]>
Commit | Line | Data |
---|---|---|
0793a61d | 1 | /* |
57c0c15b | 2 | * Performance events core code: |
0793a61d | 3 | * |
98144511 | 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
e7e7ee2e IM |
5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
7b732a75 | 8 | * |
57c0c15b | 9 | * For licensing details see kernel-base/COPYING |
0793a61d TG |
10 | */ |
11 | ||
12 | #include <linux/fs.h> | |
b9cacc7b | 13 | #include <linux/mm.h> |
0793a61d TG |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> | |
2e80a82a | 16 | #include <linux/idr.h> |
04289bb9 | 17 | #include <linux/file.h> |
0793a61d | 18 | #include <linux/poll.h> |
5a0e3ad6 | 19 | #include <linux/slab.h> |
76e1d904 | 20 | #include <linux/hash.h> |
12351ef8 | 21 | #include <linux/tick.h> |
0793a61d | 22 | #include <linux/sysfs.h> |
22a4f650 | 23 | #include <linux/dcache.h> |
0793a61d | 24 | #include <linux/percpu.h> |
22a4f650 | 25 | #include <linux/ptrace.h> |
c277443c | 26 | #include <linux/reboot.h> |
b9cacc7b | 27 | #include <linux/vmstat.h> |
abe43400 | 28 | #include <linux/device.h> |
6e5fdeed | 29 | #include <linux/export.h> |
906010b2 | 30 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
31 | #include <linux/hardirq.h> |
32 | #include <linux/rculist.h> | |
0793a61d TG |
33 | #include <linux/uaccess.h> |
34 | #include <linux/syscalls.h> | |
35 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 36 | #include <linux/kernel_stat.h> |
cdd6c482 | 37 | #include <linux/perf_event.h> |
6fb2915d | 38 | #include <linux/ftrace_event.h> |
3c502e7a | 39 | #include <linux/hw_breakpoint.h> |
c5ebcedb | 40 | #include <linux/mm_types.h> |
877c6856 | 41 | #include <linux/cgroup.h> |
0793a61d | 42 | |
76369139 FW |
43 | #include "internal.h" |
44 | ||
4e193bd4 TB |
45 | #include <asm/irq_regs.h> |
46 | ||
fe4b04fa | 47 | struct remote_function_call { |
e7e7ee2e IM |
48 | struct task_struct *p; |
49 | int (*func)(void *info); | |
50 | void *info; | |
51 | int ret; | |
fe4b04fa PZ |
52 | }; |
53 | ||
54 | static void remote_function(void *data) | |
55 | { | |
56 | struct remote_function_call *tfc = data; | |
57 | struct task_struct *p = tfc->p; | |
58 | ||
59 | if (p) { | |
60 | tfc->ret = -EAGAIN; | |
61 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
62 | return; | |
63 | } | |
64 | ||
65 | tfc->ret = tfc->func(tfc->info); | |
66 | } | |
67 | ||
68 | /** | |
69 | * task_function_call - call a function on the cpu on which a task runs | |
70 | * @p: the task to evaluate | |
71 | * @func: the function to be called | |
72 | * @info: the function call argument | |
73 | * | |
74 | * Calls the function @func when the task is currently running. This might | |
75 | * be on the current CPU, which just calls the function directly | |
76 | * | |
77 | * returns: @func return value, or | |
78 | * -ESRCH - when the process isn't running | |
79 | * -EAGAIN - when the process moved away | |
80 | */ | |
81 | static int | |
82 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
83 | { | |
84 | struct remote_function_call data = { | |
e7e7ee2e IM |
85 | .p = p, |
86 | .func = func, | |
87 | .info = info, | |
88 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
89 | }; |
90 | ||
91 | if (task_curr(p)) | |
92 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
93 | ||
94 | return data.ret; | |
95 | } | |
96 | ||
97 | /** | |
98 | * cpu_function_call - call a function on the cpu | |
99 | * @func: the function to be called | |
100 | * @info: the function call argument | |
101 | * | |
102 | * Calls the function @func on the remote cpu. | |
103 | * | |
104 | * returns: @func return value or -ENXIO when the cpu is offline | |
105 | */ | |
106 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
107 | { | |
108 | struct remote_function_call data = { | |
e7e7ee2e IM |
109 | .p = NULL, |
110 | .func = func, | |
111 | .info = info, | |
112 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
113 | }; |
114 | ||
115 | smp_call_function_single(cpu, remote_function, &data, 1); | |
116 | ||
117 | return data.ret; | |
118 | } | |
119 | ||
e5d1367f SE |
120 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
121 | PERF_FLAG_FD_OUTPUT |\ | |
122 | PERF_FLAG_PID_CGROUP) | |
123 | ||
bce38cd5 SE |
124 | /* |
125 | * branch priv levels that need permission checks | |
126 | */ | |
127 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
128 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
129 | PERF_SAMPLE_BRANCH_HV) | |
130 | ||
0b3fcf17 SE |
131 | enum event_type_t { |
132 | EVENT_FLEXIBLE = 0x1, | |
133 | EVENT_PINNED = 0x2, | |
134 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
135 | }; | |
136 | ||
e5d1367f SE |
137 | /* |
138 | * perf_sched_events : >0 events exist | |
139 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
140 | */ | |
c5905afb | 141 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 142 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 143 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
ba8a75c1 | 144 | static DEFINE_PER_CPU(atomic_t, perf_freq_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 | */ |
14c63f17 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; | |
e5302920 | 186 | do_div(tmp, 100); |
14c63f17 DH |
187 | atomic_set(&perf_sample_allowed_ns, tmp); |
188 | } | |
163ec435 | 189 | |
9e630205 SE |
190 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
191 | ||
163ec435 PZ |
192 | int perf_proc_update_handler(struct ctl_table *table, int write, |
193 | void __user *buffer, size_t *lenp, | |
194 | loff_t *ppos) | |
195 | { | |
196 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
197 | ||
198 | if (ret || !write) | |
199 | return ret; | |
200 | ||
201 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
202 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
203 | update_perf_cpu_limits(); | |
204 | ||
205 | return 0; | |
206 | } | |
207 | ||
208 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
209 | ||
210 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
211 | void __user *buffer, size_t *lenp, | |
212 | loff_t *ppos) | |
213 | { | |
214 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
215 | ||
216 | if (ret || !write) | |
217 | return ret; | |
218 | ||
219 | update_perf_cpu_limits(); | |
163ec435 PZ |
220 | |
221 | return 0; | |
222 | } | |
1ccd1549 | 223 | |
14c63f17 DH |
224 | /* |
225 | * perf samples are done in some very critical code paths (NMIs). | |
226 | * If they take too much CPU time, the system can lock up and not | |
227 | * get any real work done. This will drop the sample rate when | |
228 | * we detect that events are taking too long. | |
229 | */ | |
230 | #define NR_ACCUMULATED_SAMPLES 128 | |
231 | DEFINE_PER_CPU(u64, running_sample_length); | |
232 | ||
233 | void perf_sample_event_took(u64 sample_len_ns) | |
234 | { | |
235 | u64 avg_local_sample_len; | |
e5302920 | 236 | u64 local_samples_len; |
14c63f17 DH |
237 | |
238 | if (atomic_read(&perf_sample_allowed_ns) == 0) | |
239 | return; | |
240 | ||
241 | /* decay the counter by 1 average sample */ | |
242 | local_samples_len = __get_cpu_var(running_sample_length); | |
243 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; | |
244 | local_samples_len += sample_len_ns; | |
245 | __get_cpu_var(running_sample_length) = local_samples_len; | |
246 | ||
247 | /* | |
248 | * note: this will be biased artifically low until we have | |
249 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
250 | * from having to maintain a count. | |
251 | */ | |
252 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
253 | ||
254 | if (avg_local_sample_len <= atomic_read(&perf_sample_allowed_ns)) | |
255 | return; | |
256 | ||
257 | if (max_samples_per_tick <= 1) | |
258 | return; | |
259 | ||
260 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
261 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
262 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
263 | ||
264 | printk_ratelimited(KERN_WARNING | |
265 | "perf samples too long (%lld > %d), lowering " | |
266 | "kernel.perf_event_max_sample_rate to %d\n", | |
267 | avg_local_sample_len, | |
268 | atomic_read(&perf_sample_allowed_ns), | |
269 | sysctl_perf_event_sample_rate); | |
270 | ||
271 | update_perf_cpu_limits(); | |
272 | } | |
273 | ||
cdd6c482 | 274 | static atomic64_t perf_event_id; |
a96bbc16 | 275 | |
0b3fcf17 SE |
276 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
277 | enum event_type_t event_type); | |
278 | ||
279 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
280 | enum event_type_t event_type, |
281 | struct task_struct *task); | |
282 | ||
283 | static void update_context_time(struct perf_event_context *ctx); | |
284 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 285 | |
cdd6c482 | 286 | void __weak perf_event_print_debug(void) { } |
0793a61d | 287 | |
84c79910 | 288 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 289 | { |
84c79910 | 290 | return "pmu"; |
0793a61d TG |
291 | } |
292 | ||
0b3fcf17 SE |
293 | static inline u64 perf_clock(void) |
294 | { | |
295 | return local_clock(); | |
296 | } | |
297 | ||
e5d1367f SE |
298 | static inline struct perf_cpu_context * |
299 | __get_cpu_context(struct perf_event_context *ctx) | |
300 | { | |
301 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
302 | } | |
303 | ||
facc4307 PZ |
304 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
305 | struct perf_event_context *ctx) | |
306 | { | |
307 | raw_spin_lock(&cpuctx->ctx.lock); | |
308 | if (ctx) | |
309 | raw_spin_lock(&ctx->lock); | |
310 | } | |
311 | ||
312 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
313 | struct perf_event_context *ctx) | |
314 | { | |
315 | if (ctx) | |
316 | raw_spin_unlock(&ctx->lock); | |
317 | raw_spin_unlock(&cpuctx->ctx.lock); | |
318 | } | |
319 | ||
e5d1367f SE |
320 | #ifdef CONFIG_CGROUP_PERF |
321 | ||
877c6856 LZ |
322 | /* |
323 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
324 | * This is a per-cpu dynamically allocated data structure. | |
325 | */ | |
326 | struct perf_cgroup_info { | |
327 | u64 time; | |
328 | u64 timestamp; | |
329 | }; | |
330 | ||
331 | struct perf_cgroup { | |
332 | struct cgroup_subsys_state css; | |
86e213e1 | 333 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
334 | }; |
335 | ||
3f7cce3c SE |
336 | /* |
337 | * Must ensure cgroup is pinned (css_get) before calling | |
338 | * this function. In other words, we cannot call this function | |
339 | * if there is no cgroup event for the current CPU context. | |
340 | */ | |
e5d1367f SE |
341 | static inline struct perf_cgroup * |
342 | perf_cgroup_from_task(struct task_struct *task) | |
343 | { | |
344 | return container_of(task_subsys_state(task, perf_subsys_id), | |
345 | struct perf_cgroup, css); | |
346 | } | |
347 | ||
348 | static inline bool | |
349 | perf_cgroup_match(struct perf_event *event) | |
350 | { | |
351 | struct perf_event_context *ctx = event->ctx; | |
352 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
353 | ||
ef824fa1 TH |
354 | /* @event doesn't care about cgroup */ |
355 | if (!event->cgrp) | |
356 | return true; | |
357 | ||
358 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
359 | if (!cpuctx->cgrp) | |
360 | return false; | |
361 | ||
362 | /* | |
363 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
364 | * also enabled for all its descendant cgroups. If @cpuctx's | |
365 | * cgroup is a descendant of @event's (the test covers identity | |
366 | * case), it's a match. | |
367 | */ | |
368 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
369 | event->cgrp->css.cgroup); | |
e5d1367f SE |
370 | } |
371 | ||
9c5da09d | 372 | static inline bool perf_tryget_cgroup(struct perf_event *event) |
e5d1367f | 373 | { |
9c5da09d | 374 | return css_tryget(&event->cgrp->css); |
e5d1367f SE |
375 | } |
376 | ||
377 | static inline void perf_put_cgroup(struct perf_event *event) | |
378 | { | |
379 | css_put(&event->cgrp->css); | |
380 | } | |
381 | ||
382 | static inline void perf_detach_cgroup(struct perf_event *event) | |
383 | { | |
384 | perf_put_cgroup(event); | |
385 | event->cgrp = NULL; | |
386 | } | |
387 | ||
388 | static inline int is_cgroup_event(struct perf_event *event) | |
389 | { | |
390 | return event->cgrp != NULL; | |
391 | } | |
392 | ||
393 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
394 | { | |
395 | struct perf_cgroup_info *t; | |
396 | ||
397 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
398 | return t->time; | |
399 | } | |
400 | ||
401 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
402 | { | |
403 | struct perf_cgroup_info *info; | |
404 | u64 now; | |
405 | ||
406 | now = perf_clock(); | |
407 | ||
408 | info = this_cpu_ptr(cgrp->info); | |
409 | ||
410 | info->time += now - info->timestamp; | |
411 | info->timestamp = now; | |
412 | } | |
413 | ||
414 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
415 | { | |
416 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
417 | if (cgrp_out) | |
418 | __update_cgrp_time(cgrp_out); | |
419 | } | |
420 | ||
421 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
422 | { | |
3f7cce3c SE |
423 | struct perf_cgroup *cgrp; |
424 | ||
e5d1367f | 425 | /* |
3f7cce3c SE |
426 | * ensure we access cgroup data only when needed and |
427 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 428 | */ |
3f7cce3c | 429 | if (!is_cgroup_event(event)) |
e5d1367f SE |
430 | return; |
431 | ||
3f7cce3c SE |
432 | cgrp = perf_cgroup_from_task(current); |
433 | /* | |
434 | * Do not update time when cgroup is not active | |
435 | */ | |
436 | if (cgrp == event->cgrp) | |
437 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
438 | } |
439 | ||
440 | static inline void | |
3f7cce3c SE |
441 | perf_cgroup_set_timestamp(struct task_struct *task, |
442 | struct perf_event_context *ctx) | |
e5d1367f SE |
443 | { |
444 | struct perf_cgroup *cgrp; | |
445 | struct perf_cgroup_info *info; | |
446 | ||
3f7cce3c SE |
447 | /* |
448 | * ctx->lock held by caller | |
449 | * ensure we do not access cgroup data | |
450 | * unless we have the cgroup pinned (css_get) | |
451 | */ | |
452 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
453 | return; |
454 | ||
455 | cgrp = perf_cgroup_from_task(task); | |
456 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 457 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
458 | } |
459 | ||
460 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
461 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
462 | ||
463 | /* | |
464 | * reschedule events based on the cgroup constraint of task. | |
465 | * | |
466 | * mode SWOUT : schedule out everything | |
467 | * mode SWIN : schedule in based on cgroup for next | |
468 | */ | |
469 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
470 | { | |
471 | struct perf_cpu_context *cpuctx; | |
472 | struct pmu *pmu; | |
473 | unsigned long flags; | |
474 | ||
475 | /* | |
476 | * disable interrupts to avoid geting nr_cgroup | |
477 | * changes via __perf_event_disable(). Also | |
478 | * avoids preemption. | |
479 | */ | |
480 | local_irq_save(flags); | |
481 | ||
482 | /* | |
483 | * we reschedule only in the presence of cgroup | |
484 | * constrained events. | |
485 | */ | |
486 | rcu_read_lock(); | |
487 | ||
488 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 489 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
490 | if (cpuctx->unique_pmu != pmu) |
491 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 492 | |
e5d1367f SE |
493 | /* |
494 | * perf_cgroup_events says at least one | |
495 | * context on this CPU has cgroup events. | |
496 | * | |
497 | * ctx->nr_cgroups reports the number of cgroup | |
498 | * events for a context. | |
499 | */ | |
500 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
501 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
502 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
503 | |
504 | if (mode & PERF_CGROUP_SWOUT) { | |
505 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
506 | /* | |
507 | * must not be done before ctxswout due | |
508 | * to event_filter_match() in event_sched_out() | |
509 | */ | |
510 | cpuctx->cgrp = NULL; | |
511 | } | |
512 | ||
513 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 514 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
515 | /* |
516 | * set cgrp before ctxsw in to allow | |
517 | * event_filter_match() to not have to pass | |
518 | * task around | |
e5d1367f SE |
519 | */ |
520 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
521 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
522 | } | |
facc4307 PZ |
523 | perf_pmu_enable(cpuctx->ctx.pmu); |
524 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 525 | } |
e5d1367f SE |
526 | } |
527 | ||
528 | rcu_read_unlock(); | |
529 | ||
530 | local_irq_restore(flags); | |
531 | } | |
532 | ||
a8d757ef SE |
533 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
534 | struct task_struct *next) | |
e5d1367f | 535 | { |
a8d757ef SE |
536 | struct perf_cgroup *cgrp1; |
537 | struct perf_cgroup *cgrp2 = NULL; | |
538 | ||
539 | /* | |
540 | * we come here when we know perf_cgroup_events > 0 | |
541 | */ | |
542 | cgrp1 = perf_cgroup_from_task(task); | |
543 | ||
544 | /* | |
545 | * next is NULL when called from perf_event_enable_on_exec() | |
546 | * that will systematically cause a cgroup_switch() | |
547 | */ | |
548 | if (next) | |
549 | cgrp2 = perf_cgroup_from_task(next); | |
550 | ||
551 | /* | |
552 | * only schedule out current cgroup events if we know | |
553 | * that we are switching to a different cgroup. Otherwise, | |
554 | * do no touch the cgroup events. | |
555 | */ | |
556 | if (cgrp1 != cgrp2) | |
557 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
558 | } |
559 | ||
a8d757ef SE |
560 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
561 | struct task_struct *task) | |
e5d1367f | 562 | { |
a8d757ef SE |
563 | struct perf_cgroup *cgrp1; |
564 | struct perf_cgroup *cgrp2 = NULL; | |
565 | ||
566 | /* | |
567 | * we come here when we know perf_cgroup_events > 0 | |
568 | */ | |
569 | cgrp1 = perf_cgroup_from_task(task); | |
570 | ||
571 | /* prev can never be NULL */ | |
572 | cgrp2 = perf_cgroup_from_task(prev); | |
573 | ||
574 | /* | |
575 | * only need to schedule in cgroup events if we are changing | |
576 | * cgroup during ctxsw. Cgroup events were not scheduled | |
577 | * out of ctxsw out if that was not the case. | |
578 | */ | |
579 | if (cgrp1 != cgrp2) | |
580 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
581 | } |
582 | ||
583 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
584 | struct perf_event_attr *attr, | |
585 | struct perf_event *group_leader) | |
586 | { | |
587 | struct perf_cgroup *cgrp; | |
588 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
589 | struct fd f = fdget(fd); |
590 | int ret = 0; | |
e5d1367f | 591 | |
2903ff01 | 592 | if (!f.file) |
e5d1367f SE |
593 | return -EBADF; |
594 | ||
2903ff01 | 595 | css = cgroup_css_from_dir(f.file, perf_subsys_id); |
3db272c0 LZ |
596 | if (IS_ERR(css)) { |
597 | ret = PTR_ERR(css); | |
598 | goto out; | |
599 | } | |
e5d1367f SE |
600 | |
601 | cgrp = container_of(css, struct perf_cgroup, css); | |
602 | event->cgrp = cgrp; | |
603 | ||
f75e18cb | 604 | /* must be done before we fput() the file */ |
9c5da09d SQ |
605 | if (!perf_tryget_cgroup(event)) { |
606 | event->cgrp = NULL; | |
607 | ret = -ENOENT; | |
608 | goto out; | |
609 | } | |
f75e18cb | 610 | |
e5d1367f SE |
611 | /* |
612 | * all events in a group must monitor | |
613 | * the same cgroup because a task belongs | |
614 | * to only one perf cgroup at a time | |
615 | */ | |
616 | if (group_leader && group_leader->cgrp != cgrp) { | |
617 | perf_detach_cgroup(event); | |
618 | ret = -EINVAL; | |
e5d1367f | 619 | } |
3db272c0 | 620 | out: |
2903ff01 | 621 | fdput(f); |
e5d1367f SE |
622 | return ret; |
623 | } | |
624 | ||
625 | static inline void | |
626 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
627 | { | |
628 | struct perf_cgroup_info *t; | |
629 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
630 | event->shadow_ctx_time = now - t->timestamp; | |
631 | } | |
632 | ||
633 | static inline void | |
634 | perf_cgroup_defer_enabled(struct perf_event *event) | |
635 | { | |
636 | /* | |
637 | * when the current task's perf cgroup does not match | |
638 | * the event's, we need to remember to call the | |
639 | * perf_mark_enable() function the first time a task with | |
640 | * a matching perf cgroup is scheduled in. | |
641 | */ | |
642 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
643 | event->cgrp_defer_enabled = 1; | |
644 | } | |
645 | ||
646 | static inline void | |
647 | perf_cgroup_mark_enabled(struct perf_event *event, | |
648 | struct perf_event_context *ctx) | |
649 | { | |
650 | struct perf_event *sub; | |
651 | u64 tstamp = perf_event_time(event); | |
652 | ||
653 | if (!event->cgrp_defer_enabled) | |
654 | return; | |
655 | ||
656 | event->cgrp_defer_enabled = 0; | |
657 | ||
658 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
659 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
660 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
661 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
662 | sub->cgrp_defer_enabled = 0; | |
663 | } | |
664 | } | |
665 | } | |
666 | #else /* !CONFIG_CGROUP_PERF */ | |
667 | ||
668 | static inline bool | |
669 | perf_cgroup_match(struct perf_event *event) | |
670 | { | |
671 | return true; | |
672 | } | |
673 | ||
674 | static inline void perf_detach_cgroup(struct perf_event *event) | |
675 | {} | |
676 | ||
677 | static inline int is_cgroup_event(struct perf_event *event) | |
678 | { | |
679 | return 0; | |
680 | } | |
681 | ||
682 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
683 | { | |
684 | return 0; | |
685 | } | |
686 | ||
687 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
688 | { | |
689 | } | |
690 | ||
691 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
692 | { | |
693 | } | |
694 | ||
a8d757ef SE |
695 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
696 | struct task_struct *next) | |
e5d1367f SE |
697 | { |
698 | } | |
699 | ||
a8d757ef SE |
700 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
701 | struct task_struct *task) | |
e5d1367f SE |
702 | { |
703 | } | |
704 | ||
705 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
706 | struct perf_event_attr *attr, | |
707 | struct perf_event *group_leader) | |
708 | { | |
709 | return -EINVAL; | |
710 | } | |
711 | ||
712 | static inline void | |
3f7cce3c SE |
713 | perf_cgroup_set_timestamp(struct task_struct *task, |
714 | struct perf_event_context *ctx) | |
e5d1367f SE |
715 | { |
716 | } | |
717 | ||
718 | void | |
719 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
720 | { | |
721 | } | |
722 | ||
723 | static inline void | |
724 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
725 | { | |
726 | } | |
727 | ||
728 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
729 | { | |
730 | return 0; | |
731 | } | |
732 | ||
733 | static inline void | |
734 | perf_cgroup_defer_enabled(struct perf_event *event) | |
735 | { | |
736 | } | |
737 | ||
738 | static inline void | |
739 | perf_cgroup_mark_enabled(struct perf_event *event, | |
740 | struct perf_event_context *ctx) | |
741 | { | |
742 | } | |
743 | #endif | |
744 | ||
9e630205 SE |
745 | /* |
746 | * set default to be dependent on timer tick just | |
747 | * like original code | |
748 | */ | |
749 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
750 | /* | |
751 | * function must be called with interrupts disbled | |
752 | */ | |
753 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
754 | { | |
755 | struct perf_cpu_context *cpuctx; | |
756 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
757 | int rotations = 0; | |
758 | ||
759 | WARN_ON(!irqs_disabled()); | |
760 | ||
761 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
762 | ||
763 | rotations = perf_rotate_context(cpuctx); | |
764 | ||
765 | /* | |
766 | * arm timer if needed | |
767 | */ | |
768 | if (rotations) { | |
769 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
770 | ret = HRTIMER_RESTART; | |
771 | } | |
772 | ||
773 | return ret; | |
774 | } | |
775 | ||
776 | /* CPU is going down */ | |
777 | void perf_cpu_hrtimer_cancel(int cpu) | |
778 | { | |
779 | struct perf_cpu_context *cpuctx; | |
780 | struct pmu *pmu; | |
781 | unsigned long flags; | |
782 | ||
783 | if (WARN_ON(cpu != smp_processor_id())) | |
784 | return; | |
785 | ||
786 | local_irq_save(flags); | |
787 | ||
788 | rcu_read_lock(); | |
789 | ||
790 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
791 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
792 | ||
793 | if (pmu->task_ctx_nr == perf_sw_context) | |
794 | continue; | |
795 | ||
796 | hrtimer_cancel(&cpuctx->hrtimer); | |
797 | } | |
798 | ||
799 | rcu_read_unlock(); | |
800 | ||
801 | local_irq_restore(flags); | |
802 | } | |
803 | ||
804 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
805 | { | |
806 | struct hrtimer *hr = &cpuctx->hrtimer; | |
807 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 808 | int timer; |
9e630205 SE |
809 | |
810 | /* no multiplexing needed for SW PMU */ | |
811 | if (pmu->task_ctx_nr == perf_sw_context) | |
812 | return; | |
813 | ||
62b85639 SE |
814 | /* |
815 | * check default is sane, if not set then force to | |
816 | * default interval (1/tick) | |
817 | */ | |
818 | timer = pmu->hrtimer_interval_ms; | |
819 | if (timer < 1) | |
820 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
821 | ||
822 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
823 | |
824 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
825 | hr->function = perf_cpu_hrtimer_handler; | |
826 | } | |
827 | ||
828 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
829 | { | |
830 | struct hrtimer *hr = &cpuctx->hrtimer; | |
831 | struct pmu *pmu = cpuctx->ctx.pmu; | |
832 | ||
833 | /* not for SW PMU */ | |
834 | if (pmu->task_ctx_nr == perf_sw_context) | |
835 | return; | |
836 | ||
837 | if (hrtimer_active(hr)) | |
838 | return; | |
839 | ||
840 | if (!hrtimer_callback_running(hr)) | |
841 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
842 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
843 | } | |
844 | ||
33696fc0 | 845 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 846 | { |
33696fc0 PZ |
847 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
848 | if (!(*count)++) | |
849 | pmu->pmu_disable(pmu); | |
9e35ad38 | 850 | } |
9e35ad38 | 851 | |
33696fc0 | 852 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 853 | { |
33696fc0 PZ |
854 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
855 | if (!--(*count)) | |
856 | pmu->pmu_enable(pmu); | |
9e35ad38 | 857 | } |
9e35ad38 | 858 | |
e9d2b064 PZ |
859 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
860 | ||
861 | /* | |
862 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
863 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
864 | * disabled, while rotate_context is called from IRQ context. | |
865 | */ | |
108b02cf | 866 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 867 | { |
108b02cf | 868 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 869 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 870 | |
e9d2b064 | 871 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 872 | |
d84153d6 | 873 | if (list_empty(&cpuctx->rotation_list)) |
e9d2b064 | 874 | list_add(&cpuctx->rotation_list, head); |
9e35ad38 | 875 | } |
9e35ad38 | 876 | |
cdd6c482 | 877 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 878 | { |
e5289d4a | 879 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
880 | } |
881 | ||
cdd6c482 | 882 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 883 | { |
564c2b21 PM |
884 | if (atomic_dec_and_test(&ctx->refcount)) { |
885 | if (ctx->parent_ctx) | |
886 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
887 | if (ctx->task) |
888 | put_task_struct(ctx->task); | |
cb796ff3 | 889 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 890 | } |
a63eaf34 PM |
891 | } |
892 | ||
cdd6c482 | 893 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
894 | { |
895 | if (ctx->parent_ctx) { | |
896 | put_ctx(ctx->parent_ctx); | |
897 | ctx->parent_ctx = NULL; | |
898 | } | |
899 | } | |
900 | ||
6844c09d ACM |
901 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
902 | { | |
903 | /* | |
904 | * only top level events have the pid namespace they were created in | |
905 | */ | |
906 | if (event->parent) | |
907 | event = event->parent; | |
908 | ||
909 | return task_tgid_nr_ns(p, event->ns); | |
910 | } | |
911 | ||
912 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
913 | { | |
914 | /* | |
915 | * only top level events have the pid namespace they were created in | |
916 | */ | |
917 | if (event->parent) | |
918 | event = event->parent; | |
919 | ||
920 | return task_pid_nr_ns(p, event->ns); | |
921 | } | |
922 | ||
7f453c24 | 923 | /* |
cdd6c482 | 924 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
925 | * to userspace. |
926 | */ | |
cdd6c482 | 927 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 928 | { |
cdd6c482 | 929 | u64 id = event->id; |
7f453c24 | 930 | |
cdd6c482 IM |
931 | if (event->parent) |
932 | id = event->parent->id; | |
7f453c24 PZ |
933 | |
934 | return id; | |
935 | } | |
936 | ||
25346b93 | 937 | /* |
cdd6c482 | 938 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
939 | * This has to cope with with the fact that until it is locked, |
940 | * the context could get moved to another task. | |
941 | */ | |
cdd6c482 | 942 | static struct perf_event_context * |
8dc85d54 | 943 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 944 | { |
cdd6c482 | 945 | struct perf_event_context *ctx; |
25346b93 | 946 | |
9ed6060d | 947 | retry: |
058ebd0e PZ |
948 | /* |
949 | * One of the few rules of preemptible RCU is that one cannot do | |
950 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
951 | * part of the read side critical section was preemptible -- see | |
952 | * rcu_read_unlock_special(). | |
953 | * | |
954 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
955 | * side critical section is non-preemptible. | |
956 | */ | |
957 | preempt_disable(); | |
958 | rcu_read_lock(); | |
8dc85d54 | 959 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
960 | if (ctx) { |
961 | /* | |
962 | * If this context is a clone of another, it might | |
963 | * get swapped for another underneath us by | |
cdd6c482 | 964 | * perf_event_task_sched_out, though the |
25346b93 PM |
965 | * rcu_read_lock() protects us from any context |
966 | * getting freed. Lock the context and check if it | |
967 | * got swapped before we could get the lock, and retry | |
968 | * if so. If we locked the right context, then it | |
969 | * can't get swapped on us any more. | |
970 | */ | |
e625cce1 | 971 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 972 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 973 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
974 | rcu_read_unlock(); |
975 | preempt_enable(); | |
25346b93 PM |
976 | goto retry; |
977 | } | |
b49a9e7e PZ |
978 | |
979 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 980 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
981 | ctx = NULL; |
982 | } | |
25346b93 PM |
983 | } |
984 | rcu_read_unlock(); | |
058ebd0e | 985 | preempt_enable(); |
25346b93 PM |
986 | return ctx; |
987 | } | |
988 | ||
989 | /* | |
990 | * Get the context for a task and increment its pin_count so it | |
991 | * can't get swapped to another task. This also increments its | |
992 | * reference count so that the context can't get freed. | |
993 | */ | |
8dc85d54 PZ |
994 | static struct perf_event_context * |
995 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 996 | { |
cdd6c482 | 997 | struct perf_event_context *ctx; |
25346b93 PM |
998 | unsigned long flags; |
999 | ||
8dc85d54 | 1000 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1001 | if (ctx) { |
1002 | ++ctx->pin_count; | |
e625cce1 | 1003 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1004 | } |
1005 | return ctx; | |
1006 | } | |
1007 | ||
cdd6c482 | 1008 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1009 | { |
1010 | unsigned long flags; | |
1011 | ||
e625cce1 | 1012 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1013 | --ctx->pin_count; |
e625cce1 | 1014 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1015 | } |
1016 | ||
f67218c3 PZ |
1017 | /* |
1018 | * Update the record of the current time in a context. | |
1019 | */ | |
1020 | static void update_context_time(struct perf_event_context *ctx) | |
1021 | { | |
1022 | u64 now = perf_clock(); | |
1023 | ||
1024 | ctx->time += now - ctx->timestamp; | |
1025 | ctx->timestamp = now; | |
1026 | } | |
1027 | ||
4158755d SE |
1028 | static u64 perf_event_time(struct perf_event *event) |
1029 | { | |
1030 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1031 | |
1032 | if (is_cgroup_event(event)) | |
1033 | return perf_cgroup_event_time(event); | |
1034 | ||
4158755d SE |
1035 | return ctx ? ctx->time : 0; |
1036 | } | |
1037 | ||
f67218c3 PZ |
1038 | /* |
1039 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1040 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1041 | */ |
1042 | static void update_event_times(struct perf_event *event) | |
1043 | { | |
1044 | struct perf_event_context *ctx = event->ctx; | |
1045 | u64 run_end; | |
1046 | ||
1047 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1048 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1049 | return; | |
e5d1367f SE |
1050 | /* |
1051 | * in cgroup mode, time_enabled represents | |
1052 | * the time the event was enabled AND active | |
1053 | * tasks were in the monitored cgroup. This is | |
1054 | * independent of the activity of the context as | |
1055 | * there may be a mix of cgroup and non-cgroup events. | |
1056 | * | |
1057 | * That is why we treat cgroup events differently | |
1058 | * here. | |
1059 | */ | |
1060 | if (is_cgroup_event(event)) | |
46cd6a7f | 1061 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1062 | else if (ctx->is_active) |
1063 | run_end = ctx->time; | |
acd1d7c1 PZ |
1064 | else |
1065 | run_end = event->tstamp_stopped; | |
1066 | ||
1067 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1068 | |
1069 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1070 | run_end = event->tstamp_stopped; | |
1071 | else | |
4158755d | 1072 | run_end = perf_event_time(event); |
f67218c3 PZ |
1073 | |
1074 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1075 | |
f67218c3 PZ |
1076 | } |
1077 | ||
96c21a46 PZ |
1078 | /* |
1079 | * Update total_time_enabled and total_time_running for all events in a group. | |
1080 | */ | |
1081 | static void update_group_times(struct perf_event *leader) | |
1082 | { | |
1083 | struct perf_event *event; | |
1084 | ||
1085 | update_event_times(leader); | |
1086 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1087 | update_event_times(event); | |
1088 | } | |
1089 | ||
889ff015 FW |
1090 | static struct list_head * |
1091 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1092 | { | |
1093 | if (event->attr.pinned) | |
1094 | return &ctx->pinned_groups; | |
1095 | else | |
1096 | return &ctx->flexible_groups; | |
1097 | } | |
1098 | ||
fccc714b | 1099 | /* |
cdd6c482 | 1100 | * Add a event from the lists for its context. |
fccc714b PZ |
1101 | * Must be called with ctx->mutex and ctx->lock held. |
1102 | */ | |
04289bb9 | 1103 | static void |
cdd6c482 | 1104 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1105 | { |
8a49542c PZ |
1106 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1107 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1108 | |
1109 | /* | |
8a49542c PZ |
1110 | * If we're a stand alone event or group leader, we go to the context |
1111 | * list, group events are kept attached to the group so that | |
1112 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1113 | */ |
8a49542c | 1114 | if (event->group_leader == event) { |
889ff015 FW |
1115 | struct list_head *list; |
1116 | ||
d6f962b5 FW |
1117 | if (is_software_event(event)) |
1118 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1119 | ||
889ff015 FW |
1120 | list = ctx_group_list(event, ctx); |
1121 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1122 | } |
592903cd | 1123 | |
08309379 | 1124 | if (is_cgroup_event(event)) |
e5d1367f | 1125 | ctx->nr_cgroups++; |
e5d1367f | 1126 | |
d010b332 SE |
1127 | if (has_branch_stack(event)) |
1128 | ctx->nr_branch_stack++; | |
1129 | ||
cdd6c482 | 1130 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1131 | if (!ctx->nr_events) |
108b02cf | 1132 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1133 | ctx->nr_events++; |
1134 | if (event->attr.inherit_stat) | |
bfbd3381 | 1135 | ctx->nr_stat++; |
04289bb9 IM |
1136 | } |
1137 | ||
0231bb53 JO |
1138 | /* |
1139 | * Initialize event state based on the perf_event_attr::disabled. | |
1140 | */ | |
1141 | static inline void perf_event__state_init(struct perf_event *event) | |
1142 | { | |
1143 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1144 | PERF_EVENT_STATE_INACTIVE; | |
1145 | } | |
1146 | ||
c320c7b7 ACM |
1147 | /* |
1148 | * Called at perf_event creation and when events are attached/detached from a | |
1149 | * group. | |
1150 | */ | |
1151 | static void perf_event__read_size(struct perf_event *event) | |
1152 | { | |
1153 | int entry = sizeof(u64); /* value */ | |
1154 | int size = 0; | |
1155 | int nr = 1; | |
1156 | ||
1157 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1158 | size += sizeof(u64); | |
1159 | ||
1160 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1161 | size += sizeof(u64); | |
1162 | ||
1163 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1164 | entry += sizeof(u64); | |
1165 | ||
1166 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1167 | nr += event->group_leader->nr_siblings; | |
1168 | size += sizeof(u64); | |
1169 | } | |
1170 | ||
1171 | size += entry * nr; | |
1172 | event->read_size = size; | |
1173 | } | |
1174 | ||
1175 | static void perf_event__header_size(struct perf_event *event) | |
1176 | { | |
1177 | struct perf_sample_data *data; | |
1178 | u64 sample_type = event->attr.sample_type; | |
1179 | u16 size = 0; | |
1180 | ||
1181 | perf_event__read_size(event); | |
1182 | ||
1183 | if (sample_type & PERF_SAMPLE_IP) | |
1184 | size += sizeof(data->ip); | |
1185 | ||
6844c09d ACM |
1186 | if (sample_type & PERF_SAMPLE_ADDR) |
1187 | size += sizeof(data->addr); | |
1188 | ||
1189 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1190 | size += sizeof(data->period); | |
1191 | ||
c3feedf2 AK |
1192 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1193 | size += sizeof(data->weight); | |
1194 | ||
6844c09d ACM |
1195 | if (sample_type & PERF_SAMPLE_READ) |
1196 | size += event->read_size; | |
1197 | ||
d6be9ad6 SE |
1198 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1199 | size += sizeof(data->data_src.val); | |
1200 | ||
6844c09d ACM |
1201 | event->header_size = size; |
1202 | } | |
1203 | ||
1204 | static void perf_event__id_header_size(struct perf_event *event) | |
1205 | { | |
1206 | struct perf_sample_data *data; | |
1207 | u64 sample_type = event->attr.sample_type; | |
1208 | u16 size = 0; | |
1209 | ||
c320c7b7 ACM |
1210 | if (sample_type & PERF_SAMPLE_TID) |
1211 | size += sizeof(data->tid_entry); | |
1212 | ||
1213 | if (sample_type & PERF_SAMPLE_TIME) | |
1214 | size += sizeof(data->time); | |
1215 | ||
c320c7b7 ACM |
1216 | if (sample_type & PERF_SAMPLE_ID) |
1217 | size += sizeof(data->id); | |
1218 | ||
1219 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1220 | size += sizeof(data->stream_id); | |
1221 | ||
1222 | if (sample_type & PERF_SAMPLE_CPU) | |
1223 | size += sizeof(data->cpu_entry); | |
1224 | ||
6844c09d | 1225 | event->id_header_size = size; |
c320c7b7 ACM |
1226 | } |
1227 | ||
8a49542c PZ |
1228 | static void perf_group_attach(struct perf_event *event) |
1229 | { | |
c320c7b7 | 1230 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1231 | |
74c3337c PZ |
1232 | /* |
1233 | * We can have double attach due to group movement in perf_event_open. | |
1234 | */ | |
1235 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1236 | return; | |
1237 | ||
8a49542c PZ |
1238 | event->attach_state |= PERF_ATTACH_GROUP; |
1239 | ||
1240 | if (group_leader == event) | |
1241 | return; | |
1242 | ||
1243 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1244 | !is_software_event(event)) | |
1245 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1246 | ||
1247 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1248 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1249 | |
1250 | perf_event__header_size(group_leader); | |
1251 | ||
1252 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1253 | perf_event__header_size(pos); | |
8a49542c PZ |
1254 | } |
1255 | ||
a63eaf34 | 1256 | /* |
cdd6c482 | 1257 | * Remove a event from the lists for its context. |
fccc714b | 1258 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1259 | */ |
04289bb9 | 1260 | static void |
cdd6c482 | 1261 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1262 | { |
68cacd29 | 1263 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1264 | /* |
1265 | * We can have double detach due to exit/hot-unplug + close. | |
1266 | */ | |
1267 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1268 | return; |
8a49542c PZ |
1269 | |
1270 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1271 | ||
68cacd29 | 1272 | if (is_cgroup_event(event)) { |
e5d1367f | 1273 | ctx->nr_cgroups--; |
68cacd29 SE |
1274 | cpuctx = __get_cpu_context(ctx); |
1275 | /* | |
1276 | * if there are no more cgroup events | |
1277 | * then cler cgrp to avoid stale pointer | |
1278 | * in update_cgrp_time_from_cpuctx() | |
1279 | */ | |
1280 | if (!ctx->nr_cgroups) | |
1281 | cpuctx->cgrp = NULL; | |
1282 | } | |
e5d1367f | 1283 | |
d010b332 SE |
1284 | if (has_branch_stack(event)) |
1285 | ctx->nr_branch_stack--; | |
1286 | ||
cdd6c482 IM |
1287 | ctx->nr_events--; |
1288 | if (event->attr.inherit_stat) | |
bfbd3381 | 1289 | ctx->nr_stat--; |
8bc20959 | 1290 | |
cdd6c482 | 1291 | list_del_rcu(&event->event_entry); |
04289bb9 | 1292 | |
8a49542c PZ |
1293 | if (event->group_leader == event) |
1294 | list_del_init(&event->group_entry); | |
5c148194 | 1295 | |
96c21a46 | 1296 | update_group_times(event); |
b2e74a26 SE |
1297 | |
1298 | /* | |
1299 | * If event was in error state, then keep it | |
1300 | * that way, otherwise bogus counts will be | |
1301 | * returned on read(). The only way to get out | |
1302 | * of error state is by explicit re-enabling | |
1303 | * of the event | |
1304 | */ | |
1305 | if (event->state > PERF_EVENT_STATE_OFF) | |
1306 | event->state = PERF_EVENT_STATE_OFF; | |
050735b0 PZ |
1307 | } |
1308 | ||
8a49542c | 1309 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1310 | { |
1311 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1312 | struct list_head *list = NULL; |
1313 | ||
1314 | /* | |
1315 | * We can have double detach due to exit/hot-unplug + close. | |
1316 | */ | |
1317 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1318 | return; | |
1319 | ||
1320 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1321 | ||
1322 | /* | |
1323 | * If this is a sibling, remove it from its group. | |
1324 | */ | |
1325 | if (event->group_leader != event) { | |
1326 | list_del_init(&event->group_entry); | |
1327 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1328 | goto out; |
8a49542c PZ |
1329 | } |
1330 | ||
1331 | if (!list_empty(&event->group_entry)) | |
1332 | list = &event->group_entry; | |
2e2af50b | 1333 | |
04289bb9 | 1334 | /* |
cdd6c482 IM |
1335 | * If this was a group event with sibling events then |
1336 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1337 | * to whatever list we are on. |
04289bb9 | 1338 | */ |
cdd6c482 | 1339 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1340 | if (list) |
1341 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1342 | sibling->group_leader = sibling; |
d6f962b5 FW |
1343 | |
1344 | /* Inherit group flags from the previous leader */ | |
1345 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1346 | } |
c320c7b7 ACM |
1347 | |
1348 | out: | |
1349 | perf_event__header_size(event->group_leader); | |
1350 | ||
1351 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1352 | perf_event__header_size(tmp); | |
04289bb9 IM |
1353 | } |
1354 | ||
fa66f07a SE |
1355 | static inline int |
1356 | event_filter_match(struct perf_event *event) | |
1357 | { | |
e5d1367f SE |
1358 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1359 | && perf_cgroup_match(event); | |
fa66f07a SE |
1360 | } |
1361 | ||
9ffcfa6f SE |
1362 | static void |
1363 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1364 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1365 | struct perf_event_context *ctx) |
3b6f9e5c | 1366 | { |
4158755d | 1367 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1368 | u64 delta; |
1369 | /* | |
1370 | * An event which could not be activated because of | |
1371 | * filter mismatch still needs to have its timings | |
1372 | * maintained, otherwise bogus information is return | |
1373 | * via read() for time_enabled, time_running: | |
1374 | */ | |
1375 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1376 | && !event_filter_match(event)) { | |
e5d1367f | 1377 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1378 | event->tstamp_running += delta; |
4158755d | 1379 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1380 | } |
1381 | ||
cdd6c482 | 1382 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1383 | return; |
3b6f9e5c | 1384 | |
cdd6c482 IM |
1385 | event->state = PERF_EVENT_STATE_INACTIVE; |
1386 | if (event->pending_disable) { | |
1387 | event->pending_disable = 0; | |
1388 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1389 | } |
4158755d | 1390 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1391 | event->pmu->del(event, 0); |
cdd6c482 | 1392 | event->oncpu = -1; |
3b6f9e5c | 1393 | |
cdd6c482 | 1394 | if (!is_software_event(event)) |
3b6f9e5c PM |
1395 | cpuctx->active_oncpu--; |
1396 | ctx->nr_active--; | |
0f5a2601 PZ |
1397 | if (event->attr.freq && event->attr.sample_freq) |
1398 | ctx->nr_freq--; | |
cdd6c482 | 1399 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c PM |
1400 | cpuctx->exclusive = 0; |
1401 | } | |
1402 | ||
d859e29f | 1403 | static void |
cdd6c482 | 1404 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1405 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1406 | struct perf_event_context *ctx) |
d859e29f | 1407 | { |
cdd6c482 | 1408 | struct perf_event *event; |
fa66f07a | 1409 | int state = group_event->state; |
d859e29f | 1410 | |
cdd6c482 | 1411 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1412 | |
1413 | /* | |
1414 | * Schedule out siblings (if any): | |
1415 | */ | |
cdd6c482 IM |
1416 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1417 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1418 | |
fa66f07a | 1419 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1420 | cpuctx->exclusive = 0; |
1421 | } | |
1422 | ||
0793a61d | 1423 | /* |
cdd6c482 | 1424 | * Cross CPU call to remove a performance event |
0793a61d | 1425 | * |
cdd6c482 | 1426 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1427 | * remove it from the context list. |
1428 | */ | |
fe4b04fa | 1429 | static int __perf_remove_from_context(void *info) |
0793a61d | 1430 | { |
cdd6c482 IM |
1431 | struct perf_event *event = info; |
1432 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1433 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1434 | |
e625cce1 | 1435 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1436 | event_sched_out(event, cpuctx, ctx); |
cdd6c482 | 1437 | list_del_event(event, ctx); |
64ce3126 PZ |
1438 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1439 | ctx->is_active = 0; | |
1440 | cpuctx->task_ctx = NULL; | |
1441 | } | |
e625cce1 | 1442 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1443 | |
1444 | return 0; | |
0793a61d TG |
1445 | } |
1446 | ||
1447 | ||
1448 | /* | |
cdd6c482 | 1449 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1450 | * |
cdd6c482 | 1451 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1452 | * call when the task is on a CPU. |
c93f7669 | 1453 | * |
cdd6c482 IM |
1454 | * If event->ctx is a cloned context, callers must make sure that |
1455 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1456 | * remains valid. This is OK when called from perf_release since |
1457 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1458 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1459 | * context has been detached from its task. |
0793a61d | 1460 | */ |
fe4b04fa | 1461 | static void perf_remove_from_context(struct perf_event *event) |
0793a61d | 1462 | { |
cdd6c482 | 1463 | struct perf_event_context *ctx = event->ctx; |
0793a61d TG |
1464 | struct task_struct *task = ctx->task; |
1465 | ||
fe4b04fa PZ |
1466 | lockdep_assert_held(&ctx->mutex); |
1467 | ||
0793a61d TG |
1468 | if (!task) { |
1469 | /* | |
cdd6c482 | 1470 | * Per cpu events are removed via an smp call and |
af901ca1 | 1471 | * the removal is always successful. |
0793a61d | 1472 | */ |
fe4b04fa | 1473 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
0793a61d TG |
1474 | return; |
1475 | } | |
1476 | ||
1477 | retry: | |
fe4b04fa PZ |
1478 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1479 | return; | |
0793a61d | 1480 | |
e625cce1 | 1481 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1482 | /* |
fe4b04fa PZ |
1483 | * If we failed to find a running task, but find the context active now |
1484 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1485 | */ |
fe4b04fa | 1486 | if (ctx->is_active) { |
e625cce1 | 1487 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1488 | goto retry; |
1489 | } | |
1490 | ||
1491 | /* | |
fe4b04fa PZ |
1492 | * Since the task isn't running, its safe to remove the event, us |
1493 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1494 | */ |
fe4b04fa | 1495 | list_del_event(event, ctx); |
e625cce1 | 1496 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1497 | } |
1498 | ||
d859e29f | 1499 | /* |
cdd6c482 | 1500 | * Cross CPU call to disable a performance event |
d859e29f | 1501 | */ |
500ad2d8 | 1502 | int __perf_event_disable(void *info) |
d859e29f | 1503 | { |
cdd6c482 | 1504 | struct perf_event *event = info; |
cdd6c482 | 1505 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1506 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1507 | |
1508 | /* | |
cdd6c482 IM |
1509 | * If this is a per-task event, need to check whether this |
1510 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1511 | * |
1512 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1513 | * flipping contexts around. | |
d859e29f | 1514 | */ |
665c2142 | 1515 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1516 | return -EINVAL; |
d859e29f | 1517 | |
e625cce1 | 1518 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1519 | |
1520 | /* | |
cdd6c482 | 1521 | * If the event is on, turn it off. |
d859e29f PM |
1522 | * If it is in error state, leave it in error state. |
1523 | */ | |
cdd6c482 | 1524 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1525 | update_context_time(ctx); |
e5d1367f | 1526 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1527 | update_group_times(event); |
1528 | if (event == event->group_leader) | |
1529 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1530 | else |
cdd6c482 IM |
1531 | event_sched_out(event, cpuctx, ctx); |
1532 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1533 | } |
1534 | ||
e625cce1 | 1535 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1536 | |
1537 | return 0; | |
d859e29f PM |
1538 | } |
1539 | ||
1540 | /* | |
cdd6c482 | 1541 | * Disable a event. |
c93f7669 | 1542 | * |
cdd6c482 IM |
1543 | * If event->ctx is a cloned context, callers must make sure that |
1544 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1545 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1546 | * perf_event_for_each_child or perf_event_for_each because they |
1547 | * hold the top-level event's child_mutex, so any descendant that | |
1548 | * goes to exit will block in sync_child_event. | |
1549 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1550 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1551 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1552 | */ |
44234adc | 1553 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1554 | { |
cdd6c482 | 1555 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1556 | struct task_struct *task = ctx->task; |
1557 | ||
1558 | if (!task) { | |
1559 | /* | |
cdd6c482 | 1560 | * Disable the event on the cpu that it's on |
d859e29f | 1561 | */ |
fe4b04fa | 1562 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1563 | return; |
1564 | } | |
1565 | ||
9ed6060d | 1566 | retry: |
fe4b04fa PZ |
1567 | if (!task_function_call(task, __perf_event_disable, event)) |
1568 | return; | |
d859e29f | 1569 | |
e625cce1 | 1570 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1571 | /* |
cdd6c482 | 1572 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1573 | */ |
cdd6c482 | 1574 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1575 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1576 | /* |
1577 | * Reload the task pointer, it might have been changed by | |
1578 | * a concurrent perf_event_context_sched_out(). | |
1579 | */ | |
1580 | task = ctx->task; | |
d859e29f PM |
1581 | goto retry; |
1582 | } | |
1583 | ||
1584 | /* | |
1585 | * Since we have the lock this context can't be scheduled | |
1586 | * in, so we can change the state safely. | |
1587 | */ | |
cdd6c482 IM |
1588 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1589 | update_group_times(event); | |
1590 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1591 | } |
e625cce1 | 1592 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1593 | } |
dcfce4a0 | 1594 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1595 | |
e5d1367f SE |
1596 | static void perf_set_shadow_time(struct perf_event *event, |
1597 | struct perf_event_context *ctx, | |
1598 | u64 tstamp) | |
1599 | { | |
1600 | /* | |
1601 | * use the correct time source for the time snapshot | |
1602 | * | |
1603 | * We could get by without this by leveraging the | |
1604 | * fact that to get to this function, the caller | |
1605 | * has most likely already called update_context_time() | |
1606 | * and update_cgrp_time_xx() and thus both timestamp | |
1607 | * are identical (or very close). Given that tstamp is, | |
1608 | * already adjusted for cgroup, we could say that: | |
1609 | * tstamp - ctx->timestamp | |
1610 | * is equivalent to | |
1611 | * tstamp - cgrp->timestamp. | |
1612 | * | |
1613 | * Then, in perf_output_read(), the calculation would | |
1614 | * work with no changes because: | |
1615 | * - event is guaranteed scheduled in | |
1616 | * - no scheduled out in between | |
1617 | * - thus the timestamp would be the same | |
1618 | * | |
1619 | * But this is a bit hairy. | |
1620 | * | |
1621 | * So instead, we have an explicit cgroup call to remain | |
1622 | * within the time time source all along. We believe it | |
1623 | * is cleaner and simpler to understand. | |
1624 | */ | |
1625 | if (is_cgroup_event(event)) | |
1626 | perf_cgroup_set_shadow_time(event, tstamp); | |
1627 | else | |
1628 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1629 | } | |
1630 | ||
4fe757dd PZ |
1631 | #define MAX_INTERRUPTS (~0ULL) |
1632 | ||
1633 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1634 | ||
235c7fc7 | 1635 | static int |
9ffcfa6f | 1636 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1637 | struct perf_cpu_context *cpuctx, |
6e37738a | 1638 | struct perf_event_context *ctx) |
235c7fc7 | 1639 | { |
4158755d SE |
1640 | u64 tstamp = perf_event_time(event); |
1641 | ||
cdd6c482 | 1642 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1643 | return 0; |
1644 | ||
cdd6c482 | 1645 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1646 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1647 | |
1648 | /* | |
1649 | * Unthrottle events, since we scheduled we might have missed several | |
1650 | * ticks already, also for a heavily scheduling task there is little | |
1651 | * guarantee it'll get a tick in a timely manner. | |
1652 | */ | |
1653 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1654 | perf_log_throttle(event, 1); | |
1655 | event->hw.interrupts = 0; | |
1656 | } | |
1657 | ||
235c7fc7 IM |
1658 | /* |
1659 | * The new state must be visible before we turn it on in the hardware: | |
1660 | */ | |
1661 | smp_wmb(); | |
1662 | ||
a4eaf7f1 | 1663 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1664 | event->state = PERF_EVENT_STATE_INACTIVE; |
1665 | event->oncpu = -1; | |
235c7fc7 IM |
1666 | return -EAGAIN; |
1667 | } | |
1668 | ||
4158755d | 1669 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1670 | |
e5d1367f | 1671 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1672 | |
cdd6c482 | 1673 | if (!is_software_event(event)) |
3b6f9e5c | 1674 | cpuctx->active_oncpu++; |
235c7fc7 | 1675 | ctx->nr_active++; |
0f5a2601 PZ |
1676 | if (event->attr.freq && event->attr.sample_freq) |
1677 | ctx->nr_freq++; | |
235c7fc7 | 1678 | |
cdd6c482 | 1679 | if (event->attr.exclusive) |
3b6f9e5c PM |
1680 | cpuctx->exclusive = 1; |
1681 | ||
235c7fc7 IM |
1682 | return 0; |
1683 | } | |
1684 | ||
6751b71e | 1685 | static int |
cdd6c482 | 1686 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1687 | struct perf_cpu_context *cpuctx, |
6e37738a | 1688 | struct perf_event_context *ctx) |
6751b71e | 1689 | { |
6bde9b6c | 1690 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1691 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1692 | u64 now = ctx->time; |
1693 | bool simulate = false; | |
6751b71e | 1694 | |
cdd6c482 | 1695 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1696 | return 0; |
1697 | ||
ad5133b7 | 1698 | pmu->start_txn(pmu); |
6bde9b6c | 1699 | |
9ffcfa6f | 1700 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1701 | pmu->cancel_txn(pmu); |
9e630205 | 1702 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1703 | return -EAGAIN; |
90151c35 | 1704 | } |
6751b71e PM |
1705 | |
1706 | /* | |
1707 | * Schedule in siblings as one group (if any): | |
1708 | */ | |
cdd6c482 | 1709 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1710 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1711 | partial_group = event; |
6751b71e PM |
1712 | goto group_error; |
1713 | } | |
1714 | } | |
1715 | ||
9ffcfa6f | 1716 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1717 | return 0; |
9ffcfa6f | 1718 | |
6751b71e PM |
1719 | group_error: |
1720 | /* | |
1721 | * Groups can be scheduled in as one unit only, so undo any | |
1722 | * partial group before returning: | |
d7842da4 SE |
1723 | * The events up to the failed event are scheduled out normally, |
1724 | * tstamp_stopped will be updated. | |
1725 | * | |
1726 | * The failed events and the remaining siblings need to have | |
1727 | * their timings updated as if they had gone thru event_sched_in() | |
1728 | * and event_sched_out(). This is required to get consistent timings | |
1729 | * across the group. This also takes care of the case where the group | |
1730 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1731 | * the time the event was actually stopped, such that time delta | |
1732 | * calculation in update_event_times() is correct. | |
6751b71e | 1733 | */ |
cdd6c482 IM |
1734 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1735 | if (event == partial_group) | |
d7842da4 SE |
1736 | simulate = true; |
1737 | ||
1738 | if (simulate) { | |
1739 | event->tstamp_running += now - event->tstamp_stopped; | |
1740 | event->tstamp_stopped = now; | |
1741 | } else { | |
1742 | event_sched_out(event, cpuctx, ctx); | |
1743 | } | |
6751b71e | 1744 | } |
9ffcfa6f | 1745 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1746 | |
ad5133b7 | 1747 | pmu->cancel_txn(pmu); |
90151c35 | 1748 | |
9e630205 SE |
1749 | perf_cpu_hrtimer_restart(cpuctx); |
1750 | ||
6751b71e PM |
1751 | return -EAGAIN; |
1752 | } | |
1753 | ||
3b6f9e5c | 1754 | /* |
cdd6c482 | 1755 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1756 | */ |
cdd6c482 | 1757 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1758 | struct perf_cpu_context *cpuctx, |
1759 | int can_add_hw) | |
1760 | { | |
1761 | /* | |
cdd6c482 | 1762 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1763 | */ |
d6f962b5 | 1764 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1765 | return 1; |
1766 | /* | |
1767 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1768 | * events can go on. |
3b6f9e5c PM |
1769 | */ |
1770 | if (cpuctx->exclusive) | |
1771 | return 0; | |
1772 | /* | |
1773 | * If this group is exclusive and there are already | |
cdd6c482 | 1774 | * events on the CPU, it can't go on. |
3b6f9e5c | 1775 | */ |
cdd6c482 | 1776 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1777 | return 0; |
1778 | /* | |
1779 | * Otherwise, try to add it if all previous groups were able | |
1780 | * to go on. | |
1781 | */ | |
1782 | return can_add_hw; | |
1783 | } | |
1784 | ||
cdd6c482 IM |
1785 | static void add_event_to_ctx(struct perf_event *event, |
1786 | struct perf_event_context *ctx) | |
53cfbf59 | 1787 | { |
4158755d SE |
1788 | u64 tstamp = perf_event_time(event); |
1789 | ||
cdd6c482 | 1790 | list_add_event(event, ctx); |
8a49542c | 1791 | perf_group_attach(event); |
4158755d SE |
1792 | event->tstamp_enabled = tstamp; |
1793 | event->tstamp_running = tstamp; | |
1794 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1795 | } |
1796 | ||
2c29ef0f PZ |
1797 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1798 | static void | |
1799 | ctx_sched_in(struct perf_event_context *ctx, | |
1800 | struct perf_cpu_context *cpuctx, | |
1801 | enum event_type_t event_type, | |
1802 | struct task_struct *task); | |
fe4b04fa | 1803 | |
dce5855b PZ |
1804 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1805 | struct perf_event_context *ctx, | |
1806 | struct task_struct *task) | |
1807 | { | |
1808 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1809 | if (ctx) | |
1810 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1811 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1812 | if (ctx) | |
1813 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1814 | } | |
1815 | ||
0793a61d | 1816 | /* |
cdd6c482 | 1817 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1818 | * |
1819 | * Must be called with ctx->mutex held | |
0793a61d | 1820 | */ |
fe4b04fa | 1821 | static int __perf_install_in_context(void *info) |
0793a61d | 1822 | { |
cdd6c482 IM |
1823 | struct perf_event *event = info; |
1824 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1825 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1826 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1827 | struct task_struct *task = current; | |
1828 | ||
b58f6b0d | 1829 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1830 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1831 | |
1832 | /* | |
2c29ef0f | 1833 | * If there was an active task_ctx schedule it out. |
0793a61d | 1834 | */ |
b58f6b0d | 1835 | if (task_ctx) |
2c29ef0f | 1836 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1837 | |
1838 | /* | |
1839 | * If the context we're installing events in is not the | |
1840 | * active task_ctx, flip them. | |
1841 | */ | |
1842 | if (ctx->task && task_ctx != ctx) { | |
1843 | if (task_ctx) | |
1844 | raw_spin_unlock(&task_ctx->lock); | |
1845 | raw_spin_lock(&ctx->lock); | |
1846 | task_ctx = ctx; | |
1847 | } | |
1848 | ||
1849 | if (task_ctx) { | |
1850 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1851 | task = task_ctx->task; |
1852 | } | |
b58f6b0d | 1853 | |
2c29ef0f | 1854 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1855 | |
4af4998b | 1856 | update_context_time(ctx); |
e5d1367f SE |
1857 | /* |
1858 | * update cgrp time only if current cgrp | |
1859 | * matches event->cgrp. Must be done before | |
1860 | * calling add_event_to_ctx() | |
1861 | */ | |
1862 | update_cgrp_time_from_event(event); | |
0793a61d | 1863 | |
cdd6c482 | 1864 | add_event_to_ctx(event, ctx); |
0793a61d | 1865 | |
d859e29f | 1866 | /* |
2c29ef0f | 1867 | * Schedule everything back in |
d859e29f | 1868 | */ |
dce5855b | 1869 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1870 | |
1871 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1872 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa | 1873 | |
d84153d6 FW |
1874 | if (atomic_read(&__get_cpu_var(perf_freq_events))) |
1875 | tick_nohz_full_kick(); | |
1876 | ||
fe4b04fa | 1877 | return 0; |
0793a61d TG |
1878 | } |
1879 | ||
1880 | /* | |
cdd6c482 | 1881 | * Attach a performance event to a context |
0793a61d | 1882 | * |
cdd6c482 IM |
1883 | * First we add the event to the list with the hardware enable bit |
1884 | * in event->hw_config cleared. | |
0793a61d | 1885 | * |
cdd6c482 | 1886 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1887 | * call to enable it in the task context. The task might have been |
1888 | * scheduled away, but we check this in the smp call again. | |
1889 | */ | |
1890 | static void | |
cdd6c482 IM |
1891 | perf_install_in_context(struct perf_event_context *ctx, |
1892 | struct perf_event *event, | |
0793a61d TG |
1893 | int cpu) |
1894 | { | |
1895 | struct task_struct *task = ctx->task; | |
1896 | ||
fe4b04fa PZ |
1897 | lockdep_assert_held(&ctx->mutex); |
1898 | ||
c3f00c70 | 1899 | event->ctx = ctx; |
0cda4c02 YZ |
1900 | if (event->cpu != -1) |
1901 | event->cpu = cpu; | |
c3f00c70 | 1902 | |
0793a61d TG |
1903 | if (!task) { |
1904 | /* | |
cdd6c482 | 1905 | * Per cpu events are installed via an smp call and |
af901ca1 | 1906 | * the install is always successful. |
0793a61d | 1907 | */ |
fe4b04fa | 1908 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1909 | return; |
1910 | } | |
1911 | ||
0793a61d | 1912 | retry: |
fe4b04fa PZ |
1913 | if (!task_function_call(task, __perf_install_in_context, event)) |
1914 | return; | |
0793a61d | 1915 | |
e625cce1 | 1916 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1917 | /* |
fe4b04fa PZ |
1918 | * If we failed to find a running task, but find the context active now |
1919 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1920 | */ |
fe4b04fa | 1921 | if (ctx->is_active) { |
e625cce1 | 1922 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1923 | goto retry; |
1924 | } | |
1925 | ||
1926 | /* | |
fe4b04fa PZ |
1927 | * Since the task isn't running, its safe to add the event, us holding |
1928 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1929 | */ |
fe4b04fa | 1930 | add_event_to_ctx(event, ctx); |
e625cce1 | 1931 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1932 | } |
1933 | ||
fa289bec | 1934 | /* |
cdd6c482 | 1935 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1936 | * Enabling the leader of a group effectively enables all |
1937 | * the group members that aren't explicitly disabled, so we | |
1938 | * have to update their ->tstamp_enabled also. | |
1939 | * Note: this works for group members as well as group leaders | |
1940 | * since the non-leader members' sibling_lists will be empty. | |
1941 | */ | |
1d9b482e | 1942 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1943 | { |
cdd6c482 | 1944 | struct perf_event *sub; |
4158755d | 1945 | u64 tstamp = perf_event_time(event); |
fa289bec | 1946 | |
cdd6c482 | 1947 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1948 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1949 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1950 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1951 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1952 | } |
fa289bec PM |
1953 | } |
1954 | ||
d859e29f | 1955 | /* |
cdd6c482 | 1956 | * Cross CPU call to enable a performance event |
d859e29f | 1957 | */ |
fe4b04fa | 1958 | static int __perf_event_enable(void *info) |
04289bb9 | 1959 | { |
cdd6c482 | 1960 | struct perf_event *event = info; |
cdd6c482 IM |
1961 | struct perf_event_context *ctx = event->ctx; |
1962 | struct perf_event *leader = event->group_leader; | |
108b02cf | 1963 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 1964 | int err; |
04289bb9 | 1965 | |
06f41796 JO |
1966 | /* |
1967 | * There's a time window between 'ctx->is_active' check | |
1968 | * in perf_event_enable function and this place having: | |
1969 | * - IRQs on | |
1970 | * - ctx->lock unlocked | |
1971 | * | |
1972 | * where the task could be killed and 'ctx' deactivated | |
1973 | * by perf_event_exit_task. | |
1974 | */ | |
1975 | if (!ctx->is_active) | |
fe4b04fa | 1976 | return -EINVAL; |
3cbed429 | 1977 | |
e625cce1 | 1978 | raw_spin_lock(&ctx->lock); |
4af4998b | 1979 | update_context_time(ctx); |
d859e29f | 1980 | |
cdd6c482 | 1981 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 1982 | goto unlock; |
e5d1367f SE |
1983 | |
1984 | /* | |
1985 | * set current task's cgroup time reference point | |
1986 | */ | |
3f7cce3c | 1987 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 1988 | |
1d9b482e | 1989 | __perf_event_mark_enabled(event); |
04289bb9 | 1990 | |
e5d1367f SE |
1991 | if (!event_filter_match(event)) { |
1992 | if (is_cgroup_event(event)) | |
1993 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 1994 | goto unlock; |
e5d1367f | 1995 | } |
f4c4176f | 1996 | |
04289bb9 | 1997 | /* |
cdd6c482 | 1998 | * If the event is in a group and isn't the group leader, |
d859e29f | 1999 | * then don't put it on unless the group is on. |
04289bb9 | 2000 | */ |
cdd6c482 | 2001 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2002 | goto unlock; |
3b6f9e5c | 2003 | |
cdd6c482 | 2004 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2005 | err = -EEXIST; |
e758a33d | 2006 | } else { |
cdd6c482 | 2007 | if (event == leader) |
6e37738a | 2008 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2009 | else |
6e37738a | 2010 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2011 | } |
d859e29f PM |
2012 | |
2013 | if (err) { | |
2014 | /* | |
cdd6c482 | 2015 | * If this event can't go on and it's part of a |
d859e29f PM |
2016 | * group, then the whole group has to come off. |
2017 | */ | |
9e630205 | 2018 | if (leader != event) { |
d859e29f | 2019 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
2020 | perf_cpu_hrtimer_restart(cpuctx); |
2021 | } | |
0d48696f | 2022 | if (leader->attr.pinned) { |
53cfbf59 | 2023 | update_group_times(leader); |
cdd6c482 | 2024 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2025 | } |
d859e29f PM |
2026 | } |
2027 | ||
9ed6060d | 2028 | unlock: |
e625cce1 | 2029 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2030 | |
2031 | return 0; | |
d859e29f PM |
2032 | } |
2033 | ||
2034 | /* | |
cdd6c482 | 2035 | * Enable a event. |
c93f7669 | 2036 | * |
cdd6c482 IM |
2037 | * If event->ctx is a cloned context, callers must make sure that |
2038 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2039 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2040 | * perf_event_for_each_child or perf_event_for_each as described |
2041 | * for perf_event_disable. | |
d859e29f | 2042 | */ |
44234adc | 2043 | void perf_event_enable(struct perf_event *event) |
d859e29f | 2044 | { |
cdd6c482 | 2045 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2046 | struct task_struct *task = ctx->task; |
2047 | ||
2048 | if (!task) { | |
2049 | /* | |
cdd6c482 | 2050 | * Enable the event on the cpu that it's on |
d859e29f | 2051 | */ |
fe4b04fa | 2052 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2053 | return; |
2054 | } | |
2055 | ||
e625cce1 | 2056 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2057 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2058 | goto out; |
2059 | ||
2060 | /* | |
cdd6c482 IM |
2061 | * If the event is in error state, clear that first. |
2062 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2063 | * know that it has gone back into error state, as distinct |
2064 | * from the task having been scheduled away before the | |
2065 | * cross-call arrived. | |
2066 | */ | |
cdd6c482 IM |
2067 | if (event->state == PERF_EVENT_STATE_ERROR) |
2068 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2069 | |
9ed6060d | 2070 | retry: |
fe4b04fa | 2071 | if (!ctx->is_active) { |
1d9b482e | 2072 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2073 | goto out; |
2074 | } | |
2075 | ||
e625cce1 | 2076 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2077 | |
2078 | if (!task_function_call(task, __perf_event_enable, event)) | |
2079 | return; | |
d859e29f | 2080 | |
e625cce1 | 2081 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2082 | |
2083 | /* | |
cdd6c482 | 2084 | * If the context is active and the event is still off, |
d859e29f PM |
2085 | * we need to retry the cross-call. |
2086 | */ | |
fe4b04fa PZ |
2087 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2088 | /* | |
2089 | * task could have been flipped by a concurrent | |
2090 | * perf_event_context_sched_out() | |
2091 | */ | |
2092 | task = ctx->task; | |
d859e29f | 2093 | goto retry; |
fe4b04fa | 2094 | } |
fa289bec | 2095 | |
9ed6060d | 2096 | out: |
e625cce1 | 2097 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2098 | } |
dcfce4a0 | 2099 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2100 | |
26ca5c11 | 2101 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2102 | { |
2023b359 | 2103 | /* |
cdd6c482 | 2104 | * not supported on inherited events |
2023b359 | 2105 | */ |
2e939d1d | 2106 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2107 | return -EINVAL; |
2108 | ||
cdd6c482 IM |
2109 | atomic_add(refresh, &event->event_limit); |
2110 | perf_event_enable(event); | |
2023b359 PZ |
2111 | |
2112 | return 0; | |
79f14641 | 2113 | } |
26ca5c11 | 2114 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2115 | |
5b0311e1 FW |
2116 | static void ctx_sched_out(struct perf_event_context *ctx, |
2117 | struct perf_cpu_context *cpuctx, | |
2118 | enum event_type_t event_type) | |
235c7fc7 | 2119 | { |
cdd6c482 | 2120 | struct perf_event *event; |
db24d33e | 2121 | int is_active = ctx->is_active; |
235c7fc7 | 2122 | |
db24d33e | 2123 | ctx->is_active &= ~event_type; |
cdd6c482 | 2124 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2125 | return; |
2126 | ||
4af4998b | 2127 | update_context_time(ctx); |
e5d1367f | 2128 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2129 | if (!ctx->nr_active) |
facc4307 | 2130 | return; |
5b0311e1 | 2131 | |
075e0b00 | 2132 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2133 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2134 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2135 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2136 | } |
889ff015 | 2137 | |
db24d33e | 2138 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2139 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2140 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2141 | } |
1b9a644f | 2142 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2143 | } |
2144 | ||
564c2b21 PM |
2145 | /* |
2146 | * Test whether two contexts are equivalent, i.e. whether they | |
2147 | * have both been cloned from the same version of the same context | |
cdd6c482 IM |
2148 | * and they both have the same number of enabled events. |
2149 | * If the number of enabled events is the same, then the set | |
2150 | * of enabled events should be the same, because these are both | |
2151 | * inherited contexts, therefore we can't access individual events | |
564c2b21 | 2152 | * in them directly with an fd; we can only enable/disable all |
cdd6c482 | 2153 | * events via prctl, or enable/disable all events in a family |
564c2b21 PM |
2154 | * via ioctl, which will have the same effect on both contexts. |
2155 | */ | |
cdd6c482 IM |
2156 | static int context_equiv(struct perf_event_context *ctx1, |
2157 | struct perf_event_context *ctx2) | |
564c2b21 PM |
2158 | { |
2159 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx | |
ad3a37de | 2160 | && ctx1->parent_gen == ctx2->parent_gen |
25346b93 | 2161 | && !ctx1->pin_count && !ctx2->pin_count; |
564c2b21 PM |
2162 | } |
2163 | ||
cdd6c482 IM |
2164 | static void __perf_event_sync_stat(struct perf_event *event, |
2165 | struct perf_event *next_event) | |
bfbd3381 PZ |
2166 | { |
2167 | u64 value; | |
2168 | ||
cdd6c482 | 2169 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2170 | return; |
2171 | ||
2172 | /* | |
cdd6c482 | 2173 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2174 | * because we're in the middle of a context switch and have IRQs |
2175 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2176 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2177 | * don't need to use it. |
2178 | */ | |
cdd6c482 IM |
2179 | switch (event->state) { |
2180 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2181 | event->pmu->read(event); |
2182 | /* fall-through */ | |
bfbd3381 | 2183 | |
cdd6c482 IM |
2184 | case PERF_EVENT_STATE_INACTIVE: |
2185 | update_event_times(event); | |
bfbd3381 PZ |
2186 | break; |
2187 | ||
2188 | default: | |
2189 | break; | |
2190 | } | |
2191 | ||
2192 | /* | |
cdd6c482 | 2193 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2194 | * values when we flip the contexts. |
2195 | */ | |
e7850595 PZ |
2196 | value = local64_read(&next_event->count); |
2197 | value = local64_xchg(&event->count, value); | |
2198 | local64_set(&next_event->count, value); | |
bfbd3381 | 2199 | |
cdd6c482 IM |
2200 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2201 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2202 | |
bfbd3381 | 2203 | /* |
19d2e755 | 2204 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2205 | */ |
cdd6c482 IM |
2206 | perf_event_update_userpage(event); |
2207 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2208 | } |
2209 | ||
2210 | #define list_next_entry(pos, member) \ | |
2211 | list_entry(pos->member.next, typeof(*pos), member) | |
2212 | ||
cdd6c482 IM |
2213 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2214 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2215 | { |
cdd6c482 | 2216 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2217 | |
2218 | if (!ctx->nr_stat) | |
2219 | return; | |
2220 | ||
02ffdbc8 PZ |
2221 | update_context_time(ctx); |
2222 | ||
cdd6c482 IM |
2223 | event = list_first_entry(&ctx->event_list, |
2224 | struct perf_event, event_entry); | |
bfbd3381 | 2225 | |
cdd6c482 IM |
2226 | next_event = list_first_entry(&next_ctx->event_list, |
2227 | struct perf_event, event_entry); | |
bfbd3381 | 2228 | |
cdd6c482 IM |
2229 | while (&event->event_entry != &ctx->event_list && |
2230 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2231 | |
cdd6c482 | 2232 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2233 | |
cdd6c482 IM |
2234 | event = list_next_entry(event, event_entry); |
2235 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2236 | } |
2237 | } | |
2238 | ||
fe4b04fa PZ |
2239 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2240 | struct task_struct *next) | |
0793a61d | 2241 | { |
8dc85d54 | 2242 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 IM |
2243 | struct perf_event_context *next_ctx; |
2244 | struct perf_event_context *parent; | |
108b02cf | 2245 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2246 | int do_switch = 1; |
0793a61d | 2247 | |
108b02cf PZ |
2248 | if (likely(!ctx)) |
2249 | return; | |
10989fb2 | 2250 | |
108b02cf PZ |
2251 | cpuctx = __get_cpu_context(ctx); |
2252 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2253 | return; |
2254 | ||
c93f7669 PM |
2255 | rcu_read_lock(); |
2256 | parent = rcu_dereference(ctx->parent_ctx); | |
8dc85d54 | 2257 | next_ctx = next->perf_event_ctxp[ctxn]; |
c93f7669 PM |
2258 | if (parent && next_ctx && |
2259 | rcu_dereference(next_ctx->parent_ctx) == parent) { | |
2260 | /* | |
2261 | * Looks like the two contexts are clones, so we might be | |
2262 | * able to optimize the context switch. We lock both | |
2263 | * contexts and check that they are clones under the | |
2264 | * lock (including re-checking that neither has been | |
2265 | * uncloned in the meantime). It doesn't matter which | |
2266 | * order we take the locks because no other cpu could | |
2267 | * be trying to lock both of these tasks. | |
2268 | */ | |
e625cce1 TG |
2269 | raw_spin_lock(&ctx->lock); |
2270 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2271 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2272 | /* |
2273 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2274 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2275 | */ |
8dc85d54 PZ |
2276 | task->perf_event_ctxp[ctxn] = next_ctx; |
2277 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2278 | ctx->task = next; |
2279 | next_ctx->task = task; | |
2280 | do_switch = 0; | |
bfbd3381 | 2281 | |
cdd6c482 | 2282 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2283 | } |
e625cce1 TG |
2284 | raw_spin_unlock(&next_ctx->lock); |
2285 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2286 | } |
c93f7669 | 2287 | rcu_read_unlock(); |
564c2b21 | 2288 | |
c93f7669 | 2289 | if (do_switch) { |
facc4307 | 2290 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2291 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2292 | cpuctx->task_ctx = NULL; |
facc4307 | 2293 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2294 | } |
0793a61d TG |
2295 | } |
2296 | ||
8dc85d54 PZ |
2297 | #define for_each_task_context_nr(ctxn) \ |
2298 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2299 | ||
2300 | /* | |
2301 | * Called from scheduler to remove the events of the current task, | |
2302 | * with interrupts disabled. | |
2303 | * | |
2304 | * We stop each event and update the event value in event->count. | |
2305 | * | |
2306 | * This does not protect us against NMI, but disable() | |
2307 | * sets the disabled bit in the control field of event _before_ | |
2308 | * accessing the event control register. If a NMI hits, then it will | |
2309 | * not restart the event. | |
2310 | */ | |
ab0cce56 JO |
2311 | void __perf_event_task_sched_out(struct task_struct *task, |
2312 | struct task_struct *next) | |
8dc85d54 PZ |
2313 | { |
2314 | int ctxn; | |
2315 | ||
8dc85d54 PZ |
2316 | for_each_task_context_nr(ctxn) |
2317 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2318 | |
2319 | /* | |
2320 | * if cgroup events exist on this CPU, then we need | |
2321 | * to check if we have to switch out PMU state. | |
2322 | * cgroup event are system-wide mode only | |
2323 | */ | |
2324 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2325 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2326 | } |
2327 | ||
04dc2dbb | 2328 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2329 | { |
108b02cf | 2330 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2331 | |
a63eaf34 PM |
2332 | if (!cpuctx->task_ctx) |
2333 | return; | |
012b84da IM |
2334 | |
2335 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2336 | return; | |
2337 | ||
04dc2dbb | 2338 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2339 | cpuctx->task_ctx = NULL; |
2340 | } | |
2341 | ||
5b0311e1 FW |
2342 | /* |
2343 | * Called with IRQs disabled | |
2344 | */ | |
2345 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2346 | enum event_type_t event_type) | |
2347 | { | |
2348 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2349 | } |
2350 | ||
235c7fc7 | 2351 | static void |
5b0311e1 | 2352 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2353 | struct perf_cpu_context *cpuctx) |
0793a61d | 2354 | { |
cdd6c482 | 2355 | struct perf_event *event; |
0793a61d | 2356 | |
889ff015 FW |
2357 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2358 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2359 | continue; |
5632ab12 | 2360 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2361 | continue; |
2362 | ||
e5d1367f SE |
2363 | /* may need to reset tstamp_enabled */ |
2364 | if (is_cgroup_event(event)) | |
2365 | perf_cgroup_mark_enabled(event, ctx); | |
2366 | ||
8c9ed8e1 | 2367 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2368 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2369 | |
2370 | /* | |
2371 | * If this pinned group hasn't been scheduled, | |
2372 | * put it in error state. | |
2373 | */ | |
cdd6c482 IM |
2374 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2375 | update_group_times(event); | |
2376 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2377 | } |
3b6f9e5c | 2378 | } |
5b0311e1 FW |
2379 | } |
2380 | ||
2381 | static void | |
2382 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2383 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2384 | { |
2385 | struct perf_event *event; | |
2386 | int can_add_hw = 1; | |
3b6f9e5c | 2387 | |
889ff015 FW |
2388 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2389 | /* Ignore events in OFF or ERROR state */ | |
2390 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2391 | continue; |
04289bb9 IM |
2392 | /* |
2393 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2394 | * of events: |
04289bb9 | 2395 | */ |
5632ab12 | 2396 | if (!event_filter_match(event)) |
0793a61d TG |
2397 | continue; |
2398 | ||
e5d1367f SE |
2399 | /* may need to reset tstamp_enabled */ |
2400 | if (is_cgroup_event(event)) | |
2401 | perf_cgroup_mark_enabled(event, ctx); | |
2402 | ||
9ed6060d | 2403 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2404 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2405 | can_add_hw = 0; |
9ed6060d | 2406 | } |
0793a61d | 2407 | } |
5b0311e1 FW |
2408 | } |
2409 | ||
2410 | static void | |
2411 | ctx_sched_in(struct perf_event_context *ctx, | |
2412 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2413 | enum event_type_t event_type, |
2414 | struct task_struct *task) | |
5b0311e1 | 2415 | { |
e5d1367f | 2416 | u64 now; |
db24d33e | 2417 | int is_active = ctx->is_active; |
e5d1367f | 2418 | |
db24d33e | 2419 | ctx->is_active |= event_type; |
5b0311e1 | 2420 | if (likely(!ctx->nr_events)) |
facc4307 | 2421 | return; |
5b0311e1 | 2422 | |
e5d1367f SE |
2423 | now = perf_clock(); |
2424 | ctx->timestamp = now; | |
3f7cce3c | 2425 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2426 | /* |
2427 | * First go through the list and put on any pinned groups | |
2428 | * in order to give them the best chance of going on. | |
2429 | */ | |
db24d33e | 2430 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2431 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2432 | |
2433 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2434 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2435 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2436 | } |
2437 | ||
329c0e01 | 2438 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2439 | enum event_type_t event_type, |
2440 | struct task_struct *task) | |
329c0e01 FW |
2441 | { |
2442 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2443 | ||
e5d1367f | 2444 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2445 | } |
2446 | ||
e5d1367f SE |
2447 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2448 | struct task_struct *task) | |
235c7fc7 | 2449 | { |
108b02cf | 2450 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2451 | |
108b02cf | 2452 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2453 | if (cpuctx->task_ctx == ctx) |
2454 | return; | |
2455 | ||
facc4307 | 2456 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2457 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2458 | /* |
2459 | * We want to keep the following priority order: | |
2460 | * cpu pinned (that don't need to move), task pinned, | |
2461 | * cpu flexible, task flexible. | |
2462 | */ | |
2463 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2464 | ||
1d5f003f GN |
2465 | if (ctx->nr_events) |
2466 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2467 | |
86b47c25 GN |
2468 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2469 | ||
facc4307 PZ |
2470 | perf_pmu_enable(ctx->pmu); |
2471 | perf_ctx_unlock(cpuctx, ctx); | |
2472 | ||
b5ab4cd5 PZ |
2473 | /* |
2474 | * Since these rotations are per-cpu, we need to ensure the | |
2475 | * cpu-context we got scheduled on is actually rotating. | |
2476 | */ | |
108b02cf | 2477 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2478 | } |
2479 | ||
d010b332 SE |
2480 | /* |
2481 | * When sampling the branck stack in system-wide, it may be necessary | |
2482 | * to flush the stack on context switch. This happens when the branch | |
2483 | * stack does not tag its entries with the pid of the current task. | |
2484 | * Otherwise it becomes impossible to associate a branch entry with a | |
2485 | * task. This ambiguity is more likely to appear when the branch stack | |
2486 | * supports priv level filtering and the user sets it to monitor only | |
2487 | * at the user level (which could be a useful measurement in system-wide | |
2488 | * mode). In that case, the risk is high of having a branch stack with | |
2489 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2490 | * entries or stashing them somewhere in the PMU specific code layer. | |
2491 | * | |
2492 | * This function provides the context switch callback to the lower code | |
2493 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2494 | * with at least one active event using taken branch sampling. | |
2495 | */ | |
2496 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2497 | struct task_struct *task) | |
2498 | { | |
2499 | struct perf_cpu_context *cpuctx; | |
2500 | struct pmu *pmu; | |
2501 | unsigned long flags; | |
2502 | ||
2503 | /* no need to flush branch stack if not changing task */ | |
2504 | if (prev == task) | |
2505 | return; | |
2506 | ||
2507 | local_irq_save(flags); | |
2508 | ||
2509 | rcu_read_lock(); | |
2510 | ||
2511 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2512 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2513 | ||
2514 | /* | |
2515 | * check if the context has at least one | |
2516 | * event using PERF_SAMPLE_BRANCH_STACK | |
2517 | */ | |
2518 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2519 | && pmu->flush_branch_stack) { | |
2520 | ||
2521 | pmu = cpuctx->ctx.pmu; | |
2522 | ||
2523 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2524 | ||
2525 | perf_pmu_disable(pmu); | |
2526 | ||
2527 | pmu->flush_branch_stack(); | |
2528 | ||
2529 | perf_pmu_enable(pmu); | |
2530 | ||
2531 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2532 | } | |
2533 | } | |
2534 | ||
2535 | rcu_read_unlock(); | |
2536 | ||
2537 | local_irq_restore(flags); | |
2538 | } | |
2539 | ||
8dc85d54 PZ |
2540 | /* |
2541 | * Called from scheduler to add the events of the current task | |
2542 | * with interrupts disabled. | |
2543 | * | |
2544 | * We restore the event value and then enable it. | |
2545 | * | |
2546 | * This does not protect us against NMI, but enable() | |
2547 | * sets the enabled bit in the control field of event _before_ | |
2548 | * accessing the event control register. If a NMI hits, then it will | |
2549 | * keep the event running. | |
2550 | */ | |
ab0cce56 JO |
2551 | void __perf_event_task_sched_in(struct task_struct *prev, |
2552 | struct task_struct *task) | |
8dc85d54 PZ |
2553 | { |
2554 | struct perf_event_context *ctx; | |
2555 | int ctxn; | |
2556 | ||
2557 | for_each_task_context_nr(ctxn) { | |
2558 | ctx = task->perf_event_ctxp[ctxn]; | |
2559 | if (likely(!ctx)) | |
2560 | continue; | |
2561 | ||
e5d1367f | 2562 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2563 | } |
e5d1367f SE |
2564 | /* |
2565 | * if cgroup events exist on this CPU, then we need | |
2566 | * to check if we have to switch in PMU state. | |
2567 | * cgroup event are system-wide mode only | |
2568 | */ | |
2569 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2570 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2571 | |
2572 | /* check for system-wide branch_stack events */ | |
2573 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2574 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2575 | } |
2576 | ||
abd50713 PZ |
2577 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2578 | { | |
2579 | u64 frequency = event->attr.sample_freq; | |
2580 | u64 sec = NSEC_PER_SEC; | |
2581 | u64 divisor, dividend; | |
2582 | ||
2583 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2584 | ||
2585 | count_fls = fls64(count); | |
2586 | nsec_fls = fls64(nsec); | |
2587 | frequency_fls = fls64(frequency); | |
2588 | sec_fls = 30; | |
2589 | ||
2590 | /* | |
2591 | * We got @count in @nsec, with a target of sample_freq HZ | |
2592 | * the target period becomes: | |
2593 | * | |
2594 | * @count * 10^9 | |
2595 | * period = ------------------- | |
2596 | * @nsec * sample_freq | |
2597 | * | |
2598 | */ | |
2599 | ||
2600 | /* | |
2601 | * Reduce accuracy by one bit such that @a and @b converge | |
2602 | * to a similar magnitude. | |
2603 | */ | |
fe4b04fa | 2604 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2605 | do { \ |
2606 | if (a##_fls > b##_fls) { \ | |
2607 | a >>= 1; \ | |
2608 | a##_fls--; \ | |
2609 | } else { \ | |
2610 | b >>= 1; \ | |
2611 | b##_fls--; \ | |
2612 | } \ | |
2613 | } while (0) | |
2614 | ||
2615 | /* | |
2616 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2617 | * the other, so that finally we can do a u64/u64 division. | |
2618 | */ | |
2619 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2620 | REDUCE_FLS(nsec, frequency); | |
2621 | REDUCE_FLS(sec, count); | |
2622 | } | |
2623 | ||
2624 | if (count_fls + sec_fls > 64) { | |
2625 | divisor = nsec * frequency; | |
2626 | ||
2627 | while (count_fls + sec_fls > 64) { | |
2628 | REDUCE_FLS(count, sec); | |
2629 | divisor >>= 1; | |
2630 | } | |
2631 | ||
2632 | dividend = count * sec; | |
2633 | } else { | |
2634 | dividend = count * sec; | |
2635 | ||
2636 | while (nsec_fls + frequency_fls > 64) { | |
2637 | REDUCE_FLS(nsec, frequency); | |
2638 | dividend >>= 1; | |
2639 | } | |
2640 | ||
2641 | divisor = nsec * frequency; | |
2642 | } | |
2643 | ||
f6ab91ad PZ |
2644 | if (!divisor) |
2645 | return dividend; | |
2646 | ||
abd50713 PZ |
2647 | return div64_u64(dividend, divisor); |
2648 | } | |
2649 | ||
e050e3f0 SE |
2650 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2651 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2652 | ||
f39d47ff | 2653 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2654 | { |
cdd6c482 | 2655 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2656 | s64 period, sample_period; |
bd2b5b12 PZ |
2657 | s64 delta; |
2658 | ||
abd50713 | 2659 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2660 | |
2661 | delta = (s64)(period - hwc->sample_period); | |
2662 | delta = (delta + 7) / 8; /* low pass filter */ | |
2663 | ||
2664 | sample_period = hwc->sample_period + delta; | |
2665 | ||
2666 | if (!sample_period) | |
2667 | sample_period = 1; | |
2668 | ||
bd2b5b12 | 2669 | hwc->sample_period = sample_period; |
abd50713 | 2670 | |
e7850595 | 2671 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2672 | if (disable) |
2673 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2674 | ||
e7850595 | 2675 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2676 | |
2677 | if (disable) | |
2678 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2679 | } |
bd2b5b12 PZ |
2680 | } |
2681 | ||
e050e3f0 SE |
2682 | /* |
2683 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2684 | * events. At the same time, make sure, having freq events does not change | |
2685 | * the rate of unthrottling as that would introduce bias. | |
2686 | */ | |
2687 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2688 | int needs_unthr) | |
60db5e09 | 2689 | { |
cdd6c482 IM |
2690 | struct perf_event *event; |
2691 | struct hw_perf_event *hwc; | |
e050e3f0 | 2692 | u64 now, period = TICK_NSEC; |
abd50713 | 2693 | s64 delta; |
60db5e09 | 2694 | |
e050e3f0 SE |
2695 | /* |
2696 | * only need to iterate over all events iff: | |
2697 | * - context have events in frequency mode (needs freq adjust) | |
2698 | * - there are events to unthrottle on this cpu | |
2699 | */ | |
2700 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2701 | return; |
2702 | ||
e050e3f0 | 2703 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2704 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2705 | |
03541f8b | 2706 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2707 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2708 | continue; |
2709 | ||
5632ab12 | 2710 | if (!event_filter_match(event)) |
5d27c23d PZ |
2711 | continue; |
2712 | ||
cdd6c482 | 2713 | hwc = &event->hw; |
6a24ed6c | 2714 | |
e050e3f0 SE |
2715 | if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) { |
2716 | hwc->interrupts = 0; | |
cdd6c482 | 2717 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2718 | event->pmu->start(event, 0); |
a78ac325 PZ |
2719 | } |
2720 | ||
cdd6c482 | 2721 | if (!event->attr.freq || !event->attr.sample_freq) |
60db5e09 PZ |
2722 | continue; |
2723 | ||
e050e3f0 SE |
2724 | /* |
2725 | * stop the event and update event->count | |
2726 | */ | |
2727 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2728 | ||
e7850595 | 2729 | now = local64_read(&event->count); |
abd50713 PZ |
2730 | delta = now - hwc->freq_count_stamp; |
2731 | hwc->freq_count_stamp = now; | |
60db5e09 | 2732 | |
e050e3f0 SE |
2733 | /* |
2734 | * restart the event | |
2735 | * reload only if value has changed | |
f39d47ff SE |
2736 | * we have stopped the event so tell that |
2737 | * to perf_adjust_period() to avoid stopping it | |
2738 | * twice. | |
e050e3f0 | 2739 | */ |
abd50713 | 2740 | if (delta > 0) |
f39d47ff | 2741 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2742 | |
2743 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
60db5e09 | 2744 | } |
e050e3f0 | 2745 | |
f39d47ff | 2746 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2747 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2748 | } |
2749 | ||
235c7fc7 | 2750 | /* |
cdd6c482 | 2751 | * Round-robin a context's events: |
235c7fc7 | 2752 | */ |
cdd6c482 | 2753 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2754 | { |
dddd3379 TG |
2755 | /* |
2756 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2757 | * disabled by the inheritance code. | |
2758 | */ | |
2759 | if (!ctx->rotate_disable) | |
2760 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2761 | } |
2762 | ||
b5ab4cd5 | 2763 | /* |
e9d2b064 PZ |
2764 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2765 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2766 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2767 | */ |
9e630205 | 2768 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2769 | { |
8dc85d54 | 2770 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2771 | int rotate = 0, remove = 1; |
7fc23a53 | 2772 | |
b5ab4cd5 | 2773 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2774 | remove = 0; |
b5ab4cd5 PZ |
2775 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2776 | rotate = 1; | |
2777 | } | |
235c7fc7 | 2778 | |
8dc85d54 | 2779 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2780 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2781 | remove = 0; |
b5ab4cd5 PZ |
2782 | if (ctx->nr_events != ctx->nr_active) |
2783 | rotate = 1; | |
2784 | } | |
9717e6cd | 2785 | |
e050e3f0 | 2786 | if (!rotate) |
0f5a2601 PZ |
2787 | goto done; |
2788 | ||
facc4307 | 2789 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2790 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2791 | |
e050e3f0 SE |
2792 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2793 | if (ctx) | |
2794 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2795 | |
e050e3f0 SE |
2796 | rotate_ctx(&cpuctx->ctx); |
2797 | if (ctx) | |
2798 | rotate_ctx(ctx); | |
235c7fc7 | 2799 | |
e050e3f0 | 2800 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2801 | |
0f5a2601 PZ |
2802 | perf_pmu_enable(cpuctx->ctx.pmu); |
2803 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2804 | done: |
e9d2b064 PZ |
2805 | if (remove) |
2806 | list_del_init(&cpuctx->rotation_list); | |
9e630205 SE |
2807 | |
2808 | return rotate; | |
e9d2b064 PZ |
2809 | } |
2810 | ||
026249ef FW |
2811 | #ifdef CONFIG_NO_HZ_FULL |
2812 | bool perf_event_can_stop_tick(void) | |
2813 | { | |
d84153d6 FW |
2814 | if (atomic_read(&__get_cpu_var(perf_freq_events)) || |
2815 | __this_cpu_read(perf_throttled_count)) | |
026249ef | 2816 | return false; |
d84153d6 FW |
2817 | else |
2818 | return true; | |
026249ef FW |
2819 | } |
2820 | #endif | |
2821 | ||
e9d2b064 PZ |
2822 | void perf_event_task_tick(void) |
2823 | { | |
2824 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2825 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2826 | struct perf_event_context *ctx; |
2827 | int throttled; | |
b5ab4cd5 | 2828 | |
e9d2b064 PZ |
2829 | WARN_ON(!irqs_disabled()); |
2830 | ||
e050e3f0 SE |
2831 | __this_cpu_inc(perf_throttled_seq); |
2832 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2833 | ||
e9d2b064 | 2834 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2835 | ctx = &cpuctx->ctx; |
2836 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2837 | ||
2838 | ctx = cpuctx->task_ctx; | |
2839 | if (ctx) | |
2840 | perf_adjust_freq_unthr_context(ctx, throttled); | |
e9d2b064 | 2841 | } |
0793a61d TG |
2842 | } |
2843 | ||
889ff015 FW |
2844 | static int event_enable_on_exec(struct perf_event *event, |
2845 | struct perf_event_context *ctx) | |
2846 | { | |
2847 | if (!event->attr.enable_on_exec) | |
2848 | return 0; | |
2849 | ||
2850 | event->attr.enable_on_exec = 0; | |
2851 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2852 | return 0; | |
2853 | ||
1d9b482e | 2854 | __perf_event_mark_enabled(event); |
889ff015 FW |
2855 | |
2856 | return 1; | |
2857 | } | |
2858 | ||
57e7986e | 2859 | /* |
cdd6c482 | 2860 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2861 | * This expects task == current. |
2862 | */ | |
8dc85d54 | 2863 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2864 | { |
cdd6c482 | 2865 | struct perf_event *event; |
57e7986e PM |
2866 | unsigned long flags; |
2867 | int enabled = 0; | |
889ff015 | 2868 | int ret; |
57e7986e PM |
2869 | |
2870 | local_irq_save(flags); | |
cdd6c482 | 2871 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2872 | goto out; |
2873 | ||
e566b76e SE |
2874 | /* |
2875 | * We must ctxsw out cgroup events to avoid conflict | |
2876 | * when invoking perf_task_event_sched_in() later on | |
2877 | * in this function. Otherwise we end up trying to | |
2878 | * ctxswin cgroup events which are already scheduled | |
2879 | * in. | |
2880 | */ | |
a8d757ef | 2881 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2882 | |
e625cce1 | 2883 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2884 | task_ctx_sched_out(ctx); |
57e7986e | 2885 | |
b79387ef | 2886 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2887 | ret = event_enable_on_exec(event, ctx); |
2888 | if (ret) | |
2889 | enabled = 1; | |
57e7986e PM |
2890 | } |
2891 | ||
2892 | /* | |
cdd6c482 | 2893 | * Unclone this context if we enabled any event. |
57e7986e | 2894 | */ |
71a851b4 PZ |
2895 | if (enabled) |
2896 | unclone_ctx(ctx); | |
57e7986e | 2897 | |
e625cce1 | 2898 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2899 | |
e566b76e SE |
2900 | /* |
2901 | * Also calls ctxswin for cgroup events, if any: | |
2902 | */ | |
e5d1367f | 2903 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2904 | out: |
57e7986e PM |
2905 | local_irq_restore(flags); |
2906 | } | |
2907 | ||
0793a61d | 2908 | /* |
cdd6c482 | 2909 | * Cross CPU call to read the hardware event |
0793a61d | 2910 | */ |
cdd6c482 | 2911 | static void __perf_event_read(void *info) |
0793a61d | 2912 | { |
cdd6c482 IM |
2913 | struct perf_event *event = info; |
2914 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2915 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2916 | |
e1ac3614 PM |
2917 | /* |
2918 | * If this is a task context, we need to check whether it is | |
2919 | * the current task context of this cpu. If not it has been | |
2920 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2921 | * event->count would have been updated to a recent sample |
2922 | * when the event was scheduled out. | |
e1ac3614 PM |
2923 | */ |
2924 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2925 | return; | |
2926 | ||
e625cce1 | 2927 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2928 | if (ctx->is_active) { |
542e72fc | 2929 | update_context_time(ctx); |
e5d1367f SE |
2930 | update_cgrp_time_from_event(event); |
2931 | } | |
cdd6c482 | 2932 | update_event_times(event); |
542e72fc PZ |
2933 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2934 | event->pmu->read(event); | |
e625cce1 | 2935 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2936 | } |
2937 | ||
b5e58793 PZ |
2938 | static inline u64 perf_event_count(struct perf_event *event) |
2939 | { | |
e7850595 | 2940 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
2941 | } |
2942 | ||
cdd6c482 | 2943 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
2944 | { |
2945 | /* | |
cdd6c482 IM |
2946 | * If event is enabled and currently active on a CPU, update the |
2947 | * value in the event structure: | |
0793a61d | 2948 | */ |
cdd6c482 IM |
2949 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2950 | smp_call_function_single(event->oncpu, | |
2951 | __perf_event_read, event, 1); | |
2952 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
2953 | struct perf_event_context *ctx = event->ctx; |
2954 | unsigned long flags; | |
2955 | ||
e625cce1 | 2956 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
2957 | /* |
2958 | * may read while context is not active | |
2959 | * (e.g., thread is blocked), in that case | |
2960 | * we cannot update context time | |
2961 | */ | |
e5d1367f | 2962 | if (ctx->is_active) { |
c530ccd9 | 2963 | update_context_time(ctx); |
e5d1367f SE |
2964 | update_cgrp_time_from_event(event); |
2965 | } | |
cdd6c482 | 2966 | update_event_times(event); |
e625cce1 | 2967 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
2968 | } |
2969 | ||
b5e58793 | 2970 | return perf_event_count(event); |
0793a61d TG |
2971 | } |
2972 | ||
a63eaf34 | 2973 | /* |
cdd6c482 | 2974 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 2975 | */ |
eb184479 | 2976 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 2977 | { |
e625cce1 | 2978 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 2979 | mutex_init(&ctx->mutex); |
889ff015 FW |
2980 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2981 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
2982 | INIT_LIST_HEAD(&ctx->event_list); |
2983 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
2984 | } |
2985 | ||
2986 | static struct perf_event_context * | |
2987 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
2988 | { | |
2989 | struct perf_event_context *ctx; | |
2990 | ||
2991 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
2992 | if (!ctx) | |
2993 | return NULL; | |
2994 | ||
2995 | __perf_event_init_context(ctx); | |
2996 | if (task) { | |
2997 | ctx->task = task; | |
2998 | get_task_struct(task); | |
0793a61d | 2999 | } |
eb184479 PZ |
3000 | ctx->pmu = pmu; |
3001 | ||
3002 | return ctx; | |
a63eaf34 PM |
3003 | } |
3004 | ||
2ebd4ffb MH |
3005 | static struct task_struct * |
3006 | find_lively_task_by_vpid(pid_t vpid) | |
3007 | { | |
3008 | struct task_struct *task; | |
3009 | int err; | |
0793a61d TG |
3010 | |
3011 | rcu_read_lock(); | |
2ebd4ffb | 3012 | if (!vpid) |
0793a61d TG |
3013 | task = current; |
3014 | else | |
2ebd4ffb | 3015 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3016 | if (task) |
3017 | get_task_struct(task); | |
3018 | rcu_read_unlock(); | |
3019 | ||
3020 | if (!task) | |
3021 | return ERR_PTR(-ESRCH); | |
3022 | ||
0793a61d | 3023 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3024 | err = -EACCES; |
3025 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3026 | goto errout; | |
3027 | ||
2ebd4ffb MH |
3028 | return task; |
3029 | errout: | |
3030 | put_task_struct(task); | |
3031 | return ERR_PTR(err); | |
3032 | ||
3033 | } | |
3034 | ||
fe4b04fa PZ |
3035 | /* |
3036 | * Returns a matching context with refcount and pincount. | |
3037 | */ | |
108b02cf | 3038 | static struct perf_event_context * |
38a81da2 | 3039 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 3040 | { |
cdd6c482 | 3041 | struct perf_event_context *ctx; |
22a4f650 | 3042 | struct perf_cpu_context *cpuctx; |
25346b93 | 3043 | unsigned long flags; |
8dc85d54 | 3044 | int ctxn, err; |
0793a61d | 3045 | |
22a4ec72 | 3046 | if (!task) { |
cdd6c482 | 3047 | /* Must be root to operate on a CPU event: */ |
0764771d | 3048 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3049 | return ERR_PTR(-EACCES); |
3050 | ||
0793a61d | 3051 | /* |
cdd6c482 | 3052 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3053 | * offline CPU and activate it when the CPU comes up, but |
3054 | * that's for later. | |
3055 | */ | |
f6325e30 | 3056 | if (!cpu_online(cpu)) |
0793a61d TG |
3057 | return ERR_PTR(-ENODEV); |
3058 | ||
108b02cf | 3059 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3060 | ctx = &cpuctx->ctx; |
c93f7669 | 3061 | get_ctx(ctx); |
fe4b04fa | 3062 | ++ctx->pin_count; |
0793a61d | 3063 | |
0793a61d TG |
3064 | return ctx; |
3065 | } | |
3066 | ||
8dc85d54 PZ |
3067 | err = -EINVAL; |
3068 | ctxn = pmu->task_ctx_nr; | |
3069 | if (ctxn < 0) | |
3070 | goto errout; | |
3071 | ||
9ed6060d | 3072 | retry: |
8dc85d54 | 3073 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3074 | if (ctx) { |
71a851b4 | 3075 | unclone_ctx(ctx); |
fe4b04fa | 3076 | ++ctx->pin_count; |
e625cce1 | 3077 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 3078 | } else { |
eb184479 | 3079 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3080 | err = -ENOMEM; |
3081 | if (!ctx) | |
3082 | goto errout; | |
eb184479 | 3083 | |
dbe08d82 ON |
3084 | err = 0; |
3085 | mutex_lock(&task->perf_event_mutex); | |
3086 | /* | |
3087 | * If it has already passed perf_event_exit_task(). | |
3088 | * we must see PF_EXITING, it takes this mutex too. | |
3089 | */ | |
3090 | if (task->flags & PF_EXITING) | |
3091 | err = -ESRCH; | |
3092 | else if (task->perf_event_ctxp[ctxn]) | |
3093 | err = -EAGAIN; | |
fe4b04fa | 3094 | else { |
9137fb28 | 3095 | get_ctx(ctx); |
fe4b04fa | 3096 | ++ctx->pin_count; |
dbe08d82 | 3097 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3098 | } |
dbe08d82 ON |
3099 | mutex_unlock(&task->perf_event_mutex); |
3100 | ||
3101 | if (unlikely(err)) { | |
9137fb28 | 3102 | put_ctx(ctx); |
dbe08d82 ON |
3103 | |
3104 | if (err == -EAGAIN) | |
3105 | goto retry; | |
3106 | goto errout; | |
a63eaf34 PM |
3107 | } |
3108 | } | |
3109 | ||
0793a61d | 3110 | return ctx; |
c93f7669 | 3111 | |
9ed6060d | 3112 | errout: |
c93f7669 | 3113 | return ERR_PTR(err); |
0793a61d TG |
3114 | } |
3115 | ||
6fb2915d LZ |
3116 | static void perf_event_free_filter(struct perf_event *event); |
3117 | ||
cdd6c482 | 3118 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3119 | { |
cdd6c482 | 3120 | struct perf_event *event; |
592903cd | 3121 | |
cdd6c482 IM |
3122 | event = container_of(head, struct perf_event, rcu_head); |
3123 | if (event->ns) | |
3124 | put_pid_ns(event->ns); | |
6fb2915d | 3125 | perf_event_free_filter(event); |
cdd6c482 | 3126 | kfree(event); |
592903cd PZ |
3127 | } |
3128 | ||
76369139 | 3129 | static void ring_buffer_put(struct ring_buffer *rb); |
9bb5d40c | 3130 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb); |
925d519a | 3131 | |
4beb31f3 FW |
3132 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
3133 | { | |
3134 | if (event->parent) | |
3135 | return; | |
3136 | ||
3137 | if (has_branch_stack(event)) { | |
3138 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
3139 | atomic_dec(&per_cpu(perf_branch_stack_events, cpu)); | |
3140 | } | |
3141 | if (is_cgroup_event(event)) | |
3142 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
ba8a75c1 FW |
3143 | |
3144 | if (event->attr.freq) | |
3145 | atomic_dec(&per_cpu(perf_freq_events, cpu)); | |
4beb31f3 FW |
3146 | } |
3147 | ||
3148 | static void unaccount_event(struct perf_event *event) | |
3149 | { | |
3150 | if (event->parent) | |
3151 | return; | |
3152 | ||
3153 | if (event->attach_state & PERF_ATTACH_TASK) | |
3154 | static_key_slow_dec_deferred(&perf_sched_events); | |
3155 | if (event->attr.mmap || event->attr.mmap_data) | |
3156 | atomic_dec(&nr_mmap_events); | |
3157 | if (event->attr.comm) | |
3158 | atomic_dec(&nr_comm_events); | |
3159 | if (event->attr.task) | |
3160 | atomic_dec(&nr_task_events); | |
3161 | if (is_cgroup_event(event)) | |
3162 | static_key_slow_dec_deferred(&perf_sched_events); | |
3163 | if (has_branch_stack(event)) | |
3164 | static_key_slow_dec_deferred(&perf_sched_events); | |
3165 | ||
3166 | unaccount_event_cpu(event, event->cpu); | |
3167 | } | |
3168 | ||
766d6c07 FW |
3169 | static void __free_event(struct perf_event *event) |
3170 | { | |
3171 | if (!event->parent) { | |
3172 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) | |
3173 | put_callchain_buffers(); | |
3174 | } | |
3175 | ||
3176 | if (event->destroy) | |
3177 | event->destroy(event); | |
3178 | ||
3179 | if (event->ctx) | |
3180 | put_ctx(event->ctx); | |
3181 | ||
3182 | call_rcu(&event->rcu_head, free_event_rcu); | |
3183 | } | |
cdd6c482 | 3184 | static void free_event(struct perf_event *event) |
f1600952 | 3185 | { |
e360adbe | 3186 | irq_work_sync(&event->pending); |
925d519a | 3187 | |
4beb31f3 | 3188 | unaccount_event(event); |
9ee318a7 | 3189 | |
76369139 | 3190 | if (event->rb) { |
9bb5d40c PZ |
3191 | struct ring_buffer *rb; |
3192 | ||
3193 | /* | |
3194 | * Can happen when we close an event with re-directed output. | |
3195 | * | |
3196 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3197 | * over us; possibly making our ring_buffer_put() the last. | |
3198 | */ | |
3199 | mutex_lock(&event->mmap_mutex); | |
3200 | rb = event->rb; | |
3201 | if (rb) { | |
3202 | rcu_assign_pointer(event->rb, NULL); | |
3203 | ring_buffer_detach(event, rb); | |
3204 | ring_buffer_put(rb); /* could be last */ | |
3205 | } | |
3206 | mutex_unlock(&event->mmap_mutex); | |
a4be7c27 PZ |
3207 | } |
3208 | ||
e5d1367f SE |
3209 | if (is_cgroup_event(event)) |
3210 | perf_detach_cgroup(event); | |
3211 | ||
0c67b408 | 3212 | |
766d6c07 | 3213 | __free_event(event); |
f1600952 PZ |
3214 | } |
3215 | ||
a66a3052 | 3216 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 3217 | { |
cdd6c482 | 3218 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 3219 | |
ad3a37de | 3220 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
3221 | /* |
3222 | * There are two ways this annotation is useful: | |
3223 | * | |
3224 | * 1) there is a lock recursion from perf_event_exit_task | |
3225 | * see the comment there. | |
3226 | * | |
3227 | * 2) there is a lock-inversion with mmap_sem through | |
3228 | * perf_event_read_group(), which takes faults while | |
3229 | * holding ctx->mutex, however this is called after | |
3230 | * the last filedesc died, so there is no possibility | |
3231 | * to trigger the AB-BA case. | |
3232 | */ | |
3233 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
050735b0 | 3234 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 3235 | perf_group_detach(event); |
050735b0 | 3236 | raw_spin_unlock_irq(&ctx->lock); |
e03a9a55 | 3237 | perf_remove_from_context(event); |
d859e29f | 3238 | mutex_unlock(&ctx->mutex); |
0793a61d | 3239 | |
cdd6c482 | 3240 | free_event(event); |
0793a61d TG |
3241 | |
3242 | return 0; | |
3243 | } | |
a66a3052 | 3244 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 3245 | |
a66a3052 PZ |
3246 | /* |
3247 | * Called when the last reference to the file is gone. | |
3248 | */ | |
a6fa941d | 3249 | static void put_event(struct perf_event *event) |
fb0459d7 | 3250 | { |
8882135b | 3251 | struct task_struct *owner; |
fb0459d7 | 3252 | |
a6fa941d AV |
3253 | if (!atomic_long_dec_and_test(&event->refcount)) |
3254 | return; | |
fb0459d7 | 3255 | |
8882135b PZ |
3256 | rcu_read_lock(); |
3257 | owner = ACCESS_ONCE(event->owner); | |
3258 | /* | |
3259 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3260 | * !owner it means the list deletion is complete and we can indeed | |
3261 | * free this event, otherwise we need to serialize on | |
3262 | * owner->perf_event_mutex. | |
3263 | */ | |
3264 | smp_read_barrier_depends(); | |
3265 | if (owner) { | |
3266 | /* | |
3267 | * Since delayed_put_task_struct() also drops the last | |
3268 | * task reference we can safely take a new reference | |
3269 | * while holding the rcu_read_lock(). | |
3270 | */ | |
3271 | get_task_struct(owner); | |
3272 | } | |
3273 | rcu_read_unlock(); | |
3274 | ||
3275 | if (owner) { | |
3276 | mutex_lock(&owner->perf_event_mutex); | |
3277 | /* | |
3278 | * We have to re-check the event->owner field, if it is cleared | |
3279 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3280 | * ensured they're done, and we can proceed with freeing the | |
3281 | * event. | |
3282 | */ | |
3283 | if (event->owner) | |
3284 | list_del_init(&event->owner_entry); | |
3285 | mutex_unlock(&owner->perf_event_mutex); | |
3286 | put_task_struct(owner); | |
3287 | } | |
3288 | ||
a6fa941d AV |
3289 | perf_event_release_kernel(event); |
3290 | } | |
3291 | ||
3292 | static int perf_release(struct inode *inode, struct file *file) | |
3293 | { | |
3294 | put_event(file->private_data); | |
3295 | return 0; | |
fb0459d7 | 3296 | } |
fb0459d7 | 3297 | |
59ed446f | 3298 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3299 | { |
cdd6c482 | 3300 | struct perf_event *child; |
e53c0994 PZ |
3301 | u64 total = 0; |
3302 | ||
59ed446f PZ |
3303 | *enabled = 0; |
3304 | *running = 0; | |
3305 | ||
6f10581a | 3306 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3307 | total += perf_event_read(event); |
59ed446f PZ |
3308 | *enabled += event->total_time_enabled + |
3309 | atomic64_read(&event->child_total_time_enabled); | |
3310 | *running += event->total_time_running + | |
3311 | atomic64_read(&event->child_total_time_running); | |
3312 | ||
3313 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3314 | total += perf_event_read(child); |
59ed446f PZ |
3315 | *enabled += child->total_time_enabled; |
3316 | *running += child->total_time_running; | |
3317 | } | |
6f10581a | 3318 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3319 | |
3320 | return total; | |
3321 | } | |
fb0459d7 | 3322 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3323 | |
cdd6c482 | 3324 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3325 | u64 read_format, char __user *buf) |
3326 | { | |
cdd6c482 | 3327 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3328 | int n = 0, size = 0, ret = -EFAULT; |
3329 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3330 | u64 values[5]; |
59ed446f | 3331 | u64 count, enabled, running; |
abf4868b | 3332 | |
6f10581a | 3333 | mutex_lock(&ctx->mutex); |
59ed446f | 3334 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3335 | |
3336 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3337 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3338 | values[n++] = enabled; | |
3339 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3340 | values[n++] = running; | |
abf4868b PZ |
3341 | values[n++] = count; |
3342 | if (read_format & PERF_FORMAT_ID) | |
3343 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3344 | |
3345 | size = n * sizeof(u64); | |
3346 | ||
3347 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3348 | goto unlock; |
3dab77fb | 3349 | |
6f10581a | 3350 | ret = size; |
3dab77fb | 3351 | |
65abc865 | 3352 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3353 | n = 0; |
3dab77fb | 3354 | |
59ed446f | 3355 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3356 | if (read_format & PERF_FORMAT_ID) |
3357 | values[n++] = primary_event_id(sub); | |
3358 | ||
3359 | size = n * sizeof(u64); | |
3360 | ||
184d3da8 | 3361 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3362 | ret = -EFAULT; |
3363 | goto unlock; | |
3364 | } | |
abf4868b PZ |
3365 | |
3366 | ret += size; | |
3dab77fb | 3367 | } |
6f10581a PZ |
3368 | unlock: |
3369 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3370 | |
abf4868b | 3371 | return ret; |
3dab77fb PZ |
3372 | } |
3373 | ||
cdd6c482 | 3374 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3375 | u64 read_format, char __user *buf) |
3376 | { | |
59ed446f | 3377 | u64 enabled, running; |
3dab77fb PZ |
3378 | u64 values[4]; |
3379 | int n = 0; | |
3380 | ||
59ed446f PZ |
3381 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3382 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3383 | values[n++] = enabled; | |
3384 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3385 | values[n++] = running; | |
3dab77fb | 3386 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3387 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3388 | |
3389 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3390 | return -EFAULT; | |
3391 | ||
3392 | return n * sizeof(u64); | |
3393 | } | |
3394 | ||
0793a61d | 3395 | /* |
cdd6c482 | 3396 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3397 | */ |
3398 | static ssize_t | |
cdd6c482 | 3399 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3400 | { |
cdd6c482 | 3401 | u64 read_format = event->attr.read_format; |
3dab77fb | 3402 | int ret; |
0793a61d | 3403 | |
3b6f9e5c | 3404 | /* |
cdd6c482 | 3405 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3406 | * error state (i.e. because it was pinned but it couldn't be |
3407 | * scheduled on to the CPU at some point). | |
3408 | */ | |
cdd6c482 | 3409 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3410 | return 0; |
3411 | ||
c320c7b7 | 3412 | if (count < event->read_size) |
3dab77fb PZ |
3413 | return -ENOSPC; |
3414 | ||
cdd6c482 | 3415 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3416 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3417 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3418 | else |
cdd6c482 | 3419 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3420 | |
3dab77fb | 3421 | return ret; |
0793a61d TG |
3422 | } |
3423 | ||
0793a61d TG |
3424 | static ssize_t |
3425 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3426 | { | |
cdd6c482 | 3427 | struct perf_event *event = file->private_data; |
0793a61d | 3428 | |
cdd6c482 | 3429 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3430 | } |
3431 | ||
3432 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3433 | { | |
cdd6c482 | 3434 | struct perf_event *event = file->private_data; |
76369139 | 3435 | struct ring_buffer *rb; |
c33a0bc4 | 3436 | unsigned int events = POLL_HUP; |
c7138f37 | 3437 | |
10c6db11 | 3438 | /* |
9bb5d40c PZ |
3439 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3440 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3441 | */ |
3442 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3443 | rb = event->rb; |
3444 | if (rb) | |
76369139 | 3445 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 PZ |
3446 | mutex_unlock(&event->mmap_mutex); |
3447 | ||
cdd6c482 | 3448 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3449 | |
0793a61d TG |
3450 | return events; |
3451 | } | |
3452 | ||
cdd6c482 | 3453 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3454 | { |
cdd6c482 | 3455 | (void)perf_event_read(event); |
e7850595 | 3456 | local64_set(&event->count, 0); |
cdd6c482 | 3457 | perf_event_update_userpage(event); |
3df5edad PZ |
3458 | } |
3459 | ||
c93f7669 | 3460 | /* |
cdd6c482 IM |
3461 | * Holding the top-level event's child_mutex means that any |
3462 | * descendant process that has inherited this event will block | |
3463 | * in sync_child_event if it goes to exit, thus satisfying the | |
3464 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3465 | */ |
cdd6c482 IM |
3466 | static void perf_event_for_each_child(struct perf_event *event, |
3467 | void (*func)(struct perf_event *)) | |
3df5edad | 3468 | { |
cdd6c482 | 3469 | struct perf_event *child; |
3df5edad | 3470 | |
cdd6c482 IM |
3471 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3472 | mutex_lock(&event->child_mutex); | |
3473 | func(event); | |
3474 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3475 | func(child); |
cdd6c482 | 3476 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3477 | } |
3478 | ||
cdd6c482 IM |
3479 | static void perf_event_for_each(struct perf_event *event, |
3480 | void (*func)(struct perf_event *)) | |
3df5edad | 3481 | { |
cdd6c482 IM |
3482 | struct perf_event_context *ctx = event->ctx; |
3483 | struct perf_event *sibling; | |
3df5edad | 3484 | |
75f937f2 PZ |
3485 | WARN_ON_ONCE(ctx->parent_ctx); |
3486 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3487 | event = event->group_leader; |
75f937f2 | 3488 | |
cdd6c482 | 3489 | perf_event_for_each_child(event, func); |
cdd6c482 | 3490 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3491 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3492 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3493 | } |
3494 | ||
cdd6c482 | 3495 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3496 | { |
cdd6c482 | 3497 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3498 | int ret = 0; |
3499 | u64 value; | |
3500 | ||
6c7e550f | 3501 | if (!is_sampling_event(event)) |
08247e31 PZ |
3502 | return -EINVAL; |
3503 | ||
ad0cf347 | 3504 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3505 | return -EFAULT; |
3506 | ||
3507 | if (!value) | |
3508 | return -EINVAL; | |
3509 | ||
e625cce1 | 3510 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3511 | if (event->attr.freq) { |
3512 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3513 | ret = -EINVAL; |
3514 | goto unlock; | |
3515 | } | |
3516 | ||
cdd6c482 | 3517 | event->attr.sample_freq = value; |
08247e31 | 3518 | } else { |
cdd6c482 IM |
3519 | event->attr.sample_period = value; |
3520 | event->hw.sample_period = value; | |
08247e31 PZ |
3521 | } |
3522 | unlock: | |
e625cce1 | 3523 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3524 | |
3525 | return ret; | |
3526 | } | |
3527 | ||
ac9721f3 PZ |
3528 | static const struct file_operations perf_fops; |
3529 | ||
2903ff01 | 3530 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3531 | { |
2903ff01 AV |
3532 | struct fd f = fdget(fd); |
3533 | if (!f.file) | |
3534 | return -EBADF; | |
ac9721f3 | 3535 | |
2903ff01 AV |
3536 | if (f.file->f_op != &perf_fops) { |
3537 | fdput(f); | |
3538 | return -EBADF; | |
ac9721f3 | 3539 | } |
2903ff01 AV |
3540 | *p = f; |
3541 | return 0; | |
ac9721f3 PZ |
3542 | } |
3543 | ||
3544 | static int perf_event_set_output(struct perf_event *event, | |
3545 | struct perf_event *output_event); | |
6fb2915d | 3546 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3547 | |
d859e29f PM |
3548 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3549 | { | |
cdd6c482 IM |
3550 | struct perf_event *event = file->private_data; |
3551 | void (*func)(struct perf_event *); | |
3df5edad | 3552 | u32 flags = arg; |
d859e29f PM |
3553 | |
3554 | switch (cmd) { | |
cdd6c482 IM |
3555 | case PERF_EVENT_IOC_ENABLE: |
3556 | func = perf_event_enable; | |
d859e29f | 3557 | break; |
cdd6c482 IM |
3558 | case PERF_EVENT_IOC_DISABLE: |
3559 | func = perf_event_disable; | |
79f14641 | 3560 | break; |
cdd6c482 IM |
3561 | case PERF_EVENT_IOC_RESET: |
3562 | func = perf_event_reset; | |
6de6a7b9 | 3563 | break; |
3df5edad | 3564 | |
cdd6c482 IM |
3565 | case PERF_EVENT_IOC_REFRESH: |
3566 | return perf_event_refresh(event, arg); | |
08247e31 | 3567 | |
cdd6c482 IM |
3568 | case PERF_EVENT_IOC_PERIOD: |
3569 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3570 | |
cf4957f1 JO |
3571 | case PERF_EVENT_IOC_ID: |
3572 | { | |
3573 | u64 id = primary_event_id(event); | |
3574 | ||
3575 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
3576 | return -EFAULT; | |
3577 | return 0; | |
3578 | } | |
3579 | ||
cdd6c482 | 3580 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3581 | { |
ac9721f3 | 3582 | int ret; |
ac9721f3 | 3583 | if (arg != -1) { |
2903ff01 AV |
3584 | struct perf_event *output_event; |
3585 | struct fd output; | |
3586 | ret = perf_fget_light(arg, &output); | |
3587 | if (ret) | |
3588 | return ret; | |
3589 | output_event = output.file->private_data; | |
3590 | ret = perf_event_set_output(event, output_event); | |
3591 | fdput(output); | |
3592 | } else { | |
3593 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3594 | } |
ac9721f3 PZ |
3595 | return ret; |
3596 | } | |
a4be7c27 | 3597 | |
6fb2915d LZ |
3598 | case PERF_EVENT_IOC_SET_FILTER: |
3599 | return perf_event_set_filter(event, (void __user *)arg); | |
3600 | ||
d859e29f | 3601 | default: |
3df5edad | 3602 | return -ENOTTY; |
d859e29f | 3603 | } |
3df5edad PZ |
3604 | |
3605 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3606 | perf_event_for_each(event, func); |
3df5edad | 3607 | else |
cdd6c482 | 3608 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3609 | |
3610 | return 0; | |
d859e29f PM |
3611 | } |
3612 | ||
cdd6c482 | 3613 | int perf_event_task_enable(void) |
771d7cde | 3614 | { |
cdd6c482 | 3615 | struct perf_event *event; |
771d7cde | 3616 | |
cdd6c482 IM |
3617 | mutex_lock(¤t->perf_event_mutex); |
3618 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3619 | perf_event_for_each_child(event, perf_event_enable); | |
3620 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3621 | |
3622 | return 0; | |
3623 | } | |
3624 | ||
cdd6c482 | 3625 | int perf_event_task_disable(void) |
771d7cde | 3626 | { |
cdd6c482 | 3627 | struct perf_event *event; |
771d7cde | 3628 | |
cdd6c482 IM |
3629 | mutex_lock(¤t->perf_event_mutex); |
3630 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3631 | perf_event_for_each_child(event, perf_event_disable); | |
3632 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3633 | |
3634 | return 0; | |
3635 | } | |
3636 | ||
cdd6c482 | 3637 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3638 | { |
a4eaf7f1 PZ |
3639 | if (event->hw.state & PERF_HES_STOPPED) |
3640 | return 0; | |
3641 | ||
cdd6c482 | 3642 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3643 | return 0; |
3644 | ||
35edc2a5 | 3645 | return event->pmu->event_idx(event); |
194002b2 PZ |
3646 | } |
3647 | ||
c4794295 | 3648 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3649 | u64 *now, |
7f310a5d EM |
3650 | u64 *enabled, |
3651 | u64 *running) | |
c4794295 | 3652 | { |
e3f3541c | 3653 | u64 ctx_time; |
c4794295 | 3654 | |
e3f3541c PZ |
3655 | *now = perf_clock(); |
3656 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3657 | *enabled = ctx_time - event->tstamp_enabled; |
3658 | *running = ctx_time - event->tstamp_running; | |
3659 | } | |
3660 | ||
c7206205 | 3661 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3662 | { |
3663 | } | |
3664 | ||
38ff667b PZ |
3665 | /* |
3666 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3667 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3668 | * code calls this from NMI context. | |
3669 | */ | |
cdd6c482 | 3670 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3671 | { |
cdd6c482 | 3672 | struct perf_event_mmap_page *userpg; |
76369139 | 3673 | struct ring_buffer *rb; |
e3f3541c | 3674 | u64 enabled, running, now; |
38ff667b PZ |
3675 | |
3676 | rcu_read_lock(); | |
0d641208 EM |
3677 | /* |
3678 | * compute total_time_enabled, total_time_running | |
3679 | * based on snapshot values taken when the event | |
3680 | * was last scheduled in. | |
3681 | * | |
3682 | * we cannot simply called update_context_time() | |
3683 | * because of locking issue as we can be called in | |
3684 | * NMI context | |
3685 | */ | |
e3f3541c | 3686 | calc_timer_values(event, &now, &enabled, &running); |
76369139 FW |
3687 | rb = rcu_dereference(event->rb); |
3688 | if (!rb) | |
38ff667b PZ |
3689 | goto unlock; |
3690 | ||
76369139 | 3691 | userpg = rb->user_page; |
37d81828 | 3692 | |
7b732a75 PZ |
3693 | /* |
3694 | * Disable preemption so as to not let the corresponding user-space | |
3695 | * spin too long if we get preempted. | |
3696 | */ | |
3697 | preempt_disable(); | |
37d81828 | 3698 | ++userpg->lock; |
92f22a38 | 3699 | barrier(); |
cdd6c482 | 3700 | userpg->index = perf_event_index(event); |
b5e58793 | 3701 | userpg->offset = perf_event_count(event); |
365a4038 | 3702 | if (userpg->index) |
e7850595 | 3703 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3704 | |
0d641208 | 3705 | userpg->time_enabled = enabled + |
cdd6c482 | 3706 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3707 | |
0d641208 | 3708 | userpg->time_running = running + |
cdd6c482 | 3709 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3710 | |
c7206205 | 3711 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3712 | |
92f22a38 | 3713 | barrier(); |
37d81828 | 3714 | ++userpg->lock; |
7b732a75 | 3715 | preempt_enable(); |
38ff667b | 3716 | unlock: |
7b732a75 | 3717 | rcu_read_unlock(); |
37d81828 PM |
3718 | } |
3719 | ||
906010b2 PZ |
3720 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3721 | { | |
3722 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3723 | struct ring_buffer *rb; |
906010b2 PZ |
3724 | int ret = VM_FAULT_SIGBUS; |
3725 | ||
3726 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3727 | if (vmf->pgoff == 0) | |
3728 | ret = 0; | |
3729 | return ret; | |
3730 | } | |
3731 | ||
3732 | rcu_read_lock(); | |
76369139 FW |
3733 | rb = rcu_dereference(event->rb); |
3734 | if (!rb) | |
906010b2 PZ |
3735 | goto unlock; |
3736 | ||
3737 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3738 | goto unlock; | |
3739 | ||
76369139 | 3740 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3741 | if (!vmf->page) |
3742 | goto unlock; | |
3743 | ||
3744 | get_page(vmf->page); | |
3745 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3746 | vmf->page->index = vmf->pgoff; | |
3747 | ||
3748 | ret = 0; | |
3749 | unlock: | |
3750 | rcu_read_unlock(); | |
3751 | ||
3752 | return ret; | |
3753 | } | |
3754 | ||
10c6db11 PZ |
3755 | static void ring_buffer_attach(struct perf_event *event, |
3756 | struct ring_buffer *rb) | |
3757 | { | |
3758 | unsigned long flags; | |
3759 | ||
3760 | if (!list_empty(&event->rb_entry)) | |
3761 | return; | |
3762 | ||
3763 | spin_lock_irqsave(&rb->event_lock, flags); | |
9bb5d40c PZ |
3764 | if (list_empty(&event->rb_entry)) |
3765 | list_add(&event->rb_entry, &rb->event_list); | |
10c6db11 PZ |
3766 | spin_unlock_irqrestore(&rb->event_lock, flags); |
3767 | } | |
3768 | ||
9bb5d40c | 3769 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb) |
10c6db11 PZ |
3770 | { |
3771 | unsigned long flags; | |
3772 | ||
3773 | if (list_empty(&event->rb_entry)) | |
3774 | return; | |
3775 | ||
3776 | spin_lock_irqsave(&rb->event_lock, flags); | |
3777 | list_del_init(&event->rb_entry); | |
3778 | wake_up_all(&event->waitq); | |
3779 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3780 | } | |
3781 | ||
3782 | static void ring_buffer_wakeup(struct perf_event *event) | |
3783 | { | |
3784 | struct ring_buffer *rb; | |
3785 | ||
3786 | rcu_read_lock(); | |
3787 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
3788 | if (rb) { |
3789 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
3790 | wake_up_all(&event->waitq); | |
3791 | } | |
10c6db11 PZ |
3792 | rcu_read_unlock(); |
3793 | } | |
3794 | ||
76369139 | 3795 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3796 | { |
76369139 | 3797 | struct ring_buffer *rb; |
906010b2 | 3798 | |
76369139 FW |
3799 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3800 | rb_free(rb); | |
7b732a75 PZ |
3801 | } |
3802 | ||
76369139 | 3803 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3804 | { |
76369139 | 3805 | struct ring_buffer *rb; |
7b732a75 | 3806 | |
ac9721f3 | 3807 | rcu_read_lock(); |
76369139 FW |
3808 | rb = rcu_dereference(event->rb); |
3809 | if (rb) { | |
3810 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3811 | rb = NULL; | |
ac9721f3 PZ |
3812 | } |
3813 | rcu_read_unlock(); | |
3814 | ||
76369139 | 3815 | return rb; |
ac9721f3 PZ |
3816 | } |
3817 | ||
76369139 | 3818 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3819 | { |
76369139 | 3820 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3821 | return; |
7b732a75 | 3822 | |
9bb5d40c | 3823 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 3824 | |
76369139 | 3825 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3826 | } |
3827 | ||
3828 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3829 | { | |
cdd6c482 | 3830 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3831 | |
cdd6c482 | 3832 | atomic_inc(&event->mmap_count); |
9bb5d40c | 3833 | atomic_inc(&event->rb->mmap_count); |
7b732a75 PZ |
3834 | } |
3835 | ||
9bb5d40c PZ |
3836 | /* |
3837 | * A buffer can be mmap()ed multiple times; either directly through the same | |
3838 | * event, or through other events by use of perf_event_set_output(). | |
3839 | * | |
3840 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
3841 | * the buffer here, where we still have a VM context. This means we need | |
3842 | * to detach all events redirecting to us. | |
3843 | */ | |
7b732a75 PZ |
3844 | static void perf_mmap_close(struct vm_area_struct *vma) |
3845 | { | |
cdd6c482 | 3846 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3847 | |
9bb5d40c PZ |
3848 | struct ring_buffer *rb = event->rb; |
3849 | struct user_struct *mmap_user = rb->mmap_user; | |
3850 | int mmap_locked = rb->mmap_locked; | |
3851 | unsigned long size = perf_data_size(rb); | |
789f90fc | 3852 | |
9bb5d40c PZ |
3853 | atomic_dec(&rb->mmap_count); |
3854 | ||
3855 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
3856 | return; | |
3857 | ||
3858 | /* Detach current event from the buffer. */ | |
3859 | rcu_assign_pointer(event->rb, NULL); | |
3860 | ring_buffer_detach(event, rb); | |
3861 | mutex_unlock(&event->mmap_mutex); | |
3862 | ||
3863 | /* If there's still other mmap()s of this buffer, we're done. */ | |
3864 | if (atomic_read(&rb->mmap_count)) { | |
3865 | ring_buffer_put(rb); /* can't be last */ | |
3866 | return; | |
3867 | } | |
ac9721f3 | 3868 | |
9bb5d40c PZ |
3869 | /* |
3870 | * No other mmap()s, detach from all other events that might redirect | |
3871 | * into the now unreachable buffer. Somewhat complicated by the | |
3872 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
3873 | */ | |
3874 | again: | |
3875 | rcu_read_lock(); | |
3876 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
3877 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
3878 | /* | |
3879 | * This event is en-route to free_event() which will | |
3880 | * detach it and remove it from the list. | |
3881 | */ | |
3882 | continue; | |
3883 | } | |
3884 | rcu_read_unlock(); | |
789f90fc | 3885 | |
9bb5d40c PZ |
3886 | mutex_lock(&event->mmap_mutex); |
3887 | /* | |
3888 | * Check we didn't race with perf_event_set_output() which can | |
3889 | * swizzle the rb from under us while we were waiting to | |
3890 | * acquire mmap_mutex. | |
3891 | * | |
3892 | * If we find a different rb; ignore this event, a next | |
3893 | * iteration will no longer find it on the list. We have to | |
3894 | * still restart the iteration to make sure we're not now | |
3895 | * iterating the wrong list. | |
3896 | */ | |
3897 | if (event->rb == rb) { | |
3898 | rcu_assign_pointer(event->rb, NULL); | |
3899 | ring_buffer_detach(event, rb); | |
3900 | ring_buffer_put(rb); /* can't be last, we still have one */ | |
26cb63ad | 3901 | } |
cdd6c482 | 3902 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 3903 | put_event(event); |
ac9721f3 | 3904 | |
9bb5d40c PZ |
3905 | /* |
3906 | * Restart the iteration; either we're on the wrong list or | |
3907 | * destroyed its integrity by doing a deletion. | |
3908 | */ | |
3909 | goto again; | |
7b732a75 | 3910 | } |
9bb5d40c PZ |
3911 | rcu_read_unlock(); |
3912 | ||
3913 | /* | |
3914 | * It could be there's still a few 0-ref events on the list; they'll | |
3915 | * get cleaned up by free_event() -- they'll also still have their | |
3916 | * ref on the rb and will free it whenever they are done with it. | |
3917 | * | |
3918 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
3919 | * undo the VM accounting. | |
3920 | */ | |
3921 | ||
3922 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
3923 | vma->vm_mm->pinned_vm -= mmap_locked; | |
3924 | free_uid(mmap_user); | |
3925 | ||
3926 | ring_buffer_put(rb); /* could be last */ | |
37d81828 PM |
3927 | } |
3928 | ||
f0f37e2f | 3929 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3930 | .open = perf_mmap_open, |
3931 | .close = perf_mmap_close, | |
3932 | .fault = perf_mmap_fault, | |
3933 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3934 | }; |
3935 | ||
3936 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3937 | { | |
cdd6c482 | 3938 | struct perf_event *event = file->private_data; |
22a4f650 | 3939 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3940 | struct user_struct *user = current_user(); |
22a4f650 | 3941 | unsigned long locked, lock_limit; |
76369139 | 3942 | struct ring_buffer *rb; |
7b732a75 PZ |
3943 | unsigned long vma_size; |
3944 | unsigned long nr_pages; | |
789f90fc | 3945 | long user_extra, extra; |
d57e34fd | 3946 | int ret = 0, flags = 0; |
37d81828 | 3947 | |
c7920614 PZ |
3948 | /* |
3949 | * Don't allow mmap() of inherited per-task counters. This would | |
3950 | * create a performance issue due to all children writing to the | |
76369139 | 3951 | * same rb. |
c7920614 PZ |
3952 | */ |
3953 | if (event->cpu == -1 && event->attr.inherit) | |
3954 | return -EINVAL; | |
3955 | ||
43a21ea8 | 3956 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3957 | return -EINVAL; |
7b732a75 PZ |
3958 | |
3959 | vma_size = vma->vm_end - vma->vm_start; | |
3960 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3961 | ||
7730d865 | 3962 | /* |
76369139 | 3963 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3964 | * can do bitmasks instead of modulo. |
3965 | */ | |
3966 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3967 | return -EINVAL; |
3968 | ||
7b732a75 | 3969 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3970 | return -EINVAL; |
3971 | ||
7b732a75 PZ |
3972 | if (vma->vm_pgoff != 0) |
3973 | return -EINVAL; | |
37d81828 | 3974 | |
cdd6c482 | 3975 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 3976 | again: |
cdd6c482 | 3977 | mutex_lock(&event->mmap_mutex); |
76369139 | 3978 | if (event->rb) { |
9bb5d40c | 3979 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 3980 | ret = -EINVAL; |
9bb5d40c PZ |
3981 | goto unlock; |
3982 | } | |
3983 | ||
3984 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
3985 | /* | |
3986 | * Raced against perf_mmap_close() through | |
3987 | * perf_event_set_output(). Try again, hope for better | |
3988 | * luck. | |
3989 | */ | |
3990 | mutex_unlock(&event->mmap_mutex); | |
3991 | goto again; | |
3992 | } | |
3993 | ||
ebb3c4c4 PZ |
3994 | goto unlock; |
3995 | } | |
3996 | ||
789f90fc | 3997 | user_extra = nr_pages + 1; |
cdd6c482 | 3998 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
3999 | |
4000 | /* | |
4001 | * Increase the limit linearly with more CPUs: | |
4002 | */ | |
4003 | user_lock_limit *= num_online_cpus(); | |
4004 | ||
789f90fc | 4005 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4006 | |
789f90fc PZ |
4007 | extra = 0; |
4008 | if (user_locked > user_lock_limit) | |
4009 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4010 | |
78d7d407 | 4011 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4012 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4013 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4014 | |
459ec28a IM |
4015 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4016 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4017 | ret = -EPERM; |
4018 | goto unlock; | |
4019 | } | |
7b732a75 | 4020 | |
76369139 | 4021 | WARN_ON(event->rb); |
906010b2 | 4022 | |
d57e34fd | 4023 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4024 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4025 | |
4ec8363d VW |
4026 | rb = rb_alloc(nr_pages, |
4027 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4028 | event->cpu, flags); | |
4029 | ||
76369139 | 4030 | if (!rb) { |
ac9721f3 | 4031 | ret = -ENOMEM; |
ebb3c4c4 | 4032 | goto unlock; |
ac9721f3 | 4033 | } |
26cb63ad | 4034 | |
9bb5d40c | 4035 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
4036 | rb->mmap_locked = extra; |
4037 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 4038 | |
ac9721f3 | 4039 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
4040 | vma->vm_mm->pinned_vm += extra; |
4041 | ||
9bb5d40c | 4042 | ring_buffer_attach(event, rb); |
26cb63ad | 4043 | rcu_assign_pointer(event->rb, rb); |
ac9721f3 | 4044 | |
9a0f05cb PZ |
4045 | perf_event_update_userpage(event); |
4046 | ||
ebb3c4c4 | 4047 | unlock: |
ac9721f3 PZ |
4048 | if (!ret) |
4049 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 4050 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4051 | |
9bb5d40c PZ |
4052 | /* |
4053 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4054 | * vma. | |
4055 | */ | |
26cb63ad | 4056 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4057 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
4058 | |
4059 | return ret; | |
37d81828 PM |
4060 | } |
4061 | ||
3c446b3d PZ |
4062 | static int perf_fasync(int fd, struct file *filp, int on) |
4063 | { | |
496ad9aa | 4064 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4065 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4066 | int retval; |
4067 | ||
4068 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4069 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4070 | mutex_unlock(&inode->i_mutex); |
4071 | ||
4072 | if (retval < 0) | |
4073 | return retval; | |
4074 | ||
4075 | return 0; | |
4076 | } | |
4077 | ||
0793a61d | 4078 | static const struct file_operations perf_fops = { |
3326c1ce | 4079 | .llseek = no_llseek, |
0793a61d TG |
4080 | .release = perf_release, |
4081 | .read = perf_read, | |
4082 | .poll = perf_poll, | |
d859e29f PM |
4083 | .unlocked_ioctl = perf_ioctl, |
4084 | .compat_ioctl = perf_ioctl, | |
37d81828 | 4085 | .mmap = perf_mmap, |
3c446b3d | 4086 | .fasync = perf_fasync, |
0793a61d TG |
4087 | }; |
4088 | ||
925d519a | 4089 | /* |
cdd6c482 | 4090 | * Perf event wakeup |
925d519a PZ |
4091 | * |
4092 | * If there's data, ensure we set the poll() state and publish everything | |
4093 | * to user-space before waking everybody up. | |
4094 | */ | |
4095 | ||
cdd6c482 | 4096 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4097 | { |
10c6db11 | 4098 | ring_buffer_wakeup(event); |
4c9e2542 | 4099 | |
cdd6c482 IM |
4100 | if (event->pending_kill) { |
4101 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4102 | event->pending_kill = 0; | |
4c9e2542 | 4103 | } |
925d519a PZ |
4104 | } |
4105 | ||
e360adbe | 4106 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4107 | { |
cdd6c482 IM |
4108 | struct perf_event *event = container_of(entry, |
4109 | struct perf_event, pending); | |
79f14641 | 4110 | |
cdd6c482 IM |
4111 | if (event->pending_disable) { |
4112 | event->pending_disable = 0; | |
4113 | __perf_event_disable(event); | |
79f14641 PZ |
4114 | } |
4115 | ||
cdd6c482 IM |
4116 | if (event->pending_wakeup) { |
4117 | event->pending_wakeup = 0; | |
4118 | perf_event_wakeup(event); | |
79f14641 PZ |
4119 | } |
4120 | } | |
4121 | ||
39447b38 ZY |
4122 | /* |
4123 | * We assume there is only KVM supporting the callbacks. | |
4124 | * Later on, we might change it to a list if there is | |
4125 | * another virtualization implementation supporting the callbacks. | |
4126 | */ | |
4127 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4128 | ||
4129 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4130 | { | |
4131 | perf_guest_cbs = cbs; | |
4132 | return 0; | |
4133 | } | |
4134 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4135 | ||
4136 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4137 | { | |
4138 | perf_guest_cbs = NULL; | |
4139 | return 0; | |
4140 | } | |
4141 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4142 | ||
4018994f JO |
4143 | static void |
4144 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4145 | struct pt_regs *regs, u64 mask) | |
4146 | { | |
4147 | int bit; | |
4148 | ||
4149 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4150 | sizeof(mask) * BITS_PER_BYTE) { | |
4151 | u64 val; | |
4152 | ||
4153 | val = perf_reg_value(regs, bit); | |
4154 | perf_output_put(handle, val); | |
4155 | } | |
4156 | } | |
4157 | ||
4158 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
4159 | struct pt_regs *regs) | |
4160 | { | |
4161 | if (!user_mode(regs)) { | |
4162 | if (current->mm) | |
4163 | regs = task_pt_regs(current); | |
4164 | else | |
4165 | regs = NULL; | |
4166 | } | |
4167 | ||
4168 | if (regs) { | |
4169 | regs_user->regs = regs; | |
4170 | regs_user->abi = perf_reg_abi(current); | |
4171 | } | |
4172 | } | |
4173 | ||
c5ebcedb JO |
4174 | /* |
4175 | * Get remaining task size from user stack pointer. | |
4176 | * | |
4177 | * It'd be better to take stack vma map and limit this more | |
4178 | * precisly, but there's no way to get it safely under interrupt, | |
4179 | * so using TASK_SIZE as limit. | |
4180 | */ | |
4181 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4182 | { | |
4183 | unsigned long addr = perf_user_stack_pointer(regs); | |
4184 | ||
4185 | if (!addr || addr >= TASK_SIZE) | |
4186 | return 0; | |
4187 | ||
4188 | return TASK_SIZE - addr; | |
4189 | } | |
4190 | ||
4191 | static u16 | |
4192 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4193 | struct pt_regs *regs) | |
4194 | { | |
4195 | u64 task_size; | |
4196 | ||
4197 | /* No regs, no stack pointer, no dump. */ | |
4198 | if (!regs) | |
4199 | return 0; | |
4200 | ||
4201 | /* | |
4202 | * Check if we fit in with the requested stack size into the: | |
4203 | * - TASK_SIZE | |
4204 | * If we don't, we limit the size to the TASK_SIZE. | |
4205 | * | |
4206 | * - remaining sample size | |
4207 | * If we don't, we customize the stack size to | |
4208 | * fit in to the remaining sample size. | |
4209 | */ | |
4210 | ||
4211 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4212 | stack_size = min(stack_size, (u16) task_size); | |
4213 | ||
4214 | /* Current header size plus static size and dynamic size. */ | |
4215 | header_size += 2 * sizeof(u64); | |
4216 | ||
4217 | /* Do we fit in with the current stack dump size? */ | |
4218 | if ((u16) (header_size + stack_size) < header_size) { | |
4219 | /* | |
4220 | * If we overflow the maximum size for the sample, | |
4221 | * we customize the stack dump size to fit in. | |
4222 | */ | |
4223 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4224 | stack_size = round_up(stack_size, sizeof(u64)); | |
4225 | } | |
4226 | ||
4227 | return stack_size; | |
4228 | } | |
4229 | ||
4230 | static void | |
4231 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4232 | struct pt_regs *regs) | |
4233 | { | |
4234 | /* Case of a kernel thread, nothing to dump */ | |
4235 | if (!regs) { | |
4236 | u64 size = 0; | |
4237 | perf_output_put(handle, size); | |
4238 | } else { | |
4239 | unsigned long sp; | |
4240 | unsigned int rem; | |
4241 | u64 dyn_size; | |
4242 | ||
4243 | /* | |
4244 | * We dump: | |
4245 | * static size | |
4246 | * - the size requested by user or the best one we can fit | |
4247 | * in to the sample max size | |
4248 | * data | |
4249 | * - user stack dump data | |
4250 | * dynamic size | |
4251 | * - the actual dumped size | |
4252 | */ | |
4253 | ||
4254 | /* Static size. */ | |
4255 | perf_output_put(handle, dump_size); | |
4256 | ||
4257 | /* Data. */ | |
4258 | sp = perf_user_stack_pointer(regs); | |
4259 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4260 | dyn_size = dump_size - rem; | |
4261 | ||
4262 | perf_output_skip(handle, rem); | |
4263 | ||
4264 | /* Dynamic size. */ | |
4265 | perf_output_put(handle, dyn_size); | |
4266 | } | |
4267 | } | |
4268 | ||
c980d109 ACM |
4269 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4270 | struct perf_sample_data *data, | |
4271 | struct perf_event *event) | |
6844c09d ACM |
4272 | { |
4273 | u64 sample_type = event->attr.sample_type; | |
4274 | ||
4275 | data->type = sample_type; | |
4276 | header->size += event->id_header_size; | |
4277 | ||
4278 | if (sample_type & PERF_SAMPLE_TID) { | |
4279 | /* namespace issues */ | |
4280 | data->tid_entry.pid = perf_event_pid(event, current); | |
4281 | data->tid_entry.tid = perf_event_tid(event, current); | |
4282 | } | |
4283 | ||
4284 | if (sample_type & PERF_SAMPLE_TIME) | |
4285 | data->time = perf_clock(); | |
4286 | ||
4287 | if (sample_type & PERF_SAMPLE_ID) | |
4288 | data->id = primary_event_id(event); | |
4289 | ||
4290 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4291 | data->stream_id = event->id; | |
4292 | ||
4293 | if (sample_type & PERF_SAMPLE_CPU) { | |
4294 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4295 | data->cpu_entry.reserved = 0; | |
4296 | } | |
4297 | } | |
4298 | ||
76369139 FW |
4299 | void perf_event_header__init_id(struct perf_event_header *header, |
4300 | struct perf_sample_data *data, | |
4301 | struct perf_event *event) | |
c980d109 ACM |
4302 | { |
4303 | if (event->attr.sample_id_all) | |
4304 | __perf_event_header__init_id(header, data, event); | |
4305 | } | |
4306 | ||
4307 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4308 | struct perf_sample_data *data) | |
4309 | { | |
4310 | u64 sample_type = data->type; | |
4311 | ||
4312 | if (sample_type & PERF_SAMPLE_TID) | |
4313 | perf_output_put(handle, data->tid_entry); | |
4314 | ||
4315 | if (sample_type & PERF_SAMPLE_TIME) | |
4316 | perf_output_put(handle, data->time); | |
4317 | ||
4318 | if (sample_type & PERF_SAMPLE_ID) | |
4319 | perf_output_put(handle, data->id); | |
4320 | ||
4321 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4322 | perf_output_put(handle, data->stream_id); | |
4323 | ||
4324 | if (sample_type & PERF_SAMPLE_CPU) | |
4325 | perf_output_put(handle, data->cpu_entry); | |
4326 | } | |
4327 | ||
76369139 FW |
4328 | void perf_event__output_id_sample(struct perf_event *event, |
4329 | struct perf_output_handle *handle, | |
4330 | struct perf_sample_data *sample) | |
c980d109 ACM |
4331 | { |
4332 | if (event->attr.sample_id_all) | |
4333 | __perf_event__output_id_sample(handle, sample); | |
4334 | } | |
4335 | ||
3dab77fb | 4336 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4337 | struct perf_event *event, |
4338 | u64 enabled, u64 running) | |
3dab77fb | 4339 | { |
cdd6c482 | 4340 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4341 | u64 values[4]; |
4342 | int n = 0; | |
4343 | ||
b5e58793 | 4344 | values[n++] = perf_event_count(event); |
3dab77fb | 4345 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4346 | values[n++] = enabled + |
cdd6c482 | 4347 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4348 | } |
4349 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4350 | values[n++] = running + |
cdd6c482 | 4351 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4352 | } |
4353 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4354 | values[n++] = primary_event_id(event); |
3dab77fb | 4355 | |
76369139 | 4356 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4357 | } |
4358 | ||
4359 | /* | |
cdd6c482 | 4360 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4361 | */ |
4362 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4363 | struct perf_event *event, |
4364 | u64 enabled, u64 running) | |
3dab77fb | 4365 | { |
cdd6c482 IM |
4366 | struct perf_event *leader = event->group_leader, *sub; |
4367 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4368 | u64 values[5]; |
4369 | int n = 0; | |
4370 | ||
4371 | values[n++] = 1 + leader->nr_siblings; | |
4372 | ||
4373 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4374 | values[n++] = enabled; |
3dab77fb PZ |
4375 | |
4376 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4377 | values[n++] = running; |
3dab77fb | 4378 | |
cdd6c482 | 4379 | if (leader != event) |
3dab77fb PZ |
4380 | leader->pmu->read(leader); |
4381 | ||
b5e58793 | 4382 | values[n++] = perf_event_count(leader); |
3dab77fb | 4383 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4384 | values[n++] = primary_event_id(leader); |
3dab77fb | 4385 | |
76369139 | 4386 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4387 | |
65abc865 | 4388 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4389 | n = 0; |
4390 | ||
cdd6c482 | 4391 | if (sub != event) |
3dab77fb PZ |
4392 | sub->pmu->read(sub); |
4393 | ||
b5e58793 | 4394 | values[n++] = perf_event_count(sub); |
3dab77fb | 4395 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4396 | values[n++] = primary_event_id(sub); |
3dab77fb | 4397 | |
76369139 | 4398 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4399 | } |
4400 | } | |
4401 | ||
eed01528 SE |
4402 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4403 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4404 | ||
3dab77fb | 4405 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4406 | struct perf_event *event) |
3dab77fb | 4407 | { |
e3f3541c | 4408 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4409 | u64 read_format = event->attr.read_format; |
4410 | ||
4411 | /* | |
4412 | * compute total_time_enabled, total_time_running | |
4413 | * based on snapshot values taken when the event | |
4414 | * was last scheduled in. | |
4415 | * | |
4416 | * we cannot simply called update_context_time() | |
4417 | * because of locking issue as we are called in | |
4418 | * NMI context | |
4419 | */ | |
c4794295 | 4420 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4421 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4422 | |
cdd6c482 | 4423 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4424 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4425 | else |
eed01528 | 4426 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4427 | } |
4428 | ||
5622f295 MM |
4429 | void perf_output_sample(struct perf_output_handle *handle, |
4430 | struct perf_event_header *header, | |
4431 | struct perf_sample_data *data, | |
cdd6c482 | 4432 | struct perf_event *event) |
5622f295 MM |
4433 | { |
4434 | u64 sample_type = data->type; | |
4435 | ||
4436 | perf_output_put(handle, *header); | |
4437 | ||
4438 | if (sample_type & PERF_SAMPLE_IP) | |
4439 | perf_output_put(handle, data->ip); | |
4440 | ||
4441 | if (sample_type & PERF_SAMPLE_TID) | |
4442 | perf_output_put(handle, data->tid_entry); | |
4443 | ||
4444 | if (sample_type & PERF_SAMPLE_TIME) | |
4445 | perf_output_put(handle, data->time); | |
4446 | ||
4447 | if (sample_type & PERF_SAMPLE_ADDR) | |
4448 | perf_output_put(handle, data->addr); | |
4449 | ||
4450 | if (sample_type & PERF_SAMPLE_ID) | |
4451 | perf_output_put(handle, data->id); | |
4452 | ||
4453 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4454 | perf_output_put(handle, data->stream_id); | |
4455 | ||
4456 | if (sample_type & PERF_SAMPLE_CPU) | |
4457 | perf_output_put(handle, data->cpu_entry); | |
4458 | ||
4459 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4460 | perf_output_put(handle, data->period); | |
4461 | ||
4462 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4463 | perf_output_read(handle, event); |
5622f295 MM |
4464 | |
4465 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4466 | if (data->callchain) { | |
4467 | int size = 1; | |
4468 | ||
4469 | if (data->callchain) | |
4470 | size += data->callchain->nr; | |
4471 | ||
4472 | size *= sizeof(u64); | |
4473 | ||
76369139 | 4474 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4475 | } else { |
4476 | u64 nr = 0; | |
4477 | perf_output_put(handle, nr); | |
4478 | } | |
4479 | } | |
4480 | ||
4481 | if (sample_type & PERF_SAMPLE_RAW) { | |
4482 | if (data->raw) { | |
4483 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4484 | __output_copy(handle, data->raw->data, |
4485 | data->raw->size); | |
5622f295 MM |
4486 | } else { |
4487 | struct { | |
4488 | u32 size; | |
4489 | u32 data; | |
4490 | } raw = { | |
4491 | .size = sizeof(u32), | |
4492 | .data = 0, | |
4493 | }; | |
4494 | perf_output_put(handle, raw); | |
4495 | } | |
4496 | } | |
a7ac67ea | 4497 | |
bce38cd5 SE |
4498 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
4499 | if (data->br_stack) { | |
4500 | size_t size; | |
4501 | ||
4502 | size = data->br_stack->nr | |
4503 | * sizeof(struct perf_branch_entry); | |
4504 | ||
4505 | perf_output_put(handle, data->br_stack->nr); | |
4506 | perf_output_copy(handle, data->br_stack->entries, size); | |
4507 | } else { | |
4508 | /* | |
4509 | * we always store at least the value of nr | |
4510 | */ | |
4511 | u64 nr = 0; | |
4512 | perf_output_put(handle, nr); | |
4513 | } | |
4514 | } | |
4018994f JO |
4515 | |
4516 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4517 | u64 abi = data->regs_user.abi; | |
4518 | ||
4519 | /* | |
4520 | * If there are no regs to dump, notice it through | |
4521 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4522 | */ | |
4523 | perf_output_put(handle, abi); | |
4524 | ||
4525 | if (abi) { | |
4526 | u64 mask = event->attr.sample_regs_user; | |
4527 | perf_output_sample_regs(handle, | |
4528 | data->regs_user.regs, | |
4529 | mask); | |
4530 | } | |
4531 | } | |
c5ebcedb | 4532 | |
a5cdd40c | 4533 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
4534 | perf_output_sample_ustack(handle, |
4535 | data->stack_user_size, | |
4536 | data->regs_user.regs); | |
a5cdd40c | 4537 | } |
c3feedf2 AK |
4538 | |
4539 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
4540 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
4541 | |
4542 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
4543 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c PZ |
4544 | |
4545 | if (!event->attr.watermark) { | |
4546 | int wakeup_events = event->attr.wakeup_events; | |
4547 | ||
4548 | if (wakeup_events) { | |
4549 | struct ring_buffer *rb = handle->rb; | |
4550 | int events = local_inc_return(&rb->events); | |
4551 | ||
4552 | if (events >= wakeup_events) { | |
4553 | local_sub(wakeup_events, &rb->events); | |
4554 | local_inc(&rb->wakeup); | |
4555 | } | |
4556 | } | |
4557 | } | |
5622f295 MM |
4558 | } |
4559 | ||
4560 | void perf_prepare_sample(struct perf_event_header *header, | |
4561 | struct perf_sample_data *data, | |
cdd6c482 | 4562 | struct perf_event *event, |
5622f295 | 4563 | struct pt_regs *regs) |
7b732a75 | 4564 | { |
cdd6c482 | 4565 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4566 | |
cdd6c482 | 4567 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4568 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4569 | |
4570 | header->misc = 0; | |
4571 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4572 | |
c980d109 | 4573 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4574 | |
c320c7b7 | 4575 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4576 | data->ip = perf_instruction_pointer(regs); |
4577 | ||
b23f3325 | 4578 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4579 | int size = 1; |
394ee076 | 4580 | |
e6dab5ff | 4581 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4582 | |
4583 | if (data->callchain) | |
4584 | size += data->callchain->nr; | |
4585 | ||
4586 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4587 | } |
4588 | ||
3a43ce68 | 4589 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4590 | int size = sizeof(u32); |
4591 | ||
4592 | if (data->raw) | |
4593 | size += data->raw->size; | |
4594 | else | |
4595 | size += sizeof(u32); | |
4596 | ||
4597 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4598 | header->size += size; |
7f453c24 | 4599 | } |
bce38cd5 SE |
4600 | |
4601 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4602 | int size = sizeof(u64); /* nr */ | |
4603 | if (data->br_stack) { | |
4604 | size += data->br_stack->nr | |
4605 | * sizeof(struct perf_branch_entry); | |
4606 | } | |
4607 | header->size += size; | |
4608 | } | |
4018994f JO |
4609 | |
4610 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4611 | /* regs dump ABI info */ | |
4612 | int size = sizeof(u64); | |
4613 | ||
4614 | perf_sample_regs_user(&data->regs_user, regs); | |
4615 | ||
4616 | if (data->regs_user.regs) { | |
4617 | u64 mask = event->attr.sample_regs_user; | |
4618 | size += hweight64(mask) * sizeof(u64); | |
4619 | } | |
4620 | ||
4621 | header->size += size; | |
4622 | } | |
c5ebcedb JO |
4623 | |
4624 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4625 | /* | |
4626 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4627 | * processed as the last one or have additional check added | |
4628 | * in case new sample type is added, because we could eat | |
4629 | * up the rest of the sample size. | |
4630 | */ | |
4631 | struct perf_regs_user *uregs = &data->regs_user; | |
4632 | u16 stack_size = event->attr.sample_stack_user; | |
4633 | u16 size = sizeof(u64); | |
4634 | ||
4635 | if (!uregs->abi) | |
4636 | perf_sample_regs_user(uregs, regs); | |
4637 | ||
4638 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4639 | uregs->regs); | |
4640 | ||
4641 | /* | |
4642 | * If there is something to dump, add space for the dump | |
4643 | * itself and for the field that tells the dynamic size, | |
4644 | * which is how many have been actually dumped. | |
4645 | */ | |
4646 | if (stack_size) | |
4647 | size += sizeof(u64) + stack_size; | |
4648 | ||
4649 | data->stack_user_size = stack_size; | |
4650 | header->size += size; | |
4651 | } | |
5622f295 | 4652 | } |
7f453c24 | 4653 | |
a8b0ca17 | 4654 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4655 | struct perf_sample_data *data, |
4656 | struct pt_regs *regs) | |
4657 | { | |
4658 | struct perf_output_handle handle; | |
4659 | struct perf_event_header header; | |
689802b2 | 4660 | |
927c7a9e FW |
4661 | /* protect the callchain buffers */ |
4662 | rcu_read_lock(); | |
4663 | ||
cdd6c482 | 4664 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4665 | |
a7ac67ea | 4666 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4667 | goto exit; |
0322cd6e | 4668 | |
cdd6c482 | 4669 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4670 | |
8a057d84 | 4671 | perf_output_end(&handle); |
927c7a9e FW |
4672 | |
4673 | exit: | |
4674 | rcu_read_unlock(); | |
0322cd6e PZ |
4675 | } |
4676 | ||
38b200d6 | 4677 | /* |
cdd6c482 | 4678 | * read event_id |
38b200d6 PZ |
4679 | */ |
4680 | ||
4681 | struct perf_read_event { | |
4682 | struct perf_event_header header; | |
4683 | ||
4684 | u32 pid; | |
4685 | u32 tid; | |
38b200d6 PZ |
4686 | }; |
4687 | ||
4688 | static void | |
cdd6c482 | 4689 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4690 | struct task_struct *task) |
4691 | { | |
4692 | struct perf_output_handle handle; | |
c980d109 | 4693 | struct perf_sample_data sample; |
dfc65094 | 4694 | struct perf_read_event read_event = { |
38b200d6 | 4695 | .header = { |
cdd6c482 | 4696 | .type = PERF_RECORD_READ, |
38b200d6 | 4697 | .misc = 0, |
c320c7b7 | 4698 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4699 | }, |
cdd6c482 IM |
4700 | .pid = perf_event_pid(event, task), |
4701 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4702 | }; |
3dab77fb | 4703 | int ret; |
38b200d6 | 4704 | |
c980d109 | 4705 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4706 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4707 | if (ret) |
4708 | return; | |
4709 | ||
dfc65094 | 4710 | perf_output_put(&handle, read_event); |
cdd6c482 | 4711 | perf_output_read(&handle, event); |
c980d109 | 4712 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4713 | |
38b200d6 PZ |
4714 | perf_output_end(&handle); |
4715 | } | |
4716 | ||
52d857a8 JO |
4717 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
4718 | ||
4719 | static void | |
4720 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
4721 | perf_event_aux_output_cb output, |
4722 | void *data) | |
4723 | { | |
4724 | struct perf_event *event; | |
4725 | ||
4726 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
4727 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
4728 | continue; | |
4729 | if (!event_filter_match(event)) | |
4730 | continue; | |
67516844 | 4731 | output(event, data); |
52d857a8 JO |
4732 | } |
4733 | } | |
4734 | ||
4735 | static void | |
67516844 | 4736 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
4737 | struct perf_event_context *task_ctx) |
4738 | { | |
4739 | struct perf_cpu_context *cpuctx; | |
4740 | struct perf_event_context *ctx; | |
4741 | struct pmu *pmu; | |
4742 | int ctxn; | |
4743 | ||
4744 | rcu_read_lock(); | |
4745 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
4746 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
4747 | if (cpuctx->unique_pmu != pmu) | |
4748 | goto next; | |
67516844 | 4749 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
4750 | if (task_ctx) |
4751 | goto next; | |
4752 | ctxn = pmu->task_ctx_nr; | |
4753 | if (ctxn < 0) | |
4754 | goto next; | |
4755 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4756 | if (ctx) | |
67516844 | 4757 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
4758 | next: |
4759 | put_cpu_ptr(pmu->pmu_cpu_context); | |
4760 | } | |
4761 | ||
4762 | if (task_ctx) { | |
4763 | preempt_disable(); | |
67516844 | 4764 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
4765 | preempt_enable(); |
4766 | } | |
4767 | rcu_read_unlock(); | |
4768 | } | |
4769 | ||
60313ebe | 4770 | /* |
9f498cc5 PZ |
4771 | * task tracking -- fork/exit |
4772 | * | |
3af9e859 | 4773 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
60313ebe PZ |
4774 | */ |
4775 | ||
9f498cc5 | 4776 | struct perf_task_event { |
3a80b4a3 | 4777 | struct task_struct *task; |
cdd6c482 | 4778 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4779 | |
4780 | struct { | |
4781 | struct perf_event_header header; | |
4782 | ||
4783 | u32 pid; | |
4784 | u32 ppid; | |
9f498cc5 PZ |
4785 | u32 tid; |
4786 | u32 ptid; | |
393b2ad8 | 4787 | u64 time; |
cdd6c482 | 4788 | } event_id; |
60313ebe PZ |
4789 | }; |
4790 | ||
67516844 JO |
4791 | static int perf_event_task_match(struct perf_event *event) |
4792 | { | |
4793 | return event->attr.comm || event->attr.mmap || | |
4794 | event->attr.mmap_data || event->attr.task; | |
4795 | } | |
4796 | ||
cdd6c482 | 4797 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 4798 | void *data) |
60313ebe | 4799 | { |
52d857a8 | 4800 | struct perf_task_event *task_event = data; |
60313ebe | 4801 | struct perf_output_handle handle; |
c980d109 | 4802 | struct perf_sample_data sample; |
9f498cc5 | 4803 | struct task_struct *task = task_event->task; |
c980d109 | 4804 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4805 | |
67516844 JO |
4806 | if (!perf_event_task_match(event)) |
4807 | return; | |
4808 | ||
c980d109 | 4809 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4810 | |
c980d109 | 4811 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4812 | task_event->event_id.header.size); |
ef60777c | 4813 | if (ret) |
c980d109 | 4814 | goto out; |
60313ebe | 4815 | |
cdd6c482 IM |
4816 | task_event->event_id.pid = perf_event_pid(event, task); |
4817 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4818 | |
cdd6c482 IM |
4819 | task_event->event_id.tid = perf_event_tid(event, task); |
4820 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4821 | |
cdd6c482 | 4822 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4823 | |
c980d109 ACM |
4824 | perf_event__output_id_sample(event, &handle, &sample); |
4825 | ||
60313ebe | 4826 | perf_output_end(&handle); |
c980d109 ACM |
4827 | out: |
4828 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4829 | } |
4830 | ||
cdd6c482 IM |
4831 | static void perf_event_task(struct task_struct *task, |
4832 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4833 | int new) |
60313ebe | 4834 | { |
9f498cc5 | 4835 | struct perf_task_event task_event; |
60313ebe | 4836 | |
cdd6c482 IM |
4837 | if (!atomic_read(&nr_comm_events) && |
4838 | !atomic_read(&nr_mmap_events) && | |
4839 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4840 | return; |
4841 | ||
9f498cc5 | 4842 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4843 | .task = task, |
4844 | .task_ctx = task_ctx, | |
cdd6c482 | 4845 | .event_id = { |
60313ebe | 4846 | .header = { |
cdd6c482 | 4847 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4848 | .misc = 0, |
cdd6c482 | 4849 | .size = sizeof(task_event.event_id), |
60313ebe | 4850 | }, |
573402db PZ |
4851 | /* .pid */ |
4852 | /* .ppid */ | |
9f498cc5 PZ |
4853 | /* .tid */ |
4854 | /* .ptid */ | |
6f93d0a7 | 4855 | .time = perf_clock(), |
60313ebe PZ |
4856 | }, |
4857 | }; | |
4858 | ||
67516844 | 4859 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
4860 | &task_event, |
4861 | task_ctx); | |
9f498cc5 PZ |
4862 | } |
4863 | ||
cdd6c482 | 4864 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4865 | { |
cdd6c482 | 4866 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4867 | } |
4868 | ||
8d1b2d93 PZ |
4869 | /* |
4870 | * comm tracking | |
4871 | */ | |
4872 | ||
4873 | struct perf_comm_event { | |
22a4f650 IM |
4874 | struct task_struct *task; |
4875 | char *comm; | |
8d1b2d93 PZ |
4876 | int comm_size; |
4877 | ||
4878 | struct { | |
4879 | struct perf_event_header header; | |
4880 | ||
4881 | u32 pid; | |
4882 | u32 tid; | |
cdd6c482 | 4883 | } event_id; |
8d1b2d93 PZ |
4884 | }; |
4885 | ||
67516844 JO |
4886 | static int perf_event_comm_match(struct perf_event *event) |
4887 | { | |
4888 | return event->attr.comm; | |
4889 | } | |
4890 | ||
cdd6c482 | 4891 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 4892 | void *data) |
8d1b2d93 | 4893 | { |
52d857a8 | 4894 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 4895 | struct perf_output_handle handle; |
c980d109 | 4896 | struct perf_sample_data sample; |
cdd6c482 | 4897 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4898 | int ret; |
4899 | ||
67516844 JO |
4900 | if (!perf_event_comm_match(event)) |
4901 | return; | |
4902 | ||
c980d109 ACM |
4903 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
4904 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4905 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4906 | |
4907 | if (ret) | |
c980d109 | 4908 | goto out; |
8d1b2d93 | 4909 | |
cdd6c482 IM |
4910 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4911 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4912 | |
cdd6c482 | 4913 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4914 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4915 | comm_event->comm_size); |
c980d109 ACM |
4916 | |
4917 | perf_event__output_id_sample(event, &handle, &sample); | |
4918 | ||
8d1b2d93 | 4919 | perf_output_end(&handle); |
c980d109 ACM |
4920 | out: |
4921 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4922 | } |
4923 | ||
cdd6c482 | 4924 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 4925 | { |
413ee3b4 | 4926 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4927 | unsigned int size; |
8d1b2d93 | 4928 | |
413ee3b4 | 4929 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4930 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4931 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4932 | |
4933 | comm_event->comm = comm; | |
4934 | comm_event->comm_size = size; | |
4935 | ||
cdd6c482 | 4936 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 4937 | |
67516844 | 4938 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
4939 | comm_event, |
4940 | NULL); | |
8d1b2d93 PZ |
4941 | } |
4942 | ||
cdd6c482 | 4943 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4944 | { |
9ee318a7 | 4945 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4946 | struct perf_event_context *ctx; |
4947 | int ctxn; | |
9ee318a7 | 4948 | |
c79aa0d9 | 4949 | rcu_read_lock(); |
8dc85d54 PZ |
4950 | for_each_task_context_nr(ctxn) { |
4951 | ctx = task->perf_event_ctxp[ctxn]; | |
4952 | if (!ctx) | |
4953 | continue; | |
9ee318a7 | 4954 | |
8dc85d54 PZ |
4955 | perf_event_enable_on_exec(ctx); |
4956 | } | |
c79aa0d9 | 4957 | rcu_read_unlock(); |
9ee318a7 | 4958 | |
cdd6c482 | 4959 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4960 | return; |
a63eaf34 | 4961 | |
9ee318a7 | 4962 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4963 | .task = task, |
573402db PZ |
4964 | /* .comm */ |
4965 | /* .comm_size */ | |
cdd6c482 | 4966 | .event_id = { |
573402db | 4967 | .header = { |
cdd6c482 | 4968 | .type = PERF_RECORD_COMM, |
573402db PZ |
4969 | .misc = 0, |
4970 | /* .size */ | |
4971 | }, | |
4972 | /* .pid */ | |
4973 | /* .tid */ | |
8d1b2d93 PZ |
4974 | }, |
4975 | }; | |
4976 | ||
cdd6c482 | 4977 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
4978 | } |
4979 | ||
0a4a9391 PZ |
4980 | /* |
4981 | * mmap tracking | |
4982 | */ | |
4983 | ||
4984 | struct perf_mmap_event { | |
089dd79d PZ |
4985 | struct vm_area_struct *vma; |
4986 | ||
4987 | const char *file_name; | |
4988 | int file_size; | |
0a4a9391 PZ |
4989 | |
4990 | struct { | |
4991 | struct perf_event_header header; | |
4992 | ||
4993 | u32 pid; | |
4994 | u32 tid; | |
4995 | u64 start; | |
4996 | u64 len; | |
4997 | u64 pgoff; | |
cdd6c482 | 4998 | } event_id; |
0a4a9391 PZ |
4999 | }; |
5000 | ||
67516844 JO |
5001 | static int perf_event_mmap_match(struct perf_event *event, |
5002 | void *data) | |
5003 | { | |
5004 | struct perf_mmap_event *mmap_event = data; | |
5005 | struct vm_area_struct *vma = mmap_event->vma; | |
5006 | int executable = vma->vm_flags & VM_EXEC; | |
5007 | ||
5008 | return (!executable && event->attr.mmap_data) || | |
5009 | (executable && event->attr.mmap); | |
5010 | } | |
5011 | ||
cdd6c482 | 5012 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5013 | void *data) |
0a4a9391 | 5014 | { |
52d857a8 | 5015 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5016 | struct perf_output_handle handle; |
c980d109 | 5017 | struct perf_sample_data sample; |
cdd6c482 | 5018 | int size = mmap_event->event_id.header.size; |
c980d109 | 5019 | int ret; |
0a4a9391 | 5020 | |
67516844 JO |
5021 | if (!perf_event_mmap_match(event, data)) |
5022 | return; | |
5023 | ||
c980d109 ACM |
5024 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5025 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5026 | mmap_event->event_id.header.size); |
0a4a9391 | 5027 | if (ret) |
c980d109 | 5028 | goto out; |
0a4a9391 | 5029 | |
cdd6c482 IM |
5030 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5031 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5032 | |
cdd6c482 | 5033 | perf_output_put(&handle, mmap_event->event_id); |
76369139 | 5034 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5035 | mmap_event->file_size); |
c980d109 ACM |
5036 | |
5037 | perf_event__output_id_sample(event, &handle, &sample); | |
5038 | ||
78d613eb | 5039 | perf_output_end(&handle); |
c980d109 ACM |
5040 | out: |
5041 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5042 | } |
5043 | ||
cdd6c482 | 5044 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5045 | { |
089dd79d PZ |
5046 | struct vm_area_struct *vma = mmap_event->vma; |
5047 | struct file *file = vma->vm_file; | |
0a4a9391 PZ |
5048 | unsigned int size; |
5049 | char tmp[16]; | |
5050 | char *buf = NULL; | |
089dd79d | 5051 | const char *name; |
0a4a9391 | 5052 | |
413ee3b4 AB |
5053 | memset(tmp, 0, sizeof(tmp)); |
5054 | ||
0a4a9391 | 5055 | if (file) { |
413ee3b4 | 5056 | /* |
76369139 | 5057 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
5058 | * need to add enough zero bytes after the string to handle |
5059 | * the 64bit alignment we do later. | |
5060 | */ | |
5061 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
5062 | if (!buf) { |
5063 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
5064 | goto got_name; | |
5065 | } | |
d3d21c41 | 5066 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
5067 | if (IS_ERR(name)) { |
5068 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
5069 | goto got_name; | |
5070 | } | |
5071 | } else { | |
413ee3b4 AB |
5072 | if (arch_vma_name(mmap_event->vma)) { |
5073 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
c97847d2 CG |
5074 | sizeof(tmp) - 1); |
5075 | tmp[sizeof(tmp) - 1] = '\0'; | |
089dd79d | 5076 | goto got_name; |
413ee3b4 | 5077 | } |
089dd79d PZ |
5078 | |
5079 | if (!vma->vm_mm) { | |
5080 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
5081 | goto got_name; | |
3af9e859 EM |
5082 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
5083 | vma->vm_end >= vma->vm_mm->brk) { | |
5084 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
5085 | goto got_name; | |
5086 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
5087 | vma->vm_end >= vma->vm_mm->start_stack) { | |
5088 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
5089 | goto got_name; | |
089dd79d PZ |
5090 | } |
5091 | ||
0a4a9391 PZ |
5092 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
5093 | goto got_name; | |
5094 | } | |
5095 | ||
5096 | got_name: | |
888fcee0 | 5097 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
5098 | |
5099 | mmap_event->file_name = name; | |
5100 | mmap_event->file_size = size; | |
5101 | ||
2fe85427 SE |
5102 | if (!(vma->vm_flags & VM_EXEC)) |
5103 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5104 | ||
cdd6c482 | 5105 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5106 | |
67516844 | 5107 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5108 | mmap_event, |
5109 | NULL); | |
665c2142 | 5110 | |
0a4a9391 PZ |
5111 | kfree(buf); |
5112 | } | |
5113 | ||
3af9e859 | 5114 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5115 | { |
9ee318a7 PZ |
5116 | struct perf_mmap_event mmap_event; |
5117 | ||
cdd6c482 | 5118 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5119 | return; |
5120 | ||
5121 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5122 | .vma = vma, |
573402db PZ |
5123 | /* .file_name */ |
5124 | /* .file_size */ | |
cdd6c482 | 5125 | .event_id = { |
573402db | 5126 | .header = { |
cdd6c482 | 5127 | .type = PERF_RECORD_MMAP, |
39447b38 | 5128 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5129 | /* .size */ |
5130 | }, | |
5131 | /* .pid */ | |
5132 | /* .tid */ | |
089dd79d PZ |
5133 | .start = vma->vm_start, |
5134 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5135 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 PZ |
5136 | }, |
5137 | }; | |
5138 | ||
cdd6c482 | 5139 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5140 | } |
5141 | ||
a78ac325 PZ |
5142 | /* |
5143 | * IRQ throttle logging | |
5144 | */ | |
5145 | ||
cdd6c482 | 5146 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5147 | { |
5148 | struct perf_output_handle handle; | |
c980d109 | 5149 | struct perf_sample_data sample; |
a78ac325 PZ |
5150 | int ret; |
5151 | ||
5152 | struct { | |
5153 | struct perf_event_header header; | |
5154 | u64 time; | |
cca3f454 | 5155 | u64 id; |
7f453c24 | 5156 | u64 stream_id; |
a78ac325 PZ |
5157 | } throttle_event = { |
5158 | .header = { | |
cdd6c482 | 5159 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5160 | .misc = 0, |
5161 | .size = sizeof(throttle_event), | |
5162 | }, | |
def0a9b2 | 5163 | .time = perf_clock(), |
cdd6c482 IM |
5164 | .id = primary_event_id(event), |
5165 | .stream_id = event->id, | |
a78ac325 PZ |
5166 | }; |
5167 | ||
966ee4d6 | 5168 | if (enable) |
cdd6c482 | 5169 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5170 | |
c980d109 ACM |
5171 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5172 | ||
5173 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5174 | throttle_event.header.size); |
a78ac325 PZ |
5175 | if (ret) |
5176 | return; | |
5177 | ||
5178 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5179 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5180 | perf_output_end(&handle); |
5181 | } | |
5182 | ||
f6c7d5fe | 5183 | /* |
cdd6c482 | 5184 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5185 | */ |
5186 | ||
a8b0ca17 | 5187 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5188 | int throttle, struct perf_sample_data *data, |
5189 | struct pt_regs *regs) | |
f6c7d5fe | 5190 | { |
cdd6c482 IM |
5191 | int events = atomic_read(&event->event_limit); |
5192 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5193 | u64 seq; |
79f14641 PZ |
5194 | int ret = 0; |
5195 | ||
96398826 PZ |
5196 | /* |
5197 | * Non-sampling counters might still use the PMI to fold short | |
5198 | * hardware counters, ignore those. | |
5199 | */ | |
5200 | if (unlikely(!is_sampling_event(event))) | |
5201 | return 0; | |
5202 | ||
e050e3f0 SE |
5203 | seq = __this_cpu_read(perf_throttled_seq); |
5204 | if (seq != hwc->interrupts_seq) { | |
5205 | hwc->interrupts_seq = seq; | |
5206 | hwc->interrupts = 1; | |
5207 | } else { | |
5208 | hwc->interrupts++; | |
5209 | if (unlikely(throttle | |
5210 | && hwc->interrupts >= max_samples_per_tick)) { | |
5211 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5212 | hwc->interrupts = MAX_INTERRUPTS; |
5213 | perf_log_throttle(event, 0); | |
d84153d6 | 5214 | tick_nohz_full_kick(); |
a78ac325 PZ |
5215 | ret = 1; |
5216 | } | |
e050e3f0 | 5217 | } |
60db5e09 | 5218 | |
cdd6c482 | 5219 | if (event->attr.freq) { |
def0a9b2 | 5220 | u64 now = perf_clock(); |
abd50713 | 5221 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5222 | |
abd50713 | 5223 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5224 | |
abd50713 | 5225 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5226 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5227 | } |
5228 | ||
2023b359 PZ |
5229 | /* |
5230 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5231 | * events |
2023b359 PZ |
5232 | */ |
5233 | ||
cdd6c482 IM |
5234 | event->pending_kill = POLL_IN; |
5235 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5236 | ret = 1; |
cdd6c482 | 5237 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5238 | event->pending_disable = 1; |
5239 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5240 | } |
5241 | ||
453f19ee | 5242 | if (event->overflow_handler) |
a8b0ca17 | 5243 | event->overflow_handler(event, data, regs); |
453f19ee | 5244 | else |
a8b0ca17 | 5245 | perf_event_output(event, data, regs); |
453f19ee | 5246 | |
f506b3dc | 5247 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5248 | event->pending_wakeup = 1; |
5249 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5250 | } |
5251 | ||
79f14641 | 5252 | return ret; |
f6c7d5fe PZ |
5253 | } |
5254 | ||
a8b0ca17 | 5255 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5256 | struct perf_sample_data *data, |
5257 | struct pt_regs *regs) | |
850bc73f | 5258 | { |
a8b0ca17 | 5259 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5260 | } |
5261 | ||
15dbf27c | 5262 | /* |
cdd6c482 | 5263 | * Generic software event infrastructure |
15dbf27c PZ |
5264 | */ |
5265 | ||
b28ab83c PZ |
5266 | struct swevent_htable { |
5267 | struct swevent_hlist *swevent_hlist; | |
5268 | struct mutex hlist_mutex; | |
5269 | int hlist_refcount; | |
5270 | ||
5271 | /* Recursion avoidance in each contexts */ | |
5272 | int recursion[PERF_NR_CONTEXTS]; | |
5273 | }; | |
5274 | ||
5275 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5276 | ||
7b4b6658 | 5277 | /* |
cdd6c482 IM |
5278 | * We directly increment event->count and keep a second value in |
5279 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5280 | * is kept in the range [-sample_period, 0] so that we can use the |
5281 | * sign as trigger. | |
5282 | */ | |
5283 | ||
ab573844 | 5284 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5285 | { |
cdd6c482 | 5286 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5287 | u64 period = hwc->last_period; |
5288 | u64 nr, offset; | |
5289 | s64 old, val; | |
5290 | ||
5291 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5292 | |
5293 | again: | |
e7850595 | 5294 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5295 | if (val < 0) |
5296 | return 0; | |
15dbf27c | 5297 | |
7b4b6658 PZ |
5298 | nr = div64_u64(period + val, period); |
5299 | offset = nr * period; | |
5300 | val -= offset; | |
e7850595 | 5301 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5302 | goto again; |
15dbf27c | 5303 | |
7b4b6658 | 5304 | return nr; |
15dbf27c PZ |
5305 | } |
5306 | ||
0cff784a | 5307 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5308 | struct perf_sample_data *data, |
5622f295 | 5309 | struct pt_regs *regs) |
15dbf27c | 5310 | { |
cdd6c482 | 5311 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5312 | int throttle = 0; |
15dbf27c | 5313 | |
0cff784a PZ |
5314 | if (!overflow) |
5315 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5316 | |
7b4b6658 PZ |
5317 | if (hwc->interrupts == MAX_INTERRUPTS) |
5318 | return; | |
15dbf27c | 5319 | |
7b4b6658 | 5320 | for (; overflow; overflow--) { |
a8b0ca17 | 5321 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5322 | data, regs)) { |
7b4b6658 PZ |
5323 | /* |
5324 | * We inhibit the overflow from happening when | |
5325 | * hwc->interrupts == MAX_INTERRUPTS. | |
5326 | */ | |
5327 | break; | |
5328 | } | |
cf450a73 | 5329 | throttle = 1; |
7b4b6658 | 5330 | } |
15dbf27c PZ |
5331 | } |
5332 | ||
a4eaf7f1 | 5333 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5334 | struct perf_sample_data *data, |
5622f295 | 5335 | struct pt_regs *regs) |
7b4b6658 | 5336 | { |
cdd6c482 | 5337 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5338 | |
e7850595 | 5339 | local64_add(nr, &event->count); |
d6d020e9 | 5340 | |
0cff784a PZ |
5341 | if (!regs) |
5342 | return; | |
5343 | ||
6c7e550f | 5344 | if (!is_sampling_event(event)) |
7b4b6658 | 5345 | return; |
d6d020e9 | 5346 | |
5d81e5cf AV |
5347 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5348 | data->period = nr; | |
5349 | return perf_swevent_overflow(event, 1, data, regs); | |
5350 | } else | |
5351 | data->period = event->hw.last_period; | |
5352 | ||
0cff784a | 5353 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5354 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5355 | |
e7850595 | 5356 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5357 | return; |
df1a132b | 5358 | |
a8b0ca17 | 5359 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5360 | } |
5361 | ||
f5ffe02e FW |
5362 | static int perf_exclude_event(struct perf_event *event, |
5363 | struct pt_regs *regs) | |
5364 | { | |
a4eaf7f1 | 5365 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5366 | return 1; |
a4eaf7f1 | 5367 | |
f5ffe02e FW |
5368 | if (regs) { |
5369 | if (event->attr.exclude_user && user_mode(regs)) | |
5370 | return 1; | |
5371 | ||
5372 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5373 | return 1; | |
5374 | } | |
5375 | ||
5376 | return 0; | |
5377 | } | |
5378 | ||
cdd6c482 | 5379 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5380 | enum perf_type_id type, |
6fb2915d LZ |
5381 | u32 event_id, |
5382 | struct perf_sample_data *data, | |
5383 | struct pt_regs *regs) | |
15dbf27c | 5384 | { |
cdd6c482 | 5385 | if (event->attr.type != type) |
a21ca2ca | 5386 | return 0; |
f5ffe02e | 5387 | |
cdd6c482 | 5388 | if (event->attr.config != event_id) |
15dbf27c PZ |
5389 | return 0; |
5390 | ||
f5ffe02e FW |
5391 | if (perf_exclude_event(event, regs)) |
5392 | return 0; | |
15dbf27c PZ |
5393 | |
5394 | return 1; | |
5395 | } | |
5396 | ||
76e1d904 FW |
5397 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5398 | { | |
5399 | u64 val = event_id | (type << 32); | |
5400 | ||
5401 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5402 | } | |
5403 | ||
49f135ed FW |
5404 | static inline struct hlist_head * |
5405 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5406 | { |
49f135ed FW |
5407 | u64 hash = swevent_hash(type, event_id); |
5408 | ||
5409 | return &hlist->heads[hash]; | |
5410 | } | |
76e1d904 | 5411 | |
49f135ed FW |
5412 | /* For the read side: events when they trigger */ |
5413 | static inline struct hlist_head * | |
b28ab83c | 5414 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5415 | { |
5416 | struct swevent_hlist *hlist; | |
76e1d904 | 5417 | |
b28ab83c | 5418 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5419 | if (!hlist) |
5420 | return NULL; | |
5421 | ||
49f135ed FW |
5422 | return __find_swevent_head(hlist, type, event_id); |
5423 | } | |
5424 | ||
5425 | /* For the event head insertion and removal in the hlist */ | |
5426 | static inline struct hlist_head * | |
b28ab83c | 5427 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5428 | { |
5429 | struct swevent_hlist *hlist; | |
5430 | u32 event_id = event->attr.config; | |
5431 | u64 type = event->attr.type; | |
5432 | ||
5433 | /* | |
5434 | * Event scheduling is always serialized against hlist allocation | |
5435 | * and release. Which makes the protected version suitable here. | |
5436 | * The context lock guarantees that. | |
5437 | */ | |
b28ab83c | 5438 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5439 | lockdep_is_held(&event->ctx->lock)); |
5440 | if (!hlist) | |
5441 | return NULL; | |
5442 | ||
5443 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5444 | } |
5445 | ||
5446 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5447 | u64 nr, |
76e1d904 FW |
5448 | struct perf_sample_data *data, |
5449 | struct pt_regs *regs) | |
15dbf27c | 5450 | { |
b28ab83c | 5451 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5452 | struct perf_event *event; |
76e1d904 | 5453 | struct hlist_head *head; |
15dbf27c | 5454 | |
76e1d904 | 5455 | rcu_read_lock(); |
b28ab83c | 5456 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5457 | if (!head) |
5458 | goto end; | |
5459 | ||
b67bfe0d | 5460 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5461 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5462 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5463 | } |
76e1d904 FW |
5464 | end: |
5465 | rcu_read_unlock(); | |
15dbf27c PZ |
5466 | } |
5467 | ||
4ed7c92d | 5468 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5469 | { |
b28ab83c | 5470 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5471 | |
b28ab83c | 5472 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5473 | } |
645e8cc0 | 5474 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5475 | |
fa9f90be | 5476 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5477 | { |
b28ab83c | 5478 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5479 | |
b28ab83c | 5480 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5481 | } |
15dbf27c | 5482 | |
a8b0ca17 | 5483 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5484 | { |
a4234bfc | 5485 | struct perf_sample_data data; |
4ed7c92d PZ |
5486 | int rctx; |
5487 | ||
1c024eca | 5488 | preempt_disable_notrace(); |
4ed7c92d PZ |
5489 | rctx = perf_swevent_get_recursion_context(); |
5490 | if (rctx < 0) | |
5491 | return; | |
a4234bfc | 5492 | |
fd0d000b | 5493 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5494 | |
a8b0ca17 | 5495 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5496 | |
5497 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5498 | preempt_enable_notrace(); |
b8e83514 PZ |
5499 | } |
5500 | ||
cdd6c482 | 5501 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5502 | { |
15dbf27c PZ |
5503 | } |
5504 | ||
a4eaf7f1 | 5505 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5506 | { |
b28ab83c | 5507 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5508 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5509 | struct hlist_head *head; |
5510 | ||
6c7e550f | 5511 | if (is_sampling_event(event)) { |
7b4b6658 | 5512 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5513 | perf_swevent_set_period(event); |
7b4b6658 | 5514 | } |
76e1d904 | 5515 | |
a4eaf7f1 PZ |
5516 | hwc->state = !(flags & PERF_EF_START); |
5517 | ||
b28ab83c | 5518 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
5519 | if (WARN_ON_ONCE(!head)) |
5520 | return -EINVAL; | |
5521 | ||
5522 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5523 | ||
15dbf27c PZ |
5524 | return 0; |
5525 | } | |
5526 | ||
a4eaf7f1 | 5527 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5528 | { |
76e1d904 | 5529 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5530 | } |
5531 | ||
a4eaf7f1 | 5532 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5533 | { |
a4eaf7f1 | 5534 | event->hw.state = 0; |
d6d020e9 | 5535 | } |
aa9c4c0f | 5536 | |
a4eaf7f1 | 5537 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5538 | { |
a4eaf7f1 | 5539 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5540 | } |
5541 | ||
49f135ed FW |
5542 | /* Deref the hlist from the update side */ |
5543 | static inline struct swevent_hlist * | |
b28ab83c | 5544 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5545 | { |
b28ab83c PZ |
5546 | return rcu_dereference_protected(swhash->swevent_hlist, |
5547 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5548 | } |
5549 | ||
b28ab83c | 5550 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5551 | { |
b28ab83c | 5552 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5553 | |
49f135ed | 5554 | if (!hlist) |
76e1d904 FW |
5555 | return; |
5556 | ||
b28ab83c | 5557 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5558 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5559 | } |
5560 | ||
5561 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5562 | { | |
b28ab83c | 5563 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5564 | |
b28ab83c | 5565 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5566 | |
b28ab83c PZ |
5567 | if (!--swhash->hlist_refcount) |
5568 | swevent_hlist_release(swhash); | |
76e1d904 | 5569 | |
b28ab83c | 5570 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5571 | } |
5572 | ||
5573 | static void swevent_hlist_put(struct perf_event *event) | |
5574 | { | |
5575 | int cpu; | |
5576 | ||
5577 | if (event->cpu != -1) { | |
5578 | swevent_hlist_put_cpu(event, event->cpu); | |
5579 | return; | |
5580 | } | |
5581 | ||
5582 | for_each_possible_cpu(cpu) | |
5583 | swevent_hlist_put_cpu(event, cpu); | |
5584 | } | |
5585 | ||
5586 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5587 | { | |
b28ab83c | 5588 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5589 | int err = 0; |
5590 | ||
b28ab83c | 5591 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5592 | |
b28ab83c | 5593 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5594 | struct swevent_hlist *hlist; |
5595 | ||
5596 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5597 | if (!hlist) { | |
5598 | err = -ENOMEM; | |
5599 | goto exit; | |
5600 | } | |
b28ab83c | 5601 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5602 | } |
b28ab83c | 5603 | swhash->hlist_refcount++; |
9ed6060d | 5604 | exit: |
b28ab83c | 5605 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5606 | |
5607 | return err; | |
5608 | } | |
5609 | ||
5610 | static int swevent_hlist_get(struct perf_event *event) | |
5611 | { | |
5612 | int err; | |
5613 | int cpu, failed_cpu; | |
5614 | ||
5615 | if (event->cpu != -1) | |
5616 | return swevent_hlist_get_cpu(event, event->cpu); | |
5617 | ||
5618 | get_online_cpus(); | |
5619 | for_each_possible_cpu(cpu) { | |
5620 | err = swevent_hlist_get_cpu(event, cpu); | |
5621 | if (err) { | |
5622 | failed_cpu = cpu; | |
5623 | goto fail; | |
5624 | } | |
5625 | } | |
5626 | put_online_cpus(); | |
5627 | ||
5628 | return 0; | |
9ed6060d | 5629 | fail: |
76e1d904 FW |
5630 | for_each_possible_cpu(cpu) { |
5631 | if (cpu == failed_cpu) | |
5632 | break; | |
5633 | swevent_hlist_put_cpu(event, cpu); | |
5634 | } | |
5635 | ||
5636 | put_online_cpus(); | |
5637 | return err; | |
5638 | } | |
5639 | ||
c5905afb | 5640 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5641 | |
b0a873eb PZ |
5642 | static void sw_perf_event_destroy(struct perf_event *event) |
5643 | { | |
5644 | u64 event_id = event->attr.config; | |
95476b64 | 5645 | |
b0a873eb PZ |
5646 | WARN_ON(event->parent); |
5647 | ||
c5905afb | 5648 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5649 | swevent_hlist_put(event); |
5650 | } | |
5651 | ||
5652 | static int perf_swevent_init(struct perf_event *event) | |
5653 | { | |
8176cced | 5654 | u64 event_id = event->attr.config; |
b0a873eb PZ |
5655 | |
5656 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5657 | return -ENOENT; | |
5658 | ||
2481c5fa SE |
5659 | /* |
5660 | * no branch sampling for software events | |
5661 | */ | |
5662 | if (has_branch_stack(event)) | |
5663 | return -EOPNOTSUPP; | |
5664 | ||
b0a873eb PZ |
5665 | switch (event_id) { |
5666 | case PERF_COUNT_SW_CPU_CLOCK: | |
5667 | case PERF_COUNT_SW_TASK_CLOCK: | |
5668 | return -ENOENT; | |
5669 | ||
5670 | default: | |
5671 | break; | |
5672 | } | |
5673 | ||
ce677831 | 5674 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5675 | return -ENOENT; |
5676 | ||
5677 | if (!event->parent) { | |
5678 | int err; | |
5679 | ||
5680 | err = swevent_hlist_get(event); | |
5681 | if (err) | |
5682 | return err; | |
5683 | ||
c5905afb | 5684 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5685 | event->destroy = sw_perf_event_destroy; |
5686 | } | |
5687 | ||
5688 | return 0; | |
5689 | } | |
5690 | ||
35edc2a5 PZ |
5691 | static int perf_swevent_event_idx(struct perf_event *event) |
5692 | { | |
5693 | return 0; | |
5694 | } | |
5695 | ||
b0a873eb | 5696 | static struct pmu perf_swevent = { |
89a1e187 | 5697 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5698 | |
b0a873eb | 5699 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5700 | .add = perf_swevent_add, |
5701 | .del = perf_swevent_del, | |
5702 | .start = perf_swevent_start, | |
5703 | .stop = perf_swevent_stop, | |
1c024eca | 5704 | .read = perf_swevent_read, |
35edc2a5 PZ |
5705 | |
5706 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5707 | }; |
5708 | ||
b0a873eb PZ |
5709 | #ifdef CONFIG_EVENT_TRACING |
5710 | ||
1c024eca PZ |
5711 | static int perf_tp_filter_match(struct perf_event *event, |
5712 | struct perf_sample_data *data) | |
5713 | { | |
5714 | void *record = data->raw->data; | |
5715 | ||
5716 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5717 | return 1; | |
5718 | return 0; | |
5719 | } | |
5720 | ||
5721 | static int perf_tp_event_match(struct perf_event *event, | |
5722 | struct perf_sample_data *data, | |
5723 | struct pt_regs *regs) | |
5724 | { | |
a0f7d0f7 FW |
5725 | if (event->hw.state & PERF_HES_STOPPED) |
5726 | return 0; | |
580d607c PZ |
5727 | /* |
5728 | * All tracepoints are from kernel-space. | |
5729 | */ | |
5730 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5731 | return 0; |
5732 | ||
5733 | if (!perf_tp_filter_match(event, data)) | |
5734 | return 0; | |
5735 | ||
5736 | return 1; | |
5737 | } | |
5738 | ||
5739 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5740 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5741 | struct task_struct *task) | |
95476b64 FW |
5742 | { |
5743 | struct perf_sample_data data; | |
1c024eca | 5744 | struct perf_event *event; |
1c024eca | 5745 | |
95476b64 FW |
5746 | struct perf_raw_record raw = { |
5747 | .size = entry_size, | |
5748 | .data = record, | |
5749 | }; | |
5750 | ||
fd0d000b | 5751 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5752 | data.raw = &raw; |
5753 | ||
b67bfe0d | 5754 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 5755 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 5756 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5757 | } |
ecc55f84 | 5758 | |
e6dab5ff AV |
5759 | /* |
5760 | * If we got specified a target task, also iterate its context and | |
5761 | * deliver this event there too. | |
5762 | */ | |
5763 | if (task && task != current) { | |
5764 | struct perf_event_context *ctx; | |
5765 | struct trace_entry *entry = record; | |
5766 | ||
5767 | rcu_read_lock(); | |
5768 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5769 | if (!ctx) | |
5770 | goto unlock; | |
5771 | ||
5772 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5773 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5774 | continue; | |
5775 | if (event->attr.config != entry->type) | |
5776 | continue; | |
5777 | if (perf_tp_event_match(event, &data, regs)) | |
5778 | perf_swevent_event(event, count, &data, regs); | |
5779 | } | |
5780 | unlock: | |
5781 | rcu_read_unlock(); | |
5782 | } | |
5783 | ||
ecc55f84 | 5784 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5785 | } |
5786 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5787 | ||
cdd6c482 | 5788 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5789 | { |
1c024eca | 5790 | perf_trace_destroy(event); |
e077df4f PZ |
5791 | } |
5792 | ||
b0a873eb | 5793 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5794 | { |
76e1d904 FW |
5795 | int err; |
5796 | ||
b0a873eb PZ |
5797 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5798 | return -ENOENT; | |
5799 | ||
2481c5fa SE |
5800 | /* |
5801 | * no branch sampling for tracepoint events | |
5802 | */ | |
5803 | if (has_branch_stack(event)) | |
5804 | return -EOPNOTSUPP; | |
5805 | ||
1c024eca PZ |
5806 | err = perf_trace_init(event); |
5807 | if (err) | |
b0a873eb | 5808 | return err; |
e077df4f | 5809 | |
cdd6c482 | 5810 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5811 | |
b0a873eb PZ |
5812 | return 0; |
5813 | } | |
5814 | ||
5815 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5816 | .task_ctx_nr = perf_sw_context, |
5817 | ||
b0a873eb | 5818 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5819 | .add = perf_trace_add, |
5820 | .del = perf_trace_del, | |
5821 | .start = perf_swevent_start, | |
5822 | .stop = perf_swevent_stop, | |
b0a873eb | 5823 | .read = perf_swevent_read, |
35edc2a5 PZ |
5824 | |
5825 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5826 | }; |
5827 | ||
5828 | static inline void perf_tp_register(void) | |
5829 | { | |
2e80a82a | 5830 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5831 | } |
6fb2915d LZ |
5832 | |
5833 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5834 | { | |
5835 | char *filter_str; | |
5836 | int ret; | |
5837 | ||
5838 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5839 | return -EINVAL; | |
5840 | ||
5841 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5842 | if (IS_ERR(filter_str)) | |
5843 | return PTR_ERR(filter_str); | |
5844 | ||
5845 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5846 | ||
5847 | kfree(filter_str); | |
5848 | return ret; | |
5849 | } | |
5850 | ||
5851 | static void perf_event_free_filter(struct perf_event *event) | |
5852 | { | |
5853 | ftrace_profile_free_filter(event); | |
5854 | } | |
5855 | ||
e077df4f | 5856 | #else |
6fb2915d | 5857 | |
b0a873eb | 5858 | static inline void perf_tp_register(void) |
e077df4f | 5859 | { |
e077df4f | 5860 | } |
6fb2915d LZ |
5861 | |
5862 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5863 | { | |
5864 | return -ENOENT; | |
5865 | } | |
5866 | ||
5867 | static void perf_event_free_filter(struct perf_event *event) | |
5868 | { | |
5869 | } | |
5870 | ||
07b139c8 | 5871 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5872 | |
24f1e32c | 5873 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5874 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5875 | { |
f5ffe02e FW |
5876 | struct perf_sample_data sample; |
5877 | struct pt_regs *regs = data; | |
5878 | ||
fd0d000b | 5879 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 5880 | |
a4eaf7f1 | 5881 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5882 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5883 | } |
5884 | #endif | |
5885 | ||
b0a873eb PZ |
5886 | /* |
5887 | * hrtimer based swevent callback | |
5888 | */ | |
f29ac756 | 5889 | |
b0a873eb | 5890 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5891 | { |
b0a873eb PZ |
5892 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5893 | struct perf_sample_data data; | |
5894 | struct pt_regs *regs; | |
5895 | struct perf_event *event; | |
5896 | u64 period; | |
f29ac756 | 5897 | |
b0a873eb | 5898 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5899 | |
5900 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5901 | return HRTIMER_NORESTART; | |
5902 | ||
b0a873eb | 5903 | event->pmu->read(event); |
f344011c | 5904 | |
fd0d000b | 5905 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
5906 | regs = get_irq_regs(); |
5907 | ||
5908 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5909 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 5910 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
5911 | ret = HRTIMER_NORESTART; |
5912 | } | |
24f1e32c | 5913 | |
b0a873eb PZ |
5914 | period = max_t(u64, 10000, event->hw.sample_period); |
5915 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5916 | |
b0a873eb | 5917 | return ret; |
f29ac756 PZ |
5918 | } |
5919 | ||
b0a873eb | 5920 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5921 | { |
b0a873eb | 5922 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5923 | s64 period; |
5924 | ||
5925 | if (!is_sampling_event(event)) | |
5926 | return; | |
f5ffe02e | 5927 | |
5d508e82 FBH |
5928 | period = local64_read(&hwc->period_left); |
5929 | if (period) { | |
5930 | if (period < 0) | |
5931 | period = 10000; | |
fa407f35 | 5932 | |
5d508e82 FBH |
5933 | local64_set(&hwc->period_left, 0); |
5934 | } else { | |
5935 | period = max_t(u64, 10000, hwc->sample_period); | |
5936 | } | |
5937 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 5938 | ns_to_ktime(period), 0, |
b5ab4cd5 | 5939 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 5940 | } |
b0a873eb PZ |
5941 | |
5942 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 5943 | { |
b0a873eb PZ |
5944 | struct hw_perf_event *hwc = &event->hw; |
5945 | ||
6c7e550f | 5946 | if (is_sampling_event(event)) { |
b0a873eb | 5947 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 5948 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
5949 | |
5950 | hrtimer_cancel(&hwc->hrtimer); | |
5951 | } | |
24f1e32c FW |
5952 | } |
5953 | ||
ba3dd36c PZ |
5954 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
5955 | { | |
5956 | struct hw_perf_event *hwc = &event->hw; | |
5957 | ||
5958 | if (!is_sampling_event(event)) | |
5959 | return; | |
5960 | ||
5961 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5962 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
5963 | ||
5964 | /* | |
5965 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
5966 | * mapping and avoid the whole period adjust feedback stuff. | |
5967 | */ | |
5968 | if (event->attr.freq) { | |
5969 | long freq = event->attr.sample_freq; | |
5970 | ||
5971 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
5972 | hwc->sample_period = event->attr.sample_period; | |
5973 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 5974 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
5975 | event->attr.freq = 0; |
5976 | } | |
5977 | } | |
5978 | ||
b0a873eb PZ |
5979 | /* |
5980 | * Software event: cpu wall time clock | |
5981 | */ | |
5982 | ||
5983 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 5984 | { |
b0a873eb PZ |
5985 | s64 prev; |
5986 | u64 now; | |
5987 | ||
a4eaf7f1 | 5988 | now = local_clock(); |
b0a873eb PZ |
5989 | prev = local64_xchg(&event->hw.prev_count, now); |
5990 | local64_add(now - prev, &event->count); | |
24f1e32c | 5991 | } |
24f1e32c | 5992 | |
a4eaf7f1 | 5993 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 5994 | { |
a4eaf7f1 | 5995 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 5996 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
5997 | } |
5998 | ||
a4eaf7f1 | 5999 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6000 | { |
b0a873eb PZ |
6001 | perf_swevent_cancel_hrtimer(event); |
6002 | cpu_clock_event_update(event); | |
6003 | } | |
f29ac756 | 6004 | |
a4eaf7f1 PZ |
6005 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6006 | { | |
6007 | if (flags & PERF_EF_START) | |
6008 | cpu_clock_event_start(event, flags); | |
6009 | ||
6010 | return 0; | |
6011 | } | |
6012 | ||
6013 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6014 | { | |
6015 | cpu_clock_event_stop(event, flags); | |
6016 | } | |
6017 | ||
b0a873eb PZ |
6018 | static void cpu_clock_event_read(struct perf_event *event) |
6019 | { | |
6020 | cpu_clock_event_update(event); | |
6021 | } | |
f344011c | 6022 | |
b0a873eb PZ |
6023 | static int cpu_clock_event_init(struct perf_event *event) |
6024 | { | |
6025 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6026 | return -ENOENT; | |
6027 | ||
6028 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
6029 | return -ENOENT; | |
6030 | ||
2481c5fa SE |
6031 | /* |
6032 | * no branch sampling for software events | |
6033 | */ | |
6034 | if (has_branch_stack(event)) | |
6035 | return -EOPNOTSUPP; | |
6036 | ||
ba3dd36c PZ |
6037 | perf_swevent_init_hrtimer(event); |
6038 | ||
b0a873eb | 6039 | return 0; |
f29ac756 PZ |
6040 | } |
6041 | ||
b0a873eb | 6042 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
6043 | .task_ctx_nr = perf_sw_context, |
6044 | ||
b0a873eb | 6045 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
6046 | .add = cpu_clock_event_add, |
6047 | .del = cpu_clock_event_del, | |
6048 | .start = cpu_clock_event_start, | |
6049 | .stop = cpu_clock_event_stop, | |
b0a873eb | 6050 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
6051 | |
6052 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
6053 | }; |
6054 | ||
6055 | /* | |
6056 | * Software event: task time clock | |
6057 | */ | |
6058 | ||
6059 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 6060 | { |
b0a873eb PZ |
6061 | u64 prev; |
6062 | s64 delta; | |
5c92d124 | 6063 | |
b0a873eb PZ |
6064 | prev = local64_xchg(&event->hw.prev_count, now); |
6065 | delta = now - prev; | |
6066 | local64_add(delta, &event->count); | |
6067 | } | |
5c92d124 | 6068 | |
a4eaf7f1 | 6069 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6070 | { |
a4eaf7f1 | 6071 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 6072 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6073 | } |
6074 | ||
a4eaf7f1 | 6075 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
6076 | { |
6077 | perf_swevent_cancel_hrtimer(event); | |
6078 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
6079 | } |
6080 | ||
6081 | static int task_clock_event_add(struct perf_event *event, int flags) | |
6082 | { | |
6083 | if (flags & PERF_EF_START) | |
6084 | task_clock_event_start(event, flags); | |
b0a873eb | 6085 | |
a4eaf7f1 PZ |
6086 | return 0; |
6087 | } | |
6088 | ||
6089 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6090 | { | |
6091 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6092 | } |
6093 | ||
6094 | static void task_clock_event_read(struct perf_event *event) | |
6095 | { | |
768a06e2 PZ |
6096 | u64 now = perf_clock(); |
6097 | u64 delta = now - event->ctx->timestamp; | |
6098 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6099 | |
6100 | task_clock_event_update(event, time); | |
6101 | } | |
6102 | ||
6103 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6104 | { |
b0a873eb PZ |
6105 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6106 | return -ENOENT; | |
6107 | ||
6108 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6109 | return -ENOENT; | |
6110 | ||
2481c5fa SE |
6111 | /* |
6112 | * no branch sampling for software events | |
6113 | */ | |
6114 | if (has_branch_stack(event)) | |
6115 | return -EOPNOTSUPP; | |
6116 | ||
ba3dd36c PZ |
6117 | perf_swevent_init_hrtimer(event); |
6118 | ||
b0a873eb | 6119 | return 0; |
6fb2915d LZ |
6120 | } |
6121 | ||
b0a873eb | 6122 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6123 | .task_ctx_nr = perf_sw_context, |
6124 | ||
b0a873eb | 6125 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6126 | .add = task_clock_event_add, |
6127 | .del = task_clock_event_del, | |
6128 | .start = task_clock_event_start, | |
6129 | .stop = task_clock_event_stop, | |
b0a873eb | 6130 | .read = task_clock_event_read, |
35edc2a5 PZ |
6131 | |
6132 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 6133 | }; |
6fb2915d | 6134 | |
ad5133b7 | 6135 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6136 | { |
e077df4f | 6137 | } |
6fb2915d | 6138 | |
ad5133b7 | 6139 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6140 | { |
ad5133b7 | 6141 | return 0; |
6fb2915d LZ |
6142 | } |
6143 | ||
ad5133b7 | 6144 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6145 | { |
ad5133b7 | 6146 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6147 | } |
6148 | ||
ad5133b7 PZ |
6149 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6150 | { | |
6151 | perf_pmu_enable(pmu); | |
6152 | return 0; | |
6153 | } | |
e077df4f | 6154 | |
ad5133b7 | 6155 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6156 | { |
ad5133b7 | 6157 | perf_pmu_enable(pmu); |
24f1e32c FW |
6158 | } |
6159 | ||
35edc2a5 PZ |
6160 | static int perf_event_idx_default(struct perf_event *event) |
6161 | { | |
6162 | return event->hw.idx + 1; | |
6163 | } | |
6164 | ||
8dc85d54 PZ |
6165 | /* |
6166 | * Ensures all contexts with the same task_ctx_nr have the same | |
6167 | * pmu_cpu_context too. | |
6168 | */ | |
6169 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 6170 | { |
8dc85d54 | 6171 | struct pmu *pmu; |
b326e956 | 6172 | |
8dc85d54 PZ |
6173 | if (ctxn < 0) |
6174 | return NULL; | |
24f1e32c | 6175 | |
8dc85d54 PZ |
6176 | list_for_each_entry(pmu, &pmus, entry) { |
6177 | if (pmu->task_ctx_nr == ctxn) | |
6178 | return pmu->pmu_cpu_context; | |
6179 | } | |
24f1e32c | 6180 | |
8dc85d54 | 6181 | return NULL; |
24f1e32c FW |
6182 | } |
6183 | ||
51676957 | 6184 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6185 | { |
51676957 PZ |
6186 | int cpu; |
6187 | ||
6188 | for_each_possible_cpu(cpu) { | |
6189 | struct perf_cpu_context *cpuctx; | |
6190 | ||
6191 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6192 | ||
3f1f3320 PZ |
6193 | if (cpuctx->unique_pmu == old_pmu) |
6194 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6195 | } |
6196 | } | |
6197 | ||
6198 | static void free_pmu_context(struct pmu *pmu) | |
6199 | { | |
6200 | struct pmu *i; | |
f5ffe02e | 6201 | |
8dc85d54 | 6202 | mutex_lock(&pmus_lock); |
0475f9ea | 6203 | /* |
8dc85d54 | 6204 | * Like a real lame refcount. |
0475f9ea | 6205 | */ |
51676957 PZ |
6206 | list_for_each_entry(i, &pmus, entry) { |
6207 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6208 | update_pmu_context(i, pmu); | |
8dc85d54 | 6209 | goto out; |
51676957 | 6210 | } |
8dc85d54 | 6211 | } |
d6d020e9 | 6212 | |
51676957 | 6213 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6214 | out: |
6215 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6216 | } |
2e80a82a | 6217 | static struct idr pmu_idr; |
d6d020e9 | 6218 | |
abe43400 PZ |
6219 | static ssize_t |
6220 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6221 | { | |
6222 | struct pmu *pmu = dev_get_drvdata(dev); | |
6223 | ||
6224 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6225 | } | |
6226 | ||
62b85639 SE |
6227 | static ssize_t |
6228 | perf_event_mux_interval_ms_show(struct device *dev, | |
6229 | struct device_attribute *attr, | |
6230 | char *page) | |
6231 | { | |
6232 | struct pmu *pmu = dev_get_drvdata(dev); | |
6233 | ||
6234 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6235 | } | |
6236 | ||
6237 | static ssize_t | |
6238 | perf_event_mux_interval_ms_store(struct device *dev, | |
6239 | struct device_attribute *attr, | |
6240 | const char *buf, size_t count) | |
6241 | { | |
6242 | struct pmu *pmu = dev_get_drvdata(dev); | |
6243 | int timer, cpu, ret; | |
6244 | ||
6245 | ret = kstrtoint(buf, 0, &timer); | |
6246 | if (ret) | |
6247 | return ret; | |
6248 | ||
6249 | if (timer < 1) | |
6250 | return -EINVAL; | |
6251 | ||
6252 | /* same value, noting to do */ | |
6253 | if (timer == pmu->hrtimer_interval_ms) | |
6254 | return count; | |
6255 | ||
6256 | pmu->hrtimer_interval_ms = timer; | |
6257 | ||
6258 | /* update all cpuctx for this PMU */ | |
6259 | for_each_possible_cpu(cpu) { | |
6260 | struct perf_cpu_context *cpuctx; | |
6261 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6262 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6263 | ||
6264 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6265 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6266 | } | |
6267 | ||
6268 | return count; | |
6269 | } | |
6270 | ||
abe43400 | 6271 | static struct device_attribute pmu_dev_attrs[] = { |
62b85639 SE |
6272 | __ATTR_RO(type), |
6273 | __ATTR_RW(perf_event_mux_interval_ms), | |
6274 | __ATTR_NULL, | |
abe43400 PZ |
6275 | }; |
6276 | ||
6277 | static int pmu_bus_running; | |
6278 | static struct bus_type pmu_bus = { | |
6279 | .name = "event_source", | |
6280 | .dev_attrs = pmu_dev_attrs, | |
6281 | }; | |
6282 | ||
6283 | static void pmu_dev_release(struct device *dev) | |
6284 | { | |
6285 | kfree(dev); | |
6286 | } | |
6287 | ||
6288 | static int pmu_dev_alloc(struct pmu *pmu) | |
6289 | { | |
6290 | int ret = -ENOMEM; | |
6291 | ||
6292 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6293 | if (!pmu->dev) | |
6294 | goto out; | |
6295 | ||
0c9d42ed | 6296 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6297 | device_initialize(pmu->dev); |
6298 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6299 | if (ret) | |
6300 | goto free_dev; | |
6301 | ||
6302 | dev_set_drvdata(pmu->dev, pmu); | |
6303 | pmu->dev->bus = &pmu_bus; | |
6304 | pmu->dev->release = pmu_dev_release; | |
6305 | ret = device_add(pmu->dev); | |
6306 | if (ret) | |
6307 | goto free_dev; | |
6308 | ||
6309 | out: | |
6310 | return ret; | |
6311 | ||
6312 | free_dev: | |
6313 | put_device(pmu->dev); | |
6314 | goto out; | |
6315 | } | |
6316 | ||
547e9fd7 | 6317 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6318 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6319 | |
03d8e80b | 6320 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6321 | { |
108b02cf | 6322 | int cpu, ret; |
24f1e32c | 6323 | |
b0a873eb | 6324 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6325 | ret = -ENOMEM; |
6326 | pmu->pmu_disable_count = alloc_percpu(int); | |
6327 | if (!pmu->pmu_disable_count) | |
6328 | goto unlock; | |
f29ac756 | 6329 | |
2e80a82a PZ |
6330 | pmu->type = -1; |
6331 | if (!name) | |
6332 | goto skip_type; | |
6333 | pmu->name = name; | |
6334 | ||
6335 | if (type < 0) { | |
0e9c3be2 TH |
6336 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6337 | if (type < 0) { | |
6338 | ret = type; | |
2e80a82a PZ |
6339 | goto free_pdc; |
6340 | } | |
6341 | } | |
6342 | pmu->type = type; | |
6343 | ||
abe43400 PZ |
6344 | if (pmu_bus_running) { |
6345 | ret = pmu_dev_alloc(pmu); | |
6346 | if (ret) | |
6347 | goto free_idr; | |
6348 | } | |
6349 | ||
2e80a82a | 6350 | skip_type: |
8dc85d54 PZ |
6351 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6352 | if (pmu->pmu_cpu_context) | |
6353 | goto got_cpu_context; | |
f29ac756 | 6354 | |
c4814202 | 6355 | ret = -ENOMEM; |
108b02cf PZ |
6356 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6357 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6358 | goto free_dev; |
f344011c | 6359 | |
108b02cf PZ |
6360 | for_each_possible_cpu(cpu) { |
6361 | struct perf_cpu_context *cpuctx; | |
6362 | ||
6363 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6364 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6365 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6366 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6367 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6368 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6369 | |
6370 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6371 | ||
e9d2b064 | 6372 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
3f1f3320 | 6373 | cpuctx->unique_pmu = pmu; |
108b02cf | 6374 | } |
76e1d904 | 6375 | |
8dc85d54 | 6376 | got_cpu_context: |
ad5133b7 PZ |
6377 | if (!pmu->start_txn) { |
6378 | if (pmu->pmu_enable) { | |
6379 | /* | |
6380 | * If we have pmu_enable/pmu_disable calls, install | |
6381 | * transaction stubs that use that to try and batch | |
6382 | * hardware accesses. | |
6383 | */ | |
6384 | pmu->start_txn = perf_pmu_start_txn; | |
6385 | pmu->commit_txn = perf_pmu_commit_txn; | |
6386 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6387 | } else { | |
6388 | pmu->start_txn = perf_pmu_nop_void; | |
6389 | pmu->commit_txn = perf_pmu_nop_int; | |
6390 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6391 | } |
5c92d124 | 6392 | } |
15dbf27c | 6393 | |
ad5133b7 PZ |
6394 | if (!pmu->pmu_enable) { |
6395 | pmu->pmu_enable = perf_pmu_nop_void; | |
6396 | pmu->pmu_disable = perf_pmu_nop_void; | |
6397 | } | |
6398 | ||
35edc2a5 PZ |
6399 | if (!pmu->event_idx) |
6400 | pmu->event_idx = perf_event_idx_default; | |
6401 | ||
b0a873eb | 6402 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6403 | ret = 0; |
6404 | unlock: | |
b0a873eb PZ |
6405 | mutex_unlock(&pmus_lock); |
6406 | ||
33696fc0 | 6407 | return ret; |
108b02cf | 6408 | |
abe43400 PZ |
6409 | free_dev: |
6410 | device_del(pmu->dev); | |
6411 | put_device(pmu->dev); | |
6412 | ||
2e80a82a PZ |
6413 | free_idr: |
6414 | if (pmu->type >= PERF_TYPE_MAX) | |
6415 | idr_remove(&pmu_idr, pmu->type); | |
6416 | ||
108b02cf PZ |
6417 | free_pdc: |
6418 | free_percpu(pmu->pmu_disable_count); | |
6419 | goto unlock; | |
f29ac756 PZ |
6420 | } |
6421 | ||
b0a873eb | 6422 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6423 | { |
b0a873eb PZ |
6424 | mutex_lock(&pmus_lock); |
6425 | list_del_rcu(&pmu->entry); | |
6426 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6427 | |
0475f9ea | 6428 | /* |
cde8e884 PZ |
6429 | * We dereference the pmu list under both SRCU and regular RCU, so |
6430 | * synchronize against both of those. | |
0475f9ea | 6431 | */ |
b0a873eb | 6432 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6433 | synchronize_rcu(); |
d6d020e9 | 6434 | |
33696fc0 | 6435 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6436 | if (pmu->type >= PERF_TYPE_MAX) |
6437 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6438 | device_del(pmu->dev); |
6439 | put_device(pmu->dev); | |
51676957 | 6440 | free_pmu_context(pmu); |
b0a873eb | 6441 | } |
d6d020e9 | 6442 | |
b0a873eb PZ |
6443 | struct pmu *perf_init_event(struct perf_event *event) |
6444 | { | |
6445 | struct pmu *pmu = NULL; | |
6446 | int idx; | |
940c5b29 | 6447 | int ret; |
b0a873eb PZ |
6448 | |
6449 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6450 | |
6451 | rcu_read_lock(); | |
6452 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6453 | rcu_read_unlock(); | |
940c5b29 | 6454 | if (pmu) { |
7e5b2a01 | 6455 | event->pmu = pmu; |
940c5b29 LM |
6456 | ret = pmu->event_init(event); |
6457 | if (ret) | |
6458 | pmu = ERR_PTR(ret); | |
2e80a82a | 6459 | goto unlock; |
940c5b29 | 6460 | } |
2e80a82a | 6461 | |
b0a873eb | 6462 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 6463 | event->pmu = pmu; |
940c5b29 | 6464 | ret = pmu->event_init(event); |
b0a873eb | 6465 | if (!ret) |
e5f4d339 | 6466 | goto unlock; |
76e1d904 | 6467 | |
b0a873eb PZ |
6468 | if (ret != -ENOENT) { |
6469 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6470 | goto unlock; |
f344011c | 6471 | } |
5c92d124 | 6472 | } |
e5f4d339 PZ |
6473 | pmu = ERR_PTR(-ENOENT); |
6474 | unlock: | |
b0a873eb | 6475 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6476 | |
4aeb0b42 | 6477 | return pmu; |
5c92d124 IM |
6478 | } |
6479 | ||
4beb31f3 FW |
6480 | static void account_event_cpu(struct perf_event *event, int cpu) |
6481 | { | |
6482 | if (event->parent) | |
6483 | return; | |
6484 | ||
6485 | if (has_branch_stack(event)) { | |
6486 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6487 | atomic_inc(&per_cpu(perf_branch_stack_events, cpu)); | |
6488 | } | |
6489 | if (is_cgroup_event(event)) | |
6490 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
ba8a75c1 FW |
6491 | |
6492 | if (event->attr.freq) | |
6493 | atomic_inc(&per_cpu(perf_freq_events, cpu)); | |
4beb31f3 FW |
6494 | } |
6495 | ||
766d6c07 FW |
6496 | static void account_event(struct perf_event *event) |
6497 | { | |
4beb31f3 FW |
6498 | if (event->parent) |
6499 | return; | |
6500 | ||
766d6c07 FW |
6501 | if (event->attach_state & PERF_ATTACH_TASK) |
6502 | static_key_slow_inc(&perf_sched_events.key); | |
6503 | if (event->attr.mmap || event->attr.mmap_data) | |
6504 | atomic_inc(&nr_mmap_events); | |
6505 | if (event->attr.comm) | |
6506 | atomic_inc(&nr_comm_events); | |
6507 | if (event->attr.task) | |
6508 | atomic_inc(&nr_task_events); | |
4beb31f3 | 6509 | if (has_branch_stack(event)) |
766d6c07 | 6510 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 6511 | if (is_cgroup_event(event)) |
766d6c07 | 6512 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
6513 | |
6514 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
6515 | } |
6516 | ||
0793a61d | 6517 | /* |
cdd6c482 | 6518 | * Allocate and initialize a event structure |
0793a61d | 6519 | */ |
cdd6c482 | 6520 | static struct perf_event * |
c3f00c70 | 6521 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6522 | struct task_struct *task, |
6523 | struct perf_event *group_leader, | |
6524 | struct perf_event *parent_event, | |
4dc0da86 AK |
6525 | perf_overflow_handler_t overflow_handler, |
6526 | void *context) | |
0793a61d | 6527 | { |
51b0fe39 | 6528 | struct pmu *pmu; |
cdd6c482 IM |
6529 | struct perf_event *event; |
6530 | struct hw_perf_event *hwc; | |
90983b16 | 6531 | long err = -EINVAL; |
0793a61d | 6532 | |
66832eb4 ON |
6533 | if ((unsigned)cpu >= nr_cpu_ids) { |
6534 | if (!task || cpu != -1) | |
6535 | return ERR_PTR(-EINVAL); | |
6536 | } | |
6537 | ||
c3f00c70 | 6538 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6539 | if (!event) |
d5d2bc0d | 6540 | return ERR_PTR(-ENOMEM); |
0793a61d | 6541 | |
04289bb9 | 6542 | /* |
cdd6c482 | 6543 | * Single events are their own group leaders, with an |
04289bb9 IM |
6544 | * empty sibling list: |
6545 | */ | |
6546 | if (!group_leader) | |
cdd6c482 | 6547 | group_leader = event; |
04289bb9 | 6548 | |
cdd6c482 IM |
6549 | mutex_init(&event->child_mutex); |
6550 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6551 | |
cdd6c482 IM |
6552 | INIT_LIST_HEAD(&event->group_entry); |
6553 | INIT_LIST_HEAD(&event->event_entry); | |
6554 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
6555 | INIT_LIST_HEAD(&event->rb_entry); |
6556 | ||
cdd6c482 | 6557 | init_waitqueue_head(&event->waitq); |
e360adbe | 6558 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6559 | |
cdd6c482 | 6560 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6561 | |
a6fa941d | 6562 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6563 | event->cpu = cpu; |
6564 | event->attr = *attr; | |
6565 | event->group_leader = group_leader; | |
6566 | event->pmu = NULL; | |
cdd6c482 | 6567 | event->oncpu = -1; |
a96bbc16 | 6568 | |
cdd6c482 | 6569 | event->parent = parent_event; |
b84fbc9f | 6570 | |
17cf22c3 | 6571 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6572 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6573 | |
cdd6c482 | 6574 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6575 | |
d580ff86 PZ |
6576 | if (task) { |
6577 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6578 | |
6579 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6580 | event->hw.tp_target = task; | |
d580ff86 PZ |
6581 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6582 | /* | |
6583 | * hw_breakpoint is a bit difficult here.. | |
6584 | */ | |
f22c1bb6 | 6585 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6586 | event->hw.bp_target = task; |
6587 | #endif | |
6588 | } | |
6589 | ||
4dc0da86 | 6590 | if (!overflow_handler && parent_event) { |
b326e956 | 6591 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6592 | context = parent_event->overflow_handler_context; |
6593 | } | |
66832eb4 | 6594 | |
b326e956 | 6595 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6596 | event->overflow_handler_context = context; |
97eaf530 | 6597 | |
0231bb53 | 6598 | perf_event__state_init(event); |
a86ed508 | 6599 | |
4aeb0b42 | 6600 | pmu = NULL; |
b8e83514 | 6601 | |
cdd6c482 | 6602 | hwc = &event->hw; |
bd2b5b12 | 6603 | hwc->sample_period = attr->sample_period; |
0d48696f | 6604 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6605 | hwc->sample_period = 1; |
eced1dfc | 6606 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6607 | |
e7850595 | 6608 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6609 | |
2023b359 | 6610 | /* |
cdd6c482 | 6611 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6612 | */ |
3dab77fb | 6613 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 6614 | goto err_ns; |
2023b359 | 6615 | |
b0a873eb | 6616 | pmu = perf_init_event(event); |
4aeb0b42 | 6617 | if (!pmu) |
90983b16 FW |
6618 | goto err_ns; |
6619 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 6620 | err = PTR_ERR(pmu); |
90983b16 | 6621 | goto err_ns; |
621a01ea | 6622 | } |
d5d2bc0d | 6623 | |
cdd6c482 | 6624 | if (!event->parent) { |
90983b16 FW |
6625 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6626 | err = get_callchain_buffers(); | |
6627 | if (err) | |
6628 | goto err_pmu; | |
6629 | } | |
f344011c | 6630 | } |
9ee318a7 | 6631 | |
cdd6c482 | 6632 | return event; |
90983b16 FW |
6633 | |
6634 | err_pmu: | |
6635 | if (event->destroy) | |
6636 | event->destroy(event); | |
6637 | err_ns: | |
6638 | if (event->ns) | |
6639 | put_pid_ns(event->ns); | |
6640 | kfree(event); | |
6641 | ||
6642 | return ERR_PTR(err); | |
0793a61d TG |
6643 | } |
6644 | ||
cdd6c482 IM |
6645 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6646 | struct perf_event_attr *attr) | |
974802ea | 6647 | { |
974802ea | 6648 | u32 size; |
cdf8073d | 6649 | int ret; |
974802ea PZ |
6650 | |
6651 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6652 | return -EFAULT; | |
6653 | ||
6654 | /* | |
6655 | * zero the full structure, so that a short copy will be nice. | |
6656 | */ | |
6657 | memset(attr, 0, sizeof(*attr)); | |
6658 | ||
6659 | ret = get_user(size, &uattr->size); | |
6660 | if (ret) | |
6661 | return ret; | |
6662 | ||
6663 | if (size > PAGE_SIZE) /* silly large */ | |
6664 | goto err_size; | |
6665 | ||
6666 | if (!size) /* abi compat */ | |
6667 | size = PERF_ATTR_SIZE_VER0; | |
6668 | ||
6669 | if (size < PERF_ATTR_SIZE_VER0) | |
6670 | goto err_size; | |
6671 | ||
6672 | /* | |
6673 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6674 | * ensure all the unknown bits are 0 - i.e. new |
6675 | * user-space does not rely on any kernel feature | |
6676 | * extensions we dont know about yet. | |
974802ea PZ |
6677 | */ |
6678 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6679 | unsigned char __user *addr; |
6680 | unsigned char __user *end; | |
6681 | unsigned char val; | |
974802ea | 6682 | |
cdf8073d IS |
6683 | addr = (void __user *)uattr + sizeof(*attr); |
6684 | end = (void __user *)uattr + size; | |
974802ea | 6685 | |
cdf8073d | 6686 | for (; addr < end; addr++) { |
974802ea PZ |
6687 | ret = get_user(val, addr); |
6688 | if (ret) | |
6689 | return ret; | |
6690 | if (val) | |
6691 | goto err_size; | |
6692 | } | |
b3e62e35 | 6693 | size = sizeof(*attr); |
974802ea PZ |
6694 | } |
6695 | ||
6696 | ret = copy_from_user(attr, uattr, size); | |
6697 | if (ret) | |
6698 | return -EFAULT; | |
6699 | ||
cd757645 | 6700 | if (attr->__reserved_1) |
974802ea PZ |
6701 | return -EINVAL; |
6702 | ||
6703 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6704 | return -EINVAL; | |
6705 | ||
6706 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6707 | return -EINVAL; | |
6708 | ||
bce38cd5 SE |
6709 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6710 | u64 mask = attr->branch_sample_type; | |
6711 | ||
6712 | /* only using defined bits */ | |
6713 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6714 | return -EINVAL; | |
6715 | ||
6716 | /* at least one branch bit must be set */ | |
6717 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6718 | return -EINVAL; | |
6719 | ||
bce38cd5 SE |
6720 | /* propagate priv level, when not set for branch */ |
6721 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6722 | ||
6723 | /* exclude_kernel checked on syscall entry */ | |
6724 | if (!attr->exclude_kernel) | |
6725 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6726 | ||
6727 | if (!attr->exclude_user) | |
6728 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6729 | ||
6730 | if (!attr->exclude_hv) | |
6731 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6732 | /* | |
6733 | * adjust user setting (for HW filter setup) | |
6734 | */ | |
6735 | attr->branch_sample_type = mask; | |
6736 | } | |
e712209a SE |
6737 | /* privileged levels capture (kernel, hv): check permissions */ |
6738 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
6739 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
6740 | return -EACCES; | |
bce38cd5 | 6741 | } |
4018994f | 6742 | |
c5ebcedb | 6743 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 6744 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
6745 | if (ret) |
6746 | return ret; | |
6747 | } | |
6748 | ||
6749 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
6750 | if (!arch_perf_have_user_stack_dump()) | |
6751 | return -ENOSYS; | |
6752 | ||
6753 | /* | |
6754 | * We have __u32 type for the size, but so far | |
6755 | * we can only use __u16 as maximum due to the | |
6756 | * __u16 sample size limit. | |
6757 | */ | |
6758 | if (attr->sample_stack_user >= USHRT_MAX) | |
6759 | ret = -EINVAL; | |
6760 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
6761 | ret = -EINVAL; | |
6762 | } | |
4018994f | 6763 | |
974802ea PZ |
6764 | out: |
6765 | return ret; | |
6766 | ||
6767 | err_size: | |
6768 | put_user(sizeof(*attr), &uattr->size); | |
6769 | ret = -E2BIG; | |
6770 | goto out; | |
6771 | } | |
6772 | ||
ac9721f3 PZ |
6773 | static int |
6774 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6775 | { |
76369139 | 6776 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6777 | int ret = -EINVAL; |
6778 | ||
ac9721f3 | 6779 | if (!output_event) |
a4be7c27 PZ |
6780 | goto set; |
6781 | ||
ac9721f3 PZ |
6782 | /* don't allow circular references */ |
6783 | if (event == output_event) | |
a4be7c27 PZ |
6784 | goto out; |
6785 | ||
0f139300 PZ |
6786 | /* |
6787 | * Don't allow cross-cpu buffers | |
6788 | */ | |
6789 | if (output_event->cpu != event->cpu) | |
6790 | goto out; | |
6791 | ||
6792 | /* | |
76369139 | 6793 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6794 | */ |
6795 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6796 | goto out; | |
6797 | ||
a4be7c27 | 6798 | set: |
cdd6c482 | 6799 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6800 | /* Can't redirect output if we've got an active mmap() */ |
6801 | if (atomic_read(&event->mmap_count)) | |
6802 | goto unlock; | |
a4be7c27 | 6803 | |
9bb5d40c PZ |
6804 | old_rb = event->rb; |
6805 | ||
ac9721f3 | 6806 | if (output_event) { |
76369139 FW |
6807 | /* get the rb we want to redirect to */ |
6808 | rb = ring_buffer_get(output_event); | |
6809 | if (!rb) | |
ac9721f3 | 6810 | goto unlock; |
a4be7c27 PZ |
6811 | } |
6812 | ||
10c6db11 PZ |
6813 | if (old_rb) |
6814 | ring_buffer_detach(event, old_rb); | |
9bb5d40c PZ |
6815 | |
6816 | if (rb) | |
6817 | ring_buffer_attach(event, rb); | |
6818 | ||
6819 | rcu_assign_pointer(event->rb, rb); | |
6820 | ||
6821 | if (old_rb) { | |
6822 | ring_buffer_put(old_rb); | |
6823 | /* | |
6824 | * Since we detached before setting the new rb, so that we | |
6825 | * could attach the new rb, we could have missed a wakeup. | |
6826 | * Provide it now. | |
6827 | */ | |
6828 | wake_up_all(&event->waitq); | |
6829 | } | |
6830 | ||
a4be7c27 | 6831 | ret = 0; |
ac9721f3 PZ |
6832 | unlock: |
6833 | mutex_unlock(&event->mmap_mutex); | |
6834 | ||
a4be7c27 | 6835 | out: |
a4be7c27 PZ |
6836 | return ret; |
6837 | } | |
6838 | ||
0793a61d | 6839 | /** |
cdd6c482 | 6840 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6841 | * |
cdd6c482 | 6842 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6843 | * @pid: target pid |
9f66a381 | 6844 | * @cpu: target cpu |
cdd6c482 | 6845 | * @group_fd: group leader event fd |
0793a61d | 6846 | */ |
cdd6c482 IM |
6847 | SYSCALL_DEFINE5(perf_event_open, |
6848 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 6849 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 6850 | { |
b04243ef PZ |
6851 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6852 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
6853 | struct perf_event_attr attr; |
6854 | struct perf_event_context *ctx; | |
6855 | struct file *event_file = NULL; | |
2903ff01 | 6856 | struct fd group = {NULL, 0}; |
38a81da2 | 6857 | struct task_struct *task = NULL; |
89a1e187 | 6858 | struct pmu *pmu; |
ea635c64 | 6859 | int event_fd; |
b04243ef | 6860 | int move_group = 0; |
dc86cabe | 6861 | int err; |
0793a61d | 6862 | |
2743a5b0 | 6863 | /* for future expandability... */ |
e5d1367f | 6864 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
6865 | return -EINVAL; |
6866 | ||
dc86cabe IM |
6867 | err = perf_copy_attr(attr_uptr, &attr); |
6868 | if (err) | |
6869 | return err; | |
eab656ae | 6870 | |
0764771d PZ |
6871 | if (!attr.exclude_kernel) { |
6872 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6873 | return -EACCES; | |
6874 | } | |
6875 | ||
df58ab24 | 6876 | if (attr.freq) { |
cdd6c482 | 6877 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
6878 | return -EINVAL; |
6879 | } | |
6880 | ||
e5d1367f SE |
6881 | /* |
6882 | * In cgroup mode, the pid argument is used to pass the fd | |
6883 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6884 | * designates the cpu on which to monitor threads from that | |
6885 | * cgroup. | |
6886 | */ | |
6887 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6888 | return -EINVAL; | |
6889 | ||
ab72a702 | 6890 | event_fd = get_unused_fd(); |
ea635c64 AV |
6891 | if (event_fd < 0) |
6892 | return event_fd; | |
6893 | ||
ac9721f3 | 6894 | if (group_fd != -1) { |
2903ff01 AV |
6895 | err = perf_fget_light(group_fd, &group); |
6896 | if (err) | |
d14b12d7 | 6897 | goto err_fd; |
2903ff01 | 6898 | group_leader = group.file->private_data; |
ac9721f3 PZ |
6899 | if (flags & PERF_FLAG_FD_OUTPUT) |
6900 | output_event = group_leader; | |
6901 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6902 | group_leader = NULL; | |
6903 | } | |
6904 | ||
e5d1367f | 6905 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6906 | task = find_lively_task_by_vpid(pid); |
6907 | if (IS_ERR(task)) { | |
6908 | err = PTR_ERR(task); | |
6909 | goto err_group_fd; | |
6910 | } | |
6911 | } | |
6912 | ||
fbfc623f YZ |
6913 | get_online_cpus(); |
6914 | ||
4dc0da86 AK |
6915 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6916 | NULL, NULL); | |
d14b12d7 SE |
6917 | if (IS_ERR(event)) { |
6918 | err = PTR_ERR(event); | |
c6be5a5c | 6919 | goto err_task; |
d14b12d7 SE |
6920 | } |
6921 | ||
e5d1367f SE |
6922 | if (flags & PERF_FLAG_PID_CGROUP) { |
6923 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
766d6c07 FW |
6924 | if (err) { |
6925 | __free_event(event); | |
6926 | goto err_task; | |
6927 | } | |
e5d1367f SE |
6928 | } |
6929 | ||
766d6c07 FW |
6930 | account_event(event); |
6931 | ||
89a1e187 PZ |
6932 | /* |
6933 | * Special case software events and allow them to be part of | |
6934 | * any hardware group. | |
6935 | */ | |
6936 | pmu = event->pmu; | |
b04243ef PZ |
6937 | |
6938 | if (group_leader && | |
6939 | (is_software_event(event) != is_software_event(group_leader))) { | |
6940 | if (is_software_event(event)) { | |
6941 | /* | |
6942 | * If event and group_leader are not both a software | |
6943 | * event, and event is, then group leader is not. | |
6944 | * | |
6945 | * Allow the addition of software events to !software | |
6946 | * groups, this is safe because software events never | |
6947 | * fail to schedule. | |
6948 | */ | |
6949 | pmu = group_leader->pmu; | |
6950 | } else if (is_software_event(group_leader) && | |
6951 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
6952 | /* | |
6953 | * In case the group is a pure software group, and we | |
6954 | * try to add a hardware event, move the whole group to | |
6955 | * the hardware context. | |
6956 | */ | |
6957 | move_group = 1; | |
6958 | } | |
6959 | } | |
89a1e187 PZ |
6960 | |
6961 | /* | |
6962 | * Get the target context (task or percpu): | |
6963 | */ | |
e2d37cd2 | 6964 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
6965 | if (IS_ERR(ctx)) { |
6966 | err = PTR_ERR(ctx); | |
c6be5a5c | 6967 | goto err_alloc; |
89a1e187 PZ |
6968 | } |
6969 | ||
fd1edb3a PZ |
6970 | if (task) { |
6971 | put_task_struct(task); | |
6972 | task = NULL; | |
6973 | } | |
6974 | ||
ccff286d | 6975 | /* |
cdd6c482 | 6976 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 6977 | */ |
ac9721f3 | 6978 | if (group_leader) { |
dc86cabe | 6979 | err = -EINVAL; |
04289bb9 | 6980 | |
04289bb9 | 6981 | /* |
ccff286d IM |
6982 | * Do not allow a recursive hierarchy (this new sibling |
6983 | * becoming part of another group-sibling): | |
6984 | */ | |
6985 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 6986 | goto err_context; |
ccff286d IM |
6987 | /* |
6988 | * Do not allow to attach to a group in a different | |
6989 | * task or CPU context: | |
04289bb9 | 6990 | */ |
b04243ef PZ |
6991 | if (move_group) { |
6992 | if (group_leader->ctx->type != ctx->type) | |
6993 | goto err_context; | |
6994 | } else { | |
6995 | if (group_leader->ctx != ctx) | |
6996 | goto err_context; | |
6997 | } | |
6998 | ||
3b6f9e5c PM |
6999 | /* |
7000 | * Only a group leader can be exclusive or pinned | |
7001 | */ | |
0d48696f | 7002 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 7003 | goto err_context; |
ac9721f3 PZ |
7004 | } |
7005 | ||
7006 | if (output_event) { | |
7007 | err = perf_event_set_output(event, output_event); | |
7008 | if (err) | |
c3f00c70 | 7009 | goto err_context; |
ac9721f3 | 7010 | } |
0793a61d | 7011 | |
ea635c64 AV |
7012 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
7013 | if (IS_ERR(event_file)) { | |
7014 | err = PTR_ERR(event_file); | |
c3f00c70 | 7015 | goto err_context; |
ea635c64 | 7016 | } |
9b51f66d | 7017 | |
b04243ef PZ |
7018 | if (move_group) { |
7019 | struct perf_event_context *gctx = group_leader->ctx; | |
7020 | ||
7021 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 7022 | perf_remove_from_context(group_leader); |
0231bb53 JO |
7023 | |
7024 | /* | |
7025 | * Removing from the context ends up with disabled | |
7026 | * event. What we want here is event in the initial | |
7027 | * startup state, ready to be add into new context. | |
7028 | */ | |
7029 | perf_event__state_init(group_leader); | |
b04243ef PZ |
7030 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7031 | group_entry) { | |
fe4b04fa | 7032 | perf_remove_from_context(sibling); |
0231bb53 | 7033 | perf_event__state_init(sibling); |
b04243ef PZ |
7034 | put_ctx(gctx); |
7035 | } | |
7036 | mutex_unlock(&gctx->mutex); | |
7037 | put_ctx(gctx); | |
ea635c64 | 7038 | } |
9b51f66d | 7039 | |
ad3a37de | 7040 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 7041 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
7042 | |
7043 | if (move_group) { | |
0cda4c02 | 7044 | synchronize_rcu(); |
e2d37cd2 | 7045 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
7046 | get_ctx(ctx); |
7047 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
7048 | group_entry) { | |
e2d37cd2 | 7049 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
7050 | get_ctx(ctx); |
7051 | } | |
7052 | } | |
7053 | ||
e2d37cd2 | 7054 | perf_install_in_context(ctx, event, event->cpu); |
ad3a37de | 7055 | ++ctx->generation; |
fe4b04fa | 7056 | perf_unpin_context(ctx); |
d859e29f | 7057 | mutex_unlock(&ctx->mutex); |
9b51f66d | 7058 | |
fbfc623f YZ |
7059 | put_online_cpus(); |
7060 | ||
cdd6c482 | 7061 | event->owner = current; |
8882135b | 7062 | |
cdd6c482 IM |
7063 | mutex_lock(¤t->perf_event_mutex); |
7064 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
7065 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 7066 | |
c320c7b7 ACM |
7067 | /* |
7068 | * Precalculate sample_data sizes | |
7069 | */ | |
7070 | perf_event__header_size(event); | |
6844c09d | 7071 | perf_event__id_header_size(event); |
c320c7b7 | 7072 | |
8a49542c PZ |
7073 | /* |
7074 | * Drop the reference on the group_event after placing the | |
7075 | * new event on the sibling_list. This ensures destruction | |
7076 | * of the group leader will find the pointer to itself in | |
7077 | * perf_group_detach(). | |
7078 | */ | |
2903ff01 | 7079 | fdput(group); |
ea635c64 AV |
7080 | fd_install(event_fd, event_file); |
7081 | return event_fd; | |
0793a61d | 7082 | |
c3f00c70 | 7083 | err_context: |
fe4b04fa | 7084 | perf_unpin_context(ctx); |
ea635c64 | 7085 | put_ctx(ctx); |
c6be5a5c | 7086 | err_alloc: |
ea635c64 | 7087 | free_event(event); |
e7d0bc04 | 7088 | err_task: |
fbfc623f | 7089 | put_online_cpus(); |
e7d0bc04 PZ |
7090 | if (task) |
7091 | put_task_struct(task); | |
89a1e187 | 7092 | err_group_fd: |
2903ff01 | 7093 | fdput(group); |
ea635c64 AV |
7094 | err_fd: |
7095 | put_unused_fd(event_fd); | |
dc86cabe | 7096 | return err; |
0793a61d TG |
7097 | } |
7098 | ||
fb0459d7 AV |
7099 | /** |
7100 | * perf_event_create_kernel_counter | |
7101 | * | |
7102 | * @attr: attributes of the counter to create | |
7103 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 7104 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
7105 | */ |
7106 | struct perf_event * | |
7107 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 7108 | struct task_struct *task, |
4dc0da86 AK |
7109 | perf_overflow_handler_t overflow_handler, |
7110 | void *context) | |
fb0459d7 | 7111 | { |
fb0459d7 | 7112 | struct perf_event_context *ctx; |
c3f00c70 | 7113 | struct perf_event *event; |
fb0459d7 | 7114 | int err; |
d859e29f | 7115 | |
fb0459d7 AV |
7116 | /* |
7117 | * Get the target context (task or percpu): | |
7118 | */ | |
d859e29f | 7119 | |
4dc0da86 AK |
7120 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
7121 | overflow_handler, context); | |
c3f00c70 PZ |
7122 | if (IS_ERR(event)) { |
7123 | err = PTR_ERR(event); | |
7124 | goto err; | |
7125 | } | |
d859e29f | 7126 | |
766d6c07 FW |
7127 | account_event(event); |
7128 | ||
38a81da2 | 7129 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
7130 | if (IS_ERR(ctx)) { |
7131 | err = PTR_ERR(ctx); | |
c3f00c70 | 7132 | goto err_free; |
d859e29f | 7133 | } |
fb0459d7 | 7134 | |
fb0459d7 AV |
7135 | WARN_ON_ONCE(ctx->parent_ctx); |
7136 | mutex_lock(&ctx->mutex); | |
7137 | perf_install_in_context(ctx, event, cpu); | |
7138 | ++ctx->generation; | |
fe4b04fa | 7139 | perf_unpin_context(ctx); |
fb0459d7 AV |
7140 | mutex_unlock(&ctx->mutex); |
7141 | ||
fb0459d7 AV |
7142 | return event; |
7143 | ||
c3f00c70 PZ |
7144 | err_free: |
7145 | free_event(event); | |
7146 | err: | |
c6567f64 | 7147 | return ERR_PTR(err); |
9b51f66d | 7148 | } |
fb0459d7 | 7149 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7150 | |
0cda4c02 YZ |
7151 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7152 | { | |
7153 | struct perf_event_context *src_ctx; | |
7154 | struct perf_event_context *dst_ctx; | |
7155 | struct perf_event *event, *tmp; | |
7156 | LIST_HEAD(events); | |
7157 | ||
7158 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7159 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7160 | ||
7161 | mutex_lock(&src_ctx->mutex); | |
7162 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
7163 | event_entry) { | |
7164 | perf_remove_from_context(event); | |
9a545de0 | 7165 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 YZ |
7166 | put_ctx(src_ctx); |
7167 | list_add(&event->event_entry, &events); | |
7168 | } | |
7169 | mutex_unlock(&src_ctx->mutex); | |
7170 | ||
7171 | synchronize_rcu(); | |
7172 | ||
7173 | mutex_lock(&dst_ctx->mutex); | |
7174 | list_for_each_entry_safe(event, tmp, &events, event_entry) { | |
7175 | list_del(&event->event_entry); | |
7176 | if (event->state >= PERF_EVENT_STATE_OFF) | |
7177 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 7178 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
7179 | perf_install_in_context(dst_ctx, event, dst_cpu); |
7180 | get_ctx(dst_ctx); | |
7181 | } | |
7182 | mutex_unlock(&dst_ctx->mutex); | |
7183 | } | |
7184 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7185 | ||
cdd6c482 | 7186 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7187 | struct task_struct *child) |
d859e29f | 7188 | { |
cdd6c482 | 7189 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7190 | u64 child_val; |
d859e29f | 7191 | |
cdd6c482 IM |
7192 | if (child_event->attr.inherit_stat) |
7193 | perf_event_read_event(child_event, child); | |
38b200d6 | 7194 | |
b5e58793 | 7195 | child_val = perf_event_count(child_event); |
d859e29f PM |
7196 | |
7197 | /* | |
7198 | * Add back the child's count to the parent's count: | |
7199 | */ | |
a6e6dea6 | 7200 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7201 | atomic64_add(child_event->total_time_enabled, |
7202 | &parent_event->child_total_time_enabled); | |
7203 | atomic64_add(child_event->total_time_running, | |
7204 | &parent_event->child_total_time_running); | |
d859e29f PM |
7205 | |
7206 | /* | |
cdd6c482 | 7207 | * Remove this event from the parent's list |
d859e29f | 7208 | */ |
cdd6c482 IM |
7209 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7210 | mutex_lock(&parent_event->child_mutex); | |
7211 | list_del_init(&child_event->child_list); | |
7212 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
7213 | |
7214 | /* | |
cdd6c482 | 7215 | * Release the parent event, if this was the last |
d859e29f PM |
7216 | * reference to it. |
7217 | */ | |
a6fa941d | 7218 | put_event(parent_event); |
d859e29f PM |
7219 | } |
7220 | ||
9b51f66d | 7221 | static void |
cdd6c482 IM |
7222 | __perf_event_exit_task(struct perf_event *child_event, |
7223 | struct perf_event_context *child_ctx, | |
38b200d6 | 7224 | struct task_struct *child) |
9b51f66d | 7225 | { |
38b435b1 PZ |
7226 | if (child_event->parent) { |
7227 | raw_spin_lock_irq(&child_ctx->lock); | |
7228 | perf_group_detach(child_event); | |
7229 | raw_spin_unlock_irq(&child_ctx->lock); | |
7230 | } | |
9b51f66d | 7231 | |
fe4b04fa | 7232 | perf_remove_from_context(child_event); |
0cc0c027 | 7233 | |
9b51f66d | 7234 | /* |
38b435b1 | 7235 | * It can happen that the parent exits first, and has events |
9b51f66d | 7236 | * that are still around due to the child reference. These |
38b435b1 | 7237 | * events need to be zapped. |
9b51f66d | 7238 | */ |
38b435b1 | 7239 | if (child_event->parent) { |
cdd6c482 IM |
7240 | sync_child_event(child_event, child); |
7241 | free_event(child_event); | |
4bcf349a | 7242 | } |
9b51f66d IM |
7243 | } |
7244 | ||
8dc85d54 | 7245 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7246 | { |
cdd6c482 IM |
7247 | struct perf_event *child_event, *tmp; |
7248 | struct perf_event_context *child_ctx; | |
a63eaf34 | 7249 | unsigned long flags; |
9b51f66d | 7250 | |
8dc85d54 | 7251 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7252 | perf_event_task(child, NULL, 0); |
9b51f66d | 7253 | return; |
9f498cc5 | 7254 | } |
9b51f66d | 7255 | |
a63eaf34 | 7256 | local_irq_save(flags); |
ad3a37de PM |
7257 | /* |
7258 | * We can't reschedule here because interrupts are disabled, | |
7259 | * and either child is current or it is a task that can't be | |
7260 | * scheduled, so we are now safe from rescheduling changing | |
7261 | * our context. | |
7262 | */ | |
806839b2 | 7263 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7264 | |
7265 | /* | |
7266 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7267 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7268 | * incremented the context's refcount before we do put_ctx below. |
7269 | */ | |
e625cce1 | 7270 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7271 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7272 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
7273 | /* |
7274 | * If this context is a clone; unclone it so it can't get | |
7275 | * swapped to another process while we're removing all | |
cdd6c482 | 7276 | * the events from it. |
71a851b4 PZ |
7277 | */ |
7278 | unclone_ctx(child_ctx); | |
5e942bb3 | 7279 | update_context_time(child_ctx); |
e625cce1 | 7280 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
7281 | |
7282 | /* | |
cdd6c482 IM |
7283 | * Report the task dead after unscheduling the events so that we |
7284 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7285 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7286 | */ |
cdd6c482 | 7287 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 7288 | |
66fff224 PZ |
7289 | /* |
7290 | * We can recurse on the same lock type through: | |
7291 | * | |
cdd6c482 IM |
7292 | * __perf_event_exit_task() |
7293 | * sync_child_event() | |
a6fa941d AV |
7294 | * put_event() |
7295 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
7296 | * |
7297 | * But since its the parent context it won't be the same instance. | |
7298 | */ | |
a0507c84 | 7299 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 7300 | |
8bc20959 | 7301 | again: |
889ff015 FW |
7302 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
7303 | group_entry) | |
7304 | __perf_event_exit_task(child_event, child_ctx, child); | |
7305 | ||
7306 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 7307 | group_entry) |
cdd6c482 | 7308 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
7309 | |
7310 | /* | |
cdd6c482 | 7311 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
7312 | * its siblings to the list, but we obtained 'tmp' before that which |
7313 | * will still point to the list head terminating the iteration. | |
7314 | */ | |
889ff015 FW |
7315 | if (!list_empty(&child_ctx->pinned_groups) || |
7316 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 7317 | goto again; |
a63eaf34 PM |
7318 | |
7319 | mutex_unlock(&child_ctx->mutex); | |
7320 | ||
7321 | put_ctx(child_ctx); | |
9b51f66d IM |
7322 | } |
7323 | ||
8dc85d54 PZ |
7324 | /* |
7325 | * When a child task exits, feed back event values to parent events. | |
7326 | */ | |
7327 | void perf_event_exit_task(struct task_struct *child) | |
7328 | { | |
8882135b | 7329 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
7330 | int ctxn; |
7331 | ||
8882135b PZ |
7332 | mutex_lock(&child->perf_event_mutex); |
7333 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
7334 | owner_entry) { | |
7335 | list_del_init(&event->owner_entry); | |
7336 | ||
7337 | /* | |
7338 | * Ensure the list deletion is visible before we clear | |
7339 | * the owner, closes a race against perf_release() where | |
7340 | * we need to serialize on the owner->perf_event_mutex. | |
7341 | */ | |
7342 | smp_wmb(); | |
7343 | event->owner = NULL; | |
7344 | } | |
7345 | mutex_unlock(&child->perf_event_mutex); | |
7346 | ||
8dc85d54 PZ |
7347 | for_each_task_context_nr(ctxn) |
7348 | perf_event_exit_task_context(child, ctxn); | |
7349 | } | |
7350 | ||
889ff015 FW |
7351 | static void perf_free_event(struct perf_event *event, |
7352 | struct perf_event_context *ctx) | |
7353 | { | |
7354 | struct perf_event *parent = event->parent; | |
7355 | ||
7356 | if (WARN_ON_ONCE(!parent)) | |
7357 | return; | |
7358 | ||
7359 | mutex_lock(&parent->child_mutex); | |
7360 | list_del_init(&event->child_list); | |
7361 | mutex_unlock(&parent->child_mutex); | |
7362 | ||
a6fa941d | 7363 | put_event(parent); |
889ff015 | 7364 | |
8a49542c | 7365 | perf_group_detach(event); |
889ff015 FW |
7366 | list_del_event(event, ctx); |
7367 | free_event(event); | |
7368 | } | |
7369 | ||
bbbee908 PZ |
7370 | /* |
7371 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 7372 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 7373 | */ |
cdd6c482 | 7374 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 7375 | { |
8dc85d54 | 7376 | struct perf_event_context *ctx; |
cdd6c482 | 7377 | struct perf_event *event, *tmp; |
8dc85d54 | 7378 | int ctxn; |
bbbee908 | 7379 | |
8dc85d54 PZ |
7380 | for_each_task_context_nr(ctxn) { |
7381 | ctx = task->perf_event_ctxp[ctxn]; | |
7382 | if (!ctx) | |
7383 | continue; | |
bbbee908 | 7384 | |
8dc85d54 | 7385 | mutex_lock(&ctx->mutex); |
bbbee908 | 7386 | again: |
8dc85d54 PZ |
7387 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
7388 | group_entry) | |
7389 | perf_free_event(event, ctx); | |
bbbee908 | 7390 | |
8dc85d54 PZ |
7391 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
7392 | group_entry) | |
7393 | perf_free_event(event, ctx); | |
bbbee908 | 7394 | |
8dc85d54 PZ |
7395 | if (!list_empty(&ctx->pinned_groups) || |
7396 | !list_empty(&ctx->flexible_groups)) | |
7397 | goto again; | |
bbbee908 | 7398 | |
8dc85d54 | 7399 | mutex_unlock(&ctx->mutex); |
bbbee908 | 7400 | |
8dc85d54 PZ |
7401 | put_ctx(ctx); |
7402 | } | |
889ff015 FW |
7403 | } |
7404 | ||
4e231c79 PZ |
7405 | void perf_event_delayed_put(struct task_struct *task) |
7406 | { | |
7407 | int ctxn; | |
7408 | ||
7409 | for_each_task_context_nr(ctxn) | |
7410 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7411 | } | |
7412 | ||
97dee4f3 PZ |
7413 | /* |
7414 | * inherit a event from parent task to child task: | |
7415 | */ | |
7416 | static struct perf_event * | |
7417 | inherit_event(struct perf_event *parent_event, | |
7418 | struct task_struct *parent, | |
7419 | struct perf_event_context *parent_ctx, | |
7420 | struct task_struct *child, | |
7421 | struct perf_event *group_leader, | |
7422 | struct perf_event_context *child_ctx) | |
7423 | { | |
7424 | struct perf_event *child_event; | |
cee010ec | 7425 | unsigned long flags; |
97dee4f3 PZ |
7426 | |
7427 | /* | |
7428 | * Instead of creating recursive hierarchies of events, | |
7429 | * we link inherited events back to the original parent, | |
7430 | * which has a filp for sure, which we use as the reference | |
7431 | * count: | |
7432 | */ | |
7433 | if (parent_event->parent) | |
7434 | parent_event = parent_event->parent; | |
7435 | ||
7436 | child_event = perf_event_alloc(&parent_event->attr, | |
7437 | parent_event->cpu, | |
d580ff86 | 7438 | child, |
97dee4f3 | 7439 | group_leader, parent_event, |
4dc0da86 | 7440 | NULL, NULL); |
97dee4f3 PZ |
7441 | if (IS_ERR(child_event)) |
7442 | return child_event; | |
a6fa941d AV |
7443 | |
7444 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7445 | free_event(child_event); | |
7446 | return NULL; | |
7447 | } | |
7448 | ||
97dee4f3 PZ |
7449 | get_ctx(child_ctx); |
7450 | ||
7451 | /* | |
7452 | * Make the child state follow the state of the parent event, | |
7453 | * not its attr.disabled bit. We hold the parent's mutex, | |
7454 | * so we won't race with perf_event_{en, dis}able_family. | |
7455 | */ | |
7456 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7457 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7458 | else | |
7459 | child_event->state = PERF_EVENT_STATE_OFF; | |
7460 | ||
7461 | if (parent_event->attr.freq) { | |
7462 | u64 sample_period = parent_event->hw.sample_period; | |
7463 | struct hw_perf_event *hwc = &child_event->hw; | |
7464 | ||
7465 | hwc->sample_period = sample_period; | |
7466 | hwc->last_period = sample_period; | |
7467 | ||
7468 | local64_set(&hwc->period_left, sample_period); | |
7469 | } | |
7470 | ||
7471 | child_event->ctx = child_ctx; | |
7472 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7473 | child_event->overflow_handler_context |
7474 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7475 | |
614b6780 TG |
7476 | /* |
7477 | * Precalculate sample_data sizes | |
7478 | */ | |
7479 | perf_event__header_size(child_event); | |
6844c09d | 7480 | perf_event__id_header_size(child_event); |
614b6780 | 7481 | |
97dee4f3 PZ |
7482 | /* |
7483 | * Link it up in the child's context: | |
7484 | */ | |
cee010ec | 7485 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7486 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7487 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7488 | |
97dee4f3 PZ |
7489 | /* |
7490 | * Link this into the parent event's child list | |
7491 | */ | |
7492 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7493 | mutex_lock(&parent_event->child_mutex); | |
7494 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7495 | mutex_unlock(&parent_event->child_mutex); | |
7496 | ||
7497 | return child_event; | |
7498 | } | |
7499 | ||
7500 | static int inherit_group(struct perf_event *parent_event, | |
7501 | struct task_struct *parent, | |
7502 | struct perf_event_context *parent_ctx, | |
7503 | struct task_struct *child, | |
7504 | struct perf_event_context *child_ctx) | |
7505 | { | |
7506 | struct perf_event *leader; | |
7507 | struct perf_event *sub; | |
7508 | struct perf_event *child_ctr; | |
7509 | ||
7510 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7511 | child, NULL, child_ctx); | |
7512 | if (IS_ERR(leader)) | |
7513 | return PTR_ERR(leader); | |
7514 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7515 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7516 | child, leader, child_ctx); | |
7517 | if (IS_ERR(child_ctr)) | |
7518 | return PTR_ERR(child_ctr); | |
7519 | } | |
7520 | return 0; | |
889ff015 FW |
7521 | } |
7522 | ||
7523 | static int | |
7524 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7525 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7526 | struct task_struct *child, int ctxn, |
889ff015 FW |
7527 | int *inherited_all) |
7528 | { | |
7529 | int ret; | |
8dc85d54 | 7530 | struct perf_event_context *child_ctx; |
889ff015 FW |
7531 | |
7532 | if (!event->attr.inherit) { | |
7533 | *inherited_all = 0; | |
7534 | return 0; | |
bbbee908 PZ |
7535 | } |
7536 | ||
fe4b04fa | 7537 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7538 | if (!child_ctx) { |
7539 | /* | |
7540 | * This is executed from the parent task context, so | |
7541 | * inherit events that have been marked for cloning. | |
7542 | * First allocate and initialize a context for the | |
7543 | * child. | |
7544 | */ | |
bbbee908 | 7545 | |
734df5ab | 7546 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
7547 | if (!child_ctx) |
7548 | return -ENOMEM; | |
bbbee908 | 7549 | |
8dc85d54 | 7550 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7551 | } |
7552 | ||
7553 | ret = inherit_group(event, parent, parent_ctx, | |
7554 | child, child_ctx); | |
7555 | ||
7556 | if (ret) | |
7557 | *inherited_all = 0; | |
7558 | ||
7559 | return ret; | |
bbbee908 PZ |
7560 | } |
7561 | ||
9b51f66d | 7562 | /* |
cdd6c482 | 7563 | * Initialize the perf_event context in task_struct |
9b51f66d | 7564 | */ |
8dc85d54 | 7565 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7566 | { |
889ff015 | 7567 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7568 | struct perf_event_context *cloned_ctx; |
7569 | struct perf_event *event; | |
9b51f66d | 7570 | struct task_struct *parent = current; |
564c2b21 | 7571 | int inherited_all = 1; |
dddd3379 | 7572 | unsigned long flags; |
6ab423e0 | 7573 | int ret = 0; |
9b51f66d | 7574 | |
8dc85d54 | 7575 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7576 | return 0; |
7577 | ||
ad3a37de | 7578 | /* |
25346b93 PM |
7579 | * If the parent's context is a clone, pin it so it won't get |
7580 | * swapped under us. | |
ad3a37de | 7581 | */ |
8dc85d54 | 7582 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 7583 | |
ad3a37de PM |
7584 | /* |
7585 | * No need to check if parent_ctx != NULL here; since we saw | |
7586 | * it non-NULL earlier, the only reason for it to become NULL | |
7587 | * is if we exit, and since we're currently in the middle of | |
7588 | * a fork we can't be exiting at the same time. | |
7589 | */ | |
ad3a37de | 7590 | |
9b51f66d IM |
7591 | /* |
7592 | * Lock the parent list. No need to lock the child - not PID | |
7593 | * hashed yet and not running, so nobody can access it. | |
7594 | */ | |
d859e29f | 7595 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7596 | |
7597 | /* | |
7598 | * We dont have to disable NMIs - we are only looking at | |
7599 | * the list, not manipulating it: | |
7600 | */ | |
889ff015 | 7601 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7602 | ret = inherit_task_group(event, parent, parent_ctx, |
7603 | child, ctxn, &inherited_all); | |
889ff015 FW |
7604 | if (ret) |
7605 | break; | |
7606 | } | |
b93f7978 | 7607 | |
dddd3379 TG |
7608 | /* |
7609 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7610 | * to allocations, but we need to prevent rotation because | |
7611 | * rotate_ctx() will change the list from interrupt context. | |
7612 | */ | |
7613 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7614 | parent_ctx->rotate_disable = 1; | |
7615 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7616 | ||
889ff015 | 7617 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7618 | ret = inherit_task_group(event, parent, parent_ctx, |
7619 | child, ctxn, &inherited_all); | |
889ff015 | 7620 | if (ret) |
9b51f66d | 7621 | break; |
564c2b21 PM |
7622 | } |
7623 | ||
dddd3379 TG |
7624 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7625 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7626 | |
8dc85d54 | 7627 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7628 | |
05cbaa28 | 7629 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7630 | /* |
7631 | * Mark the child context as a clone of the parent | |
7632 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7633 | * |
7634 | * Note that if the parent is a clone, the holding of | |
7635 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7636 | */ |
c5ed5145 | 7637 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7638 | if (cloned_ctx) { |
7639 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7640 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7641 | } else { |
7642 | child_ctx->parent_ctx = parent_ctx; | |
7643 | child_ctx->parent_gen = parent_ctx->generation; | |
7644 | } | |
7645 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7646 | } |
7647 | ||
c5ed5145 | 7648 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7649 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7650 | |
25346b93 | 7651 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7652 | put_ctx(parent_ctx); |
ad3a37de | 7653 | |
6ab423e0 | 7654 | return ret; |
9b51f66d IM |
7655 | } |
7656 | ||
8dc85d54 PZ |
7657 | /* |
7658 | * Initialize the perf_event context in task_struct | |
7659 | */ | |
7660 | int perf_event_init_task(struct task_struct *child) | |
7661 | { | |
7662 | int ctxn, ret; | |
7663 | ||
8550d7cb ON |
7664 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7665 | mutex_init(&child->perf_event_mutex); | |
7666 | INIT_LIST_HEAD(&child->perf_event_list); | |
7667 | ||
8dc85d54 PZ |
7668 | for_each_task_context_nr(ctxn) { |
7669 | ret = perf_event_init_context(child, ctxn); | |
7670 | if (ret) | |
7671 | return ret; | |
7672 | } | |
7673 | ||
7674 | return 0; | |
7675 | } | |
7676 | ||
220b140b PM |
7677 | static void __init perf_event_init_all_cpus(void) |
7678 | { | |
b28ab83c | 7679 | struct swevent_htable *swhash; |
220b140b | 7680 | int cpu; |
220b140b PM |
7681 | |
7682 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7683 | swhash = &per_cpu(swevent_htable, cpu); |
7684 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7685 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7686 | } |
7687 | } | |
7688 | ||
0db0628d | 7689 | static void perf_event_init_cpu(int cpu) |
0793a61d | 7690 | { |
108b02cf | 7691 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7692 | |
b28ab83c | 7693 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7694 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7695 | struct swevent_hlist *hlist; |
7696 | ||
b28ab83c PZ |
7697 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7698 | WARN_ON(!hlist); | |
7699 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7700 | } |
b28ab83c | 7701 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7702 | } |
7703 | ||
c277443c | 7704 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7705 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7706 | { |
e9d2b064 PZ |
7707 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7708 | ||
7709 | WARN_ON(!irqs_disabled()); | |
7710 | ||
7711 | list_del_init(&cpuctx->rotation_list); | |
7712 | } | |
7713 | ||
108b02cf | 7714 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7715 | { |
108b02cf | 7716 | struct perf_event_context *ctx = __info; |
cdd6c482 | 7717 | struct perf_event *event, *tmp; |
0793a61d | 7718 | |
108b02cf | 7719 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7720 | |
889ff015 | 7721 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 7722 | __perf_remove_from_context(event); |
889ff015 | 7723 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 7724 | __perf_remove_from_context(event); |
0793a61d | 7725 | } |
108b02cf PZ |
7726 | |
7727 | static void perf_event_exit_cpu_context(int cpu) | |
7728 | { | |
7729 | struct perf_event_context *ctx; | |
7730 | struct pmu *pmu; | |
7731 | int idx; | |
7732 | ||
7733 | idx = srcu_read_lock(&pmus_srcu); | |
7734 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7735 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7736 | |
7737 | mutex_lock(&ctx->mutex); | |
7738 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7739 | mutex_unlock(&ctx->mutex); | |
7740 | } | |
7741 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7742 | } |
7743 | ||
cdd6c482 | 7744 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7745 | { |
b28ab83c | 7746 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7747 | |
b28ab83c PZ |
7748 | mutex_lock(&swhash->hlist_mutex); |
7749 | swevent_hlist_release(swhash); | |
7750 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 7751 | |
108b02cf | 7752 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7753 | } |
7754 | #else | |
cdd6c482 | 7755 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7756 | #endif |
7757 | ||
c277443c PZ |
7758 | static int |
7759 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7760 | { | |
7761 | int cpu; | |
7762 | ||
7763 | for_each_online_cpu(cpu) | |
7764 | perf_event_exit_cpu(cpu); | |
7765 | ||
7766 | return NOTIFY_OK; | |
7767 | } | |
7768 | ||
7769 | /* | |
7770 | * Run the perf reboot notifier at the very last possible moment so that | |
7771 | * the generic watchdog code runs as long as possible. | |
7772 | */ | |
7773 | static struct notifier_block perf_reboot_notifier = { | |
7774 | .notifier_call = perf_reboot, | |
7775 | .priority = INT_MIN, | |
7776 | }; | |
7777 | ||
0db0628d | 7778 | static int |
0793a61d TG |
7779 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
7780 | { | |
7781 | unsigned int cpu = (long)hcpu; | |
7782 | ||
4536e4d1 | 7783 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7784 | |
7785 | case CPU_UP_PREPARE: | |
5e11637e | 7786 | case CPU_DOWN_FAILED: |
cdd6c482 | 7787 | perf_event_init_cpu(cpu); |
0793a61d TG |
7788 | break; |
7789 | ||
5e11637e | 7790 | case CPU_UP_CANCELED: |
0793a61d | 7791 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7792 | perf_event_exit_cpu(cpu); |
0793a61d | 7793 | break; |
0793a61d TG |
7794 | default: |
7795 | break; | |
7796 | } | |
7797 | ||
7798 | return NOTIFY_OK; | |
7799 | } | |
7800 | ||
cdd6c482 | 7801 | void __init perf_event_init(void) |
0793a61d | 7802 | { |
3c502e7a JW |
7803 | int ret; |
7804 | ||
2e80a82a PZ |
7805 | idr_init(&pmu_idr); |
7806 | ||
220b140b | 7807 | perf_event_init_all_cpus(); |
b0a873eb | 7808 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7809 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7810 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7811 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7812 | perf_tp_register(); |
7813 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7814 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7815 | |
7816 | ret = init_hw_breakpoint(); | |
7817 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7818 | |
7819 | /* do not patch jump label more than once per second */ | |
7820 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7821 | |
7822 | /* | |
7823 | * Build time assertion that we keep the data_head at the intended | |
7824 | * location. IOW, validation we got the __reserved[] size right. | |
7825 | */ | |
7826 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
7827 | != 1024); | |
0793a61d | 7828 | } |
abe43400 PZ |
7829 | |
7830 | static int __init perf_event_sysfs_init(void) | |
7831 | { | |
7832 | struct pmu *pmu; | |
7833 | int ret; | |
7834 | ||
7835 | mutex_lock(&pmus_lock); | |
7836 | ||
7837 | ret = bus_register(&pmu_bus); | |
7838 | if (ret) | |
7839 | goto unlock; | |
7840 | ||
7841 | list_for_each_entry(pmu, &pmus, entry) { | |
7842 | if (!pmu->name || pmu->type < 0) | |
7843 | continue; | |
7844 | ||
7845 | ret = pmu_dev_alloc(pmu); | |
7846 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
7847 | } | |
7848 | pmu_bus_running = 1; | |
7849 | ret = 0; | |
7850 | ||
7851 | unlock: | |
7852 | mutex_unlock(&pmus_lock); | |
7853 | ||
7854 | return ret; | |
7855 | } | |
7856 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
7857 | |
7858 | #ifdef CONFIG_CGROUP_PERF | |
92fb9748 | 7859 | static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont) |
e5d1367f SE |
7860 | { |
7861 | struct perf_cgroup *jc; | |
e5d1367f | 7862 | |
1b15d055 | 7863 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
7864 | if (!jc) |
7865 | return ERR_PTR(-ENOMEM); | |
7866 | ||
e5d1367f SE |
7867 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7868 | if (!jc->info) { | |
7869 | kfree(jc); | |
7870 | return ERR_PTR(-ENOMEM); | |
7871 | } | |
7872 | ||
e5d1367f SE |
7873 | return &jc->css; |
7874 | } | |
7875 | ||
92fb9748 | 7876 | static void perf_cgroup_css_free(struct cgroup *cont) |
e5d1367f SE |
7877 | { |
7878 | struct perf_cgroup *jc; | |
7879 | jc = container_of(cgroup_subsys_state(cont, perf_subsys_id), | |
7880 | struct perf_cgroup, css); | |
7881 | free_percpu(jc->info); | |
7882 | kfree(jc); | |
7883 | } | |
7884 | ||
7885 | static int __perf_cgroup_move(void *info) | |
7886 | { | |
7887 | struct task_struct *task = info; | |
7888 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
7889 | return 0; | |
7890 | } | |
7891 | ||
761b3ef5 | 7892 | static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
e5d1367f | 7893 | { |
bb9d97b6 TH |
7894 | struct task_struct *task; |
7895 | ||
7896 | cgroup_taskset_for_each(task, cgrp, tset) | |
7897 | task_function_call(task, __perf_cgroup_move, task); | |
e5d1367f SE |
7898 | } |
7899 | ||
761b3ef5 LZ |
7900 | static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, |
7901 | struct task_struct *task) | |
e5d1367f SE |
7902 | { |
7903 | /* | |
7904 | * cgroup_exit() is called in the copy_process() failure path. | |
7905 | * Ignore this case since the task hasn't ran yet, this avoids | |
7906 | * trying to poke a half freed task state from generic code. | |
7907 | */ | |
7908 | if (!(task->flags & PF_EXITING)) | |
7909 | return; | |
7910 | ||
bb9d97b6 | 7911 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7912 | } |
7913 | ||
7914 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
7915 | .name = "perf_event", |
7916 | .subsys_id = perf_subsys_id, | |
92fb9748 TH |
7917 | .css_alloc = perf_cgroup_css_alloc, |
7918 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 7919 | .exit = perf_cgroup_exit, |
bb9d97b6 | 7920 | .attach = perf_cgroup_attach, |
e5d1367f SE |
7921 | }; |
7922 | #endif /* CONFIG_CGROUP_PERF */ |