put_event(struct perf_event *event) |
fb0459d7 | 3312 | { |
8882135b | 3313 | struct task_struct *owner; |
fb0459d7 | 3314 | |
a6fa941d AV |
3315 | if (!atomic_long_dec_and_test(&event->refcount)) |
3316 | return; | |
fb0459d7 | 3317 | |
8882135b PZ |
3318 | rcu_read_lock(); |
3319 | owner = ACCESS_ONCE(event->owner); | |
3320 | /* | |
3321 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3322 | * !owner it means the list deletion is complete and we can indeed | |
3323 | * free this event, otherwise we need to serialize on | |
3324 | * owner->perf_event_mutex. | |
3325 | */ | |
3326 | smp_read_barrier_depends(); | |
3327 | if (owner) { | |
3328 | /* | |
3329 | * Since delayed_put_task_struct() also drops the last | |
3330 | * task reference we can safely take a new reference | |
3331 | * while holding the rcu_read_lock(). | |
3332 | */ | |
3333 | get_task_struct(owner); | |
3334 | } | |
3335 | rcu_read_unlock(); | |
3336 | ||
3337 | if (owner) { | |
3338 | mutex_lock(&owner->perf_event_mutex); | |
3339 | /* | |
3340 | * We have to re-check the event->owner field, if it is cleared | |
3341 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3342 | * ensured they're done, and we can proceed with freeing the | |
3343 | * event. | |
3344 | */ | |
3345 | if (event->owner) | |
3346 | list_del_init(&event->owner_entry); | |
3347 | mutex_unlock(&owner->perf_event_mutex); | |
3348 | put_task_struct(owner); | |
3349 | } | |
3350 | ||
a6fa941d AV |
3351 | perf_event_release_kernel(event); |
3352 | } | |
3353 | ||
3354 | static int perf_release(struct inode *inode, struct file *file) | |
3355 | { | |
3356 | put_event(file->private_data); | |
3357 | return 0; | |
fb0459d7 | 3358 | } |
fb0459d7 | 3359 | |
59ed446f | 3360 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3361 | { |
cdd6c482 | 3362 | struct perf_event *child; |
e53c0994 PZ |
3363 | u64 total = 0; |
3364 | ||
59ed446f PZ |
3365 | *enabled = 0; |
3366 | *running = 0; | |
3367 | ||
6f10581a | 3368 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3369 | total += perf_event_read(event); |
59ed446f PZ |
3370 | *enabled += event->total_time_enabled + |
3371 | atomic64_read(&event->child_total_time_enabled); | |
3372 | *running += event->total_time_running + | |
3373 | atomic64_read(&event->child_total_time_running); | |
3374 | ||
3375 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3376 | total += perf_event_read(child); |
59ed446f PZ |
3377 | *enabled += child->total_time_enabled; |
3378 | *running += child->total_time_running; | |
3379 | } | |
6f10581a | 3380 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3381 | |
3382 | return total; | |
3383 | } | |
fb0459d7 | 3384 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3385 | |
cdd6c482 | 3386 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3387 | u64 read_format, char __user *buf) |
3388 | { | |
cdd6c482 | 3389 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3390 | int n = 0, size = 0, ret = -EFAULT; |
3391 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3392 | u64 values[5]; |
59ed446f | 3393 | u64 count, enabled, running; |
abf4868b | 3394 | |
6f10581a | 3395 | mutex_lock(&ctx->mutex); |
59ed446f | 3396 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3397 | |
3398 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3399 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3400 | values[n++] = enabled; | |
3401 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3402 | values[n++] = running; | |
abf4868b PZ |
3403 | values[n++] = count; |
3404 | if (read_format & PERF_FORMAT_ID) | |
3405 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3406 | |
3407 | size = n * sizeof(u64); | |
3408 | ||
3409 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3410 | goto unlock; |
3dab77fb | 3411 | |
6f10581a | 3412 | ret = size; |
3dab77fb | 3413 | |
65abc865 | 3414 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3415 | n = 0; |
3dab77fb | 3416 | |
59ed446f | 3417 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3418 | if (read_format & PERF_FORMAT_ID) |
3419 | values[n++] = primary_event_id(sub); | |
3420 | ||
3421 | size = n * sizeof(u64); | |
3422 | ||
184d3da8 | 3423 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3424 | ret = -EFAULT; |
3425 | goto unlock; | |
3426 | } | |
abf4868b PZ |
3427 | |
3428 | ret += size; | |
3dab77fb | 3429 | } |
6f10581a PZ |
3430 | unlock: |
3431 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3432 | |
abf4868b | 3433 | return ret; |
3dab77fb PZ |
3434 | } |
3435 | ||
cdd6c482 | 3436 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3437 | u64 read_format, char __user *buf) |
3438 | { | |
59ed446f | 3439 | u64 enabled, running; |
3dab77fb PZ |
3440 | u64 values[4]; |
3441 | int n = 0; | |
3442 | ||
59ed446f PZ |
3443 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3444 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3445 | values[n++] = enabled; | |
3446 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3447 | values[n++] = running; | |
3dab77fb | 3448 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3449 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3450 | |
3451 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3452 | return -EFAULT; | |
3453 | ||
3454 | return n * sizeof(u64); | |
3455 | } | |
3456 | ||
0793a61d | 3457 | /* |
cdd6c482 | 3458 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3459 | */ |
3460 | static ssize_t | |
cdd6c482 | 3461 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3462 | { |
cdd6c482 | 3463 | u64 read_format = event->attr.read_format; |
3dab77fb | 3464 | int ret; |
0793a61d | 3465 | |
3b6f9e5c | 3466 | /* |
cdd6c482 | 3467 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3468 | * error state (i.e. because it was pinned but it couldn't be |
3469 | * scheduled on to the CPU at some point). | |
3470 | */ | |
cdd6c482 | 3471 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3472 | return 0; |
3473 | ||
c320c7b7 | 3474 | if (count < event->read_size) |
3dab77fb PZ |
3475 | return -ENOSPC; |
3476 | ||
cdd6c482 | 3477 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3478 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3479 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3480 | else |
cdd6c482 | 3481 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3482 | |
3dab77fb | 3483 | return ret; |
0793a61d TG |
3484 | } |
3485 | ||
0793a61d TG |
3486 | static ssize_t |
3487 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3488 | { | |
cdd6c482 | 3489 | struct perf_event *event = file->private_data; |
0793a61d | 3490 | |
cdd6c482 | 3491 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3492 | } |
3493 | ||
3494 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3495 | { | |
cdd6c482 | 3496 | struct perf_event *event = file->private_data; |
76369139 | 3497 | struct ring_buffer *rb; |
c33a0bc4 | 3498 | unsigned int events = POLL_HUP; |
c7138f37 | 3499 | |
10c6db11 | 3500 | /* |
9bb5d40c PZ |
3501 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3502 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3503 | */ |
3504 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3505 | rb = event->rb; |
3506 | if (rb) | |
76369139 | 3507 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 PZ |
3508 | mutex_unlock(&event->mmap_mutex); |
3509 | ||
cdd6c482 | 3510 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3511 | |
0793a61d TG |
3512 | return events; |
3513 | } | |
3514 | ||
cdd6c482 | 3515 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3516 | { |
cdd6c482 | 3517 | (void)perf_event_read(event); |
e7850595 | 3518 | local64_set(&event->count, 0); |
cdd6c482 | 3519 | perf_event_update_userpage(event); |
3df5edad PZ |
3520 | } |
3521 | ||
c93f7669 | 3522 | /* |
cdd6c482 IM |
3523 | * Holding the top-level event's child_mutex means that any |
3524 | * descendant process that has inherited this event will block | |
3525 | * in sync_child_event if it goes to exit, thus satisfying the | |
3526 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3527 | */ |
cdd6c482 IM |
3528 | static void perf_event_for_each_child(struct perf_event *event, |
3529 | void (*func)(struct perf_event *)) | |
3df5edad | 3530 | { |
cdd6c482 | 3531 | struct perf_event *child; |
3df5edad | 3532 | |
cdd6c482 IM |
3533 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3534 | mutex_lock(&event->child_mutex); | |
3535 | func(event); | |
3536 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3537 | func(child); |
cdd6c482 | 3538 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3539 | } |
3540 | ||
cdd6c482 IM |
3541 | static void perf_event_for_each(struct perf_event *event, |
3542 | void (*func)(struct perf_event *)) | |
3df5edad | 3543 | { |
cdd6c482 IM |
3544 | struct perf_event_context *ctx = event->ctx; |
3545 | struct perf_event *sibling; | |
3df5edad | 3546 | |
75f937f2 PZ |
3547 | WARN_ON_ONCE(ctx->parent_ctx); |
3548 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3549 | event = event->group_leader; |
75f937f2 | 3550 | |
cdd6c482 | 3551 | perf_event_for_each_child(event, func); |
cdd6c482 | 3552 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3553 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3554 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3555 | } |
3556 | ||
cdd6c482 | 3557 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3558 | { |
cdd6c482 | 3559 | struct perf_event_context *ctx = event->ctx; |
bad7192b | 3560 | int ret = 0, active; |
08247e31 PZ |
3561 | u64 value; |
3562 | ||
6c7e550f | 3563 | if (!is_sampling_event(event)) |
08247e31 PZ |
3564 | return -EINVAL; |
3565 | ||
ad0cf347 | 3566 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3567 | return -EFAULT; |
3568 | ||
3569 | if (!value) | |
3570 | return -EINVAL; | |
3571 | ||
e625cce1 | 3572 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3573 | if (event->attr.freq) { |
3574 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3575 | ret = -EINVAL; |
3576 | goto unlock; | |
3577 | } | |
3578 | ||
cdd6c482 | 3579 | event->attr.sample_freq = value; |
08247e31 | 3580 | } else { |
cdd6c482 IM |
3581 | event->attr.sample_period = value; |
3582 | event->hw.sample_period = value; | |
08247e31 | 3583 | } |
bad7192b PZ |
3584 | |
3585 | active = (event->state == PERF_EVENT_STATE_ACTIVE); | |
3586 | if (active) { | |
3587 | perf_pmu_disable(ctx->pmu); | |
3588 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3589 | } | |
3590 | ||
3591 | local64_set(&event->hw.period_left, 0); | |
3592 | ||
3593 | if (active) { | |
3594 | event->pmu->start(event, PERF_EF_RELOAD); | |
3595 | perf_pmu_enable(ctx->pmu); | |
3596 | } | |
3597 | ||
08247e31 | 3598 | unlock: |
e625cce1 | 3599 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3600 | |
3601 | return ret; | |
3602 | } | |
3603 | ||
ac9721f3 PZ |
3604 | static const struct file_operations perf_fops; |
3605 | ||
2903ff01 | 3606 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3607 | { |
2903ff01 AV |
3608 | struct fd f = fdget(fd); |
3609 | if (!f.file) | |
3610 | return -EBADF; | |
ac9721f3 | 3611 | |
2903ff01 AV |
3612 | if (f.file->f_op != &perf_fops) { |
3613 | fdput(f); | |
3614 | return -EBADF; | |
ac9721f3 | 3615 | } |
2903ff01 AV |
3616 | *p = f; |
3617 | return 0; | |
ac9721f3 PZ |
3618 | } |
3619 | ||
3620 | static int perf_event_set_output(struct perf_event *event, | |
3621 | struct perf_event *output_event); | |
6fb2915d | 3622 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3623 | |
d859e29f PM |
3624 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3625 | { | |
cdd6c482 IM |
3626 | struct perf_event *event = file->private_data; |
3627 | void (*func)(struct perf_event *); | |
3df5edad | 3628 | u32 flags = arg; |
d859e29f PM |
3629 | |
3630 | switch (cmd) { | |
cdd6c482 IM |
3631 | case PERF_EVENT_IOC_ENABLE: |
3632 | func = perf_event_enable; | |
d859e29f | 3633 | break; |
cdd6c482 IM |
3634 | case PERF_EVENT_IOC_DISABLE: |
3635 | func = perf_event_disable; | |
79f14641 | 3636 | break; |
cdd6c482 IM |
3637 | case PERF_EVENT_IOC_RESET: |
3638 | func = perf_event_reset; | |
6de6a7b9 | 3639 | break; |
3df5edad | 3640 | |
cdd6c482 IM |
3641 | case PERF_EVENT_IOC_REFRESH: |
3642 | return perf_event_refresh(event, arg); | |
08247e31 | 3643 | |
cdd6c482 IM |
3644 | case PERF_EVENT_IOC_PERIOD: |
3645 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3646 | |
cf4957f1 JO |
3647 | case PERF_EVENT_IOC_ID: |
3648 | { | |
3649 | u64 id = primary_event_id(event); | |
3650 | ||
3651 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
3652 | return -EFAULT; | |
3653 | return 0; | |
3654 | } | |
3655 | ||
cdd6c482 | 3656 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3657 | { |
ac9721f3 | 3658 | int ret; |
ac9721f3 | 3659 | if (arg != -1) { |
2903ff01 AV |
3660 | struct perf_event *output_event; |
3661 | struct fd output; | |
3662 | ret = perf_fget_light(arg, &output); | |
3663 | if (ret) | |
3664 | return ret; | |
3665 | output_event = output.file->private_data; | |
3666 | ret = perf_event_set_output(event, output_event); | |
3667 | fdput(output); | |
3668 | } else { | |
3669 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3670 | } |
ac9721f3 PZ |
3671 | return ret; |
3672 | } | |
a4be7c27 | 3673 | |
6fb2915d LZ |
3674 | case PERF_EVENT_IOC_SET_FILTER: |
3675 | return perf_event_set_filter(event, (void __user *)arg); | |
3676 | ||
d859e29f | 3677 | default: |
3df5edad | 3678 | return -ENOTTY; |
d859e29f | 3679 | } |
3df5edad PZ |
3680 | |
3681 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3682 | perf_event_for_each(event, func); |
3df5edad | 3683 | else |
cdd6c482 | 3684 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3685 | |
3686 | return 0; | |
d859e29f PM |
3687 | } |
3688 | ||
cdd6c482 | 3689 | int perf_event_task_enable(void) |
771d7cde | 3690 | { |
cdd6c482 | 3691 | struct perf_event *event; |
771d7cde | 3692 | |
cdd6c482 IM |
3693 | mutex_lock(¤t->perf_event_mutex); |
3694 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3695 | perf_event_for_each_child(event, perf_event_enable); | |
3696 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3697 | |
3698 | return 0; | |
3699 | } | |
3700 | ||
cdd6c482 | 3701 | int perf_event_task_disable(void) |
771d7cde | 3702 | { |
cdd6c482 | 3703 | struct perf_event *event; |
771d7cde | 3704 | |
cdd6c482 IM |
3705 | mutex_lock(¤t->perf_event_mutex); |
3706 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3707 | perf_event_for_each_child(event, perf_event_disable); | |
3708 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3709 | |
3710 | return 0; | |
3711 | } | |
3712 | ||
cdd6c482 | 3713 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3714 | { |
a4eaf7f1 PZ |
3715 | if (event->hw.state & PERF_HES_STOPPED) |
3716 | return 0; | |
3717 | ||
cdd6c482 | 3718 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3719 | return 0; |
3720 | ||
35edc2a5 | 3721 | return event->pmu->event_idx(event); |
194002b2 PZ |
3722 | } |
3723 | ||
c4794295 | 3724 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3725 | u64 *now, |
7f310a5d EM |
3726 | u64 *enabled, |
3727 | u64 *running) | |
c4794295 | 3728 | { |
e3f3541c | 3729 | u64 ctx_time; |
c4794295 | 3730 | |
e3f3541c PZ |
3731 | *now = perf_clock(); |
3732 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3733 | *enabled = ctx_time - event->tstamp_enabled; |
3734 | *running = ctx_time - event->tstamp_running; | |
3735 | } | |
3736 | ||
fa731587 PZ |
3737 | static void perf_event_init_userpage(struct perf_event *event) |
3738 | { | |
3739 | struct perf_event_mmap_page *userpg; | |
3740 | struct ring_buffer *rb; | |
3741 | ||
3742 | rcu_read_lock(); | |
3743 | rb = rcu_dereference(event->rb); | |
3744 | if (!rb) | |
3745 | goto unlock; | |
3746 | ||
3747 | userpg = rb->user_page; | |
3748 | ||
3749 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
3750 | userpg->cap_bit0_is_deprecated = 1; | |
3751 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
3752 | ||
3753 | unlock: | |
3754 | rcu_read_unlock(); | |
3755 | } | |
3756 | ||
c7206205 | 3757 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3758 | { |
3759 | } | |
3760 | ||
38ff667b PZ |
3761 | /* |
3762 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3763 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3764 | * code calls this from NMI context. | |
3765 | */ | |
cdd6c482 | 3766 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3767 | { |
cdd6c482 | 3768 | struct perf_event_mmap_page *userpg; |
76369139 | 3769 | struct ring_buffer *rb; |
e3f3541c | 3770 | u64 enabled, running, now; |
38ff667b PZ |
3771 | |
3772 | rcu_read_lock(); | |
5ec4c599 PZ |
3773 | rb = rcu_dereference(event->rb); |
3774 | if (!rb) | |
3775 | goto unlock; | |
3776 | ||
0d641208 EM |
3777 | /* |
3778 | * compute total_time_enabled, total_time_running | |
3779 | * based on snapshot values taken when the event | |
3780 | * was last scheduled in. | |
3781 | * | |
3782 | * we cannot simply called update_context_time() | |
3783 | * because of locking issue as we can be called in | |
3784 | * NMI context | |
3785 | */ | |
e3f3541c | 3786 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 3787 | |
76369139 | 3788 | userpg = rb->user_page; |
7b732a75 PZ |
3789 | /* |
3790 | * Disable preemption so as to not let the corresponding user-space | |
3791 | * spin too long if we get preempted. | |
3792 | */ | |
3793 | preempt_disable(); | |
37d81828 | 3794 | ++userpg->lock; |
92f22a38 | 3795 | barrier(); |
cdd6c482 | 3796 | userpg->index = perf_event_index(event); |
b5e58793 | 3797 | userpg->offset = perf_event_count(event); |
365a4038 | 3798 | if (userpg->index) |
e7850595 | 3799 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3800 | |
0d641208 | 3801 | userpg->time_enabled = enabled + |
cdd6c482 | 3802 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3803 | |
0d641208 | 3804 | userpg->time_running = running + |
cdd6c482 | 3805 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3806 | |
c7206205 | 3807 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3808 | |
92f22a38 | 3809 | barrier(); |
37d81828 | 3810 | ++userpg->lock; |
7b732a75 | 3811 | preempt_enable(); |
38ff667b | 3812 | unlock: |
7b732a75 | 3813 | rcu_read_unlock(); |
37d81828 PM |
3814 | } |
3815 | ||
906010b2 PZ |
3816 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3817 | { | |
3818 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3819 | struct ring_buffer *rb; |
906010b2 PZ |
3820 | int ret = VM_FAULT_SIGBUS; |
3821 | ||
3822 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3823 | if (vmf->pgoff == 0) | |
3824 | ret = 0; | |
3825 | return ret; | |
3826 | } | |
3827 | ||
3828 | rcu_read_lock(); | |
76369139 FW |
3829 | rb = rcu_dereference(event->rb); |
3830 | if (!rb) | |
906010b2 PZ |
3831 | goto unlock; |
3832 | ||
3833 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3834 | goto unlock; | |
3835 | ||
76369139 | 3836 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3837 | if (!vmf->page) |
3838 | goto unlock; | |
3839 | ||
3840 | get_page(vmf->page); | |
3841 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3842 | vmf->page->index = vmf->pgoff; | |
3843 | ||
3844 | ret = 0; | |
3845 | unlock: | |
3846 | rcu_read_unlock(); | |
3847 | ||
3848 | return ret; | |
3849 | } | |
3850 | ||
10c6db11 PZ |
3851 | static void ring_buffer_attach(struct perf_event *event, |
3852 | struct ring_buffer *rb) | |
3853 | { | |
3854 | unsigned long flags; | |
3855 | ||
3856 | if (!list_empty(&event->rb_entry)) | |
3857 | return; | |
3858 | ||
3859 | spin_lock_irqsave(&rb->event_lock, flags); | |
9bb5d40c PZ |
3860 | if (list_empty(&event->rb_entry)) |
3861 | list_add(&event->rb_entry, &rb->event_list); | |
10c6db11 PZ |
3862 | spin_unlock_irqrestore(&rb->event_lock, flags); |
3863 | } | |
3864 | ||
9bb5d40c | 3865 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb) |
10c6db11 PZ |
3866 | { |
3867 | unsigned long flags; | |
3868 | ||
3869 | if (list_empty(&event->rb_entry)) | |
3870 | return; | |
3871 | ||
3872 | spin_lock_irqsave(&rb->event_lock, flags); | |
3873 | list_del_init(&event->rb_entry); | |
3874 | wake_up_all(&event->waitq); | |
3875 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3876 | } | |
3877 | ||
3878 | static void ring_buffer_wakeup(struct perf_event *event) | |
3879 | { | |
3880 | struct ring_buffer *rb; | |
3881 | ||
3882 | rcu_read_lock(); | |
3883 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
3884 | if (rb) { |
3885 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
3886 | wake_up_all(&event->waitq); | |
3887 | } | |
10c6db11 PZ |
3888 | rcu_read_unlock(); |
3889 | } | |
3890 | ||
76369139 | 3891 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3892 | { |
76369139 | 3893 | struct ring_buffer *rb; |
906010b2 | 3894 | |
76369139 FW |
3895 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3896 | rb_free(rb); | |
7b732a75 PZ |
3897 | } |
3898 | ||
76369139 | 3899 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3900 | { |
76369139 | 3901 | struct ring_buffer *rb; |
7b732a75 | 3902 | |
ac9721f3 | 3903 | rcu_read_lock(); |
76369139 FW |
3904 | rb = rcu_dereference(event->rb); |
3905 | if (rb) { | |
3906 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3907 | rb = NULL; | |
ac9721f3 PZ |
3908 | } |
3909 | rcu_read_unlock(); | |
3910 | ||
76369139 | 3911 | return rb; |
ac9721f3 PZ |
3912 | } |
3913 | ||
76369139 | 3914 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3915 | { |
76369139 | 3916 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3917 | return; |
7b732a75 | 3918 | |
9bb5d40c | 3919 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 3920 | |
76369139 | 3921 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3922 | } |
3923 | ||
3924 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3925 | { | |
cdd6c482 | 3926 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3927 | |
cdd6c482 | 3928 | atomic_inc(&event->mmap_count); |
9bb5d40c | 3929 | atomic_inc(&event->rb->mmap_count); |
7b732a75 PZ |
3930 | } |
3931 | ||
9bb5d40c PZ |
3932 | /* |
3933 | * A buffer can be mmap()ed multiple times; either directly through the same | |
3934 | * event, or through other events by use of perf_event_set_output(). | |
3935 | * | |
3936 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
3937 | * the buffer here, where we still have a VM context. This means we need | |
3938 | * to detach all events redirecting to us. | |
3939 | */ | |
7b732a75 PZ |
3940 | static void perf_mmap_close(struct vm_area_struct *vma) |
3941 | { | |
cdd6c482 | 3942 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3943 | |
9bb5d40c PZ |
3944 | struct ring_buffer *rb = event->rb; |
3945 | struct user_struct *mmap_user = rb->mmap_user; | |
3946 | int mmap_locked = rb->mmap_locked; | |
3947 | unsigned long size = perf_data_size(rb); | |
789f90fc | 3948 | |
9bb5d40c PZ |
3949 | atomic_dec(&rb->mmap_count); |
3950 | ||
3951 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
3952 | return; | |
3953 | ||
3954 | /* Detach current event from the buffer. */ | |
3955 | rcu_assign_pointer(event->rb, NULL); | |
3956 | ring_buffer_detach(event, rb); | |
3957 | mutex_unlock(&event->mmap_mutex); | |
3958 | ||
3959 | /* If there's still other mmap()s of this buffer, we're done. */ | |
3960 | if (atomic_read(&rb->mmap_count)) { | |
3961 | ring_buffer_put(rb); /* can't be last */ | |
3962 | return; | |
3963 | } | |
ac9721f3 | 3964 | |
9bb5d40c PZ |
3965 | /* |
3966 | * No other mmap()s, detach from all other events that might redirect | |
3967 | * into the now unreachable buffer. Somewhat complicated by the | |
3968 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
3969 | */ | |
3970 | again: | |
3971 | rcu_read_lock(); | |
3972 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
3973 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
3974 | /* | |
3975 | * This event is en-route to free_event() which will | |
3976 | * detach it and remove it from the list. | |
3977 | */ | |
3978 | continue; | |
3979 | } | |
3980 | rcu_read_unlock(); | |
789f90fc | 3981 | |
9bb5d40c PZ |
3982 | mutex_lock(&event->mmap_mutex); |
3983 | /* | |
3984 | * Check we didn't race with perf_event_set_output() which can | |
3985 | * swizzle the rb from under us while we were waiting to | |
3986 | * acquire mmap_mutex. | |
3987 | * | |
3988 | * If we find a different rb; ignore this event, a next | |
3989 | * iteration will no longer find it on the list. We have to | |
3990 | * still restart the iteration to make sure we're not now | |
3991 | * iterating the wrong list. | |
3992 | */ | |
3993 | if (event->rb == rb) { | |
3994 | rcu_assign_pointer(event->rb, NULL); | |
3995 | ring_buffer_detach(event, rb); | |
3996 | ring_buffer_put(rb); /* can't be last, we still have one */ | |
26cb63ad | 3997 | } |
cdd6c482 | 3998 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 3999 | put_event(event); |
ac9721f3 | 4000 | |
9bb5d40c PZ |
4001 | /* |
4002 | * Restart the iteration; either we're on the wrong list or | |
4003 | * destroyed its integrity by doing a deletion. | |
4004 | */ | |
4005 | goto again; | |
7b732a75 | 4006 | } |
9bb5d40c PZ |
4007 | rcu_read_unlock(); |
4008 | ||
4009 | /* | |
4010 | * It could be there's still a few 0-ref events on the list; they'll | |
4011 | * get cleaned up by free_event() -- they'll also still have their | |
4012 | * ref on the rb and will free it whenever they are done with it. | |
4013 | * | |
4014 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
4015 | * undo the VM accounting. | |
4016 | */ | |
4017 | ||
4018 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
4019 | vma->vm_mm->pinned_vm -= mmap_locked; | |
4020 | free_uid(mmap_user); | |
4021 | ||
4022 | ring_buffer_put(rb); /* could be last */ | |
37d81828 PM |
4023 | } |
4024 | ||
f0f37e2f | 4025 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
4026 | .open = perf_mmap_open, |
4027 | .close = perf_mmap_close, | |
4028 | .fault = perf_mmap_fault, | |
4029 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
4030 | }; |
4031 | ||
4032 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
4033 | { | |
cdd6c482 | 4034 | struct perf_event *event = file->private_data; |
22a4f650 | 4035 | unsigned long user_locked, user_lock_limit; |
789f90fc | 4036 | struct user_struct *user = current_user(); |
22a4f650 | 4037 | unsigned long locked, lock_limit; |
76369139 | 4038 | struct ring_buffer *rb; |
7b732a75 PZ |
4039 | unsigned long vma_size; |
4040 | unsigned long nr_pages; | |
789f90fc | 4041 | long user_extra, extra; |
d57e34fd | 4042 | int ret = 0, flags = 0; |
37d81828 | 4043 | |
c7920614 PZ |
4044 | /* |
4045 | * Don't allow mmap() of inherited per-task counters. This would | |
4046 | * create a performance issue due to all children writing to the | |
76369139 | 4047 | * same rb. |
c7920614 PZ |
4048 | */ |
4049 | if (event->cpu == -1 && event->attr.inherit) | |
4050 | return -EINVAL; | |
4051 | ||
43a21ea8 | 4052 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 4053 | return -EINVAL; |
7b732a75 PZ |
4054 | |
4055 | vma_size = vma->vm_end - vma->vm_start; | |
4056 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
4057 | ||
7730d865 | 4058 | /* |
76369139 | 4059 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
4060 | * can do bitmasks instead of modulo. |
4061 | */ | |
4062 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
4063 | return -EINVAL; |
4064 | ||
7b732a75 | 4065 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
4066 | return -EINVAL; |
4067 | ||
7b732a75 PZ |
4068 | if (vma->vm_pgoff != 0) |
4069 | return -EINVAL; | |
37d81828 | 4070 | |
cdd6c482 | 4071 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 4072 | again: |
cdd6c482 | 4073 | mutex_lock(&event->mmap_mutex); |
76369139 | 4074 | if (event->rb) { |
9bb5d40c | 4075 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4076 | ret = -EINVAL; |
9bb5d40c PZ |
4077 | goto unlock; |
4078 | } | |
4079 | ||
4080 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4081 | /* | |
4082 | * Raced against perf_mmap_close() through | |
4083 | * perf_event_set_output(). Try again, hope for better | |
4084 | * luck. | |
4085 | */ | |
4086 | mutex_unlock(&event->mmap_mutex); | |
4087 | goto again; | |
4088 | } | |
4089 | ||
ebb3c4c4 PZ |
4090 | goto unlock; |
4091 | } | |
4092 | ||
789f90fc | 4093 | user_extra = nr_pages + 1; |
cdd6c482 | 4094 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4095 | |
4096 | /* | |
4097 | * Increase the limit linearly with more CPUs: | |
4098 | */ | |
4099 | user_lock_limit *= num_online_cpus(); | |
4100 | ||
789f90fc | 4101 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4102 | |
789f90fc PZ |
4103 | extra = 0; |
4104 | if (user_locked > user_lock_limit) | |
4105 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4106 | |
78d7d407 | 4107 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4108 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4109 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4110 | |
459ec28a IM |
4111 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4112 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4113 | ret = -EPERM; |
4114 | goto unlock; | |
4115 | } | |
7b732a75 | 4116 | |
76369139 | 4117 | WARN_ON(event->rb); |
906010b2 | 4118 | |
d57e34fd | 4119 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4120 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4121 | |
4ec8363d VW |
4122 | rb = rb_alloc(nr_pages, |
4123 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4124 | event->cpu, flags); | |
4125 | ||
76369139 | 4126 | if (!rb) { |
ac9721f3 | 4127 | ret = -ENOMEM; |
ebb3c4c4 | 4128 | goto unlock; |
ac9721f3 | 4129 | } |
26cb63ad | 4130 | |
9bb5d40c | 4131 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
4132 | rb->mmap_locked = extra; |
4133 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 4134 | |
ac9721f3 | 4135 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
4136 | vma->vm_mm->pinned_vm += extra; |
4137 | ||
9bb5d40c | 4138 | ring_buffer_attach(event, rb); |
26cb63ad | 4139 | rcu_assign_pointer(event->rb, rb); |
ac9721f3 | 4140 | |
fa731587 | 4141 | perf_event_init_userpage(event); |
9a0f05cb PZ |
4142 | perf_event_update_userpage(event); |
4143 | ||
ebb3c4c4 | 4144 | unlock: |
ac9721f3 PZ |
4145 | if (!ret) |
4146 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 4147 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4148 | |
9bb5d40c PZ |
4149 | /* |
4150 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4151 | * vma. | |
4152 | */ | |
26cb63ad | 4153 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4154 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
4155 | |
4156 | return ret; | |
37d81828 PM |
4157 | } |
4158 | ||
3c446b3d PZ |
4159 | static int perf_fasync(int fd, struct file *filp, int on) |
4160 | { | |
496ad9aa | 4161 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4162 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4163 | int retval; |
4164 | ||
4165 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4166 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4167 | mutex_unlock(&inode->i_mutex); |
4168 | ||
4169 | if (retval < 0) | |
4170 | return retval; | |
4171 | ||
4172 | return 0; | |
4173 | } | |
4174 | ||
0793a61d | 4175 | static const struct file_operations perf_fops = { |
3326c1ce | 4176 | .llseek = no_llseek, |
0793a61d TG |
4177 | .release = perf_release, |
4178 | .read = perf_read, | |
4179 | .poll = perf_poll, | |
d859e29f PM |
4180 | .unlocked_ioctl = perf_ioctl, |
4181 | .compat_ioctl = perf_ioctl, | |
37d81828 | 4182 | .mmap = perf_mmap, |
3c446b3d | 4183 | .fasync = perf_fasync, |
0793a61d TG |
4184 | }; |
4185 | ||
925d519a | 4186 | /* |
cdd6c482 | 4187 | * Perf event wakeup |
925d519a PZ |
4188 | * |
4189 | * If there's data, ensure we set the poll() state and publish everything | |
4190 | * to user-space before waking everybody up. | |
4191 | */ | |
4192 | ||
cdd6c482 | 4193 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4194 | { |
10c6db11 | 4195 | ring_buffer_wakeup(event); |
4c9e2542 | 4196 | |
cdd6c482 IM |
4197 | if (event->pending_kill) { |
4198 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4199 | event->pending_kill = 0; | |
4c9e2542 | 4200 | } |
925d519a PZ |
4201 | } |
4202 | ||
e360adbe | 4203 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4204 | { |
cdd6c482 IM |
4205 | struct perf_event *event = container_of(entry, |
4206 | struct perf_event, pending); | |
79f14641 | 4207 | |
cdd6c482 IM |
4208 | if (event->pending_disable) { |
4209 | event->pending_disable = 0; | |
4210 | __perf_event_disable(event); | |
79f14641 PZ |
4211 | } |
4212 | ||
cdd6c482 IM |
4213 | if (event->pending_wakeup) { |
4214 | event->pending_wakeup = 0; | |
4215 | perf_event_wakeup(event); | |
79f14641 PZ |
4216 | } |
4217 | } | |
4218 | ||
39447b38 ZY |
4219 | /* |
4220 | * We assume there is only KVM supporting the callbacks. | |
4221 | * Later on, we might change it to a list if there is | |
4222 | * another virtualization implementation supporting the callbacks. | |
4223 | */ | |
4224 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4225 | ||
4226 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4227 | { | |
4228 | perf_guest_cbs = cbs; | |
4229 | return 0; | |
4230 | } | |
4231 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4232 | ||
4233 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4234 | { | |
4235 | perf_guest_cbs = NULL; | |
4236 | return 0; | |
4237 | } | |
4238 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4239 | ||
4018994f JO |
4240 | static void |
4241 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4242 | struct pt_regs *regs, u64 mask) | |
4243 | { | |
4244 | int bit; | |
4245 | ||
4246 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4247 | sizeof(mask) * BITS_PER_BYTE) { | |
4248 | u64 val; | |
4249 | ||
4250 | val = perf_reg_value(regs, bit); | |
4251 | perf_output_put(handle, val); | |
4252 | } | |
4253 | } | |
4254 | ||
4255 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
4256 | struct pt_regs *regs) | |
4257 | { | |
4258 | if (!user_mode(regs)) { | |
4259 | if (current->mm) | |
4260 | regs = task_pt_regs(current); | |
4261 | else | |
4262 | regs = NULL; | |
4263 | } | |
4264 | ||
4265 | if (regs) { | |
4266 | regs_user->regs = regs; | |
4267 | regs_user->abi = perf_reg_abi(current); | |
4268 | } | |
4269 | } | |
4270 | ||
c5ebcedb JO |
4271 | /* |
4272 | * Get remaining task size from user stack pointer. | |
4273 | * | |
4274 | * It'd be better to take stack vma map and limit this more | |
4275 | * precisly, but there's no way to get it safely under interrupt, | |
4276 | * so using TASK_SIZE as limit. | |
4277 | */ | |
4278 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4279 | { | |
4280 | unsigned long addr = perf_user_stack_pointer(regs); | |
4281 | ||
4282 | if (!addr || addr >= TASK_SIZE) | |
4283 | return 0; | |
4284 | ||
4285 | return TASK_SIZE - addr; | |
4286 | } | |
4287 | ||
4288 | static u16 | |
4289 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4290 | struct pt_regs *regs) | |
4291 | { | |
4292 | u64 task_size; | |
4293 | ||
4294 | /* No regs, no stack pointer, no dump. */ | |
4295 | if (!regs) | |
4296 | return 0; | |
4297 | ||
4298 | /* | |
4299 | * Check if we fit in with the requested stack size into the: | |
4300 | * - TASK_SIZE | |
4301 | * If we don't, we limit the size to the TASK_SIZE. | |
4302 | * | |
4303 | * - remaining sample size | |
4304 | * If we don't, we customize the stack size to | |
4305 | * fit in to the remaining sample size. | |
4306 | */ | |
4307 | ||
4308 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4309 | stack_size = min(stack_size, (u16) task_size); | |
4310 | ||
4311 | /* Current header size plus static size and dynamic size. */ | |
4312 | header_size += 2 * sizeof(u64); | |
4313 | ||
4314 | /* Do we fit in with the current stack dump size? */ | |
4315 | if ((u16) (header_size + stack_size) < header_size) { | |
4316 | /* | |
4317 | * If we overflow the maximum size for the sample, | |
4318 | * we customize the stack dump size to fit in. | |
4319 | */ | |
4320 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4321 | stack_size = round_up(stack_size, sizeof(u64)); | |
4322 | } | |
4323 | ||
4324 | return stack_size; | |
4325 | } | |
4326 | ||
4327 | static void | |
4328 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4329 | struct pt_regs *regs) | |
4330 | { | |
4331 | /* Case of a kernel thread, nothing to dump */ | |
4332 | if (!regs) { | |
4333 | u64 size = 0; | |
4334 | perf_output_put(handle, size); | |
4335 | } else { | |
4336 | unsigned long sp; | |
4337 | unsigned int rem; | |
4338 | u64 dyn_size; | |
4339 | ||
4340 | /* | |
4341 | * We dump: | |
4342 | * static size | |
4343 | * - the size requested by user or the best one we can fit | |
4344 | * in to the sample max size | |
4345 | * data | |
4346 | * - user stack dump data | |
4347 | * dynamic size | |
4348 | * - the actual dumped size | |
4349 | */ | |
4350 | ||
4351 | /* Static size. */ | |
4352 | perf_output_put(handle, dump_size); | |
4353 | ||
4354 | /* Data. */ | |
4355 | sp = perf_user_stack_pointer(regs); | |
4356 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4357 | dyn_size = dump_size - rem; | |
4358 | ||
4359 | perf_output_skip(handle, rem); | |
4360 | ||
4361 | /* Dynamic size. */ | |
4362 | perf_output_put(handle, dyn_size); | |
4363 | } | |
4364 | } | |
4365 | ||
c980d109 ACM |
4366 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4367 | struct perf_sample_data *data, | |
4368 | struct perf_event *event) | |
6844c09d ACM |
4369 | { |
4370 | u64 sample_type = event->attr.sample_type; | |
4371 | ||
4372 | data->type = sample_type; | |
4373 | header->size += event->id_header_size; | |
4374 | ||
4375 | if (sample_type & PERF_SAMPLE_TID) { | |
4376 | /* namespace issues */ | |
4377 | data->tid_entry.pid = perf_event_pid(event, current); | |
4378 | data->tid_entry.tid = perf_event_tid(event, current); | |
4379 | } | |
4380 | ||
4381 | if (sample_type & PERF_SAMPLE_TIME) | |
4382 | data->time = perf_clock(); | |
4383 | ||
ff3d527c | 4384 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4385 | data->id = primary_event_id(event); |
4386 | ||
4387 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4388 | data->stream_id = event->id; | |
4389 | ||
4390 | if (sample_type & PERF_SAMPLE_CPU) { | |
4391 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4392 | data->cpu_entry.reserved = 0; | |
4393 | } | |
4394 | } | |
4395 | ||
76369139 FW |
4396 | void perf_event_header__init_id(struct perf_event_header *header, |
4397 | struct perf_sample_data *data, | |
4398 | struct perf_event *event) | |
c980d109 ACM |
4399 | { |
4400 | if (event->attr.sample_id_all) | |
4401 | __perf_event_header__init_id(header, data, event); | |
4402 | } | |
4403 | ||
4404 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4405 | struct perf_sample_data *data) | |
4406 | { | |
4407 | u64 sample_type = data->type; | |
4408 | ||
4409 | if (sample_type & PERF_SAMPLE_TID) | |
4410 | perf_output_put(handle, data->tid_entry); | |
4411 | ||
4412 | if (sample_type & PERF_SAMPLE_TIME) | |
4413 | perf_output_put(handle, data->time); | |
4414 | ||
4415 | if (sample_type & PERF_SAMPLE_ID) | |
4416 | perf_output_put(handle, data->id); | |
4417 | ||
4418 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4419 | perf_output_put(handle, data->stream_id); | |
4420 | ||
4421 | if (sample_type & PERF_SAMPLE_CPU) | |
4422 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
4423 | |
4424 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
4425 | perf_output_put(handle, data->id); | |
c980d109 ACM |
4426 | } |
4427 | ||
76369139 FW |
4428 | void perf_event__output_id_sample(struct perf_event *event, |
4429 | struct perf_output_handle *handle, | |
4430 | struct perf_sample_data *sample) | |
c980d109 ACM |
4431 | { |
4432 | if (event->attr.sample_id_all) | |
4433 | __perf_event__output_id_sample(handle, sample); | |
4434 | } | |
4435 | ||
3dab77fb | 4436 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4437 | struct perf_event *event, |
4438 | u64 enabled, u64 running) | |
3dab77fb | 4439 | { |
cdd6c482 | 4440 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4441 | u64 values[4]; |
4442 | int n = 0; | |
4443 | ||
b5e58793 | 4444 | values[n++] = perf_event_count(event); |
3dab77fb | 4445 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4446 | values[n++] = enabled + |
cdd6c482 | 4447 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4448 | } |
4449 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4450 | values[n++] = running + |
cdd6c482 | 4451 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4452 | } |
4453 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4454 | values[n++] = primary_event_id(event); |
3dab77fb | 4455 | |
76369139 | 4456 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4457 | } |
4458 | ||
4459 | /* | |
cdd6c482 | 4460 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4461 | */ |
4462 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4463 | struct perf_event *event, |
4464 | u64 enabled, u64 running) | |
3dab77fb | 4465 | { |
cdd6c482 IM |
4466 | struct perf_event *leader = event->group_leader, *sub; |
4467 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4468 | u64 values[5]; |
4469 | int n = 0; | |
4470 | ||
4471 | values[n++] = 1 + leader->nr_siblings; | |
4472 | ||
4473 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4474 | values[n++] = enabled; |
3dab77fb PZ |
4475 | |
4476 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4477 | values[n++] = running; |
3dab77fb | 4478 | |
cdd6c482 | 4479 | if (leader != event) |
3dab77fb PZ |
4480 | leader->pmu->read(leader); |
4481 | ||
b5e58793 | 4482 | values[n++] = perf_event_count(leader); |
3dab77fb | 4483 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4484 | values[n++] = primary_event_id(leader); |
3dab77fb | 4485 | |
76369139 | 4486 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4487 | |
65abc865 | 4488 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4489 | n = 0; |
4490 | ||
6f5ab001 JO |
4491 | if ((sub != event) && |
4492 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
4493 | sub->pmu->read(sub); |
4494 | ||
b5e58793 | 4495 | values[n++] = perf_event_count(sub); |
3dab77fb | 4496 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4497 | values[n++] = primary_event_id(sub); |
3dab77fb | 4498 | |
76369139 | 4499 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4500 | } |
4501 | } | |
4502 | ||
eed01528 SE |
4503 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4504 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4505 | ||
3dab77fb | 4506 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4507 | struct perf_event *event) |
3dab77fb | 4508 | { |
e3f3541c | 4509 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4510 | u64 read_format = event->attr.read_format; |
4511 | ||
4512 | /* | |
4513 | * compute total_time_enabled, total_time_running | |
4514 | * based on snapshot values taken when the event | |
4515 | * was last scheduled in. | |
4516 | * | |
4517 | * we cannot simply called update_context_time() | |
4518 | * because of locking issue as we are called in | |
4519 | * NMI context | |
4520 | */ | |
c4794295 | 4521 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4522 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4523 | |
cdd6c482 | 4524 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4525 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4526 | else |
eed01528 | 4527 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4528 | } |
4529 | ||
5622f295 MM |
4530 | void perf_output_sample(struct perf_output_handle *handle, |
4531 | struct perf_event_header *header, | |
4532 | struct perf_sample_data *data, | |
cdd6c482 | 4533 | struct perf_event *event) |
5622f295 MM |
4534 | { |
4535 | u64 sample_type = data->type; | |
4536 | ||
4537 | perf_output_put(handle, *header); | |
4538 | ||
ff3d527c AH |
4539 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
4540 | perf_output_put(handle, data->id); | |
4541 | ||
5622f295 MM |
4542 | if (sample_type & PERF_SAMPLE_IP) |
4543 | perf_output_put(handle, data->ip); | |
4544 | ||
4545 | if (sample_type & PERF_SAMPLE_TID) | |
4546 | perf_output_put(handle, data->tid_entry); | |
4547 | ||
4548 | if (sample_type & PERF_SAMPLE_TIME) | |
4549 | perf_output_put(handle, data->time); | |
4550 | ||
4551 | if (sample_type & PERF_SAMPLE_ADDR) | |
4552 | perf_output_put(handle, data->addr); | |
4553 | ||
4554 | if (sample_type & PERF_SAMPLE_ID) | |
4555 | perf_output_put(handle, data->id); | |
4556 | ||
4557 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4558 | perf_output_put(handle, data->stream_id); | |
4559 | ||
4560 | if (sample_type & PERF_SAMPLE_CPU) | |
4561 | perf_output_put(handle, data->cpu_entry); | |
4562 | ||
4563 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4564 | perf_output_put(handle, data->period); | |
4565 | ||
4566 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4567 | perf_output_read(handle, event); |
5622f295 MM |
4568 | |
4569 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4570 | if (data->callchain) { | |
4571 | int size = 1; | |
4572 | ||
4573 | if (data->callchain) | |
4574 | size += data->callchain->nr; | |
4575 | ||
4576 | size *= sizeof(u64); | |
4577 | ||
76369139 | 4578 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4579 | } else { |
4580 | u64 nr = 0; | |
4581 | perf_output_put(handle, nr); | |
4582 | } | |
4583 | } | |
4584 | ||
4585 | if (sample_type & PERF_SAMPLE_RAW) { | |
4586 | if (data->raw) { | |
4587 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4588 | __output_copy(handle, data->raw->data, |
4589 | data->raw->size); | |
5622f295 MM |
4590 | } else { |
4591 | struct { | |
4592 | u32 size; | |
4593 | u32 data; | |
4594 | } raw = { | |
4595 | .size = sizeof(u32), | |
4596 | .data = 0, | |
4597 | }; | |
4598 | perf_output_put(handle, raw); | |
4599 | } | |
4600 | } | |
a7ac67ea | 4601 | |
bce38cd5 SE |
4602 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
4603 | if (data->br_stack) { | |
4604 | size_t size; | |
4605 | ||
4606 | size = data->br_stack->nr | |
4607 | * sizeof(struct perf_branch_entry); | |
4608 | ||
4609 | perf_output_put(handle, data->br_stack->nr); | |
4610 | perf_output_copy(handle, data->br_stack->entries, size); | |
4611 | } else { | |
4612 | /* | |
4613 | * we always store at least the value of nr | |
4614 | */ | |
4615 | u64 nr = 0; | |
4616 | perf_output_put(handle, nr); | |
4617 | } | |
4618 | } | |
4018994f JO |
4619 | |
4620 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4621 | u64 abi = data->regs_user.abi; | |
4622 | ||
4623 | /* | |
4624 | * If there are no regs to dump, notice it through | |
4625 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4626 | */ | |
4627 | perf_output_put(handle, abi); | |
4628 | ||
4629 | if (abi) { | |
4630 | u64 mask = event->attr.sample_regs_user; | |
4631 | perf_output_sample_regs(handle, | |
4632 | data->regs_user.regs, | |
4633 | mask); | |
4634 | } | |
4635 | } | |
c5ebcedb | 4636 | |
a5cdd40c | 4637 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
4638 | perf_output_sample_ustack(handle, |
4639 | data->stack_user_size, | |
4640 | data->regs_user.regs); | |
a5cdd40c | 4641 | } |
c3feedf2 AK |
4642 | |
4643 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
4644 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
4645 | |
4646 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
4647 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c | 4648 | |
fdfbbd07 AK |
4649 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
4650 | perf_output_put(handle, data->txn); | |
4651 | ||
a5cdd40c PZ |
4652 | if (!event->attr.watermark) { |
4653 | int wakeup_events = event->attr.wakeup_events; | |
4654 | ||
4655 | if (wakeup_events) { | |
4656 | struct ring_buffer *rb = handle->rb; | |
4657 | int events = local_inc_return(&rb->events); | |
4658 | ||
4659 | if (events >= wakeup_events) { | |
4660 | local_sub(wakeup_events, &rb->events); | |
4661 | local_inc(&rb->wakeup); | |
4662 | } | |
4663 | } | |
4664 | } | |
5622f295 MM |
4665 | } |
4666 | ||
4667 | void perf_prepare_sample(struct perf_event_header *header, | |
4668 | struct perf_sample_data *data, | |
cdd6c482 | 4669 | struct perf_event *event, |
5622f295 | 4670 | struct pt_regs *regs) |
7b732a75 | 4671 | { |
cdd6c482 | 4672 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4673 | |
cdd6c482 | 4674 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4675 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4676 | |
4677 | header->misc = 0; | |
4678 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4679 | |
c980d109 | 4680 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4681 | |
c320c7b7 | 4682 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4683 | data->ip = perf_instruction_pointer(regs); |
4684 | ||
b23f3325 | 4685 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4686 | int size = 1; |
394ee076 | 4687 | |
e6dab5ff | 4688 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4689 | |
4690 | if (data->callchain) | |
4691 | size += data->callchain->nr; | |
4692 | ||
4693 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4694 | } |
4695 | ||
3a43ce68 | 4696 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4697 | int size = sizeof(u32); |
4698 | ||
4699 | if (data->raw) | |
4700 | size += data->raw->size; | |
4701 | else | |
4702 | size += sizeof(u32); | |
4703 | ||
4704 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4705 | header->size += size; |
7f453c24 | 4706 | } |
bce38cd5 SE |
4707 | |
4708 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4709 | int size = sizeof(u64); /* nr */ | |
4710 | if (data->br_stack) { | |
4711 | size += data->br_stack->nr | |
4712 | * sizeof(struct perf_branch_entry); | |
4713 | } | |
4714 | header->size += size; | |
4715 | } | |
4018994f JO |
4716 | |
4717 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4718 | /* regs dump ABI info */ | |
4719 | int size = sizeof(u64); | |
4720 | ||
4721 | perf_sample_regs_user(&data->regs_user, regs); | |
4722 | ||
4723 | if (data->regs_user.regs) { | |
4724 | u64 mask = event->attr.sample_regs_user; | |
4725 | size += hweight64(mask) * sizeof(u64); | |
4726 | } | |
4727 | ||
4728 | header->size += size; | |
4729 | } | |
c5ebcedb JO |
4730 | |
4731 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4732 | /* | |
4733 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4734 | * processed as the last one or have additional check added | |
4735 | * in case new sample type is added, because we could eat | |
4736 | * up the rest of the sample size. | |
4737 | */ | |
4738 | struct perf_regs_user *uregs = &data->regs_user; | |
4739 | u16 stack_size = event->attr.sample_stack_user; | |
4740 | u16 size = sizeof(u64); | |
4741 | ||
4742 | if (!uregs->abi) | |
4743 | perf_sample_regs_user(uregs, regs); | |
4744 | ||
4745 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4746 | uregs->regs); | |
4747 | ||
4748 | /* | |
4749 | * If there is something to dump, add space for the dump | |
4750 | * itself and for the field that tells the dynamic size, | |
4751 | * which is how many have been actually dumped. | |
4752 | */ | |
4753 | if (stack_size) | |
4754 | size += sizeof(u64) + stack_size; | |
4755 | ||
4756 | data->stack_user_size = stack_size; | |
4757 | header->size += size; | |
4758 | } | |
5622f295 | 4759 | } |
7f453c24 | 4760 | |
a8b0ca17 | 4761 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4762 | struct perf_sample_data *data, |
4763 | struct pt_regs *regs) | |
4764 | { | |
4765 | struct perf_output_handle handle; | |
4766 | struct perf_event_header header; | |
689802b2 | 4767 | |
927c7a9e FW |
4768 | /* protect the callchain buffers */ |
4769 | rcu_read_lock(); | |
4770 | ||
cdd6c482 | 4771 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4772 | |
a7ac67ea | 4773 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4774 | goto exit; |
0322cd6e | 4775 | |
cdd6c482 | 4776 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4777 | |
8a057d84 | 4778 | perf_output_end(&handle); |
927c7a9e FW |
4779 | |
4780 | exit: | |
4781 | rcu_read_unlock(); | |
0322cd6e PZ |
4782 | } |
4783 | ||
38b200d6 | 4784 | /* |
cdd6c482 | 4785 | * read event_id |
38b200d6 PZ |
4786 | */ |
4787 | ||
4788 | struct perf_read_event { | |
4789 | struct perf_event_header header; | |
4790 | ||
4791 | u32 pid; | |
4792 | u32 tid; | |
38b200d6 PZ |
4793 | }; |
4794 | ||
4795 | static void | |
cdd6c482 | 4796 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4797 | struct task_struct *task) |
4798 | { | |
4799 | struct perf_output_handle handle; | |
c980d109 | 4800 | struct perf_sample_data sample; |
dfc65094 | 4801 | struct perf_read_event read_event = { |
38b200d6 | 4802 | .header = { |
cdd6c482 | 4803 | .type = PERF_RECORD_READ, |
38b200d6 | 4804 | .misc = 0, |
c320c7b7 | 4805 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4806 | }, |
cdd6c482 IM |
4807 | .pid = perf_event_pid(event, task), |
4808 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4809 | }; |
3dab77fb | 4810 | int ret; |
38b200d6 | 4811 | |
c980d109 | 4812 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4813 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4814 | if (ret) |
4815 | return; | |
4816 | ||
dfc65094 | 4817 | perf_output_put(&handle, read_event); |
cdd6c482 | 4818 | perf_output_read(&handle, event); |
c980d109 | 4819 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4820 | |
38b200d6 PZ |
4821 | perf_output_end(&handle); |
4822 | } | |
4823 | ||
52d857a8 JO |
4824 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
4825 | ||
4826 | static void | |
4827 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
4828 | perf_event_aux_output_cb output, |
4829 | void *data) | |
4830 | { | |
4831 | struct perf_event *event; | |
4832 | ||
4833 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
4834 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
4835 | continue; | |
4836 | if (!event_filter_match(event)) | |
4837 | continue; | |
67516844 | 4838 | output(event, data); |
52d857a8 JO |
4839 | } |
4840 | } | |
4841 | ||
4842 | static void | |
67516844 | 4843 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
4844 | struct perf_event_context *task_ctx) |
4845 | { | |
4846 | struct perf_cpu_context *cpuctx; | |
4847 | struct perf_event_context *ctx; | |
4848 | struct pmu *pmu; | |
4849 | int ctxn; | |
4850 | ||
4851 | rcu_read_lock(); | |
4852 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
4853 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
4854 | if (cpuctx->unique_pmu != pmu) | |
4855 | goto next; | |
67516844 | 4856 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
4857 | if (task_ctx) |
4858 | goto next; | |
4859 | ctxn = pmu->task_ctx_nr; | |
4860 | if (ctxn < 0) | |
4861 | goto next; | |
4862 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4863 | if (ctx) | |
67516844 | 4864 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
4865 | next: |
4866 | put_cpu_ptr(pmu->pmu_cpu_context); | |
4867 | } | |
4868 | ||
4869 | if (task_ctx) { | |
4870 | preempt_disable(); | |
67516844 | 4871 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
4872 | preempt_enable(); |
4873 | } | |
4874 | rcu_read_unlock(); | |
4875 | } | |
4876 | ||
60313ebe | 4877 | /* |
9f498cc5 PZ |
4878 | * task tracking -- fork/exit |
4879 | * | |
13d7a241 | 4880 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
4881 | */ |
4882 | ||
9f498cc5 | 4883 | struct perf_task_event { |
3a80b4a3 | 4884 | struct task_struct *task; |
cdd6c482 | 4885 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4886 | |
4887 | struct { | |
4888 | struct perf_event_header header; | |
4889 | ||
4890 | u32 pid; | |
4891 | u32 ppid; | |
9f498cc5 PZ |
4892 | u32 tid; |
4893 | u32 ptid; | |
393b2ad8 | 4894 | u64 time; |
cdd6c482 | 4895 | } event_id; |
60313ebe PZ |
4896 | }; |
4897 | ||
67516844 JO |
4898 | static int perf_event_task_match(struct perf_event *event) |
4899 | { | |
13d7a241 SE |
4900 | return event->attr.comm || event->attr.mmap || |
4901 | event->attr.mmap2 || event->attr.mmap_data || | |
4902 | event->attr.task; | |
67516844 JO |
4903 | } |
4904 | ||
cdd6c482 | 4905 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 4906 | void *data) |
60313ebe | 4907 | { |
52d857a8 | 4908 | struct perf_task_event *task_event = data; |
60313ebe | 4909 | struct perf_output_handle handle; |
c980d109 | 4910 | struct perf_sample_data sample; |
9f498cc5 | 4911 | struct task_struct *task = task_event->task; |
c980d109 | 4912 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4913 | |
67516844 JO |
4914 | if (!perf_event_task_match(event)) |
4915 | return; | |
4916 | ||
c980d109 | 4917 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4918 | |
c980d109 | 4919 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4920 | task_event->event_id.header.size); |
ef60777c | 4921 | if (ret) |
c980d109 | 4922 | goto out; |
60313ebe | 4923 | |
cdd6c482 IM |
4924 | task_event->event_id.pid = perf_event_pid(event, task); |
4925 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4926 | |
cdd6c482 IM |
4927 | task_event->event_id.tid = perf_event_tid(event, task); |
4928 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4929 | |
cdd6c482 | 4930 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4931 | |
c980d109 ACM |
4932 | perf_event__output_id_sample(event, &handle, &sample); |
4933 | ||
60313ebe | 4934 | perf_output_end(&handle); |
c980d109 ACM |
4935 | out: |
4936 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4937 | } |
4938 | ||
cdd6c482 IM |
4939 | static void perf_event_task(struct task_struct *task, |
4940 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4941 | int new) |
60313ebe | 4942 | { |
9f498cc5 | 4943 | struct perf_task_event task_event; |
60313ebe | 4944 | |
cdd6c482 IM |
4945 | if (!atomic_read(&nr_comm_events) && |
4946 | !atomic_read(&nr_mmap_events) && | |
4947 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4948 | return; |
4949 | ||
9f498cc5 | 4950 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4951 | .task = task, |
4952 | .task_ctx = task_ctx, | |
cdd6c482 | 4953 | .event_id = { |
60313ebe | 4954 | .header = { |
cdd6c482 | 4955 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4956 | .misc = 0, |
cdd6c482 | 4957 | .size = sizeof(task_event.event_id), |
60313ebe | 4958 | }, |
573402db PZ |
4959 | /* .pid */ |
4960 | /* .ppid */ | |
9f498cc5 PZ |
4961 | /* .tid */ |
4962 | /* .ptid */ | |
6f93d0a7 | 4963 | .time = perf_clock(), |
60313ebe PZ |
4964 | }, |
4965 | }; | |
4966 | ||
67516844 | 4967 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
4968 | &task_event, |
4969 | task_ctx); | |
9f498cc5 PZ |
4970 | } |
4971 | ||
cdd6c482 | 4972 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4973 | { |
cdd6c482 | 4974 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4975 | } |
4976 | ||
8d1b2d93 PZ |
4977 | /* |
4978 | * comm tracking | |
4979 | */ | |
4980 | ||
4981 | struct perf_comm_event { | |
22a4f650 IM |
4982 | struct task_struct *task; |
4983 | char *comm; | |
8d1b2d93 PZ |
4984 | int comm_size; |
4985 | ||
4986 | struct { | |
4987 | struct perf_event_header header; | |
4988 | ||
4989 | u32 pid; | |
4990 | u32 tid; | |
cdd6c482 | 4991 | } event_id; |
8d1b2d93 PZ |
4992 | }; |
4993 | ||
67516844 JO |
4994 | static int perf_event_comm_match(struct perf_event *event) |
4995 | { | |
4996 | return event->attr.comm; | |
4997 | } | |
4998 | ||
cdd6c482 | 4999 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 5000 | void *data) |
8d1b2d93 | 5001 | { |
52d857a8 | 5002 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 5003 | struct perf_output_handle handle; |
c980d109 | 5004 | struct perf_sample_data sample; |
cdd6c482 | 5005 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
5006 | int ret; |
5007 | ||
67516844 JO |
5008 | if (!perf_event_comm_match(event)) |
5009 | return; | |
5010 | ||
c980d109 ACM |
5011 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
5012 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5013 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
5014 | |
5015 | if (ret) | |
c980d109 | 5016 | goto out; |
8d1b2d93 | 5017 | |
cdd6c482 IM |
5018 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
5019 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 5020 | |
cdd6c482 | 5021 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 5022 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 5023 | comm_event->comm_size); |
c980d109 ACM |
5024 | |
5025 | perf_event__output_id_sample(event, &handle, &sample); | |
5026 | ||
8d1b2d93 | 5027 | perf_output_end(&handle); |
c980d109 ACM |
5028 | out: |
5029 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
5030 | } |
5031 | ||
cdd6c482 | 5032 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 5033 | { |
413ee3b4 | 5034 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 5035 | unsigned int size; |
8d1b2d93 | 5036 | |
413ee3b4 | 5037 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 5038 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 5039 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
5040 | |
5041 | comm_event->comm = comm; | |
5042 | comm_event->comm_size = size; | |
5043 | ||
cdd6c482 | 5044 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 5045 | |
67516844 | 5046 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
5047 | comm_event, |
5048 | NULL); | |
8d1b2d93 PZ |
5049 | } |
5050 | ||
cdd6c482 | 5051 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 5052 | { |
9ee318a7 | 5053 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
5054 | struct perf_event_context *ctx; |
5055 | int ctxn; | |
9ee318a7 | 5056 | |
c79aa0d9 | 5057 | rcu_read_lock(); |
8dc85d54 PZ |
5058 | for_each_task_context_nr(ctxn) { |
5059 | ctx = task->perf_event_ctxp[ctxn]; | |
5060 | if (!ctx) | |
5061 | continue; | |
9ee318a7 | 5062 | |
8dc85d54 PZ |
5063 | perf_event_enable_on_exec(ctx); |
5064 | } | |
c79aa0d9 | 5065 | rcu_read_unlock(); |
9ee318a7 | 5066 | |
cdd6c482 | 5067 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 5068 | return; |
a63eaf34 | 5069 | |
9ee318a7 | 5070 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 5071 | .task = task, |
573402db PZ |
5072 | /* .comm */ |
5073 | /* .comm_size */ | |
cdd6c482 | 5074 | .event_id = { |
573402db | 5075 | .header = { |
cdd6c482 | 5076 | .type = PERF_RECORD_COMM, |
573402db PZ |
5077 | .misc = 0, |
5078 | /* .size */ | |
5079 | }, | |
5080 | /* .pid */ | |
5081 | /* .tid */ | |
8d1b2d93 PZ |
5082 | }, |
5083 | }; | |
5084 | ||
cdd6c482 | 5085 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5086 | } |
5087 | ||
0a4a9391 PZ |
5088 | /* |
5089 | * mmap tracking | |
5090 | */ | |
5091 | ||
5092 | struct perf_mmap_event { | |
089dd79d PZ |
5093 | struct vm_area_struct *vma; |
5094 | ||
5095 | const char *file_name; | |
5096 | int file_size; | |
13d7a241 SE |
5097 | int maj, min; |
5098 | u64 ino; | |
5099 | u64 ino_generation; | |
0a4a9391 PZ |
5100 | |
5101 | struct { | |
5102 | struct perf_event_header header; | |
5103 | ||
5104 | u32 pid; | |
5105 | u32 tid; | |
5106 | u64 start; | |
5107 | u64 len; | |
5108 | u64 pgoff; | |
cdd6c482 | 5109 | } event_id; |
0a4a9391 PZ |
5110 | }; |
5111 | ||
67516844 JO |
5112 | static int perf_event_mmap_match(struct perf_event *event, |
5113 | void *data) | |
5114 | { | |
5115 | struct perf_mmap_event *mmap_event = data; | |
5116 | struct vm_area_struct *vma = mmap_event->vma; | |
5117 | int executable = vma->vm_flags & VM_EXEC; | |
5118 | ||
5119 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5120 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5121 | } |
5122 | ||
cdd6c482 | 5123 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5124 | void *data) |
0a4a9391 | 5125 | { |
52d857a8 | 5126 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5127 | struct perf_output_handle handle; |
c980d109 | 5128 | struct perf_sample_data sample; |
cdd6c482 | 5129 | int size = mmap_event->event_id.header.size; |
c980d109 | 5130 | int ret; |
0a4a9391 | 5131 | |
67516844 JO |
5132 | if (!perf_event_mmap_match(event, data)) |
5133 | return; | |
5134 | ||
13d7a241 SE |
5135 | if (event->attr.mmap2) { |
5136 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5137 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5138 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5139 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5140 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
13d7a241 SE |
5141 | } |
5142 | ||
c980d109 ACM |
5143 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5144 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5145 | mmap_event->event_id.header.size); |
0a4a9391 | 5146 | if (ret) |
c980d109 | 5147 | goto out; |
0a4a9391 | 5148 | |
cdd6c482 IM |
5149 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5150 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5151 | |
cdd6c482 | 5152 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5153 | |
5154 | if (event->attr.mmap2) { | |
5155 | perf_output_put(&handle, mmap_event->maj); | |
5156 | perf_output_put(&handle, mmap_event->min); | |
5157 | perf_output_put(&handle, mmap_event->ino); | |
5158 | perf_output_put(&handle, mmap_event->ino_generation); | |
5159 | } | |
5160 | ||
76369139 | 5161 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5162 | mmap_event->file_size); |
c980d109 ACM |
5163 | |
5164 | perf_event__output_id_sample(event, &handle, &sample); | |
5165 | ||
78d613eb | 5166 | perf_output_end(&handle); |
c980d109 ACM |
5167 | out: |
5168 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5169 | } |
5170 | ||
cdd6c482 | 5171 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5172 | { |
089dd79d PZ |
5173 | struct vm_area_struct *vma = mmap_event->vma; |
5174 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5175 | int maj = 0, min = 0; |
5176 | u64 ino = 0, gen = 0; | |
0a4a9391 PZ |
5177 | unsigned int size; |
5178 | char tmp[16]; | |
5179 | char *buf = NULL; | |
2c42cfbf | 5180 | char *name; |
413ee3b4 | 5181 | |
0a4a9391 | 5182 | if (file) { |
13d7a241 SE |
5183 | struct inode *inode; |
5184 | dev_t dev; | |
3ea2f2b9 | 5185 | |
2c42cfbf | 5186 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
0a4a9391 | 5187 | if (!buf) { |
c7e548b4 ON |
5188 | name = "//enomem"; |
5189 | goto cpy_name; | |
0a4a9391 | 5190 | } |
413ee3b4 | 5191 | /* |
3ea2f2b9 | 5192 | * d_path() works from the end of the rb backwards, so we |
413ee3b4 AB |
5193 | * need to add enough zero bytes after the string to handle |
5194 | * the 64bit alignment we do later. | |
5195 | */ | |
3ea2f2b9 | 5196 | name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64)); |
0a4a9391 | 5197 | if (IS_ERR(name)) { |
c7e548b4 ON |
5198 | name = "//toolong"; |
5199 | goto cpy_name; | |
0a4a9391 | 5200 | } |
13d7a241 SE |
5201 | inode = file_inode(vma->vm_file); |
5202 | dev = inode->i_sb->s_dev; | |
5203 | ino = inode->i_ino; | |
5204 | gen = inode->i_generation; | |
5205 | maj = MAJOR(dev); | |
5206 | min = MINOR(dev); | |
c7e548b4 | 5207 | goto got_name; |
0a4a9391 | 5208 | } else { |
2c42cfbf | 5209 | name = (char *)arch_vma_name(vma); |
c7e548b4 ON |
5210 | if (name) |
5211 | goto cpy_name; | |
089dd79d | 5212 | |
32c5fb7e | 5213 | if (vma->vm_start <= vma->vm_mm->start_brk && |
3af9e859 | 5214 | vma->vm_end >= vma->vm_mm->brk) { |
c7e548b4 ON |
5215 | name = "[heap]"; |
5216 | goto cpy_name; | |
32c5fb7e ON |
5217 | } |
5218 | if (vma->vm_start <= vma->vm_mm->start_stack && | |
3af9e859 | 5219 | vma->vm_end >= vma->vm_mm->start_stack) { |
c7e548b4 ON |
5220 | name = "[stack]"; |
5221 | goto cpy_name; | |
089dd79d PZ |
5222 | } |
5223 | ||
c7e548b4 ON |
5224 | name = "//anon"; |
5225 | goto cpy_name; | |
0a4a9391 PZ |
5226 | } |
5227 | ||
c7e548b4 ON |
5228 | cpy_name: |
5229 | strlcpy(tmp, name, sizeof(tmp)); | |
5230 | name = tmp; | |
0a4a9391 | 5231 | got_name: |
2c42cfbf PZ |
5232 | /* |
5233 | * Since our buffer works in 8 byte units we need to align our string | |
5234 | * size to a multiple of 8. However, we must guarantee the tail end is | |
5235 | * zero'd out to avoid leaking random bits to userspace. | |
5236 | */ | |
5237 | size = strlen(name)+1; | |
5238 | while (!IS_ALIGNED(size, sizeof(u64))) | |
5239 | name[size++] = '\0'; | |
0a4a9391 PZ |
5240 | |
5241 | mmap_event->file_name = name; | |
5242 | mmap_event->file_size = size; | |
13d7a241 SE |
5243 | mmap_event->maj = maj; |
5244 | mmap_event->min = min; | |
5245 | mmap_event->ino = ino; | |
5246 | mmap_event->ino_generation = gen; | |
0a4a9391 | 5247 | |
2fe85427 SE |
5248 | if (!(vma->vm_flags & VM_EXEC)) |
5249 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5250 | ||
cdd6c482 | 5251 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5252 | |
67516844 | 5253 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5254 | mmap_event, |
5255 | NULL); | |
665c2142 | 5256 | |
0a4a9391 PZ |
5257 | kfree(buf); |
5258 | } | |
5259 | ||
3af9e859 | 5260 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5261 | { |
9ee318a7 PZ |
5262 | struct perf_mmap_event mmap_event; |
5263 | ||
cdd6c482 | 5264 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5265 | return; |
5266 | ||
5267 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5268 | .vma = vma, |
573402db PZ |
5269 | /* .file_name */ |
5270 | /* .file_size */ | |
cdd6c482 | 5271 | .event_id = { |
573402db | 5272 | .header = { |
cdd6c482 | 5273 | .type = PERF_RECORD_MMAP, |
39447b38 | 5274 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5275 | /* .size */ |
5276 | }, | |
5277 | /* .pid */ | |
5278 | /* .tid */ | |
089dd79d PZ |
5279 | .start = vma->vm_start, |
5280 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5281 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5282 | }, |
13d7a241 SE |
5283 | /* .maj (attr_mmap2 only) */ |
5284 | /* .min (attr_mmap2 only) */ | |
5285 | /* .ino (attr_mmap2 only) */ | |
5286 | /* .ino_generation (attr_mmap2 only) */ | |
0a4a9391 PZ |
5287 | }; |
5288 | ||
cdd6c482 | 5289 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5290 | } |
5291 | ||
a78ac325 PZ |
5292 | /* |
5293 | * IRQ throttle logging | |
5294 | */ | |
5295 | ||
cdd6c482 | 5296 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5297 | { |
5298 | struct perf_output_handle handle; | |
c980d109 | 5299 | struct perf_sample_data sample; |
a78ac325 PZ |
5300 | int ret; |
5301 | ||
5302 | struct { | |
5303 | struct perf_event_header header; | |
5304 | u64 time; | |
cca3f454 | 5305 | u64 id; |
7f453c24 | 5306 | u64 stream_id; |
a78ac325 PZ |
5307 | } throttle_event = { |
5308 | .header = { | |
cdd6c482 | 5309 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5310 | .misc = 0, |
5311 | .size = sizeof(throttle_event), | |
5312 | }, | |
def0a9b2 | 5313 | .time = perf_clock(), |
cdd6c482 IM |
5314 | .id = primary_event_id(event), |
5315 | .stream_id = event->id, | |
a78ac325 PZ |
5316 | }; |
5317 | ||
966ee4d6 | 5318 | if (enable) |
cdd6c482 | 5319 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5320 | |
c980d109 ACM |
5321 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5322 | ||
5323 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5324 | throttle_event.header.size); |
a78ac325 PZ |
5325 | if (ret) |
5326 | return; | |
5327 | ||
5328 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5329 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5330 | perf_output_end(&handle); |
5331 | } | |
5332 | ||
f6c7d5fe | 5333 | /* |
cdd6c482 | 5334 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5335 | */ |
5336 | ||
a8b0ca17 | 5337 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5338 | int throttle, struct perf_sample_data *data, |
5339 | struct pt_regs *regs) | |
f6c7d5fe | 5340 | { |
cdd6c482 IM |
5341 | int events = atomic_read(&event->event_limit); |
5342 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5343 | u64 seq; |
79f14641 PZ |
5344 | int ret = 0; |
5345 | ||
96398826 PZ |
5346 | /* |
5347 | * Non-sampling counters might still use the PMI to fold short | |
5348 | * hardware counters, ignore those. | |
5349 | */ | |
5350 | if (unlikely(!is_sampling_event(event))) | |
5351 | return 0; | |
5352 | ||
e050e3f0 SE |
5353 | seq = __this_cpu_read(perf_throttled_seq); |
5354 | if (seq != hwc->interrupts_seq) { | |
5355 | hwc->interrupts_seq = seq; | |
5356 | hwc->interrupts = 1; | |
5357 | } else { | |
5358 | hwc->interrupts++; | |
5359 | if (unlikely(throttle | |
5360 | && hwc->interrupts >= max_samples_per_tick)) { | |
5361 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5362 | hwc->interrupts = MAX_INTERRUPTS; |
5363 | perf_log_throttle(event, 0); | |
d84153d6 | 5364 | tick_nohz_full_kick(); |
a78ac325 PZ |
5365 | ret = 1; |
5366 | } | |
e050e3f0 | 5367 | } |
60db5e09 | 5368 | |
cdd6c482 | 5369 | if (event->attr.freq) { |
def0a9b2 | 5370 | u64 now = perf_clock(); |
abd50713 | 5371 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5372 | |
abd50713 | 5373 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5374 | |
abd50713 | 5375 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5376 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5377 | } |
5378 | ||
2023b359 PZ |
5379 | /* |
5380 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5381 | * events |
2023b359 PZ |
5382 | */ |
5383 | ||
cdd6c482 IM |
5384 | event->pending_kill = POLL_IN; |
5385 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5386 | ret = 1; |
cdd6c482 | 5387 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5388 | event->pending_disable = 1; |
5389 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5390 | } |
5391 | ||
453f19ee | 5392 | if (event->overflow_handler) |
a8b0ca17 | 5393 | event->overflow_handler(event, data, regs); |
453f19ee | 5394 | else |
a8b0ca17 | 5395 | perf_event_output(event, data, regs); |
453f19ee | 5396 | |
f506b3dc | 5397 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5398 | event->pending_wakeup = 1; |
5399 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5400 | } |
5401 | ||
79f14641 | 5402 | return ret; |
f6c7d5fe PZ |
5403 | } |
5404 | ||
a8b0ca17 | 5405 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5406 | struct perf_sample_data *data, |
5407 | struct pt_regs *regs) | |
850bc73f | 5408 | { |
a8b0ca17 | 5409 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5410 | } |
5411 | ||
15dbf27c | 5412 | /* |
cdd6c482 | 5413 | * Generic software event infrastructure |
15dbf27c PZ |
5414 | */ |
5415 | ||
b28ab83c PZ |
5416 | struct swevent_htable { |
5417 | struct swevent_hlist *swevent_hlist; | |
5418 | struct mutex hlist_mutex; | |
5419 | int hlist_refcount; | |
5420 | ||
5421 | /* Recursion avoidance in each contexts */ | |
5422 | int recursion[PERF_NR_CONTEXTS]; | |
5423 | }; | |
5424 | ||
5425 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5426 | ||
7b4b6658 | 5427 | /* |
cdd6c482 IM |
5428 | * We directly increment event->count and keep a second value in |
5429 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5430 | * is kept in the range [-sample_period, 0] so that we can use the |
5431 | * sign as trigger. | |
5432 | */ | |
5433 | ||
ab573844 | 5434 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5435 | { |
cdd6c482 | 5436 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5437 | u64 period = hwc->last_period; |
5438 | u64 nr, offset; | |
5439 | s64 old, val; | |
5440 | ||
5441 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5442 | |
5443 | again: | |
e7850595 | 5444 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5445 | if (val < 0) |
5446 | return 0; | |
15dbf27c | 5447 | |
7b4b6658 PZ |
5448 | nr = div64_u64(period + val, period); |
5449 | offset = nr * period; | |
5450 | val -= offset; | |
e7850595 | 5451 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5452 | goto again; |
15dbf27c | 5453 | |
7b4b6658 | 5454 | return nr; |
15dbf27c PZ |
5455 | } |
5456 | ||
0cff784a | 5457 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5458 | struct perf_sample_data *data, |
5622f295 | 5459 | struct pt_regs *regs) |
15dbf27c | 5460 | { |
cdd6c482 | 5461 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5462 | int throttle = 0; |
15dbf27c | 5463 | |
0cff784a PZ |
5464 | if (!overflow) |
5465 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5466 | |
7b4b6658 PZ |
5467 | if (hwc->interrupts == MAX_INTERRUPTS) |
5468 | return; | |
15dbf27c | 5469 | |
7b4b6658 | 5470 | for (; overflow; overflow--) { |
a8b0ca17 | 5471 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5472 | data, regs)) { |
7b4b6658 PZ |
5473 | /* |
5474 | * We inhibit the overflow from happening when | |
5475 | * hwc->interrupts == MAX_INTERRUPTS. | |
5476 | */ | |
5477 | break; | |
5478 | } | |
cf450a73 | 5479 | throttle = 1; |
7b4b6658 | 5480 | } |
15dbf27c PZ |
5481 | } |
5482 | ||
a4eaf7f1 | 5483 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5484 | struct perf_sample_data *data, |
5622f295 | 5485 | struct pt_regs *regs) |
7b4b6658 | 5486 | { |
cdd6c482 | 5487 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5488 | |
e7850595 | 5489 | local64_add(nr, &event->count); |
d6d020e9 | 5490 | |
0cff784a PZ |
5491 | if (!regs) |
5492 | return; | |
5493 | ||
6c7e550f | 5494 | if (!is_sampling_event(event)) |
7b4b6658 | 5495 | return; |
d6d020e9 | 5496 | |
5d81e5cf AV |
5497 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5498 | data->period = nr; | |
5499 | return perf_swevent_overflow(event, 1, data, regs); | |
5500 | } else | |
5501 | data->period = event->hw.last_period; | |
5502 | ||
0cff784a | 5503 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5504 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5505 | |
e7850595 | 5506 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5507 | return; |
df1a132b | 5508 | |
a8b0ca17 | 5509 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5510 | } |
5511 | ||
f5ffe02e FW |
5512 | static int perf_exclude_event(struct perf_event *event, |
5513 | struct pt_regs *regs) | |
5514 | { | |
a4eaf7f1 | 5515 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5516 | return 1; |
a4eaf7f1 | 5517 | |
f5ffe02e FW |
5518 | if (regs) { |
5519 | if (event->attr.exclude_user && user_mode(regs)) | |
5520 | return 1; | |
5521 | ||
5522 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5523 | return 1; | |
5524 | } | |
5525 | ||
5526 | return 0; | |
5527 | } | |
5528 | ||
cdd6c482 | 5529 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5530 | enum perf_type_id type, |
6fb2915d LZ |
5531 | u32 event_id, |
5532 | struct perf_sample_data *data, | |
5533 | struct pt_regs *regs) | |
15dbf27c | 5534 | { |
cdd6c482 | 5535 | if (event->attr.type != type) |
a21ca2ca | 5536 | return 0; |
f5ffe02e | 5537 | |
cdd6c482 | 5538 | if (event->attr.config != event_id) |
15dbf27c PZ |
5539 | return 0; |
5540 | ||
f5ffe02e FW |
5541 | if (perf_exclude_event(event, regs)) |
5542 | return 0; | |
15dbf27c PZ |
5543 | |
5544 | return 1; | |
5545 | } | |
5546 | ||
76e1d904 FW |
5547 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5548 | { | |
5549 | u64 val = event_id | (type << 32); | |
5550 | ||
5551 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5552 | } | |
5553 | ||
49f135ed FW |
5554 | static inline struct hlist_head * |
5555 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5556 | { |
49f135ed FW |
5557 | u64 hash = swevent_hash(type, event_id); |
5558 | ||
5559 | return &hlist->heads[hash]; | |
5560 | } | |
76e1d904 | 5561 | |
49f135ed FW |
5562 | /* For the read side: events when they trigger */ |
5563 | static inline struct hlist_head * | |
b28ab83c | 5564 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5565 | { |
5566 | struct swevent_hlist *hlist; | |
76e1d904 | 5567 | |
b28ab83c | 5568 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5569 | if (!hlist) |
5570 | return NULL; | |
5571 | ||
49f135ed FW |
5572 | return __find_swevent_head(hlist, type, event_id); |
5573 | } | |
5574 | ||
5575 | /* For the event head insertion and removal in the hlist */ | |
5576 | static inline struct hlist_head * | |
b28ab83c | 5577 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5578 | { |
5579 | struct swevent_hlist *hlist; | |
5580 | u32 event_id = event->attr.config; | |
5581 | u64 type = event->attr.type; | |
5582 | ||
5583 | /* | |
5584 | * Event scheduling is always serialized against hlist allocation | |
5585 | * and release. Which makes the protected version suitable here. | |
5586 | * The context lock guarantees that. | |
5587 | */ | |
b28ab83c | 5588 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5589 | lockdep_is_held(&event->ctx->lock)); |
5590 | if (!hlist) | |
5591 | return NULL; | |
5592 | ||
5593 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5594 | } |
5595 | ||
5596 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5597 | u64 nr, |
76e1d904 FW |
5598 | struct perf_sample_data *data, |
5599 | struct pt_regs *regs) | |
15dbf27c | 5600 | { |
b28ab83c | 5601 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5602 | struct perf_event *event; |
76e1d904 | 5603 | struct hlist_head *head; |
15dbf27c | 5604 | |
76e1d904 | 5605 | rcu_read_lock(); |
b28ab83c | 5606 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5607 | if (!head) |
5608 | goto end; | |
5609 | ||
b67bfe0d | 5610 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5611 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5612 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5613 | } |
76e1d904 FW |
5614 | end: |
5615 | rcu_read_unlock(); | |
15dbf27c PZ |
5616 | } |
5617 | ||
4ed7c92d | 5618 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5619 | { |
b28ab83c | 5620 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5621 | |
b28ab83c | 5622 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5623 | } |
645e8cc0 | 5624 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5625 | |
fa9f90be | 5626 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5627 | { |
b28ab83c | 5628 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5629 | |
b28ab83c | 5630 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5631 | } |
15dbf27c | 5632 | |
a8b0ca17 | 5633 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5634 | { |
a4234bfc | 5635 | struct perf_sample_data data; |
4ed7c92d PZ |
5636 | int rctx; |
5637 | ||
1c024eca | 5638 | preempt_disable_notrace(); |
4ed7c92d PZ |
5639 | rctx = perf_swevent_get_recursion_context(); |
5640 | if (rctx < 0) | |
5641 | return; | |
a4234bfc | 5642 | |
fd0d000b | 5643 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5644 | |
a8b0ca17 | 5645 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5646 | |
5647 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5648 | preempt_enable_notrace(); |
b8e83514 PZ |
5649 | } |
5650 | ||
cdd6c482 | 5651 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5652 | { |
15dbf27c PZ |
5653 | } |
5654 | ||
a4eaf7f1 | 5655 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5656 | { |
b28ab83c | 5657 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5658 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5659 | struct hlist_head *head; |
5660 | ||
6c7e550f | 5661 | if (is_sampling_event(event)) { |
7b4b6658 | 5662 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5663 | perf_swevent_set_period(event); |
7b4b6658 | 5664 | } |
76e1d904 | 5665 | |
a4eaf7f1 PZ |
5666 | hwc->state = !(flags & PERF_EF_START); |
5667 | ||
b28ab83c | 5668 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
5669 | if (WARN_ON_ONCE(!head)) |
5670 | return -EINVAL; | |
5671 | ||
5672 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5673 | ||
15dbf27c PZ |
5674 | return 0; |
5675 | } | |
5676 | ||
a4eaf7f1 | 5677 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5678 | { |
76e1d904 | 5679 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5680 | } |
5681 | ||
a4eaf7f1 | 5682 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5683 | { |
a4eaf7f1 | 5684 | event->hw.state = 0; |
d6d020e9 | 5685 | } |
aa9c4c0f | 5686 | |
a4eaf7f1 | 5687 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5688 | { |
a4eaf7f1 | 5689 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5690 | } |
5691 | ||
49f135ed FW |
5692 | /* Deref the hlist from the update side */ |
5693 | static inline struct swevent_hlist * | |
b28ab83c | 5694 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5695 | { |
b28ab83c PZ |
5696 | return rcu_dereference_protected(swhash->swevent_hlist, |
5697 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5698 | } |
5699 | ||
b28ab83c | 5700 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5701 | { |
b28ab83c | 5702 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5703 | |
49f135ed | 5704 | if (!hlist) |
76e1d904 FW |
5705 | return; |
5706 | ||
b28ab83c | 5707 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5708 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5709 | } |
5710 | ||
5711 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5712 | { | |
b28ab83c | 5713 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5714 | |
b28ab83c | 5715 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5716 | |
b28ab83c PZ |
5717 | if (!--swhash->hlist_refcount) |
5718 | swevent_hlist_release(swhash); | |
76e1d904 | 5719 | |
b28ab83c | 5720 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5721 | } |
5722 | ||
5723 | static void swevent_hlist_put(struct perf_event *event) | |
5724 | { | |
5725 | int cpu; | |
5726 | ||
76e1d904 FW |
5727 | for_each_possible_cpu(cpu) |
5728 | swevent_hlist_put_cpu(event, cpu); | |
5729 | } | |
5730 | ||
5731 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5732 | { | |
b28ab83c | 5733 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5734 | int err = 0; |
5735 | ||
b28ab83c | 5736 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5737 | |
b28ab83c | 5738 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5739 | struct swevent_hlist *hlist; |
5740 | ||
5741 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5742 | if (!hlist) { | |
5743 | err = -ENOMEM; | |
5744 | goto exit; | |
5745 | } | |
b28ab83c | 5746 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5747 | } |
b28ab83c | 5748 | swhash->hlist_refcount++; |
9ed6060d | 5749 | exit: |
b28ab83c | 5750 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5751 | |
5752 | return err; | |
5753 | } | |
5754 | ||
5755 | static int swevent_hlist_get(struct perf_event *event) | |
5756 | { | |
5757 | int err; | |
5758 | int cpu, failed_cpu; | |
5759 | ||
76e1d904 FW |
5760 | get_online_cpus(); |
5761 | for_each_possible_cpu(cpu) { | |
5762 | err = swevent_hlist_get_cpu(event, cpu); | |
5763 | if (err) { | |
5764 | failed_cpu = cpu; | |
5765 | goto fail; | |
5766 | } | |
5767 | } | |
5768 | put_online_cpus(); | |
5769 | ||
5770 | return 0; | |
9ed6060d | 5771 | fail: |
76e1d904 FW |
5772 | for_each_possible_cpu(cpu) { |
5773 | if (cpu == failed_cpu) | |
5774 | break; | |
5775 | swevent_hlist_put_cpu(event, cpu); | |
5776 | } | |
5777 | ||
5778 | put_online_cpus(); | |
5779 | return err; | |
5780 | } | |
5781 | ||
c5905afb | 5782 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5783 | |
b0a873eb PZ |
5784 | static void sw_perf_event_destroy(struct perf_event *event) |
5785 | { | |
5786 | u64 event_id = event->attr.config; | |
95476b64 | 5787 | |
b0a873eb PZ |
5788 | WARN_ON(event->parent); |
5789 | ||
c5905afb | 5790 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5791 | swevent_hlist_put(event); |
5792 | } | |
5793 | ||
5794 | static int perf_swevent_init(struct perf_event *event) | |
5795 | { | |
8176cced | 5796 | u64 event_id = event->attr.config; |
b0a873eb PZ |
5797 | |
5798 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5799 | return -ENOENT; | |
5800 | ||
2481c5fa SE |
5801 | /* |
5802 | * no branch sampling for software events | |
5803 | */ | |
5804 | if (has_branch_stack(event)) | |
5805 | return -EOPNOTSUPP; | |
5806 | ||
b0a873eb PZ |
5807 | switch (event_id) { |
5808 | case PERF_COUNT_SW_CPU_CLOCK: | |
5809 | case PERF_COUNT_SW_TASK_CLOCK: | |
5810 | return -ENOENT; | |
5811 | ||
5812 | default: | |
5813 | break; | |
5814 | } | |
5815 | ||
ce677831 | 5816 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5817 | return -ENOENT; |
5818 | ||
5819 | if (!event->parent) { | |
5820 | int err; | |
5821 | ||
5822 | err = swevent_hlist_get(event); | |
5823 | if (err) | |
5824 | return err; | |
5825 | ||
c5905afb | 5826 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5827 | event->destroy = sw_perf_event_destroy; |
5828 | } | |
5829 | ||
5830 | return 0; | |
5831 | } | |
5832 | ||
35edc2a5 PZ |
5833 | static int perf_swevent_event_idx(struct perf_event *event) |
5834 | { | |
5835 | return 0; | |
5836 | } | |
5837 | ||
b0a873eb | 5838 | static struct pmu perf_swevent = { |
89a1e187 | 5839 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5840 | |
b0a873eb | 5841 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5842 | .add = perf_swevent_add, |
5843 | .del = perf_swevent_del, | |
5844 | .start = perf_swevent_start, | |
5845 | .stop = perf_swevent_stop, | |
1c024eca | 5846 | .read = perf_swevent_read, |
35edc2a5 PZ |
5847 | |
5848 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5849 | }; |
5850 | ||
b0a873eb PZ |
5851 | #ifdef CONFIG_EVENT_TRACING |
5852 | ||
1c024eca PZ |
5853 | static int perf_tp_filter_match(struct perf_event *event, |
5854 | struct perf_sample_data *data) | |
5855 | { | |
5856 | void *record = data->raw->data; | |
5857 | ||
5858 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5859 | return 1; | |
5860 | return 0; | |
5861 | } | |
5862 | ||
5863 | static int perf_tp_event_match(struct perf_event *event, | |
5864 | struct perf_sample_data *data, | |
5865 | struct pt_regs *regs) | |
5866 | { | |
a0f7d0f7 FW |
5867 | if (event->hw.state & PERF_HES_STOPPED) |
5868 | return 0; | |
580d607c PZ |
5869 | /* |
5870 | * All tracepoints are from kernel-space. | |
5871 | */ | |
5872 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5873 | return 0; |
5874 | ||
5875 | if (!perf_tp_filter_match(event, data)) | |
5876 | return 0; | |
5877 | ||
5878 | return 1; | |
5879 | } | |
5880 | ||
5881 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5882 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5883 | struct task_struct *task) | |
95476b64 FW |
5884 | { |
5885 | struct perf_sample_data data; | |
1c024eca | 5886 | struct perf_event *event; |
1c024eca | 5887 | |
95476b64 FW |
5888 | struct perf_raw_record raw = { |
5889 | .size = entry_size, | |
5890 | .data = record, | |
5891 | }; | |
5892 | ||
fd0d000b | 5893 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5894 | data.raw = &raw; |
5895 | ||
b67bfe0d | 5896 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 5897 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 5898 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5899 | } |
ecc55f84 | 5900 | |
e6dab5ff AV |
5901 | /* |
5902 | * If we got specified a target task, also iterate its context and | |
5903 | * deliver this event there too. | |
5904 | */ | |
5905 | if (task && task != current) { | |
5906 | struct perf_event_context *ctx; | |
5907 | struct trace_entry *entry = record; | |
5908 | ||
5909 | rcu_read_lock(); | |
5910 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5911 | if (!ctx) | |
5912 | goto unlock; | |
5913 | ||
5914 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5915 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5916 | continue; | |
5917 | if (event->attr.config != entry->type) | |
5918 | continue; | |
5919 | if (perf_tp_event_match(event, &data, regs)) | |
5920 | perf_swevent_event(event, count, &data, regs); | |
5921 | } | |
5922 | unlock: | |
5923 | rcu_read_unlock(); | |
5924 | } | |
5925 | ||
ecc55f84 | 5926 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5927 | } |
5928 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5929 | ||
cdd6c482 | 5930 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5931 | { |
1c024eca | 5932 | perf_trace_destroy(event); |
e077df4f PZ |
5933 | } |
5934 | ||
b0a873eb | 5935 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5936 | { |
76e1d904 FW |
5937 | int err; |
5938 | ||
b0a873eb PZ |
5939 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5940 | return -ENOENT; | |
5941 | ||
2481c5fa SE |
5942 | /* |
5943 | * no branch sampling for tracepoint events | |
5944 | */ | |
5945 | if (has_branch_stack(event)) | |
5946 | return -EOPNOTSUPP; | |
5947 | ||
1c024eca PZ |
5948 | err = perf_trace_init(event); |
5949 | if (err) | |
b0a873eb | 5950 | return err; |
e077df4f | 5951 | |
cdd6c482 | 5952 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5953 | |
b0a873eb PZ |
5954 | return 0; |
5955 | } | |
5956 | ||
5957 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5958 | .task_ctx_nr = perf_sw_context, |
5959 | ||
b0a873eb | 5960 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5961 | .add = perf_trace_add, |
5962 | .del = perf_trace_del, | |
5963 | .start = perf_swevent_start, | |
5964 | .stop = perf_swevent_stop, | |
b0a873eb | 5965 | .read = perf_swevent_read, |
35edc2a5 PZ |
5966 | |
5967 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5968 | }; |
5969 | ||
5970 | static inline void perf_tp_register(void) | |
5971 | { | |
2e80a82a | 5972 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5973 | } |
6fb2915d LZ |
5974 | |
5975 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5976 | { | |
5977 | char *filter_str; | |
5978 | int ret; | |
5979 | ||
5980 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5981 | return -EINVAL; | |
5982 | ||
5983 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5984 | if (IS_ERR(filter_str)) | |
5985 | return PTR_ERR(filter_str); | |
5986 | ||
5987 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5988 | ||
5989 | kfree(filter_str); | |
5990 | return ret; | |
5991 | } | |
5992 | ||
5993 | static void perf_event_free_filter(struct perf_event *event) | |
5994 | { | |
5995 | ftrace_profile_free_filter(event); | |
5996 | } | |
5997 | ||
e077df4f | 5998 | #else |
6fb2915d | 5999 | |
b0a873eb | 6000 | static inline void perf_tp_register(void) |
e077df4f | 6001 | { |
e077df4f | 6002 | } |
6fb2915d LZ |
6003 | |
6004 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6005 | { | |
6006 | return -ENOENT; | |
6007 | } | |
6008 | ||
6009 | static void perf_event_free_filter(struct perf_event *event) | |
6010 | { | |
6011 | } | |
6012 | ||
07b139c8 | 6013 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 6014 | |
24f1e32c | 6015 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 6016 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 6017 | { |
f5ffe02e FW |
6018 | struct perf_sample_data sample; |
6019 | struct pt_regs *regs = data; | |
6020 | ||
fd0d000b | 6021 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 6022 | |
a4eaf7f1 | 6023 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 6024 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
6025 | } |
6026 | #endif | |
6027 | ||
b0a873eb PZ |
6028 | /* |
6029 | * hrtimer based swevent callback | |
6030 | */ | |
f29ac756 | 6031 | |
b0a873eb | 6032 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 6033 | { |
b0a873eb PZ |
6034 | enum hrtimer_restart ret = HRTIMER_RESTART; |
6035 | struct perf_sample_data data; | |
6036 | struct pt_regs *regs; | |
6037 | struct perf_event *event; | |
6038 | u64 period; | |
f29ac756 | 6039 | |
b0a873eb | 6040 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
6041 | |
6042 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
6043 | return HRTIMER_NORESTART; | |
6044 | ||
b0a873eb | 6045 | event->pmu->read(event); |
f344011c | 6046 | |
fd0d000b | 6047 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
6048 | regs = get_irq_regs(); |
6049 | ||
6050 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 6051 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 6052 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
6053 | ret = HRTIMER_NORESTART; |
6054 | } | |
24f1e32c | 6055 | |
b0a873eb PZ |
6056 | period = max_t(u64, 10000, event->hw.sample_period); |
6057 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 6058 | |
b0a873eb | 6059 | return ret; |
f29ac756 PZ |
6060 | } |
6061 | ||
b0a873eb | 6062 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 6063 | { |
b0a873eb | 6064 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
6065 | s64 period; |
6066 | ||
6067 | if (!is_sampling_event(event)) | |
6068 | return; | |
f5ffe02e | 6069 | |
5d508e82 FBH |
6070 | period = local64_read(&hwc->period_left); |
6071 | if (period) { | |
6072 | if (period < 0) | |
6073 | period = 10000; | |
fa407f35 | 6074 | |
5d508e82 FBH |
6075 | local64_set(&hwc->period_left, 0); |
6076 | } else { | |
6077 | period = max_t(u64, 10000, hwc->sample_period); | |
6078 | } | |
6079 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 6080 | ns_to_ktime(period), 0, |
b5ab4cd5 | 6081 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 6082 | } |
b0a873eb PZ |
6083 | |
6084 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6085 | { |
b0a873eb PZ |
6086 | struct hw_perf_event *hwc = &event->hw; |
6087 | ||
6c7e550f | 6088 | if (is_sampling_event(event)) { |
b0a873eb | 6089 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6090 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6091 | |
6092 | hrtimer_cancel(&hwc->hrtimer); | |
6093 | } | |
24f1e32c FW |
6094 | } |
6095 | ||
ba3dd36c PZ |
6096 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6097 | { | |
6098 | struct hw_perf_event *hwc = &event->hw; | |
6099 | ||
6100 | if (!is_sampling_event(event)) | |
6101 | return; | |
6102 | ||
6103 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6104 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6105 | ||
6106 | /* | |
6107 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6108 | * mapping and avoid the whole period adjust feedback stuff. | |
6109 | */ | |
6110 | if (event->attr.freq) { | |
6111 | long freq = event->attr.sample_freq; | |
6112 | ||
6113 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
6114 | hwc->sample_period = event->attr.sample_period; | |
6115 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 6116 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
6117 | event->attr.freq = 0; |
6118 | } | |
6119 | } | |
6120 | ||
b0a873eb PZ |
6121 | /* |
6122 | * Software event: cpu wall time clock | |
6123 | */ | |
6124 | ||
6125 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 6126 | { |
b0a873eb PZ |
6127 | s64 prev; |
6128 | u64 now; | |
6129 | ||
a4eaf7f1 | 6130 | now = local_clock(); |
b0a873eb PZ |
6131 | prev = local64_xchg(&event->hw.prev_count, now); |
6132 | local64_add(now - prev, &event->count); | |
24f1e32c | 6133 | } |
24f1e32c | 6134 | |
a4eaf7f1 | 6135 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6136 | { |
a4eaf7f1 | 6137 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 6138 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6139 | } |
6140 | ||
a4eaf7f1 | 6141 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6142 | { |
b0a873eb PZ |
6143 | perf_swevent_cancel_hrtimer(event); |
6144 | cpu_clock_event_update(event); | |
6145 | } | |
f29ac756 | 6146 | |
a4eaf7f1 PZ |
6147 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6148 | { | |
6149 | if (flags & PERF_EF_START) | |
6150 | cpu_clock_event_start(event, flags); | |
6151 | ||
6152 | return 0; | |
6153 | } | |
6154 | ||
6155 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6156 | { | |
6157 | cpu_clock_event_stop(event, flags); | |
6158 | } | |
6159 | ||
b0a873eb PZ |
6160 | static void cpu_clock_event_read(struct perf_event *event) |
6161 | { | |
6162 | cpu_clock_event_update(event); | |
6163 | } | |
f344011c | 6164 | |
b0a873eb PZ |
6165 | static int cpu_clock_event_init(struct perf_event *event) |
6166 | { | |
6167 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6168 | return -ENOENT; | |
6169 | ||
6170 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
6171 | return -ENOENT; | |
6172 | ||
2481c5fa SE |
6173 | /* |
6174 | * no branch sampling for software events | |
6175 | */ | |
6176 | if (has_branch_stack(event)) | |
6177 | return -EOPNOTSUPP; | |
6178 | ||
ba3dd36c PZ |
6179 | perf_swevent_init_hrtimer(event); |
6180 | ||
b0a873eb | 6181 | return 0; |
f29ac756 PZ |
6182 | } |
6183 | ||
b0a873eb | 6184 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
6185 | .task_ctx_nr = perf_sw_context, |
6186 | ||
b0a873eb | 6187 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
6188 | .add = cpu_clock_event_add, |
6189 | .del = cpu_clock_event_del, | |
6190 | .start = cpu_clock_event_start, | |
6191 | .stop = cpu_clock_event_stop, | |
b0a873eb | 6192 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
6193 | |
6194 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
6195 | }; |
6196 | ||
6197 | /* | |
6198 | * Software event: task time clock | |
6199 | */ | |
6200 | ||
6201 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 6202 | { |
b0a873eb PZ |
6203 | u64 prev; |
6204 | s64 delta; | |
5c92d124 | 6205 | |
b0a873eb PZ |
6206 | prev = local64_xchg(&event->hw.prev_count, now); |
6207 | delta = now - prev; | |
6208 | local64_add(delta, &event->count); | |
6209 | } | |
5c92d124 | 6210 | |
a4eaf7f1 | 6211 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6212 | { |
a4eaf7f1 | 6213 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 6214 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6215 | } |
6216 | ||
a4eaf7f1 | 6217 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
6218 | { |
6219 | perf_swevent_cancel_hrtimer(event); | |
6220 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
6221 | } |
6222 | ||
6223 | static int task_clock_event_add(struct perf_event *event, int flags) | |
6224 | { | |
6225 | if (flags & PERF_EF_START) | |
6226 | task_clock_event_start(event, flags); | |
b0a873eb | 6227 | |
a4eaf7f1 PZ |
6228 | return 0; |
6229 | } | |
6230 | ||
6231 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6232 | { | |
6233 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6234 | } |
6235 | ||
6236 | static void task_clock_event_read(struct perf_event *event) | |
6237 | { | |
768a06e2 PZ |
6238 | u64 now = perf_clock(); |
6239 | u64 delta = now - event->ctx->timestamp; | |
6240 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6241 | |
6242 | task_clock_event_update(event, time); | |
6243 | } | |
6244 | ||
6245 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6246 | { |
b0a873eb PZ |
6247 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6248 | return -ENOENT; | |
6249 | ||
6250 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6251 | return -ENOENT; | |
6252 | ||
2481c5fa SE |
6253 | /* |
6254 | * no branch sampling for software events | |
6255 | */ | |
6256 | if (has_branch_stack(event)) | |
6257 | return -EOPNOTSUPP; | |
6258 | ||
ba3dd36c PZ |
6259 | perf_swevent_init_hrtimer(event); |
6260 | ||
b0a873eb | 6261 | return 0; |
6fb2915d LZ |
6262 | } |
6263 | ||
b0a873eb | 6264 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6265 | .task_ctx_nr = perf_sw_context, |
6266 | ||
b0a873eb | 6267 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6268 | .add = task_clock_event_add, |
6269 | .del = task_clock_event_del, | |
6270 | .start = task_clock_event_start, | |
6271 | .stop = task_clock_event_stop, | |
b0a873eb | 6272 | .read = task_clock_event_read, |
35edc2a5 PZ |
6273 | |
6274 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 6275 | }; |
6fb2915d | 6276 | |
ad5133b7 | 6277 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6278 | { |
e077df4f | 6279 | } |
6fb2915d | 6280 | |
ad5133b7 | 6281 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6282 | { |
ad5133b7 | 6283 | return 0; |
6fb2915d LZ |
6284 | } |
6285 | ||
ad5133b7 | 6286 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6287 | { |
ad5133b7 | 6288 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6289 | } |
6290 | ||
ad5133b7 PZ |
6291 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6292 | { | |
6293 | perf_pmu_enable(pmu); | |
6294 | return 0; | |
6295 | } | |
e077df4f | 6296 | |
ad5133b7 | 6297 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6298 | { |
ad5133b7 | 6299 | perf_pmu_enable(pmu); |
24f1e32c FW |
6300 | } |
6301 | ||
35edc2a5 PZ |
6302 | static int perf_event_idx_default(struct perf_event *event) |
6303 | { | |
6304 | return event->hw.idx + 1; | |
6305 | } | |
6306 | ||
8dc85d54 PZ |
6307 | /* |
6308 | * Ensures all contexts with the same task_ctx_nr have the same | |
6309 | * pmu_cpu_context too. | |
6310 | */ | |
6311 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 6312 | { |
8dc85d54 | 6313 | struct pmu *pmu; |
b326e956 | 6314 | |
8dc85d54 PZ |
6315 | if (ctxn < 0) |
6316 | return NULL; | |
24f1e32c | 6317 | |
8dc85d54 PZ |
6318 | list_for_each_entry(pmu, &pmus, entry) { |
6319 | if (pmu->task_ctx_nr == ctxn) | |
6320 | return pmu->pmu_cpu_context; | |
6321 | } | |
24f1e32c | 6322 | |
8dc85d54 | 6323 | return NULL; |
24f1e32c FW |
6324 | } |
6325 | ||
51676957 | 6326 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6327 | { |
51676957 PZ |
6328 | int cpu; |
6329 | ||
6330 | for_each_possible_cpu(cpu) { | |
6331 | struct perf_cpu_context *cpuctx; | |
6332 | ||
6333 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6334 | ||
3f1f3320 PZ |
6335 | if (cpuctx->unique_pmu == old_pmu) |
6336 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6337 | } |
6338 | } | |
6339 | ||
6340 | static void free_pmu_context(struct pmu *pmu) | |
6341 | { | |
6342 | struct pmu *i; | |
f5ffe02e | 6343 | |
8dc85d54 | 6344 | mutex_lock(&pmus_lock); |
0475f9ea | 6345 | /* |
8dc85d54 | 6346 | * Like a real lame refcount. |
0475f9ea | 6347 | */ |
51676957 PZ |
6348 | list_for_each_entry(i, &pmus, entry) { |
6349 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6350 | update_pmu_context(i, pmu); | |
8dc85d54 | 6351 | goto out; |
51676957 | 6352 | } |
8dc85d54 | 6353 | } |
d6d020e9 | 6354 | |
51676957 | 6355 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6356 | out: |
6357 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6358 | } |
2e80a82a | 6359 | static struct idr pmu_idr; |
d6d020e9 | 6360 | |
abe43400 PZ |
6361 | static ssize_t |
6362 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6363 | { | |
6364 | struct pmu *pmu = dev_get_drvdata(dev); | |
6365 | ||
6366 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6367 | } | |
90826ca7 | 6368 | static DEVICE_ATTR_RO(type); |
abe43400 | 6369 | |
62b85639 SE |
6370 | static ssize_t |
6371 | perf_event_mux_interval_ms_show(struct device *dev, | |
6372 | struct device_attribute *attr, | |
6373 | char *page) | |
6374 | { | |
6375 | struct pmu *pmu = dev_get_drvdata(dev); | |
6376 | ||
6377 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6378 | } | |
6379 | ||
6380 | static ssize_t | |
6381 | perf_event_mux_interval_ms_store(struct device *dev, | |
6382 | struct device_attribute *attr, | |
6383 | const char *buf, size_t count) | |
6384 | { | |
6385 | struct pmu *pmu = dev_get_drvdata(dev); | |
6386 | int timer, cpu, ret; | |
6387 | ||
6388 | ret = kstrtoint(buf, 0, &timer); | |
6389 | if (ret) | |
6390 | return ret; | |
6391 | ||
6392 | if (timer < 1) | |
6393 | return -EINVAL; | |
6394 | ||
6395 | /* same value, noting to do */ | |
6396 | if (timer == pmu->hrtimer_interval_ms) | |
6397 | return count; | |
6398 | ||
6399 | pmu->hrtimer_interval_ms = timer; | |
6400 | ||
6401 | /* update all cpuctx for this PMU */ | |
6402 | for_each_possible_cpu(cpu) { | |
6403 | struct perf_cpu_context *cpuctx; | |
6404 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6405 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6406 | ||
6407 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6408 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6409 | } | |
6410 | ||
6411 | return count; | |
6412 | } | |
90826ca7 | 6413 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
62b85639 | 6414 | |
90826ca7 GKH |
6415 | static struct attribute *pmu_dev_attrs[] = { |
6416 | &dev_attr_type.attr, | |
6417 | &dev_attr_perf_event_mux_interval_ms.attr, | |
6418 | NULL, | |
abe43400 | 6419 | }; |
90826ca7 | 6420 | ATTRIBUTE_GROUPS(pmu_dev); |
abe43400 PZ |
6421 | |
6422 | static int pmu_bus_running; | |
6423 | static struct bus_type pmu_bus = { | |
6424 | .name = "event_source", | |
90826ca7 | 6425 | .dev_groups = pmu_dev_groups, |
abe43400 PZ |
6426 | }; |
6427 | ||
6428 | static void pmu_dev_release(struct device *dev) | |
6429 | { | |
6430 | kfree(dev); | |
6431 | } | |
6432 | ||
6433 | static int pmu_dev_alloc(struct pmu *pmu) | |
6434 | { | |
6435 | int ret = -ENOMEM; | |
6436 | ||
6437 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6438 | if (!pmu->dev) | |
6439 | goto out; | |
6440 | ||
0c9d42ed | 6441 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6442 | device_initialize(pmu->dev); |
6443 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6444 | if (ret) | |
6445 | goto free_dev; | |
6446 | ||
6447 | dev_set_drvdata(pmu->dev, pmu); | |
6448 | pmu->dev->bus = &pmu_bus; | |
6449 | pmu->dev->release = pmu_dev_release; | |
6450 | ret = device_add(pmu->dev); | |
6451 | if (ret) | |
6452 | goto free_dev; | |
6453 | ||
6454 | out: | |
6455 | return ret; | |
6456 | ||
6457 | free_dev: | |
6458 | put_device(pmu->dev); | |
6459 | goto out; | |
6460 | } | |
6461 | ||
547e9fd7 | 6462 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6463 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6464 | |
03d8e80b | 6465 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6466 | { |
108b02cf | 6467 | int cpu, ret; |
24f1e32c | 6468 | |
b0a873eb | 6469 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6470 | ret = -ENOMEM; |
6471 | pmu->pmu_disable_count = alloc_percpu(int); | |
6472 | if (!pmu->pmu_disable_count) | |
6473 | goto unlock; | |
f29ac756 | 6474 | |
2e80a82a PZ |
6475 | pmu->type = -1; |
6476 | if (!name) | |
6477 | goto skip_type; | |
6478 | pmu->name = name; | |
6479 | ||
6480 | if (type < 0) { | |
0e9c3be2 TH |
6481 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6482 | if (type < 0) { | |
6483 | ret = type; | |
2e80a82a PZ |
6484 | goto free_pdc; |
6485 | } | |
6486 | } | |
6487 | pmu->type = type; | |
6488 | ||
abe43400 PZ |
6489 | if (pmu_bus_running) { |
6490 | ret = pmu_dev_alloc(pmu); | |
6491 | if (ret) | |
6492 | goto free_idr; | |
6493 | } | |
6494 | ||
2e80a82a | 6495 | skip_type: |
8dc85d54 PZ |
6496 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6497 | if (pmu->pmu_cpu_context) | |
6498 | goto got_cpu_context; | |
f29ac756 | 6499 | |
c4814202 | 6500 | ret = -ENOMEM; |
108b02cf PZ |
6501 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6502 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6503 | goto free_dev; |
f344011c | 6504 | |
108b02cf PZ |
6505 | for_each_possible_cpu(cpu) { |
6506 | struct perf_cpu_context *cpuctx; | |
6507 | ||
6508 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6509 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6510 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6511 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6512 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6513 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6514 | |
6515 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6516 | ||
e9d2b064 | 6517 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
3f1f3320 | 6518 | cpuctx->unique_pmu = pmu; |
108b02cf | 6519 | } |
76e1d904 | 6520 | |
8dc85d54 | 6521 | got_cpu_context: |
ad5133b7 PZ |
6522 | if (!pmu->start_txn) { |
6523 | if (pmu->pmu_enable) { | |
6524 | /* | |
6525 | * If we have pmu_enable/pmu_disable calls, install | |
6526 | * transaction stubs that use that to try and batch | |
6527 | * hardware accesses. | |
6528 | */ | |
6529 | pmu->start_txn = perf_pmu_start_txn; | |
6530 | pmu->commit_txn = perf_pmu_commit_txn; | |
6531 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6532 | } else { | |
6533 | pmu->start_txn = perf_pmu_nop_void; | |
6534 | pmu->commit_txn = perf_pmu_nop_int; | |
6535 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6536 | } |
5c92d124 | 6537 | } |
15dbf27c | 6538 | |
ad5133b7 PZ |
6539 | if (!pmu->pmu_enable) { |
6540 | pmu->pmu_enable = perf_pmu_nop_void; | |
6541 | pmu->pmu_disable = perf_pmu_nop_void; | |
6542 | } | |
6543 | ||
35edc2a5 PZ |
6544 | if (!pmu->event_idx) |
6545 | pmu->event_idx = perf_event_idx_default; | |
6546 | ||
b0a873eb | 6547 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6548 | ret = 0; |
6549 | unlock: | |
b0a873eb PZ |
6550 | mutex_unlock(&pmus_lock); |
6551 | ||
33696fc0 | 6552 | return ret; |
108b02cf | 6553 | |
abe43400 PZ |
6554 | free_dev: |
6555 | device_del(pmu->dev); | |
6556 | put_device(pmu->dev); | |
6557 | ||
2e80a82a PZ |
6558 | free_idr: |
6559 | if (pmu->type >= PERF_TYPE_MAX) | |
6560 | idr_remove(&pmu_idr, pmu->type); | |
6561 | ||
108b02cf PZ |
6562 | free_pdc: |
6563 | free_percpu(pmu->pmu_disable_count); | |
6564 | goto unlock; | |
f29ac756 PZ |
6565 | } |
6566 | ||
b0a873eb | 6567 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6568 | { |
b0a873eb PZ |
6569 | mutex_lock(&pmus_lock); |
6570 | list_del_rcu(&pmu->entry); | |
6571 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6572 | |
0475f9ea | 6573 | /* |
cde8e884 PZ |
6574 | * We dereference the pmu list under both SRCU and regular RCU, so |
6575 | * synchronize against both of those. | |
0475f9ea | 6576 | */ |
b0a873eb | 6577 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6578 | synchronize_rcu(); |
d6d020e9 | 6579 | |
33696fc0 | 6580 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6581 | if (pmu->type >= PERF_TYPE_MAX) |
6582 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6583 | device_del(pmu->dev); |
6584 | put_device(pmu->dev); | |
51676957 | 6585 | free_pmu_context(pmu); |
b0a873eb | 6586 | } |
d6d020e9 | 6587 | |
b0a873eb PZ |
6588 | struct pmu *perf_init_event(struct perf_event *event) |
6589 | { | |
6590 | struct pmu *pmu = NULL; | |
6591 | int idx; | |
940c5b29 | 6592 | int ret; |
b0a873eb PZ |
6593 | |
6594 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6595 | |
6596 | rcu_read_lock(); | |
6597 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6598 | rcu_read_unlock(); | |
940c5b29 | 6599 | if (pmu) { |
7e5b2a01 | 6600 | event->pmu = pmu; |
940c5b29 LM |
6601 | ret = pmu->event_init(event); |
6602 | if (ret) | |
6603 | pmu = ERR_PTR(ret); | |
2e80a82a | 6604 | goto unlock; |
940c5b29 | 6605 | } |
2e80a82a | 6606 | |
b0a873eb | 6607 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 6608 | event->pmu = pmu; |
940c5b29 | 6609 | ret = pmu->event_init(event); |
b0a873eb | 6610 | if (!ret) |
e5f4d339 | 6611 | goto unlock; |
76e1d904 | 6612 | |
b0a873eb PZ |
6613 | if (ret != -ENOENT) { |
6614 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6615 | goto unlock; |
f344011c | 6616 | } |
5c92d124 | 6617 | } |
e5f4d339 PZ |
6618 | pmu = ERR_PTR(-ENOENT); |
6619 | unlock: | |
b0a873eb | 6620 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6621 | |
4aeb0b42 | 6622 | return pmu; |
5c92d124 IM |
6623 | } |
6624 | ||
4beb31f3 FW |
6625 | static void account_event_cpu(struct perf_event *event, int cpu) |
6626 | { | |
6627 | if (event->parent) | |
6628 | return; | |
6629 | ||
6630 | if (has_branch_stack(event)) { | |
6631 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6632 | atomic_inc(&per_cpu(perf_branch_stack_events, cpu)); | |
6633 | } | |
6634 | if (is_cgroup_event(event)) | |
6635 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
6636 | } | |
6637 | ||
766d6c07 FW |
6638 | static void account_event(struct perf_event *event) |
6639 | { | |
4beb31f3 FW |
6640 | if (event->parent) |
6641 | return; | |
6642 | ||
766d6c07 FW |
6643 | if (event->attach_state & PERF_ATTACH_TASK) |
6644 | static_key_slow_inc(&perf_sched_events.key); | |
6645 | if (event->attr.mmap || event->attr.mmap_data) | |
6646 | atomic_inc(&nr_mmap_events); | |
6647 | if (event->attr.comm) | |
6648 | atomic_inc(&nr_comm_events); | |
6649 | if (event->attr.task) | |
6650 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
6651 | if (event->attr.freq) { |
6652 | if (atomic_inc_return(&nr_freq_events) == 1) | |
6653 | tick_nohz_full_kick_all(); | |
6654 | } | |
4beb31f3 | 6655 | if (has_branch_stack(event)) |
766d6c07 | 6656 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 6657 | if (is_cgroup_event(event)) |
766d6c07 | 6658 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
6659 | |
6660 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
6661 | } |
6662 | ||
0793a61d | 6663 | /* |
cdd6c482 | 6664 | * Allocate and initialize a event structure |
0793a61d | 6665 | */ |
cdd6c482 | 6666 | static struct perf_event * |
c3f00c70 | 6667 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6668 | struct task_struct *task, |
6669 | struct perf_event *group_leader, | |
6670 | struct perf_event *parent_event, | |
4dc0da86 AK |
6671 | perf_overflow_handler_t overflow_handler, |
6672 | void *context) | |
0793a61d | 6673 | { |
51b0fe39 | 6674 | struct pmu *pmu; |
cdd6c482 IM |
6675 | struct perf_event *event; |
6676 | struct hw_perf_event *hwc; | |
90983b16 | 6677 | long err = -EINVAL; |
0793a61d | 6678 | |
66832eb4 ON |
6679 | if ((unsigned)cpu >= nr_cpu_ids) { |
6680 | if (!task || cpu != -1) | |
6681 | return ERR_PTR(-EINVAL); | |
6682 | } | |
6683 | ||
c3f00c70 | 6684 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6685 | if (!event) |
d5d2bc0d | 6686 | return ERR_PTR(-ENOMEM); |
0793a61d | 6687 | |
04289bb9 | 6688 | /* |
cdd6c482 | 6689 | * Single events are their own group leaders, with an |
04289bb9 IM |
6690 | * empty sibling list: |
6691 | */ | |
6692 | if (!group_leader) | |
cdd6c482 | 6693 | group_leader = event; |
04289bb9 | 6694 | |
cdd6c482 IM |
6695 | mutex_init(&event->child_mutex); |
6696 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6697 | |
cdd6c482 IM |
6698 | INIT_LIST_HEAD(&event->group_entry); |
6699 | INIT_LIST_HEAD(&event->event_entry); | |
6700 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 | 6701 | INIT_LIST_HEAD(&event->rb_entry); |
71ad88ef | 6702 | INIT_LIST_HEAD(&event->active_entry); |
f3ae75de SE |
6703 | INIT_HLIST_NODE(&event->hlist_entry); |
6704 | ||
10c6db11 | 6705 | |
cdd6c482 | 6706 | init_waitqueue_head(&event->waitq); |
e360adbe | 6707 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6708 | |
cdd6c482 | 6709 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6710 | |
a6fa941d | 6711 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6712 | event->cpu = cpu; |
6713 | event->attr = *attr; | |
6714 | event->group_leader = group_leader; | |
6715 | event->pmu = NULL; | |
cdd6c482 | 6716 | event->oncpu = -1; |
a96bbc16 | 6717 | |
cdd6c482 | 6718 | event->parent = parent_event; |
b84fbc9f | 6719 | |
17cf22c3 | 6720 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6721 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6722 | |
cdd6c482 | 6723 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6724 | |
d580ff86 PZ |
6725 | if (task) { |
6726 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6727 | |
6728 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6729 | event->hw.tp_target = task; | |
d580ff86 PZ |
6730 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6731 | /* | |
6732 | * hw_breakpoint is a bit difficult here.. | |
6733 | */ | |
f22c1bb6 | 6734 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6735 | event->hw.bp_target = task; |
6736 | #endif | |
6737 | } | |
6738 | ||
4dc0da86 | 6739 | if (!overflow_handler && parent_event) { |
b326e956 | 6740 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6741 | context = parent_event->overflow_handler_context; |
6742 | } | |
66832eb4 | 6743 | |
b326e956 | 6744 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6745 | event->overflow_handler_context = context; |
97eaf530 | 6746 | |
0231bb53 | 6747 | perf_event__state_init(event); |
a86ed508 | 6748 | |
4aeb0b42 | 6749 | pmu = NULL; |
b8e83514 | 6750 | |
cdd6c482 | 6751 | hwc = &event->hw; |
bd2b5b12 | 6752 | hwc->sample_period = attr->sample_period; |
0d48696f | 6753 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6754 | hwc->sample_period = 1; |
eced1dfc | 6755 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6756 | |
e7850595 | 6757 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6758 | |
2023b359 | 6759 | /* |
cdd6c482 | 6760 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6761 | */ |
3dab77fb | 6762 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 6763 | goto err_ns; |
2023b359 | 6764 | |
b0a873eb | 6765 | pmu = perf_init_event(event); |
4aeb0b42 | 6766 | if (!pmu) |
90983b16 FW |
6767 | goto err_ns; |
6768 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 6769 | err = PTR_ERR(pmu); |
90983b16 | 6770 | goto err_ns; |
621a01ea | 6771 | } |
d5d2bc0d | 6772 | |
cdd6c482 | 6773 | if (!event->parent) { |
927c7a9e FW |
6774 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6775 | err = get_callchain_buffers(); | |
90983b16 FW |
6776 | if (err) |
6777 | goto err_pmu; | |
d010b332 | 6778 | } |
f344011c | 6779 | } |
9ee318a7 | 6780 | |
cdd6c482 | 6781 | return event; |
90983b16 FW |
6782 | |
6783 | err_pmu: | |
6784 | if (event->destroy) | |
6785 | event->destroy(event); | |
6786 | err_ns: | |
6787 | if (event->ns) | |
6788 | put_pid_ns(event->ns); | |
6789 | kfree(event); | |
6790 | ||
6791 | return ERR_PTR(err); | |
0793a61d TG |
6792 | } |
6793 | ||
cdd6c482 IM |
6794 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6795 | struct perf_event_attr *attr) | |
974802ea | 6796 | { |
974802ea | 6797 | u32 size; |
cdf8073d | 6798 | int ret; |
974802ea PZ |
6799 | |
6800 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6801 | return -EFAULT; | |
6802 | ||
6803 | /* | |
6804 | * zero the full structure, so that a short copy will be nice. | |
6805 | */ | |
6806 | memset(attr, 0, sizeof(*attr)); | |
6807 | ||
6808 | ret = get_user(size, &uattr->size); | |
6809 | if (ret) | |
6810 | return ret; | |
6811 | ||
6812 | if (size > PAGE_SIZE) /* silly large */ | |
6813 | goto err_size; | |
6814 | ||
6815 | if (!size) /* abi compat */ | |
6816 | size = PERF_ATTR_SIZE_VER0; | |
6817 | ||
6818 | if (size < PERF_ATTR_SIZE_VER0) | |
6819 | goto err_size; | |
6820 | ||
6821 | /* | |
6822 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6823 | * ensure all the unknown bits are 0 - i.e. new |
6824 | * user-space does not rely on any kernel feature | |
6825 | * extensions we dont know about yet. | |
974802ea PZ |
6826 | */ |
6827 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6828 | unsigned char __user *addr; |
6829 | unsigned char __user *end; | |
6830 | unsigned char val; | |
974802ea | 6831 | |
cdf8073d IS |
6832 | addr = (void __user *)uattr + sizeof(*attr); |
6833 | end = (void __user *)uattr + size; | |
974802ea | 6834 | |
cdf8073d | 6835 | for (; addr < end; addr++) { |
974802ea PZ |
6836 | ret = get_user(val, addr); |
6837 | if (ret) | |
6838 | return ret; | |
6839 | if (val) | |
6840 | goto err_size; | |
6841 | } | |
b3e62e35 | 6842 | size = sizeof(*attr); |
974802ea PZ |
6843 | } |
6844 | ||
6845 | ret = copy_from_user(attr, uattr, size); | |
6846 | if (ret) | |
6847 | return -EFAULT; | |
6848 | ||
3090ffb5 SE |
6849 | /* disabled for now */ |
6850 | if (attr->mmap2) | |
6851 | return -EINVAL; | |
6852 | ||
cd757645 | 6853 | if (attr->__reserved_1) |
974802ea PZ |
6854 | return -EINVAL; |
6855 | ||
6856 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6857 | return -EINVAL; | |
6858 | ||
6859 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6860 | return -EINVAL; | |
6861 | ||
bce38cd5 SE |
6862 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6863 | u64 mask = attr->branch_sample_type; | |
6864 | ||
6865 | /* only using defined bits */ | |
6866 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6867 | return -EINVAL; | |
6868 | ||
6869 | /* at least one branch bit must be set */ | |
6870 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6871 | return -EINVAL; | |
6872 | ||
bce38cd5 SE |
6873 | /* propagate priv level, when not set for branch */ |
6874 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6875 | ||
6876 | /* exclude_kernel checked on syscall entry */ | |
6877 | if (!attr->exclude_kernel) | |
6878 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6879 | ||
6880 | if (!attr->exclude_user) | |
6881 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6882 | ||
6883 | if (!attr->exclude_hv) | |
6884 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6885 | /* | |
6886 | * adjust user setting (for HW filter setup) | |
6887 | */ | |
6888 | attr->branch_sample_type = mask; | |
6889 | } | |
e712209a SE |
6890 | /* privileged levels capture (kernel, hv): check permissions */ |
6891 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
6892 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
6893 | return -EACCES; | |
bce38cd5 | 6894 | } |
4018994f | 6895 | |
c5ebcedb | 6896 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 6897 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
6898 | if (ret) |
6899 | return ret; | |
6900 | } | |
6901 | ||
6902 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
6903 | if (!arch_perf_have_user_stack_dump()) | |
6904 | return -ENOSYS; | |
6905 | ||
6906 | /* | |
6907 | * We have __u32 type for the size, but so far | |
6908 | * we can only use __u16 as maximum due to the | |
6909 | * __u16 sample size limit. | |
6910 | */ | |
6911 | if (attr->sample_stack_user >= USHRT_MAX) | |
6912 | ret = -EINVAL; | |
6913 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
6914 | ret = -EINVAL; | |
6915 | } | |
4018994f | 6916 | |
974802ea PZ |
6917 | out: |
6918 | return ret; | |
6919 | ||
6920 | err_size: | |
6921 | put_user(sizeof(*attr), &uattr->size); | |
6922 | ret = -E2BIG; | |
6923 | goto out; | |
6924 | } | |
6925 | ||
ac9721f3 PZ |
6926 | static int |
6927 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6928 | { |
76369139 | 6929 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6930 | int ret = -EINVAL; |
6931 | ||
ac9721f3 | 6932 | if (!output_event) |
a4be7c27 PZ |
6933 | goto set; |
6934 | ||
ac9721f3 PZ |
6935 | /* don't allow circular references */ |
6936 | if (event == output_event) | |
a4be7c27 PZ |
6937 | goto out; |
6938 | ||
0f139300 PZ |
6939 | /* |
6940 | * Don't allow cross-cpu buffers | |
6941 | */ | |
6942 | if (output_event->cpu != event->cpu) | |
6943 | goto out; | |
6944 | ||
6945 | /* | |
76369139 | 6946 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6947 | */ |
6948 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6949 | goto out; | |
6950 | ||
a4be7c27 | 6951 | set: |
cdd6c482 | 6952 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6953 | /* Can't redirect output if we've got an active mmap() */ |
6954 | if (atomic_read(&event->mmap_count)) | |
6955 | goto unlock; | |
a4be7c27 | 6956 | |
9bb5d40c PZ |
6957 | old_rb = event->rb; |
6958 | ||
ac9721f3 | 6959 | if (output_event) { |
76369139 FW |
6960 | /* get the rb we want to redirect to */ |
6961 | rb = ring_buffer_get(output_event); | |
6962 | if (!rb) | |
ac9721f3 | 6963 | goto unlock; |
a4be7c27 PZ |
6964 | } |
6965 | ||
10c6db11 PZ |
6966 | if (old_rb) |
6967 | ring_buffer_detach(event, old_rb); | |
9bb5d40c PZ |
6968 | |
6969 | if (rb) | |
6970 | ring_buffer_attach(event, rb); | |
6971 | ||
6972 | rcu_assign_pointer(event->rb, rb); | |
6973 | ||
6974 | if (old_rb) { | |
6975 | ring_buffer_put(old_rb); | |
6976 | /* | |
6977 | * Since we detached before setting the new rb, so that we | |
6978 | * could attach the new rb, we could have missed a wakeup. | |
6979 | * Provide it now. | |
6980 | */ | |
6981 | wake_up_all(&event->waitq); | |
6982 | } | |
6983 | ||
a4be7c27 | 6984 | ret = 0; |
ac9721f3 PZ |
6985 | unlock: |
6986 | mutex_unlock(&event->mmap_mutex); | |
6987 | ||
a4be7c27 | 6988 | out: |
a4be7c27 PZ |
6989 | return ret; |
6990 | } | |
6991 | ||
0793a61d | 6992 | /** |
cdd6c482 | 6993 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6994 | * |
cdd6c482 | 6995 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6996 | * @pid: target pid |
9f66a381 | 6997 | * @cpu: target cpu |
cdd6c482 | 6998 | * @group_fd: group leader event fd |
0793a61d | 6999 | */ |
cdd6c482 IM |
7000 | SYSCALL_DEFINE5(perf_event_open, |
7001 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 7002 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 7003 | { |
b04243ef PZ |
7004 | struct perf_event *group_leader = NULL, *output_event = NULL; |
7005 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
7006 | struct perf_event_attr attr; |
7007 | struct perf_event_context *ctx; | |
7008 | struct file *event_file = NULL; | |
2903ff01 | 7009 | struct fd group = {NULL, 0}; |
38a81da2 | 7010 | struct task_struct *task = NULL; |
89a1e187 | 7011 | struct pmu *pmu; |
ea635c64 | 7012 | int event_fd; |
b04243ef | 7013 | int move_group = 0; |
dc86cabe | 7014 | int err; |
a21b0b35 | 7015 | int f_flags = O_RDWR; |
0793a61d | 7016 | |
2743a5b0 | 7017 | /* for future expandability... */ |
e5d1367f | 7018 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
7019 | return -EINVAL; |
7020 | ||
dc86cabe IM |
7021 | err = perf_copy_attr(attr_uptr, &attr); |
7022 | if (err) | |
7023 | return err; | |
eab656ae | 7024 | |
0764771d PZ |
7025 | if (!attr.exclude_kernel) { |
7026 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
7027 | return -EACCES; | |
7028 | } | |
7029 | ||
df58ab24 | 7030 | if (attr.freq) { |
cdd6c482 | 7031 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
7032 | return -EINVAL; |
7033 | } | |
7034 | ||
e5d1367f SE |
7035 | /* |
7036 | * In cgroup mode, the pid argument is used to pass the fd | |
7037 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
7038 | * designates the cpu on which to monitor threads from that | |
7039 | * cgroup. | |
7040 | */ | |
7041 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
7042 | return -EINVAL; | |
7043 | ||
a21b0b35 YD |
7044 | if (flags & PERF_FLAG_FD_CLOEXEC) |
7045 | f_flags |= O_CLOEXEC; | |
7046 | ||
7047 | event_fd = get_unused_fd_flags(f_flags); | |
ea635c64 AV |
7048 | if (event_fd < 0) |
7049 | return event_fd; | |
7050 | ||
ac9721f3 | 7051 | if (group_fd != -1) { |
2903ff01 AV |
7052 | err = perf_fget_light(group_fd, &group); |
7053 | if (err) | |
d14b12d7 | 7054 | goto err_fd; |
2903ff01 | 7055 | group_leader = group.file->private_data; |
ac9721f3 PZ |
7056 | if (flags & PERF_FLAG_FD_OUTPUT) |
7057 | output_event = group_leader; | |
7058 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
7059 | group_leader = NULL; | |
7060 | } | |
7061 | ||
e5d1367f | 7062 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
7063 | task = find_lively_task_by_vpid(pid); |
7064 | if (IS_ERR(task)) { | |
7065 | err = PTR_ERR(task); | |
7066 | goto err_group_fd; | |
7067 | } | |
7068 | } | |
7069 | ||
fbfc623f YZ |
7070 | get_online_cpus(); |
7071 | ||
4dc0da86 AK |
7072 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
7073 | NULL, NULL); | |
d14b12d7 SE |
7074 | if (IS_ERR(event)) { |
7075 | err = PTR_ERR(event); | |
c6be5a5c | 7076 | goto err_task; |
d14b12d7 SE |
7077 | } |
7078 | ||
e5d1367f SE |
7079 | if (flags & PERF_FLAG_PID_CGROUP) { |
7080 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
766d6c07 FW |
7081 | if (err) { |
7082 | __free_event(event); | |
7083 | goto err_task; | |
7084 | } | |
e5d1367f SE |
7085 | } |
7086 | ||
766d6c07 FW |
7087 | account_event(event); |
7088 | ||
89a1e187 PZ |
7089 | /* |
7090 | * Special case software events and allow them to be part of | |
7091 | * any hardware group. | |
7092 | */ | |
7093 | pmu = event->pmu; | |
b04243ef PZ |
7094 | |
7095 | if (group_leader && | |
7096 | (is_software_event(event) != is_software_event(group_leader))) { | |
7097 | if (is_software_event(event)) { | |
7098 | /* | |
7099 | * If event and group_leader are not both a software | |
7100 | * event, and event is, then group leader is not. | |
7101 | * | |
7102 | * Allow the addition of software events to !software | |
7103 | * groups, this is safe because software events never | |
7104 | * fail to schedule. | |
7105 | */ | |
7106 | pmu = group_leader->pmu; | |
7107 | } else if (is_software_event(group_leader) && | |
7108 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
7109 | /* | |
7110 | * In case the group is a pure software group, and we | |
7111 | * try to add a hardware event, move the whole group to | |
7112 | * the hardware context. | |
7113 | */ | |
7114 | move_group = 1; | |
7115 | } | |
7116 | } | |
89a1e187 PZ |
7117 | |
7118 | /* | |
7119 | * Get the target context (task or percpu): | |
7120 | */ | |
e2d37cd2 | 7121 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
7122 | if (IS_ERR(ctx)) { |
7123 | err = PTR_ERR(ctx); | |
c6be5a5c | 7124 | goto err_alloc; |
89a1e187 PZ |
7125 | } |
7126 | ||
fd1edb3a PZ |
7127 | if (task) { |
7128 | put_task_struct(task); | |
7129 | task = NULL; | |
7130 | } | |
7131 | ||
ccff286d | 7132 | /* |
cdd6c482 | 7133 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 7134 | */ |
ac9721f3 | 7135 | if (group_leader) { |
dc86cabe | 7136 | err = -EINVAL; |
04289bb9 | 7137 | |
04289bb9 | 7138 | /* |
ccff286d IM |
7139 | * Do not allow a recursive hierarchy (this new sibling |
7140 | * becoming part of another group-sibling): | |
7141 | */ | |
7142 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 7143 | goto err_context; |
ccff286d IM |
7144 | /* |
7145 | * Do not allow to attach to a group in a different | |
7146 | * task or CPU context: | |
04289bb9 | 7147 | */ |
b04243ef PZ |
7148 | if (move_group) { |
7149 | if (group_leader->ctx->type != ctx->type) | |
7150 | goto err_context; | |
7151 | } else { | |
7152 | if (group_leader->ctx != ctx) | |
7153 | goto err_context; | |
7154 | } | |
7155 | ||
3b6f9e5c PM |
7156 | /* |
7157 | * Only a group leader can be exclusive or pinned | |
7158 | */ | |
0d48696f | 7159 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 7160 | goto err_context; |
ac9721f3 PZ |
7161 | } |
7162 | ||
7163 | if (output_event) { | |
7164 | err = perf_event_set_output(event, output_event); | |
7165 | if (err) | |
c3f00c70 | 7166 | goto err_context; |
ac9721f3 | 7167 | } |
0793a61d | 7168 | |
a21b0b35 YD |
7169 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
7170 | f_flags); | |
ea635c64 AV |
7171 | if (IS_ERR(event_file)) { |
7172 | err = PTR_ERR(event_file); | |
c3f00c70 | 7173 | goto err_context; |
ea635c64 | 7174 | } |
9b51f66d | 7175 | |
b04243ef PZ |
7176 | if (move_group) { |
7177 | struct perf_event_context *gctx = group_leader->ctx; | |
7178 | ||
7179 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 7180 | perf_remove_from_context(group_leader); |
0231bb53 JO |
7181 | |
7182 | /* | |
7183 | * Removing from the context ends up with disabled | |
7184 | * event. What we want here is event in the initial | |
7185 | * startup state, ready to be add into new context. | |
7186 | */ | |
7187 | perf_event__state_init(group_leader); | |
b04243ef PZ |
7188 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7189 | group_entry) { | |
fe4b04fa | 7190 | perf_remove_from_context(sibling); |
0231bb53 | 7191 | perf_event__state_init(sibling); |
b04243ef PZ |
7192 | put_ctx(gctx); |
7193 | } | |
7194 | mutex_unlock(&gctx->mutex); | |
7195 | put_ctx(gctx); | |
ea635c64 | 7196 | } |
9b51f66d | 7197 | |
ad3a37de | 7198 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 7199 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
7200 | |
7201 | if (move_group) { | |
0cda4c02 | 7202 | synchronize_rcu(); |
e2d37cd2 | 7203 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
7204 | get_ctx(ctx); |
7205 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
7206 | group_entry) { | |
e2d37cd2 | 7207 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
7208 | get_ctx(ctx); |
7209 | } | |
7210 | } | |
7211 | ||
e2d37cd2 | 7212 | perf_install_in_context(ctx, event, event->cpu); |
fe4b04fa | 7213 | perf_unpin_context(ctx); |
d859e29f | 7214 | mutex_unlock(&ctx->mutex); |
9b51f66d | 7215 | |
fbfc623f YZ |
7216 | put_online_cpus(); |
7217 | ||
cdd6c482 | 7218 | event->owner = current; |
8882135b | 7219 | |
cdd6c482 IM |
7220 | mutex_lock(¤t->perf_event_mutex); |
7221 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
7222 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 7223 | |
c320c7b7 ACM |
7224 | /* |
7225 | * Precalculate sample_data sizes | |
7226 | */ | |
7227 | perf_event__header_size(event); | |
6844c09d | 7228 | perf_event__id_header_size(event); |
c320c7b7 | 7229 | |
8a49542c PZ |
7230 | /* |
7231 | * Drop the reference on the group_event after placing the | |
7232 | * new event on the sibling_list. This ensures destruction | |
7233 | * of the group leader will find the pointer to itself in | |
7234 | * perf_group_detach(). | |
7235 | */ | |
2903ff01 | 7236 | fdput(group); |
ea635c64 AV |
7237 | fd_install(event_fd, event_file); |
7238 | return event_fd; | |
0793a61d | 7239 | |
c3f00c70 | 7240 | err_context: |
fe4b04fa | 7241 | perf_unpin_context(ctx); |
ea635c64 | 7242 | put_ctx(ctx); |
c6be5a5c | 7243 | err_alloc: |
ea635c64 | 7244 | free_event(event); |
e7d0bc04 | 7245 | err_task: |
fbfc623f | 7246 | put_online_cpus(); |
e7d0bc04 PZ |
7247 | if (task) |
7248 | put_task_struct(task); | |
89a1e187 | 7249 | err_group_fd: |
2903ff01 | 7250 | fdput(group); |
ea635c64 AV |
7251 | err_fd: |
7252 | put_unused_fd(event_fd); | |
dc86cabe | 7253 | return err; |
0793a61d TG |
7254 | } |
7255 | ||
fb0459d7 AV |
7256 | /** |
7257 | * perf_event_create_kernel_counter | |
7258 | * | |
7259 | * @attr: attributes of the counter to create | |
7260 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 7261 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
7262 | */ |
7263 | struct perf_event * | |
7264 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 7265 | struct task_struct *task, |
4dc0da86 AK |
7266 | perf_overflow_handler_t overflow_handler, |
7267 | void *context) | |
fb0459d7 | 7268 | { |
fb0459d7 | 7269 | struct perf_event_context *ctx; |
c3f00c70 | 7270 | struct perf_event *event; |
fb0459d7 | 7271 | int err; |
d859e29f | 7272 | |
fb0459d7 AV |
7273 | /* |
7274 | * Get the target context (task or percpu): | |
7275 | */ | |
d859e29f | 7276 | |
4dc0da86 AK |
7277 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
7278 | overflow_handler, context); | |
c3f00c70 PZ |
7279 | if (IS_ERR(event)) { |
7280 | err = PTR_ERR(event); | |
7281 | goto err; | |
7282 | } | |
d859e29f | 7283 | |
766d6c07 FW |
7284 | account_event(event); |
7285 | ||
38a81da2 | 7286 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
7287 | if (IS_ERR(ctx)) { |
7288 | err = PTR_ERR(ctx); | |
c3f00c70 | 7289 | goto err_free; |
d859e29f | 7290 | } |
fb0459d7 | 7291 | |
fb0459d7 AV |
7292 | WARN_ON_ONCE(ctx->parent_ctx); |
7293 | mutex_lock(&ctx->mutex); | |
7294 | perf_install_in_context(ctx, event, cpu); | |
fe4b04fa | 7295 | perf_unpin_context(ctx); |
fb0459d7 AV |
7296 | mutex_unlock(&ctx->mutex); |
7297 | ||
fb0459d7 AV |
7298 | return event; |
7299 | ||
c3f00c70 PZ |
7300 | err_free: |
7301 | free_event(event); | |
7302 | err: | |
c6567f64 | 7303 | return ERR_PTR(err); |
9b51f66d | 7304 | } |
fb0459d7 | 7305 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7306 | |
0cda4c02 YZ |
7307 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7308 | { | |
7309 | struct perf_event_context *src_ctx; | |
7310 | struct perf_event_context *dst_ctx; | |
7311 | struct perf_event *event, *tmp; | |
7312 | LIST_HEAD(events); | |
7313 | ||
7314 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7315 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7316 | ||
7317 | mutex_lock(&src_ctx->mutex); | |
7318 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
7319 | event_entry) { | |
7320 | perf_remove_from_context(event); | |
9a545de0 | 7321 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 7322 | put_ctx(src_ctx); |
9886167d | 7323 | list_add(&event->migrate_entry, &events); |
0cda4c02 YZ |
7324 | } |
7325 | mutex_unlock(&src_ctx->mutex); | |
7326 | ||
7327 | synchronize_rcu(); | |
7328 | ||
7329 | mutex_lock(&dst_ctx->mutex); | |
9886167d PZ |
7330 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
7331 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
7332 | if (event->state >= PERF_EVENT_STATE_OFF) |
7333 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 7334 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
7335 | perf_install_in_context(dst_ctx, event, dst_cpu); |
7336 | get_ctx(dst_ctx); | |
7337 | } | |
7338 | mutex_unlock(&dst_ctx->mutex); | |
7339 | } | |
7340 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7341 | ||
cdd6c482 | 7342 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7343 | struct task_struct *child) |
d859e29f | 7344 | { |
cdd6c482 | 7345 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7346 | u64 child_val; |
d859e29f | 7347 | |
cdd6c482 IM |
7348 | if (child_event->attr.inherit_stat) |
7349 | perf_event_read_event(child_event, child); | |
38b200d6 | 7350 | |
b5e58793 | 7351 | child_val = perf_event_count(child_event); |
d859e29f PM |
7352 | |
7353 | /* | |
7354 | * Add back the child's count to the parent's count: | |
7355 | */ | |
a6e6dea6 | 7356 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7357 | atomic64_add(child_event->total_time_enabled, |
7358 | &parent_event->child_total_time_enabled); | |
7359 | atomic64_add(child_event->total_time_running, | |
7360 | &parent_event->child_total_time_running); | |
d859e29f PM |
7361 | |
7362 | /* | |
cdd6c482 | 7363 | * Remove this event from the parent's list |
d859e29f | 7364 | */ |
cdd6c482 IM |
7365 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7366 | mutex_lock(&parent_event->child_mutex); | |
7367 | list_del_init(&child_event->child_list); | |
7368 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
7369 | |
7370 | /* | |
cdd6c482 | 7371 | * Release the parent event, if this was the last |
d859e29f PM |
7372 | * reference to it. |
7373 | */ | |
a6fa941d | 7374 | put_event(parent_event); |
d859e29f PM |
7375 | } |
7376 | ||
9b51f66d | 7377 | static void |
cdd6c482 IM |
7378 | __perf_event_exit_task(struct perf_event *child_event, |
7379 | struct perf_event_context *child_ctx, | |
38b200d6 | 7380 | struct task_struct *child) |
9b51f66d | 7381 | { |
38b435b1 PZ |
7382 | if (child_event->parent) { |
7383 | raw_spin_lock_irq(&child_ctx->lock); | |
7384 | perf_group_detach(child_event); | |
7385 | raw_spin_unlock_irq(&child_ctx->lock); | |
7386 | } | |
9b51f66d | 7387 | |
fe4b04fa | 7388 | perf_remove_from_context(child_event); |
0cc0c027 | 7389 | |
9b51f66d | 7390 | /* |
38b435b1 | 7391 | * It can happen that the parent exits first, and has events |
9b51f66d | 7392 | * that are still around due to the child reference. These |
38b435b1 | 7393 | * events need to be zapped. |
9b51f66d | 7394 | */ |
38b435b1 | 7395 | if (child_event->parent) { |
cdd6c482 IM |
7396 | sync_child_event(child_event, child); |
7397 | free_event(child_event); | |
4bcf349a | 7398 | } |
9b51f66d IM |
7399 | } |
7400 | ||
8dc85d54 | 7401 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7402 | { |
cdd6c482 IM |
7403 | struct perf_event *child_event, *tmp; |
7404 | struct perf_event_context *child_ctx; | |
a63eaf34 | 7405 | unsigned long flags; |
9b51f66d | 7406 | |
8dc85d54 | 7407 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7408 | perf_event_task(child, NULL, 0); |
9b51f66d | 7409 | return; |
9f498cc5 | 7410 | } |
9b51f66d | 7411 | |
a63eaf34 | 7412 | local_irq_save(flags); |
ad3a37de PM |
7413 | /* |
7414 | * We can't reschedule here because interrupts are disabled, | |
7415 | * and either child is current or it is a task that can't be | |
7416 | * scheduled, so we are now safe from rescheduling changing | |
7417 | * our context. | |
7418 | */ | |
806839b2 | 7419 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7420 | |
7421 | /* | |
7422 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7423 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7424 | * incremented the context's refcount before we do put_ctx below. |
7425 | */ | |
e625cce1 | 7426 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7427 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7428 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
7429 | /* |
7430 | * If this context is a clone; unclone it so it can't get | |
7431 | * swapped to another process while we're removing all | |
cdd6c482 | 7432 | * the events from it. |
71a851b4 PZ |
7433 | */ |
7434 | unclone_ctx(child_ctx); | |
5e942bb3 | 7435 | update_context_time(child_ctx); |
e625cce1 | 7436 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
7437 | |
7438 | /* | |
cdd6c482 IM |
7439 | * Report the task dead after unscheduling the events so that we |
7440 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7441 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7442 | */ |
cdd6c482 | 7443 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 7444 | |
66fff224 PZ |
7445 | /* |
7446 | * We can recurse on the same lock type through: | |
7447 | * | |
cdd6c482 IM |
7448 | * __perf_event_exit_task() |
7449 | * sync_child_event() | |
a6fa941d AV |
7450 | * put_event() |
7451 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
7452 | * |
7453 | * But since its the parent context it won't be the same instance. | |
7454 | */ | |
a0507c84 | 7455 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 7456 | |
8bc20959 | 7457 | again: |
889ff015 FW |
7458 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
7459 | group_entry) | |
7460 | __perf_event_exit_task(child_event, child_ctx, child); | |
7461 | ||
7462 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 7463 | group_entry) |
cdd6c482 | 7464 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
7465 | |
7466 | /* | |
cdd6c482 | 7467 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
7468 | * its siblings to the list, but we obtained 'tmp' before that which |
7469 | * will still point to the list head terminating the iteration. | |
7470 | */ | |
889ff015 FW |
7471 | if (!list_empty(&child_ctx->pinned_groups) || |
7472 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 7473 | goto again; |
a63eaf34 PM |
7474 | |
7475 | mutex_unlock(&child_ctx->mutex); | |
7476 | ||
7477 | put_ctx(child_ctx); | |
9b51f66d IM |
7478 | } |
7479 | ||
8dc85d54 PZ |
7480 | /* |
7481 | * When a child task exits, feed back event values to parent events. | |
7482 | */ | |
7483 | void perf_event_exit_task(struct task_struct *child) | |
7484 | { | |
8882135b | 7485 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
7486 | int ctxn; |
7487 | ||
8882135b PZ |
7488 | mutex_lock(&child->perf_event_mutex); |
7489 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
7490 | owner_entry) { | |
7491 | list_del_init(&event->owner_entry); | |
7492 | ||
7493 | /* | |
7494 | * Ensure the list deletion is visible before we clear | |
7495 | * the owner, closes a race against perf_release() where | |
7496 | * we need to serialize on the owner->perf_event_mutex. | |
7497 | */ | |
7498 | smp_wmb(); | |
7499 | event->owner = NULL; | |
7500 | } | |
7501 | mutex_unlock(&child->perf_event_mutex); | |
7502 | ||
8dc85d54 PZ |
7503 | for_each_task_context_nr(ctxn) |
7504 | perf_event_exit_task_context(child, ctxn); | |
7505 | } | |
7506 | ||
889ff015 FW |
7507 | static void perf_free_event(struct perf_event *event, |
7508 | struct perf_event_context *ctx) | |
7509 | { | |
7510 | struct perf_event *parent = event->parent; | |
7511 | ||
7512 | if (WARN_ON_ONCE(!parent)) | |
7513 | return; | |
7514 | ||
7515 | mutex_lock(&parent->child_mutex); | |
7516 | list_del_init(&event->child_list); | |
7517 | mutex_unlock(&parent->child_mutex); | |
7518 | ||
a6fa941d | 7519 | put_event(parent); |
889ff015 | 7520 | |
8a49542c | 7521 | perf_group_detach(event); |
889ff015 FW |
7522 | list_del_event(event, ctx); |
7523 | free_event(event); | |
7524 | } | |
7525 | ||
bbbee908 PZ |
7526 | /* |
7527 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 7528 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 7529 | */ |
cdd6c482 | 7530 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 7531 | { |
8dc85d54 | 7532 | struct perf_event_context *ctx; |
cdd6c482 | 7533 | struct perf_event *event, *tmp; |
8dc85d54 | 7534 | int ctxn; |
bbbee908 | 7535 | |
8dc85d54 PZ |
7536 | for_each_task_context_nr(ctxn) { |
7537 | ctx = task->perf_event_ctxp[ctxn]; | |
7538 | if (!ctx) | |
7539 | continue; | |
bbbee908 | 7540 | |
8dc85d54 | 7541 | mutex_lock(&ctx->mutex); |
bbbee908 | 7542 | again: |
8dc85d54 PZ |
7543 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
7544 | group_entry) | |
7545 | perf_free_event(event, ctx); | |
bbbee908 | 7546 | |
8dc85d54 PZ |
7547 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
7548 | group_entry) | |
7549 | perf_free_event(event, ctx); | |
bbbee908 | 7550 | |
8dc85d54 PZ |
7551 | if (!list_empty(&ctx->pinned_groups) || |
7552 | !list_empty(&ctx->flexible_groups)) | |
7553 | goto again; | |
bbbee908 | 7554 | |
8dc85d54 | 7555 | mutex_unlock(&ctx->mutex); |
bbbee908 | 7556 | |
8dc85d54 PZ |
7557 | put_ctx(ctx); |
7558 | } | |
889ff015 FW |
7559 | } |
7560 | ||
4e231c79 PZ |
7561 | void perf_event_delayed_put(struct task_struct *task) |
7562 | { | |
7563 | int ctxn; | |
7564 | ||
7565 | for_each_task_context_nr(ctxn) | |
7566 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7567 | } | |
7568 | ||
97dee4f3 PZ |
7569 | /* |
7570 | * inherit a event from parent task to child task: | |
7571 | */ | |
7572 | static struct perf_event * | |
7573 | inherit_event(struct perf_event *parent_event, | |
7574 | struct task_struct *parent, | |
7575 | struct perf_event_context *parent_ctx, | |
7576 | struct task_struct *child, | |
7577 | struct perf_event *group_leader, | |
7578 | struct perf_event_context *child_ctx) | |
7579 | { | |
7580 | struct perf_event *child_event; | |
cee010ec | 7581 | unsigned long flags; |
97dee4f3 PZ |
7582 | |
7583 | /* | |
7584 | * Instead of creating recursive hierarchies of events, | |
7585 | * we link inherited events back to the original parent, | |
7586 | * which has a filp for sure, which we use as the reference | |
7587 | * count: | |
7588 | */ | |
7589 | if (parent_event->parent) | |
7590 | parent_event = parent_event->parent; | |
7591 | ||
7592 | child_event = perf_event_alloc(&parent_event->attr, | |
7593 | parent_event->cpu, | |
d580ff86 | 7594 | child, |
97dee4f3 | 7595 | group_leader, parent_event, |
4dc0da86 | 7596 | NULL, NULL); |
97dee4f3 PZ |
7597 | if (IS_ERR(child_event)) |
7598 | return child_event; | |
a6fa941d AV |
7599 | |
7600 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7601 | free_event(child_event); | |
7602 | return NULL; | |
7603 | } | |
7604 | ||
97dee4f3 PZ |
7605 | get_ctx(child_ctx); |
7606 | ||
7607 | /* | |
7608 | * Make the child state follow the state of the parent event, | |
7609 | * not its attr.disabled bit. We hold the parent's mutex, | |
7610 | * so we won't race with perf_event_{en, dis}able_family. | |
7611 | */ | |
7612 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7613 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7614 | else | |
7615 | child_event->state = PERF_EVENT_STATE_OFF; | |
7616 | ||
7617 | if (parent_event->attr.freq) { | |
7618 | u64 sample_period = parent_event->hw.sample_period; | |
7619 | struct hw_perf_event *hwc = &child_event->hw; | |
7620 | ||
7621 | hwc->sample_period = sample_period; | |
7622 | hwc->last_period = sample_period; | |
7623 | ||
7624 | local64_set(&hwc->period_left, sample_period); | |
7625 | } | |
7626 | ||
7627 | child_event->ctx = child_ctx; | |
7628 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7629 | child_event->overflow_handler_context |
7630 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7631 | |
614b6780 TG |
7632 | /* |
7633 | * Precalculate sample_data sizes | |
7634 | */ | |
7635 | perf_event__header_size(child_event); | |
6844c09d | 7636 | perf_event__id_header_size(child_event); |
614b6780 | 7637 | |
97dee4f3 PZ |
7638 | /* |
7639 | * Link it up in the child's context: | |
7640 | */ | |
cee010ec | 7641 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7642 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7643 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7644 | |
97dee4f3 PZ |
7645 | /* |
7646 | * Link this into the parent event's child list | |
7647 | */ | |
7648 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7649 | mutex_lock(&parent_event->child_mutex); | |
7650 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7651 | mutex_unlock(&parent_event->child_mutex); | |
7652 | ||
7653 | return child_event; | |
7654 | } | |
7655 | ||
7656 | static int inherit_group(struct perf_event *parent_event, | |
7657 | struct task_struct *parent, | |
7658 | struct perf_event_context *parent_ctx, | |
7659 | struct task_struct *child, | |
7660 | struct perf_event_context *child_ctx) | |
7661 | { | |
7662 | struct perf_event *leader; | |
7663 | struct perf_event *sub; | |
7664 | struct perf_event *child_ctr; | |
7665 | ||
7666 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7667 | child, NULL, child_ctx); | |
7668 | if (IS_ERR(leader)) | |
7669 | return PTR_ERR(leader); | |
7670 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7671 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7672 | child, leader, child_ctx); | |
7673 | if (IS_ERR(child_ctr)) | |
7674 | return PTR_ERR(child_ctr); | |
7675 | } | |
7676 | return 0; | |
889ff015 FW |
7677 | } |
7678 | ||
7679 | static int | |
7680 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7681 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7682 | struct task_struct *child, int ctxn, |
889ff015 FW |
7683 | int *inherited_all) |
7684 | { | |
7685 | int ret; | |
8dc85d54 | 7686 | struct perf_event_context *child_ctx; |
889ff015 FW |
7687 | |
7688 | if (!event->attr.inherit) { | |
7689 | *inherited_all = 0; | |
7690 | return 0; | |
bbbee908 PZ |
7691 | } |
7692 | ||
fe4b04fa | 7693 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7694 | if (!child_ctx) { |
7695 | /* | |
7696 | * This is executed from the parent task context, so | |
7697 | * inherit events that have been marked for cloning. | |
7698 | * First allocate and initialize a context for the | |
7699 | * child. | |
7700 | */ | |
bbbee908 | 7701 | |
734df5ab | 7702 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
7703 | if (!child_ctx) |
7704 | return -ENOMEM; | |
bbbee908 | 7705 | |
8dc85d54 | 7706 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7707 | } |
7708 | ||
7709 | ret = inherit_group(event, parent, parent_ctx, | |
7710 | child, child_ctx); | |
7711 | ||
7712 | if (ret) | |
7713 | *inherited_all = 0; | |
7714 | ||
7715 | return ret; | |
bbbee908 PZ |
7716 | } |
7717 | ||
9b51f66d | 7718 | /* |
cdd6c482 | 7719 | * Initialize the perf_event context in task_struct |
9b51f66d | 7720 | */ |
8dc85d54 | 7721 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7722 | { |
889ff015 | 7723 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7724 | struct perf_event_context *cloned_ctx; |
7725 | struct perf_event *event; | |
9b51f66d | 7726 | struct task_struct *parent = current; |
564c2b21 | 7727 | int inherited_all = 1; |
dddd3379 | 7728 | unsigned long flags; |
6ab423e0 | 7729 | int ret = 0; |
9b51f66d | 7730 | |
8dc85d54 | 7731 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7732 | return 0; |
7733 | ||
ad3a37de | 7734 | /* |
25346b93 PM |
7735 | * If the parent's context is a clone, pin it so it won't get |
7736 | * swapped under us. | |
ad3a37de | 7737 | */ |
8dc85d54 | 7738 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 7739 | |
ad3a37de PM |
7740 | /* |
7741 | * No need to check if parent_ctx != NULL here; since we saw | |
7742 | * it non-NULL earlier, the only reason for it to become NULL | |
7743 | * is if we exit, and since we're currently in the middle of | |
7744 | * a fork we can't be exiting at the same time. | |
7745 | */ | |
ad3a37de | 7746 | |
9b51f66d IM |
7747 | /* |
7748 | * Lock the parent list. No need to lock the child - not PID | |
7749 | * hashed yet and not running, so nobody can access it. | |
7750 | */ | |
d859e29f | 7751 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7752 | |
7753 | /* | |
7754 | * We dont have to disable NMIs - we are only looking at | |
7755 | * the list, not manipulating it: | |
7756 | */ | |
889ff015 | 7757 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7758 | ret = inherit_task_group(event, parent, parent_ctx, |
7759 | child, ctxn, &inherited_all); | |
889ff015 FW |
7760 | if (ret) |
7761 | break; | |
7762 | } | |
b93f7978 | 7763 | |
dddd3379 TG |
7764 | /* |
7765 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7766 | * to allocations, but we need to prevent rotation because | |
7767 | * rotate_ctx() will change the list from interrupt context. | |
7768 | */ | |
7769 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7770 | parent_ctx->rotate_disable = 1; | |
7771 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7772 | ||
889ff015 | 7773 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7774 | ret = inherit_task_group(event, parent, parent_ctx, |
7775 | child, ctxn, &inherited_all); | |
889ff015 | 7776 | if (ret) |
9b51f66d | 7777 | break; |
564c2b21 PM |
7778 | } |
7779 | ||
dddd3379 TG |
7780 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7781 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7782 | |
8dc85d54 | 7783 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7784 | |
05cbaa28 | 7785 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7786 | /* |
7787 | * Mark the child context as a clone of the parent | |
7788 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7789 | * |
7790 | * Note that if the parent is a clone, the holding of | |
7791 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7792 | */ |
c5ed5145 | 7793 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7794 | if (cloned_ctx) { |
7795 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7796 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7797 | } else { |
7798 | child_ctx->parent_ctx = parent_ctx; | |
7799 | child_ctx->parent_gen = parent_ctx->generation; | |
7800 | } | |
7801 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7802 | } |
7803 | ||
c5ed5145 | 7804 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7805 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7806 | |
25346b93 | 7807 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7808 | put_ctx(parent_ctx); |
ad3a37de | 7809 | |
6ab423e0 | 7810 | return ret; |
9b51f66d IM |
7811 | } |
7812 | ||
8dc85d54 PZ |
7813 | /* |
7814 | * Initialize the perf_event context in task_struct | |
7815 | */ | |
7816 | int perf_event_init_task(struct task_struct *child) | |
7817 | { | |
7818 | int ctxn, ret; | |
7819 | ||
8550d7cb ON |
7820 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7821 | mutex_init(&child->perf_event_mutex); | |
7822 | INIT_LIST_HEAD(&child->perf_event_list); | |
7823 | ||
8dc85d54 PZ |
7824 | for_each_task_context_nr(ctxn) { |
7825 | ret = perf_event_init_context(child, ctxn); | |
7826 | if (ret) | |
7827 | return ret; | |
7828 | } | |
7829 | ||
7830 | return 0; | |
7831 | } | |
7832 | ||
220b140b PM |
7833 | static void __init perf_event_init_all_cpus(void) |
7834 | { | |
b28ab83c | 7835 | struct swevent_htable *swhash; |
220b140b | 7836 | int cpu; |
220b140b PM |
7837 | |
7838 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7839 | swhash = &per_cpu(swevent_htable, cpu); |
7840 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7841 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7842 | } |
7843 | } | |
7844 | ||
0db0628d | 7845 | static void perf_event_init_cpu(int cpu) |
0793a61d | 7846 | { |
108b02cf | 7847 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7848 | |
b28ab83c | 7849 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7850 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7851 | struct swevent_hlist *hlist; |
7852 | ||
b28ab83c PZ |
7853 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7854 | WARN_ON(!hlist); | |
7855 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7856 | } |
b28ab83c | 7857 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7858 | } |
7859 | ||
c277443c | 7860 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7861 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7862 | { |
e9d2b064 PZ |
7863 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7864 | ||
7865 | WARN_ON(!irqs_disabled()); | |
7866 | ||
7867 | list_del_init(&cpuctx->rotation_list); | |
7868 | } | |
7869 | ||
108b02cf | 7870 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7871 | { |
108b02cf | 7872 | struct perf_event_context *ctx = __info; |
cdd6c482 | 7873 | struct perf_event *event, *tmp; |
0793a61d | 7874 | |
108b02cf | 7875 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7876 | |
889ff015 | 7877 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 7878 | __perf_remove_from_context(event); |
889ff015 | 7879 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 7880 | __perf_remove_from_context(event); |
0793a61d | 7881 | } |
108b02cf PZ |
7882 | |
7883 | static void perf_event_exit_cpu_context(int cpu) | |
7884 | { | |
7885 | struct perf_event_context *ctx; | |
7886 | struct pmu *pmu; | |
7887 | int idx; | |
7888 | ||
7889 | idx = srcu_read_lock(&pmus_srcu); | |
7890 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7891 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7892 | |
7893 | mutex_lock(&ctx->mutex); | |
7894 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7895 | mutex_unlock(&ctx->mutex); | |
7896 | } | |
7897 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7898 | } |
7899 | ||
cdd6c482 | 7900 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7901 | { |
b28ab83c | 7902 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7903 | |
b28ab83c PZ |
7904 | mutex_lock(&swhash->hlist_mutex); |
7905 | swevent_hlist_release(swhash); | |
7906 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 7907 | |
108b02cf | 7908 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7909 | } |
7910 | #else | |
cdd6c482 | 7911 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7912 | #endif |
7913 | ||
c277443c PZ |
7914 | static int |
7915 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7916 | { | |
7917 | int cpu; | |
7918 | ||
7919 | for_each_online_cpu(cpu) | |
7920 | perf_event_exit_cpu(cpu); | |
7921 | ||
7922 | return NOTIFY_OK; | |
7923 | } | |
7924 | ||
7925 | /* | |
7926 | * Run the perf reboot notifier at the very last possible moment so that | |
7927 | * the generic watchdog code runs as long as possible. | |
7928 | */ | |
7929 | static struct notifier_block perf_reboot_notifier = { | |
7930 | .notifier_call = perf_reboot, | |
7931 | .priority = INT_MIN, | |
7932 | }; | |
7933 | ||
0db0628d | 7934 | static int |
0793a61d TG |
7935 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
7936 | { | |
7937 | unsigned int cpu = (long)hcpu; | |
7938 | ||
4536e4d1 | 7939 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7940 | |
7941 | case CPU_UP_PREPARE: | |
5e11637e | 7942 | case CPU_DOWN_FAILED: |
cdd6c482 | 7943 | perf_event_init_cpu(cpu); |
0793a61d TG |
7944 | break; |
7945 | ||
5e11637e | 7946 | case CPU_UP_CANCELED: |
0793a61d | 7947 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7948 | perf_event_exit_cpu(cpu); |
0793a61d | 7949 | break; |
0793a61d TG |
7950 | default: |
7951 | break; | |
7952 | } | |
7953 | ||
7954 | return NOTIFY_OK; | |
7955 | } | |
7956 | ||
cdd6c482 | 7957 | void __init perf_event_init(void) |
0793a61d | 7958 | { |
3c502e7a JW |
7959 | int ret; |
7960 | ||
2e80a82a PZ |
7961 | idr_init(&pmu_idr); |
7962 | ||
220b140b | 7963 | perf_event_init_all_cpus(); |
b0a873eb | 7964 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7965 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7966 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7967 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7968 | perf_tp_register(); |
7969 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7970 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7971 | |
7972 | ret = init_hw_breakpoint(); | |
7973 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7974 | |
7975 | /* do not patch jump label more than once per second */ | |
7976 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7977 | |
7978 | /* | |
7979 | * Build time assertion that we keep the data_head at the intended | |
7980 | * location. IOW, validation we got the __reserved[] size right. | |
7981 | */ | |
7982 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
7983 | != 1024); | |
0793a61d | 7984 | } |
abe43400 PZ |
7985 | |
7986 | static int __init perf_event_sysfs_init(void) | |
7987 | { | |
7988 | struct pmu *pmu; | |
7989 | int ret; | |
7990 | ||
7991 | mutex_lock(&pmus_lock); | |
7992 | ||
7993 | ret = bus_register(&pmu_bus); | |
7994 | if (ret) | |
7995 | goto unlock; | |
7996 | ||
7997 | list_for_each_entry(pmu, &pmus, entry) { | |
7998 | if (!pmu->name || pmu->type < 0) | |
7999 | continue; | |
8000 | ||
8001 | ret = pmu_dev_alloc(pmu); | |
8002 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
8003 | } | |
8004 | pmu_bus_running = 1; | |
8005 | ret = 0; | |
8006 | ||
8007 | unlock: | |
8008 | mutex_unlock(&pmus_lock); | |
8009 | ||
8010 | return ret; | |
8011 | } | |
8012 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
8013 | |
8014 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
8015 | static struct cgroup_subsys_state * |
8016 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
8017 | { |
8018 | struct perf_cgroup *jc; | |
e5d1367f | 8019 | |
1b15d055 | 8020 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
8021 | if (!jc) |
8022 | return ERR_PTR(-ENOMEM); | |
8023 | ||
e5d1367f SE |
8024 | jc->info = alloc_percpu(struct perf_cgroup_info); |
8025 | if (!jc->info) { | |
8026 | kfree(jc); | |
8027 | return ERR_PTR(-ENOMEM); | |
8028 | } | |
8029 | ||
e5d1367f SE |
8030 | return &jc->css; |
8031 | } | |
8032 | ||
eb95419b | 8033 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 8034 | { |
eb95419b TH |
8035 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
8036 | ||
e5d1367f SE |
8037 | free_percpu(jc->info); |
8038 | kfree(jc); | |
8039 | } | |
8040 | ||
8041 | static int __perf_cgroup_move(void *info) | |
8042 | { | |
8043 | struct task_struct *task = info; | |
8044 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
8045 | return 0; | |
8046 | } | |
8047 | ||
eb95419b TH |
8048 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
8049 | struct cgroup_taskset *tset) | |
e5d1367f | 8050 | { |
bb9d97b6 TH |
8051 | struct task_struct *task; |
8052 | ||
d99c8727 | 8053 | cgroup_taskset_for_each(task, css, tset) |
bb9d97b6 | 8054 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8055 | } |
8056 | ||
eb95419b TH |
8057 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
8058 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 8059 | struct task_struct *task) |
e5d1367f SE |
8060 | { |
8061 | /* | |
8062 | * cgroup_exit() is called in the copy_process() failure path. | |
8063 | * Ignore this case since the task hasn't ran yet, this avoids | |
8064 | * trying to poke a half freed task state from generic code. | |
8065 | */ | |
8066 | if (!(task->flags & PF_EXITING)) | |
8067 | return; | |
8068 | ||
bb9d97b6 | 8069 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8070 | } |
8071 | ||
8072 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
8073 | .name = "perf_event", |
8074 | .subsys_id = perf_subsys_id, | |
92fb9748 TH |
8075 | .css_alloc = perf_cgroup_css_alloc, |
8076 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 8077 | .exit = perf_cgroup_exit, |
bb9d97b6 | 8078 | .attach = perf_cgroup_attach, |
e5d1367f SE |
8079 | }; |
8080 | #endif /* CONFIG_CGROUP_PERF */ |