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> |
85887973 | 42 | #include <linux/compat.h> |
0793a61d | 43 | |
76369139 FW |
44 | #include "internal.h" |
45 | ||
4e193bd4 TB |
46 | #include <asm/irq_regs.h> |
47 | ||
fe4b04fa | 48 | struct remote_function_call { |
e7e7ee2e IM |
49 | struct task_struct *p; |
50 | int (*func)(void *info); | |
51 | void *info; | |
52 | int ret; | |
fe4b04fa PZ |
53 | }; |
54 | ||
55 | static void remote_function(void *data) | |
56 | { | |
57 | struct remote_function_call *tfc = data; | |
58 | struct task_struct *p = tfc->p; | |
59 | ||
60 | if (p) { | |
61 | tfc->ret = -EAGAIN; | |
62 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
63 | return; | |
64 | } | |
65 | ||
66 | tfc->ret = tfc->func(tfc->info); | |
67 | } | |
68 | ||
69 | /** | |
70 | * task_function_call - call a function on the cpu on which a task runs | |
71 | * @p: the task to evaluate | |
72 | * @func: the function to be called | |
73 | * @info: the function call argument | |
74 | * | |
75 | * Calls the function @func when the task is currently running. This might | |
76 | * be on the current CPU, which just calls the function directly | |
77 | * | |
78 | * returns: @func return value, or | |
79 | * -ESRCH - when the process isn't running | |
80 | * -EAGAIN - when the process moved away | |
81 | */ | |
82 | static int | |
83 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
84 | { | |
85 | struct remote_function_call data = { | |
e7e7ee2e IM |
86 | .p = p, |
87 | .func = func, | |
88 | .info = info, | |
89 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
90 | }; |
91 | ||
92 | if (task_curr(p)) | |
93 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
94 | ||
95 | return data.ret; | |
96 | } | |
97 | ||
98 | /** | |
99 | * cpu_function_call - call a function on the cpu | |
100 | * @func: the function to be called | |
101 | * @info: the function call argument | |
102 | * | |
103 | * Calls the function @func on the remote cpu. | |
104 | * | |
105 | * returns: @func return value or -ENXIO when the cpu is offline | |
106 | */ | |
107 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
108 | { | |
109 | struct remote_function_call data = { | |
e7e7ee2e IM |
110 | .p = NULL, |
111 | .func = func, | |
112 | .info = info, | |
113 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
114 | }; |
115 | ||
116 | smp_call_function_single(cpu, remote_function, &data, 1); | |
117 | ||
118 | return data.ret; | |
119 | } | |
120 | ||
e5d1367f SE |
121 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
122 | PERF_FLAG_FD_OUTPUT |\ | |
123 | PERF_FLAG_PID_CGROUP) | |
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; | |
9ee318a7 | 149 | |
108b02cf PZ |
150 | static LIST_HEAD(pmus); |
151 | static DEFINE_MUTEX(pmus_lock); | |
152 | static struct srcu_struct pmus_srcu; | |
153 | ||
0764771d | 154 | /* |
cdd6c482 | 155 | * perf event paranoia level: |
0fbdea19 IM |
156 | * -1 - not paranoid at all |
157 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 158 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 159 | * 2 - disallow kernel profiling for unpriv |
0764771d | 160 | */ |
cdd6c482 | 161 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 162 | |
20443384 FW |
163 | /* Minimum for 512 kiB + 1 user control page */ |
164 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
165 | |
166 | /* | |
cdd6c482 | 167 | * max perf event sample rate |
df58ab24 | 168 | */ |
3cd49fd7 DH |
169 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
170 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
171 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
172 | ||
173 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
174 | ||
175 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
176 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
177 | ||
178 | static atomic_t perf_sample_allowed_ns __read_mostly = | |
179 | ATOMIC_INIT( DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100); | |
180 | ||
181 | void update_perf_cpu_limits(void) | |
182 | { | |
183 | u64 tmp = perf_sample_period_ns; | |
184 | ||
185 | tmp *= sysctl_perf_cpu_time_max_percent; | |
a4a108e8 | 186 | do_div(tmp, 100); |
3cd49fd7 DH |
187 | atomic_set(&perf_sample_allowed_ns, tmp); |
188 | } | |
163ec435 PZ |
189 | |
190 | int perf_proc_update_handler(struct ctl_table *table, int write, | |
191 | void __user *buffer, size_t *lenp, | |
192 | loff_t *ppos) | |
193 | { | |
02f98e3e | 194 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
163ec435 PZ |
195 | |
196 | if (ret || !write) | |
197 | return ret; | |
198 | ||
199 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
3cd49fd7 DH |
200 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
201 | update_perf_cpu_limits(); | |
202 | ||
203 | return 0; | |
204 | } | |
205 | ||
206 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
207 | ||
208 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
209 | void __user *buffer, size_t *lenp, | |
210 | loff_t *ppos) | |
211 | { | |
212 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
213 | ||
214 | if (ret || !write) | |
215 | return ret; | |
216 | ||
217 | update_perf_cpu_limits(); | |
163ec435 PZ |
218 | |
219 | return 0; | |
220 | } | |
1ccd1549 | 221 | |
3cd49fd7 DH |
222 | /* |
223 | * perf samples are done in some very critical code paths (NMIs). | |
224 | * If they take too much CPU time, the system can lock up and not | |
225 | * get any real work done. This will drop the sample rate when | |
226 | * we detect that events are taking too long. | |
227 | */ | |
228 | #define NR_ACCUMULATED_SAMPLES 128 | |
229 | DEFINE_PER_CPU(u64, running_sample_length); | |
230 | ||
231 | void perf_sample_event_took(u64 sample_len_ns) | |
232 | { | |
233 | u64 avg_local_sample_len; | |
a4a108e8 | 234 | u64 local_samples_len; |
3cd49fd7 DH |
235 | |
236 | if (atomic_read(&perf_sample_allowed_ns) == 0) | |
237 | return; | |
238 | ||
239 | /* decay the counter by 1 average sample */ | |
240 | local_samples_len = __get_cpu_var(running_sample_length); | |
241 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; | |
242 | local_samples_len += sample_len_ns; | |
243 | __get_cpu_var(running_sample_length) = local_samples_len; | |
244 | ||
245 | /* | |
246 | * note: this will be biased artifically low until we have | |
247 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
248 | * from having to maintain a count. | |
249 | */ | |
250 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
251 | ||
252 | if (avg_local_sample_len <= atomic_read(&perf_sample_allowed_ns)) | |
253 | return; | |
254 | ||
255 | if (max_samples_per_tick <= 1) | |
256 | return; | |
257 | ||
258 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
259 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
260 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
261 | ||
262 | printk_ratelimited(KERN_WARNING | |
263 | "perf samples too long (%lld > %d), lowering " | |
264 | "kernel.perf_event_max_sample_rate to %d\n", | |
265 | avg_local_sample_len, | |
266 | atomic_read(&perf_sample_allowed_ns), | |
267 | sysctl_perf_event_sample_rate); | |
268 | ||
269 | update_perf_cpu_limits(); | |
270 | } | |
271 | ||
cdd6c482 | 272 | static atomic64_t perf_event_id; |
a96bbc16 | 273 | |
0b3fcf17 SE |
274 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
275 | enum event_type_t event_type); | |
276 | ||
277 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
278 | enum event_type_t event_type, |
279 | struct task_struct *task); | |
280 | ||
281 | static void update_context_time(struct perf_event_context *ctx); | |
282 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 283 | |
cdd6c482 | 284 | void __weak perf_event_print_debug(void) { } |
0793a61d | 285 | |
84c79910 | 286 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 287 | { |
84c79910 | 288 | return "pmu"; |
0793a61d TG |
289 | } |
290 | ||
0b3fcf17 SE |
291 | static inline u64 perf_clock(void) |
292 | { | |
293 | return local_clock(); | |
294 | } | |
295 | ||
e5d1367f SE |
296 | static inline struct perf_cpu_context * |
297 | __get_cpu_context(struct perf_event_context *ctx) | |
298 | { | |
299 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
300 | } | |
301 | ||
facc4307 PZ |
302 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
303 | struct perf_event_context *ctx) | |
304 | { | |
305 | raw_spin_lock(&cpuctx->ctx.lock); | |
306 | if (ctx) | |
307 | raw_spin_lock(&ctx->lock); | |
308 | } | |
309 | ||
310 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
311 | struct perf_event_context *ctx) | |
312 | { | |
313 | if (ctx) | |
314 | raw_spin_unlock(&ctx->lock); | |
315 | raw_spin_unlock(&cpuctx->ctx.lock); | |
316 | } | |
317 | ||
e5d1367f SE |
318 | #ifdef CONFIG_CGROUP_PERF |
319 | ||
877c6856 LZ |
320 | /* |
321 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
322 | * This is a per-cpu dynamically allocated data structure. | |
323 | */ | |
324 | struct perf_cgroup_info { | |
325 | u64 time; | |
326 | u64 timestamp; | |
327 | }; | |
328 | ||
329 | struct perf_cgroup { | |
330 | struct cgroup_subsys_state css; | |
86e213e1 | 331 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
332 | }; |
333 | ||
3f7cce3c SE |
334 | /* |
335 | * Must ensure cgroup is pinned (css_get) before calling | |
336 | * this function. In other words, we cannot call this function | |
337 | * if there is no cgroup event for the current CPU context. | |
338 | */ | |
e5d1367f SE |
339 | static inline struct perf_cgroup * |
340 | perf_cgroup_from_task(struct task_struct *task) | |
341 | { | |
342 | return container_of(task_subsys_state(task, perf_subsys_id), | |
343 | struct perf_cgroup, css); | |
344 | } | |
345 | ||
346 | static inline bool | |
347 | perf_cgroup_match(struct perf_event *event) | |
348 | { | |
349 | struct perf_event_context *ctx = event->ctx; | |
350 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
351 | ||
ef824fa1 TH |
352 | /* @event doesn't care about cgroup */ |
353 | if (!event->cgrp) | |
354 | return true; | |
355 | ||
356 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
357 | if (!cpuctx->cgrp) | |
358 | return false; | |
359 | ||
360 | /* | |
361 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
362 | * also enabled for all its descendant cgroups. If @cpuctx's | |
363 | * cgroup is a descendant of @event's (the test covers identity | |
364 | * case), it's a match. | |
365 | */ | |
366 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
367 | event->cgrp->css.cgroup); | |
e5d1367f SE |
368 | } |
369 | ||
9c5da09d | 370 | static inline bool perf_tryget_cgroup(struct perf_event *event) |
e5d1367f | 371 | { |
9c5da09d | 372 | return css_tryget(&event->cgrp->css); |
e5d1367f SE |
373 | } |
374 | ||
375 | static inline void perf_put_cgroup(struct perf_event *event) | |
376 | { | |
377 | css_put(&event->cgrp->css); | |
378 | } | |
379 | ||
380 | static inline void perf_detach_cgroup(struct perf_event *event) | |
381 | { | |
382 | perf_put_cgroup(event); | |
383 | event->cgrp = NULL; | |
384 | } | |
385 | ||
386 | static inline int is_cgroup_event(struct perf_event *event) | |
387 | { | |
388 | return event->cgrp != NULL; | |
389 | } | |
390 | ||
391 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
392 | { | |
393 | struct perf_cgroup_info *t; | |
394 | ||
395 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
396 | return t->time; | |
397 | } | |
398 | ||
399 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
400 | { | |
401 | struct perf_cgroup_info *info; | |
402 | u64 now; | |
403 | ||
404 | now = perf_clock(); | |
405 | ||
406 | info = this_cpu_ptr(cgrp->info); | |
407 | ||
408 | info->time += now - info->timestamp; | |
409 | info->timestamp = now; | |
410 | } | |
411 | ||
412 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
413 | { | |
414 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
415 | if (cgrp_out) | |
416 | __update_cgrp_time(cgrp_out); | |
417 | } | |
418 | ||
419 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
420 | { | |
3f7cce3c SE |
421 | struct perf_cgroup *cgrp; |
422 | ||
e5d1367f | 423 | /* |
3f7cce3c SE |
424 | * ensure we access cgroup data only when needed and |
425 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 426 | */ |
3f7cce3c | 427 | if (!is_cgroup_event(event)) |
e5d1367f SE |
428 | return; |
429 | ||
3f7cce3c SE |
430 | cgrp = perf_cgroup_from_task(current); |
431 | /* | |
432 | * Do not update time when cgroup is not active | |
433 | */ | |
434 | if (cgrp == event->cgrp) | |
435 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
436 | } |
437 | ||
438 | static inline void | |
3f7cce3c SE |
439 | perf_cgroup_set_timestamp(struct task_struct *task, |
440 | struct perf_event_context *ctx) | |
e5d1367f SE |
441 | { |
442 | struct perf_cgroup *cgrp; | |
443 | struct perf_cgroup_info *info; | |
444 | ||
3f7cce3c SE |
445 | /* |
446 | * ctx->lock held by caller | |
447 | * ensure we do not access cgroup data | |
448 | * unless we have the cgroup pinned (css_get) | |
449 | */ | |
450 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
451 | return; |
452 | ||
453 | cgrp = perf_cgroup_from_task(task); | |
454 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 455 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
456 | } |
457 | ||
458 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
459 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
460 | ||
461 | /* | |
462 | * reschedule events based on the cgroup constraint of task. | |
463 | * | |
464 | * mode SWOUT : schedule out everything | |
465 | * mode SWIN : schedule in based on cgroup for next | |
466 | */ | |
467 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
468 | { | |
469 | struct perf_cpu_context *cpuctx; | |
470 | struct pmu *pmu; | |
471 | unsigned long flags; | |
472 | ||
473 | /* | |
474 | * disable interrupts to avoid geting nr_cgroup | |
475 | * changes via __perf_event_disable(). Also | |
476 | * avoids preemption. | |
477 | */ | |
478 | local_irq_save(flags); | |
479 | ||
480 | /* | |
481 | * we reschedule only in the presence of cgroup | |
482 | * constrained events. | |
483 | */ | |
484 | rcu_read_lock(); | |
485 | ||
486 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 487 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
488 | if (cpuctx->unique_pmu != pmu) |
489 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 490 | |
e5d1367f SE |
491 | /* |
492 | * perf_cgroup_events says at least one | |
493 | * context on this CPU has cgroup events. | |
494 | * | |
495 | * ctx->nr_cgroups reports the number of cgroup | |
496 | * events for a context. | |
497 | */ | |
498 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
499 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
500 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
501 | |
502 | if (mode & PERF_CGROUP_SWOUT) { | |
503 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
504 | /* | |
505 | * must not be done before ctxswout due | |
506 | * to event_filter_match() in event_sched_out() | |
507 | */ | |
508 | cpuctx->cgrp = NULL; | |
509 | } | |
510 | ||
511 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 512 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
513 | /* |
514 | * set cgrp before ctxsw in to allow | |
515 | * event_filter_match() to not have to pass | |
516 | * task around | |
e5d1367f SE |
517 | */ |
518 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
519 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
520 | } | |
facc4307 PZ |
521 | perf_pmu_enable(cpuctx->ctx.pmu); |
522 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 523 | } |
e5d1367f SE |
524 | } |
525 | ||
526 | rcu_read_unlock(); | |
527 | ||
528 | local_irq_restore(flags); | |
529 | } | |
530 | ||
a8d757ef SE |
531 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
532 | struct task_struct *next) | |
e5d1367f | 533 | { |
a8d757ef SE |
534 | struct perf_cgroup *cgrp1; |
535 | struct perf_cgroup *cgrp2 = NULL; | |
536 | ||
537 | /* | |
538 | * we come here when we know perf_cgroup_events > 0 | |
539 | */ | |
540 | cgrp1 = perf_cgroup_from_task(task); | |
541 | ||
542 | /* | |
543 | * next is NULL when called from perf_event_enable_on_exec() | |
544 | * that will systematically cause a cgroup_switch() | |
545 | */ | |
546 | if (next) | |
547 | cgrp2 = perf_cgroup_from_task(next); | |
548 | ||
549 | /* | |
550 | * only schedule out current cgroup events if we know | |
551 | * that we are switching to a different cgroup. Otherwise, | |
552 | * do no touch the cgroup events. | |
553 | */ | |
554 | if (cgrp1 != cgrp2) | |
555 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
556 | } |
557 | ||
a8d757ef SE |
558 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
559 | struct task_struct *task) | |
e5d1367f | 560 | { |
a8d757ef SE |
561 | struct perf_cgroup *cgrp1; |
562 | struct perf_cgroup *cgrp2 = NULL; | |
563 | ||
564 | /* | |
565 | * we come here when we know perf_cgroup_events > 0 | |
566 | */ | |
567 | cgrp1 = perf_cgroup_from_task(task); | |
568 | ||
569 | /* prev can never be NULL */ | |
570 | cgrp2 = perf_cgroup_from_task(prev); | |
571 | ||
572 | /* | |
573 | * only need to schedule in cgroup events if we are changing | |
574 | * cgroup during ctxsw. Cgroup events were not scheduled | |
575 | * out of ctxsw out if that was not the case. | |
576 | */ | |
577 | if (cgrp1 != cgrp2) | |
578 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
579 | } |
580 | ||
581 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
582 | struct perf_event_attr *attr, | |
583 | struct perf_event *group_leader) | |
584 | { | |
585 | struct perf_cgroup *cgrp; | |
586 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
587 | struct fd f = fdget(fd); |
588 | int ret = 0; | |
e5d1367f | 589 | |
2903ff01 | 590 | if (!f.file) |
e5d1367f SE |
591 | return -EBADF; |
592 | ||
2903ff01 | 593 | css = cgroup_css_from_dir(f.file, perf_subsys_id); |
3db272c0 LZ |
594 | if (IS_ERR(css)) { |
595 | ret = PTR_ERR(css); | |
596 | goto out; | |
597 | } | |
e5d1367f SE |
598 | |
599 | cgrp = container_of(css, struct perf_cgroup, css); | |
600 | event->cgrp = cgrp; | |
601 | ||
f75e18cb | 602 | /* must be done before we fput() the file */ |
9c5da09d SQ |
603 | if (!perf_tryget_cgroup(event)) { |
604 | event->cgrp = NULL; | |
605 | ret = -ENOENT; | |
606 | goto out; | |
607 | } | |
f75e18cb | 608 | |
e5d1367f SE |
609 | /* |
610 | * all events in a group must monitor | |
611 | * the same cgroup because a task belongs | |
612 | * to only one perf cgroup at a time | |
613 | */ | |
614 | if (group_leader && group_leader->cgrp != cgrp) { | |
615 | perf_detach_cgroup(event); | |
616 | ret = -EINVAL; | |
e5d1367f | 617 | } |
3db272c0 | 618 | out: |
2903ff01 | 619 | fdput(f); |
e5d1367f SE |
620 | return ret; |
621 | } | |
622 | ||
623 | static inline void | |
624 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
625 | { | |
626 | struct perf_cgroup_info *t; | |
627 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
628 | event->shadow_ctx_time = now - t->timestamp; | |
629 | } | |
630 | ||
631 | static inline void | |
632 | perf_cgroup_defer_enabled(struct perf_event *event) | |
633 | { | |
634 | /* | |
635 | * when the current task's perf cgroup does not match | |
636 | * the event's, we need to remember to call the | |
637 | * perf_mark_enable() function the first time a task with | |
638 | * a matching perf cgroup is scheduled in. | |
639 | */ | |
640 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
641 | event->cgrp_defer_enabled = 1; | |
642 | } | |
643 | ||
644 | static inline void | |
645 | perf_cgroup_mark_enabled(struct perf_event *event, | |
646 | struct perf_event_context *ctx) | |
647 | { | |
648 | struct perf_event *sub; | |
649 | u64 tstamp = perf_event_time(event); | |
650 | ||
651 | if (!event->cgrp_defer_enabled) | |
652 | return; | |
653 | ||
654 | event->cgrp_defer_enabled = 0; | |
655 | ||
656 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
657 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
658 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
659 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
660 | sub->cgrp_defer_enabled = 0; | |
661 | } | |
662 | } | |
663 | } | |
664 | #else /* !CONFIG_CGROUP_PERF */ | |
665 | ||
666 | static inline bool | |
667 | perf_cgroup_match(struct perf_event *event) | |
668 | { | |
669 | return true; | |
670 | } | |
671 | ||
672 | static inline void perf_detach_cgroup(struct perf_event *event) | |
673 | {} | |
674 | ||
675 | static inline int is_cgroup_event(struct perf_event *event) | |
676 | { | |
677 | return 0; | |
678 | } | |
679 | ||
680 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
681 | { | |
682 | return 0; | |
683 | } | |
684 | ||
685 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
686 | { | |
687 | } | |
688 | ||
689 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
690 | { | |
691 | } | |
692 | ||
a8d757ef SE |
693 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
694 | struct task_struct *next) | |
e5d1367f SE |
695 | { |
696 | } | |
697 | ||
a8d757ef SE |
698 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
699 | struct task_struct *task) | |
e5d1367f SE |
700 | { |
701 | } | |
702 | ||
703 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
704 | struct perf_event_attr *attr, | |
705 | struct perf_event *group_leader) | |
706 | { | |
707 | return -EINVAL; | |
708 | } | |
709 | ||
710 | static inline void | |
3f7cce3c SE |
711 | perf_cgroup_set_timestamp(struct task_struct *task, |
712 | struct perf_event_context *ctx) | |
e5d1367f SE |
713 | { |
714 | } | |
715 | ||
716 | void | |
717 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
718 | { | |
719 | } | |
720 | ||
721 | static inline void | |
722 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
723 | { | |
724 | } | |
725 | ||
726 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
727 | { | |
728 | return 0; | |
729 | } | |
730 | ||
731 | static inline void | |
732 | perf_cgroup_defer_enabled(struct perf_event *event) | |
733 | { | |
734 | } | |
735 | ||
736 | static inline void | |
737 | perf_cgroup_mark_enabled(struct perf_event *event, | |
738 | struct perf_event_context *ctx) | |
739 | { | |
740 | } | |
741 | #endif | |
742 | ||
33696fc0 | 743 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 744 | { |
33696fc0 PZ |
745 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
746 | if (!(*count)++) | |
747 | pmu->pmu_disable(pmu); | |
9e35ad38 | 748 | } |
9e35ad38 | 749 | |
33696fc0 | 750 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 751 | { |
33696fc0 PZ |
752 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
753 | if (!--(*count)) | |
754 | pmu->pmu_enable(pmu); | |
9e35ad38 | 755 | } |
9e35ad38 | 756 | |
e9d2b064 PZ |
757 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
758 | ||
759 | /* | |
760 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
761 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
762 | * disabled, while rotate_context is called from IRQ context. | |
763 | */ | |
108b02cf | 764 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 765 | { |
108b02cf | 766 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 767 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 768 | |
e9d2b064 | 769 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 770 | |
12351ef8 FW |
771 | if (list_empty(&cpuctx->rotation_list)) { |
772 | int was_empty = list_empty(head); | |
e9d2b064 | 773 | list_add(&cpuctx->rotation_list, head); |
12351ef8 FW |
774 | if (was_empty) |
775 | tick_nohz_full_kick(); | |
776 | } | |
9e35ad38 | 777 | } |
9e35ad38 | 778 | |
cdd6c482 | 779 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 780 | { |
e5289d4a | 781 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
782 | } |
783 | ||
cdd6c482 | 784 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 785 | { |
564c2b21 PM |
786 | if (atomic_dec_and_test(&ctx->refcount)) { |
787 | if (ctx->parent_ctx) | |
788 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
789 | if (ctx->task) |
790 | put_task_struct(ctx->task); | |
cb796ff3 | 791 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 792 | } |
a63eaf34 PM |
793 | } |
794 | ||
cdd6c482 | 795 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
796 | { |
797 | if (ctx->parent_ctx) { | |
798 | put_ctx(ctx->parent_ctx); | |
799 | ctx->parent_ctx = NULL; | |
800 | } | |
801 | } | |
802 | ||
6844c09d ACM |
803 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
804 | { | |
805 | /* | |
806 | * only top level events have the pid namespace they were created in | |
807 | */ | |
808 | if (event->parent) | |
809 | event = event->parent; | |
810 | ||
811 | return task_tgid_nr_ns(p, event->ns); | |
812 | } | |
813 | ||
814 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
815 | { | |
816 | /* | |
817 | * only top level events have the pid namespace they were created in | |
818 | */ | |
819 | if (event->parent) | |
820 | event = event->parent; | |
821 | ||
822 | return task_pid_nr_ns(p, event->ns); | |
823 | } | |
824 | ||
7f453c24 | 825 | /* |
cdd6c482 | 826 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
827 | * to userspace. |
828 | */ | |
cdd6c482 | 829 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 830 | { |
cdd6c482 | 831 | u64 id = event->id; |
7f453c24 | 832 | |
cdd6c482 IM |
833 | if (event->parent) |
834 | id = event->parent->id; | |
7f453c24 PZ |
835 | |
836 | return id; | |
837 | } | |
838 | ||
25346b93 | 839 | /* |
cdd6c482 | 840 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
841 | * This has to cope with with the fact that until it is locked, |
842 | * the context could get moved to another task. | |
843 | */ | |
cdd6c482 | 844 | static struct perf_event_context * |
8dc85d54 | 845 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 846 | { |
cdd6c482 | 847 | struct perf_event_context *ctx; |
25346b93 | 848 | |
9ed6060d | 849 | retry: |
65e303d7 PZ |
850 | /* |
851 | * One of the few rules of preemptible RCU is that one cannot do | |
852 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
853 | * part of the read side critical section was preemptible -- see | |
854 | * rcu_read_unlock_special(). | |
855 | * | |
856 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
857 | * side critical section is non-preemptible. | |
858 | */ | |
859 | preempt_disable(); | |
860 | rcu_read_lock(); | |
8dc85d54 | 861 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
862 | if (ctx) { |
863 | /* | |
864 | * If this context is a clone of another, it might | |
865 | * get swapped for another underneath us by | |
cdd6c482 | 866 | * perf_event_task_sched_out, though the |
25346b93 PM |
867 | * rcu_read_lock() protects us from any context |
868 | * getting freed. Lock the context and check if it | |
869 | * got swapped before we could get the lock, and retry | |
870 | * if so. If we locked the right context, then it | |
871 | * can't get swapped on us any more. | |
872 | */ | |
e625cce1 | 873 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 874 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 875 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
65e303d7 PZ |
876 | rcu_read_unlock(); |
877 | preempt_enable(); | |
25346b93 PM |
878 | goto retry; |
879 | } | |
b49a9e7e PZ |
880 | |
881 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 882 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
883 | ctx = NULL; |
884 | } | |
25346b93 PM |
885 | } |
886 | rcu_read_unlock(); | |
65e303d7 | 887 | preempt_enable(); |
25346b93 PM |
888 | return ctx; |
889 | } | |
890 | ||
891 | /* | |
892 | * Get the context for a task and increment its pin_count so it | |
893 | * can't get swapped to another task. This also increments its | |
894 | * reference count so that the context can't get freed. | |
895 | */ | |
8dc85d54 PZ |
896 | static struct perf_event_context * |
897 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 898 | { |
cdd6c482 | 899 | struct perf_event_context *ctx; |
25346b93 PM |
900 | unsigned long flags; |
901 | ||
8dc85d54 | 902 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
903 | if (ctx) { |
904 | ++ctx->pin_count; | |
e625cce1 | 905 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
906 | } |
907 | return ctx; | |
908 | } | |
909 | ||
cdd6c482 | 910 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
911 | { |
912 | unsigned long flags; | |
913 | ||
e625cce1 | 914 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 915 | --ctx->pin_count; |
e625cce1 | 916 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
917 | } |
918 | ||
f67218c3 PZ |
919 | /* |
920 | * Update the record of the current time in a context. | |
921 | */ | |
922 | static void update_context_time(struct perf_event_context *ctx) | |
923 | { | |
924 | u64 now = perf_clock(); | |
925 | ||
926 | ctx->time += now - ctx->timestamp; | |
927 | ctx->timestamp = now; | |
928 | } | |
929 | ||
4158755d SE |
930 | static u64 perf_event_time(struct perf_event *event) |
931 | { | |
932 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
933 | |
934 | if (is_cgroup_event(event)) | |
935 | return perf_cgroup_event_time(event); | |
936 | ||
4158755d SE |
937 | return ctx ? ctx->time : 0; |
938 | } | |
939 | ||
f67218c3 PZ |
940 | /* |
941 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 942 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
943 | */ |
944 | static void update_event_times(struct perf_event *event) | |
945 | { | |
946 | struct perf_event_context *ctx = event->ctx; | |
947 | u64 run_end; | |
948 | ||
949 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
950 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
951 | return; | |
e5d1367f SE |
952 | /* |
953 | * in cgroup mode, time_enabled represents | |
954 | * the time the event was enabled AND active | |
955 | * tasks were in the monitored cgroup. This is | |
956 | * independent of the activity of the context as | |
957 | * there may be a mix of cgroup and non-cgroup events. | |
958 | * | |
959 | * That is why we treat cgroup events differently | |
960 | * here. | |
961 | */ | |
962 | if (is_cgroup_event(event)) | |
46cd6a7f | 963 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
964 | else if (ctx->is_active) |
965 | run_end = ctx->time; | |
acd1d7c1 PZ |
966 | else |
967 | run_end = event->tstamp_stopped; | |
968 | ||
969 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
970 | |
971 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
972 | run_end = event->tstamp_stopped; | |
973 | else | |
4158755d | 974 | run_end = perf_event_time(event); |
f67218c3 PZ |
975 | |
976 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 977 | |
f67218c3 PZ |
978 | } |
979 | ||
96c21a46 PZ |
980 | /* |
981 | * Update total_time_enabled and total_time_running for all events in a group. | |
982 | */ | |
983 | static void update_group_times(struct perf_event *leader) | |
984 | { | |
985 | struct perf_event *event; | |
986 | ||
987 | update_event_times(leader); | |
988 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
989 | update_event_times(event); | |
990 | } | |
991 | ||
889ff015 FW |
992 | static struct list_head * |
993 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
994 | { | |
995 | if (event->attr.pinned) | |
996 | return &ctx->pinned_groups; | |
997 | else | |
998 | return &ctx->flexible_groups; | |
999 | } | |
1000 | ||
fccc714b | 1001 | /* |
cdd6c482 | 1002 | * Add a event from the lists for its context. |
fccc714b PZ |
1003 | * Must be called with ctx->mutex and ctx->lock held. |
1004 | */ | |
04289bb9 | 1005 | static void |
cdd6c482 | 1006 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1007 | { |
8a49542c PZ |
1008 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1009 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1010 | |
1011 | /* | |
8a49542c PZ |
1012 | * If we're a stand alone event or group leader, we go to the context |
1013 | * list, group events are kept attached to the group so that | |
1014 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1015 | */ |
8a49542c | 1016 | if (event->group_leader == event) { |
889ff015 FW |
1017 | struct list_head *list; |
1018 | ||
d6f962b5 FW |
1019 | if (is_software_event(event)) |
1020 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1021 | ||
889ff015 FW |
1022 | list = ctx_group_list(event, ctx); |
1023 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1024 | } |
592903cd | 1025 | |
08309379 | 1026 | if (is_cgroup_event(event)) |
e5d1367f | 1027 | ctx->nr_cgroups++; |
e5d1367f | 1028 | |
d010b332 SE |
1029 | if (has_branch_stack(event)) |
1030 | ctx->nr_branch_stack++; | |
1031 | ||
cdd6c482 | 1032 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1033 | if (!ctx->nr_events) |
108b02cf | 1034 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1035 | ctx->nr_events++; |
1036 | if (event->attr.inherit_stat) | |
bfbd3381 | 1037 | ctx->nr_stat++; |
04289bb9 IM |
1038 | } |
1039 | ||
0231bb53 JO |
1040 | /* |
1041 | * Initialize event state based on the perf_event_attr::disabled. | |
1042 | */ | |
1043 | static inline void perf_event__state_init(struct perf_event *event) | |
1044 | { | |
1045 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1046 | PERF_EVENT_STATE_INACTIVE; | |
1047 | } | |
1048 | ||
c320c7b7 ACM |
1049 | /* |
1050 | * Called at perf_event creation and when events are attached/detached from a | |
1051 | * group. | |
1052 | */ | |
1053 | static void perf_event__read_size(struct perf_event *event) | |
1054 | { | |
1055 | int entry = sizeof(u64); /* value */ | |
1056 | int size = 0; | |
1057 | int nr = 1; | |
1058 | ||
1059 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1060 | size += sizeof(u64); | |
1061 | ||
1062 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1063 | size += sizeof(u64); | |
1064 | ||
1065 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1066 | entry += sizeof(u64); | |
1067 | ||
1068 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1069 | nr += event->group_leader->nr_siblings; | |
1070 | size += sizeof(u64); | |
1071 | } | |
1072 | ||
1073 | size += entry * nr; | |
1074 | event->read_size = size; | |
1075 | } | |
1076 | ||
1077 | static void perf_event__header_size(struct perf_event *event) | |
1078 | { | |
1079 | struct perf_sample_data *data; | |
1080 | u64 sample_type = event->attr.sample_type; | |
1081 | u16 size = 0; | |
1082 | ||
1083 | perf_event__read_size(event); | |
1084 | ||
1085 | if (sample_type & PERF_SAMPLE_IP) | |
1086 | size += sizeof(data->ip); | |
1087 | ||
6844c09d ACM |
1088 | if (sample_type & PERF_SAMPLE_ADDR) |
1089 | size += sizeof(data->addr); | |
1090 | ||
1091 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1092 | size += sizeof(data->period); | |
1093 | ||
c3feedf2 AK |
1094 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1095 | size += sizeof(data->weight); | |
1096 | ||
6844c09d ACM |
1097 | if (sample_type & PERF_SAMPLE_READ) |
1098 | size += event->read_size; | |
1099 | ||
d6be9ad6 SE |
1100 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1101 | size += sizeof(data->data_src.val); | |
1102 | ||
6844c09d ACM |
1103 | event->header_size = size; |
1104 | } | |
1105 | ||
1106 | static void perf_event__id_header_size(struct perf_event *event) | |
1107 | { | |
1108 | struct perf_sample_data *data; | |
1109 | u64 sample_type = event->attr.sample_type; | |
1110 | u16 size = 0; | |
1111 | ||
c320c7b7 ACM |
1112 | if (sample_type & PERF_SAMPLE_TID) |
1113 | size += sizeof(data->tid_entry); | |
1114 | ||
1115 | if (sample_type & PERF_SAMPLE_TIME) | |
1116 | size += sizeof(data->time); | |
1117 | ||
c320c7b7 ACM |
1118 | if (sample_type & PERF_SAMPLE_ID) |
1119 | size += sizeof(data->id); | |
1120 | ||
1121 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1122 | size += sizeof(data->stream_id); | |
1123 | ||
1124 | if (sample_type & PERF_SAMPLE_CPU) | |
1125 | size += sizeof(data->cpu_entry); | |
1126 | ||
6844c09d | 1127 | event->id_header_size = size; |
c320c7b7 ACM |
1128 | } |
1129 | ||
8a49542c PZ |
1130 | static void perf_group_attach(struct perf_event *event) |
1131 | { | |
c320c7b7 | 1132 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1133 | |
74c3337c PZ |
1134 | /* |
1135 | * We can have double attach due to group movement in perf_event_open. | |
1136 | */ | |
1137 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1138 | return; | |
1139 | ||
8a49542c PZ |
1140 | event->attach_state |= PERF_ATTACH_GROUP; |
1141 | ||
1142 | if (group_leader == event) | |
1143 | return; | |
1144 | ||
1145 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1146 | !is_software_event(event)) | |
1147 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1148 | ||
1149 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1150 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1151 | |
1152 | perf_event__header_size(group_leader); | |
1153 | ||
1154 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1155 | perf_event__header_size(pos); | |
8a49542c PZ |
1156 | } |
1157 | ||
a63eaf34 | 1158 | /* |
cdd6c482 | 1159 | * Remove a event from the lists for its context. |
fccc714b | 1160 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1161 | */ |
04289bb9 | 1162 | static void |
cdd6c482 | 1163 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1164 | { |
68cacd29 | 1165 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1166 | /* |
1167 | * We can have double detach due to exit/hot-unplug + close. | |
1168 | */ | |
1169 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1170 | return; |
8a49542c PZ |
1171 | |
1172 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1173 | ||
68cacd29 | 1174 | if (is_cgroup_event(event)) { |
e5d1367f | 1175 | ctx->nr_cgroups--; |
68cacd29 SE |
1176 | cpuctx = __get_cpu_context(ctx); |
1177 | /* | |
1178 | * if there are no more cgroup events | |
1179 | * then cler cgrp to avoid stale pointer | |
1180 | * in update_cgrp_time_from_cpuctx() | |
1181 | */ | |
1182 | if (!ctx->nr_cgroups) | |
1183 | cpuctx->cgrp = NULL; | |
1184 | } | |
e5d1367f | 1185 | |
d010b332 SE |
1186 | if (has_branch_stack(event)) |
1187 | ctx->nr_branch_stack--; | |
1188 | ||
cdd6c482 IM |
1189 | ctx->nr_events--; |
1190 | if (event->attr.inherit_stat) | |
bfbd3381 | 1191 | ctx->nr_stat--; |
8bc20959 | 1192 | |
cdd6c482 | 1193 | list_del_rcu(&event->event_entry); |
04289bb9 | 1194 | |
8a49542c PZ |
1195 | if (event->group_leader == event) |
1196 | list_del_init(&event->group_entry); | |
5c148194 | 1197 | |
96c21a46 | 1198 | update_group_times(event); |
b2e74a26 SE |
1199 | |
1200 | /* | |
1201 | * If event was in error state, then keep it | |
1202 | * that way, otherwise bogus counts will be | |
1203 | * returned on read(). The only way to get out | |
1204 | * of error state is by explicit re-enabling | |
1205 | * of the event | |
1206 | */ | |
1207 | if (event->state > PERF_EVENT_STATE_OFF) | |
1208 | event->state = PERF_EVENT_STATE_OFF; | |
050735b0 PZ |
1209 | } |
1210 | ||
8a49542c | 1211 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1212 | { |
1213 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1214 | struct list_head *list = NULL; |
1215 | ||
1216 | /* | |
1217 | * We can have double detach due to exit/hot-unplug + close. | |
1218 | */ | |
1219 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1220 | return; | |
1221 | ||
1222 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1223 | ||
1224 | /* | |
1225 | * If this is a sibling, remove it from its group. | |
1226 | */ | |
1227 | if (event->group_leader != event) { | |
1228 | list_del_init(&event->group_entry); | |
1229 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1230 | goto out; |
8a49542c PZ |
1231 | } |
1232 | ||
1233 | if (!list_empty(&event->group_entry)) | |
1234 | list = &event->group_entry; | |
2e2af50b | 1235 | |
04289bb9 | 1236 | /* |
cdd6c482 IM |
1237 | * If this was a group event with sibling events then |
1238 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1239 | * to whatever list we are on. |
4b9e9796 S |
1240 | * If this isn't on a list, make sure we still remove the sibling's |
1241 | * group_entry from this sibling_list; otherwise, when that sibling | |
1242 | * is later deallocated, it will try to remove itself from this | |
1243 | * sibling_list, which may well have been deallocated already, | |
1244 | * resulting in a use-after-free. | |
04289bb9 | 1245 | */ |
cdd6c482 | 1246 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1247 | if (list) |
1248 | list_move_tail(&sibling->group_entry, list); | |
4b9e9796 S |
1249 | else |
1250 | list_del_init(&sibling->group_entry); | |
04289bb9 | 1251 | sibling->group_leader = sibling; |
d6f962b5 FW |
1252 | |
1253 | /* Inherit group flags from the previous leader */ | |
1254 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1255 | } |
c320c7b7 ACM |
1256 | |
1257 | out: | |
1258 | perf_event__header_size(event->group_leader); | |
1259 | ||
1260 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1261 | perf_event__header_size(tmp); | |
04289bb9 IM |
1262 | } |
1263 | ||
fa66f07a SE |
1264 | static inline int |
1265 | event_filter_match(struct perf_event *event) | |
1266 | { | |
e5d1367f SE |
1267 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1268 | && perf_cgroup_match(event); | |
fa66f07a SE |
1269 | } |
1270 | ||
9ffcfa6f SE |
1271 | static void |
1272 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1273 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1274 | struct perf_event_context *ctx) |
3b6f9e5c | 1275 | { |
4158755d | 1276 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1277 | u64 delta; |
1278 | /* | |
1279 | * An event which could not be activated because of | |
1280 | * filter mismatch still needs to have its timings | |
1281 | * maintained, otherwise bogus information is return | |
1282 | * via read() for time_enabled, time_running: | |
1283 | */ | |
1284 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1285 | && !event_filter_match(event)) { | |
e5d1367f | 1286 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1287 | event->tstamp_running += delta; |
4158755d | 1288 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1289 | } |
1290 | ||
cdd6c482 | 1291 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1292 | return; |
3b6f9e5c | 1293 | |
cdd6c482 IM |
1294 | event->state = PERF_EVENT_STATE_INACTIVE; |
1295 | if (event->pending_disable) { | |
1296 | event->pending_disable = 0; | |
1297 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1298 | } |
4158755d | 1299 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1300 | event->pmu->del(event, 0); |
cdd6c482 | 1301 | event->oncpu = -1; |
3b6f9e5c | 1302 | |
cdd6c482 | 1303 | if (!is_software_event(event)) |
3b6f9e5c PM |
1304 | cpuctx->active_oncpu--; |
1305 | ctx->nr_active--; | |
0f5a2601 PZ |
1306 | if (event->attr.freq && event->attr.sample_freq) |
1307 | ctx->nr_freq--; | |
cdd6c482 | 1308 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c PM |
1309 | cpuctx->exclusive = 0; |
1310 | } | |
1311 | ||
d859e29f | 1312 | static void |
cdd6c482 | 1313 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1314 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1315 | struct perf_event_context *ctx) |
d859e29f | 1316 | { |
cdd6c482 | 1317 | struct perf_event *event; |
fa66f07a | 1318 | int state = group_event->state; |
d859e29f | 1319 | |
cdd6c482 | 1320 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1321 | |
1322 | /* | |
1323 | * Schedule out siblings (if any): | |
1324 | */ | |
cdd6c482 IM |
1325 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1326 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1327 | |
fa66f07a | 1328 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1329 | cpuctx->exclusive = 0; |
1330 | } | |
1331 | ||
54b3f8df PZ |
1332 | struct remove_event { |
1333 | struct perf_event *event; | |
1334 | bool detach_group; | |
1335 | }; | |
1336 | ||
0793a61d | 1337 | /* |
cdd6c482 | 1338 | * Cross CPU call to remove a performance event |
0793a61d | 1339 | * |
cdd6c482 | 1340 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1341 | * remove it from the context list. |
1342 | */ | |
fe4b04fa | 1343 | static int __perf_remove_from_context(void *info) |
0793a61d | 1344 | { |
54b3f8df PZ |
1345 | struct remove_event *re = info; |
1346 | struct perf_event *event = re->event; | |
cdd6c482 | 1347 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1348 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1349 | |
e625cce1 | 1350 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1351 | event_sched_out(event, cpuctx, ctx); |
54b3f8df PZ |
1352 | if (re->detach_group) |
1353 | perf_group_detach(event); | |
cdd6c482 | 1354 | list_del_event(event, ctx); |
64ce3126 PZ |
1355 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1356 | ctx->is_active = 0; | |
1357 | cpuctx->task_ctx = NULL; | |
1358 | } | |
e625cce1 | 1359 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1360 | |
1361 | return 0; | |
0793a61d TG |
1362 | } |
1363 | ||
1364 | ||
1365 | /* | |
cdd6c482 | 1366 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1367 | * |
cdd6c482 | 1368 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1369 | * call when the task is on a CPU. |
c93f7669 | 1370 | * |
cdd6c482 IM |
1371 | * If event->ctx is a cloned context, callers must make sure that |
1372 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1373 | * remains valid. This is OK when called from perf_release since |
1374 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1375 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1376 | * context has been detached from its task. |
0793a61d | 1377 | */ |
54b3f8df | 1378 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
0793a61d | 1379 | { |
cdd6c482 | 1380 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 1381 | struct task_struct *task = ctx->task; |
54b3f8df PZ |
1382 | struct remove_event re = { |
1383 | .event = event, | |
1384 | .detach_group = detach_group, | |
1385 | }; | |
0793a61d | 1386 | |
fe4b04fa PZ |
1387 | lockdep_assert_held(&ctx->mutex); |
1388 | ||
0793a61d TG |
1389 | if (!task) { |
1390 | /* | |
cdd6c482 | 1391 | * Per cpu events are removed via an smp call and |
af901ca1 | 1392 | * the removal is always successful. |
0793a61d | 1393 | */ |
54b3f8df | 1394 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
0793a61d TG |
1395 | return; |
1396 | } | |
1397 | ||
1398 | retry: | |
54b3f8df | 1399 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
fe4b04fa | 1400 | return; |
0793a61d | 1401 | |
e625cce1 | 1402 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1403 | /* |
fe4b04fa PZ |
1404 | * If we failed to find a running task, but find the context active now |
1405 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1406 | */ |
fe4b04fa | 1407 | if (ctx->is_active) { |
e625cce1 | 1408 | raw_spin_unlock_irq(&ctx->lock); |
a63607b9 CW |
1409 | /* |
1410 | * Reload the task pointer, it might have been changed by | |
1411 | * a concurrent perf_event_context_sched_out(). | |
1412 | */ | |
1413 | task = ctx->task; | |
0793a61d TG |
1414 | goto retry; |
1415 | } | |
1416 | ||
1417 | /* | |
fe4b04fa PZ |
1418 | * Since the task isn't running, its safe to remove the event, us |
1419 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1420 | */ |
54b3f8df PZ |
1421 | if (detach_group) |
1422 | perf_group_detach(event); | |
fe4b04fa | 1423 | list_del_event(event, ctx); |
e625cce1 | 1424 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1425 | } |
1426 | ||
d859e29f | 1427 | /* |
cdd6c482 | 1428 | * Cross CPU call to disable a performance event |
d859e29f | 1429 | */ |
500ad2d8 | 1430 | int __perf_event_disable(void *info) |
d859e29f | 1431 | { |
cdd6c482 | 1432 | struct perf_event *event = info; |
cdd6c482 | 1433 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1434 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1435 | |
1436 | /* | |
cdd6c482 IM |
1437 | * If this is a per-task event, need to check whether this |
1438 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1439 | * |
1440 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1441 | * flipping contexts around. | |
d859e29f | 1442 | */ |
665c2142 | 1443 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1444 | return -EINVAL; |
d859e29f | 1445 | |
e625cce1 | 1446 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1447 | |
1448 | /* | |
cdd6c482 | 1449 | * If the event is on, turn it off. |
d859e29f PM |
1450 | * If it is in error state, leave it in error state. |
1451 | */ | |
cdd6c482 | 1452 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1453 | update_context_time(ctx); |
e5d1367f | 1454 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1455 | update_group_times(event); |
1456 | if (event == event->group_leader) | |
1457 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1458 | else |
cdd6c482 IM |
1459 | event_sched_out(event, cpuctx, ctx); |
1460 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1461 | } |
1462 | ||
e625cce1 | 1463 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1464 | |
1465 | return 0; | |
d859e29f PM |
1466 | } |
1467 | ||
1468 | /* | |
cdd6c482 | 1469 | * Disable a event. |
c93f7669 | 1470 | * |
cdd6c482 IM |
1471 | * If event->ctx is a cloned context, callers must make sure that |
1472 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1473 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1474 | * perf_event_for_each_child or perf_event_for_each because they |
1475 | * hold the top-level event's child_mutex, so any descendant that | |
1476 | * goes to exit will block in sync_child_event. | |
1477 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1478 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1479 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1480 | */ |
44234adc | 1481 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1482 | { |
cdd6c482 | 1483 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1484 | struct task_struct *task = ctx->task; |
1485 | ||
1486 | if (!task) { | |
1487 | /* | |
cdd6c482 | 1488 | * Disable the event on the cpu that it's on |
d859e29f | 1489 | */ |
fe4b04fa | 1490 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1491 | return; |
1492 | } | |
1493 | ||
9ed6060d | 1494 | retry: |
fe4b04fa PZ |
1495 | if (!task_function_call(task, __perf_event_disable, event)) |
1496 | return; | |
d859e29f | 1497 | |
e625cce1 | 1498 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1499 | /* |
cdd6c482 | 1500 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1501 | */ |
cdd6c482 | 1502 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1503 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1504 | /* |
1505 | * Reload the task pointer, it might have been changed by | |
1506 | * a concurrent perf_event_context_sched_out(). | |
1507 | */ | |
1508 | task = ctx->task; | |
d859e29f PM |
1509 | goto retry; |
1510 | } | |
1511 | ||
1512 | /* | |
1513 | * Since we have the lock this context can't be scheduled | |
1514 | * in, so we can change the state safely. | |
1515 | */ | |
cdd6c482 IM |
1516 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1517 | update_group_times(event); | |
1518 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1519 | } |
e625cce1 | 1520 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1521 | } |
dcfce4a0 | 1522 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1523 | |
e5d1367f SE |
1524 | static void perf_set_shadow_time(struct perf_event *event, |
1525 | struct perf_event_context *ctx, | |
1526 | u64 tstamp) | |
1527 | { | |
1528 | /* | |
1529 | * use the correct time source for the time snapshot | |
1530 | * | |
1531 | * We could get by without this by leveraging the | |
1532 | * fact that to get to this function, the caller | |
1533 | * has most likely already called update_context_time() | |
1534 | * and update_cgrp_time_xx() and thus both timestamp | |
1535 | * are identical (or very close). Given that tstamp is, | |
1536 | * already adjusted for cgroup, we could say that: | |
1537 | * tstamp - ctx->timestamp | |
1538 | * is equivalent to | |
1539 | * tstamp - cgrp->timestamp. | |
1540 | * | |
1541 | * Then, in perf_output_read(), the calculation would | |
1542 | * work with no changes because: | |
1543 | * - event is guaranteed scheduled in | |
1544 | * - no scheduled out in between | |
1545 | * - thus the timestamp would be the same | |
1546 | * | |
1547 | * But this is a bit hairy. | |
1548 | * | |
1549 | * So instead, we have an explicit cgroup call to remain | |
1550 | * within the time time source all along. We believe it | |
1551 | * is cleaner and simpler to understand. | |
1552 | */ | |
1553 | if (is_cgroup_event(event)) | |
1554 | perf_cgroup_set_shadow_time(event, tstamp); | |
1555 | else | |
1556 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1557 | } | |
1558 | ||
4fe757dd PZ |
1559 | #define MAX_INTERRUPTS (~0ULL) |
1560 | ||
1561 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1562 | ||
235c7fc7 | 1563 | static int |
9ffcfa6f | 1564 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1565 | struct perf_cpu_context *cpuctx, |
6e37738a | 1566 | struct perf_event_context *ctx) |
235c7fc7 | 1567 | { |
4158755d SE |
1568 | u64 tstamp = perf_event_time(event); |
1569 | ||
cdd6c482 | 1570 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1571 | return 0; |
1572 | ||
cdd6c482 | 1573 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1574 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1575 | |
1576 | /* | |
1577 | * Unthrottle events, since we scheduled we might have missed several | |
1578 | * ticks already, also for a heavily scheduling task there is little | |
1579 | * guarantee it'll get a tick in a timely manner. | |
1580 | */ | |
1581 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1582 | perf_log_throttle(event, 1); | |
1583 | event->hw.interrupts = 0; | |
1584 | } | |
1585 | ||
235c7fc7 IM |
1586 | /* |
1587 | * The new state must be visible before we turn it on in the hardware: | |
1588 | */ | |
1589 | smp_wmb(); | |
1590 | ||
a4eaf7f1 | 1591 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1592 | event->state = PERF_EVENT_STATE_INACTIVE; |
1593 | event->oncpu = -1; | |
235c7fc7 IM |
1594 | return -EAGAIN; |
1595 | } | |
1596 | ||
4158755d | 1597 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1598 | |
e5d1367f | 1599 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1600 | |
cdd6c482 | 1601 | if (!is_software_event(event)) |
3b6f9e5c | 1602 | cpuctx->active_oncpu++; |
235c7fc7 | 1603 | ctx->nr_active++; |
0f5a2601 PZ |
1604 | if (event->attr.freq && event->attr.sample_freq) |
1605 | ctx->nr_freq++; | |
235c7fc7 | 1606 | |
cdd6c482 | 1607 | if (event->attr.exclusive) |
3b6f9e5c PM |
1608 | cpuctx->exclusive = 1; |
1609 | ||
235c7fc7 IM |
1610 | return 0; |
1611 | } | |
1612 | ||
6751b71e | 1613 | static int |
cdd6c482 | 1614 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1615 | struct perf_cpu_context *cpuctx, |
6e37738a | 1616 | struct perf_event_context *ctx) |
6751b71e | 1617 | { |
6bde9b6c | 1618 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1619 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1620 | u64 now = ctx->time; |
1621 | bool simulate = false; | |
6751b71e | 1622 | |
cdd6c482 | 1623 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1624 | return 0; |
1625 | ||
ad5133b7 | 1626 | pmu->start_txn(pmu); |
6bde9b6c | 1627 | |
9ffcfa6f | 1628 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1629 | pmu->cancel_txn(pmu); |
6751b71e | 1630 | return -EAGAIN; |
90151c35 | 1631 | } |
6751b71e PM |
1632 | |
1633 | /* | |
1634 | * Schedule in siblings as one group (if any): | |
1635 | */ | |
cdd6c482 | 1636 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1637 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1638 | partial_group = event; |
6751b71e PM |
1639 | goto group_error; |
1640 | } | |
1641 | } | |
1642 | ||
9ffcfa6f | 1643 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1644 | return 0; |
9ffcfa6f | 1645 | |
6751b71e PM |
1646 | group_error: |
1647 | /* | |
1648 | * Groups can be scheduled in as one unit only, so undo any | |
1649 | * partial group before returning: | |
d7842da4 SE |
1650 | * The events up to the failed event are scheduled out normally, |
1651 | * tstamp_stopped will be updated. | |
1652 | * | |
1653 | * The failed events and the remaining siblings need to have | |
1654 | * their timings updated as if they had gone thru event_sched_in() | |
1655 | * and event_sched_out(). This is required to get consistent timings | |
1656 | * across the group. This also takes care of the case where the group | |
1657 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1658 | * the time the event was actually stopped, such that time delta | |
1659 | * calculation in update_event_times() is correct. | |
6751b71e | 1660 | */ |
cdd6c482 IM |
1661 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1662 | if (event == partial_group) | |
d7842da4 SE |
1663 | simulate = true; |
1664 | ||
1665 | if (simulate) { | |
1666 | event->tstamp_running += now - event->tstamp_stopped; | |
1667 | event->tstamp_stopped = now; | |
1668 | } else { | |
1669 | event_sched_out(event, cpuctx, ctx); | |
1670 | } | |
6751b71e | 1671 | } |
9ffcfa6f | 1672 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1673 | |
ad5133b7 | 1674 | pmu->cancel_txn(pmu); |
90151c35 | 1675 | |
6751b71e PM |
1676 | return -EAGAIN; |
1677 | } | |
1678 | ||
3b6f9e5c | 1679 | /* |
cdd6c482 | 1680 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1681 | */ |
cdd6c482 | 1682 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1683 | struct perf_cpu_context *cpuctx, |
1684 | int can_add_hw) | |
1685 | { | |
1686 | /* | |
cdd6c482 | 1687 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1688 | */ |
d6f962b5 | 1689 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1690 | return 1; |
1691 | /* | |
1692 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1693 | * events can go on. |
3b6f9e5c PM |
1694 | */ |
1695 | if (cpuctx->exclusive) | |
1696 | return 0; | |
1697 | /* | |
1698 | * If this group is exclusive and there are already | |
cdd6c482 | 1699 | * events on the CPU, it can't go on. |
3b6f9e5c | 1700 | */ |
cdd6c482 | 1701 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1702 | return 0; |
1703 | /* | |
1704 | * Otherwise, try to add it if all previous groups were able | |
1705 | * to go on. | |
1706 | */ | |
1707 | return can_add_hw; | |
1708 | } | |
1709 | ||
cdd6c482 IM |
1710 | static void add_event_to_ctx(struct perf_event *event, |
1711 | struct perf_event_context *ctx) | |
53cfbf59 | 1712 | { |
4158755d SE |
1713 | u64 tstamp = perf_event_time(event); |
1714 | ||
cdd6c482 | 1715 | list_add_event(event, ctx); |
8a49542c | 1716 | perf_group_attach(event); |
4158755d SE |
1717 | event->tstamp_enabled = tstamp; |
1718 | event->tstamp_running = tstamp; | |
1719 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1720 | } |
1721 | ||
2c29ef0f PZ |
1722 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1723 | static void | |
1724 | ctx_sched_in(struct perf_event_context *ctx, | |
1725 | struct perf_cpu_context *cpuctx, | |
1726 | enum event_type_t event_type, | |
1727 | struct task_struct *task); | |
fe4b04fa | 1728 | |
dce5855b PZ |
1729 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1730 | struct perf_event_context *ctx, | |
1731 | struct task_struct *task) | |
1732 | { | |
1733 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1734 | if (ctx) | |
1735 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1736 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1737 | if (ctx) | |
1738 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1739 | } | |
1740 | ||
0793a61d | 1741 | /* |
cdd6c482 | 1742 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1743 | * |
1744 | * Must be called with ctx->mutex held | |
0793a61d | 1745 | */ |
fe4b04fa | 1746 | static int __perf_install_in_context(void *info) |
0793a61d | 1747 | { |
cdd6c482 IM |
1748 | struct perf_event *event = info; |
1749 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1750 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1751 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1752 | struct task_struct *task = current; | |
1753 | ||
b58f6b0d | 1754 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1755 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1756 | |
1757 | /* | |
2c29ef0f | 1758 | * If there was an active task_ctx schedule it out. |
0793a61d | 1759 | */ |
b58f6b0d | 1760 | if (task_ctx) |
2c29ef0f | 1761 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1762 | |
1763 | /* | |
1764 | * If the context we're installing events in is not the | |
1765 | * active task_ctx, flip them. | |
1766 | */ | |
1767 | if (ctx->task && task_ctx != ctx) { | |
1768 | if (task_ctx) | |
1769 | raw_spin_unlock(&task_ctx->lock); | |
1770 | raw_spin_lock(&ctx->lock); | |
1771 | task_ctx = ctx; | |
1772 | } | |
1773 | ||
1774 | if (task_ctx) { | |
1775 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1776 | task = task_ctx->task; |
1777 | } | |
b58f6b0d | 1778 | |
2c29ef0f | 1779 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1780 | |
4af4998b | 1781 | update_context_time(ctx); |
e5d1367f SE |
1782 | /* |
1783 | * update cgrp time only if current cgrp | |
1784 | * matches event->cgrp. Must be done before | |
1785 | * calling add_event_to_ctx() | |
1786 | */ | |
1787 | update_cgrp_time_from_event(event); | |
0793a61d | 1788 | |
cdd6c482 | 1789 | add_event_to_ctx(event, ctx); |
0793a61d | 1790 | |
d859e29f | 1791 | /* |
2c29ef0f | 1792 | * Schedule everything back in |
d859e29f | 1793 | */ |
dce5855b | 1794 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1795 | |
1796 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1797 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1798 | |
1799 | return 0; | |
0793a61d TG |
1800 | } |
1801 | ||
1802 | /* | |
cdd6c482 | 1803 | * Attach a performance event to a context |
0793a61d | 1804 | * |
cdd6c482 IM |
1805 | * First we add the event to the list with the hardware enable bit |
1806 | * in event->hw_config cleared. | |
0793a61d | 1807 | * |
cdd6c482 | 1808 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1809 | * call to enable it in the task context. The task might have been |
1810 | * scheduled away, but we check this in the smp call again. | |
1811 | */ | |
1812 | static void | |
cdd6c482 IM |
1813 | perf_install_in_context(struct perf_event_context *ctx, |
1814 | struct perf_event *event, | |
0793a61d TG |
1815 | int cpu) |
1816 | { | |
1817 | struct task_struct *task = ctx->task; | |
1818 | ||
fe4b04fa PZ |
1819 | lockdep_assert_held(&ctx->mutex); |
1820 | ||
c3f00c70 | 1821 | event->ctx = ctx; |
0cda4c02 YZ |
1822 | if (event->cpu != -1) |
1823 | event->cpu = cpu; | |
c3f00c70 | 1824 | |
0793a61d TG |
1825 | if (!task) { |
1826 | /* | |
cdd6c482 | 1827 | * Per cpu events are installed via an smp call and |
af901ca1 | 1828 | * the install is always successful. |
0793a61d | 1829 | */ |
fe4b04fa | 1830 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1831 | return; |
1832 | } | |
1833 | ||
0793a61d | 1834 | retry: |
fe4b04fa PZ |
1835 | if (!task_function_call(task, __perf_install_in_context, event)) |
1836 | return; | |
0793a61d | 1837 | |
e625cce1 | 1838 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1839 | /* |
fe4b04fa PZ |
1840 | * If we failed to find a running task, but find the context active now |
1841 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1842 | */ |
fe4b04fa | 1843 | if (ctx->is_active) { |
e625cce1 | 1844 | raw_spin_unlock_irq(&ctx->lock); |
a63607b9 CW |
1845 | /* |
1846 | * Reload the task pointer, it might have been changed by | |
1847 | * a concurrent perf_event_context_sched_out(). | |
1848 | */ | |
1849 | task = ctx->task; | |
0793a61d TG |
1850 | goto retry; |
1851 | } | |
1852 | ||
1853 | /* | |
fe4b04fa PZ |
1854 | * Since the task isn't running, its safe to add the event, us holding |
1855 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1856 | */ |
fe4b04fa | 1857 | add_event_to_ctx(event, ctx); |
e625cce1 | 1858 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1859 | } |
1860 | ||
fa289bec | 1861 | /* |
cdd6c482 | 1862 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1863 | * Enabling the leader of a group effectively enables all |
1864 | * the group members that aren't explicitly disabled, so we | |
1865 | * have to update their ->tstamp_enabled also. | |
1866 | * Note: this works for group members as well as group leaders | |
1867 | * since the non-leader members' sibling_lists will be empty. | |
1868 | */ | |
1d9b482e | 1869 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1870 | { |
cdd6c482 | 1871 | struct perf_event *sub; |
4158755d | 1872 | u64 tstamp = perf_event_time(event); |
fa289bec | 1873 | |
cdd6c482 | 1874 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1875 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1876 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1877 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1878 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1879 | } |
fa289bec PM |
1880 | } |
1881 | ||
d859e29f | 1882 | /* |
cdd6c482 | 1883 | * Cross CPU call to enable a performance event |
d859e29f | 1884 | */ |
fe4b04fa | 1885 | static int __perf_event_enable(void *info) |
04289bb9 | 1886 | { |
cdd6c482 | 1887 | struct perf_event *event = info; |
cdd6c482 IM |
1888 | struct perf_event_context *ctx = event->ctx; |
1889 | struct perf_event *leader = event->group_leader; | |
108b02cf | 1890 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 1891 | int err; |
04289bb9 | 1892 | |
b2412679 JO |
1893 | /* |
1894 | * There's a time window between 'ctx->is_active' check | |
1895 | * in perf_event_enable function and this place having: | |
1896 | * - IRQs on | |
1897 | * - ctx->lock unlocked | |
1898 | * | |
1899 | * where the task could be killed and 'ctx' deactivated | |
1900 | * by perf_event_exit_task. | |
1901 | */ | |
1902 | if (!ctx->is_active) | |
fe4b04fa | 1903 | return -EINVAL; |
3cbed429 | 1904 | |
e625cce1 | 1905 | raw_spin_lock(&ctx->lock); |
4af4998b | 1906 | update_context_time(ctx); |
d859e29f | 1907 | |
cdd6c482 | 1908 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 1909 | goto unlock; |
e5d1367f SE |
1910 | |
1911 | /* | |
1912 | * set current task's cgroup time reference point | |
1913 | */ | |
3f7cce3c | 1914 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 1915 | |
1d9b482e | 1916 | __perf_event_mark_enabled(event); |
04289bb9 | 1917 | |
e5d1367f SE |
1918 | if (!event_filter_match(event)) { |
1919 | if (is_cgroup_event(event)) | |
1920 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 1921 | goto unlock; |
e5d1367f | 1922 | } |
f4c4176f | 1923 | |
04289bb9 | 1924 | /* |
cdd6c482 | 1925 | * If the event is in a group and isn't the group leader, |
d859e29f | 1926 | * then don't put it on unless the group is on. |
04289bb9 | 1927 | */ |
cdd6c482 | 1928 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 1929 | goto unlock; |
3b6f9e5c | 1930 | |
cdd6c482 | 1931 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 1932 | err = -EEXIST; |
e758a33d | 1933 | } else { |
cdd6c482 | 1934 | if (event == leader) |
6e37738a | 1935 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 1936 | else |
6e37738a | 1937 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 1938 | } |
d859e29f PM |
1939 | |
1940 | if (err) { | |
1941 | /* | |
cdd6c482 | 1942 | * If this event can't go on and it's part of a |
d859e29f PM |
1943 | * group, then the whole group has to come off. |
1944 | */ | |
cdd6c482 | 1945 | if (leader != event) |
d859e29f | 1946 | group_sched_out(leader, cpuctx, ctx); |
0d48696f | 1947 | if (leader->attr.pinned) { |
53cfbf59 | 1948 | update_group_times(leader); |
cdd6c482 | 1949 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 1950 | } |
d859e29f PM |
1951 | } |
1952 | ||
9ed6060d | 1953 | unlock: |
e625cce1 | 1954 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1955 | |
1956 | return 0; | |
d859e29f PM |
1957 | } |
1958 | ||
1959 | /* | |
cdd6c482 | 1960 | * Enable a event. |
c93f7669 | 1961 | * |
cdd6c482 IM |
1962 | * If event->ctx is a cloned context, callers must make sure that |
1963 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1964 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
1965 | * perf_event_for_each_child or perf_event_for_each as described |
1966 | * for perf_event_disable. | |
d859e29f | 1967 | */ |
44234adc | 1968 | void perf_event_enable(struct perf_event *event) |
d859e29f | 1969 | { |
cdd6c482 | 1970 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1971 | struct task_struct *task = ctx->task; |
1972 | ||
1973 | if (!task) { | |
1974 | /* | |
cdd6c482 | 1975 | * Enable the event on the cpu that it's on |
d859e29f | 1976 | */ |
fe4b04fa | 1977 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
1978 | return; |
1979 | } | |
1980 | ||
e625cce1 | 1981 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 1982 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
1983 | goto out; |
1984 | ||
1985 | /* | |
cdd6c482 IM |
1986 | * If the event is in error state, clear that first. |
1987 | * That way, if we see the event in error state below, we | |
d859e29f PM |
1988 | * know that it has gone back into error state, as distinct |
1989 | * from the task having been scheduled away before the | |
1990 | * cross-call arrived. | |
1991 | */ | |
cdd6c482 IM |
1992 | if (event->state == PERF_EVENT_STATE_ERROR) |
1993 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 1994 | |
9ed6060d | 1995 | retry: |
fe4b04fa | 1996 | if (!ctx->is_active) { |
1d9b482e | 1997 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
1998 | goto out; |
1999 | } | |
2000 | ||
e625cce1 | 2001 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2002 | |
2003 | if (!task_function_call(task, __perf_event_enable, event)) | |
2004 | return; | |
d859e29f | 2005 | |
e625cce1 | 2006 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2007 | |
2008 | /* | |
cdd6c482 | 2009 | * If the context is active and the event is still off, |
d859e29f PM |
2010 | * we need to retry the cross-call. |
2011 | */ | |
fe4b04fa PZ |
2012 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2013 | /* | |
2014 | * task could have been flipped by a concurrent | |
2015 | * perf_event_context_sched_out() | |
2016 | */ | |
2017 | task = ctx->task; | |
d859e29f | 2018 | goto retry; |
fe4b04fa | 2019 | } |
fa289bec | 2020 | |
9ed6060d | 2021 | out: |
e625cce1 | 2022 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2023 | } |
dcfce4a0 | 2024 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2025 | |
26ca5c11 | 2026 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2027 | { |
2023b359 | 2028 | /* |
cdd6c482 | 2029 | * not supported on inherited events |
2023b359 | 2030 | */ |
2e939d1d | 2031 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2032 | return -EINVAL; |
2033 | ||
cdd6c482 IM |
2034 | atomic_add(refresh, &event->event_limit); |
2035 | perf_event_enable(event); | |
2023b359 PZ |
2036 | |
2037 | return 0; | |
79f14641 | 2038 | } |
26ca5c11 | 2039 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2040 | |
5b0311e1 FW |
2041 | static void ctx_sched_out(struct perf_event_context *ctx, |
2042 | struct perf_cpu_context *cpuctx, | |
2043 | enum event_type_t event_type) | |
235c7fc7 | 2044 | { |
cdd6c482 | 2045 | struct perf_event *event; |
db24d33e | 2046 | int is_active = ctx->is_active; |
235c7fc7 | 2047 | |
db24d33e | 2048 | ctx->is_active &= ~event_type; |
cdd6c482 | 2049 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2050 | return; |
2051 | ||
4af4998b | 2052 | update_context_time(ctx); |
e5d1367f | 2053 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2054 | if (!ctx->nr_active) |
facc4307 | 2055 | return; |
5b0311e1 | 2056 | |
075e0b00 | 2057 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2058 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2059 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2060 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2061 | } |
889ff015 | 2062 | |
db24d33e | 2063 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2064 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2065 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2066 | } |
1b9a644f | 2067 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2068 | } |
2069 | ||
564c2b21 PM |
2070 | /* |
2071 | * Test whether two contexts are equivalent, i.e. whether they | |
2072 | * have both been cloned from the same version of the same context | |
cdd6c482 IM |
2073 | * and they both have the same number of enabled events. |
2074 | * If the number of enabled events is the same, then the set | |
2075 | * of enabled events should be the same, because these are both | |
2076 | * inherited contexts, therefore we can't access individual events | |
564c2b21 | 2077 | * in them directly with an fd; we can only enable/disable all |
cdd6c482 | 2078 | * events via prctl, or enable/disable all events in a family |
564c2b21 PM |
2079 | * via ioctl, which will have the same effect on both contexts. |
2080 | */ | |
cdd6c482 IM |
2081 | static int context_equiv(struct perf_event_context *ctx1, |
2082 | struct perf_event_context *ctx2) | |
564c2b21 PM |
2083 | { |
2084 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx | |
ad3a37de | 2085 | && ctx1->parent_gen == ctx2->parent_gen |
25346b93 | 2086 | && !ctx1->pin_count && !ctx2->pin_count; |
564c2b21 PM |
2087 | } |
2088 | ||
cdd6c482 IM |
2089 | static void __perf_event_sync_stat(struct perf_event *event, |
2090 | struct perf_event *next_event) | |
bfbd3381 PZ |
2091 | { |
2092 | u64 value; | |
2093 | ||
cdd6c482 | 2094 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2095 | return; |
2096 | ||
2097 | /* | |
cdd6c482 | 2098 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2099 | * because we're in the middle of a context switch and have IRQs |
2100 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2101 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2102 | * don't need to use it. |
2103 | */ | |
cdd6c482 IM |
2104 | switch (event->state) { |
2105 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2106 | event->pmu->read(event); |
2107 | /* fall-through */ | |
bfbd3381 | 2108 | |
cdd6c482 IM |
2109 | case PERF_EVENT_STATE_INACTIVE: |
2110 | update_event_times(event); | |
bfbd3381 PZ |
2111 | break; |
2112 | ||
2113 | default: | |
2114 | break; | |
2115 | } | |
2116 | ||
2117 | /* | |
cdd6c482 | 2118 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2119 | * values when we flip the contexts. |
2120 | */ | |
e7850595 PZ |
2121 | value = local64_read(&next_event->count); |
2122 | value = local64_xchg(&event->count, value); | |
2123 | local64_set(&next_event->count, value); | |
bfbd3381 | 2124 | |
cdd6c482 IM |
2125 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2126 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2127 | |
bfbd3381 | 2128 | /* |
19d2e755 | 2129 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2130 | */ |
cdd6c482 IM |
2131 | perf_event_update_userpage(event); |
2132 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2133 | } |
2134 | ||
cdd6c482 IM |
2135 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2136 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2137 | { |
cdd6c482 | 2138 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2139 | |
2140 | if (!ctx->nr_stat) | |
2141 | return; | |
2142 | ||
02ffdbc8 PZ |
2143 | update_context_time(ctx); |
2144 | ||
cdd6c482 IM |
2145 | event = list_first_entry(&ctx->event_list, |
2146 | struct perf_event, event_entry); | |
bfbd3381 | 2147 | |
cdd6c482 IM |
2148 | next_event = list_first_entry(&next_ctx->event_list, |
2149 | struct perf_event, event_entry); | |
bfbd3381 | 2150 | |
cdd6c482 IM |
2151 | while (&event->event_entry != &ctx->event_list && |
2152 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2153 | |
cdd6c482 | 2154 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2155 | |
cdd6c482 IM |
2156 | event = list_next_entry(event, event_entry); |
2157 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2158 | } |
2159 | } | |
2160 | ||
fe4b04fa PZ |
2161 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2162 | struct task_struct *next) | |
0793a61d | 2163 | { |
8dc85d54 | 2164 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 IM |
2165 | struct perf_event_context *next_ctx; |
2166 | struct perf_event_context *parent; | |
108b02cf | 2167 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2168 | int do_switch = 1; |
0793a61d | 2169 | |
108b02cf PZ |
2170 | if (likely(!ctx)) |
2171 | return; | |
10989fb2 | 2172 | |
108b02cf PZ |
2173 | cpuctx = __get_cpu_context(ctx); |
2174 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2175 | return; |
2176 | ||
c93f7669 PM |
2177 | rcu_read_lock(); |
2178 | parent = rcu_dereference(ctx->parent_ctx); | |
8dc85d54 | 2179 | next_ctx = next->perf_event_ctxp[ctxn]; |
c93f7669 PM |
2180 | if (parent && next_ctx && |
2181 | rcu_dereference(next_ctx->parent_ctx) == parent) { | |
2182 | /* | |
2183 | * Looks like the two contexts are clones, so we might be | |
2184 | * able to optimize the context switch. We lock both | |
2185 | * contexts and check that they are clones under the | |
2186 | * lock (including re-checking that neither has been | |
2187 | * uncloned in the meantime). It doesn't matter which | |
2188 | * order we take the locks because no other cpu could | |
2189 | * be trying to lock both of these tasks. | |
2190 | */ | |
e625cce1 TG |
2191 | raw_spin_lock(&ctx->lock); |
2192 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2193 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2194 | /* |
2195 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2196 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2197 | */ |
8dc85d54 PZ |
2198 | task->perf_event_ctxp[ctxn] = next_ctx; |
2199 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2200 | ctx->task = next; |
2201 | next_ctx->task = task; | |
2202 | do_switch = 0; | |
bfbd3381 | 2203 | |
cdd6c482 | 2204 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2205 | } |
e625cce1 TG |
2206 | raw_spin_unlock(&next_ctx->lock); |
2207 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2208 | } |
c93f7669 | 2209 | rcu_read_unlock(); |
564c2b21 | 2210 | |
c93f7669 | 2211 | if (do_switch) { |
facc4307 | 2212 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2213 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2214 | cpuctx->task_ctx = NULL; |
facc4307 | 2215 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2216 | } |
0793a61d TG |
2217 | } |
2218 | ||
8dc85d54 PZ |
2219 | #define for_each_task_context_nr(ctxn) \ |
2220 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2221 | ||
2222 | /* | |
2223 | * Called from scheduler to remove the events of the current task, | |
2224 | * with interrupts disabled. | |
2225 | * | |
2226 | * We stop each event and update the event value in event->count. | |
2227 | * | |
2228 | * This does not protect us against NMI, but disable() | |
2229 | * sets the disabled bit in the control field of event _before_ | |
2230 | * accessing the event control register. If a NMI hits, then it will | |
2231 | * not restart the event. | |
2232 | */ | |
ab0cce56 JO |
2233 | void __perf_event_task_sched_out(struct task_struct *task, |
2234 | struct task_struct *next) | |
8dc85d54 PZ |
2235 | { |
2236 | int ctxn; | |
2237 | ||
8dc85d54 PZ |
2238 | for_each_task_context_nr(ctxn) |
2239 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2240 | |
2241 | /* | |
2242 | * if cgroup events exist on this CPU, then we need | |
2243 | * to check if we have to switch out PMU state. | |
2244 | * cgroup event are system-wide mode only | |
2245 | */ | |
2246 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2247 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2248 | } |
2249 | ||
04dc2dbb | 2250 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2251 | { |
108b02cf | 2252 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2253 | |
a63eaf34 PM |
2254 | if (!cpuctx->task_ctx) |
2255 | return; | |
012b84da IM |
2256 | |
2257 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2258 | return; | |
2259 | ||
04dc2dbb | 2260 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2261 | cpuctx->task_ctx = NULL; |
2262 | } | |
2263 | ||
5b0311e1 FW |
2264 | /* |
2265 | * Called with IRQs disabled | |
2266 | */ | |
2267 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2268 | enum event_type_t event_type) | |
2269 | { | |
2270 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2271 | } |
2272 | ||
235c7fc7 | 2273 | static void |
5b0311e1 | 2274 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2275 | struct perf_cpu_context *cpuctx) |
0793a61d | 2276 | { |
cdd6c482 | 2277 | struct perf_event *event; |
0793a61d | 2278 | |
889ff015 FW |
2279 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2280 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2281 | continue; |
5632ab12 | 2282 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2283 | continue; |
2284 | ||
e5d1367f SE |
2285 | /* may need to reset tstamp_enabled */ |
2286 | if (is_cgroup_event(event)) | |
2287 | perf_cgroup_mark_enabled(event, ctx); | |
2288 | ||
8c9ed8e1 | 2289 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2290 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2291 | |
2292 | /* | |
2293 | * If this pinned group hasn't been scheduled, | |
2294 | * put it in error state. | |
2295 | */ | |
cdd6c482 IM |
2296 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2297 | update_group_times(event); | |
2298 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2299 | } |
3b6f9e5c | 2300 | } |
5b0311e1 FW |
2301 | } |
2302 | ||
2303 | static void | |
2304 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2305 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2306 | { |
2307 | struct perf_event *event; | |
2308 | int can_add_hw = 1; | |
3b6f9e5c | 2309 | |
889ff015 FW |
2310 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2311 | /* Ignore events in OFF or ERROR state */ | |
2312 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2313 | continue; |
04289bb9 IM |
2314 | /* |
2315 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2316 | * of events: |
04289bb9 | 2317 | */ |
5632ab12 | 2318 | if (!event_filter_match(event)) |
0793a61d TG |
2319 | continue; |
2320 | ||
e5d1367f SE |
2321 | /* may need to reset tstamp_enabled */ |
2322 | if (is_cgroup_event(event)) | |
2323 | perf_cgroup_mark_enabled(event, ctx); | |
2324 | ||
9ed6060d | 2325 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2326 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2327 | can_add_hw = 0; |
9ed6060d | 2328 | } |
0793a61d | 2329 | } |
5b0311e1 FW |
2330 | } |
2331 | ||
2332 | static void | |
2333 | ctx_sched_in(struct perf_event_context *ctx, | |
2334 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2335 | enum event_type_t event_type, |
2336 | struct task_struct *task) | |
5b0311e1 | 2337 | { |
e5d1367f | 2338 | u64 now; |
db24d33e | 2339 | int is_active = ctx->is_active; |
e5d1367f | 2340 | |
db24d33e | 2341 | ctx->is_active |= event_type; |
5b0311e1 | 2342 | if (likely(!ctx->nr_events)) |
facc4307 | 2343 | return; |
5b0311e1 | 2344 | |
e5d1367f SE |
2345 | now = perf_clock(); |
2346 | ctx->timestamp = now; | |
3f7cce3c | 2347 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2348 | /* |
2349 | * First go through the list and put on any pinned groups | |
2350 | * in order to give them the best chance of going on. | |
2351 | */ | |
db24d33e | 2352 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2353 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2354 | |
2355 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2356 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2357 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2358 | } |
2359 | ||
329c0e01 | 2360 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2361 | enum event_type_t event_type, |
2362 | struct task_struct *task) | |
329c0e01 FW |
2363 | { |
2364 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2365 | ||
e5d1367f | 2366 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2367 | } |
2368 | ||
e5d1367f SE |
2369 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2370 | struct task_struct *task) | |
235c7fc7 | 2371 | { |
108b02cf | 2372 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2373 | |
108b02cf | 2374 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2375 | if (cpuctx->task_ctx == ctx) |
2376 | return; | |
2377 | ||
facc4307 | 2378 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2379 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2380 | /* |
2381 | * We want to keep the following priority order: | |
2382 | * cpu pinned (that don't need to move), task pinned, | |
2383 | * cpu flexible, task flexible. | |
2384 | */ | |
2385 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2386 | ||
1d5f003f GN |
2387 | if (ctx->nr_events) |
2388 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2389 | |
86b47c25 GN |
2390 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2391 | ||
facc4307 PZ |
2392 | perf_pmu_enable(ctx->pmu); |
2393 | perf_ctx_unlock(cpuctx, ctx); | |
2394 | ||
b5ab4cd5 PZ |
2395 | /* |
2396 | * Since these rotations are per-cpu, we need to ensure the | |
2397 | * cpu-context we got scheduled on is actually rotating. | |
2398 | */ | |
108b02cf | 2399 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2400 | } |
2401 | ||
d010b332 SE |
2402 | /* |
2403 | * When sampling the branck stack in system-wide, it may be necessary | |
2404 | * to flush the stack on context switch. This happens when the branch | |
2405 | * stack does not tag its entries with the pid of the current task. | |
2406 | * Otherwise it becomes impossible to associate a branch entry with a | |
2407 | * task. This ambiguity is more likely to appear when the branch stack | |
2408 | * supports priv level filtering and the user sets it to monitor only | |
2409 | * at the user level (which could be a useful measurement in system-wide | |
2410 | * mode). In that case, the risk is high of having a branch stack with | |
2411 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2412 | * entries or stashing them somewhere in the PMU specific code layer. | |
2413 | * | |
2414 | * This function provides the context switch callback to the lower code | |
2415 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2416 | * with at least one active event using taken branch sampling. | |
2417 | */ | |
2418 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2419 | struct task_struct *task) | |
2420 | { | |
2421 | struct perf_cpu_context *cpuctx; | |
2422 | struct pmu *pmu; | |
2423 | unsigned long flags; | |
2424 | ||
2425 | /* no need to flush branch stack if not changing task */ | |
2426 | if (prev == task) | |
2427 | return; | |
2428 | ||
2429 | local_irq_save(flags); | |
2430 | ||
2431 | rcu_read_lock(); | |
2432 | ||
2433 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2434 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2435 | ||
2436 | /* | |
2437 | * check if the context has at least one | |
2438 | * event using PERF_SAMPLE_BRANCH_STACK | |
2439 | */ | |
2440 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2441 | && pmu->flush_branch_stack) { | |
2442 | ||
2443 | pmu = cpuctx->ctx.pmu; | |
2444 | ||
2445 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2446 | ||
2447 | perf_pmu_disable(pmu); | |
2448 | ||
2449 | pmu->flush_branch_stack(); | |
2450 | ||
2451 | perf_pmu_enable(pmu); | |
2452 | ||
2453 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2454 | } | |
2455 | } | |
2456 | ||
2457 | rcu_read_unlock(); | |
2458 | ||
2459 | local_irq_restore(flags); | |
2460 | } | |
2461 | ||
8dc85d54 PZ |
2462 | /* |
2463 | * Called from scheduler to add the events of the current task | |
2464 | * with interrupts disabled. | |
2465 | * | |
2466 | * We restore the event value and then enable it. | |
2467 | * | |
2468 | * This does not protect us against NMI, but enable() | |
2469 | * sets the enabled bit in the control field of event _before_ | |
2470 | * accessing the event control register. If a NMI hits, then it will | |
2471 | * keep the event running. | |
2472 | */ | |
ab0cce56 JO |
2473 | void __perf_event_task_sched_in(struct task_struct *prev, |
2474 | struct task_struct *task) | |
8dc85d54 PZ |
2475 | { |
2476 | struct perf_event_context *ctx; | |
2477 | int ctxn; | |
2478 | ||
2479 | for_each_task_context_nr(ctxn) { | |
2480 | ctx = task->perf_event_ctxp[ctxn]; | |
2481 | if (likely(!ctx)) | |
2482 | continue; | |
2483 | ||
e5d1367f | 2484 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2485 | } |
e5d1367f SE |
2486 | /* |
2487 | * if cgroup events exist on this CPU, then we need | |
2488 | * to check if we have to switch in PMU state. | |
2489 | * cgroup event are system-wide mode only | |
2490 | */ | |
2491 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2492 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2493 | |
2494 | /* check for system-wide branch_stack events */ | |
2495 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2496 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2497 | } |
2498 | ||
abd50713 PZ |
2499 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2500 | { | |
2501 | u64 frequency = event->attr.sample_freq; | |
2502 | u64 sec = NSEC_PER_SEC; | |
2503 | u64 divisor, dividend; | |
2504 | ||
2505 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2506 | ||
2507 | count_fls = fls64(count); | |
2508 | nsec_fls = fls64(nsec); | |
2509 | frequency_fls = fls64(frequency); | |
2510 | sec_fls = 30; | |
2511 | ||
2512 | /* | |
2513 | * We got @count in @nsec, with a target of sample_freq HZ | |
2514 | * the target period becomes: | |
2515 | * | |
2516 | * @count * 10^9 | |
2517 | * period = ------------------- | |
2518 | * @nsec * sample_freq | |
2519 | * | |
2520 | */ | |
2521 | ||
2522 | /* | |
2523 | * Reduce accuracy by one bit such that @a and @b converge | |
2524 | * to a similar magnitude. | |
2525 | */ | |
fe4b04fa | 2526 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2527 | do { \ |
2528 | if (a##_fls > b##_fls) { \ | |
2529 | a >>= 1; \ | |
2530 | a##_fls--; \ | |
2531 | } else { \ | |
2532 | b >>= 1; \ | |
2533 | b##_fls--; \ | |
2534 | } \ | |
2535 | } while (0) | |
2536 | ||
2537 | /* | |
2538 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2539 | * the other, so that finally we can do a u64/u64 division. | |
2540 | */ | |
2541 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2542 | REDUCE_FLS(nsec, frequency); | |
2543 | REDUCE_FLS(sec, count); | |
2544 | } | |
2545 | ||
2546 | if (count_fls + sec_fls > 64) { | |
2547 | divisor = nsec * frequency; | |
2548 | ||
2549 | while (count_fls + sec_fls > 64) { | |
2550 | REDUCE_FLS(count, sec); | |
2551 | divisor >>= 1; | |
2552 | } | |
2553 | ||
2554 | dividend = count * sec; | |
2555 | } else { | |
2556 | dividend = count * sec; | |
2557 | ||
2558 | while (nsec_fls + frequency_fls > 64) { | |
2559 | REDUCE_FLS(nsec, frequency); | |
2560 | dividend >>= 1; | |
2561 | } | |
2562 | ||
2563 | divisor = nsec * frequency; | |
2564 | } | |
2565 | ||
f6ab91ad PZ |
2566 | if (!divisor) |
2567 | return dividend; | |
2568 | ||
abd50713 PZ |
2569 | return div64_u64(dividend, divisor); |
2570 | } | |
2571 | ||
e050e3f0 SE |
2572 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2573 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2574 | ||
f39d47ff | 2575 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2576 | { |
cdd6c482 | 2577 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2578 | s64 period, sample_period; |
bd2b5b12 PZ |
2579 | s64 delta; |
2580 | ||
abd50713 | 2581 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2582 | |
2583 | delta = (s64)(period - hwc->sample_period); | |
2584 | delta = (delta + 7) / 8; /* low pass filter */ | |
2585 | ||
2586 | sample_period = hwc->sample_period + delta; | |
2587 | ||
2588 | if (!sample_period) | |
2589 | sample_period = 1; | |
2590 | ||
bd2b5b12 | 2591 | hwc->sample_period = sample_period; |
abd50713 | 2592 | |
e7850595 | 2593 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2594 | if (disable) |
2595 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2596 | ||
e7850595 | 2597 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2598 | |
2599 | if (disable) | |
2600 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2601 | } |
bd2b5b12 PZ |
2602 | } |
2603 | ||
e050e3f0 SE |
2604 | /* |
2605 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2606 | * events. At the same time, make sure, having freq events does not change | |
2607 | * the rate of unthrottling as that would introduce bias. | |
2608 | */ | |
2609 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2610 | int needs_unthr) | |
60db5e09 | 2611 | { |
cdd6c482 IM |
2612 | struct perf_event *event; |
2613 | struct hw_perf_event *hwc; | |
e050e3f0 | 2614 | u64 now, period = TICK_NSEC; |
abd50713 | 2615 | s64 delta; |
60db5e09 | 2616 | |
e050e3f0 SE |
2617 | /* |
2618 | * only need to iterate over all events iff: | |
2619 | * - context have events in frequency mode (needs freq adjust) | |
2620 | * - there are events to unthrottle on this cpu | |
2621 | */ | |
2622 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2623 | return; |
2624 | ||
e050e3f0 | 2625 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2626 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2627 | |
03541f8b | 2628 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2629 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2630 | continue; |
2631 | ||
5632ab12 | 2632 | if (!event_filter_match(event)) |
5d27c23d PZ |
2633 | continue; |
2634 | ||
cdd6c482 | 2635 | hwc = &event->hw; |
6a24ed6c | 2636 | |
e050e3f0 SE |
2637 | if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) { |
2638 | hwc->interrupts = 0; | |
cdd6c482 | 2639 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2640 | event->pmu->start(event, 0); |
a78ac325 PZ |
2641 | } |
2642 | ||
cdd6c482 | 2643 | if (!event->attr.freq || !event->attr.sample_freq) |
60db5e09 PZ |
2644 | continue; |
2645 | ||
e050e3f0 SE |
2646 | /* |
2647 | * stop the event and update event->count | |
2648 | */ | |
2649 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2650 | ||
e7850595 | 2651 | now = local64_read(&event->count); |
abd50713 PZ |
2652 | delta = now - hwc->freq_count_stamp; |
2653 | hwc->freq_count_stamp = now; | |
60db5e09 | 2654 | |
e050e3f0 SE |
2655 | /* |
2656 | * restart the event | |
2657 | * reload only if value has changed | |
f39d47ff SE |
2658 | * we have stopped the event so tell that |
2659 | * to perf_adjust_period() to avoid stopping it | |
2660 | * twice. | |
e050e3f0 | 2661 | */ |
abd50713 | 2662 | if (delta > 0) |
f39d47ff | 2663 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2664 | |
2665 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
60db5e09 | 2666 | } |
e050e3f0 | 2667 | |
f39d47ff | 2668 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2669 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2670 | } |
2671 | ||
235c7fc7 | 2672 | /* |
cdd6c482 | 2673 | * Round-robin a context's events: |
235c7fc7 | 2674 | */ |
cdd6c482 | 2675 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2676 | { |
dddd3379 TG |
2677 | /* |
2678 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2679 | * disabled by the inheritance code. | |
2680 | */ | |
2681 | if (!ctx->rotate_disable) | |
2682 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2683 | } |
2684 | ||
b5ab4cd5 | 2685 | /* |
e9d2b064 PZ |
2686 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2687 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2688 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2689 | */ |
e9d2b064 | 2690 | static void perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2691 | { |
8dc85d54 | 2692 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2693 | int rotate = 0, remove = 1; |
7fc23a53 | 2694 | |
b5ab4cd5 | 2695 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2696 | remove = 0; |
b5ab4cd5 PZ |
2697 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2698 | rotate = 1; | |
2699 | } | |
235c7fc7 | 2700 | |
8dc85d54 | 2701 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2702 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2703 | remove = 0; |
b5ab4cd5 PZ |
2704 | if (ctx->nr_events != ctx->nr_active) |
2705 | rotate = 1; | |
2706 | } | |
9717e6cd | 2707 | |
e050e3f0 | 2708 | if (!rotate) |
0f5a2601 PZ |
2709 | goto done; |
2710 | ||
facc4307 | 2711 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2712 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2713 | |
e050e3f0 SE |
2714 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2715 | if (ctx) | |
2716 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2717 | |
e050e3f0 SE |
2718 | rotate_ctx(&cpuctx->ctx); |
2719 | if (ctx) | |
2720 | rotate_ctx(ctx); | |
235c7fc7 | 2721 | |
e050e3f0 | 2722 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2723 | |
0f5a2601 PZ |
2724 | perf_pmu_enable(cpuctx->ctx.pmu); |
2725 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2726 | done: |
e9d2b064 PZ |
2727 | if (remove) |
2728 | list_del_init(&cpuctx->rotation_list); | |
e9d2b064 PZ |
2729 | } |
2730 | ||
026249ef FW |
2731 | #ifdef CONFIG_NO_HZ_FULL |
2732 | bool perf_event_can_stop_tick(void) | |
2733 | { | |
2734 | if (list_empty(&__get_cpu_var(rotation_list))) | |
2735 | return true; | |
2736 | else | |
2737 | return false; | |
2738 | } | |
2739 | #endif | |
2740 | ||
e9d2b064 PZ |
2741 | void perf_event_task_tick(void) |
2742 | { | |
2743 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2744 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2745 | struct perf_event_context *ctx; |
2746 | int throttled; | |
b5ab4cd5 | 2747 | |
e9d2b064 PZ |
2748 | WARN_ON(!irqs_disabled()); |
2749 | ||
e050e3f0 SE |
2750 | __this_cpu_inc(perf_throttled_seq); |
2751 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2752 | ||
e9d2b064 | 2753 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2754 | ctx = &cpuctx->ctx; |
2755 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2756 | ||
2757 | ctx = cpuctx->task_ctx; | |
2758 | if (ctx) | |
2759 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2760 | ||
e9d2b064 PZ |
2761 | if (cpuctx->jiffies_interval == 1 || |
2762 | !(jiffies % cpuctx->jiffies_interval)) | |
2763 | perf_rotate_context(cpuctx); | |
2764 | } | |
0793a61d TG |
2765 | } |
2766 | ||
889ff015 FW |
2767 | static int event_enable_on_exec(struct perf_event *event, |
2768 | struct perf_event_context *ctx) | |
2769 | { | |
2770 | if (!event->attr.enable_on_exec) | |
2771 | return 0; | |
2772 | ||
2773 | event->attr.enable_on_exec = 0; | |
2774 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2775 | return 0; | |
2776 | ||
1d9b482e | 2777 | __perf_event_mark_enabled(event); |
889ff015 FW |
2778 | |
2779 | return 1; | |
2780 | } | |
2781 | ||
57e7986e | 2782 | /* |
cdd6c482 | 2783 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2784 | * This expects task == current. |
2785 | */ | |
8dc85d54 | 2786 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2787 | { |
cdd6c482 | 2788 | struct perf_event *event; |
57e7986e PM |
2789 | unsigned long flags; |
2790 | int enabled = 0; | |
889ff015 | 2791 | int ret; |
57e7986e PM |
2792 | |
2793 | local_irq_save(flags); | |
cdd6c482 | 2794 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2795 | goto out; |
2796 | ||
e566b76e SE |
2797 | /* |
2798 | * We must ctxsw out cgroup events to avoid conflict | |
2799 | * when invoking perf_task_event_sched_in() later on | |
2800 | * in this function. Otherwise we end up trying to | |
2801 | * ctxswin cgroup events which are already scheduled | |
2802 | * in. | |
2803 | */ | |
a8d757ef | 2804 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2805 | |
e625cce1 | 2806 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2807 | task_ctx_sched_out(ctx); |
57e7986e | 2808 | |
b79387ef | 2809 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2810 | ret = event_enable_on_exec(event, ctx); |
2811 | if (ret) | |
2812 | enabled = 1; | |
57e7986e PM |
2813 | } |
2814 | ||
2815 | /* | |
cdd6c482 | 2816 | * Unclone this context if we enabled any event. |
57e7986e | 2817 | */ |
71a851b4 PZ |
2818 | if (enabled) |
2819 | unclone_ctx(ctx); | |
57e7986e | 2820 | |
e625cce1 | 2821 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2822 | |
e566b76e SE |
2823 | /* |
2824 | * Also calls ctxswin for cgroup events, if any: | |
2825 | */ | |
e5d1367f | 2826 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2827 | out: |
57e7986e PM |
2828 | local_irq_restore(flags); |
2829 | } | |
2830 | ||
0793a61d | 2831 | /* |
cdd6c482 | 2832 | * Cross CPU call to read the hardware event |
0793a61d | 2833 | */ |
cdd6c482 | 2834 | static void __perf_event_read(void *info) |
0793a61d | 2835 | { |
cdd6c482 IM |
2836 | struct perf_event *event = info; |
2837 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2838 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2839 | |
e1ac3614 PM |
2840 | /* |
2841 | * If this is a task context, we need to check whether it is | |
2842 | * the current task context of this cpu. If not it has been | |
2843 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2844 | * event->count would have been updated to a recent sample |
2845 | * when the event was scheduled out. | |
e1ac3614 PM |
2846 | */ |
2847 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2848 | return; | |
2849 | ||
e625cce1 | 2850 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2851 | if (ctx->is_active) { |
542e72fc | 2852 | update_context_time(ctx); |
e5d1367f SE |
2853 | update_cgrp_time_from_event(event); |
2854 | } | |
cdd6c482 | 2855 | update_event_times(event); |
542e72fc PZ |
2856 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2857 | event->pmu->read(event); | |
e625cce1 | 2858 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2859 | } |
2860 | ||
b5e58793 PZ |
2861 | static inline u64 perf_event_count(struct perf_event *event) |
2862 | { | |
e7850595 | 2863 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
2864 | } |
2865 | ||
cdd6c482 | 2866 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
2867 | { |
2868 | /* | |
cdd6c482 IM |
2869 | * If event is enabled and currently active on a CPU, update the |
2870 | * value in the event structure: | |
0793a61d | 2871 | */ |
cdd6c482 IM |
2872 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2873 | smp_call_function_single(event->oncpu, | |
2874 | __perf_event_read, event, 1); | |
2875 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
2876 | struct perf_event_context *ctx = event->ctx; |
2877 | unsigned long flags; | |
2878 | ||
e625cce1 | 2879 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
2880 | /* |
2881 | * may read while context is not active | |
2882 | * (e.g., thread is blocked), in that case | |
2883 | * we cannot update context time | |
2884 | */ | |
e5d1367f | 2885 | if (ctx->is_active) { |
c530ccd9 | 2886 | update_context_time(ctx); |
e5d1367f SE |
2887 | update_cgrp_time_from_event(event); |
2888 | } | |
cdd6c482 | 2889 | update_event_times(event); |
e625cce1 | 2890 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
2891 | } |
2892 | ||
b5e58793 | 2893 | return perf_event_count(event); |
0793a61d TG |
2894 | } |
2895 | ||
a63eaf34 | 2896 | /* |
cdd6c482 | 2897 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 2898 | */ |
eb184479 | 2899 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 2900 | { |
e625cce1 | 2901 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 2902 | mutex_init(&ctx->mutex); |
889ff015 FW |
2903 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2904 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
2905 | INIT_LIST_HEAD(&ctx->event_list); |
2906 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
2907 | } |
2908 | ||
2909 | static struct perf_event_context * | |
2910 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
2911 | { | |
2912 | struct perf_event_context *ctx; | |
2913 | ||
2914 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
2915 | if (!ctx) | |
2916 | return NULL; | |
2917 | ||
2918 | __perf_event_init_context(ctx); | |
2919 | if (task) { | |
2920 | ctx->task = task; | |
2921 | get_task_struct(task); | |
0793a61d | 2922 | } |
eb184479 PZ |
2923 | ctx->pmu = pmu; |
2924 | ||
2925 | return ctx; | |
a63eaf34 PM |
2926 | } |
2927 | ||
2ebd4ffb MH |
2928 | static struct task_struct * |
2929 | find_lively_task_by_vpid(pid_t vpid) | |
2930 | { | |
2931 | struct task_struct *task; | |
2932 | int err; | |
0793a61d TG |
2933 | |
2934 | rcu_read_lock(); | |
2ebd4ffb | 2935 | if (!vpid) |
0793a61d TG |
2936 | task = current; |
2937 | else | |
2ebd4ffb | 2938 | task = find_task_by_vpid(vpid); |
0793a61d TG |
2939 | if (task) |
2940 | get_task_struct(task); | |
2941 | rcu_read_unlock(); | |
2942 | ||
2943 | if (!task) | |
2944 | return ERR_PTR(-ESRCH); | |
2945 | ||
0793a61d | 2946 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
2947 | err = -EACCES; |
2948 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
2949 | goto errout; | |
2950 | ||
2ebd4ffb MH |
2951 | return task; |
2952 | errout: | |
2953 | put_task_struct(task); | |
2954 | return ERR_PTR(err); | |
2955 | ||
2956 | } | |
2957 | ||
fe4b04fa PZ |
2958 | /* |
2959 | * Returns a matching context with refcount and pincount. | |
2960 | */ | |
108b02cf | 2961 | static struct perf_event_context * |
38a81da2 | 2962 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 2963 | { |
cdd6c482 | 2964 | struct perf_event_context *ctx; |
22a4f650 | 2965 | struct perf_cpu_context *cpuctx; |
25346b93 | 2966 | unsigned long flags; |
8dc85d54 | 2967 | int ctxn, err; |
0793a61d | 2968 | |
22a4ec72 | 2969 | if (!task) { |
cdd6c482 | 2970 | /* Must be root to operate on a CPU event: */ |
0764771d | 2971 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
2972 | return ERR_PTR(-EACCES); |
2973 | ||
0793a61d | 2974 | /* |
cdd6c482 | 2975 | * We could be clever and allow to attach a event to an |
0793a61d TG |
2976 | * offline CPU and activate it when the CPU comes up, but |
2977 | * that's for later. | |
2978 | */ | |
f6325e30 | 2979 | if (!cpu_online(cpu)) |
0793a61d TG |
2980 | return ERR_PTR(-ENODEV); |
2981 | ||
108b02cf | 2982 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 2983 | ctx = &cpuctx->ctx; |
c93f7669 | 2984 | get_ctx(ctx); |
fe4b04fa | 2985 | ++ctx->pin_count; |
0793a61d | 2986 | |
0793a61d TG |
2987 | return ctx; |
2988 | } | |
2989 | ||
8dc85d54 PZ |
2990 | err = -EINVAL; |
2991 | ctxn = pmu->task_ctx_nr; | |
2992 | if (ctxn < 0) | |
2993 | goto errout; | |
2994 | ||
9ed6060d | 2995 | retry: |
8dc85d54 | 2996 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 2997 | if (ctx) { |
71a851b4 | 2998 | unclone_ctx(ctx); |
fe4b04fa | 2999 | ++ctx->pin_count; |
e625cce1 | 3000 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 3001 | } else { |
eb184479 | 3002 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3003 | err = -ENOMEM; |
3004 | if (!ctx) | |
3005 | goto errout; | |
eb184479 | 3006 | |
dbe08d82 ON |
3007 | err = 0; |
3008 | mutex_lock(&task->perf_event_mutex); | |
3009 | /* | |
3010 | * If it has already passed perf_event_exit_task(). | |
3011 | * we must see PF_EXITING, it takes this mutex too. | |
3012 | */ | |
3013 | if (task->flags & PF_EXITING) | |
3014 | err = -ESRCH; | |
3015 | else if (task->perf_event_ctxp[ctxn]) | |
3016 | err = -EAGAIN; | |
fe4b04fa | 3017 | else { |
9137fb28 | 3018 | get_ctx(ctx); |
fe4b04fa | 3019 | ++ctx->pin_count; |
dbe08d82 | 3020 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3021 | } |
dbe08d82 ON |
3022 | mutex_unlock(&task->perf_event_mutex); |
3023 | ||
3024 | if (unlikely(err)) { | |
9137fb28 | 3025 | put_ctx(ctx); |
dbe08d82 ON |
3026 | |
3027 | if (err == -EAGAIN) | |
3028 | goto retry; | |
3029 | goto errout; | |
a63eaf34 PM |
3030 | } |
3031 | } | |
3032 | ||
0793a61d | 3033 | return ctx; |
c93f7669 | 3034 | |
9ed6060d | 3035 | errout: |
c93f7669 | 3036 | return ERR_PTR(err); |
0793a61d TG |
3037 | } |
3038 | ||
6fb2915d LZ |
3039 | static void perf_event_free_filter(struct perf_event *event); |
3040 | ||
cdd6c482 | 3041 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3042 | { |
cdd6c482 | 3043 | struct perf_event *event; |
592903cd | 3044 | |
cdd6c482 IM |
3045 | event = container_of(head, struct perf_event, rcu_head); |
3046 | if (event->ns) | |
3047 | put_pid_ns(event->ns); | |
6fb2915d | 3048 | perf_event_free_filter(event); |
cdd6c482 | 3049 | kfree(event); |
592903cd PZ |
3050 | } |
3051 | ||
9bb5d40c PZ |
3052 | static void ring_buffer_put(struct ring_buffer *rb); |
3053 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb); | |
925d519a | 3054 | |
cdd6c482 | 3055 | static void free_event(struct perf_event *event) |
f1600952 | 3056 | { |
e360adbe | 3057 | irq_work_sync(&event->pending); |
925d519a | 3058 | |
cdd6c482 | 3059 | if (!event->parent) { |
82cd6def | 3060 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 3061 | static_key_slow_dec_deferred(&perf_sched_events); |
3af9e859 | 3062 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
3063 | atomic_dec(&nr_mmap_events); |
3064 | if (event->attr.comm) | |
3065 | atomic_dec(&nr_comm_events); | |
3066 | if (event->attr.task) | |
3067 | atomic_dec(&nr_task_events); | |
927c7a9e FW |
3068 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3069 | put_callchain_buffers(); | |
08309379 PZ |
3070 | if (is_cgroup_event(event)) { |
3071 | atomic_dec(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 3072 | static_key_slow_dec_deferred(&perf_sched_events); |
08309379 | 3073 | } |
d010b332 SE |
3074 | |
3075 | if (has_branch_stack(event)) { | |
3076 | static_key_slow_dec_deferred(&perf_sched_events); | |
3077 | /* is system-wide event */ | |
9bb5d40c | 3078 | if (!(event->attach_state & PERF_ATTACH_TASK)) { |
d010b332 SE |
3079 | atomic_dec(&per_cpu(perf_branch_stack_events, |
3080 | event->cpu)); | |
9bb5d40c | 3081 | } |
d010b332 | 3082 | } |
f344011c | 3083 | } |
9ee318a7 | 3084 | |
76369139 | 3085 | if (event->rb) { |
9bb5d40c PZ |
3086 | struct ring_buffer *rb; |
3087 | ||
3088 | /* | |
3089 | * Can happen when we close an event with re-directed output. | |
3090 | * | |
3091 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3092 | * over us; possibly making our ring_buffer_put() the last. | |
3093 | */ | |
3094 | mutex_lock(&event->mmap_mutex); | |
3095 | rb = event->rb; | |
3096 | if (rb) { | |
3097 | rcu_assign_pointer(event->rb, NULL); | |
3098 | ring_buffer_detach(event, rb); | |
3099 | ring_buffer_put(rb); /* could be last */ | |
3100 | } | |
3101 | mutex_unlock(&event->mmap_mutex); | |
a4be7c27 PZ |
3102 | } |
3103 | ||
e5d1367f SE |
3104 | if (is_cgroup_event(event)) |
3105 | perf_detach_cgroup(event); | |
3106 | ||
cdd6c482 IM |
3107 | if (event->destroy) |
3108 | event->destroy(event); | |
e077df4f | 3109 | |
0c67b408 PZ |
3110 | if (event->ctx) |
3111 | put_ctx(event->ctx); | |
3112 | ||
cdd6c482 | 3113 | call_rcu(&event->rcu_head, free_event_rcu); |
f1600952 PZ |
3114 | } |
3115 | ||
a66a3052 | 3116 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 3117 | { |
cdd6c482 | 3118 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 3119 | |
ad3a37de | 3120 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
3121 | /* |
3122 | * There are two ways this annotation is useful: | |
3123 | * | |
3124 | * 1) there is a lock recursion from perf_event_exit_task | |
3125 | * see the comment there. | |
3126 | * | |
3127 | * 2) there is a lock-inversion with mmap_sem through | |
3128 | * perf_event_read_group(), which takes faults while | |
3129 | * holding ctx->mutex, however this is called after | |
3130 | * the last filedesc died, so there is no possibility | |
3131 | * to trigger the AB-BA case. | |
3132 | */ | |
3133 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
54b3f8df | 3134 | perf_remove_from_context(event, true); |
d859e29f | 3135 | mutex_unlock(&ctx->mutex); |
0793a61d | 3136 | |
cdd6c482 | 3137 | free_event(event); |
0793a61d TG |
3138 | |
3139 | return 0; | |
3140 | } | |
a66a3052 | 3141 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 3142 | |
a66a3052 PZ |
3143 | /* |
3144 | * Called when the last reference to the file is gone. | |
3145 | */ | |
a6fa941d | 3146 | static void put_event(struct perf_event *event) |
fb0459d7 | 3147 | { |
8882135b | 3148 | struct task_struct *owner; |
fb0459d7 | 3149 | |
a6fa941d AV |
3150 | if (!atomic_long_dec_and_test(&event->refcount)) |
3151 | return; | |
fb0459d7 | 3152 | |
8882135b PZ |
3153 | rcu_read_lock(); |
3154 | owner = ACCESS_ONCE(event->owner); | |
3155 | /* | |
3156 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3157 | * !owner it means the list deletion is complete and we can indeed | |
3158 | * free this event, otherwise we need to serialize on | |
3159 | * owner->perf_event_mutex. | |
3160 | */ | |
3161 | smp_read_barrier_depends(); | |
3162 | if (owner) { | |
3163 | /* | |
3164 | * Since delayed_put_task_struct() also drops the last | |
3165 | * task reference we can safely take a new reference | |
3166 | * while holding the rcu_read_lock(). | |
3167 | */ | |
3168 | get_task_struct(owner); | |
3169 | } | |
3170 | rcu_read_unlock(); | |
3171 | ||
3172 | if (owner) { | |
3173 | mutex_lock(&owner->perf_event_mutex); | |
3174 | /* | |
3175 | * We have to re-check the event->owner field, if it is cleared | |
3176 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3177 | * ensured they're done, and we can proceed with freeing the | |
3178 | * event. | |
3179 | */ | |
3180 | if (event->owner) | |
3181 | list_del_init(&event->owner_entry); | |
3182 | mutex_unlock(&owner->perf_event_mutex); | |
3183 | put_task_struct(owner); | |
3184 | } | |
3185 | ||
a6fa941d AV |
3186 | perf_event_release_kernel(event); |
3187 | } | |
3188 | ||
3189 | static int perf_release(struct inode *inode, struct file *file) | |
3190 | { | |
3191 | put_event(file->private_data); | |
3192 | return 0; | |
fb0459d7 | 3193 | } |
fb0459d7 | 3194 | |
59ed446f | 3195 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3196 | { |
cdd6c482 | 3197 | struct perf_event *child; |
e53c0994 PZ |
3198 | u64 total = 0; |
3199 | ||
59ed446f PZ |
3200 | *enabled = 0; |
3201 | *running = 0; | |
3202 | ||
6f10581a | 3203 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3204 | total += perf_event_read(event); |
59ed446f PZ |
3205 | *enabled += event->total_time_enabled + |
3206 | atomic64_read(&event->child_total_time_enabled); | |
3207 | *running += event->total_time_running + | |
3208 | atomic64_read(&event->child_total_time_running); | |
3209 | ||
3210 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3211 | total += perf_event_read(child); |
59ed446f PZ |
3212 | *enabled += child->total_time_enabled; |
3213 | *running += child->total_time_running; | |
3214 | } | |
6f10581a | 3215 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3216 | |
3217 | return total; | |
3218 | } | |
fb0459d7 | 3219 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3220 | |
cdd6c482 | 3221 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3222 | u64 read_format, char __user *buf) |
3223 | { | |
cdd6c482 | 3224 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3225 | int n = 0, size = 0, ret = -EFAULT; |
3226 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3227 | u64 values[5]; |
59ed446f | 3228 | u64 count, enabled, running; |
abf4868b | 3229 | |
6f10581a | 3230 | mutex_lock(&ctx->mutex); |
59ed446f | 3231 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3232 | |
3233 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3234 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3235 | values[n++] = enabled; | |
3236 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3237 | values[n++] = running; | |
abf4868b PZ |
3238 | values[n++] = count; |
3239 | if (read_format & PERF_FORMAT_ID) | |
3240 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3241 | |
3242 | size = n * sizeof(u64); | |
3243 | ||
3244 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3245 | goto unlock; |
3dab77fb | 3246 | |
6f10581a | 3247 | ret = size; |
3dab77fb | 3248 | |
65abc865 | 3249 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3250 | n = 0; |
3dab77fb | 3251 | |
59ed446f | 3252 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3253 | if (read_format & PERF_FORMAT_ID) |
3254 | values[n++] = primary_event_id(sub); | |
3255 | ||
3256 | size = n * sizeof(u64); | |
3257 | ||
184d3da8 | 3258 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3259 | ret = -EFAULT; |
3260 | goto unlock; | |
3261 | } | |
abf4868b PZ |
3262 | |
3263 | ret += size; | |
3dab77fb | 3264 | } |
6f10581a PZ |
3265 | unlock: |
3266 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3267 | |
abf4868b | 3268 | return ret; |
3dab77fb PZ |
3269 | } |
3270 | ||
cdd6c482 | 3271 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3272 | u64 read_format, char __user *buf) |
3273 | { | |
59ed446f | 3274 | u64 enabled, running; |
3dab77fb PZ |
3275 | u64 values[4]; |
3276 | int n = 0; | |
3277 | ||
59ed446f PZ |
3278 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3279 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3280 | values[n++] = enabled; | |
3281 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3282 | values[n++] = running; | |
3dab77fb | 3283 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3284 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3285 | |
3286 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3287 | return -EFAULT; | |
3288 | ||
3289 | return n * sizeof(u64); | |
3290 | } | |
3291 | ||
0793a61d | 3292 | /* |
cdd6c482 | 3293 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3294 | */ |
3295 | static ssize_t | |
cdd6c482 | 3296 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3297 | { |
cdd6c482 | 3298 | u64 read_format = event->attr.read_format; |
3dab77fb | 3299 | int ret; |
0793a61d | 3300 | |
3b6f9e5c | 3301 | /* |
cdd6c482 | 3302 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3303 | * error state (i.e. because it was pinned but it couldn't be |
3304 | * scheduled on to the CPU at some point). | |
3305 | */ | |
cdd6c482 | 3306 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3307 | return 0; |
3308 | ||
c320c7b7 | 3309 | if (count < event->read_size) |
3dab77fb PZ |
3310 | return -ENOSPC; |
3311 | ||
cdd6c482 | 3312 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3313 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3314 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3315 | else |
cdd6c482 | 3316 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3317 | |
3dab77fb | 3318 | return ret; |
0793a61d TG |
3319 | } |
3320 | ||
0793a61d TG |
3321 | static ssize_t |
3322 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3323 | { | |
cdd6c482 | 3324 | struct perf_event *event = file->private_data; |
0793a61d | 3325 | |
cdd6c482 | 3326 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3327 | } |
3328 | ||
3329 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3330 | { | |
cdd6c482 | 3331 | struct perf_event *event = file->private_data; |
76369139 | 3332 | struct ring_buffer *rb; |
c33a0bc4 | 3333 | unsigned int events = POLL_HUP; |
c7138f37 | 3334 | |
10c6db11 | 3335 | /* |
9bb5d40c PZ |
3336 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3337 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3338 | */ |
3339 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3340 | rb = event->rb; |
3341 | if (rb) | |
76369139 | 3342 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 PZ |
3343 | mutex_unlock(&event->mmap_mutex); |
3344 | ||
cdd6c482 | 3345 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3346 | |
0793a61d TG |
3347 | return events; |
3348 | } | |
3349 | ||
cdd6c482 | 3350 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3351 | { |
cdd6c482 | 3352 | (void)perf_event_read(event); |
e7850595 | 3353 | local64_set(&event->count, 0); |
cdd6c482 | 3354 | perf_event_update_userpage(event); |
3df5edad PZ |
3355 | } |
3356 | ||
c93f7669 | 3357 | /* |
cdd6c482 IM |
3358 | * Holding the top-level event's child_mutex means that any |
3359 | * descendant process that has inherited this event will block | |
3360 | * in sync_child_event if it goes to exit, thus satisfying the | |
3361 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3362 | */ |
cdd6c482 IM |
3363 | static void perf_event_for_each_child(struct perf_event *event, |
3364 | void (*func)(struct perf_event *)) | |
3df5edad | 3365 | { |
cdd6c482 | 3366 | struct perf_event *child; |
3df5edad | 3367 | |
cdd6c482 IM |
3368 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3369 | mutex_lock(&event->child_mutex); | |
3370 | func(event); | |
3371 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3372 | func(child); |
cdd6c482 | 3373 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3374 | } |
3375 | ||
cdd6c482 IM |
3376 | static void perf_event_for_each(struct perf_event *event, |
3377 | void (*func)(struct perf_event *)) | |
3df5edad | 3378 | { |
cdd6c482 IM |
3379 | struct perf_event_context *ctx = event->ctx; |
3380 | struct perf_event *sibling; | |
3df5edad | 3381 | |
75f937f2 PZ |
3382 | WARN_ON_ONCE(ctx->parent_ctx); |
3383 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3384 | event = event->group_leader; |
75f937f2 | 3385 | |
cdd6c482 | 3386 | perf_event_for_each_child(event, func); |
cdd6c482 | 3387 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3388 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3389 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3390 | } |
3391 | ||
cdd6c482 | 3392 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3393 | { |
cdd6c482 | 3394 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3395 | int ret = 0; |
3396 | u64 value; | |
3397 | ||
6c7e550f | 3398 | if (!is_sampling_event(event)) |
08247e31 PZ |
3399 | return -EINVAL; |
3400 | ||
ad0cf347 | 3401 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3402 | return -EFAULT; |
3403 | ||
3404 | if (!value) | |
3405 | return -EINVAL; | |
3406 | ||
e625cce1 | 3407 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3408 | if (event->attr.freq) { |
3409 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3410 | ret = -EINVAL; |
3411 | goto unlock; | |
3412 | } | |
3413 | ||
cdd6c482 | 3414 | event->attr.sample_freq = value; |
08247e31 | 3415 | } else { |
cdd6c482 IM |
3416 | event->attr.sample_period = value; |
3417 | event->hw.sample_period = value; | |
08247e31 PZ |
3418 | } |
3419 | unlock: | |
e625cce1 | 3420 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3421 | |
3422 | return ret; | |
3423 | } | |
3424 | ||
ac9721f3 PZ |
3425 | static const struct file_operations perf_fops; |
3426 | ||
2903ff01 | 3427 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3428 | { |
2903ff01 AV |
3429 | struct fd f = fdget(fd); |
3430 | if (!f.file) | |
3431 | return -EBADF; | |
ac9721f3 | 3432 | |
2903ff01 AV |
3433 | if (f.file->f_op != &perf_fops) { |
3434 | fdput(f); | |
3435 | return -EBADF; | |
ac9721f3 | 3436 | } |
2903ff01 AV |
3437 | *p = f; |
3438 | return 0; | |
ac9721f3 PZ |
3439 | } |
3440 | ||
3441 | static int perf_event_set_output(struct perf_event *event, | |
3442 | struct perf_event *output_event); | |
6fb2915d | 3443 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3444 | |
d859e29f PM |
3445 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3446 | { | |
cdd6c482 IM |
3447 | struct perf_event *event = file->private_data; |
3448 | void (*func)(struct perf_event *); | |
3df5edad | 3449 | u32 flags = arg; |
d859e29f PM |
3450 | |
3451 | switch (cmd) { | |
cdd6c482 IM |
3452 | case PERF_EVENT_IOC_ENABLE: |
3453 | func = perf_event_enable; | |
d859e29f | 3454 | break; |
cdd6c482 IM |
3455 | case PERF_EVENT_IOC_DISABLE: |
3456 | func = perf_event_disable; | |
79f14641 | 3457 | break; |
cdd6c482 IM |
3458 | case PERF_EVENT_IOC_RESET: |
3459 | func = perf_event_reset; | |
6de6a7b9 | 3460 | break; |
3df5edad | 3461 | |
cdd6c482 IM |
3462 | case PERF_EVENT_IOC_REFRESH: |
3463 | return perf_event_refresh(event, arg); | |
08247e31 | 3464 | |
cdd6c482 IM |
3465 | case PERF_EVENT_IOC_PERIOD: |
3466 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3467 | |
cdd6c482 | 3468 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3469 | { |
ac9721f3 | 3470 | int ret; |
ac9721f3 | 3471 | if (arg != -1) { |
2903ff01 AV |
3472 | struct perf_event *output_event; |
3473 | struct fd output; | |
3474 | ret = perf_fget_light(arg, &output); | |
3475 | if (ret) | |
3476 | return ret; | |
3477 | output_event = output.file->private_data; | |
3478 | ret = perf_event_set_output(event, output_event); | |
3479 | fdput(output); | |
3480 | } else { | |
3481 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3482 | } |
ac9721f3 PZ |
3483 | return ret; |
3484 | } | |
a4be7c27 | 3485 | |
6fb2915d LZ |
3486 | case PERF_EVENT_IOC_SET_FILTER: |
3487 | return perf_event_set_filter(event, (void __user *)arg); | |
3488 | ||
d859e29f | 3489 | default: |
3df5edad | 3490 | return -ENOTTY; |
d859e29f | 3491 | } |
3df5edad PZ |
3492 | |
3493 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3494 | perf_event_for_each(event, func); |
3df5edad | 3495 | else |
cdd6c482 | 3496 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3497 | |
3498 | return 0; | |
d859e29f PM |
3499 | } |
3500 | ||
85887973 PM |
3501 | #ifdef CONFIG_COMPAT |
3502 | static long perf_compat_ioctl(struct file *file, unsigned int cmd, | |
3503 | unsigned long arg) | |
3504 | { | |
3505 | switch (_IOC_NR(cmd)) { | |
3506 | case _IOC_NR(PERF_EVENT_IOC_SET_FILTER): | |
3507 | /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */ | |
3508 | if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) { | |
3509 | cmd &= ~IOCSIZE_MASK; | |
3510 | cmd |= sizeof(void *) << IOCSIZE_SHIFT; | |
3511 | } | |
3512 | break; | |
3513 | } | |
3514 | return perf_ioctl(file, cmd, arg); | |
3515 | } | |
3516 | #else | |
3517 | # define perf_compat_ioctl NULL | |
3518 | #endif | |
3519 | ||
cdd6c482 | 3520 | int perf_event_task_enable(void) |
771d7cde | 3521 | { |
cdd6c482 | 3522 | struct perf_event *event; |
771d7cde | 3523 | |
cdd6c482 IM |
3524 | mutex_lock(¤t->perf_event_mutex); |
3525 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3526 | perf_event_for_each_child(event, perf_event_enable); | |
3527 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3528 | |
3529 | return 0; | |
3530 | } | |
3531 | ||
cdd6c482 | 3532 | int perf_event_task_disable(void) |
771d7cde | 3533 | { |
cdd6c482 | 3534 | struct perf_event *event; |
771d7cde | 3535 | |
cdd6c482 IM |
3536 | mutex_lock(¤t->perf_event_mutex); |
3537 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3538 | perf_event_for_each_child(event, perf_event_disable); | |
3539 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3540 | |
3541 | return 0; | |
3542 | } | |
3543 | ||
cdd6c482 | 3544 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3545 | { |
a4eaf7f1 PZ |
3546 | if (event->hw.state & PERF_HES_STOPPED) |
3547 | return 0; | |
3548 | ||
cdd6c482 | 3549 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3550 | return 0; |
3551 | ||
35edc2a5 | 3552 | return event->pmu->event_idx(event); |
194002b2 PZ |
3553 | } |
3554 | ||
c4794295 | 3555 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3556 | u64 *now, |
7f310a5d EM |
3557 | u64 *enabled, |
3558 | u64 *running) | |
c4794295 | 3559 | { |
e3f3541c | 3560 | u64 ctx_time; |
c4794295 | 3561 | |
e3f3541c PZ |
3562 | *now = perf_clock(); |
3563 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3564 | *enabled = ctx_time - event->tstamp_enabled; |
3565 | *running = ctx_time - event->tstamp_running; | |
3566 | } | |
3567 | ||
c7206205 | 3568 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3569 | { |
3570 | } | |
3571 | ||
38ff667b PZ |
3572 | /* |
3573 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3574 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3575 | * code calls this from NMI context. | |
3576 | */ | |
cdd6c482 | 3577 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3578 | { |
cdd6c482 | 3579 | struct perf_event_mmap_page *userpg; |
76369139 | 3580 | struct ring_buffer *rb; |
e3f3541c | 3581 | u64 enabled, running, now; |
38ff667b PZ |
3582 | |
3583 | rcu_read_lock(); | |
0d641208 EM |
3584 | /* |
3585 | * compute total_time_enabled, total_time_running | |
3586 | * based on snapshot values taken when the event | |
3587 | * was last scheduled in. | |
3588 | * | |
3589 | * we cannot simply called update_context_time() | |
3590 | * because of locking issue as we can be called in | |
3591 | * NMI context | |
3592 | */ | |
e3f3541c | 3593 | calc_timer_values(event, &now, &enabled, &running); |
76369139 FW |
3594 | rb = rcu_dereference(event->rb); |
3595 | if (!rb) | |
38ff667b PZ |
3596 | goto unlock; |
3597 | ||
76369139 | 3598 | userpg = rb->user_page; |
37d81828 | 3599 | |
7b732a75 PZ |
3600 | /* |
3601 | * Disable preemption so as to not let the corresponding user-space | |
3602 | * spin too long if we get preempted. | |
3603 | */ | |
3604 | preempt_disable(); | |
37d81828 | 3605 | ++userpg->lock; |
92f22a38 | 3606 | barrier(); |
cdd6c482 | 3607 | userpg->index = perf_event_index(event); |
b5e58793 | 3608 | userpg->offset = perf_event_count(event); |
365a4038 | 3609 | if (userpg->index) |
e7850595 | 3610 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3611 | |
0d641208 | 3612 | userpg->time_enabled = enabled + |
cdd6c482 | 3613 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3614 | |
0d641208 | 3615 | userpg->time_running = running + |
cdd6c482 | 3616 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3617 | |
c7206205 | 3618 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3619 | |
92f22a38 | 3620 | barrier(); |
37d81828 | 3621 | ++userpg->lock; |
7b732a75 | 3622 | preempt_enable(); |
38ff667b | 3623 | unlock: |
7b732a75 | 3624 | rcu_read_unlock(); |
37d81828 PM |
3625 | } |
3626 | ||
906010b2 PZ |
3627 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3628 | { | |
3629 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3630 | struct ring_buffer *rb; |
906010b2 PZ |
3631 | int ret = VM_FAULT_SIGBUS; |
3632 | ||
3633 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3634 | if (vmf->pgoff == 0) | |
3635 | ret = 0; | |
3636 | return ret; | |
3637 | } | |
3638 | ||
3639 | rcu_read_lock(); | |
76369139 FW |
3640 | rb = rcu_dereference(event->rb); |
3641 | if (!rb) | |
906010b2 PZ |
3642 | goto unlock; |
3643 | ||
3644 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3645 | goto unlock; | |
3646 | ||
76369139 | 3647 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3648 | if (!vmf->page) |
3649 | goto unlock; | |
3650 | ||
3651 | get_page(vmf->page); | |
3652 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3653 | vmf->page->index = vmf->pgoff; | |
3654 | ||
3655 | ret = 0; | |
3656 | unlock: | |
3657 | rcu_read_unlock(); | |
3658 | ||
3659 | return ret; | |
3660 | } | |
3661 | ||
10c6db11 PZ |
3662 | static void ring_buffer_attach(struct perf_event *event, |
3663 | struct ring_buffer *rb) | |
3664 | { | |
3665 | unsigned long flags; | |
3666 | ||
3667 | if (!list_empty(&event->rb_entry)) | |
3668 | return; | |
3669 | ||
3670 | spin_lock_irqsave(&rb->event_lock, flags); | |
9bb5d40c PZ |
3671 | if (list_empty(&event->rb_entry)) |
3672 | list_add(&event->rb_entry, &rb->event_list); | |
10c6db11 PZ |
3673 | spin_unlock_irqrestore(&rb->event_lock, flags); |
3674 | } | |
3675 | ||
9bb5d40c | 3676 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb) |
10c6db11 PZ |
3677 | { |
3678 | unsigned long flags; | |
3679 | ||
3680 | if (list_empty(&event->rb_entry)) | |
3681 | return; | |
3682 | ||
3683 | spin_lock_irqsave(&rb->event_lock, flags); | |
3684 | list_del_init(&event->rb_entry); | |
3685 | wake_up_all(&event->waitq); | |
3686 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3687 | } | |
3688 | ||
3689 | static void ring_buffer_wakeup(struct perf_event *event) | |
3690 | { | |
3691 | struct ring_buffer *rb; | |
3692 | ||
3693 | rcu_read_lock(); | |
3694 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
3695 | if (rb) { |
3696 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
3697 | wake_up_all(&event->waitq); | |
3698 | } | |
10c6db11 PZ |
3699 | rcu_read_unlock(); |
3700 | } | |
3701 | ||
76369139 | 3702 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3703 | { |
76369139 | 3704 | struct ring_buffer *rb; |
906010b2 | 3705 | |
76369139 FW |
3706 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3707 | rb_free(rb); | |
7b732a75 PZ |
3708 | } |
3709 | ||
76369139 | 3710 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3711 | { |
76369139 | 3712 | struct ring_buffer *rb; |
7b732a75 | 3713 | |
ac9721f3 | 3714 | rcu_read_lock(); |
76369139 FW |
3715 | rb = rcu_dereference(event->rb); |
3716 | if (rb) { | |
3717 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3718 | rb = NULL; | |
ac9721f3 PZ |
3719 | } |
3720 | rcu_read_unlock(); | |
3721 | ||
76369139 | 3722 | return rb; |
ac9721f3 PZ |
3723 | } |
3724 | ||
9bb5d40c | 3725 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3726 | { |
76369139 | 3727 | if (!atomic_dec_and_test(&rb->refcount)) |
9bb5d40c | 3728 | return; |
7b732a75 | 3729 | |
9bb5d40c | 3730 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 3731 | |
76369139 | 3732 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3733 | } |
3734 | ||
3735 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3736 | { | |
cdd6c482 | 3737 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3738 | |
cdd6c482 | 3739 | atomic_inc(&event->mmap_count); |
9bb5d40c | 3740 | atomic_inc(&event->rb->mmap_count); |
7b732a75 PZ |
3741 | } |
3742 | ||
9bb5d40c PZ |
3743 | /* |
3744 | * A buffer can be mmap()ed multiple times; either directly through the same | |
3745 | * event, or through other events by use of perf_event_set_output(). | |
3746 | * | |
3747 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
3748 | * the buffer here, where we still have a VM context. This means we need | |
3749 | * to detach all events redirecting to us. | |
3750 | */ | |
7b732a75 PZ |
3751 | static void perf_mmap_close(struct vm_area_struct *vma) |
3752 | { | |
cdd6c482 | 3753 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3754 | |
9bb5d40c PZ |
3755 | struct ring_buffer *rb = event->rb; |
3756 | struct user_struct *mmap_user = rb->mmap_user; | |
3757 | int mmap_locked = rb->mmap_locked; | |
3758 | unsigned long size = perf_data_size(rb); | |
789f90fc | 3759 | |
9bb5d40c PZ |
3760 | atomic_dec(&rb->mmap_count); |
3761 | ||
3762 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
3763 | return; | |
3764 | ||
3765 | /* Detach current event from the buffer. */ | |
3766 | rcu_assign_pointer(event->rb, NULL); | |
3767 | ring_buffer_detach(event, rb); | |
3768 | mutex_unlock(&event->mmap_mutex); | |
3769 | ||
3770 | /* If there's still other mmap()s of this buffer, we're done. */ | |
3771 | if (atomic_read(&rb->mmap_count)) { | |
3772 | ring_buffer_put(rb); /* can't be last */ | |
3773 | return; | |
3774 | } | |
ac9721f3 | 3775 | |
9bb5d40c PZ |
3776 | /* |
3777 | * No other mmap()s, detach from all other events that might redirect | |
3778 | * into the now unreachable buffer. Somewhat complicated by the | |
3779 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
3780 | */ | |
3781 | again: | |
3782 | rcu_read_lock(); | |
3783 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
3784 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
3785 | /* | |
3786 | * This event is en-route to free_event() which will | |
3787 | * detach it and remove it from the list. | |
3788 | */ | |
3789 | continue; | |
3790 | } | |
3791 | rcu_read_unlock(); | |
3792 | ||
3793 | mutex_lock(&event->mmap_mutex); | |
3794 | /* | |
3795 | * Check we didn't race with perf_event_set_output() which can | |
3796 | * swizzle the rb from under us while we were waiting to | |
3797 | * acquire mmap_mutex. | |
3798 | * | |
3799 | * If we find a different rb; ignore this event, a next | |
3800 | * iteration will no longer find it on the list. We have to | |
3801 | * still restart the iteration to make sure we're not now | |
3802 | * iterating the wrong list. | |
3803 | */ | |
3804 | if (event->rb == rb) { | |
3805 | rcu_assign_pointer(event->rb, NULL); | |
3806 | ring_buffer_detach(event, rb); | |
3807 | ring_buffer_put(rb); /* can't be last, we still have one */ | |
26cb63ad | 3808 | } |
9bb5d40c PZ |
3809 | mutex_unlock(&event->mmap_mutex); |
3810 | put_event(event); | |
3811 | ||
3812 | /* | |
3813 | * Restart the iteration; either we're on the wrong list or | |
3814 | * destroyed its integrity by doing a deletion. | |
3815 | */ | |
3816 | goto again; | |
7b732a75 | 3817 | } |
9bb5d40c PZ |
3818 | rcu_read_unlock(); |
3819 | ||
3820 | /* | |
3821 | * It could be there's still a few 0-ref events on the list; they'll | |
3822 | * get cleaned up by free_event() -- they'll also still have their | |
3823 | * ref on the rb and will free it whenever they are done with it. | |
3824 | * | |
3825 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
3826 | * undo the VM accounting. | |
3827 | */ | |
3828 | ||
3829 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
3830 | vma->vm_mm->pinned_vm -= mmap_locked; | |
3831 | free_uid(mmap_user); | |
3832 | ||
3833 | ring_buffer_put(rb); /* could be last */ | |
37d81828 PM |
3834 | } |
3835 | ||
f0f37e2f | 3836 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3837 | .open = perf_mmap_open, |
3838 | .close = perf_mmap_close, | |
3839 | .fault = perf_mmap_fault, | |
3840 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3841 | }; |
3842 | ||
3843 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3844 | { | |
cdd6c482 | 3845 | struct perf_event *event = file->private_data; |
22a4f650 | 3846 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3847 | struct user_struct *user = current_user(); |
22a4f650 | 3848 | unsigned long locked, lock_limit; |
76369139 | 3849 | struct ring_buffer *rb; |
7b732a75 PZ |
3850 | unsigned long vma_size; |
3851 | unsigned long nr_pages; | |
789f90fc | 3852 | long user_extra, extra; |
d57e34fd | 3853 | int ret = 0, flags = 0; |
37d81828 | 3854 | |
c7920614 PZ |
3855 | /* |
3856 | * Don't allow mmap() of inherited per-task counters. This would | |
3857 | * create a performance issue due to all children writing to the | |
76369139 | 3858 | * same rb. |
c7920614 PZ |
3859 | */ |
3860 | if (event->cpu == -1 && event->attr.inherit) | |
3861 | return -EINVAL; | |
3862 | ||
43a21ea8 | 3863 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3864 | return -EINVAL; |
7b732a75 PZ |
3865 | |
3866 | vma_size = vma->vm_end - vma->vm_start; | |
3867 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3868 | ||
7730d865 | 3869 | /* |
76369139 | 3870 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3871 | * can do bitmasks instead of modulo. |
3872 | */ | |
3873 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3874 | return -EINVAL; |
3875 | ||
7b732a75 | 3876 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3877 | return -EINVAL; |
3878 | ||
7b732a75 PZ |
3879 | if (vma->vm_pgoff != 0) |
3880 | return -EINVAL; | |
37d81828 | 3881 | |
cdd6c482 | 3882 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 3883 | again: |
cdd6c482 | 3884 | mutex_lock(&event->mmap_mutex); |
76369139 | 3885 | if (event->rb) { |
9bb5d40c | 3886 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 3887 | ret = -EINVAL; |
9bb5d40c PZ |
3888 | goto unlock; |
3889 | } | |
3890 | ||
3891 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
3892 | /* | |
3893 | * Raced against perf_mmap_close() through | |
3894 | * perf_event_set_output(). Try again, hope for better | |
3895 | * luck. | |
3896 | */ | |
3897 | mutex_unlock(&event->mmap_mutex); | |
3898 | goto again; | |
3899 | } | |
3900 | ||
ebb3c4c4 PZ |
3901 | goto unlock; |
3902 | } | |
3903 | ||
789f90fc | 3904 | user_extra = nr_pages + 1; |
cdd6c482 | 3905 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
3906 | |
3907 | /* | |
3908 | * Increase the limit linearly with more CPUs: | |
3909 | */ | |
3910 | user_lock_limit *= num_online_cpus(); | |
3911 | ||
789f90fc | 3912 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 3913 | |
789f90fc PZ |
3914 | extra = 0; |
3915 | if (user_locked > user_lock_limit) | |
3916 | extra = user_locked - user_lock_limit; | |
7b732a75 | 3917 | |
78d7d407 | 3918 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 3919 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 3920 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 3921 | |
459ec28a IM |
3922 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
3923 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
3924 | ret = -EPERM; |
3925 | goto unlock; | |
3926 | } | |
7b732a75 | 3927 | |
76369139 | 3928 | WARN_ON(event->rb); |
906010b2 | 3929 | |
d57e34fd | 3930 | if (vma->vm_flags & VM_WRITE) |
76369139 | 3931 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 3932 | |
4ec8363d VW |
3933 | rb = rb_alloc(nr_pages, |
3934 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
3935 | event->cpu, flags); | |
3936 | ||
76369139 | 3937 | if (!rb) { |
ac9721f3 | 3938 | ret = -ENOMEM; |
ebb3c4c4 | 3939 | goto unlock; |
ac9721f3 | 3940 | } |
26cb63ad | 3941 | |
9bb5d40c | 3942 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
3943 | rb->mmap_locked = extra; |
3944 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 3945 | |
ac9721f3 | 3946 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
3947 | vma->vm_mm->pinned_vm += extra; |
3948 | ||
9bb5d40c | 3949 | ring_buffer_attach(event, rb); |
26cb63ad | 3950 | rcu_assign_pointer(event->rb, rb); |
ac9721f3 | 3951 | |
9a0f05cb PZ |
3952 | perf_event_update_userpage(event); |
3953 | ||
ebb3c4c4 | 3954 | unlock: |
ac9721f3 PZ |
3955 | if (!ret) |
3956 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 3957 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 3958 | |
9bb5d40c PZ |
3959 | /* |
3960 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
3961 | * vma. | |
3962 | */ | |
26cb63ad | 3963 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 3964 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
3965 | |
3966 | return ret; | |
37d81828 PM |
3967 | } |
3968 | ||
3c446b3d PZ |
3969 | static int perf_fasync(int fd, struct file *filp, int on) |
3970 | { | |
496ad9aa | 3971 | struct inode *inode = file_inode(filp); |
cdd6c482 | 3972 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
3973 | int retval; |
3974 | ||
3975 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 3976 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
3977 | mutex_unlock(&inode->i_mutex); |
3978 | ||
3979 | if (retval < 0) | |
3980 | return retval; | |
3981 | ||
3982 | return 0; | |
3983 | } | |
3984 | ||
0793a61d | 3985 | static const struct file_operations perf_fops = { |
3326c1ce | 3986 | .llseek = no_llseek, |
0793a61d TG |
3987 | .release = perf_release, |
3988 | .read = perf_read, | |
3989 | .poll = perf_poll, | |
d859e29f | 3990 | .unlocked_ioctl = perf_ioctl, |
85887973 | 3991 | .compat_ioctl = perf_compat_ioctl, |
37d81828 | 3992 | .mmap = perf_mmap, |
3c446b3d | 3993 | .fasync = perf_fasync, |
0793a61d TG |
3994 | }; |
3995 | ||
925d519a | 3996 | /* |
cdd6c482 | 3997 | * Perf event wakeup |
925d519a PZ |
3998 | * |
3999 | * If there's data, ensure we set the poll() state and publish everything | |
4000 | * to user-space before waking everybody up. | |
4001 | */ | |
4002 | ||
cdd6c482 | 4003 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4004 | { |
10c6db11 | 4005 | ring_buffer_wakeup(event); |
4c9e2542 | 4006 | |
cdd6c482 IM |
4007 | if (event->pending_kill) { |
4008 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4009 | event->pending_kill = 0; | |
4c9e2542 | 4010 | } |
925d519a PZ |
4011 | } |
4012 | ||
e360adbe | 4013 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4014 | { |
cdd6c482 IM |
4015 | struct perf_event *event = container_of(entry, |
4016 | struct perf_event, pending); | |
a49a0c95 PZ |
4017 | int rctx; |
4018 | ||
4019 | rctx = perf_swevent_get_recursion_context(); | |
4020 | /* | |
4021 | * If we 'fail' here, that's OK, it means recursion is already disabled | |
4022 | * and we won't recurse 'further'. | |
4023 | */ | |
79f14641 | 4024 | |
cdd6c482 IM |
4025 | if (event->pending_disable) { |
4026 | event->pending_disable = 0; | |
4027 | __perf_event_disable(event); | |
79f14641 PZ |
4028 | } |
4029 | ||
cdd6c482 IM |
4030 | if (event->pending_wakeup) { |
4031 | event->pending_wakeup = 0; | |
4032 | perf_event_wakeup(event); | |
79f14641 | 4033 | } |
a49a0c95 PZ |
4034 | |
4035 | if (rctx >= 0) | |
4036 | perf_swevent_put_recursion_context(rctx); | |
79f14641 PZ |
4037 | } |
4038 | ||
39447b38 ZY |
4039 | /* |
4040 | * We assume there is only KVM supporting the callbacks. | |
4041 | * Later on, we might change it to a list if there is | |
4042 | * another virtualization implementation supporting the callbacks. | |
4043 | */ | |
4044 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4045 | ||
4046 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4047 | { | |
4048 | perf_guest_cbs = cbs; | |
4049 | return 0; | |
4050 | } | |
4051 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4052 | ||
4053 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4054 | { | |
4055 | perf_guest_cbs = NULL; | |
4056 | return 0; | |
4057 | } | |
4058 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4059 | ||
4018994f JO |
4060 | static void |
4061 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4062 | struct pt_regs *regs, u64 mask) | |
4063 | { | |
4064 | int bit; | |
4065 | ||
4066 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4067 | sizeof(mask) * BITS_PER_BYTE) { | |
4068 | u64 val; | |
4069 | ||
4070 | val = perf_reg_value(regs, bit); | |
4071 | perf_output_put(handle, val); | |
4072 | } | |
4073 | } | |
4074 | ||
4075 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
4076 | struct pt_regs *regs) | |
4077 | { | |
4078 | if (!user_mode(regs)) { | |
4079 | if (current->mm) | |
4080 | regs = task_pt_regs(current); | |
4081 | else | |
4082 | regs = NULL; | |
4083 | } | |
4084 | ||
4085 | if (regs) { | |
4086 | regs_user->regs = regs; | |
4087 | regs_user->abi = perf_reg_abi(current); | |
4088 | } | |
4089 | } | |
4090 | ||
c5ebcedb JO |
4091 | /* |
4092 | * Get remaining task size from user stack pointer. | |
4093 | * | |
4094 | * It'd be better to take stack vma map and limit this more | |
4095 | * precisly, but there's no way to get it safely under interrupt, | |
4096 | * so using TASK_SIZE as limit. | |
4097 | */ | |
4098 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4099 | { | |
4100 | unsigned long addr = perf_user_stack_pointer(regs); | |
4101 | ||
4102 | if (!addr || addr >= TASK_SIZE) | |
4103 | return 0; | |
4104 | ||
4105 | return TASK_SIZE - addr; | |
4106 | } | |
4107 | ||
4108 | static u16 | |
4109 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4110 | struct pt_regs *regs) | |
4111 | { | |
4112 | u64 task_size; | |
4113 | ||
4114 | /* No regs, no stack pointer, no dump. */ | |
4115 | if (!regs) | |
4116 | return 0; | |
4117 | ||
4118 | /* | |
4119 | * Check if we fit in with the requested stack size into the: | |
4120 | * - TASK_SIZE | |
4121 | * If we don't, we limit the size to the TASK_SIZE. | |
4122 | * | |
4123 | * - remaining sample size | |
4124 | * If we don't, we customize the stack size to | |
4125 | * fit in to the remaining sample size. | |
4126 | */ | |
4127 | ||
4128 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4129 | stack_size = min(stack_size, (u16) task_size); | |
4130 | ||
4131 | /* Current header size plus static size and dynamic size. */ | |
4132 | header_size += 2 * sizeof(u64); | |
4133 | ||
4134 | /* Do we fit in with the current stack dump size? */ | |
4135 | if ((u16) (header_size + stack_size) < header_size) { | |
4136 | /* | |
4137 | * If we overflow the maximum size for the sample, | |
4138 | * we customize the stack dump size to fit in. | |
4139 | */ | |
4140 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4141 | stack_size = round_up(stack_size, sizeof(u64)); | |
4142 | } | |
4143 | ||
4144 | return stack_size; | |
4145 | } | |
4146 | ||
4147 | static void | |
4148 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4149 | struct pt_regs *regs) | |
4150 | { | |
4151 | /* Case of a kernel thread, nothing to dump */ | |
4152 | if (!regs) { | |
4153 | u64 size = 0; | |
4154 | perf_output_put(handle, size); | |
4155 | } else { | |
4156 | unsigned long sp; | |
4157 | unsigned int rem; | |
4158 | u64 dyn_size; | |
4159 | ||
4160 | /* | |
4161 | * We dump: | |
4162 | * static size | |
4163 | * - the size requested by user or the best one we can fit | |
4164 | * in to the sample max size | |
4165 | * data | |
4166 | * - user stack dump data | |
4167 | * dynamic size | |
4168 | * - the actual dumped size | |
4169 | */ | |
4170 | ||
4171 | /* Static size. */ | |
4172 | perf_output_put(handle, dump_size); | |
4173 | ||
4174 | /* Data. */ | |
4175 | sp = perf_user_stack_pointer(regs); | |
4176 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4177 | dyn_size = dump_size - rem; | |
4178 | ||
4179 | perf_output_skip(handle, rem); | |
4180 | ||
4181 | /* Dynamic size. */ | |
4182 | perf_output_put(handle, dyn_size); | |
4183 | } | |
4184 | } | |
4185 | ||
c980d109 ACM |
4186 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4187 | struct perf_sample_data *data, | |
4188 | struct perf_event *event) | |
6844c09d ACM |
4189 | { |
4190 | u64 sample_type = event->attr.sample_type; | |
4191 | ||
4192 | data->type = sample_type; | |
4193 | header->size += event->id_header_size; | |
4194 | ||
4195 | if (sample_type & PERF_SAMPLE_TID) { | |
4196 | /* namespace issues */ | |
4197 | data->tid_entry.pid = perf_event_pid(event, current); | |
4198 | data->tid_entry.tid = perf_event_tid(event, current); | |
4199 | } | |
4200 | ||
4201 | if (sample_type & PERF_SAMPLE_TIME) | |
4202 | data->time = perf_clock(); | |
4203 | ||
4204 | if (sample_type & PERF_SAMPLE_ID) | |
4205 | data->id = primary_event_id(event); | |
4206 | ||
4207 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4208 | data->stream_id = event->id; | |
4209 | ||
4210 | if (sample_type & PERF_SAMPLE_CPU) { | |
4211 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4212 | data->cpu_entry.reserved = 0; | |
4213 | } | |
4214 | } | |
4215 | ||
76369139 FW |
4216 | void perf_event_header__init_id(struct perf_event_header *header, |
4217 | struct perf_sample_data *data, | |
4218 | struct perf_event *event) | |
c980d109 ACM |
4219 | { |
4220 | if (event->attr.sample_id_all) | |
4221 | __perf_event_header__init_id(header, data, event); | |
4222 | } | |
4223 | ||
4224 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4225 | struct perf_sample_data *data) | |
4226 | { | |
4227 | u64 sample_type = data->type; | |
4228 | ||
4229 | if (sample_type & PERF_SAMPLE_TID) | |
4230 | perf_output_put(handle, data->tid_entry); | |
4231 | ||
4232 | if (sample_type & PERF_SAMPLE_TIME) | |
4233 | perf_output_put(handle, data->time); | |
4234 | ||
4235 | if (sample_type & PERF_SAMPLE_ID) | |
4236 | perf_output_put(handle, data->id); | |
4237 | ||
4238 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4239 | perf_output_put(handle, data->stream_id); | |
4240 | ||
4241 | if (sample_type & PERF_SAMPLE_CPU) | |
4242 | perf_output_put(handle, data->cpu_entry); | |
4243 | } | |
4244 | ||
76369139 FW |
4245 | void perf_event__output_id_sample(struct perf_event *event, |
4246 | struct perf_output_handle *handle, | |
4247 | struct perf_sample_data *sample) | |
c980d109 ACM |
4248 | { |
4249 | if (event->attr.sample_id_all) | |
4250 | __perf_event__output_id_sample(handle, sample); | |
4251 | } | |
4252 | ||
3dab77fb | 4253 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4254 | struct perf_event *event, |
4255 | u64 enabled, u64 running) | |
3dab77fb | 4256 | { |
cdd6c482 | 4257 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4258 | u64 values[4]; |
4259 | int n = 0; | |
4260 | ||
b5e58793 | 4261 | values[n++] = perf_event_count(event); |
3dab77fb | 4262 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4263 | values[n++] = enabled + |
cdd6c482 | 4264 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4265 | } |
4266 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4267 | values[n++] = running + |
cdd6c482 | 4268 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4269 | } |
4270 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4271 | values[n++] = primary_event_id(event); |
3dab77fb | 4272 | |
76369139 | 4273 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4274 | } |
4275 | ||
4276 | /* | |
cdd6c482 | 4277 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4278 | */ |
4279 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4280 | struct perf_event *event, |
4281 | u64 enabled, u64 running) | |
3dab77fb | 4282 | { |
cdd6c482 IM |
4283 | struct perf_event *leader = event->group_leader, *sub; |
4284 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4285 | u64 values[5]; |
4286 | int n = 0; | |
4287 | ||
4288 | values[n++] = 1 + leader->nr_siblings; | |
4289 | ||
4290 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4291 | values[n++] = enabled; |
3dab77fb PZ |
4292 | |
4293 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4294 | values[n++] = running; |
3dab77fb | 4295 | |
cdd6c482 | 4296 | if (leader != event) |
3dab77fb PZ |
4297 | leader->pmu->read(leader); |
4298 | ||
b5e58793 | 4299 | values[n++] = perf_event_count(leader); |
3dab77fb | 4300 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4301 | values[n++] = primary_event_id(leader); |
3dab77fb | 4302 | |
76369139 | 4303 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4304 | |
65abc865 | 4305 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4306 | n = 0; |
4307 | ||
cdd6c482 | 4308 | if (sub != event) |
3dab77fb PZ |
4309 | sub->pmu->read(sub); |
4310 | ||
b5e58793 | 4311 | values[n++] = perf_event_count(sub); |
3dab77fb | 4312 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4313 | values[n++] = primary_event_id(sub); |
3dab77fb | 4314 | |
76369139 | 4315 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4316 | } |
4317 | } | |
4318 | ||
eed01528 SE |
4319 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4320 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4321 | ||
3dab77fb | 4322 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4323 | struct perf_event *event) |
3dab77fb | 4324 | { |
e3f3541c | 4325 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4326 | u64 read_format = event->attr.read_format; |
4327 | ||
4328 | /* | |
4329 | * compute total_time_enabled, total_time_running | |
4330 | * based on snapshot values taken when the event | |
4331 | * was last scheduled in. | |
4332 | * | |
4333 | * we cannot simply called update_context_time() | |
4334 | * because of locking issue as we are called in | |
4335 | * NMI context | |
4336 | */ | |
c4794295 | 4337 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4338 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4339 | |
cdd6c482 | 4340 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4341 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4342 | else |
eed01528 | 4343 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4344 | } |
4345 | ||
5622f295 MM |
4346 | void perf_output_sample(struct perf_output_handle *handle, |
4347 | struct perf_event_header *header, | |
4348 | struct perf_sample_data *data, | |
cdd6c482 | 4349 | struct perf_event *event) |
5622f295 MM |
4350 | { |
4351 | u64 sample_type = data->type; | |
4352 | ||
4353 | perf_output_put(handle, *header); | |
4354 | ||
4355 | if (sample_type & PERF_SAMPLE_IP) | |
4356 | perf_output_put(handle, data->ip); | |
4357 | ||
4358 | if (sample_type & PERF_SAMPLE_TID) | |
4359 | perf_output_put(handle, data->tid_entry); | |
4360 | ||
4361 | if (sample_type & PERF_SAMPLE_TIME) | |
4362 | perf_output_put(handle, data->time); | |
4363 | ||
4364 | if (sample_type & PERF_SAMPLE_ADDR) | |
4365 | perf_output_put(handle, data->addr); | |
4366 | ||
4367 | if (sample_type & PERF_SAMPLE_ID) | |
4368 | perf_output_put(handle, data->id); | |
4369 | ||
4370 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4371 | perf_output_put(handle, data->stream_id); | |
4372 | ||
4373 | if (sample_type & PERF_SAMPLE_CPU) | |
4374 | perf_output_put(handle, data->cpu_entry); | |
4375 | ||
4376 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4377 | perf_output_put(handle, data->period); | |
4378 | ||
4379 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4380 | perf_output_read(handle, event); |
5622f295 MM |
4381 | |
4382 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4383 | if (data->callchain) { | |
4384 | int size = 1; | |
4385 | ||
4386 | if (data->callchain) | |
4387 | size += data->callchain->nr; | |
4388 | ||
4389 | size *= sizeof(u64); | |
4390 | ||
76369139 | 4391 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4392 | } else { |
4393 | u64 nr = 0; | |
4394 | perf_output_put(handle, nr); | |
4395 | } | |
4396 | } | |
4397 | ||
4398 | if (sample_type & PERF_SAMPLE_RAW) { | |
4399 | if (data->raw) { | |
4400 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4401 | __output_copy(handle, data->raw->data, |
4402 | data->raw->size); | |
5622f295 MM |
4403 | } else { |
4404 | struct { | |
4405 | u32 size; | |
4406 | u32 data; | |
4407 | } raw = { | |
4408 | .size = sizeof(u32), | |
4409 | .data = 0, | |
4410 | }; | |
4411 | perf_output_put(handle, raw); | |
4412 | } | |
4413 | } | |
a7ac67ea PZ |
4414 | |
4415 | if (!event->attr.watermark) { | |
4416 | int wakeup_events = event->attr.wakeup_events; | |
4417 | ||
4418 | if (wakeup_events) { | |
4419 | struct ring_buffer *rb = handle->rb; | |
4420 | int events = local_inc_return(&rb->events); | |
4421 | ||
4422 | if (events >= wakeup_events) { | |
4423 | local_sub(wakeup_events, &rb->events); | |
4424 | local_inc(&rb->wakeup); | |
4425 | } | |
4426 | } | |
4427 | } | |
bce38cd5 SE |
4428 | |
4429 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4430 | if (data->br_stack) { | |
4431 | size_t size; | |
4432 | ||
4433 | size = data->br_stack->nr | |
4434 | * sizeof(struct perf_branch_entry); | |
4435 | ||
4436 | perf_output_put(handle, data->br_stack->nr); | |
4437 | perf_output_copy(handle, data->br_stack->entries, size); | |
4438 | } else { | |
4439 | /* | |
4440 | * we always store at least the value of nr | |
4441 | */ | |
4442 | u64 nr = 0; | |
4443 | perf_output_put(handle, nr); | |
4444 | } | |
4445 | } | |
4018994f JO |
4446 | |
4447 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4448 | u64 abi = data->regs_user.abi; | |
4449 | ||
4450 | /* | |
4451 | * If there are no regs to dump, notice it through | |
4452 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4453 | */ | |
4454 | perf_output_put(handle, abi); | |
4455 | ||
4456 | if (abi) { | |
4457 | u64 mask = event->attr.sample_regs_user; | |
4458 | perf_output_sample_regs(handle, | |
4459 | data->regs_user.regs, | |
4460 | mask); | |
4461 | } | |
4462 | } | |
c5ebcedb JO |
4463 | |
4464 | if (sample_type & PERF_SAMPLE_STACK_USER) | |
4465 | perf_output_sample_ustack(handle, | |
4466 | data->stack_user_size, | |
4467 | data->regs_user.regs); | |
c3feedf2 AK |
4468 | |
4469 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
4470 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
4471 | |
4472 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
4473 | perf_output_put(handle, data->data_src.val); | |
5622f295 MM |
4474 | } |
4475 | ||
4476 | void perf_prepare_sample(struct perf_event_header *header, | |
4477 | struct perf_sample_data *data, | |
cdd6c482 | 4478 | struct perf_event *event, |
5622f295 | 4479 | struct pt_regs *regs) |
7b732a75 | 4480 | { |
cdd6c482 | 4481 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4482 | |
cdd6c482 | 4483 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4484 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4485 | |
4486 | header->misc = 0; | |
4487 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4488 | |
c980d109 | 4489 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4490 | |
c320c7b7 | 4491 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4492 | data->ip = perf_instruction_pointer(regs); |
4493 | ||
b23f3325 | 4494 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4495 | int size = 1; |
394ee076 | 4496 | |
e6dab5ff | 4497 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4498 | |
4499 | if (data->callchain) | |
4500 | size += data->callchain->nr; | |
4501 | ||
4502 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4503 | } |
4504 | ||
3a43ce68 | 4505 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4506 | int size = sizeof(u32); |
4507 | ||
4508 | if (data->raw) | |
4509 | size += data->raw->size; | |
4510 | else | |
4511 | size += sizeof(u32); | |
4512 | ||
4513 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4514 | header->size += size; |
7f453c24 | 4515 | } |
bce38cd5 SE |
4516 | |
4517 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4518 | int size = sizeof(u64); /* nr */ | |
4519 | if (data->br_stack) { | |
4520 | size += data->br_stack->nr | |
4521 | * sizeof(struct perf_branch_entry); | |
4522 | } | |
4523 | header->size += size; | |
4524 | } | |
4018994f JO |
4525 | |
4526 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4527 | /* regs dump ABI info */ | |
4528 | int size = sizeof(u64); | |
4529 | ||
4530 | perf_sample_regs_user(&data->regs_user, regs); | |
4531 | ||
4532 | if (data->regs_user.regs) { | |
4533 | u64 mask = event->attr.sample_regs_user; | |
4534 | size += hweight64(mask) * sizeof(u64); | |
4535 | } | |
4536 | ||
4537 | header->size += size; | |
4538 | } | |
c5ebcedb JO |
4539 | |
4540 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4541 | /* | |
4542 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4543 | * processed as the last one or have additional check added | |
4544 | * in case new sample type is added, because we could eat | |
4545 | * up the rest of the sample size. | |
4546 | */ | |
4547 | struct perf_regs_user *uregs = &data->regs_user; | |
4548 | u16 stack_size = event->attr.sample_stack_user; | |
4549 | u16 size = sizeof(u64); | |
4550 | ||
4551 | if (!uregs->abi) | |
4552 | perf_sample_regs_user(uregs, regs); | |
4553 | ||
4554 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4555 | uregs->regs); | |
4556 | ||
4557 | /* | |
4558 | * If there is something to dump, add space for the dump | |
4559 | * itself and for the field that tells the dynamic size, | |
4560 | * which is how many have been actually dumped. | |
4561 | */ | |
4562 | if (stack_size) | |
4563 | size += sizeof(u64) + stack_size; | |
4564 | ||
4565 | data->stack_user_size = stack_size; | |
4566 | header->size += size; | |
4567 | } | |
5622f295 | 4568 | } |
7f453c24 | 4569 | |
a8b0ca17 | 4570 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4571 | struct perf_sample_data *data, |
4572 | struct pt_regs *regs) | |
4573 | { | |
4574 | struct perf_output_handle handle; | |
4575 | struct perf_event_header header; | |
689802b2 | 4576 | |
927c7a9e FW |
4577 | /* protect the callchain buffers */ |
4578 | rcu_read_lock(); | |
4579 | ||
cdd6c482 | 4580 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4581 | |
a7ac67ea | 4582 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4583 | goto exit; |
0322cd6e | 4584 | |
cdd6c482 | 4585 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4586 | |
8a057d84 | 4587 | perf_output_end(&handle); |
927c7a9e FW |
4588 | |
4589 | exit: | |
4590 | rcu_read_unlock(); | |
0322cd6e PZ |
4591 | } |
4592 | ||
38b200d6 | 4593 | /* |
cdd6c482 | 4594 | * read event_id |
38b200d6 PZ |
4595 | */ |
4596 | ||
4597 | struct perf_read_event { | |
4598 | struct perf_event_header header; | |
4599 | ||
4600 | u32 pid; | |
4601 | u32 tid; | |
38b200d6 PZ |
4602 | }; |
4603 | ||
4604 | static void | |
cdd6c482 | 4605 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4606 | struct task_struct *task) |
4607 | { | |
4608 | struct perf_output_handle handle; | |
c980d109 | 4609 | struct perf_sample_data sample; |
dfc65094 | 4610 | struct perf_read_event read_event = { |
38b200d6 | 4611 | .header = { |
cdd6c482 | 4612 | .type = PERF_RECORD_READ, |
38b200d6 | 4613 | .misc = 0, |
c320c7b7 | 4614 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4615 | }, |
cdd6c482 IM |
4616 | .pid = perf_event_pid(event, task), |
4617 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4618 | }; |
3dab77fb | 4619 | int ret; |
38b200d6 | 4620 | |
c980d109 | 4621 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4622 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4623 | if (ret) |
4624 | return; | |
4625 | ||
dfc65094 | 4626 | perf_output_put(&handle, read_event); |
cdd6c482 | 4627 | perf_output_read(&handle, event); |
c980d109 | 4628 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4629 | |
38b200d6 PZ |
4630 | perf_output_end(&handle); |
4631 | } | |
4632 | ||
52d857a8 JO |
4633 | typedef int (perf_event_aux_match_cb)(struct perf_event *event, void *data); |
4634 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); | |
4635 | ||
4636 | static void | |
4637 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
4638 | perf_event_aux_match_cb match, | |
4639 | perf_event_aux_output_cb output, | |
4640 | void *data) | |
4641 | { | |
4642 | struct perf_event *event; | |
4643 | ||
4644 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
4645 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
4646 | continue; | |
4647 | if (!event_filter_match(event)) | |
4648 | continue; | |
4649 | if (match(event, data)) | |
4650 | output(event, data); | |
4651 | } | |
4652 | } | |
4653 | ||
4654 | static void | |
4655 | perf_event_aux(perf_event_aux_match_cb match, | |
4656 | perf_event_aux_output_cb output, | |
4657 | void *data, | |
4658 | struct perf_event_context *task_ctx) | |
4659 | { | |
4660 | struct perf_cpu_context *cpuctx; | |
4661 | struct perf_event_context *ctx; | |
4662 | struct pmu *pmu; | |
4663 | int ctxn; | |
4664 | ||
4665 | rcu_read_lock(); | |
4666 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
4667 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
4668 | if (cpuctx->unique_pmu != pmu) | |
4669 | goto next; | |
4670 | perf_event_aux_ctx(&cpuctx->ctx, match, output, data); | |
4671 | if (task_ctx) | |
4672 | goto next; | |
4673 | ctxn = pmu->task_ctx_nr; | |
4674 | if (ctxn < 0) | |
4675 | goto next; | |
4676 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4677 | if (ctx) | |
4678 | perf_event_aux_ctx(ctx, match, output, data); | |
4679 | next: | |
4680 | put_cpu_ptr(pmu->pmu_cpu_context); | |
4681 | } | |
4682 | ||
4683 | if (task_ctx) { | |
4684 | preempt_disable(); | |
4685 | perf_event_aux_ctx(task_ctx, match, output, data); | |
4686 | preempt_enable(); | |
4687 | } | |
4688 | rcu_read_unlock(); | |
4689 | } | |
4690 | ||
60313ebe | 4691 | /* |
9f498cc5 PZ |
4692 | * task tracking -- fork/exit |
4693 | * | |
3af9e859 | 4694 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
60313ebe PZ |
4695 | */ |
4696 | ||
9f498cc5 | 4697 | struct perf_task_event { |
3a80b4a3 | 4698 | struct task_struct *task; |
cdd6c482 | 4699 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4700 | |
4701 | struct { | |
4702 | struct perf_event_header header; | |
4703 | ||
4704 | u32 pid; | |
4705 | u32 ppid; | |
9f498cc5 PZ |
4706 | u32 tid; |
4707 | u32 ptid; | |
393b2ad8 | 4708 | u64 time; |
cdd6c482 | 4709 | } event_id; |
60313ebe PZ |
4710 | }; |
4711 | ||
cdd6c482 | 4712 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 4713 | void *data) |
60313ebe | 4714 | { |
52d857a8 | 4715 | struct perf_task_event *task_event = data; |
60313ebe | 4716 | struct perf_output_handle handle; |
c980d109 | 4717 | struct perf_sample_data sample; |
9f498cc5 | 4718 | struct task_struct *task = task_event->task; |
c980d109 | 4719 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4720 | |
c980d109 | 4721 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4722 | |
c980d109 | 4723 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4724 | task_event->event_id.header.size); |
ef60777c | 4725 | if (ret) |
c980d109 | 4726 | goto out; |
60313ebe | 4727 | |
cdd6c482 IM |
4728 | task_event->event_id.pid = perf_event_pid(event, task); |
4729 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4730 | |
cdd6c482 IM |
4731 | task_event->event_id.tid = perf_event_tid(event, task); |
4732 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4733 | |
cdd6c482 | 4734 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4735 | |
c980d109 ACM |
4736 | perf_event__output_id_sample(event, &handle, &sample); |
4737 | ||
60313ebe | 4738 | perf_output_end(&handle); |
c980d109 ACM |
4739 | out: |
4740 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4741 | } |
4742 | ||
52d857a8 JO |
4743 | static int perf_event_task_match(struct perf_event *event, |
4744 | void *data __maybe_unused) | |
60313ebe | 4745 | { |
52d857a8 JO |
4746 | return event->attr.comm || event->attr.mmap || |
4747 | event->attr.mmap_data || event->attr.task; | |
60313ebe PZ |
4748 | } |
4749 | ||
cdd6c482 IM |
4750 | static void perf_event_task(struct task_struct *task, |
4751 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4752 | int new) |
60313ebe | 4753 | { |
9f498cc5 | 4754 | struct perf_task_event task_event; |
60313ebe | 4755 | |
cdd6c482 IM |
4756 | if (!atomic_read(&nr_comm_events) && |
4757 | !atomic_read(&nr_mmap_events) && | |
4758 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4759 | return; |
4760 | ||
9f498cc5 | 4761 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4762 | .task = task, |
4763 | .task_ctx = task_ctx, | |
cdd6c482 | 4764 | .event_id = { |
60313ebe | 4765 | .header = { |
cdd6c482 | 4766 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4767 | .misc = 0, |
cdd6c482 | 4768 | .size = sizeof(task_event.event_id), |
60313ebe | 4769 | }, |
573402db PZ |
4770 | /* .pid */ |
4771 | /* .ppid */ | |
9f498cc5 PZ |
4772 | /* .tid */ |
4773 | /* .ptid */ | |
6f93d0a7 | 4774 | .time = perf_clock(), |
60313ebe PZ |
4775 | }, |
4776 | }; | |
4777 | ||
52d857a8 JO |
4778 | perf_event_aux(perf_event_task_match, |
4779 | perf_event_task_output, | |
4780 | &task_event, | |
4781 | task_ctx); | |
9f498cc5 PZ |
4782 | } |
4783 | ||
cdd6c482 | 4784 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4785 | { |
cdd6c482 | 4786 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4787 | } |
4788 | ||
8d1b2d93 PZ |
4789 | /* |
4790 | * comm tracking | |
4791 | */ | |
4792 | ||
4793 | struct perf_comm_event { | |
22a4f650 IM |
4794 | struct task_struct *task; |
4795 | char *comm; | |
8d1b2d93 PZ |
4796 | int comm_size; |
4797 | ||
4798 | struct { | |
4799 | struct perf_event_header header; | |
4800 | ||
4801 | u32 pid; | |
4802 | u32 tid; | |
cdd6c482 | 4803 | } event_id; |
8d1b2d93 PZ |
4804 | }; |
4805 | ||
cdd6c482 | 4806 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 4807 | void *data) |
8d1b2d93 | 4808 | { |
52d857a8 | 4809 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 4810 | struct perf_output_handle handle; |
c980d109 | 4811 | struct perf_sample_data sample; |
cdd6c482 | 4812 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4813 | int ret; |
4814 | ||
4815 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); | |
4816 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4817 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4818 | |
4819 | if (ret) | |
c980d109 | 4820 | goto out; |
8d1b2d93 | 4821 | |
cdd6c482 IM |
4822 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4823 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4824 | |
cdd6c482 | 4825 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4826 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4827 | comm_event->comm_size); |
c980d109 ACM |
4828 | |
4829 | perf_event__output_id_sample(event, &handle, &sample); | |
4830 | ||
8d1b2d93 | 4831 | perf_output_end(&handle); |
c980d109 ACM |
4832 | out: |
4833 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4834 | } |
4835 | ||
52d857a8 JO |
4836 | static int perf_event_comm_match(struct perf_event *event, |
4837 | void *data __maybe_unused) | |
8d1b2d93 | 4838 | { |
52d857a8 | 4839 | return event->attr.comm; |
8d1b2d93 PZ |
4840 | } |
4841 | ||
cdd6c482 | 4842 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 4843 | { |
413ee3b4 | 4844 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4845 | unsigned int size; |
8d1b2d93 | 4846 | |
413ee3b4 | 4847 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4848 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4849 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4850 | |
4851 | comm_event->comm = comm; | |
4852 | comm_event->comm_size = size; | |
4853 | ||
cdd6c482 | 4854 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 4855 | |
52d857a8 JO |
4856 | perf_event_aux(perf_event_comm_match, |
4857 | perf_event_comm_output, | |
4858 | comm_event, | |
4859 | NULL); | |
8d1b2d93 PZ |
4860 | } |
4861 | ||
cdd6c482 | 4862 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4863 | { |
9ee318a7 | 4864 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4865 | struct perf_event_context *ctx; |
4866 | int ctxn; | |
9ee318a7 | 4867 | |
c79aa0d9 | 4868 | rcu_read_lock(); |
8dc85d54 PZ |
4869 | for_each_task_context_nr(ctxn) { |
4870 | ctx = task->perf_event_ctxp[ctxn]; | |
4871 | if (!ctx) | |
4872 | continue; | |
9ee318a7 | 4873 | |
8dc85d54 PZ |
4874 | perf_event_enable_on_exec(ctx); |
4875 | } | |
c79aa0d9 | 4876 | rcu_read_unlock(); |
9ee318a7 | 4877 | |
cdd6c482 | 4878 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4879 | return; |
a63eaf34 | 4880 | |
9ee318a7 | 4881 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4882 | .task = task, |
573402db PZ |
4883 | /* .comm */ |
4884 | /* .comm_size */ | |
cdd6c482 | 4885 | .event_id = { |
573402db | 4886 | .header = { |
cdd6c482 | 4887 | .type = PERF_RECORD_COMM, |
573402db PZ |
4888 | .misc = 0, |
4889 | /* .size */ | |
4890 | }, | |
4891 | /* .pid */ | |
4892 | /* .tid */ | |
8d1b2d93 PZ |
4893 | }, |
4894 | }; | |
4895 | ||
cdd6c482 | 4896 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
4897 | } |
4898 | ||
0a4a9391 PZ |
4899 | /* |
4900 | * mmap tracking | |
4901 | */ | |
4902 | ||
4903 | struct perf_mmap_event { | |
089dd79d PZ |
4904 | struct vm_area_struct *vma; |
4905 | ||
4906 | const char *file_name; | |
4907 | int file_size; | |
0a4a9391 PZ |
4908 | |
4909 | struct { | |
4910 | struct perf_event_header header; | |
4911 | ||
4912 | u32 pid; | |
4913 | u32 tid; | |
4914 | u64 start; | |
4915 | u64 len; | |
4916 | u64 pgoff; | |
cdd6c482 | 4917 | } event_id; |
0a4a9391 PZ |
4918 | }; |
4919 | ||
cdd6c482 | 4920 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 4921 | void *data) |
0a4a9391 | 4922 | { |
52d857a8 | 4923 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 4924 | struct perf_output_handle handle; |
c980d109 | 4925 | struct perf_sample_data sample; |
cdd6c482 | 4926 | int size = mmap_event->event_id.header.size; |
c980d109 | 4927 | int ret; |
0a4a9391 | 4928 | |
c980d109 ACM |
4929 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
4930 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4931 | mmap_event->event_id.header.size); |
0a4a9391 | 4932 | if (ret) |
c980d109 | 4933 | goto out; |
0a4a9391 | 4934 | |
cdd6c482 IM |
4935 | mmap_event->event_id.pid = perf_event_pid(event, current); |
4936 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 4937 | |
cdd6c482 | 4938 | perf_output_put(&handle, mmap_event->event_id); |
76369139 | 4939 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 4940 | mmap_event->file_size); |
c980d109 ACM |
4941 | |
4942 | perf_event__output_id_sample(event, &handle, &sample); | |
4943 | ||
78d613eb | 4944 | perf_output_end(&handle); |
c980d109 ACM |
4945 | out: |
4946 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
4947 | } |
4948 | ||
cdd6c482 | 4949 | static int perf_event_mmap_match(struct perf_event *event, |
52d857a8 | 4950 | void *data) |
0a4a9391 | 4951 | { |
52d857a8 JO |
4952 | struct perf_mmap_event *mmap_event = data; |
4953 | struct vm_area_struct *vma = mmap_event->vma; | |
4954 | int executable = vma->vm_flags & VM_EXEC; | |
0a4a9391 | 4955 | |
52d857a8 JO |
4956 | return (!executable && event->attr.mmap_data) || |
4957 | (executable && event->attr.mmap); | |
0a4a9391 PZ |
4958 | } |
4959 | ||
cdd6c482 | 4960 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 4961 | { |
089dd79d PZ |
4962 | struct vm_area_struct *vma = mmap_event->vma; |
4963 | struct file *file = vma->vm_file; | |
0a4a9391 PZ |
4964 | unsigned int size; |
4965 | char tmp[16]; | |
4966 | char *buf = NULL; | |
089dd79d | 4967 | const char *name; |
0a4a9391 | 4968 | |
413ee3b4 AB |
4969 | memset(tmp, 0, sizeof(tmp)); |
4970 | ||
0a4a9391 | 4971 | if (file) { |
413ee3b4 | 4972 | /* |
76369139 | 4973 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
4974 | * need to add enough zero bytes after the string to handle |
4975 | * the 64bit alignment we do later. | |
4976 | */ | |
4977 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
4978 | if (!buf) { |
4979 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
4980 | goto got_name; | |
4981 | } | |
d3d21c41 | 4982 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
4983 | if (IS_ERR(name)) { |
4984 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
4985 | goto got_name; | |
4986 | } | |
4987 | } else { | |
413ee3b4 AB |
4988 | if (arch_vma_name(mmap_event->vma)) { |
4989 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
c97847d2 CG |
4990 | sizeof(tmp) - 1); |
4991 | tmp[sizeof(tmp) - 1] = '\0'; | |
089dd79d | 4992 | goto got_name; |
413ee3b4 | 4993 | } |
089dd79d PZ |
4994 | |
4995 | if (!vma->vm_mm) { | |
4996 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
4997 | goto got_name; | |
3af9e859 EM |
4998 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
4999 | vma->vm_end >= vma->vm_mm->brk) { | |
5000 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
5001 | goto got_name; | |
5002 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
5003 | vma->vm_end >= vma->vm_mm->start_stack) { | |
5004 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
5005 | goto got_name; | |
089dd79d PZ |
5006 | } |
5007 | ||
0a4a9391 PZ |
5008 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
5009 | goto got_name; | |
5010 | } | |
5011 | ||
5012 | got_name: | |
888fcee0 | 5013 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
5014 | |
5015 | mmap_event->file_name = name; | |
5016 | mmap_event->file_size = size; | |
5017 | ||
2fe85427 SE |
5018 | if (!(vma->vm_flags & VM_EXEC)) |
5019 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5020 | ||
cdd6c482 | 5021 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5022 | |
52d857a8 JO |
5023 | perf_event_aux(perf_event_mmap_match, |
5024 | perf_event_mmap_output, | |
5025 | mmap_event, | |
5026 | NULL); | |
665c2142 | 5027 | |
0a4a9391 PZ |
5028 | kfree(buf); |
5029 | } | |
5030 | ||
3af9e859 | 5031 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5032 | { |
9ee318a7 PZ |
5033 | struct perf_mmap_event mmap_event; |
5034 | ||
cdd6c482 | 5035 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5036 | return; |
5037 | ||
5038 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5039 | .vma = vma, |
573402db PZ |
5040 | /* .file_name */ |
5041 | /* .file_size */ | |
cdd6c482 | 5042 | .event_id = { |
573402db | 5043 | .header = { |
cdd6c482 | 5044 | .type = PERF_RECORD_MMAP, |
39447b38 | 5045 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5046 | /* .size */ |
5047 | }, | |
5048 | /* .pid */ | |
5049 | /* .tid */ | |
089dd79d PZ |
5050 | .start = vma->vm_start, |
5051 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5052 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 PZ |
5053 | }, |
5054 | }; | |
5055 | ||
cdd6c482 | 5056 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5057 | } |
5058 | ||
a78ac325 PZ |
5059 | /* |
5060 | * IRQ throttle logging | |
5061 | */ | |
5062 | ||
cdd6c482 | 5063 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5064 | { |
5065 | struct perf_output_handle handle; | |
c980d109 | 5066 | struct perf_sample_data sample; |
a78ac325 PZ |
5067 | int ret; |
5068 | ||
5069 | struct { | |
5070 | struct perf_event_header header; | |
5071 | u64 time; | |
cca3f454 | 5072 | u64 id; |
7f453c24 | 5073 | u64 stream_id; |
a78ac325 PZ |
5074 | } throttle_event = { |
5075 | .header = { | |
cdd6c482 | 5076 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5077 | .misc = 0, |
5078 | .size = sizeof(throttle_event), | |
5079 | }, | |
def0a9b2 | 5080 | .time = perf_clock(), |
cdd6c482 IM |
5081 | .id = primary_event_id(event), |
5082 | .stream_id = event->id, | |
a78ac325 PZ |
5083 | }; |
5084 | ||
966ee4d6 | 5085 | if (enable) |
cdd6c482 | 5086 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5087 | |
c980d109 ACM |
5088 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5089 | ||
5090 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5091 | throttle_event.header.size); |
a78ac325 PZ |
5092 | if (ret) |
5093 | return; | |
5094 | ||
5095 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5096 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5097 | perf_output_end(&handle); |
5098 | } | |
5099 | ||
f6c7d5fe | 5100 | /* |
cdd6c482 | 5101 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5102 | */ |
5103 | ||
a8b0ca17 | 5104 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5105 | int throttle, struct perf_sample_data *data, |
5106 | struct pt_regs *regs) | |
f6c7d5fe | 5107 | { |
cdd6c482 IM |
5108 | int events = atomic_read(&event->event_limit); |
5109 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5110 | u64 seq; |
79f14641 PZ |
5111 | int ret = 0; |
5112 | ||
96398826 PZ |
5113 | /* |
5114 | * Non-sampling counters might still use the PMI to fold short | |
5115 | * hardware counters, ignore those. | |
5116 | */ | |
5117 | if (unlikely(!is_sampling_event(event))) | |
5118 | return 0; | |
5119 | ||
e050e3f0 SE |
5120 | seq = __this_cpu_read(perf_throttled_seq); |
5121 | if (seq != hwc->interrupts_seq) { | |
5122 | hwc->interrupts_seq = seq; | |
5123 | hwc->interrupts = 1; | |
5124 | } else { | |
5125 | hwc->interrupts++; | |
5126 | if (unlikely(throttle | |
5127 | && hwc->interrupts >= max_samples_per_tick)) { | |
5128 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5129 | hwc->interrupts = MAX_INTERRUPTS; |
5130 | perf_log_throttle(event, 0); | |
a78ac325 PZ |
5131 | ret = 1; |
5132 | } | |
e050e3f0 | 5133 | } |
60db5e09 | 5134 | |
cdd6c482 | 5135 | if (event->attr.freq) { |
def0a9b2 | 5136 | u64 now = perf_clock(); |
abd50713 | 5137 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5138 | |
abd50713 | 5139 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5140 | |
abd50713 | 5141 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5142 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5143 | } |
5144 | ||
2023b359 PZ |
5145 | /* |
5146 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5147 | * events |
2023b359 PZ |
5148 | */ |
5149 | ||
cdd6c482 IM |
5150 | event->pending_kill = POLL_IN; |
5151 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5152 | ret = 1; |
cdd6c482 | 5153 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5154 | event->pending_disable = 1; |
5155 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5156 | } |
5157 | ||
453f19ee | 5158 | if (event->overflow_handler) |
a8b0ca17 | 5159 | event->overflow_handler(event, data, regs); |
453f19ee | 5160 | else |
a8b0ca17 | 5161 | perf_event_output(event, data, regs); |
453f19ee | 5162 | |
f506b3dc | 5163 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5164 | event->pending_wakeup = 1; |
5165 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5166 | } |
5167 | ||
79f14641 | 5168 | return ret; |
f6c7d5fe PZ |
5169 | } |
5170 | ||
a8b0ca17 | 5171 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5172 | struct perf_sample_data *data, |
5173 | struct pt_regs *regs) | |
850bc73f | 5174 | { |
a8b0ca17 | 5175 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5176 | } |
5177 | ||
15dbf27c | 5178 | /* |
cdd6c482 | 5179 | * Generic software event infrastructure |
15dbf27c PZ |
5180 | */ |
5181 | ||
b28ab83c PZ |
5182 | struct swevent_htable { |
5183 | struct swevent_hlist *swevent_hlist; | |
5184 | struct mutex hlist_mutex; | |
5185 | int hlist_refcount; | |
5186 | ||
5187 | /* Recursion avoidance in each contexts */ | |
5188 | int recursion[PERF_NR_CONTEXTS]; | |
5189 | }; | |
5190 | ||
5191 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5192 | ||
7b4b6658 | 5193 | /* |
cdd6c482 IM |
5194 | * We directly increment event->count and keep a second value in |
5195 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5196 | * is kept in the range [-sample_period, 0] so that we can use the |
5197 | * sign as trigger. | |
5198 | */ | |
5199 | ||
cdd6c482 | 5200 | static u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5201 | { |
cdd6c482 | 5202 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5203 | u64 period = hwc->last_period; |
5204 | u64 nr, offset; | |
5205 | s64 old, val; | |
5206 | ||
5207 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5208 | |
5209 | again: | |
e7850595 | 5210 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5211 | if (val < 0) |
5212 | return 0; | |
15dbf27c | 5213 | |
7b4b6658 PZ |
5214 | nr = div64_u64(period + val, period); |
5215 | offset = nr * period; | |
5216 | val -= offset; | |
e7850595 | 5217 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5218 | goto again; |
15dbf27c | 5219 | |
7b4b6658 | 5220 | return nr; |
15dbf27c PZ |
5221 | } |
5222 | ||
0cff784a | 5223 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5224 | struct perf_sample_data *data, |
5622f295 | 5225 | struct pt_regs *regs) |
15dbf27c | 5226 | { |
cdd6c482 | 5227 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5228 | int throttle = 0; |
15dbf27c | 5229 | |
0cff784a PZ |
5230 | if (!overflow) |
5231 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5232 | |
7b4b6658 PZ |
5233 | if (hwc->interrupts == MAX_INTERRUPTS) |
5234 | return; | |
15dbf27c | 5235 | |
7b4b6658 | 5236 | for (; overflow; overflow--) { |
a8b0ca17 | 5237 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5238 | data, regs)) { |
7b4b6658 PZ |
5239 | /* |
5240 | * We inhibit the overflow from happening when | |
5241 | * hwc->interrupts == MAX_INTERRUPTS. | |
5242 | */ | |
5243 | break; | |
5244 | } | |
cf450a73 | 5245 | throttle = 1; |
7b4b6658 | 5246 | } |
15dbf27c PZ |
5247 | } |
5248 | ||
a4eaf7f1 | 5249 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5250 | struct perf_sample_data *data, |
5622f295 | 5251 | struct pt_regs *regs) |
7b4b6658 | 5252 | { |
cdd6c482 | 5253 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5254 | |
e7850595 | 5255 | local64_add(nr, &event->count); |
d6d020e9 | 5256 | |
0cff784a PZ |
5257 | if (!regs) |
5258 | return; | |
5259 | ||
6c7e550f | 5260 | if (!is_sampling_event(event)) |
7b4b6658 | 5261 | return; |
d6d020e9 | 5262 | |
5d81e5cf AV |
5263 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5264 | data->period = nr; | |
5265 | return perf_swevent_overflow(event, 1, data, regs); | |
5266 | } else | |
5267 | data->period = event->hw.last_period; | |
5268 | ||
0cff784a | 5269 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5270 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5271 | |
e7850595 | 5272 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5273 | return; |
df1a132b | 5274 | |
a8b0ca17 | 5275 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5276 | } |
5277 | ||
f5ffe02e FW |
5278 | static int perf_exclude_event(struct perf_event *event, |
5279 | struct pt_regs *regs) | |
5280 | { | |
a4eaf7f1 | 5281 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5282 | return 1; |
a4eaf7f1 | 5283 | |
f5ffe02e FW |
5284 | if (regs) { |
5285 | if (event->attr.exclude_user && user_mode(regs)) | |
5286 | return 1; | |
5287 | ||
5288 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5289 | return 1; | |
5290 | } | |
5291 | ||
5292 | return 0; | |
5293 | } | |
5294 | ||
cdd6c482 | 5295 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5296 | enum perf_type_id type, |
6fb2915d LZ |
5297 | u32 event_id, |
5298 | struct perf_sample_data *data, | |
5299 | struct pt_regs *regs) | |
15dbf27c | 5300 | { |
cdd6c482 | 5301 | if (event->attr.type != type) |
a21ca2ca | 5302 | return 0; |
f5ffe02e | 5303 | |
cdd6c482 | 5304 | if (event->attr.config != event_id) |
15dbf27c PZ |
5305 | return 0; |
5306 | ||
f5ffe02e FW |
5307 | if (perf_exclude_event(event, regs)) |
5308 | return 0; | |
15dbf27c PZ |
5309 | |
5310 | return 1; | |
5311 | } | |
5312 | ||
76e1d904 FW |
5313 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5314 | { | |
5315 | u64 val = event_id | (type << 32); | |
5316 | ||
5317 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5318 | } | |
5319 | ||
49f135ed FW |
5320 | static inline struct hlist_head * |
5321 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5322 | { |
49f135ed FW |
5323 | u64 hash = swevent_hash(type, event_id); |
5324 | ||
5325 | return &hlist->heads[hash]; | |
5326 | } | |
76e1d904 | 5327 | |
49f135ed FW |
5328 | /* For the read side: events when they trigger */ |
5329 | static inline struct hlist_head * | |
b28ab83c | 5330 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5331 | { |
5332 | struct swevent_hlist *hlist; | |
76e1d904 | 5333 | |
b28ab83c | 5334 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5335 | if (!hlist) |
5336 | return NULL; | |
5337 | ||
49f135ed FW |
5338 | return __find_swevent_head(hlist, type, event_id); |
5339 | } | |
5340 | ||
5341 | /* For the event head insertion and removal in the hlist */ | |
5342 | static inline struct hlist_head * | |
b28ab83c | 5343 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5344 | { |
5345 | struct swevent_hlist *hlist; | |
5346 | u32 event_id = event->attr.config; | |
5347 | u64 type = event->attr.type; | |
5348 | ||
5349 | /* | |
5350 | * Event scheduling is always serialized against hlist allocation | |
5351 | * and release. Which makes the protected version suitable here. | |
5352 | * The context lock guarantees that. | |
5353 | */ | |
b28ab83c | 5354 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5355 | lockdep_is_held(&event->ctx->lock)); |
5356 | if (!hlist) | |
5357 | return NULL; | |
5358 | ||
5359 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5360 | } |
5361 | ||
5362 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5363 | u64 nr, |
76e1d904 FW |
5364 | struct perf_sample_data *data, |
5365 | struct pt_regs *regs) | |
15dbf27c | 5366 | { |
b28ab83c | 5367 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5368 | struct perf_event *event; |
76e1d904 | 5369 | struct hlist_head *head; |
15dbf27c | 5370 | |
76e1d904 | 5371 | rcu_read_lock(); |
b28ab83c | 5372 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5373 | if (!head) |
5374 | goto end; | |
5375 | ||
b67bfe0d | 5376 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5377 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5378 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5379 | } |
76e1d904 FW |
5380 | end: |
5381 | rcu_read_unlock(); | |
15dbf27c PZ |
5382 | } |
5383 | ||
4ed7c92d | 5384 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5385 | { |
b28ab83c | 5386 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5387 | |
b28ab83c | 5388 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5389 | } |
645e8cc0 | 5390 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5391 | |
fa9f90be | 5392 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5393 | { |
b28ab83c | 5394 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5395 | |
b28ab83c | 5396 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5397 | } |
15dbf27c | 5398 | |
a8b0ca17 | 5399 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5400 | { |
a4234bfc | 5401 | struct perf_sample_data data; |
4ed7c92d PZ |
5402 | int rctx; |
5403 | ||
1c024eca | 5404 | preempt_disable_notrace(); |
4ed7c92d PZ |
5405 | rctx = perf_swevent_get_recursion_context(); |
5406 | if (rctx < 0) | |
5407 | return; | |
a4234bfc | 5408 | |
fd0d000b | 5409 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5410 | |
a8b0ca17 | 5411 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5412 | |
5413 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5414 | preempt_enable_notrace(); |
b8e83514 PZ |
5415 | } |
5416 | ||
cdd6c482 | 5417 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5418 | { |
15dbf27c PZ |
5419 | } |
5420 | ||
a4eaf7f1 | 5421 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5422 | { |
b28ab83c | 5423 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5424 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5425 | struct hlist_head *head; |
5426 | ||
6c7e550f | 5427 | if (is_sampling_event(event)) { |
7b4b6658 | 5428 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5429 | perf_swevent_set_period(event); |
7b4b6658 | 5430 | } |
76e1d904 | 5431 | |
a4eaf7f1 PZ |
5432 | hwc->state = !(flags & PERF_EF_START); |
5433 | ||
b28ab83c | 5434 | head = find_swevent_head(swhash, event); |
4b9e9796 | 5435 | if (WARN_ON_ONCE(!head)) |
76e1d904 FW |
5436 | return -EINVAL; |
5437 | ||
5438 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5439 | ||
15dbf27c PZ |
5440 | return 0; |
5441 | } | |
5442 | ||
a4eaf7f1 | 5443 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5444 | { |
76e1d904 | 5445 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5446 | } |
5447 | ||
a4eaf7f1 | 5448 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5449 | { |
a4eaf7f1 | 5450 | event->hw.state = 0; |
d6d020e9 | 5451 | } |
aa9c4c0f | 5452 | |
a4eaf7f1 | 5453 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5454 | { |
a4eaf7f1 | 5455 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5456 | } |
5457 | ||
49f135ed FW |
5458 | /* Deref the hlist from the update side */ |
5459 | static inline struct swevent_hlist * | |
b28ab83c | 5460 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5461 | { |
b28ab83c PZ |
5462 | return rcu_dereference_protected(swhash->swevent_hlist, |
5463 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5464 | } |
5465 | ||
b28ab83c | 5466 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5467 | { |
b28ab83c | 5468 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5469 | |
49f135ed | 5470 | if (!hlist) |
76e1d904 FW |
5471 | return; |
5472 | ||
b28ab83c | 5473 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5474 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5475 | } |
5476 | ||
5477 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5478 | { | |
b28ab83c | 5479 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5480 | |
b28ab83c | 5481 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5482 | |
b28ab83c PZ |
5483 | if (!--swhash->hlist_refcount) |
5484 | swevent_hlist_release(swhash); | |
76e1d904 | 5485 | |
b28ab83c | 5486 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5487 | } |
5488 | ||
5489 | static void swevent_hlist_put(struct perf_event *event) | |
5490 | { | |
5491 | int cpu; | |
5492 | ||
5493 | if (event->cpu != -1) { | |
5494 | swevent_hlist_put_cpu(event, event->cpu); | |
5495 | return; | |
5496 | } | |
5497 | ||
5498 | for_each_possible_cpu(cpu) | |
5499 | swevent_hlist_put_cpu(event, cpu); | |
5500 | } | |
5501 | ||
5502 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5503 | { | |
b28ab83c | 5504 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5505 | int err = 0; |
5506 | ||
b28ab83c | 5507 | mutex_lock(&swhash->hlist_mutex); |
b28ab83c | 5508 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5509 | struct swevent_hlist *hlist; |
5510 | ||
5511 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5512 | if (!hlist) { | |
5513 | err = -ENOMEM; | |
5514 | goto exit; | |
5515 | } | |
b28ab83c | 5516 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5517 | } |
b28ab83c | 5518 | swhash->hlist_refcount++; |
9ed6060d | 5519 | exit: |
b28ab83c | 5520 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5521 | |
5522 | return err; | |
5523 | } | |
5524 | ||
5525 | static int swevent_hlist_get(struct perf_event *event) | |
5526 | { | |
5527 | int err; | |
5528 | int cpu, failed_cpu; | |
5529 | ||
5530 | if (event->cpu != -1) | |
5531 | return swevent_hlist_get_cpu(event, event->cpu); | |
5532 | ||
5533 | get_online_cpus(); | |
5534 | for_each_possible_cpu(cpu) { | |
5535 | err = swevent_hlist_get_cpu(event, cpu); | |
5536 | if (err) { | |
5537 | failed_cpu = cpu; | |
5538 | goto fail; | |
5539 | } | |
5540 | } | |
5541 | put_online_cpus(); | |
5542 | ||
5543 | return 0; | |
9ed6060d | 5544 | fail: |
76e1d904 FW |
5545 | for_each_possible_cpu(cpu) { |
5546 | if (cpu == failed_cpu) | |
5547 | break; | |
5548 | swevent_hlist_put_cpu(event, cpu); | |
5549 | } | |
5550 | ||
5551 | put_online_cpus(); | |
5552 | return err; | |
5553 | } | |
5554 | ||
c5905afb | 5555 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5556 | |
b0a873eb PZ |
5557 | static void sw_perf_event_destroy(struct perf_event *event) |
5558 | { | |
5559 | u64 event_id = event->attr.config; | |
95476b64 | 5560 | |
b0a873eb PZ |
5561 | WARN_ON(event->parent); |
5562 | ||
c5905afb | 5563 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5564 | swevent_hlist_put(event); |
5565 | } | |
5566 | ||
5567 | static int perf_swevent_init(struct perf_event *event) | |
5568 | { | |
8176cced | 5569 | u64 event_id = event->attr.config; |
b0a873eb PZ |
5570 | |
5571 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5572 | return -ENOENT; | |
5573 | ||
2481c5fa SE |
5574 | /* |
5575 | * no branch sampling for software events | |
5576 | */ | |
5577 | if (has_branch_stack(event)) | |
5578 | return -EOPNOTSUPP; | |
5579 | ||
b0a873eb PZ |
5580 | switch (event_id) { |
5581 | case PERF_COUNT_SW_CPU_CLOCK: | |
5582 | case PERF_COUNT_SW_TASK_CLOCK: | |
5583 | return -ENOENT; | |
5584 | ||
5585 | default: | |
5586 | break; | |
5587 | } | |
5588 | ||
ce677831 | 5589 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5590 | return -ENOENT; |
5591 | ||
5592 | if (!event->parent) { | |
5593 | int err; | |
5594 | ||
5595 | err = swevent_hlist_get(event); | |
5596 | if (err) | |
5597 | return err; | |
5598 | ||
c5905afb | 5599 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5600 | event->destroy = sw_perf_event_destroy; |
5601 | } | |
5602 | ||
5603 | return 0; | |
5604 | } | |
5605 | ||
35edc2a5 PZ |
5606 | static int perf_swevent_event_idx(struct perf_event *event) |
5607 | { | |
5608 | return 0; | |
5609 | } | |
5610 | ||
b0a873eb | 5611 | static struct pmu perf_swevent = { |
89a1e187 | 5612 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5613 | |
b0a873eb | 5614 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5615 | .add = perf_swevent_add, |
5616 | .del = perf_swevent_del, | |
5617 | .start = perf_swevent_start, | |
5618 | .stop = perf_swevent_stop, | |
1c024eca | 5619 | .read = perf_swevent_read, |
35edc2a5 PZ |
5620 | |
5621 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5622 | }; |
5623 | ||
b0a873eb PZ |
5624 | #ifdef CONFIG_EVENT_TRACING |
5625 | ||
1c024eca PZ |
5626 | static int perf_tp_filter_match(struct perf_event *event, |
5627 | struct perf_sample_data *data) | |
5628 | { | |
5629 | void *record = data->raw->data; | |
5630 | ||
5631 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5632 | return 1; | |
5633 | return 0; | |
5634 | } | |
5635 | ||
5636 | static int perf_tp_event_match(struct perf_event *event, | |
5637 | struct perf_sample_data *data, | |
5638 | struct pt_regs *regs) | |
5639 | { | |
a0f7d0f7 FW |
5640 | if (event->hw.state & PERF_HES_STOPPED) |
5641 | return 0; | |
580d607c PZ |
5642 | /* |
5643 | * All tracepoints are from kernel-space. | |
5644 | */ | |
5645 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5646 | return 0; |
5647 | ||
5648 | if (!perf_tp_filter_match(event, data)) | |
5649 | return 0; | |
5650 | ||
5651 | return 1; | |
5652 | } | |
5653 | ||
5654 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5655 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5656 | struct task_struct *task) | |
95476b64 FW |
5657 | { |
5658 | struct perf_sample_data data; | |
1c024eca | 5659 | struct perf_event *event; |
1c024eca | 5660 | |
95476b64 FW |
5661 | struct perf_raw_record raw = { |
5662 | .size = entry_size, | |
5663 | .data = record, | |
5664 | }; | |
5665 | ||
fd0d000b | 5666 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5667 | data.raw = &raw; |
5668 | ||
b67bfe0d | 5669 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 5670 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 5671 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5672 | } |
ecc55f84 | 5673 | |
e6dab5ff AV |
5674 | /* |
5675 | * If we got specified a target task, also iterate its context and | |
5676 | * deliver this event there too. | |
5677 | */ | |
5678 | if (task && task != current) { | |
5679 | struct perf_event_context *ctx; | |
5680 | struct trace_entry *entry = record; | |
5681 | ||
5682 | rcu_read_lock(); | |
5683 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5684 | if (!ctx) | |
5685 | goto unlock; | |
5686 | ||
5687 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5688 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5689 | continue; | |
5690 | if (event->attr.config != entry->type) | |
5691 | continue; | |
5692 | if (perf_tp_event_match(event, &data, regs)) | |
5693 | perf_swevent_event(event, count, &data, regs); | |
5694 | } | |
5695 | unlock: | |
5696 | rcu_read_unlock(); | |
5697 | } | |
5698 | ||
ecc55f84 | 5699 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5700 | } |
5701 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5702 | ||
cdd6c482 | 5703 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5704 | { |
1c024eca | 5705 | perf_trace_destroy(event); |
e077df4f PZ |
5706 | } |
5707 | ||
b0a873eb | 5708 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5709 | { |
76e1d904 FW |
5710 | int err; |
5711 | ||
b0a873eb PZ |
5712 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5713 | return -ENOENT; | |
5714 | ||
2481c5fa SE |
5715 | /* |
5716 | * no branch sampling for tracepoint events | |
5717 | */ | |
5718 | if (has_branch_stack(event)) | |
5719 | return -EOPNOTSUPP; | |
5720 | ||
1c024eca PZ |
5721 | err = perf_trace_init(event); |
5722 | if (err) | |
b0a873eb | 5723 | return err; |
e077df4f | 5724 | |
cdd6c482 | 5725 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5726 | |
b0a873eb PZ |
5727 | return 0; |
5728 | } | |
5729 | ||
5730 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5731 | .task_ctx_nr = perf_sw_context, |
5732 | ||
b0a873eb | 5733 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5734 | .add = perf_trace_add, |
5735 | .del = perf_trace_del, | |
5736 | .start = perf_swevent_start, | |
5737 | .stop = perf_swevent_stop, | |
b0a873eb | 5738 | .read = perf_swevent_read, |
35edc2a5 PZ |
5739 | |
5740 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5741 | }; |
5742 | ||
5743 | static inline void perf_tp_register(void) | |
5744 | { | |
2e80a82a | 5745 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5746 | } |
6fb2915d LZ |
5747 | |
5748 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5749 | { | |
5750 | char *filter_str; | |
5751 | int ret; | |
5752 | ||
5753 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5754 | return -EINVAL; | |
5755 | ||
5756 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5757 | if (IS_ERR(filter_str)) | |
5758 | return PTR_ERR(filter_str); | |
5759 | ||
5760 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5761 | ||
5762 | kfree(filter_str); | |
5763 | return ret; | |
5764 | } | |
5765 | ||
5766 | static void perf_event_free_filter(struct perf_event *event) | |
5767 | { | |
5768 | ftrace_profile_free_filter(event); | |
5769 | } | |
5770 | ||
e077df4f | 5771 | #else |
6fb2915d | 5772 | |
b0a873eb | 5773 | static inline void perf_tp_register(void) |
e077df4f | 5774 | { |
e077df4f | 5775 | } |
6fb2915d LZ |
5776 | |
5777 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5778 | { | |
5779 | return -ENOENT; | |
5780 | } | |
5781 | ||
5782 | static void perf_event_free_filter(struct perf_event *event) | |
5783 | { | |
5784 | } | |
5785 | ||
07b139c8 | 5786 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5787 | |
24f1e32c | 5788 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5789 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5790 | { |
f5ffe02e FW |
5791 | struct perf_sample_data sample; |
5792 | struct pt_regs *regs = data; | |
5793 | ||
fd0d000b | 5794 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 5795 | |
a4eaf7f1 | 5796 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5797 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5798 | } |
5799 | #endif | |
5800 | ||
b0a873eb PZ |
5801 | /* |
5802 | * hrtimer based swevent callback | |
5803 | */ | |
f29ac756 | 5804 | |
b0a873eb | 5805 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5806 | { |
b0a873eb PZ |
5807 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5808 | struct perf_sample_data data; | |
5809 | struct pt_regs *regs; | |
5810 | struct perf_event *event; | |
5811 | u64 period; | |
f29ac756 | 5812 | |
b0a873eb | 5813 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5814 | |
5815 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5816 | return HRTIMER_NORESTART; | |
5817 | ||
b0a873eb | 5818 | event->pmu->read(event); |
f344011c | 5819 | |
fd0d000b | 5820 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
5821 | regs = get_irq_regs(); |
5822 | ||
5823 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5824 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 5825 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
5826 | ret = HRTIMER_NORESTART; |
5827 | } | |
24f1e32c | 5828 | |
b0a873eb PZ |
5829 | period = max_t(u64, 10000, event->hw.sample_period); |
5830 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5831 | |
b0a873eb | 5832 | return ret; |
f29ac756 PZ |
5833 | } |
5834 | ||
b0a873eb | 5835 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5836 | { |
b0a873eb | 5837 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5838 | s64 period; |
5839 | ||
5840 | if (!is_sampling_event(event)) | |
5841 | return; | |
f5ffe02e | 5842 | |
5d508e82 FBH |
5843 | period = local64_read(&hwc->period_left); |
5844 | if (period) { | |
5845 | if (period < 0) | |
5846 | period = 10000; | |
fa407f35 | 5847 | |
5d508e82 FBH |
5848 | local64_set(&hwc->period_left, 0); |
5849 | } else { | |
5850 | period = max_t(u64, 10000, hwc->sample_period); | |
5851 | } | |
5852 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 5853 | ns_to_ktime(period), 0, |
b5ab4cd5 | 5854 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 5855 | } |
b0a873eb PZ |
5856 | |
5857 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 5858 | { |
b0a873eb PZ |
5859 | struct hw_perf_event *hwc = &event->hw; |
5860 | ||
6c7e550f | 5861 | if (is_sampling_event(event)) { |
b0a873eb | 5862 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 5863 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
5864 | |
5865 | hrtimer_cancel(&hwc->hrtimer); | |
5866 | } | |
24f1e32c FW |
5867 | } |
5868 | ||
ba3dd36c PZ |
5869 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5870 | { | |
5871 | struct hw_perf_event *hwc = &event->hw; | |
5872 | ||
5873 | if (!is_sampling_event(event)) | |
5874 | return; | |
5875 | ||
5876 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5877 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
5878 | ||
5879 | /* | |
5880 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
5881 | * mapping and avoid the whole period adjust feedback stuff. | |
5882 | */ | |
5883 | if (event->attr.freq) { | |
5884 | long freq = event->attr.sample_freq; | |
5885 | ||
5886 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
5887 | hwc->sample_period = event->attr.sample_period; | |
5888 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 5889 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
5890 | event->attr.freq = 0; |
5891 | } | |
5892 | } | |
5893 | ||
b0a873eb PZ |
5894 | /* |
5895 | * Software event: cpu wall time clock | |
5896 | */ | |
5897 | ||
5898 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 5899 | { |
b0a873eb PZ |
5900 | s64 prev; |
5901 | u64 now; | |
5902 | ||
a4eaf7f1 | 5903 | now = local_clock(); |
b0a873eb PZ |
5904 | prev = local64_xchg(&event->hw.prev_count, now); |
5905 | local64_add(now - prev, &event->count); | |
24f1e32c | 5906 | } |
24f1e32c | 5907 | |
a4eaf7f1 | 5908 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5909 | { |
a4eaf7f1 | 5910 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 5911 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5912 | } |
5913 | ||
a4eaf7f1 | 5914 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 5915 | { |
b0a873eb PZ |
5916 | perf_swevent_cancel_hrtimer(event); |
5917 | cpu_clock_event_update(event); | |
5918 | } | |
f29ac756 | 5919 | |
a4eaf7f1 PZ |
5920 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
5921 | { | |
5922 | if (flags & PERF_EF_START) | |
5923 | cpu_clock_event_start(event, flags); | |
5924 | ||
5925 | return 0; | |
5926 | } | |
5927 | ||
5928 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
5929 | { | |
5930 | cpu_clock_event_stop(event, flags); | |
5931 | } | |
5932 | ||
b0a873eb PZ |
5933 | static void cpu_clock_event_read(struct perf_event *event) |
5934 | { | |
5935 | cpu_clock_event_update(event); | |
5936 | } | |
f344011c | 5937 | |
b0a873eb PZ |
5938 | static int cpu_clock_event_init(struct perf_event *event) |
5939 | { | |
5940 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5941 | return -ENOENT; | |
5942 | ||
5943 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
5944 | return -ENOENT; | |
5945 | ||
2481c5fa SE |
5946 | /* |
5947 | * no branch sampling for software events | |
5948 | */ | |
5949 | if (has_branch_stack(event)) | |
5950 | return -EOPNOTSUPP; | |
5951 | ||
ba3dd36c PZ |
5952 | perf_swevent_init_hrtimer(event); |
5953 | ||
b0a873eb | 5954 | return 0; |
f29ac756 PZ |
5955 | } |
5956 | ||
b0a873eb | 5957 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
5958 | .task_ctx_nr = perf_sw_context, |
5959 | ||
b0a873eb | 5960 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
5961 | .add = cpu_clock_event_add, |
5962 | .del = cpu_clock_event_del, | |
5963 | .start = cpu_clock_event_start, | |
5964 | .stop = cpu_clock_event_stop, | |
b0a873eb | 5965 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
5966 | |
5967 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5968 | }; |
5969 | ||
5970 | /* | |
5971 | * Software event: task time clock | |
5972 | */ | |
5973 | ||
5974 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 5975 | { |
b0a873eb PZ |
5976 | u64 prev; |
5977 | s64 delta; | |
5c92d124 | 5978 | |
b0a873eb PZ |
5979 | prev = local64_xchg(&event->hw.prev_count, now); |
5980 | delta = now - prev; | |
5981 | local64_add(delta, &event->count); | |
5982 | } | |
5c92d124 | 5983 | |
a4eaf7f1 | 5984 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5985 | { |
a4eaf7f1 | 5986 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 5987 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5988 | } |
5989 | ||
a4eaf7f1 | 5990 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
5991 | { |
5992 | perf_swevent_cancel_hrtimer(event); | |
5993 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
5994 | } |
5995 | ||
5996 | static int task_clock_event_add(struct perf_event *event, int flags) | |
5997 | { | |
5998 | if (flags & PERF_EF_START) | |
5999 | task_clock_event_start(event, flags); | |
b0a873eb | 6000 | |
a4eaf7f1 PZ |
6001 | return 0; |
6002 | } | |
6003 | ||
6004 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6005 | { | |
6006 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6007 | } |
6008 | ||
6009 | static void task_clock_event_read(struct perf_event *event) | |
6010 | { | |
768a06e2 PZ |
6011 | u64 now = perf_clock(); |
6012 | u64 delta = now - event->ctx->timestamp; | |
6013 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6014 | |
6015 | task_clock_event_update(event, time); | |
6016 | } | |
6017 | ||
6018 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6019 | { |
b0a873eb PZ |
6020 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6021 | return -ENOENT; | |
6022 | ||
6023 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6024 | return -ENOENT; | |
6025 | ||
2481c5fa SE |
6026 | /* |
6027 | * no branch sampling for software events | |
6028 | */ | |
6029 | if (has_branch_stack(event)) | |
6030 | return -EOPNOTSUPP; | |
6031 | ||
ba3dd36c PZ |
6032 | perf_swevent_init_hrtimer(event); |
6033 | ||
b0a873eb | 6034 | return 0; |
6fb2915d LZ |
6035 | } |
6036 | ||
b0a873eb | 6037 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6038 | .task_ctx_nr = perf_sw_context, |
6039 | ||
b0a873eb | 6040 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6041 | .add = task_clock_event_add, |
6042 | .del = task_clock_event_del, | |
6043 | .start = task_clock_event_start, | |
6044 | .stop = task_clock_event_stop, | |
b0a873eb | 6045 | .read = task_clock_event_read, |
35edc2a5 PZ |
6046 | |
6047 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 6048 | }; |
6fb2915d | 6049 | |
ad5133b7 | 6050 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6051 | { |
e077df4f | 6052 | } |
6fb2915d | 6053 | |
ad5133b7 | 6054 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6055 | { |
ad5133b7 | 6056 | return 0; |
6fb2915d LZ |
6057 | } |
6058 | ||
ad5133b7 | 6059 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6060 | { |
ad5133b7 | 6061 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6062 | } |
6063 | ||
ad5133b7 PZ |
6064 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6065 | { | |
6066 | perf_pmu_enable(pmu); | |
6067 | return 0; | |
6068 | } | |
e077df4f | 6069 | |
ad5133b7 | 6070 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6071 | { |
ad5133b7 | 6072 | perf_pmu_enable(pmu); |
24f1e32c FW |
6073 | } |
6074 | ||
35edc2a5 PZ |
6075 | static int perf_event_idx_default(struct perf_event *event) |
6076 | { | |
6077 | return event->hw.idx + 1; | |
6078 | } | |
6079 | ||
8dc85d54 PZ |
6080 | /* |
6081 | * Ensures all contexts with the same task_ctx_nr have the same | |
6082 | * pmu_cpu_context too. | |
6083 | */ | |
6084 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 6085 | { |
8dc85d54 | 6086 | struct pmu *pmu; |
b326e956 | 6087 | |
8dc85d54 PZ |
6088 | if (ctxn < 0) |
6089 | return NULL; | |
24f1e32c | 6090 | |
8dc85d54 PZ |
6091 | list_for_each_entry(pmu, &pmus, entry) { |
6092 | if (pmu->task_ctx_nr == ctxn) | |
6093 | return pmu->pmu_cpu_context; | |
6094 | } | |
24f1e32c | 6095 | |
8dc85d54 | 6096 | return NULL; |
24f1e32c FW |
6097 | } |
6098 | ||
51676957 | 6099 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6100 | { |
51676957 PZ |
6101 | int cpu; |
6102 | ||
6103 | for_each_possible_cpu(cpu) { | |
6104 | struct perf_cpu_context *cpuctx; | |
6105 | ||
6106 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6107 | ||
3f1f3320 PZ |
6108 | if (cpuctx->unique_pmu == old_pmu) |
6109 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6110 | } |
6111 | } | |
6112 | ||
6113 | static void free_pmu_context(struct pmu *pmu) | |
6114 | { | |
6115 | struct pmu *i; | |
f5ffe02e | 6116 | |
8dc85d54 | 6117 | mutex_lock(&pmus_lock); |
0475f9ea | 6118 | /* |
8dc85d54 | 6119 | * Like a real lame refcount. |
0475f9ea | 6120 | */ |
51676957 PZ |
6121 | list_for_each_entry(i, &pmus, entry) { |
6122 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6123 | update_pmu_context(i, pmu); | |
8dc85d54 | 6124 | goto out; |
51676957 | 6125 | } |
8dc85d54 | 6126 | } |
d6d020e9 | 6127 | |
51676957 | 6128 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6129 | out: |
6130 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6131 | } |
2e80a82a | 6132 | static struct idr pmu_idr; |
d6d020e9 | 6133 | |
abe43400 PZ |
6134 | static ssize_t |
6135 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6136 | { | |
6137 | struct pmu *pmu = dev_get_drvdata(dev); | |
6138 | ||
6139 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6140 | } | |
6141 | ||
6142 | static struct device_attribute pmu_dev_attrs[] = { | |
6143 | __ATTR_RO(type), | |
6144 | __ATTR_NULL, | |
6145 | }; | |
6146 | ||
6147 | static int pmu_bus_running; | |
6148 | static struct bus_type pmu_bus = { | |
6149 | .name = "event_source", | |
6150 | .dev_attrs = pmu_dev_attrs, | |
6151 | }; | |
6152 | ||
6153 | static void pmu_dev_release(struct device *dev) | |
6154 | { | |
6155 | kfree(dev); | |
6156 | } | |
6157 | ||
6158 | static int pmu_dev_alloc(struct pmu *pmu) | |
6159 | { | |
6160 | int ret = -ENOMEM; | |
6161 | ||
6162 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6163 | if (!pmu->dev) | |
6164 | goto out; | |
6165 | ||
0c9d42ed | 6166 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6167 | device_initialize(pmu->dev); |
6168 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6169 | if (ret) | |
6170 | goto free_dev; | |
6171 | ||
6172 | dev_set_drvdata(pmu->dev, pmu); | |
6173 | pmu->dev->bus = &pmu_bus; | |
6174 | pmu->dev->release = pmu_dev_release; | |
6175 | ret = device_add(pmu->dev); | |
6176 | if (ret) | |
6177 | goto free_dev; | |
6178 | ||
6179 | out: | |
6180 | return ret; | |
6181 | ||
6182 | free_dev: | |
6183 | put_device(pmu->dev); | |
6184 | goto out; | |
6185 | } | |
6186 | ||
547e9fd7 | 6187 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6188 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6189 | |
2e80a82a | 6190 | int perf_pmu_register(struct pmu *pmu, char *name, int type) |
24f1e32c | 6191 | { |
108b02cf | 6192 | int cpu, ret; |
24f1e32c | 6193 | |
b0a873eb | 6194 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6195 | ret = -ENOMEM; |
6196 | pmu->pmu_disable_count = alloc_percpu(int); | |
6197 | if (!pmu->pmu_disable_count) | |
6198 | goto unlock; | |
f29ac756 | 6199 | |
2e80a82a PZ |
6200 | pmu->type = -1; |
6201 | if (!name) | |
6202 | goto skip_type; | |
6203 | pmu->name = name; | |
6204 | ||
6205 | if (type < 0) { | |
0e9c3be2 TH |
6206 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6207 | if (type < 0) { | |
6208 | ret = type; | |
2e80a82a PZ |
6209 | goto free_pdc; |
6210 | } | |
6211 | } | |
6212 | pmu->type = type; | |
6213 | ||
abe43400 PZ |
6214 | if (pmu_bus_running) { |
6215 | ret = pmu_dev_alloc(pmu); | |
6216 | if (ret) | |
6217 | goto free_idr; | |
6218 | } | |
6219 | ||
2e80a82a | 6220 | skip_type: |
8dc85d54 PZ |
6221 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6222 | if (pmu->pmu_cpu_context) | |
6223 | goto got_cpu_context; | |
f29ac756 | 6224 | |
c4814202 | 6225 | ret = -ENOMEM; |
108b02cf PZ |
6226 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6227 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6228 | goto free_dev; |
f344011c | 6229 | |
108b02cf PZ |
6230 | for_each_possible_cpu(cpu) { |
6231 | struct perf_cpu_context *cpuctx; | |
6232 | ||
6233 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6234 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6235 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6236 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6237 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6238 | cpuctx->ctx.pmu = pmu; |
e9d2b064 PZ |
6239 | cpuctx->jiffies_interval = 1; |
6240 | INIT_LIST_HEAD(&cpuctx->rotation_list); | |
3f1f3320 | 6241 | cpuctx->unique_pmu = pmu; |
108b02cf | 6242 | } |
76e1d904 | 6243 | |
8dc85d54 | 6244 | got_cpu_context: |
ad5133b7 PZ |
6245 | if (!pmu->start_txn) { |
6246 | if (pmu->pmu_enable) { | |
6247 | /* | |
6248 | * If we have pmu_enable/pmu_disable calls, install | |
6249 | * transaction stubs that use that to try and batch | |
6250 | * hardware accesses. | |
6251 | */ | |
6252 | pmu->start_txn = perf_pmu_start_txn; | |
6253 | pmu->commit_txn = perf_pmu_commit_txn; | |
6254 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6255 | } else { | |
6256 | pmu->start_txn = perf_pmu_nop_void; | |
6257 | pmu->commit_txn = perf_pmu_nop_int; | |
6258 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6259 | } |
5c92d124 | 6260 | } |
15dbf27c | 6261 | |
ad5133b7 PZ |
6262 | if (!pmu->pmu_enable) { |
6263 | pmu->pmu_enable = perf_pmu_nop_void; | |
6264 | pmu->pmu_disable = perf_pmu_nop_void; | |
6265 | } | |
6266 | ||
35edc2a5 PZ |
6267 | if (!pmu->event_idx) |
6268 | pmu->event_idx = perf_event_idx_default; | |
6269 | ||
b0a873eb | 6270 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6271 | ret = 0; |
6272 | unlock: | |
b0a873eb PZ |
6273 | mutex_unlock(&pmus_lock); |
6274 | ||
33696fc0 | 6275 | return ret; |
108b02cf | 6276 | |
abe43400 PZ |
6277 | free_dev: |
6278 | device_del(pmu->dev); | |
6279 | put_device(pmu->dev); | |
6280 | ||
2e80a82a PZ |
6281 | free_idr: |
6282 | if (pmu->type >= PERF_TYPE_MAX) | |
6283 | idr_remove(&pmu_idr, pmu->type); | |
6284 | ||
108b02cf PZ |
6285 | free_pdc: |
6286 | free_percpu(pmu->pmu_disable_count); | |
6287 | goto unlock; | |
f29ac756 PZ |
6288 | } |
6289 | ||
b0a873eb | 6290 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6291 | { |
b0a873eb PZ |
6292 | mutex_lock(&pmus_lock); |
6293 | list_del_rcu(&pmu->entry); | |
6294 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6295 | |
0475f9ea | 6296 | /* |
cde8e884 PZ |
6297 | * We dereference the pmu list under both SRCU and regular RCU, so |
6298 | * synchronize against both of those. | |
0475f9ea | 6299 | */ |
b0a873eb | 6300 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6301 | synchronize_rcu(); |
d6d020e9 | 6302 | |
33696fc0 | 6303 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6304 | if (pmu->type >= PERF_TYPE_MAX) |
6305 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6306 | device_del(pmu->dev); |
6307 | put_device(pmu->dev); | |
51676957 | 6308 | free_pmu_context(pmu); |
b0a873eb | 6309 | } |
d6d020e9 | 6310 | |
b0a873eb PZ |
6311 | struct pmu *perf_init_event(struct perf_event *event) |
6312 | { | |
6313 | struct pmu *pmu = NULL; | |
6314 | int idx; | |
940c5b29 | 6315 | int ret; |
b0a873eb PZ |
6316 | |
6317 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6318 | |
6319 | rcu_read_lock(); | |
6320 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6321 | rcu_read_unlock(); | |
940c5b29 | 6322 | if (pmu) { |
7e5b2a01 | 6323 | event->pmu = pmu; |
940c5b29 LM |
6324 | ret = pmu->event_init(event); |
6325 | if (ret) | |
6326 | pmu = ERR_PTR(ret); | |
2e80a82a | 6327 | goto unlock; |
940c5b29 | 6328 | } |
2e80a82a | 6329 | |
b0a873eb | 6330 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 6331 | event->pmu = pmu; |
940c5b29 | 6332 | ret = pmu->event_init(event); |
b0a873eb | 6333 | if (!ret) |
e5f4d339 | 6334 | goto unlock; |
76e1d904 | 6335 | |
b0a873eb PZ |
6336 | if (ret != -ENOENT) { |
6337 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6338 | goto unlock; |
f344011c | 6339 | } |
5c92d124 | 6340 | } |
e5f4d339 PZ |
6341 | pmu = ERR_PTR(-ENOENT); |
6342 | unlock: | |
b0a873eb | 6343 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6344 | |
4aeb0b42 | 6345 | return pmu; |
5c92d124 IM |
6346 | } |
6347 | ||
0793a61d | 6348 | /* |
cdd6c482 | 6349 | * Allocate and initialize a event structure |
0793a61d | 6350 | */ |
cdd6c482 | 6351 | static struct perf_event * |
c3f00c70 | 6352 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6353 | struct task_struct *task, |
6354 | struct perf_event *group_leader, | |
6355 | struct perf_event *parent_event, | |
4dc0da86 AK |
6356 | perf_overflow_handler_t overflow_handler, |
6357 | void *context) | |
0793a61d | 6358 | { |
51b0fe39 | 6359 | struct pmu *pmu; |
cdd6c482 IM |
6360 | struct perf_event *event; |
6361 | struct hw_perf_event *hwc; | |
d5d2bc0d | 6362 | long err; |
0793a61d | 6363 | |
66832eb4 ON |
6364 | if ((unsigned)cpu >= nr_cpu_ids) { |
6365 | if (!task || cpu != -1) | |
6366 | return ERR_PTR(-EINVAL); | |
6367 | } | |
6368 | ||
c3f00c70 | 6369 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6370 | if (!event) |
d5d2bc0d | 6371 | return ERR_PTR(-ENOMEM); |
0793a61d | 6372 | |
04289bb9 | 6373 | /* |
cdd6c482 | 6374 | * Single events are their own group leaders, with an |
04289bb9 IM |
6375 | * empty sibling list: |
6376 | */ | |
6377 | if (!group_leader) | |
cdd6c482 | 6378 | group_leader = event; |
04289bb9 | 6379 | |
cdd6c482 IM |
6380 | mutex_init(&event->child_mutex); |
6381 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6382 | |
cdd6c482 IM |
6383 | INIT_LIST_HEAD(&event->group_entry); |
6384 | INIT_LIST_HEAD(&event->event_entry); | |
6385 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
6386 | INIT_LIST_HEAD(&event->rb_entry); |
6387 | ||
cdd6c482 | 6388 | init_waitqueue_head(&event->waitq); |
e360adbe | 6389 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6390 | |
cdd6c482 | 6391 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6392 | |
a6fa941d | 6393 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6394 | event->cpu = cpu; |
6395 | event->attr = *attr; | |
6396 | event->group_leader = group_leader; | |
6397 | event->pmu = NULL; | |
cdd6c482 | 6398 | event->oncpu = -1; |
a96bbc16 | 6399 | |
cdd6c482 | 6400 | event->parent = parent_event; |
b84fbc9f | 6401 | |
17cf22c3 | 6402 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6403 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6404 | |
cdd6c482 | 6405 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6406 | |
d580ff86 PZ |
6407 | if (task) { |
6408 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6409 | |
6410 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6411 | event->hw.tp_target = task; | |
d580ff86 PZ |
6412 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6413 | /* | |
6414 | * hw_breakpoint is a bit difficult here.. | |
6415 | */ | |
f22c1bb6 | 6416 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6417 | event->hw.bp_target = task; |
6418 | #endif | |
6419 | } | |
6420 | ||
4dc0da86 | 6421 | if (!overflow_handler && parent_event) { |
b326e956 | 6422 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6423 | context = parent_event->overflow_handler_context; |
6424 | } | |
66832eb4 | 6425 | |
b326e956 | 6426 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6427 | event->overflow_handler_context = context; |
97eaf530 | 6428 | |
0231bb53 | 6429 | perf_event__state_init(event); |
a86ed508 | 6430 | |
4aeb0b42 | 6431 | pmu = NULL; |
b8e83514 | 6432 | |
cdd6c482 | 6433 | hwc = &event->hw; |
bd2b5b12 | 6434 | hwc->sample_period = attr->sample_period; |
0d48696f | 6435 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6436 | hwc->sample_period = 1; |
eced1dfc | 6437 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6438 | |
e7850595 | 6439 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6440 | |
2023b359 | 6441 | /* |
cdd6c482 | 6442 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6443 | */ |
3dab77fb | 6444 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
2023b359 PZ |
6445 | goto done; |
6446 | ||
b0a873eb | 6447 | pmu = perf_init_event(event); |
974802ea | 6448 | |
d5d2bc0d PM |
6449 | done: |
6450 | err = 0; | |
4aeb0b42 | 6451 | if (!pmu) |
d5d2bc0d | 6452 | err = -EINVAL; |
4aeb0b42 RR |
6453 | else if (IS_ERR(pmu)) |
6454 | err = PTR_ERR(pmu); | |
5c92d124 | 6455 | |
d5d2bc0d | 6456 | if (err) { |
cdd6c482 IM |
6457 | if (event->ns) |
6458 | put_pid_ns(event->ns); | |
6459 | kfree(event); | |
d5d2bc0d | 6460 | return ERR_PTR(err); |
621a01ea | 6461 | } |
d5d2bc0d | 6462 | |
cdd6c482 | 6463 | if (!event->parent) { |
82cd6def | 6464 | if (event->attach_state & PERF_ATTACH_TASK) |
c5905afb | 6465 | static_key_slow_inc(&perf_sched_events.key); |
3af9e859 | 6466 | if (event->attr.mmap || event->attr.mmap_data) |
cdd6c482 IM |
6467 | atomic_inc(&nr_mmap_events); |
6468 | if (event->attr.comm) | |
6469 | atomic_inc(&nr_comm_events); | |
6470 | if (event->attr.task) | |
6471 | atomic_inc(&nr_task_events); | |
927c7a9e FW |
6472 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6473 | err = get_callchain_buffers(); | |
6474 | if (err) { | |
6475 | free_event(event); | |
6476 | return ERR_PTR(err); | |
6477 | } | |
6478 | } | |
d010b332 SE |
6479 | if (has_branch_stack(event)) { |
6480 | static_key_slow_inc(&perf_sched_events.key); | |
6481 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6482 | atomic_inc(&per_cpu(perf_branch_stack_events, | |
6483 | event->cpu)); | |
6484 | } | |
f344011c | 6485 | } |
9ee318a7 | 6486 | |
cdd6c482 | 6487 | return event; |
0793a61d TG |
6488 | } |
6489 | ||
cdd6c482 IM |
6490 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6491 | struct perf_event_attr *attr) | |
974802ea | 6492 | { |
974802ea | 6493 | u32 size; |
cdf8073d | 6494 | int ret; |
974802ea PZ |
6495 | |
6496 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6497 | return -EFAULT; | |
6498 | ||
6499 | /* | |
6500 | * zero the full structure, so that a short copy will be nice. | |
6501 | */ | |
6502 | memset(attr, 0, sizeof(*attr)); | |
6503 | ||
6504 | ret = get_user(size, &uattr->size); | |
6505 | if (ret) | |
6506 | return ret; | |
6507 | ||
6508 | if (size > PAGE_SIZE) /* silly large */ | |
6509 | goto err_size; | |
6510 | ||
6511 | if (!size) /* abi compat */ | |
6512 | size = PERF_ATTR_SIZE_VER0; | |
6513 | ||
6514 | if (size < PERF_ATTR_SIZE_VER0) | |
6515 | goto err_size; | |
6516 | ||
6517 | /* | |
6518 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6519 | * ensure all the unknown bits are 0 - i.e. new |
6520 | * user-space does not rely on any kernel feature | |
6521 | * extensions we dont know about yet. | |
974802ea PZ |
6522 | */ |
6523 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6524 | unsigned char __user *addr; |
6525 | unsigned char __user *end; | |
6526 | unsigned char val; | |
974802ea | 6527 | |
cdf8073d IS |
6528 | addr = (void __user *)uattr + sizeof(*attr); |
6529 | end = (void __user *)uattr + size; | |
974802ea | 6530 | |
cdf8073d | 6531 | for (; addr < end; addr++) { |
974802ea PZ |
6532 | ret = get_user(val, addr); |
6533 | if (ret) | |
6534 | return ret; | |
6535 | if (val) | |
6536 | goto err_size; | |
6537 | } | |
b3e62e35 | 6538 | size = sizeof(*attr); |
974802ea PZ |
6539 | } |
6540 | ||
6541 | ret = copy_from_user(attr, uattr, size); | |
6542 | if (ret) | |
6543 | return -EFAULT; | |
6544 | ||
cd757645 | 6545 | if (attr->__reserved_1) |
974802ea PZ |
6546 | return -EINVAL; |
6547 | ||
6548 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6549 | return -EINVAL; | |
6550 | ||
6551 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6552 | return -EINVAL; | |
6553 | ||
bce38cd5 SE |
6554 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6555 | u64 mask = attr->branch_sample_type; | |
6556 | ||
6557 | /* only using defined bits */ | |
6558 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6559 | return -EINVAL; | |
6560 | ||
6561 | /* at least one branch bit must be set */ | |
6562 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6563 | return -EINVAL; | |
6564 | ||
6565 | /* kernel level capture: check permissions */ | |
6566 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
6567 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6568 | return -EACCES; | |
6569 | ||
6570 | /* propagate priv level, when not set for branch */ | |
6571 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6572 | ||
6573 | /* exclude_kernel checked on syscall entry */ | |
6574 | if (!attr->exclude_kernel) | |
6575 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6576 | ||
6577 | if (!attr->exclude_user) | |
6578 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6579 | ||
6580 | if (!attr->exclude_hv) | |
6581 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6582 | /* | |
6583 | * adjust user setting (for HW filter setup) | |
6584 | */ | |
6585 | attr->branch_sample_type = mask; | |
6586 | } | |
6587 | } | |
4018994f | 6588 | |
c5ebcedb | 6589 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 6590 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
6591 | if (ret) |
6592 | return ret; | |
6593 | } | |
6594 | ||
6595 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
6596 | if (!arch_perf_have_user_stack_dump()) | |
6597 | return -ENOSYS; | |
6598 | ||
6599 | /* | |
6600 | * We have __u32 type for the size, but so far | |
6601 | * we can only use __u16 as maximum due to the | |
6602 | * __u16 sample size limit. | |
6603 | */ | |
6604 | if (attr->sample_stack_user >= USHRT_MAX) | |
6605 | ret = -EINVAL; | |
6606 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
6607 | ret = -EINVAL; | |
6608 | } | |
4018994f | 6609 | |
974802ea PZ |
6610 | out: |
6611 | return ret; | |
6612 | ||
6613 | err_size: | |
6614 | put_user(sizeof(*attr), &uattr->size); | |
6615 | ret = -E2BIG; | |
6616 | goto out; | |
6617 | } | |
6618 | ||
ac9721f3 PZ |
6619 | static int |
6620 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6621 | { |
76369139 | 6622 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6623 | int ret = -EINVAL; |
6624 | ||
ac9721f3 | 6625 | if (!output_event) |
a4be7c27 PZ |
6626 | goto set; |
6627 | ||
ac9721f3 PZ |
6628 | /* don't allow circular references */ |
6629 | if (event == output_event) | |
a4be7c27 PZ |
6630 | goto out; |
6631 | ||
0f139300 PZ |
6632 | /* |
6633 | * Don't allow cross-cpu buffers | |
6634 | */ | |
6635 | if (output_event->cpu != event->cpu) | |
6636 | goto out; | |
6637 | ||
6638 | /* | |
76369139 | 6639 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6640 | */ |
6641 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6642 | goto out; | |
6643 | ||
a4be7c27 | 6644 | set: |
cdd6c482 | 6645 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6646 | /* Can't redirect output if we've got an active mmap() */ |
6647 | if (atomic_read(&event->mmap_count)) | |
6648 | goto unlock; | |
a4be7c27 | 6649 | |
9bb5d40c PZ |
6650 | old_rb = event->rb; |
6651 | ||
ac9721f3 | 6652 | if (output_event) { |
76369139 FW |
6653 | /* get the rb we want to redirect to */ |
6654 | rb = ring_buffer_get(output_event); | |
6655 | if (!rb) | |
ac9721f3 | 6656 | goto unlock; |
a4be7c27 PZ |
6657 | } |
6658 | ||
10c6db11 PZ |
6659 | if (old_rb) |
6660 | ring_buffer_detach(event, old_rb); | |
9bb5d40c PZ |
6661 | |
6662 | if (rb) | |
6663 | ring_buffer_attach(event, rb); | |
6664 | ||
6665 | rcu_assign_pointer(event->rb, rb); | |
6666 | ||
6667 | if (old_rb) { | |
6668 | ring_buffer_put(old_rb); | |
6669 | /* | |
6670 | * Since we detached before setting the new rb, so that we | |
6671 | * could attach the new rb, we could have missed a wakeup. | |
6672 | * Provide it now. | |
6673 | */ | |
6674 | wake_up_all(&event->waitq); | |
6675 | } | |
6676 | ||
a4be7c27 | 6677 | ret = 0; |
ac9721f3 PZ |
6678 | unlock: |
6679 | mutex_unlock(&event->mmap_mutex); | |
6680 | ||
a4be7c27 | 6681 | out: |
a4be7c27 PZ |
6682 | return ret; |
6683 | } | |
6684 | ||
0793a61d | 6685 | /** |
cdd6c482 | 6686 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6687 | * |
cdd6c482 | 6688 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6689 | * @pid: target pid |
9f66a381 | 6690 | * @cpu: target cpu |
cdd6c482 | 6691 | * @group_fd: group leader event fd |
0793a61d | 6692 | */ |
cdd6c482 IM |
6693 | SYSCALL_DEFINE5(perf_event_open, |
6694 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 6695 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 6696 | { |
b04243ef PZ |
6697 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6698 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
6699 | struct perf_event_attr attr; |
6700 | struct perf_event_context *ctx; | |
6701 | struct file *event_file = NULL; | |
2903ff01 | 6702 | struct fd group = {NULL, 0}; |
38a81da2 | 6703 | struct task_struct *task = NULL; |
89a1e187 | 6704 | struct pmu *pmu; |
ea635c64 | 6705 | int event_fd; |
b04243ef | 6706 | int move_group = 0; |
dc86cabe | 6707 | int err; |
0793a61d | 6708 | |
2743a5b0 | 6709 | /* for future expandability... */ |
e5d1367f | 6710 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
6711 | return -EINVAL; |
6712 | ||
dc86cabe IM |
6713 | err = perf_copy_attr(attr_uptr, &attr); |
6714 | if (err) | |
6715 | return err; | |
eab656ae | 6716 | |
0764771d PZ |
6717 | if (!attr.exclude_kernel) { |
6718 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6719 | return -EACCES; | |
6720 | } | |
6721 | ||
df58ab24 | 6722 | if (attr.freq) { |
cdd6c482 | 6723 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 | 6724 | return -EINVAL; |
95090e8a PZ |
6725 | } else { |
6726 | if (attr.sample_period & (1ULL << 63)) | |
6727 | return -EINVAL; | |
df58ab24 PZ |
6728 | } |
6729 | ||
e5d1367f SE |
6730 | /* |
6731 | * In cgroup mode, the pid argument is used to pass the fd | |
6732 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6733 | * designates the cpu on which to monitor threads from that | |
6734 | * cgroup. | |
6735 | */ | |
6736 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6737 | return -EINVAL; | |
6738 | ||
ab72a702 | 6739 | event_fd = get_unused_fd(); |
ea635c64 AV |
6740 | if (event_fd < 0) |
6741 | return event_fd; | |
6742 | ||
ac9721f3 | 6743 | if (group_fd != -1) { |
2903ff01 AV |
6744 | err = perf_fget_light(group_fd, &group); |
6745 | if (err) | |
d14b12d7 | 6746 | goto err_fd; |
2903ff01 | 6747 | group_leader = group.file->private_data; |
ac9721f3 PZ |
6748 | if (flags & PERF_FLAG_FD_OUTPUT) |
6749 | output_event = group_leader; | |
6750 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6751 | group_leader = NULL; | |
6752 | } | |
6753 | ||
e5d1367f | 6754 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6755 | task = find_lively_task_by_vpid(pid); |
6756 | if (IS_ERR(task)) { | |
6757 | err = PTR_ERR(task); | |
6758 | goto err_group_fd; | |
6759 | } | |
6760 | } | |
6761 | ||
fbfc623f YZ |
6762 | get_online_cpus(); |
6763 | ||
4dc0da86 AK |
6764 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6765 | NULL, NULL); | |
d14b12d7 SE |
6766 | if (IS_ERR(event)) { |
6767 | err = PTR_ERR(event); | |
c6be5a5c | 6768 | goto err_task; |
d14b12d7 SE |
6769 | } |
6770 | ||
e5d1367f SE |
6771 | if (flags & PERF_FLAG_PID_CGROUP) { |
6772 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
6773 | if (err) | |
6774 | goto err_alloc; | |
08309379 PZ |
6775 | /* |
6776 | * one more event: | |
6777 | * - that has cgroup constraint on event->cpu | |
6778 | * - that may need work on context switch | |
6779 | */ | |
6780 | atomic_inc(&per_cpu(perf_cgroup_events, event->cpu)); | |
c5905afb | 6781 | static_key_slow_inc(&perf_sched_events.key); |
e5d1367f SE |
6782 | } |
6783 | ||
89a1e187 PZ |
6784 | /* |
6785 | * Special case software events and allow them to be part of | |
6786 | * any hardware group. | |
6787 | */ | |
6788 | pmu = event->pmu; | |
b04243ef PZ |
6789 | |
6790 | if (group_leader && | |
6791 | (is_software_event(event) != is_software_event(group_leader))) { | |
6792 | if (is_software_event(event)) { | |
6793 | /* | |
6794 | * If event and group_leader are not both a software | |
6795 | * event, and event is, then group leader is not. | |
6796 | * | |
6797 | * Allow the addition of software events to !software | |
6798 | * groups, this is safe because software events never | |
6799 | * fail to schedule. | |
6800 | */ | |
6801 | pmu = group_leader->pmu; | |
6802 | } else if (is_software_event(group_leader) && | |
6803 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
6804 | /* | |
6805 | * In case the group is a pure software group, and we | |
6806 | * try to add a hardware event, move the whole group to | |
6807 | * the hardware context. | |
6808 | */ | |
6809 | move_group = 1; | |
6810 | } | |
6811 | } | |
89a1e187 PZ |
6812 | |
6813 | /* | |
6814 | * Get the target context (task or percpu): | |
6815 | */ | |
e2d37cd2 | 6816 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
6817 | if (IS_ERR(ctx)) { |
6818 | err = PTR_ERR(ctx); | |
c6be5a5c | 6819 | goto err_alloc; |
89a1e187 PZ |
6820 | } |
6821 | ||
fd1edb3a PZ |
6822 | if (task) { |
6823 | put_task_struct(task); | |
6824 | task = NULL; | |
6825 | } | |
6826 | ||
ccff286d | 6827 | /* |
cdd6c482 | 6828 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 6829 | */ |
ac9721f3 | 6830 | if (group_leader) { |
dc86cabe | 6831 | err = -EINVAL; |
04289bb9 | 6832 | |
04289bb9 | 6833 | /* |
ccff286d IM |
6834 | * Do not allow a recursive hierarchy (this new sibling |
6835 | * becoming part of another group-sibling): | |
6836 | */ | |
6837 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 6838 | goto err_context; |
ccff286d IM |
6839 | /* |
6840 | * Do not allow to attach to a group in a different | |
6841 | * task or CPU context: | |
04289bb9 | 6842 | */ |
b04243ef PZ |
6843 | if (move_group) { |
6844 | if (group_leader->ctx->type != ctx->type) | |
6845 | goto err_context; | |
6846 | } else { | |
6847 | if (group_leader->ctx != ctx) | |
6848 | goto err_context; | |
6849 | } | |
6850 | ||
3b6f9e5c PM |
6851 | /* |
6852 | * Only a group leader can be exclusive or pinned | |
6853 | */ | |
0d48696f | 6854 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 6855 | goto err_context; |
ac9721f3 PZ |
6856 | } |
6857 | ||
6858 | if (output_event) { | |
6859 | err = perf_event_set_output(event, output_event); | |
6860 | if (err) | |
c3f00c70 | 6861 | goto err_context; |
ac9721f3 | 6862 | } |
0793a61d | 6863 | |
ea635c64 AV |
6864 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
6865 | if (IS_ERR(event_file)) { | |
6866 | err = PTR_ERR(event_file); | |
c3f00c70 | 6867 | goto err_context; |
ea635c64 | 6868 | } |
9b51f66d | 6869 | |
b04243ef PZ |
6870 | if (move_group) { |
6871 | struct perf_event_context *gctx = group_leader->ctx; | |
6872 | ||
6873 | mutex_lock(&gctx->mutex); | |
54b3f8df | 6874 | perf_remove_from_context(group_leader, false); |
0231bb53 JO |
6875 | |
6876 | /* | |
6877 | * Removing from the context ends up with disabled | |
6878 | * event. What we want here is event in the initial | |
6879 | * startup state, ready to be add into new context. | |
6880 | */ | |
6881 | perf_event__state_init(group_leader); | |
b04243ef PZ |
6882 | list_for_each_entry(sibling, &group_leader->sibling_list, |
6883 | group_entry) { | |
54b3f8df | 6884 | perf_remove_from_context(sibling, false); |
0231bb53 | 6885 | perf_event__state_init(sibling); |
b04243ef PZ |
6886 | put_ctx(gctx); |
6887 | } | |
6888 | mutex_unlock(&gctx->mutex); | |
6889 | put_ctx(gctx); | |
ea635c64 | 6890 | } |
9b51f66d | 6891 | |
ad3a37de | 6892 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 6893 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
6894 | |
6895 | if (move_group) { | |
0cda4c02 | 6896 | synchronize_rcu(); |
d525563b | 6897 | perf_install_in_context(ctx, group_leader, group_leader->cpu); |
b04243ef PZ |
6898 | get_ctx(ctx); |
6899 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
6900 | group_entry) { | |
d525563b | 6901 | perf_install_in_context(ctx, sibling, sibling->cpu); |
b04243ef PZ |
6902 | get_ctx(ctx); |
6903 | } | |
6904 | } | |
6905 | ||
e2d37cd2 | 6906 | perf_install_in_context(ctx, event, event->cpu); |
ad3a37de | 6907 | ++ctx->generation; |
fe4b04fa | 6908 | perf_unpin_context(ctx); |
d859e29f | 6909 | mutex_unlock(&ctx->mutex); |
9b51f66d | 6910 | |
fbfc623f YZ |
6911 | put_online_cpus(); |
6912 | ||
cdd6c482 | 6913 | event->owner = current; |
8882135b | 6914 | |
cdd6c482 IM |
6915 | mutex_lock(¤t->perf_event_mutex); |
6916 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
6917 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 6918 | |
c320c7b7 ACM |
6919 | /* |
6920 | * Precalculate sample_data sizes | |
6921 | */ | |
6922 | perf_event__header_size(event); | |
6844c09d | 6923 | perf_event__id_header_size(event); |
c320c7b7 | 6924 | |
8a49542c PZ |
6925 | /* |
6926 | * Drop the reference on the group_event after placing the | |
6927 | * new event on the sibling_list. This ensures destruction | |
6928 | * of the group leader will find the pointer to itself in | |
6929 | * perf_group_detach(). | |
6930 | */ | |
2903ff01 | 6931 | fdput(group); |
ea635c64 AV |
6932 | fd_install(event_fd, event_file); |
6933 | return event_fd; | |
0793a61d | 6934 | |
c3f00c70 | 6935 | err_context: |
fe4b04fa | 6936 | perf_unpin_context(ctx); |
ea635c64 | 6937 | put_ctx(ctx); |
c6be5a5c | 6938 | err_alloc: |
ea635c64 | 6939 | free_event(event); |
e7d0bc04 | 6940 | err_task: |
fbfc623f | 6941 | put_online_cpus(); |
e7d0bc04 PZ |
6942 | if (task) |
6943 | put_task_struct(task); | |
89a1e187 | 6944 | err_group_fd: |
2903ff01 | 6945 | fdput(group); |
ea635c64 AV |
6946 | err_fd: |
6947 | put_unused_fd(event_fd); | |
dc86cabe | 6948 | return err; |
0793a61d TG |
6949 | } |
6950 | ||
fb0459d7 AV |
6951 | /** |
6952 | * perf_event_create_kernel_counter | |
6953 | * | |
6954 | * @attr: attributes of the counter to create | |
6955 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 6956 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
6957 | */ |
6958 | struct perf_event * | |
6959 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 6960 | struct task_struct *task, |
4dc0da86 AK |
6961 | perf_overflow_handler_t overflow_handler, |
6962 | void *context) | |
fb0459d7 | 6963 | { |
fb0459d7 | 6964 | struct perf_event_context *ctx; |
c3f00c70 | 6965 | struct perf_event *event; |
fb0459d7 | 6966 | int err; |
d859e29f | 6967 | |
fb0459d7 AV |
6968 | /* |
6969 | * Get the target context (task or percpu): | |
6970 | */ | |
d859e29f | 6971 | |
4dc0da86 AK |
6972 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
6973 | overflow_handler, context); | |
c3f00c70 PZ |
6974 | if (IS_ERR(event)) { |
6975 | err = PTR_ERR(event); | |
6976 | goto err; | |
6977 | } | |
d859e29f | 6978 | |
38a81da2 | 6979 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
6980 | if (IS_ERR(ctx)) { |
6981 | err = PTR_ERR(ctx); | |
c3f00c70 | 6982 | goto err_free; |
d859e29f | 6983 | } |
fb0459d7 | 6984 | |
fb0459d7 AV |
6985 | WARN_ON_ONCE(ctx->parent_ctx); |
6986 | mutex_lock(&ctx->mutex); | |
6987 | perf_install_in_context(ctx, event, cpu); | |
6988 | ++ctx->generation; | |
fe4b04fa | 6989 | perf_unpin_context(ctx); |
fb0459d7 AV |
6990 | mutex_unlock(&ctx->mutex); |
6991 | ||
fb0459d7 AV |
6992 | return event; |
6993 | ||
c3f00c70 PZ |
6994 | err_free: |
6995 | free_event(event); | |
6996 | err: | |
c6567f64 | 6997 | return ERR_PTR(err); |
9b51f66d | 6998 | } |
fb0459d7 | 6999 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7000 | |
0cda4c02 YZ |
7001 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7002 | { | |
7003 | struct perf_event_context *src_ctx; | |
7004 | struct perf_event_context *dst_ctx; | |
7005 | struct perf_event *event, *tmp; | |
7006 | LIST_HEAD(events); | |
7007 | ||
7008 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7009 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7010 | ||
7011 | mutex_lock(&src_ctx->mutex); | |
7012 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
7013 | event_entry) { | |
54b3f8df | 7014 | perf_remove_from_context(event, false); |
0cda4c02 YZ |
7015 | put_ctx(src_ctx); |
7016 | list_add(&event->event_entry, &events); | |
7017 | } | |
7018 | mutex_unlock(&src_ctx->mutex); | |
7019 | ||
7020 | synchronize_rcu(); | |
7021 | ||
7022 | mutex_lock(&dst_ctx->mutex); | |
7023 | list_for_each_entry_safe(event, tmp, &events, event_entry) { | |
7024 | list_del(&event->event_entry); | |
7025 | if (event->state >= PERF_EVENT_STATE_OFF) | |
7026 | event->state = PERF_EVENT_STATE_INACTIVE; | |
7027 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
7028 | get_ctx(dst_ctx); | |
7029 | } | |
7030 | mutex_unlock(&dst_ctx->mutex); | |
7031 | } | |
7032 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7033 | ||
cdd6c482 | 7034 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7035 | struct task_struct *child) |
d859e29f | 7036 | { |
cdd6c482 | 7037 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7038 | u64 child_val; |
d859e29f | 7039 | |
cdd6c482 IM |
7040 | if (child_event->attr.inherit_stat) |
7041 | perf_event_read_event(child_event, child); | |
38b200d6 | 7042 | |
b5e58793 | 7043 | child_val = perf_event_count(child_event); |
d859e29f PM |
7044 | |
7045 | /* | |
7046 | * Add back the child's count to the parent's count: | |
7047 | */ | |
a6e6dea6 | 7048 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7049 | atomic64_add(child_event->total_time_enabled, |
7050 | &parent_event->child_total_time_enabled); | |
7051 | atomic64_add(child_event->total_time_running, | |
7052 | &parent_event->child_total_time_running); | |
d859e29f PM |
7053 | |
7054 | /* | |
cdd6c482 | 7055 | * Remove this event from the parent's list |
d859e29f | 7056 | */ |
cdd6c482 IM |
7057 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7058 | mutex_lock(&parent_event->child_mutex); | |
7059 | list_del_init(&child_event->child_list); | |
7060 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
7061 | |
7062 | /* | |
cdd6c482 | 7063 | * Release the parent event, if this was the last |
d859e29f PM |
7064 | * reference to it. |
7065 | */ | |
a6fa941d | 7066 | put_event(parent_event); |
d859e29f PM |
7067 | } |
7068 | ||
9b51f66d | 7069 | static void |
cdd6c482 IM |
7070 | __perf_event_exit_task(struct perf_event *child_event, |
7071 | struct perf_event_context *child_ctx, | |
38b200d6 | 7072 | struct task_struct *child) |
9b51f66d | 7073 | { |
54b3f8df | 7074 | perf_remove_from_context(child_event, !!child_event->parent); |
0cc0c027 | 7075 | |
9b51f66d | 7076 | /* |
38b435b1 | 7077 | * It can happen that the parent exits first, and has events |
9b51f66d | 7078 | * that are still around due to the child reference. These |
38b435b1 | 7079 | * events need to be zapped. |
9b51f66d | 7080 | */ |
38b435b1 | 7081 | if (child_event->parent) { |
cdd6c482 IM |
7082 | sync_child_event(child_event, child); |
7083 | free_event(child_event); | |
4bcf349a | 7084 | } |
9b51f66d IM |
7085 | } |
7086 | ||
8dc85d54 | 7087 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7088 | { |
cdd6c482 IM |
7089 | struct perf_event *child_event, *tmp; |
7090 | struct perf_event_context *child_ctx; | |
a63eaf34 | 7091 | unsigned long flags; |
9b51f66d | 7092 | |
8dc85d54 | 7093 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7094 | perf_event_task(child, NULL, 0); |
9b51f66d | 7095 | return; |
9f498cc5 | 7096 | } |
9b51f66d | 7097 | |
a63eaf34 | 7098 | local_irq_save(flags); |
ad3a37de PM |
7099 | /* |
7100 | * We can't reschedule here because interrupts are disabled, | |
7101 | * and either child is current or it is a task that can't be | |
7102 | * scheduled, so we are now safe from rescheduling changing | |
7103 | * our context. | |
7104 | */ | |
806839b2 | 7105 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7106 | |
7107 | /* | |
7108 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7109 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7110 | * incremented the context's refcount before we do put_ctx below. |
7111 | */ | |
e625cce1 | 7112 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7113 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7114 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
7115 | /* |
7116 | * If this context is a clone; unclone it so it can't get | |
7117 | * swapped to another process while we're removing all | |
cdd6c482 | 7118 | * the events from it. |
71a851b4 PZ |
7119 | */ |
7120 | unclone_ctx(child_ctx); | |
5e942bb3 | 7121 | update_context_time(child_ctx); |
e625cce1 | 7122 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
7123 | |
7124 | /* | |
cdd6c482 IM |
7125 | * Report the task dead after unscheduling the events so that we |
7126 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7127 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7128 | */ |
cdd6c482 | 7129 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 7130 | |
66fff224 PZ |
7131 | /* |
7132 | * We can recurse on the same lock type through: | |
7133 | * | |
cdd6c482 IM |
7134 | * __perf_event_exit_task() |
7135 | * sync_child_event() | |
a6fa941d AV |
7136 | * put_event() |
7137 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
7138 | * |
7139 | * But since its the parent context it won't be the same instance. | |
7140 | */ | |
a0507c84 | 7141 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 7142 | |
8bc20959 | 7143 | again: |
889ff015 FW |
7144 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
7145 | group_entry) | |
7146 | __perf_event_exit_task(child_event, child_ctx, child); | |
7147 | ||
7148 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 7149 | group_entry) |
cdd6c482 | 7150 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
7151 | |
7152 | /* | |
cdd6c482 | 7153 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
7154 | * its siblings to the list, but we obtained 'tmp' before that which |
7155 | * will still point to the list head terminating the iteration. | |
7156 | */ | |
889ff015 FW |
7157 | if (!list_empty(&child_ctx->pinned_groups) || |
7158 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 7159 | goto again; |
a63eaf34 PM |
7160 | |
7161 | mutex_unlock(&child_ctx->mutex); | |
7162 | ||
7163 | put_ctx(child_ctx); | |
9b51f66d IM |
7164 | } |
7165 | ||
8dc85d54 PZ |
7166 | /* |
7167 | * When a child task exits, feed back event values to parent events. | |
7168 | */ | |
7169 | void perf_event_exit_task(struct task_struct *child) | |
7170 | { | |
8882135b | 7171 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
7172 | int ctxn; |
7173 | ||
8882135b PZ |
7174 | mutex_lock(&child->perf_event_mutex); |
7175 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
7176 | owner_entry) { | |
7177 | list_del_init(&event->owner_entry); | |
7178 | ||
7179 | /* | |
7180 | * Ensure the list deletion is visible before we clear | |
7181 | * the owner, closes a race against perf_release() where | |
7182 | * we need to serialize on the owner->perf_event_mutex. | |
7183 | */ | |
7184 | smp_wmb(); | |
7185 | event->owner = NULL; | |
7186 | } | |
7187 | mutex_unlock(&child->perf_event_mutex); | |
7188 | ||
8dc85d54 PZ |
7189 | for_each_task_context_nr(ctxn) |
7190 | perf_event_exit_task_context(child, ctxn); | |
7191 | } | |
7192 | ||
889ff015 FW |
7193 | static void perf_free_event(struct perf_event *event, |
7194 | struct perf_event_context *ctx) | |
7195 | { | |
7196 | struct perf_event *parent = event->parent; | |
7197 | ||
7198 | if (WARN_ON_ONCE(!parent)) | |
7199 | return; | |
7200 | ||
7201 | mutex_lock(&parent->child_mutex); | |
7202 | list_del_init(&event->child_list); | |
7203 | mutex_unlock(&parent->child_mutex); | |
7204 | ||
a6fa941d | 7205 | put_event(parent); |
889ff015 | 7206 | |
8a49542c | 7207 | perf_group_detach(event); |
889ff015 FW |
7208 | list_del_event(event, ctx); |
7209 | free_event(event); | |
7210 | } | |
7211 | ||
bbbee908 PZ |
7212 | /* |
7213 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 7214 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 7215 | */ |
cdd6c482 | 7216 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 7217 | { |
8dc85d54 | 7218 | struct perf_event_context *ctx; |
cdd6c482 | 7219 | struct perf_event *event, *tmp; |
8dc85d54 | 7220 | int ctxn; |
bbbee908 | 7221 | |
8dc85d54 PZ |
7222 | for_each_task_context_nr(ctxn) { |
7223 | ctx = task->perf_event_ctxp[ctxn]; | |
7224 | if (!ctx) | |
7225 | continue; | |
bbbee908 | 7226 | |
8dc85d54 | 7227 | mutex_lock(&ctx->mutex); |
bbbee908 | 7228 | again: |
8dc85d54 PZ |
7229 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
7230 | group_entry) | |
7231 | perf_free_event(event, ctx); | |
bbbee908 | 7232 | |
8dc85d54 PZ |
7233 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
7234 | group_entry) | |
7235 | perf_free_event(event, ctx); | |
bbbee908 | 7236 | |
8dc85d54 PZ |
7237 | if (!list_empty(&ctx->pinned_groups) || |
7238 | !list_empty(&ctx->flexible_groups)) | |
7239 | goto again; | |
bbbee908 | 7240 | |
8dc85d54 | 7241 | mutex_unlock(&ctx->mutex); |
bbbee908 | 7242 | |
8dc85d54 PZ |
7243 | put_ctx(ctx); |
7244 | } | |
889ff015 FW |
7245 | } |
7246 | ||
4e231c79 PZ |
7247 | void perf_event_delayed_put(struct task_struct *task) |
7248 | { | |
7249 | int ctxn; | |
7250 | ||
7251 | for_each_task_context_nr(ctxn) | |
7252 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7253 | } | |
7254 | ||
97dee4f3 PZ |
7255 | /* |
7256 | * inherit a event from parent task to child task: | |
7257 | */ | |
7258 | static struct perf_event * | |
7259 | inherit_event(struct perf_event *parent_event, | |
7260 | struct task_struct *parent, | |
7261 | struct perf_event_context *parent_ctx, | |
7262 | struct task_struct *child, | |
7263 | struct perf_event *group_leader, | |
7264 | struct perf_event_context *child_ctx) | |
7265 | { | |
7266 | struct perf_event *child_event; | |
cee010ec | 7267 | unsigned long flags; |
97dee4f3 PZ |
7268 | |
7269 | /* | |
7270 | * Instead of creating recursive hierarchies of events, | |
7271 | * we link inherited events back to the original parent, | |
7272 | * which has a filp for sure, which we use as the reference | |
7273 | * count: | |
7274 | */ | |
7275 | if (parent_event->parent) | |
7276 | parent_event = parent_event->parent; | |
7277 | ||
7278 | child_event = perf_event_alloc(&parent_event->attr, | |
7279 | parent_event->cpu, | |
d580ff86 | 7280 | child, |
97dee4f3 | 7281 | group_leader, parent_event, |
4dc0da86 | 7282 | NULL, NULL); |
97dee4f3 PZ |
7283 | if (IS_ERR(child_event)) |
7284 | return child_event; | |
a6fa941d AV |
7285 | |
7286 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7287 | free_event(child_event); | |
7288 | return NULL; | |
7289 | } | |
7290 | ||
97dee4f3 PZ |
7291 | get_ctx(child_ctx); |
7292 | ||
7293 | /* | |
7294 | * Make the child state follow the state of the parent event, | |
7295 | * not its attr.disabled bit. We hold the parent's mutex, | |
7296 | * so we won't race with perf_event_{en, dis}able_family. | |
7297 | */ | |
7298 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7299 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7300 | else | |
7301 | child_event->state = PERF_EVENT_STATE_OFF; | |
7302 | ||
7303 | if (parent_event->attr.freq) { | |
7304 | u64 sample_period = parent_event->hw.sample_period; | |
7305 | struct hw_perf_event *hwc = &child_event->hw; | |
7306 | ||
7307 | hwc->sample_period = sample_period; | |
7308 | hwc->last_period = sample_period; | |
7309 | ||
7310 | local64_set(&hwc->period_left, sample_period); | |
7311 | } | |
7312 | ||
7313 | child_event->ctx = child_ctx; | |
7314 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7315 | child_event->overflow_handler_context |
7316 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7317 | |
614b6780 TG |
7318 | /* |
7319 | * Precalculate sample_data sizes | |
7320 | */ | |
7321 | perf_event__header_size(child_event); | |
6844c09d | 7322 | perf_event__id_header_size(child_event); |
614b6780 | 7323 | |
97dee4f3 PZ |
7324 | /* |
7325 | * Link it up in the child's context: | |
7326 | */ | |
cee010ec | 7327 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7328 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7329 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7330 | |
97dee4f3 PZ |
7331 | /* |
7332 | * Link this into the parent event's child list | |
7333 | */ | |
7334 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7335 | mutex_lock(&parent_event->child_mutex); | |
7336 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7337 | mutex_unlock(&parent_event->child_mutex); | |
7338 | ||
7339 | return child_event; | |
7340 | } | |
7341 | ||
7342 | static int inherit_group(struct perf_event *parent_event, | |
7343 | struct task_struct *parent, | |
7344 | struct perf_event_context *parent_ctx, | |
7345 | struct task_struct *child, | |
7346 | struct perf_event_context *child_ctx) | |
7347 | { | |
7348 | struct perf_event *leader; | |
7349 | struct perf_event *sub; | |
7350 | struct perf_event *child_ctr; | |
7351 | ||
7352 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7353 | child, NULL, child_ctx); | |
7354 | if (IS_ERR(leader)) | |
7355 | return PTR_ERR(leader); | |
7356 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7357 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7358 | child, leader, child_ctx); | |
7359 | if (IS_ERR(child_ctr)) | |
7360 | return PTR_ERR(child_ctr); | |
7361 | } | |
7362 | return 0; | |
889ff015 FW |
7363 | } |
7364 | ||
7365 | static int | |
7366 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7367 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7368 | struct task_struct *child, int ctxn, |
889ff015 FW |
7369 | int *inherited_all) |
7370 | { | |
7371 | int ret; | |
8dc85d54 | 7372 | struct perf_event_context *child_ctx; |
889ff015 FW |
7373 | |
7374 | if (!event->attr.inherit) { | |
7375 | *inherited_all = 0; | |
7376 | return 0; | |
bbbee908 PZ |
7377 | } |
7378 | ||
fe4b04fa | 7379 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7380 | if (!child_ctx) { |
7381 | /* | |
7382 | * This is executed from the parent task context, so | |
7383 | * inherit events that have been marked for cloning. | |
7384 | * First allocate and initialize a context for the | |
7385 | * child. | |
7386 | */ | |
bbbee908 | 7387 | |
f38bac3d | 7388 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
7389 | if (!child_ctx) |
7390 | return -ENOMEM; | |
bbbee908 | 7391 | |
8dc85d54 | 7392 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7393 | } |
7394 | ||
7395 | ret = inherit_group(event, parent, parent_ctx, | |
7396 | child, child_ctx); | |
7397 | ||
7398 | if (ret) | |
7399 | *inherited_all = 0; | |
7400 | ||
7401 | return ret; | |
bbbee908 PZ |
7402 | } |
7403 | ||
9b51f66d | 7404 | /* |
cdd6c482 | 7405 | * Initialize the perf_event context in task_struct |
9b51f66d | 7406 | */ |
8dc85d54 | 7407 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7408 | { |
889ff015 | 7409 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7410 | struct perf_event_context *cloned_ctx; |
7411 | struct perf_event *event; | |
9b51f66d | 7412 | struct task_struct *parent = current; |
564c2b21 | 7413 | int inherited_all = 1; |
dddd3379 | 7414 | unsigned long flags; |
6ab423e0 | 7415 | int ret = 0; |
9b51f66d | 7416 | |
8dc85d54 | 7417 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7418 | return 0; |
7419 | ||
ad3a37de | 7420 | /* |
25346b93 PM |
7421 | * If the parent's context is a clone, pin it so it won't get |
7422 | * swapped under us. | |
ad3a37de | 7423 | */ |
8dc85d54 | 7424 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 7425 | |
ad3a37de PM |
7426 | /* |
7427 | * No need to check if parent_ctx != NULL here; since we saw | |
7428 | * it non-NULL earlier, the only reason for it to become NULL | |
7429 | * is if we exit, and since we're currently in the middle of | |
7430 | * a fork we can't be exiting at the same time. | |
7431 | */ | |
ad3a37de | 7432 | |
9b51f66d IM |
7433 | /* |
7434 | * Lock the parent list. No need to lock the child - not PID | |
7435 | * hashed yet and not running, so nobody can access it. | |
7436 | */ | |
d859e29f | 7437 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7438 | |
7439 | /* | |
7440 | * We dont have to disable NMIs - we are only looking at | |
7441 | * the list, not manipulating it: | |
7442 | */ | |
889ff015 | 7443 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7444 | ret = inherit_task_group(event, parent, parent_ctx, |
7445 | child, ctxn, &inherited_all); | |
889ff015 FW |
7446 | if (ret) |
7447 | break; | |
7448 | } | |
b93f7978 | 7449 | |
dddd3379 TG |
7450 | /* |
7451 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7452 | * to allocations, but we need to prevent rotation because | |
7453 | * rotate_ctx() will change the list from interrupt context. | |
7454 | */ | |
7455 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7456 | parent_ctx->rotate_disable = 1; | |
7457 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7458 | ||
889ff015 | 7459 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7460 | ret = inherit_task_group(event, parent, parent_ctx, |
7461 | child, ctxn, &inherited_all); | |
889ff015 | 7462 | if (ret) |
9b51f66d | 7463 | break; |
564c2b21 PM |
7464 | } |
7465 | ||
dddd3379 TG |
7466 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7467 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7468 | |
8dc85d54 | 7469 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7470 | |
05cbaa28 | 7471 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7472 | /* |
7473 | * Mark the child context as a clone of the parent | |
7474 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7475 | * |
7476 | * Note that if the parent is a clone, the holding of | |
7477 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7478 | */ |
c5ed5145 | 7479 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7480 | if (cloned_ctx) { |
7481 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7482 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7483 | } else { |
7484 | child_ctx->parent_ctx = parent_ctx; | |
7485 | child_ctx->parent_gen = parent_ctx->generation; | |
7486 | } | |
7487 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7488 | } |
7489 | ||
c5ed5145 | 7490 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7491 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7492 | |
25346b93 | 7493 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7494 | put_ctx(parent_ctx); |
ad3a37de | 7495 | |
6ab423e0 | 7496 | return ret; |
9b51f66d IM |
7497 | } |
7498 | ||
8dc85d54 PZ |
7499 | /* |
7500 | * Initialize the perf_event context in task_struct | |
7501 | */ | |
7502 | int perf_event_init_task(struct task_struct *child) | |
7503 | { | |
7504 | int ctxn, ret; | |
7505 | ||
8550d7cb ON |
7506 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7507 | mutex_init(&child->perf_event_mutex); | |
7508 | INIT_LIST_HEAD(&child->perf_event_list); | |
7509 | ||
8dc85d54 PZ |
7510 | for_each_task_context_nr(ctxn) { |
7511 | ret = perf_event_init_context(child, ctxn); | |
bee870fc PZ |
7512 | if (ret) { |
7513 | perf_event_free_task(child); | |
8dc85d54 | 7514 | return ret; |
bee870fc | 7515 | } |
8dc85d54 PZ |
7516 | } |
7517 | ||
7518 | return 0; | |
7519 | } | |
7520 | ||
220b140b PM |
7521 | static void __init perf_event_init_all_cpus(void) |
7522 | { | |
b28ab83c | 7523 | struct swevent_htable *swhash; |
220b140b | 7524 | int cpu; |
220b140b PM |
7525 | |
7526 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7527 | swhash = &per_cpu(swevent_htable, cpu); |
7528 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7529 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7530 | } |
7531 | } | |
7532 | ||
cdd6c482 | 7533 | static void __cpuinit perf_event_init_cpu(int cpu) |
0793a61d | 7534 | { |
108b02cf | 7535 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7536 | |
b28ab83c | 7537 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7538 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7539 | struct swevent_hlist *hlist; |
7540 | ||
b28ab83c PZ |
7541 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7542 | WARN_ON(!hlist); | |
7543 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7544 | } |
b28ab83c | 7545 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7546 | } |
7547 | ||
c277443c | 7548 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7549 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7550 | { |
e9d2b064 PZ |
7551 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7552 | ||
7553 | WARN_ON(!irqs_disabled()); | |
7554 | ||
7555 | list_del_init(&cpuctx->rotation_list); | |
7556 | } | |
7557 | ||
108b02cf | 7558 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7559 | { |
54b3f8df | 7560 | struct remove_event re = { .detach_group = false }; |
108b02cf | 7561 | struct perf_event_context *ctx = __info; |
0793a61d | 7562 | |
108b02cf | 7563 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7564 | |
35d1c833 | 7565 | rcu_read_lock(); |
54b3f8df PZ |
7566 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
7567 | __perf_remove_from_context(&re); | |
35d1c833 | 7568 | rcu_read_unlock(); |
0793a61d | 7569 | } |
108b02cf PZ |
7570 | |
7571 | static void perf_event_exit_cpu_context(int cpu) | |
7572 | { | |
7573 | struct perf_event_context *ctx; | |
7574 | struct pmu *pmu; | |
7575 | int idx; | |
7576 | ||
7577 | idx = srcu_read_lock(&pmus_srcu); | |
7578 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7579 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7580 | |
7581 | mutex_lock(&ctx->mutex); | |
7582 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7583 | mutex_unlock(&ctx->mutex); | |
7584 | } | |
7585 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7586 | } |
7587 | ||
cdd6c482 | 7588 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7589 | { |
35d1c833 | 7590 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7591 | } |
7592 | #else | |
cdd6c482 | 7593 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7594 | #endif |
7595 | ||
c277443c PZ |
7596 | static int |
7597 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7598 | { | |
7599 | int cpu; | |
7600 | ||
7601 | for_each_online_cpu(cpu) | |
7602 | perf_event_exit_cpu(cpu); | |
7603 | ||
7604 | return NOTIFY_OK; | |
7605 | } | |
7606 | ||
7607 | /* | |
7608 | * Run the perf reboot notifier at the very last possible moment so that | |
7609 | * the generic watchdog code runs as long as possible. | |
7610 | */ | |
7611 | static struct notifier_block perf_reboot_notifier = { | |
7612 | .notifier_call = perf_reboot, | |
7613 | .priority = INT_MIN, | |
7614 | }; | |
7615 | ||
0793a61d TG |
7616 | static int __cpuinit |
7617 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
7618 | { | |
7619 | unsigned int cpu = (long)hcpu; | |
7620 | ||
4536e4d1 | 7621 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7622 | |
7623 | case CPU_UP_PREPARE: | |
5e11637e | 7624 | case CPU_DOWN_FAILED: |
cdd6c482 | 7625 | perf_event_init_cpu(cpu); |
0793a61d TG |
7626 | break; |
7627 | ||
5e11637e | 7628 | case CPU_UP_CANCELED: |
0793a61d | 7629 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7630 | perf_event_exit_cpu(cpu); |
0793a61d TG |
7631 | break; |
7632 | ||
7633 | default: | |
7634 | break; | |
7635 | } | |
7636 | ||
7637 | return NOTIFY_OK; | |
7638 | } | |
7639 | ||
cdd6c482 | 7640 | void __init perf_event_init(void) |
0793a61d | 7641 | { |
3c502e7a JW |
7642 | int ret; |
7643 | ||
2e80a82a PZ |
7644 | idr_init(&pmu_idr); |
7645 | ||
220b140b | 7646 | perf_event_init_all_cpus(); |
b0a873eb | 7647 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7648 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7649 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7650 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7651 | perf_tp_register(); |
7652 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7653 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7654 | |
7655 | ret = init_hw_breakpoint(); | |
7656 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7657 | |
7658 | /* do not patch jump label more than once per second */ | |
7659 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7660 | |
7661 | /* | |
7662 | * Build time assertion that we keep the data_head at the intended | |
7663 | * location. IOW, validation we got the __reserved[] size right. | |
7664 | */ | |
7665 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
7666 | != 1024); | |
0793a61d | 7667 | } |
abe43400 PZ |
7668 | |
7669 | static int __init perf_event_sysfs_init(void) | |
7670 | { | |
7671 | struct pmu *pmu; | |
7672 | int ret; | |
7673 | ||
7674 | mutex_lock(&pmus_lock); | |
7675 | ||
7676 | ret = bus_register(&pmu_bus); | |
7677 | if (ret) | |
7678 | goto unlock; | |
7679 | ||
7680 | list_for_each_entry(pmu, &pmus, entry) { | |
7681 | if (!pmu->name || pmu->type < 0) | |
7682 | continue; | |
7683 | ||
7684 | ret = pmu_dev_alloc(pmu); | |
7685 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
7686 | } | |
7687 | pmu_bus_running = 1; | |
7688 | ret = 0; | |
7689 | ||
7690 | unlock: | |
7691 | mutex_unlock(&pmus_lock); | |
7692 | ||
7693 | return ret; | |
7694 | } | |
7695 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
7696 | |
7697 | #ifdef CONFIG_CGROUP_PERF | |
92fb9748 | 7698 | static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont) |
e5d1367f SE |
7699 | { |
7700 | struct perf_cgroup *jc; | |
e5d1367f | 7701 | |
1b15d055 | 7702 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
7703 | if (!jc) |
7704 | return ERR_PTR(-ENOMEM); | |
7705 | ||
e5d1367f SE |
7706 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7707 | if (!jc->info) { | |
7708 | kfree(jc); | |
7709 | return ERR_PTR(-ENOMEM); | |
7710 | } | |
7711 | ||
e5d1367f SE |
7712 | return &jc->css; |
7713 | } | |
7714 | ||
92fb9748 | 7715 | static void perf_cgroup_css_free(struct cgroup *cont) |
e5d1367f SE |
7716 | { |
7717 | struct perf_cgroup *jc; | |
7718 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), | |
7719 | struct perf_cgroup, css); | |
7720 | free_percpu(jc->info); | |
7721 | kfree(jc); | |
7722 | } | |
7723 | ||
7724 | static int __perf_cgroup_move(void *info) | |
7725 | { | |
7726 | struct task_struct *task = info; | |
7727 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
7728 | return 0; | |
7729 | } | |
7730 | ||
761b3ef5 | 7731 | static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
e5d1367f | 7732 | { |
bb9d97b6 TH |
7733 | struct task_struct *task; |
7734 | ||
7735 | cgroup_taskset_for_each(task, cgrp, tset) | |
7736 | task_function_call(task, __perf_cgroup_move, task); | |
e5d1367f SE |
7737 | } |
7738 | ||
761b3ef5 LZ |
7739 | static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, |
7740 | struct task_struct *task) | |
e5d1367f SE |
7741 | { |
7742 | /* | |
7743 | * cgroup_exit() is called in the copy_process() failure path. | |
7744 | * Ignore this case since the task hasn't ran yet, this avoids | |
7745 | * trying to poke a half freed task state from generic code. | |
7746 | */ | |
7747 | if (!(task->flags & PF_EXITING)) | |
7748 | return; | |
7749 | ||
bb9d97b6 | 7750 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7751 | } |
7752 | ||
7753 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
7754 | .name = "perf_event", |
7755 | .subsys_id = perf_subsys_id, | |
92fb9748 TH |
7756 | .css_alloc = perf_cgroup_css_alloc, |
7757 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 7758 | .exit = perf_cgroup_exit, |
bb9d97b6 | 7759 | .attach = perf_cgroup_attach, |
e5d1367f SE |
7760 | }; |
7761 | #endif /* CONFIG_CGROUP_PERF */ |