]>
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> |
c464c76e | 42 | #include <linux/module.h> |
f972eb63 | 43 | #include <linux/mman.h> |
0793a61d | 44 | |
76369139 FW |
45 | #include "internal.h" |
46 | ||
4e193bd4 TB |
47 | #include <asm/irq_regs.h> |
48 | ||
fe4b04fa | 49 | struct remote_function_call { |
e7e7ee2e IM |
50 | struct task_struct *p; |
51 | int (*func)(void *info); | |
52 | void *info; | |
53 | int ret; | |
fe4b04fa PZ |
54 | }; |
55 | ||
56 | static void remote_function(void *data) | |
57 | { | |
58 | struct remote_function_call *tfc = data; | |
59 | struct task_struct *p = tfc->p; | |
60 | ||
61 | if (p) { | |
62 | tfc->ret = -EAGAIN; | |
63 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
64 | return; | |
65 | } | |
66 | ||
67 | tfc->ret = tfc->func(tfc->info); | |
68 | } | |
69 | ||
70 | /** | |
71 | * task_function_call - call a function on the cpu on which a task runs | |
72 | * @p: the task to evaluate | |
73 | * @func: the function to be called | |
74 | * @info: the function call argument | |
75 | * | |
76 | * Calls the function @func when the task is currently running. This might | |
77 | * be on the current CPU, which just calls the function directly | |
78 | * | |
79 | * returns: @func return value, or | |
80 | * -ESRCH - when the process isn't running | |
81 | * -EAGAIN - when the process moved away | |
82 | */ | |
83 | static int | |
84 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
85 | { | |
86 | struct remote_function_call data = { | |
e7e7ee2e IM |
87 | .p = p, |
88 | .func = func, | |
89 | .info = info, | |
90 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
91 | }; |
92 | ||
93 | if (task_curr(p)) | |
94 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
95 | ||
96 | return data.ret; | |
97 | } | |
98 | ||
99 | /** | |
100 | * cpu_function_call - call a function on the cpu | |
101 | * @func: the function to be called | |
102 | * @info: the function call argument | |
103 | * | |
104 | * Calls the function @func on the remote cpu. | |
105 | * | |
106 | * returns: @func return value or -ENXIO when the cpu is offline | |
107 | */ | |
108 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
109 | { | |
110 | struct remote_function_call data = { | |
e7e7ee2e IM |
111 | .p = NULL, |
112 | .func = func, | |
113 | .info = info, | |
114 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
115 | }; |
116 | ||
117 | smp_call_function_single(cpu, remote_function, &data, 1); | |
118 | ||
119 | return data.ret; | |
120 | } | |
121 | ||
f8697762 JO |
122 | #define EVENT_OWNER_KERNEL ((void *) -1) |
123 | ||
124 | static bool is_kernel_event(struct perf_event *event) | |
125 | { | |
126 | return event->owner == EVENT_OWNER_KERNEL; | |
127 | } | |
128 | ||
e5d1367f SE |
129 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
130 | PERF_FLAG_FD_OUTPUT |\ | |
a21b0b35 YD |
131 | PERF_FLAG_PID_CGROUP |\ |
132 | PERF_FLAG_FD_CLOEXEC) | |
e5d1367f | 133 | |
bce38cd5 SE |
134 | /* |
135 | * branch priv levels that need permission checks | |
136 | */ | |
137 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
138 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
139 | PERF_SAMPLE_BRANCH_HV) | |
140 | ||
0b3fcf17 SE |
141 | enum event_type_t { |
142 | EVENT_FLEXIBLE = 0x1, | |
143 | EVENT_PINNED = 0x2, | |
144 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
145 | }; | |
146 | ||
e5d1367f SE |
147 | /* |
148 | * perf_sched_events : >0 events exist | |
149 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
150 | */ | |
c5905afb | 151 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 152 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 153 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 154 | |
cdd6c482 IM |
155 | static atomic_t nr_mmap_events __read_mostly; |
156 | static atomic_t nr_comm_events __read_mostly; | |
157 | static atomic_t nr_task_events __read_mostly; | |
948b26b6 | 158 | static atomic_t nr_freq_events __read_mostly; |
9ee318a7 | 159 | |
108b02cf PZ |
160 | static LIST_HEAD(pmus); |
161 | static DEFINE_MUTEX(pmus_lock); | |
162 | static struct srcu_struct pmus_srcu; | |
163 | ||
0764771d | 164 | /* |
cdd6c482 | 165 | * perf event paranoia level: |
0fbdea19 IM |
166 | * -1 - not paranoid at all |
167 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 168 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 169 | * 2 - disallow kernel profiling for unpriv |
0764771d | 170 | */ |
cdd6c482 | 171 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 172 | |
20443384 FW |
173 | /* Minimum for 512 kiB + 1 user control page */ |
174 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
175 | |
176 | /* | |
cdd6c482 | 177 | * max perf event sample rate |
df58ab24 | 178 | */ |
14c63f17 DH |
179 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
180 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
181 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
182 | ||
183 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
184 | ||
185 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
186 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
187 | ||
d9494cb4 PZ |
188 | static int perf_sample_allowed_ns __read_mostly = |
189 | DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100; | |
14c63f17 DH |
190 | |
191 | void update_perf_cpu_limits(void) | |
192 | { | |
193 | u64 tmp = perf_sample_period_ns; | |
194 | ||
195 | tmp *= sysctl_perf_cpu_time_max_percent; | |
e5302920 | 196 | do_div(tmp, 100); |
d9494cb4 | 197 | ACCESS_ONCE(perf_sample_allowed_ns) = tmp; |
14c63f17 | 198 | } |
163ec435 | 199 | |
9e630205 SE |
200 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
201 | ||
163ec435 PZ |
202 | int perf_proc_update_handler(struct ctl_table *table, int write, |
203 | void __user *buffer, size_t *lenp, | |
204 | loff_t *ppos) | |
205 | { | |
723478c8 | 206 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
163ec435 PZ |
207 | |
208 | if (ret || !write) | |
209 | return ret; | |
210 | ||
211 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
212 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
213 | update_perf_cpu_limits(); | |
214 | ||
215 | return 0; | |
216 | } | |
217 | ||
218 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
219 | ||
220 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
221 | void __user *buffer, size_t *lenp, | |
222 | loff_t *ppos) | |
223 | { | |
224 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
225 | ||
226 | if (ret || !write) | |
227 | return ret; | |
228 | ||
229 | update_perf_cpu_limits(); | |
163ec435 PZ |
230 | |
231 | return 0; | |
232 | } | |
1ccd1549 | 233 | |
14c63f17 DH |
234 | /* |
235 | * perf samples are done in some very critical code paths (NMIs). | |
236 | * If they take too much CPU time, the system can lock up and not | |
237 | * get any real work done. This will drop the sample rate when | |
238 | * we detect that events are taking too long. | |
239 | */ | |
240 | #define NR_ACCUMULATED_SAMPLES 128 | |
d9494cb4 | 241 | static DEFINE_PER_CPU(u64, running_sample_length); |
14c63f17 | 242 | |
6a02ad66 | 243 | static void perf_duration_warn(struct irq_work *w) |
14c63f17 | 244 | { |
6a02ad66 | 245 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
14c63f17 | 246 | u64 avg_local_sample_len; |
e5302920 | 247 | u64 local_samples_len; |
6a02ad66 PZ |
248 | |
249 | local_samples_len = __get_cpu_var(running_sample_length); | |
250 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
251 | ||
252 | printk_ratelimited(KERN_WARNING | |
253 | "perf interrupt took too long (%lld > %lld), lowering " | |
254 | "kernel.perf_event_max_sample_rate to %d\n", | |
cd578abb | 255 | avg_local_sample_len, allowed_ns >> 1, |
6a02ad66 PZ |
256 | sysctl_perf_event_sample_rate); |
257 | } | |
258 | ||
259 | static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn); | |
260 | ||
261 | void perf_sample_event_took(u64 sample_len_ns) | |
262 | { | |
d9494cb4 | 263 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
6a02ad66 PZ |
264 | u64 avg_local_sample_len; |
265 | u64 local_samples_len; | |
14c63f17 | 266 | |
d9494cb4 | 267 | if (allowed_ns == 0) |
14c63f17 DH |
268 | return; |
269 | ||
270 | /* decay the counter by 1 average sample */ | |
271 | local_samples_len = __get_cpu_var(running_sample_length); | |
272 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; | |
273 | local_samples_len += sample_len_ns; | |
274 | __get_cpu_var(running_sample_length) = local_samples_len; | |
275 | ||
276 | /* | |
277 | * note: this will be biased artifically low until we have | |
278 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
279 | * from having to maintain a count. | |
280 | */ | |
281 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
282 | ||
d9494cb4 | 283 | if (avg_local_sample_len <= allowed_ns) |
14c63f17 DH |
284 | return; |
285 | ||
286 | if (max_samples_per_tick <= 1) | |
287 | return; | |
288 | ||
289 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
290 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
291 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
292 | ||
14c63f17 | 293 | update_perf_cpu_limits(); |
6a02ad66 | 294 | |
cd578abb PZ |
295 | if (!irq_work_queue(&perf_duration_work)) { |
296 | early_printk("perf interrupt took too long (%lld > %lld), lowering " | |
297 | "kernel.perf_event_max_sample_rate to %d\n", | |
298 | avg_local_sample_len, allowed_ns >> 1, | |
299 | sysctl_perf_event_sample_rate); | |
300 | } | |
14c63f17 DH |
301 | } |
302 | ||
cdd6c482 | 303 | static atomic64_t perf_event_id; |
a96bbc16 | 304 | |
0b3fcf17 SE |
305 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
306 | enum event_type_t event_type); | |
307 | ||
308 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
309 | enum event_type_t event_type, |
310 | struct task_struct *task); | |
311 | ||
312 | static void update_context_time(struct perf_event_context *ctx); | |
313 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 314 | |
cdd6c482 | 315 | void __weak perf_event_print_debug(void) { } |
0793a61d | 316 | |
84c79910 | 317 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 318 | { |
84c79910 | 319 | return "pmu"; |
0793a61d TG |
320 | } |
321 | ||
0b3fcf17 SE |
322 | static inline u64 perf_clock(void) |
323 | { | |
324 | return local_clock(); | |
325 | } | |
326 | ||
e5d1367f SE |
327 | static inline struct perf_cpu_context * |
328 | __get_cpu_context(struct perf_event_context *ctx) | |
329 | { | |
330 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
331 | } | |
332 | ||
facc4307 PZ |
333 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
334 | struct perf_event_context *ctx) | |
335 | { | |
336 | raw_spin_lock(&cpuctx->ctx.lock); | |
337 | if (ctx) | |
338 | raw_spin_lock(&ctx->lock); | |
339 | } | |
340 | ||
341 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
342 | struct perf_event_context *ctx) | |
343 | { | |
344 | if (ctx) | |
345 | raw_spin_unlock(&ctx->lock); | |
346 | raw_spin_unlock(&cpuctx->ctx.lock); | |
347 | } | |
348 | ||
e5d1367f SE |
349 | #ifdef CONFIG_CGROUP_PERF |
350 | ||
877c6856 LZ |
351 | /* |
352 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
353 | * This is a per-cpu dynamically allocated data structure. | |
354 | */ | |
355 | struct perf_cgroup_info { | |
356 | u64 time; | |
357 | u64 timestamp; | |
358 | }; | |
359 | ||
360 | struct perf_cgroup { | |
361 | struct cgroup_subsys_state css; | |
86e213e1 | 362 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
363 | }; |
364 | ||
3f7cce3c SE |
365 | /* |
366 | * Must ensure cgroup is pinned (css_get) before calling | |
367 | * this function. In other words, we cannot call this function | |
368 | * if there is no cgroup event for the current CPU context. | |
369 | */ | |
e5d1367f SE |
370 | static inline struct perf_cgroup * |
371 | perf_cgroup_from_task(struct task_struct *task) | |
372 | { | |
073219e9 | 373 | return container_of(task_css(task, perf_event_cgrp_id), |
8af01f56 | 374 | struct perf_cgroup, css); |
e5d1367f SE |
375 | } |
376 | ||
377 | static inline bool | |
378 | perf_cgroup_match(struct perf_event *event) | |
379 | { | |
380 | struct perf_event_context *ctx = event->ctx; | |
381 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
382 | ||
ef824fa1 TH |
383 | /* @event doesn't care about cgroup */ |
384 | if (!event->cgrp) | |
385 | return true; | |
386 | ||
387 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
388 | if (!cpuctx->cgrp) | |
389 | return false; | |
390 | ||
391 | /* | |
392 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
393 | * also enabled for all its descendant cgroups. If @cpuctx's | |
394 | * cgroup is a descendant of @event's (the test covers identity | |
395 | * case), it's a match. | |
396 | */ | |
397 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
398 | event->cgrp->css.cgroup); | |
e5d1367f SE |
399 | } |
400 | ||
e5d1367f SE |
401 | static inline void perf_put_cgroup(struct perf_event *event) |
402 | { | |
403 | css_put(&event->cgrp->css); | |
404 | } | |
405 | ||
406 | static inline void perf_detach_cgroup(struct perf_event *event) | |
407 | { | |
408 | perf_put_cgroup(event); | |
409 | event->cgrp = NULL; | |
410 | } | |
411 | ||
412 | static inline int is_cgroup_event(struct perf_event *event) | |
413 | { | |
414 | return event->cgrp != NULL; | |
415 | } | |
416 | ||
417 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
418 | { | |
419 | struct perf_cgroup_info *t; | |
420 | ||
421 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
422 | return t->time; | |
423 | } | |
424 | ||
425 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
426 | { | |
427 | struct perf_cgroup_info *info; | |
428 | u64 now; | |
429 | ||
430 | now = perf_clock(); | |
431 | ||
432 | info = this_cpu_ptr(cgrp->info); | |
433 | ||
434 | info->time += now - info->timestamp; | |
435 | info->timestamp = now; | |
436 | } | |
437 | ||
438 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
439 | { | |
440 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
441 | if (cgrp_out) | |
442 | __update_cgrp_time(cgrp_out); | |
443 | } | |
444 | ||
445 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
446 | { | |
3f7cce3c SE |
447 | struct perf_cgroup *cgrp; |
448 | ||
e5d1367f | 449 | /* |
3f7cce3c SE |
450 | * ensure we access cgroup data only when needed and |
451 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 452 | */ |
3f7cce3c | 453 | if (!is_cgroup_event(event)) |
e5d1367f SE |
454 | return; |
455 | ||
3f7cce3c SE |
456 | cgrp = perf_cgroup_from_task(current); |
457 | /* | |
458 | * Do not update time when cgroup is not active | |
459 | */ | |
460 | if (cgrp == event->cgrp) | |
461 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
462 | } |
463 | ||
464 | static inline void | |
3f7cce3c SE |
465 | perf_cgroup_set_timestamp(struct task_struct *task, |
466 | struct perf_event_context *ctx) | |
e5d1367f SE |
467 | { |
468 | struct perf_cgroup *cgrp; | |
469 | struct perf_cgroup_info *info; | |
470 | ||
3f7cce3c SE |
471 | /* |
472 | * ctx->lock held by caller | |
473 | * ensure we do not access cgroup data | |
474 | * unless we have the cgroup pinned (css_get) | |
475 | */ | |
476 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
477 | return; |
478 | ||
479 | cgrp = perf_cgroup_from_task(task); | |
480 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 481 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
482 | } |
483 | ||
484 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
485 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
486 | ||
487 | /* | |
488 | * reschedule events based on the cgroup constraint of task. | |
489 | * | |
490 | * mode SWOUT : schedule out everything | |
491 | * mode SWIN : schedule in based on cgroup for next | |
492 | */ | |
493 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
494 | { | |
495 | struct perf_cpu_context *cpuctx; | |
496 | struct pmu *pmu; | |
497 | unsigned long flags; | |
498 | ||
499 | /* | |
500 | * disable interrupts to avoid geting nr_cgroup | |
501 | * changes via __perf_event_disable(). Also | |
502 | * avoids preemption. | |
503 | */ | |
504 | local_irq_save(flags); | |
505 | ||
506 | /* | |
507 | * we reschedule only in the presence of cgroup | |
508 | * constrained events. | |
509 | */ | |
510 | rcu_read_lock(); | |
511 | ||
512 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 513 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
514 | if (cpuctx->unique_pmu != pmu) |
515 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 516 | |
e5d1367f SE |
517 | /* |
518 | * perf_cgroup_events says at least one | |
519 | * context on this CPU has cgroup events. | |
520 | * | |
521 | * ctx->nr_cgroups reports the number of cgroup | |
522 | * events for a context. | |
523 | */ | |
524 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
525 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
526 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
527 | |
528 | if (mode & PERF_CGROUP_SWOUT) { | |
529 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
530 | /* | |
531 | * must not be done before ctxswout due | |
532 | * to event_filter_match() in event_sched_out() | |
533 | */ | |
534 | cpuctx->cgrp = NULL; | |
535 | } | |
536 | ||
537 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 538 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
539 | /* |
540 | * set cgrp before ctxsw in to allow | |
541 | * event_filter_match() to not have to pass | |
542 | * task around | |
e5d1367f SE |
543 | */ |
544 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
545 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
546 | } | |
facc4307 PZ |
547 | perf_pmu_enable(cpuctx->ctx.pmu); |
548 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 549 | } |
e5d1367f SE |
550 | } |
551 | ||
552 | rcu_read_unlock(); | |
553 | ||
554 | local_irq_restore(flags); | |
555 | } | |
556 | ||
a8d757ef SE |
557 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
558 | struct task_struct *next) | |
e5d1367f | 559 | { |
a8d757ef SE |
560 | struct perf_cgroup *cgrp1; |
561 | struct perf_cgroup *cgrp2 = NULL; | |
562 | ||
563 | /* | |
564 | * we come here when we know perf_cgroup_events > 0 | |
565 | */ | |
566 | cgrp1 = perf_cgroup_from_task(task); | |
567 | ||
568 | /* | |
569 | * next is NULL when called from perf_event_enable_on_exec() | |
570 | * that will systematically cause a cgroup_switch() | |
571 | */ | |
572 | if (next) | |
573 | cgrp2 = perf_cgroup_from_task(next); | |
574 | ||
575 | /* | |
576 | * only schedule out current cgroup events if we know | |
577 | * that we are switching to a different cgroup. Otherwise, | |
578 | * do no touch the cgroup events. | |
579 | */ | |
580 | if (cgrp1 != cgrp2) | |
581 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
582 | } |
583 | ||
a8d757ef SE |
584 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
585 | struct task_struct *task) | |
e5d1367f | 586 | { |
a8d757ef SE |
587 | struct perf_cgroup *cgrp1; |
588 | struct perf_cgroup *cgrp2 = NULL; | |
589 | ||
590 | /* | |
591 | * we come here when we know perf_cgroup_events > 0 | |
592 | */ | |
593 | cgrp1 = perf_cgroup_from_task(task); | |
594 | ||
595 | /* prev can never be NULL */ | |
596 | cgrp2 = perf_cgroup_from_task(prev); | |
597 | ||
598 | /* | |
599 | * only need to schedule in cgroup events if we are changing | |
600 | * cgroup during ctxsw. Cgroup events were not scheduled | |
601 | * out of ctxsw out if that was not the case. | |
602 | */ | |
603 | if (cgrp1 != cgrp2) | |
604 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
605 | } |
606 | ||
607 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
608 | struct perf_event_attr *attr, | |
609 | struct perf_event *group_leader) | |
610 | { | |
611 | struct perf_cgroup *cgrp; | |
612 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
613 | struct fd f = fdget(fd); |
614 | int ret = 0; | |
e5d1367f | 615 | |
2903ff01 | 616 | if (!f.file) |
e5d1367f SE |
617 | return -EBADF; |
618 | ||
ec903c0c TH |
619 | css = css_tryget_online_from_dir(f.file->f_dentry, |
620 | &perf_event_cgrp_subsys); | |
3db272c0 LZ |
621 | if (IS_ERR(css)) { |
622 | ret = PTR_ERR(css); | |
623 | goto out; | |
624 | } | |
e5d1367f SE |
625 | |
626 | cgrp = container_of(css, struct perf_cgroup, css); | |
627 | event->cgrp = cgrp; | |
628 | ||
629 | /* | |
630 | * all events in a group must monitor | |
631 | * the same cgroup because a task belongs | |
632 | * to only one perf cgroup at a time | |
633 | */ | |
634 | if (group_leader && group_leader->cgrp != cgrp) { | |
635 | perf_detach_cgroup(event); | |
636 | ret = -EINVAL; | |
e5d1367f | 637 | } |
3db272c0 | 638 | out: |
2903ff01 | 639 | fdput(f); |
e5d1367f SE |
640 | return ret; |
641 | } | |
642 | ||
643 | static inline void | |
644 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
645 | { | |
646 | struct perf_cgroup_info *t; | |
647 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
648 | event->shadow_ctx_time = now - t->timestamp; | |
649 | } | |
650 | ||
651 | static inline void | |
652 | perf_cgroup_defer_enabled(struct perf_event *event) | |
653 | { | |
654 | /* | |
655 | * when the current task's perf cgroup does not match | |
656 | * the event's, we need to remember to call the | |
657 | * perf_mark_enable() function the first time a task with | |
658 | * a matching perf cgroup is scheduled in. | |
659 | */ | |
660 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
661 | event->cgrp_defer_enabled = 1; | |
662 | } | |
663 | ||
664 | static inline void | |
665 | perf_cgroup_mark_enabled(struct perf_event *event, | |
666 | struct perf_event_context *ctx) | |
667 | { | |
668 | struct perf_event *sub; | |
669 | u64 tstamp = perf_event_time(event); | |
670 | ||
671 | if (!event->cgrp_defer_enabled) | |
672 | return; | |
673 | ||
674 | event->cgrp_defer_enabled = 0; | |
675 | ||
676 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
677 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
678 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
679 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
680 | sub->cgrp_defer_enabled = 0; | |
681 | } | |
682 | } | |
683 | } | |
684 | #else /* !CONFIG_CGROUP_PERF */ | |
685 | ||
686 | static inline bool | |
687 | perf_cgroup_match(struct perf_event *event) | |
688 | { | |
689 | return true; | |
690 | } | |
691 | ||
692 | static inline void perf_detach_cgroup(struct perf_event *event) | |
693 | {} | |
694 | ||
695 | static inline int is_cgroup_event(struct perf_event *event) | |
696 | { | |
697 | return 0; | |
698 | } | |
699 | ||
700 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
701 | { | |
702 | return 0; | |
703 | } | |
704 | ||
705 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
706 | { | |
707 | } | |
708 | ||
709 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
710 | { | |
711 | } | |
712 | ||
a8d757ef SE |
713 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
714 | struct task_struct *next) | |
e5d1367f SE |
715 | { |
716 | } | |
717 | ||
a8d757ef SE |
718 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
719 | struct task_struct *task) | |
e5d1367f SE |
720 | { |
721 | } | |
722 | ||
723 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
724 | struct perf_event_attr *attr, | |
725 | struct perf_event *group_leader) | |
726 | { | |
727 | return -EINVAL; | |
728 | } | |
729 | ||
730 | static inline void | |
3f7cce3c SE |
731 | perf_cgroup_set_timestamp(struct task_struct *task, |
732 | struct perf_event_context *ctx) | |
e5d1367f SE |
733 | { |
734 | } | |
735 | ||
736 | void | |
737 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
738 | { | |
739 | } | |
740 | ||
741 | static inline void | |
742 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
743 | { | |
744 | } | |
745 | ||
746 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
747 | { | |
748 | return 0; | |
749 | } | |
750 | ||
751 | static inline void | |
752 | perf_cgroup_defer_enabled(struct perf_event *event) | |
753 | { | |
754 | } | |
755 | ||
756 | static inline void | |
757 | perf_cgroup_mark_enabled(struct perf_event *event, | |
758 | struct perf_event_context *ctx) | |
759 | { | |
760 | } | |
761 | #endif | |
762 | ||
9e630205 SE |
763 | /* |
764 | * set default to be dependent on timer tick just | |
765 | * like original code | |
766 | */ | |
767 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
768 | /* | |
769 | * function must be called with interrupts disbled | |
770 | */ | |
771 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
772 | { | |
773 | struct perf_cpu_context *cpuctx; | |
774 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
775 | int rotations = 0; | |
776 | ||
777 | WARN_ON(!irqs_disabled()); | |
778 | ||
779 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
780 | ||
781 | rotations = perf_rotate_context(cpuctx); | |
782 | ||
783 | /* | |
784 | * arm timer if needed | |
785 | */ | |
786 | if (rotations) { | |
787 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
788 | ret = HRTIMER_RESTART; | |
789 | } | |
790 | ||
791 | return ret; | |
792 | } | |
793 | ||
794 | /* CPU is going down */ | |
795 | void perf_cpu_hrtimer_cancel(int cpu) | |
796 | { | |
797 | struct perf_cpu_context *cpuctx; | |
798 | struct pmu *pmu; | |
799 | unsigned long flags; | |
800 | ||
801 | if (WARN_ON(cpu != smp_processor_id())) | |
802 | return; | |
803 | ||
804 | local_irq_save(flags); | |
805 | ||
806 | rcu_read_lock(); | |
807 | ||
808 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
809 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
810 | ||
811 | if (pmu->task_ctx_nr == perf_sw_context) | |
812 | continue; | |
813 | ||
814 | hrtimer_cancel(&cpuctx->hrtimer); | |
815 | } | |
816 | ||
817 | rcu_read_unlock(); | |
818 | ||
819 | local_irq_restore(flags); | |
820 | } | |
821 | ||
822 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
823 | { | |
824 | struct hrtimer *hr = &cpuctx->hrtimer; | |
825 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 826 | int timer; |
9e630205 SE |
827 | |
828 | /* no multiplexing needed for SW PMU */ | |
829 | if (pmu->task_ctx_nr == perf_sw_context) | |
830 | return; | |
831 | ||
62b85639 SE |
832 | /* |
833 | * check default is sane, if not set then force to | |
834 | * default interval (1/tick) | |
835 | */ | |
836 | timer = pmu->hrtimer_interval_ms; | |
837 | if (timer < 1) | |
838 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
839 | ||
840 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
841 | |
842 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
843 | hr->function = perf_cpu_hrtimer_handler; | |
844 | } | |
845 | ||
846 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
847 | { | |
848 | struct hrtimer *hr = &cpuctx->hrtimer; | |
849 | struct pmu *pmu = cpuctx->ctx.pmu; | |
850 | ||
851 | /* not for SW PMU */ | |
852 | if (pmu->task_ctx_nr == perf_sw_context) | |
853 | return; | |
854 | ||
855 | if (hrtimer_active(hr)) | |
856 | return; | |
857 | ||
858 | if (!hrtimer_callback_running(hr)) | |
859 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
860 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
861 | } | |
862 | ||
33696fc0 | 863 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 864 | { |
33696fc0 PZ |
865 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
866 | if (!(*count)++) | |
867 | pmu->pmu_disable(pmu); | |
9e35ad38 | 868 | } |
9e35ad38 | 869 | |
33696fc0 | 870 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 871 | { |
33696fc0 PZ |
872 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
873 | if (!--(*count)) | |
874 | pmu->pmu_enable(pmu); | |
9e35ad38 | 875 | } |
9e35ad38 | 876 | |
e9d2b064 PZ |
877 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
878 | ||
879 | /* | |
880 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
881 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
882 | * disabled, while rotate_context is called from IRQ context. | |
883 | */ | |
108b02cf | 884 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 885 | { |
108b02cf | 886 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 887 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 888 | |
e9d2b064 | 889 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 890 | |
d84153d6 | 891 | if (list_empty(&cpuctx->rotation_list)) |
e9d2b064 | 892 | list_add(&cpuctx->rotation_list, head); |
9e35ad38 | 893 | } |
9e35ad38 | 894 | |
cdd6c482 | 895 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 896 | { |
e5289d4a | 897 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
898 | } |
899 | ||
cdd6c482 | 900 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 901 | { |
564c2b21 PM |
902 | if (atomic_dec_and_test(&ctx->refcount)) { |
903 | if (ctx->parent_ctx) | |
904 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
905 | if (ctx->task) |
906 | put_task_struct(ctx->task); | |
cb796ff3 | 907 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 908 | } |
a63eaf34 PM |
909 | } |
910 | ||
cdd6c482 | 911 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
912 | { |
913 | if (ctx->parent_ctx) { | |
914 | put_ctx(ctx->parent_ctx); | |
915 | ctx->parent_ctx = NULL; | |
916 | } | |
5a3126d4 | 917 | ctx->generation++; |
71a851b4 PZ |
918 | } |
919 | ||
6844c09d ACM |
920 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
921 | { | |
922 | /* | |
923 | * only top level events have the pid namespace they were created in | |
924 | */ | |
925 | if (event->parent) | |
926 | event = event->parent; | |
927 | ||
928 | return task_tgid_nr_ns(p, event->ns); | |
929 | } | |
930 | ||
931 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
932 | { | |
933 | /* | |
934 | * only top level events have the pid namespace they were created in | |
935 | */ | |
936 | if (event->parent) | |
937 | event = event->parent; | |
938 | ||
939 | return task_pid_nr_ns(p, event->ns); | |
940 | } | |
941 | ||
7f453c24 | 942 | /* |
cdd6c482 | 943 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
944 | * to userspace. |
945 | */ | |
cdd6c482 | 946 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 947 | { |
cdd6c482 | 948 | u64 id = event->id; |
7f453c24 | 949 | |
cdd6c482 IM |
950 | if (event->parent) |
951 | id = event->parent->id; | |
7f453c24 PZ |
952 | |
953 | return id; | |
954 | } | |
955 | ||
25346b93 | 956 | /* |
cdd6c482 | 957 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
958 | * This has to cope with with the fact that until it is locked, |
959 | * the context could get moved to another task. | |
960 | */ | |
cdd6c482 | 961 | static struct perf_event_context * |
8dc85d54 | 962 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 963 | { |
cdd6c482 | 964 | struct perf_event_context *ctx; |
25346b93 | 965 | |
9ed6060d | 966 | retry: |
058ebd0e PZ |
967 | /* |
968 | * One of the few rules of preemptible RCU is that one cannot do | |
969 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
970 | * part of the read side critical section was preemptible -- see | |
971 | * rcu_read_unlock_special(). | |
972 | * | |
973 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
974 | * side critical section is non-preemptible. | |
975 | */ | |
976 | preempt_disable(); | |
977 | rcu_read_lock(); | |
8dc85d54 | 978 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
979 | if (ctx) { |
980 | /* | |
981 | * If this context is a clone of another, it might | |
982 | * get swapped for another underneath us by | |
cdd6c482 | 983 | * perf_event_task_sched_out, though the |
25346b93 PM |
984 | * rcu_read_lock() protects us from any context |
985 | * getting freed. Lock the context and check if it | |
986 | * got swapped before we could get the lock, and retry | |
987 | * if so. If we locked the right context, then it | |
988 | * can't get swapped on us any more. | |
989 | */ | |
e625cce1 | 990 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 991 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 992 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
993 | rcu_read_unlock(); |
994 | preempt_enable(); | |
25346b93 PM |
995 | goto retry; |
996 | } | |
b49a9e7e PZ |
997 | |
998 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 999 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
1000 | ctx = NULL; |
1001 | } | |
25346b93 PM |
1002 | } |
1003 | rcu_read_unlock(); | |
058ebd0e | 1004 | preempt_enable(); |
25346b93 PM |
1005 | return ctx; |
1006 | } | |
1007 | ||
1008 | /* | |
1009 | * Get the context for a task and increment its pin_count so it | |
1010 | * can't get swapped to another task. This also increments its | |
1011 | * reference count so that the context can't get freed. | |
1012 | */ | |
8dc85d54 PZ |
1013 | static struct perf_event_context * |
1014 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 1015 | { |
cdd6c482 | 1016 | struct perf_event_context *ctx; |
25346b93 PM |
1017 | unsigned long flags; |
1018 | ||
8dc85d54 | 1019 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1020 | if (ctx) { |
1021 | ++ctx->pin_count; | |
e625cce1 | 1022 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1023 | } |
1024 | return ctx; | |
1025 | } | |
1026 | ||
cdd6c482 | 1027 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1028 | { |
1029 | unsigned long flags; | |
1030 | ||
e625cce1 | 1031 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1032 | --ctx->pin_count; |
e625cce1 | 1033 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1034 | } |
1035 | ||
f67218c3 PZ |
1036 | /* |
1037 | * Update the record of the current time in a context. | |
1038 | */ | |
1039 | static void update_context_time(struct perf_event_context *ctx) | |
1040 | { | |
1041 | u64 now = perf_clock(); | |
1042 | ||
1043 | ctx->time += now - ctx->timestamp; | |
1044 | ctx->timestamp = now; | |
1045 | } | |
1046 | ||
4158755d SE |
1047 | static u64 perf_event_time(struct perf_event *event) |
1048 | { | |
1049 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1050 | |
1051 | if (is_cgroup_event(event)) | |
1052 | return perf_cgroup_event_time(event); | |
1053 | ||
4158755d SE |
1054 | return ctx ? ctx->time : 0; |
1055 | } | |
1056 | ||
f67218c3 PZ |
1057 | /* |
1058 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1059 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1060 | */ |
1061 | static void update_event_times(struct perf_event *event) | |
1062 | { | |
1063 | struct perf_event_context *ctx = event->ctx; | |
1064 | u64 run_end; | |
1065 | ||
1066 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1067 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1068 | return; | |
e5d1367f SE |
1069 | /* |
1070 | * in cgroup mode, time_enabled represents | |
1071 | * the time the event was enabled AND active | |
1072 | * tasks were in the monitored cgroup. This is | |
1073 | * independent of the activity of the context as | |
1074 | * there may be a mix of cgroup and non-cgroup events. | |
1075 | * | |
1076 | * That is why we treat cgroup events differently | |
1077 | * here. | |
1078 | */ | |
1079 | if (is_cgroup_event(event)) | |
46cd6a7f | 1080 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1081 | else if (ctx->is_active) |
1082 | run_end = ctx->time; | |
acd1d7c1 PZ |
1083 | else |
1084 | run_end = event->tstamp_stopped; | |
1085 | ||
1086 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1087 | |
1088 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1089 | run_end = event->tstamp_stopped; | |
1090 | else | |
4158755d | 1091 | run_end = perf_event_time(event); |
f67218c3 PZ |
1092 | |
1093 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1094 | |
f67218c3 PZ |
1095 | } |
1096 | ||
96c21a46 PZ |
1097 | /* |
1098 | * Update total_time_enabled and total_time_running for all events in a group. | |
1099 | */ | |
1100 | static void update_group_times(struct perf_event *leader) | |
1101 | { | |
1102 | struct perf_event *event; | |
1103 | ||
1104 | update_event_times(leader); | |
1105 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1106 | update_event_times(event); | |
1107 | } | |
1108 | ||
889ff015 FW |
1109 | static struct list_head * |
1110 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1111 | { | |
1112 | if (event->attr.pinned) | |
1113 | return &ctx->pinned_groups; | |
1114 | else | |
1115 | return &ctx->flexible_groups; | |
1116 | } | |
1117 | ||
fccc714b | 1118 | /* |
cdd6c482 | 1119 | * Add a event from the lists for its context. |
fccc714b PZ |
1120 | * Must be called with ctx->mutex and ctx->lock held. |
1121 | */ | |
04289bb9 | 1122 | static void |
cdd6c482 | 1123 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1124 | { |
8a49542c PZ |
1125 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1126 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1127 | |
1128 | /* | |
8a49542c PZ |
1129 | * If we're a stand alone event or group leader, we go to the context |
1130 | * list, group events are kept attached to the group so that | |
1131 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1132 | */ |
8a49542c | 1133 | if (event->group_leader == event) { |
889ff015 FW |
1134 | struct list_head *list; |
1135 | ||
d6f962b5 FW |
1136 | if (is_software_event(event)) |
1137 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1138 | ||
889ff015 FW |
1139 | list = ctx_group_list(event, ctx); |
1140 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1141 | } |
592903cd | 1142 | |
08309379 | 1143 | if (is_cgroup_event(event)) |
e5d1367f | 1144 | ctx->nr_cgroups++; |
e5d1367f | 1145 | |
d010b332 SE |
1146 | if (has_branch_stack(event)) |
1147 | ctx->nr_branch_stack++; | |
1148 | ||
cdd6c482 | 1149 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1150 | if (!ctx->nr_events) |
108b02cf | 1151 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1152 | ctx->nr_events++; |
1153 | if (event->attr.inherit_stat) | |
bfbd3381 | 1154 | ctx->nr_stat++; |
5a3126d4 PZ |
1155 | |
1156 | ctx->generation++; | |
04289bb9 IM |
1157 | } |
1158 | ||
0231bb53 JO |
1159 | /* |
1160 | * Initialize event state based on the perf_event_attr::disabled. | |
1161 | */ | |
1162 | static inline void perf_event__state_init(struct perf_event *event) | |
1163 | { | |
1164 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1165 | PERF_EVENT_STATE_INACTIVE; | |
1166 | } | |
1167 | ||
c320c7b7 ACM |
1168 | /* |
1169 | * Called at perf_event creation and when events are attached/detached from a | |
1170 | * group. | |
1171 | */ | |
1172 | static void perf_event__read_size(struct perf_event *event) | |
1173 | { | |
1174 | int entry = sizeof(u64); /* value */ | |
1175 | int size = 0; | |
1176 | int nr = 1; | |
1177 | ||
1178 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1179 | size += sizeof(u64); | |
1180 | ||
1181 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1182 | size += sizeof(u64); | |
1183 | ||
1184 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1185 | entry += sizeof(u64); | |
1186 | ||
1187 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1188 | nr += event->group_leader->nr_siblings; | |
1189 | size += sizeof(u64); | |
1190 | } | |
1191 | ||
1192 | size += entry * nr; | |
1193 | event->read_size = size; | |
1194 | } | |
1195 | ||
1196 | static void perf_event__header_size(struct perf_event *event) | |
1197 | { | |
1198 | struct perf_sample_data *data; | |
1199 | u64 sample_type = event->attr.sample_type; | |
1200 | u16 size = 0; | |
1201 | ||
1202 | perf_event__read_size(event); | |
1203 | ||
1204 | if (sample_type & PERF_SAMPLE_IP) | |
1205 | size += sizeof(data->ip); | |
1206 | ||
6844c09d ACM |
1207 | if (sample_type & PERF_SAMPLE_ADDR) |
1208 | size += sizeof(data->addr); | |
1209 | ||
1210 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1211 | size += sizeof(data->period); | |
1212 | ||
c3feedf2 AK |
1213 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1214 | size += sizeof(data->weight); | |
1215 | ||
6844c09d ACM |
1216 | if (sample_type & PERF_SAMPLE_READ) |
1217 | size += event->read_size; | |
1218 | ||
d6be9ad6 SE |
1219 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1220 | size += sizeof(data->data_src.val); | |
1221 | ||
fdfbbd07 AK |
1222 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
1223 | size += sizeof(data->txn); | |
1224 | ||
6844c09d ACM |
1225 | event->header_size = size; |
1226 | } | |
1227 | ||
1228 | static void perf_event__id_header_size(struct perf_event *event) | |
1229 | { | |
1230 | struct perf_sample_data *data; | |
1231 | u64 sample_type = event->attr.sample_type; | |
1232 | u16 size = 0; | |
1233 | ||
c320c7b7 ACM |
1234 | if (sample_type & PERF_SAMPLE_TID) |
1235 | size += sizeof(data->tid_entry); | |
1236 | ||
1237 | if (sample_type & PERF_SAMPLE_TIME) | |
1238 | size += sizeof(data->time); | |
1239 | ||
ff3d527c AH |
1240 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1241 | size += sizeof(data->id); | |
1242 | ||
c320c7b7 ACM |
1243 | if (sample_type & PERF_SAMPLE_ID) |
1244 | size += sizeof(data->id); | |
1245 | ||
1246 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1247 | size += sizeof(data->stream_id); | |
1248 | ||
1249 | if (sample_type & PERF_SAMPLE_CPU) | |
1250 | size += sizeof(data->cpu_entry); | |
1251 | ||
6844c09d | 1252 | event->id_header_size = size; |
c320c7b7 ACM |
1253 | } |
1254 | ||
8a49542c PZ |
1255 | static void perf_group_attach(struct perf_event *event) |
1256 | { | |
c320c7b7 | 1257 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1258 | |
74c3337c PZ |
1259 | /* |
1260 | * We can have double attach due to group movement in perf_event_open. | |
1261 | */ | |
1262 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1263 | return; | |
1264 | ||
8a49542c PZ |
1265 | event->attach_state |= PERF_ATTACH_GROUP; |
1266 | ||
1267 | if (group_leader == event) | |
1268 | return; | |
1269 | ||
1270 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1271 | !is_software_event(event)) | |
1272 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1273 | ||
1274 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1275 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1276 | |
1277 | perf_event__header_size(group_leader); | |
1278 | ||
1279 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1280 | perf_event__header_size(pos); | |
8a49542c PZ |
1281 | } |
1282 | ||
a63eaf34 | 1283 | /* |
cdd6c482 | 1284 | * Remove a event from the lists for its context. |
fccc714b | 1285 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1286 | */ |
04289bb9 | 1287 | static void |
cdd6c482 | 1288 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1289 | { |
68cacd29 | 1290 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1291 | /* |
1292 | * We can have double detach due to exit/hot-unplug + close. | |
1293 | */ | |
1294 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1295 | return; |
8a49542c PZ |
1296 | |
1297 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1298 | ||
68cacd29 | 1299 | if (is_cgroup_event(event)) { |
e5d1367f | 1300 | ctx->nr_cgroups--; |
68cacd29 SE |
1301 | cpuctx = __get_cpu_context(ctx); |
1302 | /* | |
1303 | * if there are no more cgroup events | |
1304 | * then cler cgrp to avoid stale pointer | |
1305 | * in update_cgrp_time_from_cpuctx() | |
1306 | */ | |
1307 | if (!ctx->nr_cgroups) | |
1308 | cpuctx->cgrp = NULL; | |
1309 | } | |
e5d1367f | 1310 | |
d010b332 SE |
1311 | if (has_branch_stack(event)) |
1312 | ctx->nr_branch_stack--; | |
1313 | ||
cdd6c482 IM |
1314 | ctx->nr_events--; |
1315 | if (event->attr.inherit_stat) | |
bfbd3381 | 1316 | ctx->nr_stat--; |
8bc20959 | 1317 | |
cdd6c482 | 1318 | list_del_rcu(&event->event_entry); |
04289bb9 | 1319 | |
8a49542c PZ |
1320 | if (event->group_leader == event) |
1321 | list_del_init(&event->group_entry); | |
5c148194 | 1322 | |
96c21a46 | 1323 | update_group_times(event); |
b2e74a26 SE |
1324 | |
1325 | /* | |
1326 | * If event was in error state, then keep it | |
1327 | * that way, otherwise bogus counts will be | |
1328 | * returned on read(). The only way to get out | |
1329 | * of error state is by explicit re-enabling | |
1330 | * of the event | |
1331 | */ | |
1332 | if (event->state > PERF_EVENT_STATE_OFF) | |
1333 | event->state = PERF_EVENT_STATE_OFF; | |
5a3126d4 PZ |
1334 | |
1335 | ctx->generation++; | |
050735b0 PZ |
1336 | } |
1337 | ||
8a49542c | 1338 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1339 | { |
1340 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1341 | struct list_head *list = NULL; |
1342 | ||
1343 | /* | |
1344 | * We can have double detach due to exit/hot-unplug + close. | |
1345 | */ | |
1346 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1347 | return; | |
1348 | ||
1349 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1350 | ||
1351 | /* | |
1352 | * If this is a sibling, remove it from its group. | |
1353 | */ | |
1354 | if (event->group_leader != event) { | |
1355 | list_del_init(&event->group_entry); | |
1356 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1357 | goto out; |
8a49542c PZ |
1358 | } |
1359 | ||
1360 | if (!list_empty(&event->group_entry)) | |
1361 | list = &event->group_entry; | |
2e2af50b | 1362 | |
04289bb9 | 1363 | /* |
cdd6c482 IM |
1364 | * If this was a group event with sibling events then |
1365 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1366 | * to whatever list we are on. |
04289bb9 | 1367 | */ |
cdd6c482 | 1368 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1369 | if (list) |
1370 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1371 | sibling->group_leader = sibling; |
d6f962b5 FW |
1372 | |
1373 | /* Inherit group flags from the previous leader */ | |
1374 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1375 | } |
c320c7b7 ACM |
1376 | |
1377 | out: | |
1378 | perf_event__header_size(event->group_leader); | |
1379 | ||
1380 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1381 | perf_event__header_size(tmp); | |
04289bb9 IM |
1382 | } |
1383 | ||
fa66f07a SE |
1384 | static inline int |
1385 | event_filter_match(struct perf_event *event) | |
1386 | { | |
e5d1367f SE |
1387 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1388 | && perf_cgroup_match(event); | |
fa66f07a SE |
1389 | } |
1390 | ||
9ffcfa6f SE |
1391 | static void |
1392 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1393 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1394 | struct perf_event_context *ctx) |
3b6f9e5c | 1395 | { |
4158755d | 1396 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1397 | u64 delta; |
1398 | /* | |
1399 | * An event which could not be activated because of | |
1400 | * filter mismatch still needs to have its timings | |
1401 | * maintained, otherwise bogus information is return | |
1402 | * via read() for time_enabled, time_running: | |
1403 | */ | |
1404 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1405 | && !event_filter_match(event)) { | |
e5d1367f | 1406 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1407 | event->tstamp_running += delta; |
4158755d | 1408 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1409 | } |
1410 | ||
cdd6c482 | 1411 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1412 | return; |
3b6f9e5c | 1413 | |
44377277 AS |
1414 | perf_pmu_disable(event->pmu); |
1415 | ||
cdd6c482 IM |
1416 | event->state = PERF_EVENT_STATE_INACTIVE; |
1417 | if (event->pending_disable) { | |
1418 | event->pending_disable = 0; | |
1419 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1420 | } |
4158755d | 1421 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1422 | event->pmu->del(event, 0); |
cdd6c482 | 1423 | event->oncpu = -1; |
3b6f9e5c | 1424 | |
cdd6c482 | 1425 | if (!is_software_event(event)) |
3b6f9e5c PM |
1426 | cpuctx->active_oncpu--; |
1427 | ctx->nr_active--; | |
0f5a2601 PZ |
1428 | if (event->attr.freq && event->attr.sample_freq) |
1429 | ctx->nr_freq--; | |
cdd6c482 | 1430 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c | 1431 | cpuctx->exclusive = 0; |
44377277 AS |
1432 | |
1433 | perf_pmu_enable(event->pmu); | |
3b6f9e5c PM |
1434 | } |
1435 | ||
d859e29f | 1436 | static void |
cdd6c482 | 1437 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1438 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1439 | struct perf_event_context *ctx) |
d859e29f | 1440 | { |
cdd6c482 | 1441 | struct perf_event *event; |
fa66f07a | 1442 | int state = group_event->state; |
d859e29f | 1443 | |
cdd6c482 | 1444 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1445 | |
1446 | /* | |
1447 | * Schedule out siblings (if any): | |
1448 | */ | |
cdd6c482 IM |
1449 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1450 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1451 | |
fa66f07a | 1452 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1453 | cpuctx->exclusive = 0; |
1454 | } | |
1455 | ||
46ce0fe9 PZ |
1456 | struct remove_event { |
1457 | struct perf_event *event; | |
1458 | bool detach_group; | |
1459 | }; | |
1460 | ||
0793a61d | 1461 | /* |
cdd6c482 | 1462 | * Cross CPU call to remove a performance event |
0793a61d | 1463 | * |
cdd6c482 | 1464 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1465 | * remove it from the context list. |
1466 | */ | |
fe4b04fa | 1467 | static int __perf_remove_from_context(void *info) |
0793a61d | 1468 | { |
46ce0fe9 PZ |
1469 | struct remove_event *re = info; |
1470 | struct perf_event *event = re->event; | |
cdd6c482 | 1471 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1472 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1473 | |
e625cce1 | 1474 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1475 | event_sched_out(event, cpuctx, ctx); |
46ce0fe9 PZ |
1476 | if (re->detach_group) |
1477 | perf_group_detach(event); | |
cdd6c482 | 1478 | list_del_event(event, ctx); |
64ce3126 PZ |
1479 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1480 | ctx->is_active = 0; | |
1481 | cpuctx->task_ctx = NULL; | |
1482 | } | |
e625cce1 | 1483 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1484 | |
1485 | return 0; | |
0793a61d TG |
1486 | } |
1487 | ||
1488 | ||
1489 | /* | |
cdd6c482 | 1490 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1491 | * |
cdd6c482 | 1492 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1493 | * call when the task is on a CPU. |
c93f7669 | 1494 | * |
cdd6c482 IM |
1495 | * If event->ctx is a cloned context, callers must make sure that |
1496 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1497 | * remains valid. This is OK when called from perf_release since |
1498 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1499 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1500 | * context has been detached from its task. |
0793a61d | 1501 | */ |
46ce0fe9 | 1502 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
0793a61d | 1503 | { |
cdd6c482 | 1504 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 1505 | struct task_struct *task = ctx->task; |
46ce0fe9 PZ |
1506 | struct remove_event re = { |
1507 | .event = event, | |
1508 | .detach_group = detach_group, | |
1509 | }; | |
0793a61d | 1510 | |
fe4b04fa PZ |
1511 | lockdep_assert_held(&ctx->mutex); |
1512 | ||
0793a61d TG |
1513 | if (!task) { |
1514 | /* | |
cdd6c482 | 1515 | * Per cpu events are removed via an smp call and |
af901ca1 | 1516 | * the removal is always successful. |
0793a61d | 1517 | */ |
46ce0fe9 | 1518 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
0793a61d TG |
1519 | return; |
1520 | } | |
1521 | ||
1522 | retry: | |
46ce0fe9 | 1523 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
fe4b04fa | 1524 | return; |
0793a61d | 1525 | |
e625cce1 | 1526 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1527 | /* |
fe4b04fa PZ |
1528 | * If we failed to find a running task, but find the context active now |
1529 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1530 | */ |
fe4b04fa | 1531 | if (ctx->is_active) { |
e625cce1 | 1532 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1533 | goto retry; |
1534 | } | |
1535 | ||
1536 | /* | |
fe4b04fa PZ |
1537 | * Since the task isn't running, its safe to remove the event, us |
1538 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1539 | */ |
46ce0fe9 PZ |
1540 | if (detach_group) |
1541 | perf_group_detach(event); | |
fe4b04fa | 1542 | list_del_event(event, ctx); |
e625cce1 | 1543 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1544 | } |
1545 | ||
d859e29f | 1546 | /* |
cdd6c482 | 1547 | * Cross CPU call to disable a performance event |
d859e29f | 1548 | */ |
500ad2d8 | 1549 | int __perf_event_disable(void *info) |
d859e29f | 1550 | { |
cdd6c482 | 1551 | struct perf_event *event = info; |
cdd6c482 | 1552 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1553 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1554 | |
1555 | /* | |
cdd6c482 IM |
1556 | * If this is a per-task event, need to check whether this |
1557 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1558 | * |
1559 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1560 | * flipping contexts around. | |
d859e29f | 1561 | */ |
665c2142 | 1562 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1563 | return -EINVAL; |
d859e29f | 1564 | |
e625cce1 | 1565 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1566 | |
1567 | /* | |
cdd6c482 | 1568 | * If the event is on, turn it off. |
d859e29f PM |
1569 | * If it is in error state, leave it in error state. |
1570 | */ | |
cdd6c482 | 1571 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1572 | update_context_time(ctx); |
e5d1367f | 1573 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1574 | update_group_times(event); |
1575 | if (event == event->group_leader) | |
1576 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1577 | else |
cdd6c482 IM |
1578 | event_sched_out(event, cpuctx, ctx); |
1579 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1580 | } |
1581 | ||
e625cce1 | 1582 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1583 | |
1584 | return 0; | |
d859e29f PM |
1585 | } |
1586 | ||
1587 | /* | |
cdd6c482 | 1588 | * Disable a event. |
c93f7669 | 1589 | * |
cdd6c482 IM |
1590 | * If event->ctx is a cloned context, callers must make sure that |
1591 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1592 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1593 | * perf_event_for_each_child or perf_event_for_each because they |
1594 | * hold the top-level event's child_mutex, so any descendant that | |
1595 | * goes to exit will block in sync_child_event. | |
1596 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1597 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1598 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1599 | */ |
44234adc | 1600 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1601 | { |
cdd6c482 | 1602 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1603 | struct task_struct *task = ctx->task; |
1604 | ||
1605 | if (!task) { | |
1606 | /* | |
cdd6c482 | 1607 | * Disable the event on the cpu that it's on |
d859e29f | 1608 | */ |
fe4b04fa | 1609 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1610 | return; |
1611 | } | |
1612 | ||
9ed6060d | 1613 | retry: |
fe4b04fa PZ |
1614 | if (!task_function_call(task, __perf_event_disable, event)) |
1615 | return; | |
d859e29f | 1616 | |
e625cce1 | 1617 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1618 | /* |
cdd6c482 | 1619 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1620 | */ |
cdd6c482 | 1621 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1622 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1623 | /* |
1624 | * Reload the task pointer, it might have been changed by | |
1625 | * a concurrent perf_event_context_sched_out(). | |
1626 | */ | |
1627 | task = ctx->task; | |
d859e29f PM |
1628 | goto retry; |
1629 | } | |
1630 | ||
1631 | /* | |
1632 | * Since we have the lock this context can't be scheduled | |
1633 | * in, so we can change the state safely. | |
1634 | */ | |
cdd6c482 IM |
1635 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1636 | update_group_times(event); | |
1637 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1638 | } |
e625cce1 | 1639 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1640 | } |
dcfce4a0 | 1641 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1642 | |
e5d1367f SE |
1643 | static void perf_set_shadow_time(struct perf_event *event, |
1644 | struct perf_event_context *ctx, | |
1645 | u64 tstamp) | |
1646 | { | |
1647 | /* | |
1648 | * use the correct time source for the time snapshot | |
1649 | * | |
1650 | * We could get by without this by leveraging the | |
1651 | * fact that to get to this function, the caller | |
1652 | * has most likely already called update_context_time() | |
1653 | * and update_cgrp_time_xx() and thus both timestamp | |
1654 | * are identical (or very close). Given that tstamp is, | |
1655 | * already adjusted for cgroup, we could say that: | |
1656 | * tstamp - ctx->timestamp | |
1657 | * is equivalent to | |
1658 | * tstamp - cgrp->timestamp. | |
1659 | * | |
1660 | * Then, in perf_output_read(), the calculation would | |
1661 | * work with no changes because: | |
1662 | * - event is guaranteed scheduled in | |
1663 | * - no scheduled out in between | |
1664 | * - thus the timestamp would be the same | |
1665 | * | |
1666 | * But this is a bit hairy. | |
1667 | * | |
1668 | * So instead, we have an explicit cgroup call to remain | |
1669 | * within the time time source all along. We believe it | |
1670 | * is cleaner and simpler to understand. | |
1671 | */ | |
1672 | if (is_cgroup_event(event)) | |
1673 | perf_cgroup_set_shadow_time(event, tstamp); | |
1674 | else | |
1675 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1676 | } | |
1677 | ||
4fe757dd PZ |
1678 | #define MAX_INTERRUPTS (~0ULL) |
1679 | ||
1680 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1681 | ||
235c7fc7 | 1682 | static int |
9ffcfa6f | 1683 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1684 | struct perf_cpu_context *cpuctx, |
6e37738a | 1685 | struct perf_event_context *ctx) |
235c7fc7 | 1686 | { |
4158755d | 1687 | u64 tstamp = perf_event_time(event); |
44377277 | 1688 | int ret = 0; |
4158755d | 1689 | |
63342411 PZ |
1690 | lockdep_assert_held(&ctx->lock); |
1691 | ||
cdd6c482 | 1692 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1693 | return 0; |
1694 | ||
cdd6c482 | 1695 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1696 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1697 | |
1698 | /* | |
1699 | * Unthrottle events, since we scheduled we might have missed several | |
1700 | * ticks already, also for a heavily scheduling task there is little | |
1701 | * guarantee it'll get a tick in a timely manner. | |
1702 | */ | |
1703 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1704 | perf_log_throttle(event, 1); | |
1705 | event->hw.interrupts = 0; | |
1706 | } | |
1707 | ||
235c7fc7 IM |
1708 | /* |
1709 | * The new state must be visible before we turn it on in the hardware: | |
1710 | */ | |
1711 | smp_wmb(); | |
1712 | ||
44377277 AS |
1713 | perf_pmu_disable(event->pmu); |
1714 | ||
a4eaf7f1 | 1715 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1716 | event->state = PERF_EVENT_STATE_INACTIVE; |
1717 | event->oncpu = -1; | |
44377277 AS |
1718 | ret = -EAGAIN; |
1719 | goto out; | |
235c7fc7 IM |
1720 | } |
1721 | ||
4158755d | 1722 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1723 | |
e5d1367f | 1724 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1725 | |
cdd6c482 | 1726 | if (!is_software_event(event)) |
3b6f9e5c | 1727 | cpuctx->active_oncpu++; |
235c7fc7 | 1728 | ctx->nr_active++; |
0f5a2601 PZ |
1729 | if (event->attr.freq && event->attr.sample_freq) |
1730 | ctx->nr_freq++; | |
235c7fc7 | 1731 | |
cdd6c482 | 1732 | if (event->attr.exclusive) |
3b6f9e5c PM |
1733 | cpuctx->exclusive = 1; |
1734 | ||
44377277 AS |
1735 | out: |
1736 | perf_pmu_enable(event->pmu); | |
1737 | ||
1738 | return ret; | |
235c7fc7 IM |
1739 | } |
1740 | ||
6751b71e | 1741 | static int |
cdd6c482 | 1742 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1743 | struct perf_cpu_context *cpuctx, |
6e37738a | 1744 | struct perf_event_context *ctx) |
6751b71e | 1745 | { |
6bde9b6c | 1746 | struct perf_event *event, *partial_group = NULL; |
4a234593 | 1747 | struct pmu *pmu = ctx->pmu; |
d7842da4 SE |
1748 | u64 now = ctx->time; |
1749 | bool simulate = false; | |
6751b71e | 1750 | |
cdd6c482 | 1751 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1752 | return 0; |
1753 | ||
ad5133b7 | 1754 | pmu->start_txn(pmu); |
6bde9b6c | 1755 | |
9ffcfa6f | 1756 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1757 | pmu->cancel_txn(pmu); |
9e630205 | 1758 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1759 | return -EAGAIN; |
90151c35 | 1760 | } |
6751b71e PM |
1761 | |
1762 | /* | |
1763 | * Schedule in siblings as one group (if any): | |
1764 | */ | |
cdd6c482 | 1765 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1766 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1767 | partial_group = event; |
6751b71e PM |
1768 | goto group_error; |
1769 | } | |
1770 | } | |
1771 | ||
9ffcfa6f | 1772 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1773 | return 0; |
9ffcfa6f | 1774 | |
6751b71e PM |
1775 | group_error: |
1776 | /* | |
1777 | * Groups can be scheduled in as one unit only, so undo any | |
1778 | * partial group before returning: | |
d7842da4 SE |
1779 | * The events up to the failed event are scheduled out normally, |
1780 | * tstamp_stopped will be updated. | |
1781 | * | |
1782 | * The failed events and the remaining siblings need to have | |
1783 | * their timings updated as if they had gone thru event_sched_in() | |
1784 | * and event_sched_out(). This is required to get consistent timings | |
1785 | * across the group. This also takes care of the case where the group | |
1786 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1787 | * the time the event was actually stopped, such that time delta | |
1788 | * calculation in update_event_times() is correct. | |
6751b71e | 1789 | */ |
cdd6c482 IM |
1790 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1791 | if (event == partial_group) | |
d7842da4 SE |
1792 | simulate = true; |
1793 | ||
1794 | if (simulate) { | |
1795 | event->tstamp_running += now - event->tstamp_stopped; | |
1796 | event->tstamp_stopped = now; | |
1797 | } else { | |
1798 | event_sched_out(event, cpuctx, ctx); | |
1799 | } | |
6751b71e | 1800 | } |
9ffcfa6f | 1801 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1802 | |
ad5133b7 | 1803 | pmu->cancel_txn(pmu); |
90151c35 | 1804 | |
9e630205 SE |
1805 | perf_cpu_hrtimer_restart(cpuctx); |
1806 | ||
6751b71e PM |
1807 | return -EAGAIN; |
1808 | } | |
1809 | ||
3b6f9e5c | 1810 | /* |
cdd6c482 | 1811 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1812 | */ |
cdd6c482 | 1813 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1814 | struct perf_cpu_context *cpuctx, |
1815 | int can_add_hw) | |
1816 | { | |
1817 | /* | |
cdd6c482 | 1818 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1819 | */ |
d6f962b5 | 1820 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1821 | return 1; |
1822 | /* | |
1823 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1824 | * events can go on. |
3b6f9e5c PM |
1825 | */ |
1826 | if (cpuctx->exclusive) | |
1827 | return 0; | |
1828 | /* | |
1829 | * If this group is exclusive and there are already | |
cdd6c482 | 1830 | * events on the CPU, it can't go on. |
3b6f9e5c | 1831 | */ |
cdd6c482 | 1832 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1833 | return 0; |
1834 | /* | |
1835 | * Otherwise, try to add it if all previous groups were able | |
1836 | * to go on. | |
1837 | */ | |
1838 | return can_add_hw; | |
1839 | } | |
1840 | ||
cdd6c482 IM |
1841 | static void add_event_to_ctx(struct perf_event *event, |
1842 | struct perf_event_context *ctx) | |
53cfbf59 | 1843 | { |
4158755d SE |
1844 | u64 tstamp = perf_event_time(event); |
1845 | ||
cdd6c482 | 1846 | list_add_event(event, ctx); |
8a49542c | 1847 | perf_group_attach(event); |
4158755d SE |
1848 | event->tstamp_enabled = tstamp; |
1849 | event->tstamp_running = tstamp; | |
1850 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1851 | } |
1852 | ||
2c29ef0f PZ |
1853 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1854 | static void | |
1855 | ctx_sched_in(struct perf_event_context *ctx, | |
1856 | struct perf_cpu_context *cpuctx, | |
1857 | enum event_type_t event_type, | |
1858 | struct task_struct *task); | |
fe4b04fa | 1859 | |
dce5855b PZ |
1860 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1861 | struct perf_event_context *ctx, | |
1862 | struct task_struct *task) | |
1863 | { | |
1864 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1865 | if (ctx) | |
1866 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1867 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1868 | if (ctx) | |
1869 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1870 | } | |
1871 | ||
0793a61d | 1872 | /* |
cdd6c482 | 1873 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1874 | * |
1875 | * Must be called with ctx->mutex held | |
0793a61d | 1876 | */ |
fe4b04fa | 1877 | static int __perf_install_in_context(void *info) |
0793a61d | 1878 | { |
cdd6c482 IM |
1879 | struct perf_event *event = info; |
1880 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1881 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1882 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1883 | struct task_struct *task = current; | |
1884 | ||
b58f6b0d | 1885 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1886 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1887 | |
1888 | /* | |
2c29ef0f | 1889 | * If there was an active task_ctx schedule it out. |
0793a61d | 1890 | */ |
b58f6b0d | 1891 | if (task_ctx) |
2c29ef0f | 1892 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1893 | |
1894 | /* | |
1895 | * If the context we're installing events in is not the | |
1896 | * active task_ctx, flip them. | |
1897 | */ | |
1898 | if (ctx->task && task_ctx != ctx) { | |
1899 | if (task_ctx) | |
1900 | raw_spin_unlock(&task_ctx->lock); | |
1901 | raw_spin_lock(&ctx->lock); | |
1902 | task_ctx = ctx; | |
1903 | } | |
1904 | ||
1905 | if (task_ctx) { | |
1906 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1907 | task = task_ctx->task; |
1908 | } | |
b58f6b0d | 1909 | |
2c29ef0f | 1910 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1911 | |
4af4998b | 1912 | update_context_time(ctx); |
e5d1367f SE |
1913 | /* |
1914 | * update cgrp time only if current cgrp | |
1915 | * matches event->cgrp. Must be done before | |
1916 | * calling add_event_to_ctx() | |
1917 | */ | |
1918 | update_cgrp_time_from_event(event); | |
0793a61d | 1919 | |
cdd6c482 | 1920 | add_event_to_ctx(event, ctx); |
0793a61d | 1921 | |
d859e29f | 1922 | /* |
2c29ef0f | 1923 | * Schedule everything back in |
d859e29f | 1924 | */ |
dce5855b | 1925 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1926 | |
1927 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1928 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1929 | |
1930 | return 0; | |
0793a61d TG |
1931 | } |
1932 | ||
1933 | /* | |
cdd6c482 | 1934 | * Attach a performance event to a context |
0793a61d | 1935 | * |
cdd6c482 IM |
1936 | * First we add the event to the list with the hardware enable bit |
1937 | * in event->hw_config cleared. | |
0793a61d | 1938 | * |
cdd6c482 | 1939 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1940 | * call to enable it in the task context. The task might have been |
1941 | * scheduled away, but we check this in the smp call again. | |
1942 | */ | |
1943 | static void | |
cdd6c482 IM |
1944 | perf_install_in_context(struct perf_event_context *ctx, |
1945 | struct perf_event *event, | |
0793a61d TG |
1946 | int cpu) |
1947 | { | |
1948 | struct task_struct *task = ctx->task; | |
1949 | ||
fe4b04fa PZ |
1950 | lockdep_assert_held(&ctx->mutex); |
1951 | ||
c3f00c70 | 1952 | event->ctx = ctx; |
0cda4c02 YZ |
1953 | if (event->cpu != -1) |
1954 | event->cpu = cpu; | |
c3f00c70 | 1955 | |
0793a61d TG |
1956 | if (!task) { |
1957 | /* | |
cdd6c482 | 1958 | * Per cpu events are installed via an smp call and |
af901ca1 | 1959 | * the install is always successful. |
0793a61d | 1960 | */ |
fe4b04fa | 1961 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1962 | return; |
1963 | } | |
1964 | ||
0793a61d | 1965 | retry: |
fe4b04fa PZ |
1966 | if (!task_function_call(task, __perf_install_in_context, event)) |
1967 | return; | |
0793a61d | 1968 | |
e625cce1 | 1969 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1970 | /* |
fe4b04fa PZ |
1971 | * If we failed to find a running task, but find the context active now |
1972 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1973 | */ |
fe4b04fa | 1974 | if (ctx->is_active) { |
e625cce1 | 1975 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1976 | goto retry; |
1977 | } | |
1978 | ||
1979 | /* | |
fe4b04fa PZ |
1980 | * Since the task isn't running, its safe to add the event, us holding |
1981 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1982 | */ |
fe4b04fa | 1983 | add_event_to_ctx(event, ctx); |
e625cce1 | 1984 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1985 | } |
1986 | ||
fa289bec | 1987 | /* |
cdd6c482 | 1988 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1989 | * Enabling the leader of a group effectively enables all |
1990 | * the group members that aren't explicitly disabled, so we | |
1991 | * have to update their ->tstamp_enabled also. | |
1992 | * Note: this works for group members as well as group leaders | |
1993 | * since the non-leader members' sibling_lists will be empty. | |
1994 | */ | |
1d9b482e | 1995 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1996 | { |
cdd6c482 | 1997 | struct perf_event *sub; |
4158755d | 1998 | u64 tstamp = perf_event_time(event); |
fa289bec | 1999 | |
cdd6c482 | 2000 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 2001 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 2002 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
2003 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
2004 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 2005 | } |
fa289bec PM |
2006 | } |
2007 | ||
d859e29f | 2008 | /* |
cdd6c482 | 2009 | * Cross CPU call to enable a performance event |
d859e29f | 2010 | */ |
fe4b04fa | 2011 | static int __perf_event_enable(void *info) |
04289bb9 | 2012 | { |
cdd6c482 | 2013 | struct perf_event *event = info; |
cdd6c482 IM |
2014 | struct perf_event_context *ctx = event->ctx; |
2015 | struct perf_event *leader = event->group_leader; | |
108b02cf | 2016 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 2017 | int err; |
04289bb9 | 2018 | |
06f41796 JO |
2019 | /* |
2020 | * There's a time window between 'ctx->is_active' check | |
2021 | * in perf_event_enable function and this place having: | |
2022 | * - IRQs on | |
2023 | * - ctx->lock unlocked | |
2024 | * | |
2025 | * where the task could be killed and 'ctx' deactivated | |
2026 | * by perf_event_exit_task. | |
2027 | */ | |
2028 | if (!ctx->is_active) | |
fe4b04fa | 2029 | return -EINVAL; |
3cbed429 | 2030 | |
e625cce1 | 2031 | raw_spin_lock(&ctx->lock); |
4af4998b | 2032 | update_context_time(ctx); |
d859e29f | 2033 | |
cdd6c482 | 2034 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 2035 | goto unlock; |
e5d1367f SE |
2036 | |
2037 | /* | |
2038 | * set current task's cgroup time reference point | |
2039 | */ | |
3f7cce3c | 2040 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 2041 | |
1d9b482e | 2042 | __perf_event_mark_enabled(event); |
04289bb9 | 2043 | |
e5d1367f SE |
2044 | if (!event_filter_match(event)) { |
2045 | if (is_cgroup_event(event)) | |
2046 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 2047 | goto unlock; |
e5d1367f | 2048 | } |
f4c4176f | 2049 | |
04289bb9 | 2050 | /* |
cdd6c482 | 2051 | * If the event is in a group and isn't the group leader, |
d859e29f | 2052 | * then don't put it on unless the group is on. |
04289bb9 | 2053 | */ |
cdd6c482 | 2054 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2055 | goto unlock; |
3b6f9e5c | 2056 | |
cdd6c482 | 2057 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2058 | err = -EEXIST; |
e758a33d | 2059 | } else { |
cdd6c482 | 2060 | if (event == leader) |
6e37738a | 2061 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2062 | else |
6e37738a | 2063 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2064 | } |
d859e29f PM |
2065 | |
2066 | if (err) { | |
2067 | /* | |
cdd6c482 | 2068 | * If this event can't go on and it's part of a |
d859e29f PM |
2069 | * group, then the whole group has to come off. |
2070 | */ | |
9e630205 | 2071 | if (leader != event) { |
d859e29f | 2072 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
2073 | perf_cpu_hrtimer_restart(cpuctx); |
2074 | } | |
0d48696f | 2075 | if (leader->attr.pinned) { |
53cfbf59 | 2076 | update_group_times(leader); |
cdd6c482 | 2077 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2078 | } |
d859e29f PM |
2079 | } |
2080 | ||
9ed6060d | 2081 | unlock: |
e625cce1 | 2082 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2083 | |
2084 | return 0; | |
d859e29f PM |
2085 | } |
2086 | ||
2087 | /* | |
cdd6c482 | 2088 | * Enable a event. |
c93f7669 | 2089 | * |
cdd6c482 IM |
2090 | * If event->ctx is a cloned context, callers must make sure that |
2091 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2092 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2093 | * perf_event_for_each_child or perf_event_for_each as described |
2094 | * for perf_event_disable. | |
d859e29f | 2095 | */ |
44234adc | 2096 | void perf_event_enable(struct perf_event *event) |
d859e29f | 2097 | { |
cdd6c482 | 2098 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2099 | struct task_struct *task = ctx->task; |
2100 | ||
2101 | if (!task) { | |
2102 | /* | |
cdd6c482 | 2103 | * Enable the event on the cpu that it's on |
d859e29f | 2104 | */ |
fe4b04fa | 2105 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2106 | return; |
2107 | } | |
2108 | ||
e625cce1 | 2109 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2110 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2111 | goto out; |
2112 | ||
2113 | /* | |
cdd6c482 IM |
2114 | * If the event is in error state, clear that first. |
2115 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2116 | * know that it has gone back into error state, as distinct |
2117 | * from the task having been scheduled away before the | |
2118 | * cross-call arrived. | |
2119 | */ | |
cdd6c482 IM |
2120 | if (event->state == PERF_EVENT_STATE_ERROR) |
2121 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2122 | |
9ed6060d | 2123 | retry: |
fe4b04fa | 2124 | if (!ctx->is_active) { |
1d9b482e | 2125 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2126 | goto out; |
2127 | } | |
2128 | ||
e625cce1 | 2129 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2130 | |
2131 | if (!task_function_call(task, __perf_event_enable, event)) | |
2132 | return; | |
d859e29f | 2133 | |
e625cce1 | 2134 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2135 | |
2136 | /* | |
cdd6c482 | 2137 | * If the context is active and the event is still off, |
d859e29f PM |
2138 | * we need to retry the cross-call. |
2139 | */ | |
fe4b04fa PZ |
2140 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2141 | /* | |
2142 | * task could have been flipped by a concurrent | |
2143 | * perf_event_context_sched_out() | |
2144 | */ | |
2145 | task = ctx->task; | |
d859e29f | 2146 | goto retry; |
fe4b04fa | 2147 | } |
fa289bec | 2148 | |
9ed6060d | 2149 | out: |
e625cce1 | 2150 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2151 | } |
dcfce4a0 | 2152 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2153 | |
26ca5c11 | 2154 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2155 | { |
2023b359 | 2156 | /* |
cdd6c482 | 2157 | * not supported on inherited events |
2023b359 | 2158 | */ |
2e939d1d | 2159 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2160 | return -EINVAL; |
2161 | ||
cdd6c482 IM |
2162 | atomic_add(refresh, &event->event_limit); |
2163 | perf_event_enable(event); | |
2023b359 PZ |
2164 | |
2165 | return 0; | |
79f14641 | 2166 | } |
26ca5c11 | 2167 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2168 | |
5b0311e1 FW |
2169 | static void ctx_sched_out(struct perf_event_context *ctx, |
2170 | struct perf_cpu_context *cpuctx, | |
2171 | enum event_type_t event_type) | |
235c7fc7 | 2172 | { |
cdd6c482 | 2173 | struct perf_event *event; |
db24d33e | 2174 | int is_active = ctx->is_active; |
235c7fc7 | 2175 | |
db24d33e | 2176 | ctx->is_active &= ~event_type; |
cdd6c482 | 2177 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2178 | return; |
2179 | ||
4af4998b | 2180 | update_context_time(ctx); |
e5d1367f | 2181 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2182 | if (!ctx->nr_active) |
facc4307 | 2183 | return; |
5b0311e1 | 2184 | |
075e0b00 | 2185 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2186 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2187 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2188 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2189 | } |
889ff015 | 2190 | |
db24d33e | 2191 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2192 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2193 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2194 | } |
1b9a644f | 2195 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2196 | } |
2197 | ||
564c2b21 | 2198 | /* |
5a3126d4 PZ |
2199 | * Test whether two contexts are equivalent, i.e. whether they have both been |
2200 | * cloned from the same version of the same context. | |
2201 | * | |
2202 | * Equivalence is measured using a generation number in the context that is | |
2203 | * incremented on each modification to it; see unclone_ctx(), list_add_event() | |
2204 | * and list_del_event(). | |
564c2b21 | 2205 | */ |
cdd6c482 IM |
2206 | static int context_equiv(struct perf_event_context *ctx1, |
2207 | struct perf_event_context *ctx2) | |
564c2b21 | 2208 | { |
5a3126d4 PZ |
2209 | /* Pinning disables the swap optimization */ |
2210 | if (ctx1->pin_count || ctx2->pin_count) | |
2211 | return 0; | |
2212 | ||
2213 | /* If ctx1 is the parent of ctx2 */ | |
2214 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) | |
2215 | return 1; | |
2216 | ||
2217 | /* If ctx2 is the parent of ctx1 */ | |
2218 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) | |
2219 | return 1; | |
2220 | ||
2221 | /* | |
2222 | * If ctx1 and ctx2 have the same parent; we flatten the parent | |
2223 | * hierarchy, see perf_event_init_context(). | |
2224 | */ | |
2225 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && | |
2226 | ctx1->parent_gen == ctx2->parent_gen) | |
2227 | return 1; | |
2228 | ||
2229 | /* Unmatched */ | |
2230 | return 0; | |
564c2b21 PM |
2231 | } |
2232 | ||
cdd6c482 IM |
2233 | static void __perf_event_sync_stat(struct perf_event *event, |
2234 | struct perf_event *next_event) | |
bfbd3381 PZ |
2235 | { |
2236 | u64 value; | |
2237 | ||
cdd6c482 | 2238 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2239 | return; |
2240 | ||
2241 | /* | |
cdd6c482 | 2242 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2243 | * because we're in the middle of a context switch and have IRQs |
2244 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2245 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2246 | * don't need to use it. |
2247 | */ | |
cdd6c482 IM |
2248 | switch (event->state) { |
2249 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2250 | event->pmu->read(event); |
2251 | /* fall-through */ | |
bfbd3381 | 2252 | |
cdd6c482 IM |
2253 | case PERF_EVENT_STATE_INACTIVE: |
2254 | update_event_times(event); | |
bfbd3381 PZ |
2255 | break; |
2256 | ||
2257 | default: | |
2258 | break; | |
2259 | } | |
2260 | ||
2261 | /* | |
cdd6c482 | 2262 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2263 | * values when we flip the contexts. |
2264 | */ | |
e7850595 PZ |
2265 | value = local64_read(&next_event->count); |
2266 | value = local64_xchg(&event->count, value); | |
2267 | local64_set(&next_event->count, value); | |
bfbd3381 | 2268 | |
cdd6c482 IM |
2269 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2270 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2271 | |
bfbd3381 | 2272 | /* |
19d2e755 | 2273 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2274 | */ |
cdd6c482 IM |
2275 | perf_event_update_userpage(event); |
2276 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2277 | } |
2278 | ||
cdd6c482 IM |
2279 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2280 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2281 | { |
cdd6c482 | 2282 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2283 | |
2284 | if (!ctx->nr_stat) | |
2285 | return; | |
2286 | ||
02ffdbc8 PZ |
2287 | update_context_time(ctx); |
2288 | ||
cdd6c482 IM |
2289 | event = list_first_entry(&ctx->event_list, |
2290 | struct perf_event, event_entry); | |
bfbd3381 | 2291 | |
cdd6c482 IM |
2292 | next_event = list_first_entry(&next_ctx->event_list, |
2293 | struct perf_event, event_entry); | |
bfbd3381 | 2294 | |
cdd6c482 IM |
2295 | while (&event->event_entry != &ctx->event_list && |
2296 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2297 | |
cdd6c482 | 2298 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2299 | |
cdd6c482 IM |
2300 | event = list_next_entry(event, event_entry); |
2301 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2302 | } |
2303 | } | |
2304 | ||
fe4b04fa PZ |
2305 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2306 | struct task_struct *next) | |
0793a61d | 2307 | { |
8dc85d54 | 2308 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 | 2309 | struct perf_event_context *next_ctx; |
5a3126d4 | 2310 | struct perf_event_context *parent, *next_parent; |
108b02cf | 2311 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2312 | int do_switch = 1; |
0793a61d | 2313 | |
108b02cf PZ |
2314 | if (likely(!ctx)) |
2315 | return; | |
10989fb2 | 2316 | |
108b02cf PZ |
2317 | cpuctx = __get_cpu_context(ctx); |
2318 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2319 | return; |
2320 | ||
c93f7669 | 2321 | rcu_read_lock(); |
8dc85d54 | 2322 | next_ctx = next->perf_event_ctxp[ctxn]; |
5a3126d4 PZ |
2323 | if (!next_ctx) |
2324 | goto unlock; | |
2325 | ||
2326 | parent = rcu_dereference(ctx->parent_ctx); | |
2327 | next_parent = rcu_dereference(next_ctx->parent_ctx); | |
2328 | ||
2329 | /* If neither context have a parent context; they cannot be clones. */ | |
1f9a7268 | 2330 | if (!parent || !next_parent) |
5a3126d4 PZ |
2331 | goto unlock; |
2332 | ||
2333 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { | |
c93f7669 PM |
2334 | /* |
2335 | * Looks like the two contexts are clones, so we might be | |
2336 | * able to optimize the context switch. We lock both | |
2337 | * contexts and check that they are clones under the | |
2338 | * lock (including re-checking that neither has been | |
2339 | * uncloned in the meantime). It doesn't matter which | |
2340 | * order we take the locks because no other cpu could | |
2341 | * be trying to lock both of these tasks. | |
2342 | */ | |
e625cce1 TG |
2343 | raw_spin_lock(&ctx->lock); |
2344 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2345 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2346 | /* |
2347 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2348 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2349 | */ |
8dc85d54 PZ |
2350 | task->perf_event_ctxp[ctxn] = next_ctx; |
2351 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2352 | ctx->task = next; |
2353 | next_ctx->task = task; | |
2354 | do_switch = 0; | |
bfbd3381 | 2355 | |
cdd6c482 | 2356 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2357 | } |
e625cce1 TG |
2358 | raw_spin_unlock(&next_ctx->lock); |
2359 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2360 | } |
5a3126d4 | 2361 | unlock: |
c93f7669 | 2362 | rcu_read_unlock(); |
564c2b21 | 2363 | |
c93f7669 | 2364 | if (do_switch) { |
facc4307 | 2365 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2366 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2367 | cpuctx->task_ctx = NULL; |
facc4307 | 2368 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2369 | } |
0793a61d TG |
2370 | } |
2371 | ||
8dc85d54 PZ |
2372 | #define for_each_task_context_nr(ctxn) \ |
2373 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2374 | ||
2375 | /* | |
2376 | * Called from scheduler to remove the events of the current task, | |
2377 | * with interrupts disabled. | |
2378 | * | |
2379 | * We stop each event and update the event value in event->count. | |
2380 | * | |
2381 | * This does not protect us against NMI, but disable() | |
2382 | * sets the disabled bit in the control field of event _before_ | |
2383 | * accessing the event control register. If a NMI hits, then it will | |
2384 | * not restart the event. | |
2385 | */ | |
ab0cce56 JO |
2386 | void __perf_event_task_sched_out(struct task_struct *task, |
2387 | struct task_struct *next) | |
8dc85d54 PZ |
2388 | { |
2389 | int ctxn; | |
2390 | ||
8dc85d54 PZ |
2391 | for_each_task_context_nr(ctxn) |
2392 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2393 | |
2394 | /* | |
2395 | * if cgroup events exist on this CPU, then we need | |
2396 | * to check if we have to switch out PMU state. | |
2397 | * cgroup event are system-wide mode only | |
2398 | */ | |
2399 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2400 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2401 | } |
2402 | ||
04dc2dbb | 2403 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2404 | { |
108b02cf | 2405 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2406 | |
a63eaf34 PM |
2407 | if (!cpuctx->task_ctx) |
2408 | return; | |
012b84da IM |
2409 | |
2410 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2411 | return; | |
2412 | ||
04dc2dbb | 2413 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2414 | cpuctx->task_ctx = NULL; |
2415 | } | |
2416 | ||
5b0311e1 FW |
2417 | /* |
2418 | * Called with IRQs disabled | |
2419 | */ | |
2420 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2421 | enum event_type_t event_type) | |
2422 | { | |
2423 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2424 | } |
2425 | ||
235c7fc7 | 2426 | static void |
5b0311e1 | 2427 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2428 | struct perf_cpu_context *cpuctx) |
0793a61d | 2429 | { |
cdd6c482 | 2430 | struct perf_event *event; |
0793a61d | 2431 | |
889ff015 FW |
2432 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2433 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2434 | continue; |
5632ab12 | 2435 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2436 | continue; |
2437 | ||
e5d1367f SE |
2438 | /* may need to reset tstamp_enabled */ |
2439 | if (is_cgroup_event(event)) | |
2440 | perf_cgroup_mark_enabled(event, ctx); | |
2441 | ||
8c9ed8e1 | 2442 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2443 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2444 | |
2445 | /* | |
2446 | * If this pinned group hasn't been scheduled, | |
2447 | * put it in error state. | |
2448 | */ | |
cdd6c482 IM |
2449 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2450 | update_group_times(event); | |
2451 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2452 | } |
3b6f9e5c | 2453 | } |
5b0311e1 FW |
2454 | } |
2455 | ||
2456 | static void | |
2457 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2458 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2459 | { |
2460 | struct perf_event *event; | |
2461 | int can_add_hw = 1; | |
3b6f9e5c | 2462 | |
889ff015 FW |
2463 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2464 | /* Ignore events in OFF or ERROR state */ | |
2465 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2466 | continue; |
04289bb9 IM |
2467 | /* |
2468 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2469 | * of events: |
04289bb9 | 2470 | */ |
5632ab12 | 2471 | if (!event_filter_match(event)) |
0793a61d TG |
2472 | continue; |
2473 | ||
e5d1367f SE |
2474 | /* may need to reset tstamp_enabled */ |
2475 | if (is_cgroup_event(event)) | |
2476 | perf_cgroup_mark_enabled(event, ctx); | |
2477 | ||
9ed6060d | 2478 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2479 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2480 | can_add_hw = 0; |
9ed6060d | 2481 | } |
0793a61d | 2482 | } |
5b0311e1 FW |
2483 | } |
2484 | ||
2485 | static void | |
2486 | ctx_sched_in(struct perf_event_context *ctx, | |
2487 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2488 | enum event_type_t event_type, |
2489 | struct task_struct *task) | |
5b0311e1 | 2490 | { |
e5d1367f | 2491 | u64 now; |
db24d33e | 2492 | int is_active = ctx->is_active; |
e5d1367f | 2493 | |
db24d33e | 2494 | ctx->is_active |= event_type; |
5b0311e1 | 2495 | if (likely(!ctx->nr_events)) |
facc4307 | 2496 | return; |
5b0311e1 | 2497 | |
e5d1367f SE |
2498 | now = perf_clock(); |
2499 | ctx->timestamp = now; | |
3f7cce3c | 2500 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2501 | /* |
2502 | * First go through the list and put on any pinned groups | |
2503 | * in order to give them the best chance of going on. | |
2504 | */ | |
db24d33e | 2505 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2506 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2507 | |
2508 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2509 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2510 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2511 | } |
2512 | ||
329c0e01 | 2513 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2514 | enum event_type_t event_type, |
2515 | struct task_struct *task) | |
329c0e01 FW |
2516 | { |
2517 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2518 | ||
e5d1367f | 2519 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2520 | } |
2521 | ||
e5d1367f SE |
2522 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2523 | struct task_struct *task) | |
235c7fc7 | 2524 | { |
108b02cf | 2525 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2526 | |
108b02cf | 2527 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2528 | if (cpuctx->task_ctx == ctx) |
2529 | return; | |
2530 | ||
facc4307 | 2531 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2532 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2533 | /* |
2534 | * We want to keep the following priority order: | |
2535 | * cpu pinned (that don't need to move), task pinned, | |
2536 | * cpu flexible, task flexible. | |
2537 | */ | |
2538 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2539 | ||
1d5f003f GN |
2540 | if (ctx->nr_events) |
2541 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2542 | |
86b47c25 GN |
2543 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2544 | ||
facc4307 PZ |
2545 | perf_pmu_enable(ctx->pmu); |
2546 | perf_ctx_unlock(cpuctx, ctx); | |
2547 | ||
b5ab4cd5 PZ |
2548 | /* |
2549 | * Since these rotations are per-cpu, we need to ensure the | |
2550 | * cpu-context we got scheduled on is actually rotating. | |
2551 | */ | |
108b02cf | 2552 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2553 | } |
2554 | ||
d010b332 SE |
2555 | /* |
2556 | * When sampling the branck stack in system-wide, it may be necessary | |
2557 | * to flush the stack on context switch. This happens when the branch | |
2558 | * stack does not tag its entries with the pid of the current task. | |
2559 | * Otherwise it becomes impossible to associate a branch entry with a | |
2560 | * task. This ambiguity is more likely to appear when the branch stack | |
2561 | * supports priv level filtering and the user sets it to monitor only | |
2562 | * at the user level (which could be a useful measurement in system-wide | |
2563 | * mode). In that case, the risk is high of having a branch stack with | |
2564 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2565 | * entries or stashing them somewhere in the PMU specific code layer. | |
2566 | * | |
2567 | * This function provides the context switch callback to the lower code | |
2568 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2569 | * with at least one active event using taken branch sampling. | |
2570 | */ | |
2571 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2572 | struct task_struct *task) | |
2573 | { | |
2574 | struct perf_cpu_context *cpuctx; | |
2575 | struct pmu *pmu; | |
2576 | unsigned long flags; | |
2577 | ||
2578 | /* no need to flush branch stack if not changing task */ | |
2579 | if (prev == task) | |
2580 | return; | |
2581 | ||
2582 | local_irq_save(flags); | |
2583 | ||
2584 | rcu_read_lock(); | |
2585 | ||
2586 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2587 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2588 | ||
2589 | /* | |
2590 | * check if the context has at least one | |
2591 | * event using PERF_SAMPLE_BRANCH_STACK | |
2592 | */ | |
2593 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2594 | && pmu->flush_branch_stack) { | |
2595 | ||
d010b332 SE |
2596 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
2597 | ||
2598 | perf_pmu_disable(pmu); | |
2599 | ||
2600 | pmu->flush_branch_stack(); | |
2601 | ||
2602 | perf_pmu_enable(pmu); | |
2603 | ||
2604 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2605 | } | |
2606 | } | |
2607 | ||
2608 | rcu_read_unlock(); | |
2609 | ||
2610 | local_irq_restore(flags); | |
2611 | } | |
2612 | ||
8dc85d54 PZ |
2613 | /* |
2614 | * Called from scheduler to add the events of the current task | |
2615 | * with interrupts disabled. | |
2616 | * | |
2617 | * We restore the event value and then enable it. | |
2618 | * | |
2619 | * This does not protect us against NMI, but enable() | |
2620 | * sets the enabled bit in the control field of event _before_ | |
2621 | * accessing the event control register. If a NMI hits, then it will | |
2622 | * keep the event running. | |
2623 | */ | |
ab0cce56 JO |
2624 | void __perf_event_task_sched_in(struct task_struct *prev, |
2625 | struct task_struct *task) | |
8dc85d54 PZ |
2626 | { |
2627 | struct perf_event_context *ctx; | |
2628 | int ctxn; | |
2629 | ||
2630 | for_each_task_context_nr(ctxn) { | |
2631 | ctx = task->perf_event_ctxp[ctxn]; | |
2632 | if (likely(!ctx)) | |
2633 | continue; | |
2634 | ||
e5d1367f | 2635 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2636 | } |
e5d1367f SE |
2637 | /* |
2638 | * if cgroup events exist on this CPU, then we need | |
2639 | * to check if we have to switch in PMU state. | |
2640 | * cgroup event are system-wide mode only | |
2641 | */ | |
2642 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2643 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2644 | |
2645 | /* check for system-wide branch_stack events */ | |
2646 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2647 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2648 | } |
2649 | ||
abd50713 PZ |
2650 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2651 | { | |
2652 | u64 frequency = event->attr.sample_freq; | |
2653 | u64 sec = NSEC_PER_SEC; | |
2654 | u64 divisor, dividend; | |
2655 | ||
2656 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2657 | ||
2658 | count_fls = fls64(count); | |
2659 | nsec_fls = fls64(nsec); | |
2660 | frequency_fls = fls64(frequency); | |
2661 | sec_fls = 30; | |
2662 | ||
2663 | /* | |
2664 | * We got @count in @nsec, with a target of sample_freq HZ | |
2665 | * the target period becomes: | |
2666 | * | |
2667 | * @count * 10^9 | |
2668 | * period = ------------------- | |
2669 | * @nsec * sample_freq | |
2670 | * | |
2671 | */ | |
2672 | ||
2673 | /* | |
2674 | * Reduce accuracy by one bit such that @a and @b converge | |
2675 | * to a similar magnitude. | |
2676 | */ | |
fe4b04fa | 2677 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2678 | do { \ |
2679 | if (a##_fls > b##_fls) { \ | |
2680 | a >>= 1; \ | |
2681 | a##_fls--; \ | |
2682 | } else { \ | |
2683 | b >>= 1; \ | |
2684 | b##_fls--; \ | |
2685 | } \ | |
2686 | } while (0) | |
2687 | ||
2688 | /* | |
2689 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2690 | * the other, so that finally we can do a u64/u64 division. | |
2691 | */ | |
2692 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2693 | REDUCE_FLS(nsec, frequency); | |
2694 | REDUCE_FLS(sec, count); | |
2695 | } | |
2696 | ||
2697 | if (count_fls + sec_fls > 64) { | |
2698 | divisor = nsec * frequency; | |
2699 | ||
2700 | while (count_fls + sec_fls > 64) { | |
2701 | REDUCE_FLS(count, sec); | |
2702 | divisor >>= 1; | |
2703 | } | |
2704 | ||
2705 | dividend = count * sec; | |
2706 | } else { | |
2707 | dividend = count * sec; | |
2708 | ||
2709 | while (nsec_fls + frequency_fls > 64) { | |
2710 | REDUCE_FLS(nsec, frequency); | |
2711 | dividend >>= 1; | |
2712 | } | |
2713 | ||
2714 | divisor = nsec * frequency; | |
2715 | } | |
2716 | ||
f6ab91ad PZ |
2717 | if (!divisor) |
2718 | return dividend; | |
2719 | ||
abd50713 PZ |
2720 | return div64_u64(dividend, divisor); |
2721 | } | |
2722 | ||
e050e3f0 SE |
2723 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2724 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2725 | ||
f39d47ff | 2726 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2727 | { |
cdd6c482 | 2728 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2729 | s64 period, sample_period; |
bd2b5b12 PZ |
2730 | s64 delta; |
2731 | ||
abd50713 | 2732 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2733 | |
2734 | delta = (s64)(period - hwc->sample_period); | |
2735 | delta = (delta + 7) / 8; /* low pass filter */ | |
2736 | ||
2737 | sample_period = hwc->sample_period + delta; | |
2738 | ||
2739 | if (!sample_period) | |
2740 | sample_period = 1; | |
2741 | ||
bd2b5b12 | 2742 | hwc->sample_period = sample_period; |
abd50713 | 2743 | |
e7850595 | 2744 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2745 | if (disable) |
2746 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2747 | ||
e7850595 | 2748 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2749 | |
2750 | if (disable) | |
2751 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2752 | } |
bd2b5b12 PZ |
2753 | } |
2754 | ||
e050e3f0 SE |
2755 | /* |
2756 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2757 | * events. At the same time, make sure, having freq events does not change | |
2758 | * the rate of unthrottling as that would introduce bias. | |
2759 | */ | |
2760 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2761 | int needs_unthr) | |
60db5e09 | 2762 | { |
cdd6c482 IM |
2763 | struct perf_event *event; |
2764 | struct hw_perf_event *hwc; | |
e050e3f0 | 2765 | u64 now, period = TICK_NSEC; |
abd50713 | 2766 | s64 delta; |
60db5e09 | 2767 | |
e050e3f0 SE |
2768 | /* |
2769 | * only need to iterate over all events iff: | |
2770 | * - context have events in frequency mode (needs freq adjust) | |
2771 | * - there are events to unthrottle on this cpu | |
2772 | */ | |
2773 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2774 | return; |
2775 | ||
e050e3f0 | 2776 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2777 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2778 | |
03541f8b | 2779 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2780 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2781 | continue; |
2782 | ||
5632ab12 | 2783 | if (!event_filter_match(event)) |
5d27c23d PZ |
2784 | continue; |
2785 | ||
44377277 AS |
2786 | perf_pmu_disable(event->pmu); |
2787 | ||
cdd6c482 | 2788 | hwc = &event->hw; |
6a24ed6c | 2789 | |
ae23bff1 | 2790 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2791 | hwc->interrupts = 0; |
cdd6c482 | 2792 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2793 | event->pmu->start(event, 0); |
a78ac325 PZ |
2794 | } |
2795 | ||
cdd6c482 | 2796 | if (!event->attr.freq || !event->attr.sample_freq) |
44377277 | 2797 | goto next; |
60db5e09 | 2798 | |
e050e3f0 SE |
2799 | /* |
2800 | * stop the event and update event->count | |
2801 | */ | |
2802 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2803 | ||
e7850595 | 2804 | now = local64_read(&event->count); |
abd50713 PZ |
2805 | delta = now - hwc->freq_count_stamp; |
2806 | hwc->freq_count_stamp = now; | |
60db5e09 | 2807 | |
e050e3f0 SE |
2808 | /* |
2809 | * restart the event | |
2810 | * reload only if value has changed | |
f39d47ff SE |
2811 | * we have stopped the event so tell that |
2812 | * to perf_adjust_period() to avoid stopping it | |
2813 | * twice. | |
e050e3f0 | 2814 | */ |
abd50713 | 2815 | if (delta > 0) |
f39d47ff | 2816 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2817 | |
2818 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
44377277 AS |
2819 | next: |
2820 | perf_pmu_enable(event->pmu); | |
60db5e09 | 2821 | } |
e050e3f0 | 2822 | |
f39d47ff | 2823 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2824 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2825 | } |
2826 | ||
235c7fc7 | 2827 | /* |
cdd6c482 | 2828 | * Round-robin a context's events: |
235c7fc7 | 2829 | */ |
cdd6c482 | 2830 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2831 | { |
dddd3379 TG |
2832 | /* |
2833 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2834 | * disabled by the inheritance code. | |
2835 | */ | |
2836 | if (!ctx->rotate_disable) | |
2837 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2838 | } |
2839 | ||
b5ab4cd5 | 2840 | /* |
e9d2b064 PZ |
2841 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2842 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2843 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2844 | */ |
9e630205 | 2845 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2846 | { |
8dc85d54 | 2847 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2848 | int rotate = 0, remove = 1; |
7fc23a53 | 2849 | |
b5ab4cd5 | 2850 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2851 | remove = 0; |
b5ab4cd5 PZ |
2852 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2853 | rotate = 1; | |
2854 | } | |
235c7fc7 | 2855 | |
8dc85d54 | 2856 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2857 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2858 | remove = 0; |
b5ab4cd5 PZ |
2859 | if (ctx->nr_events != ctx->nr_active) |
2860 | rotate = 1; | |
2861 | } | |
9717e6cd | 2862 | |
e050e3f0 | 2863 | if (!rotate) |
0f5a2601 PZ |
2864 | goto done; |
2865 | ||
facc4307 | 2866 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2867 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2868 | |
e050e3f0 SE |
2869 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2870 | if (ctx) | |
2871 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2872 | |
e050e3f0 SE |
2873 | rotate_ctx(&cpuctx->ctx); |
2874 | if (ctx) | |
2875 | rotate_ctx(ctx); | |
235c7fc7 | 2876 | |
e050e3f0 | 2877 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2878 | |
0f5a2601 PZ |
2879 | perf_pmu_enable(cpuctx->ctx.pmu); |
2880 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2881 | done: |
e9d2b064 PZ |
2882 | if (remove) |
2883 | list_del_init(&cpuctx->rotation_list); | |
9e630205 SE |
2884 | |
2885 | return rotate; | |
e9d2b064 PZ |
2886 | } |
2887 | ||
026249ef FW |
2888 | #ifdef CONFIG_NO_HZ_FULL |
2889 | bool perf_event_can_stop_tick(void) | |
2890 | { | |
948b26b6 | 2891 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 2892 | __this_cpu_read(perf_throttled_count)) |
026249ef | 2893 | return false; |
d84153d6 FW |
2894 | else |
2895 | return true; | |
026249ef FW |
2896 | } |
2897 | #endif | |
2898 | ||
e9d2b064 PZ |
2899 | void perf_event_task_tick(void) |
2900 | { | |
2901 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2902 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2903 | struct perf_event_context *ctx; |
2904 | int throttled; | |
b5ab4cd5 | 2905 | |
e9d2b064 PZ |
2906 | WARN_ON(!irqs_disabled()); |
2907 | ||
e050e3f0 SE |
2908 | __this_cpu_inc(perf_throttled_seq); |
2909 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2910 | ||
e9d2b064 | 2911 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2912 | ctx = &cpuctx->ctx; |
2913 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2914 | ||
2915 | ctx = cpuctx->task_ctx; | |
2916 | if (ctx) | |
2917 | perf_adjust_freq_unthr_context(ctx, throttled); | |
e9d2b064 | 2918 | } |
0793a61d TG |
2919 | } |
2920 | ||
889ff015 FW |
2921 | static int event_enable_on_exec(struct perf_event *event, |
2922 | struct perf_event_context *ctx) | |
2923 | { | |
2924 | if (!event->attr.enable_on_exec) | |
2925 | return 0; | |
2926 | ||
2927 | event->attr.enable_on_exec = 0; | |
2928 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2929 | return 0; | |
2930 | ||
1d9b482e | 2931 | __perf_event_mark_enabled(event); |
889ff015 FW |
2932 | |
2933 | return 1; | |
2934 | } | |
2935 | ||
57e7986e | 2936 | /* |
cdd6c482 | 2937 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2938 | * This expects task == current. |
2939 | */ | |
8dc85d54 | 2940 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2941 | { |
cdd6c482 | 2942 | struct perf_event *event; |
57e7986e PM |
2943 | unsigned long flags; |
2944 | int enabled = 0; | |
889ff015 | 2945 | int ret; |
57e7986e PM |
2946 | |
2947 | local_irq_save(flags); | |
cdd6c482 | 2948 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2949 | goto out; |
2950 | ||
e566b76e SE |
2951 | /* |
2952 | * We must ctxsw out cgroup events to avoid conflict | |
2953 | * when invoking perf_task_event_sched_in() later on | |
2954 | * in this function. Otherwise we end up trying to | |
2955 | * ctxswin cgroup events which are already scheduled | |
2956 | * in. | |
2957 | */ | |
a8d757ef | 2958 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2959 | |
e625cce1 | 2960 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2961 | task_ctx_sched_out(ctx); |
57e7986e | 2962 | |
b79387ef | 2963 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2964 | ret = event_enable_on_exec(event, ctx); |
2965 | if (ret) | |
2966 | enabled = 1; | |
57e7986e PM |
2967 | } |
2968 | ||
2969 | /* | |
cdd6c482 | 2970 | * Unclone this context if we enabled any event. |
57e7986e | 2971 | */ |
71a851b4 PZ |
2972 | if (enabled) |
2973 | unclone_ctx(ctx); | |
57e7986e | 2974 | |
e625cce1 | 2975 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2976 | |
e566b76e SE |
2977 | /* |
2978 | * Also calls ctxswin for cgroup events, if any: | |
2979 | */ | |
e5d1367f | 2980 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2981 | out: |
57e7986e PM |
2982 | local_irq_restore(flags); |
2983 | } | |
2984 | ||
e041e328 PZ |
2985 | void perf_event_exec(void) |
2986 | { | |
2987 | struct perf_event_context *ctx; | |
2988 | int ctxn; | |
2989 | ||
2990 | rcu_read_lock(); | |
2991 | for_each_task_context_nr(ctxn) { | |
2992 | ctx = current->perf_event_ctxp[ctxn]; | |
2993 | if (!ctx) | |
2994 | continue; | |
2995 | ||
2996 | perf_event_enable_on_exec(ctx); | |
2997 | } | |
2998 | rcu_read_unlock(); | |
2999 | } | |
3000 | ||
0793a61d | 3001 | /* |
cdd6c482 | 3002 | * Cross CPU call to read the hardware event |
0793a61d | 3003 | */ |
cdd6c482 | 3004 | static void __perf_event_read(void *info) |
0793a61d | 3005 | { |
cdd6c482 IM |
3006 | struct perf_event *event = info; |
3007 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 3008 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 3009 | |
e1ac3614 PM |
3010 | /* |
3011 | * If this is a task context, we need to check whether it is | |
3012 | * the current task context of this cpu. If not it has been | |
3013 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
3014 | * event->count would have been updated to a recent sample |
3015 | * when the event was scheduled out. | |
e1ac3614 PM |
3016 | */ |
3017 | if (ctx->task && cpuctx->task_ctx != ctx) | |
3018 | return; | |
3019 | ||
e625cce1 | 3020 | raw_spin_lock(&ctx->lock); |
e5d1367f | 3021 | if (ctx->is_active) { |
542e72fc | 3022 | update_context_time(ctx); |
e5d1367f SE |
3023 | update_cgrp_time_from_event(event); |
3024 | } | |
cdd6c482 | 3025 | update_event_times(event); |
542e72fc PZ |
3026 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
3027 | event->pmu->read(event); | |
e625cce1 | 3028 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
3029 | } |
3030 | ||
b5e58793 PZ |
3031 | static inline u64 perf_event_count(struct perf_event *event) |
3032 | { | |
e7850595 | 3033 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
3034 | } |
3035 | ||
cdd6c482 | 3036 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
3037 | { |
3038 | /* | |
cdd6c482 IM |
3039 | * If event is enabled and currently active on a CPU, update the |
3040 | * value in the event structure: | |
0793a61d | 3041 | */ |
cdd6c482 IM |
3042 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
3043 | smp_call_function_single(event->oncpu, | |
3044 | __perf_event_read, event, 1); | |
3045 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
3046 | struct perf_event_context *ctx = event->ctx; |
3047 | unsigned long flags; | |
3048 | ||
e625cce1 | 3049 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
3050 | /* |
3051 | * may read while context is not active | |
3052 | * (e.g., thread is blocked), in that case | |
3053 | * we cannot update context time | |
3054 | */ | |
e5d1367f | 3055 | if (ctx->is_active) { |
c530ccd9 | 3056 | update_context_time(ctx); |
e5d1367f SE |
3057 | update_cgrp_time_from_event(event); |
3058 | } | |
cdd6c482 | 3059 | update_event_times(event); |
e625cce1 | 3060 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
3061 | } |
3062 | ||
b5e58793 | 3063 | return perf_event_count(event); |
0793a61d TG |
3064 | } |
3065 | ||
a63eaf34 | 3066 | /* |
cdd6c482 | 3067 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 3068 | */ |
eb184479 | 3069 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 3070 | { |
e625cce1 | 3071 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 3072 | mutex_init(&ctx->mutex); |
889ff015 FW |
3073 | INIT_LIST_HEAD(&ctx->pinned_groups); |
3074 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
3075 | INIT_LIST_HEAD(&ctx->event_list); |
3076 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
3077 | } |
3078 | ||
3079 | static struct perf_event_context * | |
3080 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
3081 | { | |
3082 | struct perf_event_context *ctx; | |
3083 | ||
3084 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
3085 | if (!ctx) | |
3086 | return NULL; | |
3087 | ||
3088 | __perf_event_init_context(ctx); | |
3089 | if (task) { | |
3090 | ctx->task = task; | |
3091 | get_task_struct(task); | |
0793a61d | 3092 | } |
eb184479 PZ |
3093 | ctx->pmu = pmu; |
3094 | ||
3095 | return ctx; | |
a63eaf34 PM |
3096 | } |
3097 | ||
2ebd4ffb MH |
3098 | static struct task_struct * |
3099 | find_lively_task_by_vpid(pid_t vpid) | |
3100 | { | |
3101 | struct task_struct *task; | |
3102 | int err; | |
0793a61d TG |
3103 | |
3104 | rcu_read_lock(); | |
2ebd4ffb | 3105 | if (!vpid) |
0793a61d TG |
3106 | task = current; |
3107 | else | |
2ebd4ffb | 3108 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3109 | if (task) |
3110 | get_task_struct(task); | |
3111 | rcu_read_unlock(); | |
3112 | ||
3113 | if (!task) | |
3114 | return ERR_PTR(-ESRCH); | |
3115 | ||
0793a61d | 3116 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3117 | err = -EACCES; |
3118 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3119 | goto errout; | |
3120 | ||
2ebd4ffb MH |
3121 | return task; |
3122 | errout: | |
3123 | put_task_struct(task); | |
3124 | return ERR_PTR(err); | |
3125 | ||
3126 | } | |
3127 | ||
fe4b04fa PZ |
3128 | /* |
3129 | * Returns a matching context with refcount and pincount. | |
3130 | */ | |
108b02cf | 3131 | static struct perf_event_context * |
38a81da2 | 3132 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 3133 | { |
cdd6c482 | 3134 | struct perf_event_context *ctx; |
22a4f650 | 3135 | struct perf_cpu_context *cpuctx; |
25346b93 | 3136 | unsigned long flags; |
8dc85d54 | 3137 | int ctxn, err; |
0793a61d | 3138 | |
22a4ec72 | 3139 | if (!task) { |
cdd6c482 | 3140 | /* Must be root to operate on a CPU event: */ |
0764771d | 3141 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3142 | return ERR_PTR(-EACCES); |
3143 | ||
0793a61d | 3144 | /* |
cdd6c482 | 3145 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3146 | * offline CPU and activate it when the CPU comes up, but |
3147 | * that's for later. | |
3148 | */ | |
f6325e30 | 3149 | if (!cpu_online(cpu)) |
0793a61d TG |
3150 | return ERR_PTR(-ENODEV); |
3151 | ||
108b02cf | 3152 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3153 | ctx = &cpuctx->ctx; |
c93f7669 | 3154 | get_ctx(ctx); |
fe4b04fa | 3155 | ++ctx->pin_count; |
0793a61d | 3156 | |
0793a61d TG |
3157 | return ctx; |
3158 | } | |
3159 | ||
8dc85d54 PZ |
3160 | err = -EINVAL; |
3161 | ctxn = pmu->task_ctx_nr; | |
3162 | if (ctxn < 0) | |
3163 | goto errout; | |
3164 | ||
9ed6060d | 3165 | retry: |
8dc85d54 | 3166 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3167 | if (ctx) { |
71a851b4 | 3168 | unclone_ctx(ctx); |
fe4b04fa | 3169 | ++ctx->pin_count; |
e625cce1 | 3170 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 3171 | } else { |
eb184479 | 3172 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3173 | err = -ENOMEM; |
3174 | if (!ctx) | |
3175 | goto errout; | |
eb184479 | 3176 | |
dbe08d82 ON |
3177 | err = 0; |
3178 | mutex_lock(&task->perf_event_mutex); | |
3179 | /* | |
3180 | * If it has already passed perf_event_exit_task(). | |
3181 | * we must see PF_EXITING, it takes this mutex too. | |
3182 | */ | |
3183 | if (task->flags & PF_EXITING) | |
3184 | err = -ESRCH; | |
3185 | else if (task->perf_event_ctxp[ctxn]) | |
3186 | err = -EAGAIN; | |
fe4b04fa | 3187 | else { |
9137fb28 | 3188 | get_ctx(ctx); |
fe4b04fa | 3189 | ++ctx->pin_count; |
dbe08d82 | 3190 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3191 | } |
dbe08d82 ON |
3192 | mutex_unlock(&task->perf_event_mutex); |
3193 | ||
3194 | if (unlikely(err)) { | |
9137fb28 | 3195 | put_ctx(ctx); |
dbe08d82 ON |
3196 | |
3197 | if (err == -EAGAIN) | |
3198 | goto retry; | |
3199 | goto errout; | |
a63eaf34 PM |
3200 | } |
3201 | } | |
3202 | ||
0793a61d | 3203 | return ctx; |
c93f7669 | 3204 | |
9ed6060d | 3205 | errout: |
c93f7669 | 3206 | return ERR_PTR(err); |
0793a61d TG |
3207 | } |
3208 | ||
6fb2915d LZ |
3209 | static void perf_event_free_filter(struct perf_event *event); |
3210 | ||
cdd6c482 | 3211 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3212 | { |
cdd6c482 | 3213 | struct perf_event *event; |
592903cd | 3214 | |
cdd6c482 IM |
3215 | event = container_of(head, struct perf_event, rcu_head); |
3216 | if (event->ns) | |
3217 | put_pid_ns(event->ns); | |
6fb2915d | 3218 | perf_event_free_filter(event); |
cdd6c482 | 3219 | kfree(event); |
592903cd PZ |
3220 | } |
3221 | ||
76369139 | 3222 | static void ring_buffer_put(struct ring_buffer *rb); |
b69cf536 PZ |
3223 | static void ring_buffer_attach(struct perf_event *event, |
3224 | struct ring_buffer *rb); | |
925d519a | 3225 | |
4beb31f3 | 3226 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3227 | { |
4beb31f3 FW |
3228 | if (event->parent) |
3229 | return; | |
3230 | ||
3231 | if (has_branch_stack(event)) { | |
3232 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
3233 | atomic_dec(&per_cpu(perf_branch_stack_events, cpu)); | |
3234 | } | |
3235 | if (is_cgroup_event(event)) | |
3236 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3237 | } | |
925d519a | 3238 | |
4beb31f3 FW |
3239 | static void unaccount_event(struct perf_event *event) |
3240 | { | |
3241 | if (event->parent) | |
3242 | return; | |
3243 | ||
3244 | if (event->attach_state & PERF_ATTACH_TASK) | |
3245 | static_key_slow_dec_deferred(&perf_sched_events); | |
3246 | if (event->attr.mmap || event->attr.mmap_data) | |
3247 | atomic_dec(&nr_mmap_events); | |
3248 | if (event->attr.comm) | |
3249 | atomic_dec(&nr_comm_events); | |
3250 | if (event->attr.task) | |
3251 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3252 | if (event->attr.freq) |
3253 | atomic_dec(&nr_freq_events); | |
4beb31f3 FW |
3254 | if (is_cgroup_event(event)) |
3255 | static_key_slow_dec_deferred(&perf_sched_events); | |
3256 | if (has_branch_stack(event)) | |
3257 | static_key_slow_dec_deferred(&perf_sched_events); | |
3258 | ||
3259 | unaccount_event_cpu(event, event->cpu); | |
3260 | } | |
925d519a | 3261 | |
766d6c07 FW |
3262 | static void __free_event(struct perf_event *event) |
3263 | { | |
cdd6c482 | 3264 | if (!event->parent) { |
927c7a9e FW |
3265 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3266 | put_callchain_buffers(); | |
f344011c | 3267 | } |
9ee318a7 | 3268 | |
766d6c07 FW |
3269 | if (event->destroy) |
3270 | event->destroy(event); | |
3271 | ||
3272 | if (event->ctx) | |
3273 | put_ctx(event->ctx); | |
3274 | ||
c464c76e YZ |
3275 | if (event->pmu) |
3276 | module_put(event->pmu->module); | |
3277 | ||
766d6c07 FW |
3278 | call_rcu(&event->rcu_head, free_event_rcu); |
3279 | } | |
683ede43 PZ |
3280 | |
3281 | static void _free_event(struct perf_event *event) | |
f1600952 | 3282 | { |
e360adbe | 3283 | irq_work_sync(&event->pending); |
925d519a | 3284 | |
4beb31f3 | 3285 | unaccount_event(event); |
9ee318a7 | 3286 | |
76369139 | 3287 | if (event->rb) { |
9bb5d40c PZ |
3288 | /* |
3289 | * Can happen when we close an event with re-directed output. | |
3290 | * | |
3291 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3292 | * over us; possibly making our ring_buffer_put() the last. | |
3293 | */ | |
3294 | mutex_lock(&event->mmap_mutex); | |
b69cf536 | 3295 | ring_buffer_attach(event, NULL); |
9bb5d40c | 3296 | mutex_unlock(&event->mmap_mutex); |
a4be7c27 PZ |
3297 | } |
3298 | ||
e5d1367f SE |
3299 | if (is_cgroup_event(event)) |
3300 | perf_detach_cgroup(event); | |
3301 | ||
766d6c07 | 3302 | __free_event(event); |
f1600952 PZ |
3303 | } |
3304 | ||
683ede43 PZ |
3305 | /* |
3306 | * Used to free events which have a known refcount of 1, such as in error paths | |
3307 | * where the event isn't exposed yet and inherited events. | |
3308 | */ | |
3309 | static void free_event(struct perf_event *event) | |
0793a61d | 3310 | { |
683ede43 PZ |
3311 | if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, |
3312 | "unexpected event refcount: %ld; ptr=%p\n", | |
3313 | atomic_long_read(&event->refcount), event)) { | |
3314 | /* leak to avoid use-after-free */ | |
3315 | return; | |
3316 | } | |
0793a61d | 3317 | |
683ede43 | 3318 | _free_event(event); |
0793a61d TG |
3319 | } |
3320 | ||
a66a3052 | 3321 | /* |
f8697762 | 3322 | * Remove user event from the owner task. |
a66a3052 | 3323 | */ |
f8697762 | 3324 | static void perf_remove_from_owner(struct perf_event *event) |
fb0459d7 | 3325 | { |
8882135b | 3326 | struct task_struct *owner; |
fb0459d7 | 3327 | |
8882135b PZ |
3328 | rcu_read_lock(); |
3329 | owner = ACCESS_ONCE(event->owner); | |
3330 | /* | |
3331 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3332 | * !owner it means the list deletion is complete and we can indeed | |
3333 | * free this event, otherwise we need to serialize on | |
3334 | * owner->perf_event_mutex. | |
3335 | */ | |
3336 | smp_read_barrier_depends(); | |
3337 | if (owner) { | |
3338 | /* | |
3339 | * Since delayed_put_task_struct() also drops the last | |
3340 | * task reference we can safely take a new reference | |
3341 | * while holding the rcu_read_lock(). | |
3342 | */ | |
3343 | get_task_struct(owner); | |
3344 | } | |
3345 | rcu_read_unlock(); | |
3346 | ||
3347 | if (owner) { | |
3348 | mutex_lock(&owner->perf_event_mutex); | |
3349 | /* | |
3350 | * We have to re-check the event->owner field, if it is cleared | |
3351 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3352 | * ensured they're done, and we can proceed with freeing the | |
3353 | * event. | |
3354 | */ | |
3355 | if (event->owner) | |
3356 | list_del_init(&event->owner_entry); | |
3357 | mutex_unlock(&owner->perf_event_mutex); | |
3358 | put_task_struct(owner); | |
3359 | } | |
f8697762 JO |
3360 | } |
3361 | ||
3362 | /* | |
3363 | * Called when the last reference to the file is gone. | |
3364 | */ | |
3365 | static void put_event(struct perf_event *event) | |
3366 | { | |
3367 | struct perf_event_context *ctx = event->ctx; | |
3368 | ||
3369 | if (!atomic_long_dec_and_test(&event->refcount)) | |
3370 | return; | |
3371 | ||
3372 | if (!is_kernel_event(event)) | |
3373 | perf_remove_from_owner(event); | |
8882135b | 3374 | |
683ede43 PZ |
3375 | WARN_ON_ONCE(ctx->parent_ctx); |
3376 | /* | |
3377 | * There are two ways this annotation is useful: | |
3378 | * | |
3379 | * 1) there is a lock recursion from perf_event_exit_task | |
3380 | * see the comment there. | |
3381 | * | |
3382 | * 2) there is a lock-inversion with mmap_sem through | |
3383 | * perf_event_read_group(), which takes faults while | |
3384 | * holding ctx->mutex, however this is called after | |
3385 | * the last filedesc died, so there is no possibility | |
3386 | * to trigger the AB-BA case. | |
3387 | */ | |
3388 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
3389 | perf_remove_from_context(event, true); | |
3390 | mutex_unlock(&ctx->mutex); | |
3391 | ||
3392 | _free_event(event); | |
a6fa941d AV |
3393 | } |
3394 | ||
683ede43 PZ |
3395 | int perf_event_release_kernel(struct perf_event *event) |
3396 | { | |
3397 | put_event(event); | |
3398 | return 0; | |
3399 | } | |
3400 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); | |
3401 | ||
a6fa941d AV |
3402 | static int perf_release(struct inode *inode, struct file *file) |
3403 | { | |
3404 | put_event(file->private_data); | |
3405 | return 0; | |
fb0459d7 | 3406 | } |
fb0459d7 | 3407 | |
59ed446f | 3408 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3409 | { |
cdd6c482 | 3410 | struct perf_event *child; |
e53c0994 PZ |
3411 | u64 total = 0; |
3412 | ||
59ed446f PZ |
3413 | *enabled = 0; |
3414 | *running = 0; | |
3415 | ||
6f10581a | 3416 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3417 | total += perf_event_read(event); |
59ed446f PZ |
3418 | *enabled += event->total_time_enabled + |
3419 | atomic64_read(&event->child_total_time_enabled); | |
3420 | *running += event->total_time_running + | |
3421 | atomic64_read(&event->child_total_time_running); | |
3422 | ||
3423 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3424 | total += perf_event_read(child); |
59ed446f PZ |
3425 | *enabled += child->total_time_enabled; |
3426 | *running += child->total_time_running; | |
3427 | } | |
6f10581a | 3428 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3429 | |
3430 | return total; | |
3431 | } | |
fb0459d7 | 3432 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3433 | |
cdd6c482 | 3434 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3435 | u64 read_format, char __user *buf) |
3436 | { | |
cdd6c482 | 3437 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3438 | int n = 0, size = 0, ret = -EFAULT; |
3439 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3440 | u64 values[5]; |
59ed446f | 3441 | u64 count, enabled, running; |
abf4868b | 3442 | |
6f10581a | 3443 | mutex_lock(&ctx->mutex); |
59ed446f | 3444 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3445 | |
3446 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3447 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3448 | values[n++] = enabled; | |
3449 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3450 | values[n++] = running; | |
abf4868b PZ |
3451 | values[n++] = count; |
3452 | if (read_format & PERF_FORMAT_ID) | |
3453 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3454 | |
3455 | size = n * sizeof(u64); | |
3456 | ||
3457 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3458 | goto unlock; |
3dab77fb | 3459 | |
6f10581a | 3460 | ret = size; |
3dab77fb | 3461 | |
65abc865 | 3462 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3463 | n = 0; |
3dab77fb | 3464 | |
59ed446f | 3465 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3466 | if (read_format & PERF_FORMAT_ID) |
3467 | values[n++] = primary_event_id(sub); | |
3468 | ||
3469 | size = n * sizeof(u64); | |
3470 | ||
184d3da8 | 3471 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3472 | ret = -EFAULT; |
3473 | goto unlock; | |
3474 | } | |
abf4868b PZ |
3475 | |
3476 | ret += size; | |
3dab77fb | 3477 | } |
6f10581a PZ |
3478 | unlock: |
3479 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3480 | |
abf4868b | 3481 | return ret; |
3dab77fb PZ |
3482 | } |
3483 | ||
cdd6c482 | 3484 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3485 | u64 read_format, char __user *buf) |
3486 | { | |
59ed446f | 3487 | u64 enabled, running; |
3dab77fb PZ |
3488 | u64 values[4]; |
3489 | int n = 0; | |
3490 | ||
59ed446f PZ |
3491 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3492 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3493 | values[n++] = enabled; | |
3494 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3495 | values[n++] = running; | |
3dab77fb | 3496 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3497 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3498 | |
3499 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3500 | return -EFAULT; | |
3501 | ||
3502 | return n * sizeof(u64); | |
3503 | } | |
3504 | ||
0793a61d | 3505 | /* |
cdd6c482 | 3506 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3507 | */ |
3508 | static ssize_t | |
cdd6c482 | 3509 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3510 | { |
cdd6c482 | 3511 | u64 read_format = event->attr.read_format; |
3dab77fb | 3512 | int ret; |
0793a61d | 3513 | |
3b6f9e5c | 3514 | /* |
cdd6c482 | 3515 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3516 | * error state (i.e. because it was pinned but it couldn't be |
3517 | * scheduled on to the CPU at some point). | |
3518 | */ | |
cdd6c482 | 3519 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3520 | return 0; |
3521 | ||
c320c7b7 | 3522 | if (count < event->read_size) |
3dab77fb PZ |
3523 | return -ENOSPC; |
3524 | ||
cdd6c482 | 3525 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3526 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3527 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3528 | else |
cdd6c482 | 3529 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3530 | |
3dab77fb | 3531 | return ret; |
0793a61d TG |
3532 | } |
3533 | ||
0793a61d TG |
3534 | static ssize_t |
3535 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3536 | { | |
cdd6c482 | 3537 | struct perf_event *event = file->private_data; |
0793a61d | 3538 | |
cdd6c482 | 3539 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3540 | } |
3541 | ||
3542 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3543 | { | |
cdd6c482 | 3544 | struct perf_event *event = file->private_data; |
76369139 | 3545 | struct ring_buffer *rb; |
c33a0bc4 | 3546 | unsigned int events = POLL_HUP; |
c7138f37 | 3547 | |
10c6db11 | 3548 | /* |
9bb5d40c PZ |
3549 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3550 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3551 | */ |
3552 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3553 | rb = event->rb; |
3554 | if (rb) | |
76369139 | 3555 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 PZ |
3556 | mutex_unlock(&event->mmap_mutex); |
3557 | ||
cdd6c482 | 3558 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3559 | |
0793a61d TG |
3560 | return events; |
3561 | } | |
3562 | ||
cdd6c482 | 3563 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3564 | { |
cdd6c482 | 3565 | (void)perf_event_read(event); |
e7850595 | 3566 | local64_set(&event->count, 0); |
cdd6c482 | 3567 | perf_event_update_userpage(event); |
3df5edad PZ |
3568 | } |
3569 | ||
c93f7669 | 3570 | /* |
cdd6c482 IM |
3571 | * Holding the top-level event's child_mutex means that any |
3572 | * descendant process that has inherited this event will block | |
3573 | * in sync_child_event if it goes to exit, thus satisfying the | |
3574 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3575 | */ |
cdd6c482 IM |
3576 | static void perf_event_for_each_child(struct perf_event *event, |
3577 | void (*func)(struct perf_event *)) | |
3df5edad | 3578 | { |
cdd6c482 | 3579 | struct perf_event *child; |
3df5edad | 3580 | |
cdd6c482 IM |
3581 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3582 | mutex_lock(&event->child_mutex); | |
3583 | func(event); | |
3584 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3585 | func(child); |
cdd6c482 | 3586 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3587 | } |
3588 | ||
cdd6c482 IM |
3589 | static void perf_event_for_each(struct perf_event *event, |
3590 | void (*func)(struct perf_event *)) | |
3df5edad | 3591 | { |
cdd6c482 IM |
3592 | struct perf_event_context *ctx = event->ctx; |
3593 | struct perf_event *sibling; | |
3df5edad | 3594 | |
75f937f2 PZ |
3595 | WARN_ON_ONCE(ctx->parent_ctx); |
3596 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3597 | event = event->group_leader; |
75f937f2 | 3598 | |
cdd6c482 | 3599 | perf_event_for_each_child(event, func); |
cdd6c482 | 3600 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3601 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3602 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3603 | } |
3604 | ||
cdd6c482 | 3605 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3606 | { |
cdd6c482 | 3607 | struct perf_event_context *ctx = event->ctx; |
bad7192b | 3608 | int ret = 0, active; |
08247e31 PZ |
3609 | u64 value; |
3610 | ||
6c7e550f | 3611 | if (!is_sampling_event(event)) |
08247e31 PZ |
3612 | return -EINVAL; |
3613 | ||
ad0cf347 | 3614 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3615 | return -EFAULT; |
3616 | ||
3617 | if (!value) | |
3618 | return -EINVAL; | |
3619 | ||
e625cce1 | 3620 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3621 | if (event->attr.freq) { |
3622 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3623 | ret = -EINVAL; |
3624 | goto unlock; | |
3625 | } | |
3626 | ||
cdd6c482 | 3627 | event->attr.sample_freq = value; |
08247e31 | 3628 | } else { |
cdd6c482 IM |
3629 | event->attr.sample_period = value; |
3630 | event->hw.sample_period = value; | |
08247e31 | 3631 | } |
bad7192b PZ |
3632 | |
3633 | active = (event->state == PERF_EVENT_STATE_ACTIVE); | |
3634 | if (active) { | |
3635 | perf_pmu_disable(ctx->pmu); | |
3636 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3637 | } | |
3638 | ||
3639 | local64_set(&event->hw.period_left, 0); | |
3640 | ||
3641 | if (active) { | |
3642 | event->pmu->start(event, PERF_EF_RELOAD); | |
3643 | perf_pmu_enable(ctx->pmu); | |
3644 | } | |
3645 | ||
08247e31 | 3646 | unlock: |
e625cce1 | 3647 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3648 | |
3649 | return ret; | |
3650 | } | |
3651 | ||
ac9721f3 PZ |
3652 | static const struct file_operations perf_fops; |
3653 | ||
2903ff01 | 3654 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3655 | { |
2903ff01 AV |
3656 | struct fd f = fdget(fd); |
3657 | if (!f.file) | |
3658 | return -EBADF; | |
ac9721f3 | 3659 | |
2903ff01 AV |
3660 | if (f.file->f_op != &perf_fops) { |
3661 | fdput(f); | |
3662 | return -EBADF; | |
ac9721f3 | 3663 | } |
2903ff01 AV |
3664 | *p = f; |
3665 | return 0; | |
ac9721f3 PZ |
3666 | } |
3667 | ||
3668 | static int perf_event_set_output(struct perf_event *event, | |
3669 | struct perf_event *output_event); | |
6fb2915d | 3670 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3671 | |
d859e29f PM |
3672 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3673 | { | |
cdd6c482 IM |
3674 | struct perf_event *event = file->private_data; |
3675 | void (*func)(struct perf_event *); | |
3df5edad | 3676 | u32 flags = arg; |
d859e29f PM |
3677 | |
3678 | switch (cmd) { | |
cdd6c482 IM |
3679 | case PERF_EVENT_IOC_ENABLE: |
3680 | func = perf_event_enable; | |
d859e29f | 3681 | break; |
cdd6c482 IM |
3682 | case PERF_EVENT_IOC_DISABLE: |
3683 | func = perf_event_disable; | |
79f14641 | 3684 | break; |
cdd6c482 IM |
3685 | case PERF_EVENT_IOC_RESET: |
3686 | func = perf_event_reset; | |
6de6a7b9 | 3687 | break; |
3df5edad | 3688 | |
cdd6c482 IM |
3689 | case PERF_EVENT_IOC_REFRESH: |
3690 | return perf_event_refresh(event, arg); | |
08247e31 | 3691 | |
cdd6c482 IM |
3692 | case PERF_EVENT_IOC_PERIOD: |
3693 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3694 | |
cf4957f1 JO |
3695 | case PERF_EVENT_IOC_ID: |
3696 | { | |
3697 | u64 id = primary_event_id(event); | |
3698 | ||
3699 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
3700 | return -EFAULT; | |
3701 | return 0; | |
3702 | } | |
3703 | ||
cdd6c482 | 3704 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3705 | { |
ac9721f3 | 3706 | int ret; |
ac9721f3 | 3707 | if (arg != -1) { |
2903ff01 AV |
3708 | struct perf_event *output_event; |
3709 | struct fd output; | |
3710 | ret = perf_fget_light(arg, &output); | |
3711 | if (ret) | |
3712 | return ret; | |
3713 | output_event = output.file->private_data; | |
3714 | ret = perf_event_set_output(event, output_event); | |
3715 | fdput(output); | |
3716 | } else { | |
3717 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3718 | } |
ac9721f3 PZ |
3719 | return ret; |
3720 | } | |
a4be7c27 | 3721 | |
6fb2915d LZ |
3722 | case PERF_EVENT_IOC_SET_FILTER: |
3723 | return perf_event_set_filter(event, (void __user *)arg); | |
3724 | ||
d859e29f | 3725 | default: |
3df5edad | 3726 | return -ENOTTY; |
d859e29f | 3727 | } |
3df5edad PZ |
3728 | |
3729 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3730 | perf_event_for_each(event, func); |
3df5edad | 3731 | else |
cdd6c482 | 3732 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3733 | |
3734 | return 0; | |
d859e29f PM |
3735 | } |
3736 | ||
cdd6c482 | 3737 | int perf_event_task_enable(void) |
771d7cde | 3738 | { |
cdd6c482 | 3739 | struct perf_event *event; |
771d7cde | 3740 | |
cdd6c482 IM |
3741 | mutex_lock(¤t->perf_event_mutex); |
3742 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3743 | perf_event_for_each_child(event, perf_event_enable); | |
3744 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3745 | |
3746 | return 0; | |
3747 | } | |
3748 | ||
cdd6c482 | 3749 | int perf_event_task_disable(void) |
771d7cde | 3750 | { |
cdd6c482 | 3751 | struct perf_event *event; |
771d7cde | 3752 | |
cdd6c482 IM |
3753 | mutex_lock(¤t->perf_event_mutex); |
3754 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3755 | perf_event_for_each_child(event, perf_event_disable); | |
3756 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3757 | |
3758 | return 0; | |
3759 | } | |
3760 | ||
cdd6c482 | 3761 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3762 | { |
a4eaf7f1 PZ |
3763 | if (event->hw.state & PERF_HES_STOPPED) |
3764 | return 0; | |
3765 | ||
cdd6c482 | 3766 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3767 | return 0; |
3768 | ||
35edc2a5 | 3769 | return event->pmu->event_idx(event); |
194002b2 PZ |
3770 | } |
3771 | ||
c4794295 | 3772 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3773 | u64 *now, |
7f310a5d EM |
3774 | u64 *enabled, |
3775 | u64 *running) | |
c4794295 | 3776 | { |
e3f3541c | 3777 | u64 ctx_time; |
c4794295 | 3778 | |
e3f3541c PZ |
3779 | *now = perf_clock(); |
3780 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3781 | *enabled = ctx_time - event->tstamp_enabled; |
3782 | *running = ctx_time - event->tstamp_running; | |
3783 | } | |
3784 | ||
fa731587 PZ |
3785 | static void perf_event_init_userpage(struct perf_event *event) |
3786 | { | |
3787 | struct perf_event_mmap_page *userpg; | |
3788 | struct ring_buffer *rb; | |
3789 | ||
3790 | rcu_read_lock(); | |
3791 | rb = rcu_dereference(event->rb); | |
3792 | if (!rb) | |
3793 | goto unlock; | |
3794 | ||
3795 | userpg = rb->user_page; | |
3796 | ||
3797 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
3798 | userpg->cap_bit0_is_deprecated = 1; | |
3799 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
3800 | ||
3801 | unlock: | |
3802 | rcu_read_unlock(); | |
3803 | } | |
3804 | ||
c7206205 | 3805 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3806 | { |
3807 | } | |
3808 | ||
38ff667b PZ |
3809 | /* |
3810 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3811 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3812 | * code calls this from NMI context. | |
3813 | */ | |
cdd6c482 | 3814 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3815 | { |
cdd6c482 | 3816 | struct perf_event_mmap_page *userpg; |
76369139 | 3817 | struct ring_buffer *rb; |
e3f3541c | 3818 | u64 enabled, running, now; |
38ff667b PZ |
3819 | |
3820 | rcu_read_lock(); | |
5ec4c599 PZ |
3821 | rb = rcu_dereference(event->rb); |
3822 | if (!rb) | |
3823 | goto unlock; | |
3824 | ||
0d641208 EM |
3825 | /* |
3826 | * compute total_time_enabled, total_time_running | |
3827 | * based on snapshot values taken when the event | |
3828 | * was last scheduled in. | |
3829 | * | |
3830 | * we cannot simply called update_context_time() | |
3831 | * because of locking issue as we can be called in | |
3832 | * NMI context | |
3833 | */ | |
e3f3541c | 3834 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 3835 | |
76369139 | 3836 | userpg = rb->user_page; |
7b732a75 PZ |
3837 | /* |
3838 | * Disable preemption so as to not let the corresponding user-space | |
3839 | * spin too long if we get preempted. | |
3840 | */ | |
3841 | preempt_disable(); | |
37d81828 | 3842 | ++userpg->lock; |
92f22a38 | 3843 | barrier(); |
cdd6c482 | 3844 | userpg->index = perf_event_index(event); |
b5e58793 | 3845 | userpg->offset = perf_event_count(event); |
365a4038 | 3846 | if (userpg->index) |
e7850595 | 3847 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3848 | |
0d641208 | 3849 | userpg->time_enabled = enabled + |
cdd6c482 | 3850 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3851 | |
0d641208 | 3852 | userpg->time_running = running + |
cdd6c482 | 3853 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3854 | |
c7206205 | 3855 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3856 | |
92f22a38 | 3857 | barrier(); |
37d81828 | 3858 | ++userpg->lock; |
7b732a75 | 3859 | preempt_enable(); |
38ff667b | 3860 | unlock: |
7b732a75 | 3861 | rcu_read_unlock(); |
37d81828 PM |
3862 | } |
3863 | ||
906010b2 PZ |
3864 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3865 | { | |
3866 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3867 | struct ring_buffer *rb; |
906010b2 PZ |
3868 | int ret = VM_FAULT_SIGBUS; |
3869 | ||
3870 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3871 | if (vmf->pgoff == 0) | |
3872 | ret = 0; | |
3873 | return ret; | |
3874 | } | |
3875 | ||
3876 | rcu_read_lock(); | |
76369139 FW |
3877 | rb = rcu_dereference(event->rb); |
3878 | if (!rb) | |
906010b2 PZ |
3879 | goto unlock; |
3880 | ||
3881 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3882 | goto unlock; | |
3883 | ||
76369139 | 3884 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3885 | if (!vmf->page) |
3886 | goto unlock; | |
3887 | ||
3888 | get_page(vmf->page); | |
3889 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3890 | vmf->page->index = vmf->pgoff; | |
3891 | ||
3892 | ret = 0; | |
3893 | unlock: | |
3894 | rcu_read_unlock(); | |
3895 | ||
3896 | return ret; | |
3897 | } | |
3898 | ||
10c6db11 PZ |
3899 | static void ring_buffer_attach(struct perf_event *event, |
3900 | struct ring_buffer *rb) | |
3901 | { | |
b69cf536 | 3902 | struct ring_buffer *old_rb = NULL; |
10c6db11 PZ |
3903 | unsigned long flags; |
3904 | ||
b69cf536 PZ |
3905 | if (event->rb) { |
3906 | /* | |
3907 | * Should be impossible, we set this when removing | |
3908 | * event->rb_entry and wait/clear when adding event->rb_entry. | |
3909 | */ | |
3910 | WARN_ON_ONCE(event->rcu_pending); | |
10c6db11 | 3911 | |
b69cf536 PZ |
3912 | old_rb = event->rb; |
3913 | event->rcu_batches = get_state_synchronize_rcu(); | |
3914 | event->rcu_pending = 1; | |
10c6db11 | 3915 | |
b69cf536 PZ |
3916 | spin_lock_irqsave(&old_rb->event_lock, flags); |
3917 | list_del_rcu(&event->rb_entry); | |
3918 | spin_unlock_irqrestore(&old_rb->event_lock, flags); | |
3919 | } | |
10c6db11 | 3920 | |
b69cf536 PZ |
3921 | if (event->rcu_pending && rb) { |
3922 | cond_synchronize_rcu(event->rcu_batches); | |
3923 | event->rcu_pending = 0; | |
3924 | } | |
10c6db11 | 3925 | |
b69cf536 PZ |
3926 | if (rb) { |
3927 | spin_lock_irqsave(&rb->event_lock, flags); | |
3928 | list_add_rcu(&event->rb_entry, &rb->event_list); | |
3929 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3930 | } | |
3931 | ||
3932 | rcu_assign_pointer(event->rb, rb); | |
3933 | ||
3934 | if (old_rb) { | |
3935 | ring_buffer_put(old_rb); | |
3936 | /* | |
3937 | * Since we detached before setting the new rb, so that we | |
3938 | * could attach the new rb, we could have missed a wakeup. | |
3939 | * Provide it now. | |
3940 | */ | |
3941 | wake_up_all(&event->waitq); | |
3942 | } | |
10c6db11 PZ |
3943 | } |
3944 | ||
3945 | static void ring_buffer_wakeup(struct perf_event *event) | |
3946 | { | |
3947 | struct ring_buffer *rb; | |
3948 | ||
3949 | rcu_read_lock(); | |
3950 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
3951 | if (rb) { |
3952 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
3953 | wake_up_all(&event->waitq); | |
3954 | } | |
10c6db11 PZ |
3955 | rcu_read_unlock(); |
3956 | } | |
3957 | ||
76369139 | 3958 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3959 | { |
76369139 | 3960 | struct ring_buffer *rb; |
906010b2 | 3961 | |
76369139 FW |
3962 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3963 | rb_free(rb); | |
7b732a75 PZ |
3964 | } |
3965 | ||
76369139 | 3966 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3967 | { |
76369139 | 3968 | struct ring_buffer *rb; |
7b732a75 | 3969 | |
ac9721f3 | 3970 | rcu_read_lock(); |
76369139 FW |
3971 | rb = rcu_dereference(event->rb); |
3972 | if (rb) { | |
3973 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3974 | rb = NULL; | |
ac9721f3 PZ |
3975 | } |
3976 | rcu_read_unlock(); | |
3977 | ||
76369139 | 3978 | return rb; |
ac9721f3 PZ |
3979 | } |
3980 | ||
76369139 | 3981 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3982 | { |
76369139 | 3983 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3984 | return; |
7b732a75 | 3985 | |
9bb5d40c | 3986 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 3987 | |
76369139 | 3988 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3989 | } |
3990 | ||
3991 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3992 | { | |
cdd6c482 | 3993 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3994 | |
cdd6c482 | 3995 | atomic_inc(&event->mmap_count); |
9bb5d40c | 3996 | atomic_inc(&event->rb->mmap_count); |
7b732a75 PZ |
3997 | } |
3998 | ||
9bb5d40c PZ |
3999 | /* |
4000 | * A buffer can be mmap()ed multiple times; either directly through the same | |
4001 | * event, or through other events by use of perf_event_set_output(). | |
4002 | * | |
4003 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
4004 | * the buffer here, where we still have a VM context. This means we need | |
4005 | * to detach all events redirecting to us. | |
4006 | */ | |
7b732a75 PZ |
4007 | static void perf_mmap_close(struct vm_area_struct *vma) |
4008 | { | |
cdd6c482 | 4009 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4010 | |
b69cf536 | 4011 | struct ring_buffer *rb = ring_buffer_get(event); |
9bb5d40c PZ |
4012 | struct user_struct *mmap_user = rb->mmap_user; |
4013 | int mmap_locked = rb->mmap_locked; | |
4014 | unsigned long size = perf_data_size(rb); | |
789f90fc | 4015 | |
9bb5d40c PZ |
4016 | atomic_dec(&rb->mmap_count); |
4017 | ||
4018 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
b69cf536 | 4019 | goto out_put; |
9bb5d40c | 4020 | |
b69cf536 | 4021 | ring_buffer_attach(event, NULL); |
9bb5d40c PZ |
4022 | mutex_unlock(&event->mmap_mutex); |
4023 | ||
4024 | /* If there's still other mmap()s of this buffer, we're done. */ | |
b69cf536 PZ |
4025 | if (atomic_read(&rb->mmap_count)) |
4026 | goto out_put; | |
ac9721f3 | 4027 | |
9bb5d40c PZ |
4028 | /* |
4029 | * No other mmap()s, detach from all other events that might redirect | |
4030 | * into the now unreachable buffer. Somewhat complicated by the | |
4031 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
4032 | */ | |
4033 | again: | |
4034 | rcu_read_lock(); | |
4035 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
4036 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
4037 | /* | |
4038 | * This event is en-route to free_event() which will | |
4039 | * detach it and remove it from the list. | |
4040 | */ | |
4041 | continue; | |
4042 | } | |
4043 | rcu_read_unlock(); | |
789f90fc | 4044 | |
9bb5d40c PZ |
4045 | mutex_lock(&event->mmap_mutex); |
4046 | /* | |
4047 | * Check we didn't race with perf_event_set_output() which can | |
4048 | * swizzle the rb from under us while we were waiting to | |
4049 | * acquire mmap_mutex. | |
4050 | * | |
4051 | * If we find a different rb; ignore this event, a next | |
4052 | * iteration will no longer find it on the list. We have to | |
4053 | * still restart the iteration to make sure we're not now | |
4054 | * iterating the wrong list. | |
4055 | */ | |
b69cf536 PZ |
4056 | if (event->rb == rb) |
4057 | ring_buffer_attach(event, NULL); | |
4058 | ||
cdd6c482 | 4059 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 4060 | put_event(event); |
ac9721f3 | 4061 | |
9bb5d40c PZ |
4062 | /* |
4063 | * Restart the iteration; either we're on the wrong list or | |
4064 | * destroyed its integrity by doing a deletion. | |
4065 | */ | |
4066 | goto again; | |
7b732a75 | 4067 | } |
9bb5d40c PZ |
4068 | rcu_read_unlock(); |
4069 | ||
4070 | /* | |
4071 | * It could be there's still a few 0-ref events on the list; they'll | |
4072 | * get cleaned up by free_event() -- they'll also still have their | |
4073 | * ref on the rb and will free it whenever they are done with it. | |
4074 | * | |
4075 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
4076 | * undo the VM accounting. | |
4077 | */ | |
4078 | ||
4079 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
4080 | vma->vm_mm->pinned_vm -= mmap_locked; | |
4081 | free_uid(mmap_user); | |
4082 | ||
b69cf536 | 4083 | out_put: |
9bb5d40c | 4084 | ring_buffer_put(rb); /* could be last */ |
37d81828 PM |
4085 | } |
4086 | ||
f0f37e2f | 4087 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
4088 | .open = perf_mmap_open, |
4089 | .close = perf_mmap_close, | |
4090 | .fault = perf_mmap_fault, | |
4091 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
4092 | }; |
4093 | ||
4094 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
4095 | { | |
cdd6c482 | 4096 | struct perf_event *event = file->private_data; |
22a4f650 | 4097 | unsigned long user_locked, user_lock_limit; |
789f90fc | 4098 | struct user_struct *user = current_user(); |
22a4f650 | 4099 | unsigned long locked, lock_limit; |
76369139 | 4100 | struct ring_buffer *rb; |
7b732a75 PZ |
4101 | unsigned long vma_size; |
4102 | unsigned long nr_pages; | |
789f90fc | 4103 | long user_extra, extra; |
d57e34fd | 4104 | int ret = 0, flags = 0; |
37d81828 | 4105 | |
c7920614 PZ |
4106 | /* |
4107 | * Don't allow mmap() of inherited per-task counters. This would | |
4108 | * create a performance issue due to all children writing to the | |
76369139 | 4109 | * same rb. |
c7920614 PZ |
4110 | */ |
4111 | if (event->cpu == -1 && event->attr.inherit) | |
4112 | return -EINVAL; | |
4113 | ||
43a21ea8 | 4114 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 4115 | return -EINVAL; |
7b732a75 PZ |
4116 | |
4117 | vma_size = vma->vm_end - vma->vm_start; | |
4118 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
4119 | ||
7730d865 | 4120 | /* |
76369139 | 4121 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
4122 | * can do bitmasks instead of modulo. |
4123 | */ | |
4124 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
4125 | return -EINVAL; |
4126 | ||
7b732a75 | 4127 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
4128 | return -EINVAL; |
4129 | ||
7b732a75 PZ |
4130 | if (vma->vm_pgoff != 0) |
4131 | return -EINVAL; | |
37d81828 | 4132 | |
cdd6c482 | 4133 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 4134 | again: |
cdd6c482 | 4135 | mutex_lock(&event->mmap_mutex); |
76369139 | 4136 | if (event->rb) { |
9bb5d40c | 4137 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4138 | ret = -EINVAL; |
9bb5d40c PZ |
4139 | goto unlock; |
4140 | } | |
4141 | ||
4142 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4143 | /* | |
4144 | * Raced against perf_mmap_close() through | |
4145 | * perf_event_set_output(). Try again, hope for better | |
4146 | * luck. | |
4147 | */ | |
4148 | mutex_unlock(&event->mmap_mutex); | |
4149 | goto again; | |
4150 | } | |
4151 | ||
ebb3c4c4 PZ |
4152 | goto unlock; |
4153 | } | |
4154 | ||
789f90fc | 4155 | user_extra = nr_pages + 1; |
cdd6c482 | 4156 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4157 | |
4158 | /* | |
4159 | * Increase the limit linearly with more CPUs: | |
4160 | */ | |
4161 | user_lock_limit *= num_online_cpus(); | |
4162 | ||
789f90fc | 4163 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4164 | |
789f90fc PZ |
4165 | extra = 0; |
4166 | if (user_locked > user_lock_limit) | |
4167 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4168 | |
78d7d407 | 4169 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4170 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4171 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4172 | |
459ec28a IM |
4173 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4174 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4175 | ret = -EPERM; |
4176 | goto unlock; | |
4177 | } | |
7b732a75 | 4178 | |
76369139 | 4179 | WARN_ON(event->rb); |
906010b2 | 4180 | |
d57e34fd | 4181 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4182 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4183 | |
4ec8363d VW |
4184 | rb = rb_alloc(nr_pages, |
4185 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4186 | event->cpu, flags); | |
4187 | ||
76369139 | 4188 | if (!rb) { |
ac9721f3 | 4189 | ret = -ENOMEM; |
ebb3c4c4 | 4190 | goto unlock; |
ac9721f3 | 4191 | } |
26cb63ad | 4192 | |
9bb5d40c | 4193 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
4194 | rb->mmap_locked = extra; |
4195 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 4196 | |
ac9721f3 | 4197 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
4198 | vma->vm_mm->pinned_vm += extra; |
4199 | ||
9bb5d40c | 4200 | ring_buffer_attach(event, rb); |
ac9721f3 | 4201 | |
fa731587 | 4202 | perf_event_init_userpage(event); |
9a0f05cb PZ |
4203 | perf_event_update_userpage(event); |
4204 | ||
ebb3c4c4 | 4205 | unlock: |
ac9721f3 PZ |
4206 | if (!ret) |
4207 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 4208 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4209 | |
9bb5d40c PZ |
4210 | /* |
4211 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4212 | * vma. | |
4213 | */ | |
26cb63ad | 4214 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4215 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
4216 | |
4217 | return ret; | |
37d81828 PM |
4218 | } |
4219 | ||
3c446b3d PZ |
4220 | static int perf_fasync(int fd, struct file *filp, int on) |
4221 | { | |
496ad9aa | 4222 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4223 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4224 | int retval; |
4225 | ||
4226 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4227 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4228 | mutex_unlock(&inode->i_mutex); |
4229 | ||
4230 | if (retval < 0) | |
4231 | return retval; | |
4232 | ||
4233 | return 0; | |
4234 | } | |
4235 | ||
0793a61d | 4236 | static const struct file_operations perf_fops = { |
3326c1ce | 4237 | .llseek = no_llseek, |
0793a61d TG |
4238 | .release = perf_release, |
4239 | .read = perf_read, | |
4240 | .poll = perf_poll, | |
d859e29f PM |
4241 | .unlocked_ioctl = perf_ioctl, |
4242 | .compat_ioctl = perf_ioctl, | |
37d81828 | 4243 | .mmap = perf_mmap, |
3c446b3d | 4244 | .fasync = perf_fasync, |
0793a61d TG |
4245 | }; |
4246 | ||
925d519a | 4247 | /* |
cdd6c482 | 4248 | * Perf event wakeup |
925d519a PZ |
4249 | * |
4250 | * If there's data, ensure we set the poll() state and publish everything | |
4251 | * to user-space before waking everybody up. | |
4252 | */ | |
4253 | ||
cdd6c482 | 4254 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4255 | { |
10c6db11 | 4256 | ring_buffer_wakeup(event); |
4c9e2542 | 4257 | |
cdd6c482 IM |
4258 | if (event->pending_kill) { |
4259 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4260 | event->pending_kill = 0; | |
4c9e2542 | 4261 | } |
925d519a PZ |
4262 | } |
4263 | ||
e360adbe | 4264 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4265 | { |
cdd6c482 IM |
4266 | struct perf_event *event = container_of(entry, |
4267 | struct perf_event, pending); | |
79f14641 | 4268 | |
cdd6c482 IM |
4269 | if (event->pending_disable) { |
4270 | event->pending_disable = 0; | |
4271 | __perf_event_disable(event); | |
79f14641 PZ |
4272 | } |
4273 | ||
cdd6c482 IM |
4274 | if (event->pending_wakeup) { |
4275 | event->pending_wakeup = 0; | |
4276 | perf_event_wakeup(event); | |
79f14641 PZ |
4277 | } |
4278 | } | |
4279 | ||
39447b38 ZY |
4280 | /* |
4281 | * We assume there is only KVM supporting the callbacks. | |
4282 | * Later on, we might change it to a list if there is | |
4283 | * another virtualization implementation supporting the callbacks. | |
4284 | */ | |
4285 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4286 | ||
4287 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4288 | { | |
4289 | perf_guest_cbs = cbs; | |
4290 | return 0; | |
4291 | } | |
4292 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4293 | ||
4294 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4295 | { | |
4296 | perf_guest_cbs = NULL; | |
4297 | return 0; | |
4298 | } | |
4299 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4300 | ||
4018994f JO |
4301 | static void |
4302 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4303 | struct pt_regs *regs, u64 mask) | |
4304 | { | |
4305 | int bit; | |
4306 | ||
4307 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4308 | sizeof(mask) * BITS_PER_BYTE) { | |
4309 | u64 val; | |
4310 | ||
4311 | val = perf_reg_value(regs, bit); | |
4312 | perf_output_put(handle, val); | |
4313 | } | |
4314 | } | |
4315 | ||
4316 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
4317 | struct pt_regs *regs) | |
4318 | { | |
4319 | if (!user_mode(regs)) { | |
4320 | if (current->mm) | |
4321 | regs = task_pt_regs(current); | |
4322 | else | |
4323 | regs = NULL; | |
4324 | } | |
4325 | ||
4326 | if (regs) { | |
4327 | regs_user->regs = regs; | |
4328 | regs_user->abi = perf_reg_abi(current); | |
4329 | } | |
4330 | } | |
4331 | ||
c5ebcedb JO |
4332 | /* |
4333 | * Get remaining task size from user stack pointer. | |
4334 | * | |
4335 | * It'd be better to take stack vma map and limit this more | |
4336 | * precisly, but there's no way to get it safely under interrupt, | |
4337 | * so using TASK_SIZE as limit. | |
4338 | */ | |
4339 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4340 | { | |
4341 | unsigned long addr = perf_user_stack_pointer(regs); | |
4342 | ||
4343 | if (!addr || addr >= TASK_SIZE) | |
4344 | return 0; | |
4345 | ||
4346 | return TASK_SIZE - addr; | |
4347 | } | |
4348 | ||
4349 | static u16 | |
4350 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4351 | struct pt_regs *regs) | |
4352 | { | |
4353 | u64 task_size; | |
4354 | ||
4355 | /* No regs, no stack pointer, no dump. */ | |
4356 | if (!regs) | |
4357 | return 0; | |
4358 | ||
4359 | /* | |
4360 | * Check if we fit in with the requested stack size into the: | |
4361 | * - TASK_SIZE | |
4362 | * If we don't, we limit the size to the TASK_SIZE. | |
4363 | * | |
4364 | * - remaining sample size | |
4365 | * If we don't, we customize the stack size to | |
4366 | * fit in to the remaining sample size. | |
4367 | */ | |
4368 | ||
4369 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4370 | stack_size = min(stack_size, (u16) task_size); | |
4371 | ||
4372 | /* Current header size plus static size and dynamic size. */ | |
4373 | header_size += 2 * sizeof(u64); | |
4374 | ||
4375 | /* Do we fit in with the current stack dump size? */ | |
4376 | if ((u16) (header_size + stack_size) < header_size) { | |
4377 | /* | |
4378 | * If we overflow the maximum size for the sample, | |
4379 | * we customize the stack dump size to fit in. | |
4380 | */ | |
4381 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4382 | stack_size = round_up(stack_size, sizeof(u64)); | |
4383 | } | |
4384 | ||
4385 | return stack_size; | |
4386 | } | |
4387 | ||
4388 | static void | |
4389 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4390 | struct pt_regs *regs) | |
4391 | { | |
4392 | /* Case of a kernel thread, nothing to dump */ | |
4393 | if (!regs) { | |
4394 | u64 size = 0; | |
4395 | perf_output_put(handle, size); | |
4396 | } else { | |
4397 | unsigned long sp; | |
4398 | unsigned int rem; | |
4399 | u64 dyn_size; | |
4400 | ||
4401 | /* | |
4402 | * We dump: | |
4403 | * static size | |
4404 | * - the size requested by user or the best one we can fit | |
4405 | * in to the sample max size | |
4406 | * data | |
4407 | * - user stack dump data | |
4408 | * dynamic size | |
4409 | * - the actual dumped size | |
4410 | */ | |
4411 | ||
4412 | /* Static size. */ | |
4413 | perf_output_put(handle, dump_size); | |
4414 | ||
4415 | /* Data. */ | |
4416 | sp = perf_user_stack_pointer(regs); | |
4417 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4418 | dyn_size = dump_size - rem; | |
4419 | ||
4420 | perf_output_skip(handle, rem); | |
4421 | ||
4422 | /* Dynamic size. */ | |
4423 | perf_output_put(handle, dyn_size); | |
4424 | } | |
4425 | } | |
4426 | ||
c980d109 ACM |
4427 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4428 | struct perf_sample_data *data, | |
4429 | struct perf_event *event) | |
6844c09d ACM |
4430 | { |
4431 | u64 sample_type = event->attr.sample_type; | |
4432 | ||
4433 | data->type = sample_type; | |
4434 | header->size += event->id_header_size; | |
4435 | ||
4436 | if (sample_type & PERF_SAMPLE_TID) { | |
4437 | /* namespace issues */ | |
4438 | data->tid_entry.pid = perf_event_pid(event, current); | |
4439 | data->tid_entry.tid = perf_event_tid(event, current); | |
4440 | } | |
4441 | ||
4442 | if (sample_type & PERF_SAMPLE_TIME) | |
4443 | data->time = perf_clock(); | |
4444 | ||
ff3d527c | 4445 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4446 | data->id = primary_event_id(event); |
4447 | ||
4448 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4449 | data->stream_id = event->id; | |
4450 | ||
4451 | if (sample_type & PERF_SAMPLE_CPU) { | |
4452 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4453 | data->cpu_entry.reserved = 0; | |
4454 | } | |
4455 | } | |
4456 | ||
76369139 FW |
4457 | void perf_event_header__init_id(struct perf_event_header *header, |
4458 | struct perf_sample_data *data, | |
4459 | struct perf_event *event) | |
c980d109 ACM |
4460 | { |
4461 | if (event->attr.sample_id_all) | |
4462 | __perf_event_header__init_id(header, data, event); | |
4463 | } | |
4464 | ||
4465 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4466 | struct perf_sample_data *data) | |
4467 | { | |
4468 | u64 sample_type = data->type; | |
4469 | ||
4470 | if (sample_type & PERF_SAMPLE_TID) | |
4471 | perf_output_put(handle, data->tid_entry); | |
4472 | ||
4473 | if (sample_type & PERF_SAMPLE_TIME) | |
4474 | perf_output_put(handle, data->time); | |
4475 | ||
4476 | if (sample_type & PERF_SAMPLE_ID) | |
4477 | perf_output_put(handle, data->id); | |
4478 | ||
4479 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4480 | perf_output_put(handle, data->stream_id); | |
4481 | ||
4482 | if (sample_type & PERF_SAMPLE_CPU) | |
4483 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
4484 | |
4485 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
4486 | perf_output_put(handle, data->id); | |
c980d109 ACM |
4487 | } |
4488 | ||
76369139 FW |
4489 | void perf_event__output_id_sample(struct perf_event *event, |
4490 | struct perf_output_handle *handle, | |
4491 | struct perf_sample_data *sample) | |
c980d109 ACM |
4492 | { |
4493 | if (event->attr.sample_id_all) | |
4494 | __perf_event__output_id_sample(handle, sample); | |
4495 | } | |
4496 | ||
3dab77fb | 4497 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4498 | struct perf_event *event, |
4499 | u64 enabled, u64 running) | |
3dab77fb | 4500 | { |
cdd6c482 | 4501 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4502 | u64 values[4]; |
4503 | int n = 0; | |
4504 | ||
b5e58793 | 4505 | values[n++] = perf_event_count(event); |
3dab77fb | 4506 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4507 | values[n++] = enabled + |
cdd6c482 | 4508 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4509 | } |
4510 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4511 | values[n++] = running + |
cdd6c482 | 4512 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4513 | } |
4514 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4515 | values[n++] = primary_event_id(event); |
3dab77fb | 4516 | |
76369139 | 4517 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4518 | } |
4519 | ||
4520 | /* | |
cdd6c482 | 4521 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4522 | */ |
4523 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4524 | struct perf_event *event, |
4525 | u64 enabled, u64 running) | |
3dab77fb | 4526 | { |
cdd6c482 IM |
4527 | struct perf_event *leader = event->group_leader, *sub; |
4528 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4529 | u64 values[5]; |
4530 | int n = 0; | |
4531 | ||
4532 | values[n++] = 1 + leader->nr_siblings; | |
4533 | ||
4534 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4535 | values[n++] = enabled; |
3dab77fb PZ |
4536 | |
4537 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4538 | values[n++] = running; |
3dab77fb | 4539 | |
cdd6c482 | 4540 | if (leader != event) |
3dab77fb PZ |
4541 | leader->pmu->read(leader); |
4542 | ||
b5e58793 | 4543 | values[n++] = perf_event_count(leader); |
3dab77fb | 4544 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4545 | values[n++] = primary_event_id(leader); |
3dab77fb | 4546 | |
76369139 | 4547 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4548 | |
65abc865 | 4549 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4550 | n = 0; |
4551 | ||
6f5ab001 JO |
4552 | if ((sub != event) && |
4553 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
4554 | sub->pmu->read(sub); |
4555 | ||
b5e58793 | 4556 | values[n++] = perf_event_count(sub); |
3dab77fb | 4557 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4558 | values[n++] = primary_event_id(sub); |
3dab77fb | 4559 | |
76369139 | 4560 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4561 | } |
4562 | } | |
4563 | ||
eed01528 SE |
4564 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4565 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4566 | ||
3dab77fb | 4567 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4568 | struct perf_event *event) |
3dab77fb | 4569 | { |
e3f3541c | 4570 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4571 | u64 read_format = event->attr.read_format; |
4572 | ||
4573 | /* | |
4574 | * compute total_time_enabled, total_time_running | |
4575 | * based on snapshot values taken when the event | |
4576 | * was last scheduled in. | |
4577 | * | |
4578 | * we cannot simply called update_context_time() | |
4579 | * because of locking issue as we are called in | |
4580 | * NMI context | |
4581 | */ | |
c4794295 | 4582 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4583 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4584 | |
cdd6c482 | 4585 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4586 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4587 | else |
eed01528 | 4588 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4589 | } |
4590 | ||
5622f295 MM |
4591 | void perf_output_sample(struct perf_output_handle *handle, |
4592 | struct perf_event_header *header, | |
4593 | struct perf_sample_data *data, | |
cdd6c482 | 4594 | struct perf_event *event) |
5622f295 MM |
4595 | { |
4596 | u64 sample_type = data->type; | |
4597 | ||
4598 | perf_output_put(handle, *header); | |
4599 | ||
ff3d527c AH |
4600 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
4601 | perf_output_put(handle, data->id); | |
4602 | ||
5622f295 MM |
4603 | if (sample_type & PERF_SAMPLE_IP) |
4604 | perf_output_put(handle, data->ip); | |
4605 | ||
4606 | if (sample_type & PERF_SAMPLE_TID) | |
4607 | perf_output_put(handle, data->tid_entry); | |
4608 | ||
4609 | if (sample_type & PERF_SAMPLE_TIME) | |
4610 | perf_output_put(handle, data->time); | |
4611 | ||
4612 | if (sample_type & PERF_SAMPLE_ADDR) | |
4613 | perf_output_put(handle, data->addr); | |
4614 | ||
4615 | if (sample_type & PERF_SAMPLE_ID) | |
4616 | perf_output_put(handle, data->id); | |
4617 | ||
4618 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4619 | perf_output_put(handle, data->stream_id); | |
4620 | ||
4621 | if (sample_type & PERF_SAMPLE_CPU) | |
4622 | perf_output_put(handle, data->cpu_entry); | |
4623 | ||
4624 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4625 | perf_output_put(handle, data->period); | |
4626 | ||
4627 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4628 | perf_output_read(handle, event); |
5622f295 MM |
4629 | |
4630 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4631 | if (data->callchain) { | |
4632 | int size = 1; | |
4633 | ||
4634 | if (data->callchain) | |
4635 | size += data->callchain->nr; | |
4636 | ||
4637 | size *= sizeof(u64); | |
4638 | ||
76369139 | 4639 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4640 | } else { |
4641 | u64 nr = 0; | |
4642 | perf_output_put(handle, nr); | |
4643 | } | |
4644 | } | |
4645 | ||
4646 | if (sample_type & PERF_SAMPLE_RAW) { | |
4647 | if (data->raw) { | |
4648 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4649 | __output_copy(handle, data->raw->data, |
4650 | data->raw->size); | |
5622f295 MM |
4651 | } else { |
4652 | struct { | |
4653 | u32 size; | |
4654 | u32 data; | |
4655 | } raw = { | |
4656 | .size = sizeof(u32), | |
4657 | .data = 0, | |
4658 | }; | |
4659 | perf_output_put(handle, raw); | |
4660 | } | |
4661 | } | |
a7ac67ea | 4662 | |
bce38cd5 SE |
4663 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
4664 | if (data->br_stack) { | |
4665 | size_t size; | |
4666 | ||
4667 | size = data->br_stack->nr | |
4668 | * sizeof(struct perf_branch_entry); | |
4669 | ||
4670 | perf_output_put(handle, data->br_stack->nr); | |
4671 | perf_output_copy(handle, data->br_stack->entries, size); | |
4672 | } else { | |
4673 | /* | |
4674 | * we always store at least the value of nr | |
4675 | */ | |
4676 | u64 nr = 0; | |
4677 | perf_output_put(handle, nr); | |
4678 | } | |
4679 | } | |
4018994f JO |
4680 | |
4681 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4682 | u64 abi = data->regs_user.abi; | |
4683 | ||
4684 | /* | |
4685 | * If there are no regs to dump, notice it through | |
4686 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4687 | */ | |
4688 | perf_output_put(handle, abi); | |
4689 | ||
4690 | if (abi) { | |
4691 | u64 mask = event->attr.sample_regs_user; | |
4692 | perf_output_sample_regs(handle, | |
4693 | data->regs_user.regs, | |
4694 | mask); | |
4695 | } | |
4696 | } | |
c5ebcedb | 4697 | |
a5cdd40c | 4698 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
4699 | perf_output_sample_ustack(handle, |
4700 | data->stack_user_size, | |
4701 | data->regs_user.regs); | |
a5cdd40c | 4702 | } |
c3feedf2 AK |
4703 | |
4704 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
4705 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
4706 | |
4707 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
4708 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c | 4709 | |
fdfbbd07 AK |
4710 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
4711 | perf_output_put(handle, data->txn); | |
4712 | ||
a5cdd40c PZ |
4713 | if (!event->attr.watermark) { |
4714 | int wakeup_events = event->attr.wakeup_events; | |
4715 | ||
4716 | if (wakeup_events) { | |
4717 | struct ring_buffer *rb = handle->rb; | |
4718 | int events = local_inc_return(&rb->events); | |
4719 | ||
4720 | if (events >= wakeup_events) { | |
4721 | local_sub(wakeup_events, &rb->events); | |
4722 | local_inc(&rb->wakeup); | |
4723 | } | |
4724 | } | |
4725 | } | |
5622f295 MM |
4726 | } |
4727 | ||
4728 | void perf_prepare_sample(struct perf_event_header *header, | |
4729 | struct perf_sample_data *data, | |
cdd6c482 | 4730 | struct perf_event *event, |
5622f295 | 4731 | struct pt_regs *regs) |
7b732a75 | 4732 | { |
cdd6c482 | 4733 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4734 | |
cdd6c482 | 4735 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4736 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4737 | |
4738 | header->misc = 0; | |
4739 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4740 | |
c980d109 | 4741 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4742 | |
c320c7b7 | 4743 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4744 | data->ip = perf_instruction_pointer(regs); |
4745 | ||
b23f3325 | 4746 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4747 | int size = 1; |
394ee076 | 4748 | |
e6dab5ff | 4749 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4750 | |
4751 | if (data->callchain) | |
4752 | size += data->callchain->nr; | |
4753 | ||
4754 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4755 | } |
4756 | ||
3a43ce68 | 4757 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4758 | int size = sizeof(u32); |
4759 | ||
4760 | if (data->raw) | |
4761 | size += data->raw->size; | |
4762 | else | |
4763 | size += sizeof(u32); | |
4764 | ||
4765 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4766 | header->size += size; |
7f453c24 | 4767 | } |
bce38cd5 SE |
4768 | |
4769 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4770 | int size = sizeof(u64); /* nr */ | |
4771 | if (data->br_stack) { | |
4772 | size += data->br_stack->nr | |
4773 | * sizeof(struct perf_branch_entry); | |
4774 | } | |
4775 | header->size += size; | |
4776 | } | |
4018994f JO |
4777 | |
4778 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4779 | /* regs dump ABI info */ | |
4780 | int size = sizeof(u64); | |
4781 | ||
4782 | perf_sample_regs_user(&data->regs_user, regs); | |
4783 | ||
4784 | if (data->regs_user.regs) { | |
4785 | u64 mask = event->attr.sample_regs_user; | |
4786 | size += hweight64(mask) * sizeof(u64); | |
4787 | } | |
4788 | ||
4789 | header->size += size; | |
4790 | } | |
c5ebcedb JO |
4791 | |
4792 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4793 | /* | |
4794 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4795 | * processed as the last one or have additional check added | |
4796 | * in case new sample type is added, because we could eat | |
4797 | * up the rest of the sample size. | |
4798 | */ | |
4799 | struct perf_regs_user *uregs = &data->regs_user; | |
4800 | u16 stack_size = event->attr.sample_stack_user; | |
4801 | u16 size = sizeof(u64); | |
4802 | ||
4803 | if (!uregs->abi) | |
4804 | perf_sample_regs_user(uregs, regs); | |
4805 | ||
4806 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4807 | uregs->regs); | |
4808 | ||
4809 | /* | |
4810 | * If there is something to dump, add space for the dump | |
4811 | * itself and for the field that tells the dynamic size, | |
4812 | * which is how many have been actually dumped. | |
4813 | */ | |
4814 | if (stack_size) | |
4815 | size += sizeof(u64) + stack_size; | |
4816 | ||
4817 | data->stack_user_size = stack_size; | |
4818 | header->size += size; | |
4819 | } | |
5622f295 | 4820 | } |
7f453c24 | 4821 | |
a8b0ca17 | 4822 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4823 | struct perf_sample_data *data, |
4824 | struct pt_regs *regs) | |
4825 | { | |
4826 | struct perf_output_handle handle; | |
4827 | struct perf_event_header header; | |
689802b2 | 4828 | |
927c7a9e FW |
4829 | /* protect the callchain buffers */ |
4830 | rcu_read_lock(); | |
4831 | ||
cdd6c482 | 4832 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4833 | |
a7ac67ea | 4834 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4835 | goto exit; |
0322cd6e | 4836 | |
cdd6c482 | 4837 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4838 | |
8a057d84 | 4839 | perf_output_end(&handle); |
927c7a9e FW |
4840 | |
4841 | exit: | |
4842 | rcu_read_unlock(); | |
0322cd6e PZ |
4843 | } |
4844 | ||
38b200d6 | 4845 | /* |
cdd6c482 | 4846 | * read event_id |
38b200d6 PZ |
4847 | */ |
4848 | ||
4849 | struct perf_read_event { | |
4850 | struct perf_event_header header; | |
4851 | ||
4852 | u32 pid; | |
4853 | u32 tid; | |
38b200d6 PZ |
4854 | }; |
4855 | ||
4856 | static void | |
cdd6c482 | 4857 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4858 | struct task_struct *task) |
4859 | { | |
4860 | struct perf_output_handle handle; | |
c980d109 | 4861 | struct perf_sample_data sample; |
dfc65094 | 4862 | struct perf_read_event read_event = { |
38b200d6 | 4863 | .header = { |
cdd6c482 | 4864 | .type = PERF_RECORD_READ, |
38b200d6 | 4865 | .misc = 0, |
c320c7b7 | 4866 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4867 | }, |
cdd6c482 IM |
4868 | .pid = perf_event_pid(event, task), |
4869 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4870 | }; |
3dab77fb | 4871 | int ret; |
38b200d6 | 4872 | |
c980d109 | 4873 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4874 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4875 | if (ret) |
4876 | return; | |
4877 | ||
dfc65094 | 4878 | perf_output_put(&handle, read_event); |
cdd6c482 | 4879 | perf_output_read(&handle, event); |
c980d109 | 4880 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4881 | |
38b200d6 PZ |
4882 | perf_output_end(&handle); |
4883 | } | |
4884 | ||
52d857a8 JO |
4885 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
4886 | ||
4887 | static void | |
4888 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
4889 | perf_event_aux_output_cb output, |
4890 | void *data) | |
4891 | { | |
4892 | struct perf_event *event; | |
4893 | ||
4894 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
4895 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
4896 | continue; | |
4897 | if (!event_filter_match(event)) | |
4898 | continue; | |
67516844 | 4899 | output(event, data); |
52d857a8 JO |
4900 | } |
4901 | } | |
4902 | ||
4903 | static void | |
67516844 | 4904 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
4905 | struct perf_event_context *task_ctx) |
4906 | { | |
4907 | struct perf_cpu_context *cpuctx; | |
4908 | struct perf_event_context *ctx; | |
4909 | struct pmu *pmu; | |
4910 | int ctxn; | |
4911 | ||
4912 | rcu_read_lock(); | |
4913 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
4914 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
4915 | if (cpuctx->unique_pmu != pmu) | |
4916 | goto next; | |
67516844 | 4917 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
4918 | if (task_ctx) |
4919 | goto next; | |
4920 | ctxn = pmu->task_ctx_nr; | |
4921 | if (ctxn < 0) | |
4922 | goto next; | |
4923 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4924 | if (ctx) | |
67516844 | 4925 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
4926 | next: |
4927 | put_cpu_ptr(pmu->pmu_cpu_context); | |
4928 | } | |
4929 | ||
4930 | if (task_ctx) { | |
4931 | preempt_disable(); | |
67516844 | 4932 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
4933 | preempt_enable(); |
4934 | } | |
4935 | rcu_read_unlock(); | |
4936 | } | |
4937 | ||
60313ebe | 4938 | /* |
9f498cc5 PZ |
4939 | * task tracking -- fork/exit |
4940 | * | |
13d7a241 | 4941 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
4942 | */ |
4943 | ||
9f498cc5 | 4944 | struct perf_task_event { |
3a80b4a3 | 4945 | struct task_struct *task; |
cdd6c482 | 4946 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4947 | |
4948 | struct { | |
4949 | struct perf_event_header header; | |
4950 | ||
4951 | u32 pid; | |
4952 | u32 ppid; | |
9f498cc5 PZ |
4953 | u32 tid; |
4954 | u32 ptid; | |
393b2ad8 | 4955 | u64 time; |
cdd6c482 | 4956 | } event_id; |
60313ebe PZ |
4957 | }; |
4958 | ||
67516844 JO |
4959 | static int perf_event_task_match(struct perf_event *event) |
4960 | { | |
13d7a241 SE |
4961 | return event->attr.comm || event->attr.mmap || |
4962 | event->attr.mmap2 || event->attr.mmap_data || | |
4963 | event->attr.task; | |
67516844 JO |
4964 | } |
4965 | ||
cdd6c482 | 4966 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 4967 | void *data) |
60313ebe | 4968 | { |
52d857a8 | 4969 | struct perf_task_event *task_event = data; |
60313ebe | 4970 | struct perf_output_handle handle; |
c980d109 | 4971 | struct perf_sample_data sample; |
9f498cc5 | 4972 | struct task_struct *task = task_event->task; |
c980d109 | 4973 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4974 | |
67516844 JO |
4975 | if (!perf_event_task_match(event)) |
4976 | return; | |
4977 | ||
c980d109 | 4978 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4979 | |
c980d109 | 4980 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4981 | task_event->event_id.header.size); |
ef60777c | 4982 | if (ret) |
c980d109 | 4983 | goto out; |
60313ebe | 4984 | |
cdd6c482 IM |
4985 | task_event->event_id.pid = perf_event_pid(event, task); |
4986 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4987 | |
cdd6c482 IM |
4988 | task_event->event_id.tid = perf_event_tid(event, task); |
4989 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4990 | |
cdd6c482 | 4991 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4992 | |
c980d109 ACM |
4993 | perf_event__output_id_sample(event, &handle, &sample); |
4994 | ||
60313ebe | 4995 | perf_output_end(&handle); |
c980d109 ACM |
4996 | out: |
4997 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4998 | } |
4999 | ||
cdd6c482 IM |
5000 | static void perf_event_task(struct task_struct *task, |
5001 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 5002 | int new) |
60313ebe | 5003 | { |
9f498cc5 | 5004 | struct perf_task_event task_event; |
60313ebe | 5005 | |
cdd6c482 IM |
5006 | if (!atomic_read(&nr_comm_events) && |
5007 | !atomic_read(&nr_mmap_events) && | |
5008 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
5009 | return; |
5010 | ||
9f498cc5 | 5011 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
5012 | .task = task, |
5013 | .task_ctx = task_ctx, | |
cdd6c482 | 5014 | .event_id = { |
60313ebe | 5015 | .header = { |
cdd6c482 | 5016 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 5017 | .misc = 0, |
cdd6c482 | 5018 | .size = sizeof(task_event.event_id), |
60313ebe | 5019 | }, |
573402db PZ |
5020 | /* .pid */ |
5021 | /* .ppid */ | |
9f498cc5 PZ |
5022 | /* .tid */ |
5023 | /* .ptid */ | |
6f93d0a7 | 5024 | .time = perf_clock(), |
60313ebe PZ |
5025 | }, |
5026 | }; | |
5027 | ||
67516844 | 5028 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
5029 | &task_event, |
5030 | task_ctx); | |
9f498cc5 PZ |
5031 | } |
5032 | ||
cdd6c482 | 5033 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 5034 | { |
cdd6c482 | 5035 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
5036 | } |
5037 | ||
8d1b2d93 PZ |
5038 | /* |
5039 | * comm tracking | |
5040 | */ | |
5041 | ||
5042 | struct perf_comm_event { | |
22a4f650 IM |
5043 | struct task_struct *task; |
5044 | char *comm; | |
8d1b2d93 PZ |
5045 | int comm_size; |
5046 | ||
5047 | struct { | |
5048 | struct perf_event_header header; | |
5049 | ||
5050 | u32 pid; | |
5051 | u32 tid; | |
cdd6c482 | 5052 | } event_id; |
8d1b2d93 PZ |
5053 | }; |
5054 | ||
67516844 JO |
5055 | static int perf_event_comm_match(struct perf_event *event) |
5056 | { | |
5057 | return event->attr.comm; | |
5058 | } | |
5059 | ||
cdd6c482 | 5060 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 5061 | void *data) |
8d1b2d93 | 5062 | { |
52d857a8 | 5063 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 5064 | struct perf_output_handle handle; |
c980d109 | 5065 | struct perf_sample_data sample; |
cdd6c482 | 5066 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
5067 | int ret; |
5068 | ||
67516844 JO |
5069 | if (!perf_event_comm_match(event)) |
5070 | return; | |
5071 | ||
c980d109 ACM |
5072 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
5073 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5074 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
5075 | |
5076 | if (ret) | |
c980d109 | 5077 | goto out; |
8d1b2d93 | 5078 | |
cdd6c482 IM |
5079 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
5080 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 5081 | |
cdd6c482 | 5082 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 5083 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 5084 | comm_event->comm_size); |
c980d109 ACM |
5085 | |
5086 | perf_event__output_id_sample(event, &handle, &sample); | |
5087 | ||
8d1b2d93 | 5088 | perf_output_end(&handle); |
c980d109 ACM |
5089 | out: |
5090 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
5091 | } |
5092 | ||
cdd6c482 | 5093 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 5094 | { |
413ee3b4 | 5095 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 5096 | unsigned int size; |
8d1b2d93 | 5097 | |
413ee3b4 | 5098 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 5099 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 5100 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
5101 | |
5102 | comm_event->comm = comm; | |
5103 | comm_event->comm_size = size; | |
5104 | ||
cdd6c482 | 5105 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 5106 | |
67516844 | 5107 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
5108 | comm_event, |
5109 | NULL); | |
8d1b2d93 PZ |
5110 | } |
5111 | ||
82b89778 | 5112 | void perf_event_comm(struct task_struct *task, bool exec) |
8d1b2d93 | 5113 | { |
9ee318a7 PZ |
5114 | struct perf_comm_event comm_event; |
5115 | ||
cdd6c482 | 5116 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 5117 | return; |
a63eaf34 | 5118 | |
9ee318a7 | 5119 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 5120 | .task = task, |
573402db PZ |
5121 | /* .comm */ |
5122 | /* .comm_size */ | |
cdd6c482 | 5123 | .event_id = { |
573402db | 5124 | .header = { |
cdd6c482 | 5125 | .type = PERF_RECORD_COMM, |
82b89778 | 5126 | .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0, |
573402db PZ |
5127 | /* .size */ |
5128 | }, | |
5129 | /* .pid */ | |
5130 | /* .tid */ | |
8d1b2d93 PZ |
5131 | }, |
5132 | }; | |
5133 | ||
cdd6c482 | 5134 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5135 | } |
5136 | ||
0a4a9391 PZ |
5137 | /* |
5138 | * mmap tracking | |
5139 | */ | |
5140 | ||
5141 | struct perf_mmap_event { | |
089dd79d PZ |
5142 | struct vm_area_struct *vma; |
5143 | ||
5144 | const char *file_name; | |
5145 | int file_size; | |
13d7a241 SE |
5146 | int maj, min; |
5147 | u64 ino; | |
5148 | u64 ino_generation; | |
f972eb63 | 5149 | u32 prot, flags; |
0a4a9391 PZ |
5150 | |
5151 | struct { | |
5152 | struct perf_event_header header; | |
5153 | ||
5154 | u32 pid; | |
5155 | u32 tid; | |
5156 | u64 start; | |
5157 | u64 len; | |
5158 | u64 pgoff; | |
cdd6c482 | 5159 | } event_id; |
0a4a9391 PZ |
5160 | }; |
5161 | ||
67516844 JO |
5162 | static int perf_event_mmap_match(struct perf_event *event, |
5163 | void *data) | |
5164 | { | |
5165 | struct perf_mmap_event *mmap_event = data; | |
5166 | struct vm_area_struct *vma = mmap_event->vma; | |
5167 | int executable = vma->vm_flags & VM_EXEC; | |
5168 | ||
5169 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5170 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5171 | } |
5172 | ||
cdd6c482 | 5173 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5174 | void *data) |
0a4a9391 | 5175 | { |
52d857a8 | 5176 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5177 | struct perf_output_handle handle; |
c980d109 | 5178 | struct perf_sample_data sample; |
cdd6c482 | 5179 | int size = mmap_event->event_id.header.size; |
c980d109 | 5180 | int ret; |
0a4a9391 | 5181 | |
67516844 JO |
5182 | if (!perf_event_mmap_match(event, data)) |
5183 | return; | |
5184 | ||
13d7a241 SE |
5185 | if (event->attr.mmap2) { |
5186 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5187 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5188 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5189 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5190 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
f972eb63 PZ |
5191 | mmap_event->event_id.header.size += sizeof(mmap_event->prot); |
5192 | mmap_event->event_id.header.size += sizeof(mmap_event->flags); | |
13d7a241 SE |
5193 | } |
5194 | ||
c980d109 ACM |
5195 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5196 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5197 | mmap_event->event_id.header.size); |
0a4a9391 | 5198 | if (ret) |
c980d109 | 5199 | goto out; |
0a4a9391 | 5200 | |
cdd6c482 IM |
5201 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5202 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5203 | |
cdd6c482 | 5204 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5205 | |
5206 | if (event->attr.mmap2) { | |
5207 | perf_output_put(&handle, mmap_event->maj); | |
5208 | perf_output_put(&handle, mmap_event->min); | |
5209 | perf_output_put(&handle, mmap_event->ino); | |
5210 | perf_output_put(&handle, mmap_event->ino_generation); | |
f972eb63 PZ |
5211 | perf_output_put(&handle, mmap_event->prot); |
5212 | perf_output_put(&handle, mmap_event->flags); | |
13d7a241 SE |
5213 | } |
5214 | ||
76369139 | 5215 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5216 | mmap_event->file_size); |
c980d109 ACM |
5217 | |
5218 | perf_event__output_id_sample(event, &handle, &sample); | |
5219 | ||
78d613eb | 5220 | perf_output_end(&handle); |
c980d109 ACM |
5221 | out: |
5222 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5223 | } |
5224 | ||
cdd6c482 | 5225 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5226 | { |
089dd79d PZ |
5227 | struct vm_area_struct *vma = mmap_event->vma; |
5228 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5229 | int maj = 0, min = 0; |
5230 | u64 ino = 0, gen = 0; | |
f972eb63 | 5231 | u32 prot = 0, flags = 0; |
0a4a9391 PZ |
5232 | unsigned int size; |
5233 | char tmp[16]; | |
5234 | char *buf = NULL; | |
2c42cfbf | 5235 | char *name; |
413ee3b4 | 5236 | |
0a4a9391 | 5237 | if (file) { |
13d7a241 SE |
5238 | struct inode *inode; |
5239 | dev_t dev; | |
3ea2f2b9 | 5240 | |
2c42cfbf | 5241 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
0a4a9391 | 5242 | if (!buf) { |
c7e548b4 ON |
5243 | name = "//enomem"; |
5244 | goto cpy_name; | |
0a4a9391 | 5245 | } |
413ee3b4 | 5246 | /* |
3ea2f2b9 | 5247 | * d_path() works from the end of the rb backwards, so we |
413ee3b4 AB |
5248 | * need to add enough zero bytes after the string to handle |
5249 | * the 64bit alignment we do later. | |
5250 | */ | |
3ea2f2b9 | 5251 | name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64)); |
0a4a9391 | 5252 | if (IS_ERR(name)) { |
c7e548b4 ON |
5253 | name = "//toolong"; |
5254 | goto cpy_name; | |
0a4a9391 | 5255 | } |
13d7a241 SE |
5256 | inode = file_inode(vma->vm_file); |
5257 | dev = inode->i_sb->s_dev; | |
5258 | ino = inode->i_ino; | |
5259 | gen = inode->i_generation; | |
5260 | maj = MAJOR(dev); | |
5261 | min = MINOR(dev); | |
f972eb63 PZ |
5262 | |
5263 | if (vma->vm_flags & VM_READ) | |
5264 | prot |= PROT_READ; | |
5265 | if (vma->vm_flags & VM_WRITE) | |
5266 | prot |= PROT_WRITE; | |
5267 | if (vma->vm_flags & VM_EXEC) | |
5268 | prot |= PROT_EXEC; | |
5269 | ||
5270 | if (vma->vm_flags & VM_MAYSHARE) | |
5271 | flags = MAP_SHARED; | |
5272 | else | |
5273 | flags = MAP_PRIVATE; | |
5274 | ||
5275 | if (vma->vm_flags & VM_DENYWRITE) | |
5276 | flags |= MAP_DENYWRITE; | |
5277 | if (vma->vm_flags & VM_MAYEXEC) | |
5278 | flags |= MAP_EXECUTABLE; | |
5279 | if (vma->vm_flags & VM_LOCKED) | |
5280 | flags |= MAP_LOCKED; | |
5281 | if (vma->vm_flags & VM_HUGETLB) | |
5282 | flags |= MAP_HUGETLB; | |
5283 | ||
c7e548b4 | 5284 | goto got_name; |
0a4a9391 | 5285 | } else { |
fbe26abe JO |
5286 | if (vma->vm_ops && vma->vm_ops->name) { |
5287 | name = (char *) vma->vm_ops->name(vma); | |
5288 | if (name) | |
5289 | goto cpy_name; | |
5290 | } | |
5291 | ||
2c42cfbf | 5292 | name = (char *)arch_vma_name(vma); |
c7e548b4 ON |
5293 | if (name) |
5294 | goto cpy_name; | |
089dd79d | 5295 | |
32c5fb7e | 5296 | if (vma->vm_start <= vma->vm_mm->start_brk && |
3af9e859 | 5297 | vma->vm_end >= vma->vm_mm->brk) { |
c7e548b4 ON |
5298 | name = "[heap]"; |
5299 | goto cpy_name; | |
32c5fb7e ON |
5300 | } |
5301 | if (vma->vm_start <= vma->vm_mm->start_stack && | |
3af9e859 | 5302 | vma->vm_end >= vma->vm_mm->start_stack) { |
c7e548b4 ON |
5303 | name = "[stack]"; |
5304 | goto cpy_name; | |
089dd79d PZ |
5305 | } |
5306 | ||
c7e548b4 ON |
5307 | name = "//anon"; |
5308 | goto cpy_name; | |
0a4a9391 PZ |
5309 | } |
5310 | ||
c7e548b4 ON |
5311 | cpy_name: |
5312 | strlcpy(tmp, name, sizeof(tmp)); | |
5313 | name = tmp; | |
0a4a9391 | 5314 | got_name: |
2c42cfbf PZ |
5315 | /* |
5316 | * Since our buffer works in 8 byte units we need to align our string | |
5317 | * size to a multiple of 8. However, we must guarantee the tail end is | |
5318 | * zero'd out to avoid leaking random bits to userspace. | |
5319 | */ | |
5320 | size = strlen(name)+1; | |
5321 | while (!IS_ALIGNED(size, sizeof(u64))) | |
5322 | name[size++] = '\0'; | |
0a4a9391 PZ |
5323 | |
5324 | mmap_event->file_name = name; | |
5325 | mmap_event->file_size = size; | |
13d7a241 SE |
5326 | mmap_event->maj = maj; |
5327 | mmap_event->min = min; | |
5328 | mmap_event->ino = ino; | |
5329 | mmap_event->ino_generation = gen; | |
f972eb63 PZ |
5330 | mmap_event->prot = prot; |
5331 | mmap_event->flags = flags; | |
0a4a9391 | 5332 | |
2fe85427 SE |
5333 | if (!(vma->vm_flags & VM_EXEC)) |
5334 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5335 | ||
cdd6c482 | 5336 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5337 | |
67516844 | 5338 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5339 | mmap_event, |
5340 | NULL); | |
665c2142 | 5341 | |
0a4a9391 PZ |
5342 | kfree(buf); |
5343 | } | |
5344 | ||
3af9e859 | 5345 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5346 | { |
9ee318a7 PZ |
5347 | struct perf_mmap_event mmap_event; |
5348 | ||
cdd6c482 | 5349 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5350 | return; |
5351 | ||
5352 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5353 | .vma = vma, |
573402db PZ |
5354 | /* .file_name */ |
5355 | /* .file_size */ | |
cdd6c482 | 5356 | .event_id = { |
573402db | 5357 | .header = { |
cdd6c482 | 5358 | .type = PERF_RECORD_MMAP, |
39447b38 | 5359 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5360 | /* .size */ |
5361 | }, | |
5362 | /* .pid */ | |
5363 | /* .tid */ | |
089dd79d PZ |
5364 | .start = vma->vm_start, |
5365 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5366 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5367 | }, |
13d7a241 SE |
5368 | /* .maj (attr_mmap2 only) */ |
5369 | /* .min (attr_mmap2 only) */ | |
5370 | /* .ino (attr_mmap2 only) */ | |
5371 | /* .ino_generation (attr_mmap2 only) */ | |
f972eb63 PZ |
5372 | /* .prot (attr_mmap2 only) */ |
5373 | /* .flags (attr_mmap2 only) */ | |
0a4a9391 PZ |
5374 | }; |
5375 | ||
cdd6c482 | 5376 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5377 | } |
5378 | ||
a78ac325 PZ |
5379 | /* |
5380 | * IRQ throttle logging | |
5381 | */ | |
5382 | ||
cdd6c482 | 5383 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5384 | { |
5385 | struct perf_output_handle handle; | |
c980d109 | 5386 | struct perf_sample_data sample; |
a78ac325 PZ |
5387 | int ret; |
5388 | ||
5389 | struct { | |
5390 | struct perf_event_header header; | |
5391 | u64 time; | |
cca3f454 | 5392 | u64 id; |
7f453c24 | 5393 | u64 stream_id; |
a78ac325 PZ |
5394 | } throttle_event = { |
5395 | .header = { | |
cdd6c482 | 5396 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5397 | .misc = 0, |
5398 | .size = sizeof(throttle_event), | |
5399 | }, | |
def0a9b2 | 5400 | .time = perf_clock(), |
cdd6c482 IM |
5401 | .id = primary_event_id(event), |
5402 | .stream_id = event->id, | |
a78ac325 PZ |
5403 | }; |
5404 | ||
966ee4d6 | 5405 | if (enable) |
cdd6c482 | 5406 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5407 | |
c980d109 ACM |
5408 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5409 | ||
5410 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5411 | throttle_event.header.size); |
a78ac325 PZ |
5412 | if (ret) |
5413 | return; | |
5414 | ||
5415 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5416 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5417 | perf_output_end(&handle); |
5418 | } | |
5419 | ||
f6c7d5fe | 5420 | /* |
cdd6c482 | 5421 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5422 | */ |
5423 | ||
a8b0ca17 | 5424 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5425 | int throttle, struct perf_sample_data *data, |
5426 | struct pt_regs *regs) | |
f6c7d5fe | 5427 | { |
cdd6c482 IM |
5428 | int events = atomic_read(&event->event_limit); |
5429 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5430 | u64 seq; |
79f14641 PZ |
5431 | int ret = 0; |
5432 | ||
96398826 PZ |
5433 | /* |
5434 | * Non-sampling counters might still use the PMI to fold short | |
5435 | * hardware counters, ignore those. | |
5436 | */ | |
5437 | if (unlikely(!is_sampling_event(event))) | |
5438 | return 0; | |
5439 | ||
e050e3f0 SE |
5440 | seq = __this_cpu_read(perf_throttled_seq); |
5441 | if (seq != hwc->interrupts_seq) { | |
5442 | hwc->interrupts_seq = seq; | |
5443 | hwc->interrupts = 1; | |
5444 | } else { | |
5445 | hwc->interrupts++; | |
5446 | if (unlikely(throttle | |
5447 | && hwc->interrupts >= max_samples_per_tick)) { | |
5448 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5449 | hwc->interrupts = MAX_INTERRUPTS; |
5450 | perf_log_throttle(event, 0); | |
d84153d6 | 5451 | tick_nohz_full_kick(); |
a78ac325 PZ |
5452 | ret = 1; |
5453 | } | |
e050e3f0 | 5454 | } |
60db5e09 | 5455 | |
cdd6c482 | 5456 | if (event->attr.freq) { |
def0a9b2 | 5457 | u64 now = perf_clock(); |
abd50713 | 5458 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5459 | |
abd50713 | 5460 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5461 | |
abd50713 | 5462 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5463 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5464 | } |
5465 | ||
2023b359 PZ |
5466 | /* |
5467 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5468 | * events |
2023b359 PZ |
5469 | */ |
5470 | ||
cdd6c482 IM |
5471 | event->pending_kill = POLL_IN; |
5472 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5473 | ret = 1; |
cdd6c482 | 5474 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5475 | event->pending_disable = 1; |
5476 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5477 | } |
5478 | ||
453f19ee | 5479 | if (event->overflow_handler) |
a8b0ca17 | 5480 | event->overflow_handler(event, data, regs); |
453f19ee | 5481 | else |
a8b0ca17 | 5482 | perf_event_output(event, data, regs); |
453f19ee | 5483 | |
f506b3dc | 5484 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5485 | event->pending_wakeup = 1; |
5486 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5487 | } |
5488 | ||
79f14641 | 5489 | return ret; |
f6c7d5fe PZ |
5490 | } |
5491 | ||
a8b0ca17 | 5492 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5493 | struct perf_sample_data *data, |
5494 | struct pt_regs *regs) | |
850bc73f | 5495 | { |
a8b0ca17 | 5496 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5497 | } |
5498 | ||
15dbf27c | 5499 | /* |
cdd6c482 | 5500 | * Generic software event infrastructure |
15dbf27c PZ |
5501 | */ |
5502 | ||
b28ab83c PZ |
5503 | struct swevent_htable { |
5504 | struct swevent_hlist *swevent_hlist; | |
5505 | struct mutex hlist_mutex; | |
5506 | int hlist_refcount; | |
5507 | ||
5508 | /* Recursion avoidance in each contexts */ | |
5509 | int recursion[PERF_NR_CONTEXTS]; | |
39af6b16 JO |
5510 | |
5511 | /* Keeps track of cpu being initialized/exited */ | |
5512 | bool online; | |
b28ab83c PZ |
5513 | }; |
5514 | ||
5515 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5516 | ||
7b4b6658 | 5517 | /* |
cdd6c482 IM |
5518 | * We directly increment event->count and keep a second value in |
5519 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5520 | * is kept in the range [-sample_period, 0] so that we can use the |
5521 | * sign as trigger. | |
5522 | */ | |
5523 | ||
ab573844 | 5524 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5525 | { |
cdd6c482 | 5526 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5527 | u64 period = hwc->last_period; |
5528 | u64 nr, offset; | |
5529 | s64 old, val; | |
5530 | ||
5531 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5532 | |
5533 | again: | |
e7850595 | 5534 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5535 | if (val < 0) |
5536 | return 0; | |
15dbf27c | 5537 | |
7b4b6658 PZ |
5538 | nr = div64_u64(period + val, period); |
5539 | offset = nr * period; | |
5540 | val -= offset; | |
e7850595 | 5541 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5542 | goto again; |
15dbf27c | 5543 | |
7b4b6658 | 5544 | return nr; |
15dbf27c PZ |
5545 | } |
5546 | ||
0cff784a | 5547 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5548 | struct perf_sample_data *data, |
5622f295 | 5549 | struct pt_regs *regs) |
15dbf27c | 5550 | { |
cdd6c482 | 5551 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5552 | int throttle = 0; |
15dbf27c | 5553 | |
0cff784a PZ |
5554 | if (!overflow) |
5555 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5556 | |
7b4b6658 PZ |
5557 | if (hwc->interrupts == MAX_INTERRUPTS) |
5558 | return; | |
15dbf27c | 5559 | |
7b4b6658 | 5560 | for (; overflow; overflow--) { |
a8b0ca17 | 5561 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5562 | data, regs)) { |
7b4b6658 PZ |
5563 | /* |
5564 | * We inhibit the overflow from happening when | |
5565 | * hwc->interrupts == MAX_INTERRUPTS. | |
5566 | */ | |
5567 | break; | |
5568 | } | |
cf450a73 | 5569 | throttle = 1; |
7b4b6658 | 5570 | } |
15dbf27c PZ |
5571 | } |
5572 | ||
a4eaf7f1 | 5573 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5574 | struct perf_sample_data *data, |
5622f295 | 5575 | struct pt_regs *regs) |
7b4b6658 | 5576 | { |
cdd6c482 | 5577 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5578 | |
e7850595 | 5579 | local64_add(nr, &event->count); |
d6d020e9 | 5580 | |
0cff784a PZ |
5581 | if (!regs) |
5582 | return; | |
5583 | ||
6c7e550f | 5584 | if (!is_sampling_event(event)) |
7b4b6658 | 5585 | return; |
d6d020e9 | 5586 | |
5d81e5cf AV |
5587 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5588 | data->period = nr; | |
5589 | return perf_swevent_overflow(event, 1, data, regs); | |
5590 | } else | |
5591 | data->period = event->hw.last_period; | |
5592 | ||
0cff784a | 5593 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5594 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5595 | |
e7850595 | 5596 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5597 | return; |
df1a132b | 5598 | |
a8b0ca17 | 5599 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5600 | } |
5601 | ||
f5ffe02e FW |
5602 | static int perf_exclude_event(struct perf_event *event, |
5603 | struct pt_regs *regs) | |
5604 | { | |
a4eaf7f1 | 5605 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5606 | return 1; |
a4eaf7f1 | 5607 | |
f5ffe02e FW |
5608 | if (regs) { |
5609 | if (event->attr.exclude_user && user_mode(regs)) | |
5610 | return 1; | |
5611 | ||
5612 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5613 | return 1; | |
5614 | } | |
5615 | ||
5616 | return 0; | |
5617 | } | |
5618 | ||
cdd6c482 | 5619 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5620 | enum perf_type_id type, |
6fb2915d LZ |
5621 | u32 event_id, |
5622 | struct perf_sample_data *data, | |
5623 | struct pt_regs *regs) | |
15dbf27c | 5624 | { |
cdd6c482 | 5625 | if (event->attr.type != type) |
a21ca2ca | 5626 | return 0; |
f5ffe02e | 5627 | |
cdd6c482 | 5628 | if (event->attr.config != event_id) |
15dbf27c PZ |
5629 | return 0; |
5630 | ||
f5ffe02e FW |
5631 | if (perf_exclude_event(event, regs)) |
5632 | return 0; | |
15dbf27c PZ |
5633 | |
5634 | return 1; | |
5635 | } | |
5636 | ||
76e1d904 FW |
5637 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5638 | { | |
5639 | u64 val = event_id | (type << 32); | |
5640 | ||
5641 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5642 | } | |
5643 | ||
49f135ed FW |
5644 | static inline struct hlist_head * |
5645 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5646 | { |
49f135ed FW |
5647 | u64 hash = swevent_hash(type, event_id); |
5648 | ||
5649 | return &hlist->heads[hash]; | |
5650 | } | |
76e1d904 | 5651 | |
49f135ed FW |
5652 | /* For the read side: events when they trigger */ |
5653 | static inline struct hlist_head * | |
b28ab83c | 5654 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5655 | { |
5656 | struct swevent_hlist *hlist; | |
76e1d904 | 5657 | |
b28ab83c | 5658 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5659 | if (!hlist) |
5660 | return NULL; | |
5661 | ||
49f135ed FW |
5662 | return __find_swevent_head(hlist, type, event_id); |
5663 | } | |
5664 | ||
5665 | /* For the event head insertion and removal in the hlist */ | |
5666 | static inline struct hlist_head * | |
b28ab83c | 5667 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5668 | { |
5669 | struct swevent_hlist *hlist; | |
5670 | u32 event_id = event->attr.config; | |
5671 | u64 type = event->attr.type; | |
5672 | ||
5673 | /* | |
5674 | * Event scheduling is always serialized against hlist allocation | |
5675 | * and release. Which makes the protected version suitable here. | |
5676 | * The context lock guarantees that. | |
5677 | */ | |
b28ab83c | 5678 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5679 | lockdep_is_held(&event->ctx->lock)); |
5680 | if (!hlist) | |
5681 | return NULL; | |
5682 | ||
5683 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5684 | } |
5685 | ||
5686 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5687 | u64 nr, |
76e1d904 FW |
5688 | struct perf_sample_data *data, |
5689 | struct pt_regs *regs) | |
15dbf27c | 5690 | { |
b28ab83c | 5691 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5692 | struct perf_event *event; |
76e1d904 | 5693 | struct hlist_head *head; |
15dbf27c | 5694 | |
76e1d904 | 5695 | rcu_read_lock(); |
b28ab83c | 5696 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5697 | if (!head) |
5698 | goto end; | |
5699 | ||
b67bfe0d | 5700 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5701 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5702 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5703 | } |
76e1d904 FW |
5704 | end: |
5705 | rcu_read_unlock(); | |
15dbf27c PZ |
5706 | } |
5707 | ||
4ed7c92d | 5708 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5709 | { |
b28ab83c | 5710 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5711 | |
b28ab83c | 5712 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5713 | } |
645e8cc0 | 5714 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5715 | |
fa9f90be | 5716 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5717 | { |
b28ab83c | 5718 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5719 | |
b28ab83c | 5720 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5721 | } |
15dbf27c | 5722 | |
a8b0ca17 | 5723 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5724 | { |
a4234bfc | 5725 | struct perf_sample_data data; |
4ed7c92d PZ |
5726 | int rctx; |
5727 | ||
1c024eca | 5728 | preempt_disable_notrace(); |
4ed7c92d PZ |
5729 | rctx = perf_swevent_get_recursion_context(); |
5730 | if (rctx < 0) | |
5731 | return; | |
a4234bfc | 5732 | |
fd0d000b | 5733 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5734 | |
a8b0ca17 | 5735 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5736 | |
5737 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5738 | preempt_enable_notrace(); |
b8e83514 PZ |
5739 | } |
5740 | ||
cdd6c482 | 5741 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5742 | { |
15dbf27c PZ |
5743 | } |
5744 | ||
a4eaf7f1 | 5745 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5746 | { |
b28ab83c | 5747 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5748 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5749 | struct hlist_head *head; |
5750 | ||
6c7e550f | 5751 | if (is_sampling_event(event)) { |
7b4b6658 | 5752 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5753 | perf_swevent_set_period(event); |
7b4b6658 | 5754 | } |
76e1d904 | 5755 | |
a4eaf7f1 PZ |
5756 | hwc->state = !(flags & PERF_EF_START); |
5757 | ||
b28ab83c | 5758 | head = find_swevent_head(swhash, event); |
39af6b16 JO |
5759 | if (!head) { |
5760 | /* | |
5761 | * We can race with cpu hotplug code. Do not | |
5762 | * WARN if the cpu just got unplugged. | |
5763 | */ | |
5764 | WARN_ON_ONCE(swhash->online); | |
76e1d904 | 5765 | return -EINVAL; |
39af6b16 | 5766 | } |
76e1d904 FW |
5767 | |
5768 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5769 | ||
15dbf27c PZ |
5770 | return 0; |
5771 | } | |
5772 | ||
a4eaf7f1 | 5773 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5774 | { |
76e1d904 | 5775 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5776 | } |
5777 | ||
a4eaf7f1 | 5778 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5779 | { |
a4eaf7f1 | 5780 | event->hw.state = 0; |
d6d020e9 | 5781 | } |
aa9c4c0f | 5782 | |
a4eaf7f1 | 5783 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5784 | { |
a4eaf7f1 | 5785 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5786 | } |
5787 | ||
49f135ed FW |
5788 | /* Deref the hlist from the update side */ |
5789 | static inline struct swevent_hlist * | |
b28ab83c | 5790 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5791 | { |
b28ab83c PZ |
5792 | return rcu_dereference_protected(swhash->swevent_hlist, |
5793 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5794 | } |
5795 | ||
b28ab83c | 5796 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5797 | { |
b28ab83c | 5798 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5799 | |
49f135ed | 5800 | if (!hlist) |
76e1d904 FW |
5801 | return; |
5802 | ||
b28ab83c | 5803 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5804 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5805 | } |
5806 | ||
5807 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5808 | { | |
b28ab83c | 5809 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5810 | |
b28ab83c | 5811 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5812 | |
b28ab83c PZ |
5813 | if (!--swhash->hlist_refcount) |
5814 | swevent_hlist_release(swhash); | |
76e1d904 | 5815 | |
b28ab83c | 5816 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5817 | } |
5818 | ||
5819 | static void swevent_hlist_put(struct perf_event *event) | |
5820 | { | |
5821 | int cpu; | |
5822 | ||
76e1d904 FW |
5823 | for_each_possible_cpu(cpu) |
5824 | swevent_hlist_put_cpu(event, cpu); | |
5825 | } | |
5826 | ||
5827 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5828 | { | |
b28ab83c | 5829 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5830 | int err = 0; |
5831 | ||
b28ab83c | 5832 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5833 | |
b28ab83c | 5834 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5835 | struct swevent_hlist *hlist; |
5836 | ||
5837 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5838 | if (!hlist) { | |
5839 | err = -ENOMEM; | |
5840 | goto exit; | |
5841 | } | |
b28ab83c | 5842 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5843 | } |
b28ab83c | 5844 | swhash->hlist_refcount++; |
9ed6060d | 5845 | exit: |
b28ab83c | 5846 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5847 | |
5848 | return err; | |
5849 | } | |
5850 | ||
5851 | static int swevent_hlist_get(struct perf_event *event) | |
5852 | { | |
5853 | int err; | |
5854 | int cpu, failed_cpu; | |
5855 | ||
76e1d904 FW |
5856 | get_online_cpus(); |
5857 | for_each_possible_cpu(cpu) { | |
5858 | err = swevent_hlist_get_cpu(event, cpu); | |
5859 | if (err) { | |
5860 | failed_cpu = cpu; | |
5861 | goto fail; | |
5862 | } | |
5863 | } | |
5864 | put_online_cpus(); | |
5865 | ||
5866 | return 0; | |
9ed6060d | 5867 | fail: |
76e1d904 FW |
5868 | for_each_possible_cpu(cpu) { |
5869 | if (cpu == failed_cpu) | |
5870 | break; | |
5871 | swevent_hlist_put_cpu(event, cpu); | |
5872 | } | |
5873 | ||
5874 | put_online_cpus(); | |
5875 | return err; | |
5876 | } | |
5877 | ||
c5905afb | 5878 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5879 | |
b0a873eb PZ |
5880 | static void sw_perf_event_destroy(struct perf_event *event) |
5881 | { | |
5882 | u64 event_id = event->attr.config; | |
95476b64 | 5883 | |
b0a873eb PZ |
5884 | WARN_ON(event->parent); |
5885 | ||
c5905afb | 5886 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5887 | swevent_hlist_put(event); |
5888 | } | |
5889 | ||
5890 | static int perf_swevent_init(struct perf_event *event) | |
5891 | { | |
8176cced | 5892 | u64 event_id = event->attr.config; |
b0a873eb PZ |
5893 | |
5894 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5895 | return -ENOENT; | |
5896 | ||
2481c5fa SE |
5897 | /* |
5898 | * no branch sampling for software events | |
5899 | */ | |
5900 | if (has_branch_stack(event)) | |
5901 | return -EOPNOTSUPP; | |
5902 | ||
b0a873eb PZ |
5903 | switch (event_id) { |
5904 | case PERF_COUNT_SW_CPU_CLOCK: | |
5905 | case PERF_COUNT_SW_TASK_CLOCK: | |
5906 | return -ENOENT; | |
5907 | ||
5908 | default: | |
5909 | break; | |
5910 | } | |
5911 | ||
ce677831 | 5912 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5913 | return -ENOENT; |
5914 | ||
5915 | if (!event->parent) { | |
5916 | int err; | |
5917 | ||
5918 | err = swevent_hlist_get(event); | |
5919 | if (err) | |
5920 | return err; | |
5921 | ||
c5905afb | 5922 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5923 | event->destroy = sw_perf_event_destroy; |
5924 | } | |
5925 | ||
5926 | return 0; | |
5927 | } | |
5928 | ||
35edc2a5 PZ |
5929 | static int perf_swevent_event_idx(struct perf_event *event) |
5930 | { | |
5931 | return 0; | |
5932 | } | |
5933 | ||
b0a873eb | 5934 | static struct pmu perf_swevent = { |
89a1e187 | 5935 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5936 | |
b0a873eb | 5937 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5938 | .add = perf_swevent_add, |
5939 | .del = perf_swevent_del, | |
5940 | .start = perf_swevent_start, | |
5941 | .stop = perf_swevent_stop, | |
1c024eca | 5942 | .read = perf_swevent_read, |
35edc2a5 PZ |
5943 | |
5944 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5945 | }; |
5946 | ||
b0a873eb PZ |
5947 | #ifdef CONFIG_EVENT_TRACING |
5948 | ||
1c024eca PZ |
5949 | static int perf_tp_filter_match(struct perf_event *event, |
5950 | struct perf_sample_data *data) | |
5951 | { | |
5952 | void *record = data->raw->data; | |
5953 | ||
5954 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5955 | return 1; | |
5956 | return 0; | |
5957 | } | |
5958 | ||
5959 | static int perf_tp_event_match(struct perf_event *event, | |
5960 | struct perf_sample_data *data, | |
5961 | struct pt_regs *regs) | |
5962 | { | |
a0f7d0f7 FW |
5963 | if (event->hw.state & PERF_HES_STOPPED) |
5964 | return 0; | |
580d607c PZ |
5965 | /* |
5966 | * All tracepoints are from kernel-space. | |
5967 | */ | |
5968 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5969 | return 0; |
5970 | ||
5971 | if (!perf_tp_filter_match(event, data)) | |
5972 | return 0; | |
5973 | ||
5974 | return 1; | |
5975 | } | |
5976 | ||
5977 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5978 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5979 | struct task_struct *task) | |
95476b64 FW |
5980 | { |
5981 | struct perf_sample_data data; | |
1c024eca | 5982 | struct perf_event *event; |
1c024eca | 5983 | |
95476b64 FW |
5984 | struct perf_raw_record raw = { |
5985 | .size = entry_size, | |
5986 | .data = record, | |
5987 | }; | |
5988 | ||
fd0d000b | 5989 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5990 | data.raw = &raw; |
5991 | ||
b67bfe0d | 5992 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 5993 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 5994 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5995 | } |
ecc55f84 | 5996 | |
e6dab5ff AV |
5997 | /* |
5998 | * If we got specified a target task, also iterate its context and | |
5999 | * deliver this event there too. | |
6000 | */ | |
6001 | if (task && task != current) { | |
6002 | struct perf_event_context *ctx; | |
6003 | struct trace_entry *entry = record; | |
6004 | ||
6005 | rcu_read_lock(); | |
6006 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
6007 | if (!ctx) | |
6008 | goto unlock; | |
6009 | ||
6010 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
6011 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6012 | continue; | |
6013 | if (event->attr.config != entry->type) | |
6014 | continue; | |
6015 | if (perf_tp_event_match(event, &data, regs)) | |
6016 | perf_swevent_event(event, count, &data, regs); | |
6017 | } | |
6018 | unlock: | |
6019 | rcu_read_unlock(); | |
6020 | } | |
6021 | ||
ecc55f84 | 6022 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
6023 | } |
6024 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
6025 | ||
cdd6c482 | 6026 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 6027 | { |
1c024eca | 6028 | perf_trace_destroy(event); |
e077df4f PZ |
6029 | } |
6030 | ||
b0a873eb | 6031 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 6032 | { |
76e1d904 FW |
6033 | int err; |
6034 | ||
b0a873eb PZ |
6035 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
6036 | return -ENOENT; | |
6037 | ||
2481c5fa SE |
6038 | /* |
6039 | * no branch sampling for tracepoint events | |
6040 | */ | |
6041 | if (has_branch_stack(event)) | |
6042 | return -EOPNOTSUPP; | |
6043 | ||
1c024eca PZ |
6044 | err = perf_trace_init(event); |
6045 | if (err) | |
b0a873eb | 6046 | return err; |
e077df4f | 6047 | |
cdd6c482 | 6048 | event->destroy = tp_perf_event_destroy; |
e077df4f | 6049 | |
b0a873eb PZ |
6050 | return 0; |
6051 | } | |
6052 | ||
6053 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
6054 | .task_ctx_nr = perf_sw_context, |
6055 | ||
b0a873eb | 6056 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
6057 | .add = perf_trace_add, |
6058 | .del = perf_trace_del, | |
6059 | .start = perf_swevent_start, | |
6060 | .stop = perf_swevent_stop, | |
b0a873eb | 6061 | .read = perf_swevent_read, |
35edc2a5 PZ |
6062 | |
6063 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
6064 | }; |
6065 | ||
6066 | static inline void perf_tp_register(void) | |
6067 | { | |
2e80a82a | 6068 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 6069 | } |
6fb2915d LZ |
6070 | |
6071 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6072 | { | |
6073 | char *filter_str; | |
6074 | int ret; | |
6075 | ||
6076 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6077 | return -EINVAL; | |
6078 | ||
6079 | filter_str = strndup_user(arg, PAGE_SIZE); | |
6080 | if (IS_ERR(filter_str)) | |
6081 | return PTR_ERR(filter_str); | |
6082 | ||
6083 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
6084 | ||
6085 | kfree(filter_str); | |
6086 | return ret; | |
6087 | } | |
6088 | ||
6089 | static void perf_event_free_filter(struct perf_event *event) | |
6090 | { | |
6091 | ftrace_profile_free_filter(event); | |
6092 | } | |
6093 | ||
e077df4f | 6094 | #else |
6fb2915d | 6095 | |
b0a873eb | 6096 | static inline void perf_tp_register(void) |
e077df4f | 6097 | { |
e077df4f | 6098 | } |
6fb2915d LZ |
6099 | |
6100 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6101 | { | |
6102 | return -ENOENT; | |
6103 | } | |
6104 | ||
6105 | static void perf_event_free_filter(struct perf_event *event) | |
6106 | { | |
6107 | } | |
6108 | ||
07b139c8 | 6109 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 6110 | |
24f1e32c | 6111 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 6112 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 6113 | { |
f5ffe02e FW |
6114 | struct perf_sample_data sample; |
6115 | struct pt_regs *regs = data; | |
6116 | ||
fd0d000b | 6117 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 6118 | |
a4eaf7f1 | 6119 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 6120 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
6121 | } |
6122 | #endif | |
6123 | ||
b0a873eb PZ |
6124 | /* |
6125 | * hrtimer based swevent callback | |
6126 | */ | |
f29ac756 | 6127 | |
b0a873eb | 6128 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 6129 | { |
b0a873eb PZ |
6130 | enum hrtimer_restart ret = HRTIMER_RESTART; |
6131 | struct perf_sample_data data; | |
6132 | struct pt_regs *regs; | |
6133 | struct perf_event *event; | |
6134 | u64 period; | |
f29ac756 | 6135 | |
b0a873eb | 6136 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
6137 | |
6138 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
6139 | return HRTIMER_NORESTART; | |
6140 | ||
b0a873eb | 6141 | event->pmu->read(event); |
f344011c | 6142 | |
fd0d000b | 6143 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
6144 | regs = get_irq_regs(); |
6145 | ||
6146 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 6147 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 6148 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
6149 | ret = HRTIMER_NORESTART; |
6150 | } | |
24f1e32c | 6151 | |
b0a873eb PZ |
6152 | period = max_t(u64, 10000, event->hw.sample_period); |
6153 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 6154 | |
b0a873eb | 6155 | return ret; |
f29ac756 PZ |
6156 | } |
6157 | ||
b0a873eb | 6158 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 6159 | { |
b0a873eb | 6160 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
6161 | s64 period; |
6162 | ||
6163 | if (!is_sampling_event(event)) | |
6164 | return; | |
f5ffe02e | 6165 | |
5d508e82 FBH |
6166 | period = local64_read(&hwc->period_left); |
6167 | if (period) { | |
6168 | if (period < 0) | |
6169 | period = 10000; | |
fa407f35 | 6170 | |
5d508e82 FBH |
6171 | local64_set(&hwc->period_left, 0); |
6172 | } else { | |
6173 | period = max_t(u64, 10000, hwc->sample_period); | |
6174 | } | |
6175 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 6176 | ns_to_ktime(period), 0, |
b5ab4cd5 | 6177 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 6178 | } |
b0a873eb PZ |
6179 | |
6180 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6181 | { |
b0a873eb PZ |
6182 | struct hw_perf_event *hwc = &event->hw; |
6183 | ||
6c7e550f | 6184 | if (is_sampling_event(event)) { |
b0a873eb | 6185 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6186 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6187 | |
6188 | hrtimer_cancel(&hwc->hrtimer); | |
6189 | } | |
24f1e32c FW |
6190 | } |
6191 | ||
ba3dd36c PZ |
6192 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6193 | { | |
6194 | struct hw_perf_event *hwc = &event->hw; | |
6195 | ||
6196 | if (!is_sampling_event(event)) | |
6197 | return; | |
6198 | ||
6199 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6200 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6201 | ||
6202 | /* | |
6203 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6204 | * mapping and avoid the whole period adjust feedback stuff. | |
6205 | */ | |
6206 | if (event->attr.freq) { | |
6207 | long freq = event->attr.sample_freq; | |
6208 | ||
6209 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
6210 | hwc->sample_period = event->attr.sample_period; | |
6211 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 6212 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
6213 | event->attr.freq = 0; |
6214 | } | |
6215 | } | |
6216 | ||
b0a873eb PZ |
6217 | /* |
6218 | * Software event: cpu wall time clock | |
6219 | */ | |
6220 | ||
6221 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 6222 | { |
b0a873eb PZ |
6223 | s64 prev; |
6224 | u64 now; | |
6225 | ||
a4eaf7f1 | 6226 | now = local_clock(); |
b0a873eb PZ |
6227 | prev = local64_xchg(&event->hw.prev_count, now); |
6228 | local64_add(now - prev, &event->count); | |
24f1e32c | 6229 | } |
24f1e32c | 6230 | |
a4eaf7f1 | 6231 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6232 | { |
a4eaf7f1 | 6233 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 6234 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6235 | } |
6236 | ||
a4eaf7f1 | 6237 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6238 | { |
b0a873eb PZ |
6239 | perf_swevent_cancel_hrtimer(event); |
6240 | cpu_clock_event_update(event); | |
6241 | } | |
f29ac756 | 6242 | |
a4eaf7f1 PZ |
6243 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6244 | { | |
6245 | if (flags & PERF_EF_START) | |
6246 | cpu_clock_event_start(event, flags); | |
6247 | ||
6248 | return 0; | |
6249 | } | |
6250 | ||
6251 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6252 | { | |
6253 | cpu_clock_event_stop(event, flags); | |
6254 | } | |
6255 | ||
b0a873eb PZ |
6256 | static void cpu_clock_event_read(struct perf_event *event) |
6257 | { | |
6258 | cpu_clock_event_update(event); | |
6259 | } | |
f344011c | 6260 | |
b0a873eb PZ |
6261 | static int cpu_clock_event_init(struct perf_event *event) |
6262 | { | |
6263 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6264 | return -ENOENT; | |
6265 | ||
6266 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
6267 | return -ENOENT; | |
6268 | ||
2481c5fa SE |
6269 | /* |
6270 | * no branch sampling for software events | |
6271 | */ | |
6272 | if (has_branch_stack(event)) | |
6273 | return -EOPNOTSUPP; | |
6274 | ||
ba3dd36c PZ |
6275 | perf_swevent_init_hrtimer(event); |
6276 | ||
b0a873eb | 6277 | return 0; |
f29ac756 PZ |
6278 | } |
6279 | ||
b0a873eb | 6280 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
6281 | .task_ctx_nr = perf_sw_context, |
6282 | ||
b0a873eb | 6283 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
6284 | .add = cpu_clock_event_add, |
6285 | .del = cpu_clock_event_del, | |
6286 | .start = cpu_clock_event_start, | |
6287 | .stop = cpu_clock_event_stop, | |
b0a873eb | 6288 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
6289 | |
6290 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
6291 | }; |
6292 | ||
6293 | /* | |
6294 | * Software event: task time clock | |
6295 | */ | |
6296 | ||
6297 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 6298 | { |
b0a873eb PZ |
6299 | u64 prev; |
6300 | s64 delta; | |
5c92d124 | 6301 | |
b0a873eb PZ |
6302 | prev = local64_xchg(&event->hw.prev_count, now); |
6303 | delta = now - prev; | |
6304 | local64_add(delta, &event->count); | |
6305 | } | |
5c92d124 | 6306 | |
a4eaf7f1 | 6307 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6308 | { |
a4eaf7f1 | 6309 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 6310 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6311 | } |
6312 | ||
a4eaf7f1 | 6313 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
6314 | { |
6315 | perf_swevent_cancel_hrtimer(event); | |
6316 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
6317 | } |
6318 | ||
6319 | static int task_clock_event_add(struct perf_event *event, int flags) | |
6320 | { | |
6321 | if (flags & PERF_EF_START) | |
6322 | task_clock_event_start(event, flags); | |
b0a873eb | 6323 | |
a4eaf7f1 PZ |
6324 | return 0; |
6325 | } | |
6326 | ||
6327 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6328 | { | |
6329 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6330 | } |
6331 | ||
6332 | static void task_clock_event_read(struct perf_event *event) | |
6333 | { | |
768a06e2 PZ |
6334 | u64 now = perf_clock(); |
6335 | u64 delta = now - event->ctx->timestamp; | |
6336 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6337 | |
6338 | task_clock_event_update(event, time); | |
6339 | } | |
6340 | ||
6341 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6342 | { |
b0a873eb PZ |
6343 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6344 | return -ENOENT; | |
6345 | ||
6346 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6347 | return -ENOENT; | |
6348 | ||
2481c5fa SE |
6349 | /* |
6350 | * no branch sampling for software events | |
6351 | */ | |
6352 | if (has_branch_stack(event)) | |
6353 | return -EOPNOTSUPP; | |
6354 | ||
ba3dd36c PZ |
6355 | perf_swevent_init_hrtimer(event); |
6356 | ||
b0a873eb | 6357 | return 0; |
6fb2915d LZ |
6358 | } |
6359 | ||
b0a873eb | 6360 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6361 | .task_ctx_nr = perf_sw_context, |
6362 | ||
b0a873eb | 6363 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6364 | .add = task_clock_event_add, |
6365 | .del = task_clock_event_del, | |
6366 | .start = task_clock_event_start, | |
6367 | .stop = task_clock_event_stop, | |
b0a873eb | 6368 | .read = task_clock_event_read, |
35edc2a5 PZ |
6369 | |
6370 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 6371 | }; |
6fb2915d | 6372 | |
ad5133b7 | 6373 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6374 | { |
e077df4f | 6375 | } |
6fb2915d | 6376 | |
ad5133b7 | 6377 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6378 | { |
ad5133b7 | 6379 | return 0; |
6fb2915d LZ |
6380 | } |
6381 | ||
ad5133b7 | 6382 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6383 | { |
ad5133b7 | 6384 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6385 | } |
6386 | ||
ad5133b7 PZ |
6387 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6388 | { | |
6389 | perf_pmu_enable(pmu); | |
6390 | return 0; | |
6391 | } | |
e077df4f | 6392 | |
ad5133b7 | 6393 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6394 | { |
ad5133b7 | 6395 | perf_pmu_enable(pmu); |
24f1e32c FW |
6396 | } |
6397 | ||
35edc2a5 PZ |
6398 | static int perf_event_idx_default(struct perf_event *event) |
6399 | { | |
6400 | return event->hw.idx + 1; | |
6401 | } | |
6402 | ||
8dc85d54 PZ |
6403 | /* |
6404 | * Ensures all contexts with the same task_ctx_nr have the same | |
6405 | * pmu_cpu_context too. | |
6406 | */ | |
9e317041 | 6407 | static struct perf_cpu_context __percpu *find_pmu_context(int ctxn) |
24f1e32c | 6408 | { |
8dc85d54 | 6409 | struct pmu *pmu; |
b326e956 | 6410 | |
8dc85d54 PZ |
6411 | if (ctxn < 0) |
6412 | return NULL; | |
24f1e32c | 6413 | |
8dc85d54 PZ |
6414 | list_for_each_entry(pmu, &pmus, entry) { |
6415 | if (pmu->task_ctx_nr == ctxn) | |
6416 | return pmu->pmu_cpu_context; | |
6417 | } | |
24f1e32c | 6418 | |
8dc85d54 | 6419 | return NULL; |
24f1e32c FW |
6420 | } |
6421 | ||
51676957 | 6422 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6423 | { |
51676957 PZ |
6424 | int cpu; |
6425 | ||
6426 | for_each_possible_cpu(cpu) { | |
6427 | struct perf_cpu_context *cpuctx; | |
6428 | ||
6429 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6430 | ||
3f1f3320 PZ |
6431 | if (cpuctx->unique_pmu == old_pmu) |
6432 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6433 | } |
6434 | } | |
6435 | ||
6436 | static void free_pmu_context(struct pmu *pmu) | |
6437 | { | |
6438 | struct pmu *i; | |
f5ffe02e | 6439 | |
8dc85d54 | 6440 | mutex_lock(&pmus_lock); |
0475f9ea | 6441 | /* |
8dc85d54 | 6442 | * Like a real lame refcount. |
0475f9ea | 6443 | */ |
51676957 PZ |
6444 | list_for_each_entry(i, &pmus, entry) { |
6445 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6446 | update_pmu_context(i, pmu); | |
8dc85d54 | 6447 | goto out; |
51676957 | 6448 | } |
8dc85d54 | 6449 | } |
d6d020e9 | 6450 | |
51676957 | 6451 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6452 | out: |
6453 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6454 | } |
2e80a82a | 6455 | static struct idr pmu_idr; |
d6d020e9 | 6456 | |
abe43400 PZ |
6457 | static ssize_t |
6458 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6459 | { | |
6460 | struct pmu *pmu = dev_get_drvdata(dev); | |
6461 | ||
6462 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6463 | } | |
90826ca7 | 6464 | static DEVICE_ATTR_RO(type); |
abe43400 | 6465 | |
62b85639 SE |
6466 | static ssize_t |
6467 | perf_event_mux_interval_ms_show(struct device *dev, | |
6468 | struct device_attribute *attr, | |
6469 | char *page) | |
6470 | { | |
6471 | struct pmu *pmu = dev_get_drvdata(dev); | |
6472 | ||
6473 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6474 | } | |
6475 | ||
6476 | static ssize_t | |
6477 | perf_event_mux_interval_ms_store(struct device *dev, | |
6478 | struct device_attribute *attr, | |
6479 | const char *buf, size_t count) | |
6480 | { | |
6481 | struct pmu *pmu = dev_get_drvdata(dev); | |
6482 | int timer, cpu, ret; | |
6483 | ||
6484 | ret = kstrtoint(buf, 0, &timer); | |
6485 | if (ret) | |
6486 | return ret; | |
6487 | ||
6488 | if (timer < 1) | |
6489 | return -EINVAL; | |
6490 | ||
6491 | /* same value, noting to do */ | |
6492 | if (timer == pmu->hrtimer_interval_ms) | |
6493 | return count; | |
6494 | ||
6495 | pmu->hrtimer_interval_ms = timer; | |
6496 | ||
6497 | /* update all cpuctx for this PMU */ | |
6498 | for_each_possible_cpu(cpu) { | |
6499 | struct perf_cpu_context *cpuctx; | |
6500 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6501 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6502 | ||
6503 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6504 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6505 | } | |
6506 | ||
6507 | return count; | |
6508 | } | |
90826ca7 | 6509 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
62b85639 | 6510 | |
90826ca7 GKH |
6511 | static struct attribute *pmu_dev_attrs[] = { |
6512 | &dev_attr_type.attr, | |
6513 | &dev_attr_perf_event_mux_interval_ms.attr, | |
6514 | NULL, | |
abe43400 | 6515 | }; |
90826ca7 | 6516 | ATTRIBUTE_GROUPS(pmu_dev); |
abe43400 PZ |
6517 | |
6518 | static int pmu_bus_running; | |
6519 | static struct bus_type pmu_bus = { | |
6520 | .name = "event_source", | |
90826ca7 | 6521 | .dev_groups = pmu_dev_groups, |
abe43400 PZ |
6522 | }; |
6523 | ||
6524 | static void pmu_dev_release(struct device *dev) | |
6525 | { | |
6526 | kfree(dev); | |
6527 | } | |
6528 | ||
6529 | static int pmu_dev_alloc(struct pmu *pmu) | |
6530 | { | |
6531 | int ret = -ENOMEM; | |
6532 | ||
6533 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6534 | if (!pmu->dev) | |
6535 | goto out; | |
6536 | ||
0c9d42ed | 6537 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6538 | device_initialize(pmu->dev); |
6539 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6540 | if (ret) | |
6541 | goto free_dev; | |
6542 | ||
6543 | dev_set_drvdata(pmu->dev, pmu); | |
6544 | pmu->dev->bus = &pmu_bus; | |
6545 | pmu->dev->release = pmu_dev_release; | |
6546 | ret = device_add(pmu->dev); | |
6547 | if (ret) | |
6548 | goto free_dev; | |
6549 | ||
6550 | out: | |
6551 | return ret; | |
6552 | ||
6553 | free_dev: | |
6554 | put_device(pmu->dev); | |
6555 | goto out; | |
6556 | } | |
6557 | ||
547e9fd7 | 6558 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6559 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6560 | |
03d8e80b | 6561 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6562 | { |
108b02cf | 6563 | int cpu, ret; |
24f1e32c | 6564 | |
b0a873eb | 6565 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6566 | ret = -ENOMEM; |
6567 | pmu->pmu_disable_count = alloc_percpu(int); | |
6568 | if (!pmu->pmu_disable_count) | |
6569 | goto unlock; | |
f29ac756 | 6570 | |
2e80a82a PZ |
6571 | pmu->type = -1; |
6572 | if (!name) | |
6573 | goto skip_type; | |
6574 | pmu->name = name; | |
6575 | ||
6576 | if (type < 0) { | |
0e9c3be2 TH |
6577 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6578 | if (type < 0) { | |
6579 | ret = type; | |
2e80a82a PZ |
6580 | goto free_pdc; |
6581 | } | |
6582 | } | |
6583 | pmu->type = type; | |
6584 | ||
abe43400 PZ |
6585 | if (pmu_bus_running) { |
6586 | ret = pmu_dev_alloc(pmu); | |
6587 | if (ret) | |
6588 | goto free_idr; | |
6589 | } | |
6590 | ||
2e80a82a | 6591 | skip_type: |
8dc85d54 PZ |
6592 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6593 | if (pmu->pmu_cpu_context) | |
6594 | goto got_cpu_context; | |
f29ac756 | 6595 | |
c4814202 | 6596 | ret = -ENOMEM; |
108b02cf PZ |
6597 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6598 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6599 | goto free_dev; |
f344011c | 6600 | |
108b02cf PZ |
6601 | for_each_possible_cpu(cpu) { |
6602 | struct perf_cpu_context *cpuctx; | |
6603 | ||
6604 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6605 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6606 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6607 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6608 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6609 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6610 | |
6611 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6612 | ||
e9d2b064 | 6613 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
3f1f3320 | 6614 | cpuctx->unique_pmu = pmu; |
108b02cf | 6615 | } |
76e1d904 | 6616 | |
8dc85d54 | 6617 | got_cpu_context: |
ad5133b7 PZ |
6618 | if (!pmu->start_txn) { |
6619 | if (pmu->pmu_enable) { | |
6620 | /* | |
6621 | * If we have pmu_enable/pmu_disable calls, install | |
6622 | * transaction stubs that use that to try and batch | |
6623 | * hardware accesses. | |
6624 | */ | |
6625 | pmu->start_txn = perf_pmu_start_txn; | |
6626 | pmu->commit_txn = perf_pmu_commit_txn; | |
6627 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6628 | } else { | |
6629 | pmu->start_txn = perf_pmu_nop_void; | |
6630 | pmu->commit_txn = perf_pmu_nop_int; | |
6631 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6632 | } |
5c92d124 | 6633 | } |
15dbf27c | 6634 | |
ad5133b7 PZ |
6635 | if (!pmu->pmu_enable) { |
6636 | pmu->pmu_enable = perf_pmu_nop_void; | |
6637 | pmu->pmu_disable = perf_pmu_nop_void; | |
6638 | } | |
6639 | ||
35edc2a5 PZ |
6640 | if (!pmu->event_idx) |
6641 | pmu->event_idx = perf_event_idx_default; | |
6642 | ||
b0a873eb | 6643 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6644 | ret = 0; |
6645 | unlock: | |
b0a873eb PZ |
6646 | mutex_unlock(&pmus_lock); |
6647 | ||
33696fc0 | 6648 | return ret; |
108b02cf | 6649 | |
abe43400 PZ |
6650 | free_dev: |
6651 | device_del(pmu->dev); | |
6652 | put_device(pmu->dev); | |
6653 | ||
2e80a82a PZ |
6654 | free_idr: |
6655 | if (pmu->type >= PERF_TYPE_MAX) | |
6656 | idr_remove(&pmu_idr, pmu->type); | |
6657 | ||
108b02cf PZ |
6658 | free_pdc: |
6659 | free_percpu(pmu->pmu_disable_count); | |
6660 | goto unlock; | |
f29ac756 | 6661 | } |
c464c76e | 6662 | EXPORT_SYMBOL_GPL(perf_pmu_register); |
f29ac756 | 6663 | |
b0a873eb | 6664 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6665 | { |
b0a873eb PZ |
6666 | mutex_lock(&pmus_lock); |
6667 | list_del_rcu(&pmu->entry); | |
6668 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6669 | |
0475f9ea | 6670 | /* |
cde8e884 PZ |
6671 | * We dereference the pmu list under both SRCU and regular RCU, so |
6672 | * synchronize against both of those. | |
0475f9ea | 6673 | */ |
b0a873eb | 6674 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6675 | synchronize_rcu(); |
d6d020e9 | 6676 | |
33696fc0 | 6677 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6678 | if (pmu->type >= PERF_TYPE_MAX) |
6679 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6680 | device_del(pmu->dev); |
6681 | put_device(pmu->dev); | |
51676957 | 6682 | free_pmu_context(pmu); |
b0a873eb | 6683 | } |
c464c76e | 6684 | EXPORT_SYMBOL_GPL(perf_pmu_unregister); |
d6d020e9 | 6685 | |
b0a873eb PZ |
6686 | struct pmu *perf_init_event(struct perf_event *event) |
6687 | { | |
6688 | struct pmu *pmu = NULL; | |
6689 | int idx; | |
940c5b29 | 6690 | int ret; |
b0a873eb PZ |
6691 | |
6692 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6693 | |
6694 | rcu_read_lock(); | |
6695 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6696 | rcu_read_unlock(); | |
940c5b29 | 6697 | if (pmu) { |
c464c76e YZ |
6698 | if (!try_module_get(pmu->module)) { |
6699 | pmu = ERR_PTR(-ENODEV); | |
6700 | goto unlock; | |
6701 | } | |
7e5b2a01 | 6702 | event->pmu = pmu; |
940c5b29 LM |
6703 | ret = pmu->event_init(event); |
6704 | if (ret) | |
6705 | pmu = ERR_PTR(ret); | |
2e80a82a | 6706 | goto unlock; |
940c5b29 | 6707 | } |
2e80a82a | 6708 | |
b0a873eb | 6709 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
c464c76e YZ |
6710 | if (!try_module_get(pmu->module)) { |
6711 | pmu = ERR_PTR(-ENODEV); | |
6712 | goto unlock; | |
6713 | } | |
7e5b2a01 | 6714 | event->pmu = pmu; |
940c5b29 | 6715 | ret = pmu->event_init(event); |
b0a873eb | 6716 | if (!ret) |
e5f4d339 | 6717 | goto unlock; |
76e1d904 | 6718 | |
b0a873eb PZ |
6719 | if (ret != -ENOENT) { |
6720 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6721 | goto unlock; |
f344011c | 6722 | } |
5c92d124 | 6723 | } |
e5f4d339 PZ |
6724 | pmu = ERR_PTR(-ENOENT); |
6725 | unlock: | |
b0a873eb | 6726 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6727 | |
4aeb0b42 | 6728 | return pmu; |
5c92d124 IM |
6729 | } |
6730 | ||
4beb31f3 FW |
6731 | static void account_event_cpu(struct perf_event *event, int cpu) |
6732 | { | |
6733 | if (event->parent) | |
6734 | return; | |
6735 | ||
6736 | if (has_branch_stack(event)) { | |
6737 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6738 | atomic_inc(&per_cpu(perf_branch_stack_events, cpu)); | |
6739 | } | |
6740 | if (is_cgroup_event(event)) | |
6741 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
6742 | } | |
6743 | ||
766d6c07 FW |
6744 | static void account_event(struct perf_event *event) |
6745 | { | |
4beb31f3 FW |
6746 | if (event->parent) |
6747 | return; | |
6748 | ||
766d6c07 FW |
6749 | if (event->attach_state & PERF_ATTACH_TASK) |
6750 | static_key_slow_inc(&perf_sched_events.key); | |
6751 | if (event->attr.mmap || event->attr.mmap_data) | |
6752 | atomic_inc(&nr_mmap_events); | |
6753 | if (event->attr.comm) | |
6754 | atomic_inc(&nr_comm_events); | |
6755 | if (event->attr.task) | |
6756 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
6757 | if (event->attr.freq) { |
6758 | if (atomic_inc_return(&nr_freq_events) == 1) | |
6759 | tick_nohz_full_kick_all(); | |
6760 | } | |
4beb31f3 | 6761 | if (has_branch_stack(event)) |
766d6c07 | 6762 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 6763 | if (is_cgroup_event(event)) |
766d6c07 | 6764 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
6765 | |
6766 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
6767 | } |
6768 | ||
0793a61d | 6769 | /* |
cdd6c482 | 6770 | * Allocate and initialize a event structure |
0793a61d | 6771 | */ |
cdd6c482 | 6772 | static struct perf_event * |
c3f00c70 | 6773 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6774 | struct task_struct *task, |
6775 | struct perf_event *group_leader, | |
6776 | struct perf_event *parent_event, | |
4dc0da86 AK |
6777 | perf_overflow_handler_t overflow_handler, |
6778 | void *context) | |
0793a61d | 6779 | { |
51b0fe39 | 6780 | struct pmu *pmu; |
cdd6c482 IM |
6781 | struct perf_event *event; |
6782 | struct hw_perf_event *hwc; | |
90983b16 | 6783 | long err = -EINVAL; |
0793a61d | 6784 | |
66832eb4 ON |
6785 | if ((unsigned)cpu >= nr_cpu_ids) { |
6786 | if (!task || cpu != -1) | |
6787 | return ERR_PTR(-EINVAL); | |
6788 | } | |
6789 | ||
c3f00c70 | 6790 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6791 | if (!event) |
d5d2bc0d | 6792 | return ERR_PTR(-ENOMEM); |
0793a61d | 6793 | |
04289bb9 | 6794 | /* |
cdd6c482 | 6795 | * Single events are their own group leaders, with an |
04289bb9 IM |
6796 | * empty sibling list: |
6797 | */ | |
6798 | if (!group_leader) | |
cdd6c482 | 6799 | group_leader = event; |
04289bb9 | 6800 | |
cdd6c482 IM |
6801 | mutex_init(&event->child_mutex); |
6802 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6803 | |
cdd6c482 IM |
6804 | INIT_LIST_HEAD(&event->group_entry); |
6805 | INIT_LIST_HEAD(&event->event_entry); | |
6806 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 | 6807 | INIT_LIST_HEAD(&event->rb_entry); |
71ad88ef | 6808 | INIT_LIST_HEAD(&event->active_entry); |
f3ae75de SE |
6809 | INIT_HLIST_NODE(&event->hlist_entry); |
6810 | ||
10c6db11 | 6811 | |
cdd6c482 | 6812 | init_waitqueue_head(&event->waitq); |
e360adbe | 6813 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6814 | |
cdd6c482 | 6815 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6816 | |
a6fa941d | 6817 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6818 | event->cpu = cpu; |
6819 | event->attr = *attr; | |
6820 | event->group_leader = group_leader; | |
6821 | event->pmu = NULL; | |
cdd6c482 | 6822 | event->oncpu = -1; |
a96bbc16 | 6823 | |
cdd6c482 | 6824 | event->parent = parent_event; |
b84fbc9f | 6825 | |
17cf22c3 | 6826 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6827 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6828 | |
cdd6c482 | 6829 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6830 | |
d580ff86 PZ |
6831 | if (task) { |
6832 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6833 | |
6834 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6835 | event->hw.tp_target = task; | |
d580ff86 PZ |
6836 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6837 | /* | |
6838 | * hw_breakpoint is a bit difficult here.. | |
6839 | */ | |
f22c1bb6 | 6840 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6841 | event->hw.bp_target = task; |
6842 | #endif | |
6843 | } | |
6844 | ||
4dc0da86 | 6845 | if (!overflow_handler && parent_event) { |
b326e956 | 6846 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6847 | context = parent_event->overflow_handler_context; |
6848 | } | |
66832eb4 | 6849 | |
b326e956 | 6850 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6851 | event->overflow_handler_context = context; |
97eaf530 | 6852 | |
0231bb53 | 6853 | perf_event__state_init(event); |
a86ed508 | 6854 | |
4aeb0b42 | 6855 | pmu = NULL; |
b8e83514 | 6856 | |
cdd6c482 | 6857 | hwc = &event->hw; |
bd2b5b12 | 6858 | hwc->sample_period = attr->sample_period; |
0d48696f | 6859 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6860 | hwc->sample_period = 1; |
eced1dfc | 6861 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6862 | |
e7850595 | 6863 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6864 | |
2023b359 | 6865 | /* |
cdd6c482 | 6866 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6867 | */ |
3dab77fb | 6868 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 6869 | goto err_ns; |
2023b359 | 6870 | |
b0a873eb | 6871 | pmu = perf_init_event(event); |
4aeb0b42 | 6872 | if (!pmu) |
90983b16 FW |
6873 | goto err_ns; |
6874 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 6875 | err = PTR_ERR(pmu); |
90983b16 | 6876 | goto err_ns; |
621a01ea | 6877 | } |
d5d2bc0d | 6878 | |
cdd6c482 | 6879 | if (!event->parent) { |
927c7a9e FW |
6880 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6881 | err = get_callchain_buffers(); | |
90983b16 FW |
6882 | if (err) |
6883 | goto err_pmu; | |
d010b332 | 6884 | } |
f344011c | 6885 | } |
9ee318a7 | 6886 | |
cdd6c482 | 6887 | return event; |
90983b16 FW |
6888 | |
6889 | err_pmu: | |
6890 | if (event->destroy) | |
6891 | event->destroy(event); | |
c464c76e | 6892 | module_put(pmu->module); |
90983b16 FW |
6893 | err_ns: |
6894 | if (event->ns) | |
6895 | put_pid_ns(event->ns); | |
6896 | kfree(event); | |
6897 | ||
6898 | return ERR_PTR(err); | |
0793a61d TG |
6899 | } |
6900 | ||
cdd6c482 IM |
6901 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6902 | struct perf_event_attr *attr) | |
974802ea | 6903 | { |
974802ea | 6904 | u32 size; |
cdf8073d | 6905 | int ret; |
974802ea PZ |
6906 | |
6907 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6908 | return -EFAULT; | |
6909 | ||
6910 | /* | |
6911 | * zero the full structure, so that a short copy will be nice. | |
6912 | */ | |
6913 | memset(attr, 0, sizeof(*attr)); | |
6914 | ||
6915 | ret = get_user(size, &uattr->size); | |
6916 | if (ret) | |
6917 | return ret; | |
6918 | ||
6919 | if (size > PAGE_SIZE) /* silly large */ | |
6920 | goto err_size; | |
6921 | ||
6922 | if (!size) /* abi compat */ | |
6923 | size = PERF_ATTR_SIZE_VER0; | |
6924 | ||
6925 | if (size < PERF_ATTR_SIZE_VER0) | |
6926 | goto err_size; | |
6927 | ||
6928 | /* | |
6929 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6930 | * ensure all the unknown bits are 0 - i.e. new |
6931 | * user-space does not rely on any kernel feature | |
6932 | * extensions we dont know about yet. | |
974802ea PZ |
6933 | */ |
6934 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6935 | unsigned char __user *addr; |
6936 | unsigned char __user *end; | |
6937 | unsigned char val; | |
974802ea | 6938 | |
cdf8073d IS |
6939 | addr = (void __user *)uattr + sizeof(*attr); |
6940 | end = (void __user *)uattr + size; | |
974802ea | 6941 | |
cdf8073d | 6942 | for (; addr < end; addr++) { |
974802ea PZ |
6943 | ret = get_user(val, addr); |
6944 | if (ret) | |
6945 | return ret; | |
6946 | if (val) | |
6947 | goto err_size; | |
6948 | } | |
b3e62e35 | 6949 | size = sizeof(*attr); |
974802ea PZ |
6950 | } |
6951 | ||
6952 | ret = copy_from_user(attr, uattr, size); | |
6953 | if (ret) | |
6954 | return -EFAULT; | |
6955 | ||
cd757645 | 6956 | if (attr->__reserved_1) |
974802ea PZ |
6957 | return -EINVAL; |
6958 | ||
6959 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6960 | return -EINVAL; | |
6961 | ||
6962 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6963 | return -EINVAL; | |
6964 | ||
bce38cd5 SE |
6965 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6966 | u64 mask = attr->branch_sample_type; | |
6967 | ||
6968 | /* only using defined bits */ | |
6969 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6970 | return -EINVAL; | |
6971 | ||
6972 | /* at least one branch bit must be set */ | |
6973 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6974 | return -EINVAL; | |
6975 | ||
bce38cd5 SE |
6976 | /* propagate priv level, when not set for branch */ |
6977 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6978 | ||
6979 | /* exclude_kernel checked on syscall entry */ | |
6980 | if (!attr->exclude_kernel) | |
6981 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6982 | ||
6983 | if (!attr->exclude_user) | |
6984 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6985 | ||
6986 | if (!attr->exclude_hv) | |
6987 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6988 | /* | |
6989 | * adjust user setting (for HW filter setup) | |
6990 | */ | |
6991 | attr->branch_sample_type = mask; | |
6992 | } | |
e712209a SE |
6993 | /* privileged levels capture (kernel, hv): check permissions */ |
6994 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
6995 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
6996 | return -EACCES; | |
bce38cd5 | 6997 | } |
4018994f | 6998 | |
c5ebcedb | 6999 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 7000 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
7001 | if (ret) |
7002 | return ret; | |
7003 | } | |
7004 | ||
7005 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
7006 | if (!arch_perf_have_user_stack_dump()) | |
7007 | return -ENOSYS; | |
7008 | ||
7009 | /* | |
7010 | * We have __u32 type for the size, but so far | |
7011 | * we can only use __u16 as maximum due to the | |
7012 | * __u16 sample size limit. | |
7013 | */ | |
7014 | if (attr->sample_stack_user >= USHRT_MAX) | |
7015 | ret = -EINVAL; | |
7016 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
7017 | ret = -EINVAL; | |
7018 | } | |
4018994f | 7019 | |
974802ea PZ |
7020 | out: |
7021 | return ret; | |
7022 | ||
7023 | err_size: | |
7024 | put_user(sizeof(*attr), &uattr->size); | |
7025 | ret = -E2BIG; | |
7026 | goto out; | |
7027 | } | |
7028 | ||
ac9721f3 PZ |
7029 | static int |
7030 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 7031 | { |
b69cf536 | 7032 | struct ring_buffer *rb = NULL; |
a4be7c27 PZ |
7033 | int ret = -EINVAL; |
7034 | ||
ac9721f3 | 7035 | if (!output_event) |
a4be7c27 PZ |
7036 | goto set; |
7037 | ||
ac9721f3 PZ |
7038 | /* don't allow circular references */ |
7039 | if (event == output_event) | |
a4be7c27 PZ |
7040 | goto out; |
7041 | ||
0f139300 PZ |
7042 | /* |
7043 | * Don't allow cross-cpu buffers | |
7044 | */ | |
7045 | if (output_event->cpu != event->cpu) | |
7046 | goto out; | |
7047 | ||
7048 | /* | |
76369139 | 7049 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
7050 | */ |
7051 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
7052 | goto out; | |
7053 | ||
a4be7c27 | 7054 | set: |
cdd6c482 | 7055 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
7056 | /* Can't redirect output if we've got an active mmap() */ |
7057 | if (atomic_read(&event->mmap_count)) | |
7058 | goto unlock; | |
a4be7c27 | 7059 | |
ac9721f3 | 7060 | if (output_event) { |
76369139 FW |
7061 | /* get the rb we want to redirect to */ |
7062 | rb = ring_buffer_get(output_event); | |
7063 | if (!rb) | |
ac9721f3 | 7064 | goto unlock; |
a4be7c27 PZ |
7065 | } |
7066 | ||
b69cf536 | 7067 | ring_buffer_attach(event, rb); |
9bb5d40c | 7068 | |
a4be7c27 | 7069 | ret = 0; |
ac9721f3 PZ |
7070 | unlock: |
7071 | mutex_unlock(&event->mmap_mutex); | |
7072 | ||
a4be7c27 | 7073 | out: |
a4be7c27 PZ |
7074 | return ret; |
7075 | } | |
7076 | ||
0793a61d | 7077 | /** |
cdd6c482 | 7078 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 7079 | * |
cdd6c482 | 7080 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 7081 | * @pid: target pid |
9f66a381 | 7082 | * @cpu: target cpu |
cdd6c482 | 7083 | * @group_fd: group leader event fd |
0793a61d | 7084 | */ |
cdd6c482 IM |
7085 | SYSCALL_DEFINE5(perf_event_open, |
7086 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 7087 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 7088 | { |
b04243ef PZ |
7089 | struct perf_event *group_leader = NULL, *output_event = NULL; |
7090 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
7091 | struct perf_event_attr attr; |
7092 | struct perf_event_context *ctx; | |
7093 | struct file *event_file = NULL; | |
2903ff01 | 7094 | struct fd group = {NULL, 0}; |
38a81da2 | 7095 | struct task_struct *task = NULL; |
89a1e187 | 7096 | struct pmu *pmu; |
ea635c64 | 7097 | int event_fd; |
b04243ef | 7098 | int move_group = 0; |
dc86cabe | 7099 | int err; |
a21b0b35 | 7100 | int f_flags = O_RDWR; |
0793a61d | 7101 | |
2743a5b0 | 7102 | /* for future expandability... */ |
e5d1367f | 7103 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
7104 | return -EINVAL; |
7105 | ||
dc86cabe IM |
7106 | err = perf_copy_attr(attr_uptr, &attr); |
7107 | if (err) | |
7108 | return err; | |
eab656ae | 7109 | |
0764771d PZ |
7110 | if (!attr.exclude_kernel) { |
7111 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
7112 | return -EACCES; | |
7113 | } | |
7114 | ||
df58ab24 | 7115 | if (attr.freq) { |
cdd6c482 | 7116 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 | 7117 | return -EINVAL; |
0819b2e3 PZ |
7118 | } else { |
7119 | if (attr.sample_period & (1ULL << 63)) | |
7120 | return -EINVAL; | |
df58ab24 PZ |
7121 | } |
7122 | ||
e5d1367f SE |
7123 | /* |
7124 | * In cgroup mode, the pid argument is used to pass the fd | |
7125 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
7126 | * designates the cpu on which to monitor threads from that | |
7127 | * cgroup. | |
7128 | */ | |
7129 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
7130 | return -EINVAL; | |
7131 | ||
a21b0b35 YD |
7132 | if (flags & PERF_FLAG_FD_CLOEXEC) |
7133 | f_flags |= O_CLOEXEC; | |
7134 | ||
7135 | event_fd = get_unused_fd_flags(f_flags); | |
ea635c64 AV |
7136 | if (event_fd < 0) |
7137 | return event_fd; | |
7138 | ||
ac9721f3 | 7139 | if (group_fd != -1) { |
2903ff01 AV |
7140 | err = perf_fget_light(group_fd, &group); |
7141 | if (err) | |
d14b12d7 | 7142 | goto err_fd; |
2903ff01 | 7143 | group_leader = group.file->private_data; |
ac9721f3 PZ |
7144 | if (flags & PERF_FLAG_FD_OUTPUT) |
7145 | output_event = group_leader; | |
7146 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
7147 | group_leader = NULL; | |
7148 | } | |
7149 | ||
e5d1367f | 7150 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
7151 | task = find_lively_task_by_vpid(pid); |
7152 | if (IS_ERR(task)) { | |
7153 | err = PTR_ERR(task); | |
7154 | goto err_group_fd; | |
7155 | } | |
7156 | } | |
7157 | ||
1f4ee503 PZ |
7158 | if (task && group_leader && |
7159 | group_leader->attr.inherit != attr.inherit) { | |
7160 | err = -EINVAL; | |
7161 | goto err_task; | |
7162 | } | |
7163 | ||
fbfc623f YZ |
7164 | get_online_cpus(); |
7165 | ||
4dc0da86 AK |
7166 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
7167 | NULL, NULL); | |
d14b12d7 SE |
7168 | if (IS_ERR(event)) { |
7169 | err = PTR_ERR(event); | |
1f4ee503 | 7170 | goto err_cpus; |
d14b12d7 SE |
7171 | } |
7172 | ||
e5d1367f SE |
7173 | if (flags & PERF_FLAG_PID_CGROUP) { |
7174 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
766d6c07 FW |
7175 | if (err) { |
7176 | __free_event(event); | |
1f4ee503 | 7177 | goto err_cpus; |
766d6c07 | 7178 | } |
e5d1367f SE |
7179 | } |
7180 | ||
53b25335 VW |
7181 | if (is_sampling_event(event)) { |
7182 | if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { | |
7183 | err = -ENOTSUPP; | |
7184 | goto err_alloc; | |
7185 | } | |
7186 | } | |
7187 | ||
766d6c07 FW |
7188 | account_event(event); |
7189 | ||
89a1e187 PZ |
7190 | /* |
7191 | * Special case software events and allow them to be part of | |
7192 | * any hardware group. | |
7193 | */ | |
7194 | pmu = event->pmu; | |
b04243ef PZ |
7195 | |
7196 | if (group_leader && | |
7197 | (is_software_event(event) != is_software_event(group_leader))) { | |
7198 | if (is_software_event(event)) { | |
7199 | /* | |
7200 | * If event and group_leader are not both a software | |
7201 | * event, and event is, then group leader is not. | |
7202 | * | |
7203 | * Allow the addition of software events to !software | |
7204 | * groups, this is safe because software events never | |
7205 | * fail to schedule. | |
7206 | */ | |
7207 | pmu = group_leader->pmu; | |
7208 | } else if (is_software_event(group_leader) && | |
7209 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
7210 | /* | |
7211 | * In case the group is a pure software group, and we | |
7212 | * try to add a hardware event, move the whole group to | |
7213 | * the hardware context. | |
7214 | */ | |
7215 | move_group = 1; | |
7216 | } | |
7217 | } | |
89a1e187 PZ |
7218 | |
7219 | /* | |
7220 | * Get the target context (task or percpu): | |
7221 | */ | |
e2d37cd2 | 7222 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
7223 | if (IS_ERR(ctx)) { |
7224 | err = PTR_ERR(ctx); | |
c6be5a5c | 7225 | goto err_alloc; |
89a1e187 PZ |
7226 | } |
7227 | ||
fd1edb3a PZ |
7228 | if (task) { |
7229 | put_task_struct(task); | |
7230 | task = NULL; | |
7231 | } | |
7232 | ||
ccff286d | 7233 | /* |
cdd6c482 | 7234 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 7235 | */ |
ac9721f3 | 7236 | if (group_leader) { |
dc86cabe | 7237 | err = -EINVAL; |
04289bb9 | 7238 | |
04289bb9 | 7239 | /* |
ccff286d IM |
7240 | * Do not allow a recursive hierarchy (this new sibling |
7241 | * becoming part of another group-sibling): | |
7242 | */ | |
7243 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 7244 | goto err_context; |
ccff286d IM |
7245 | /* |
7246 | * Do not allow to attach to a group in a different | |
7247 | * task or CPU context: | |
04289bb9 | 7248 | */ |
b04243ef PZ |
7249 | if (move_group) { |
7250 | if (group_leader->ctx->type != ctx->type) | |
7251 | goto err_context; | |
7252 | } else { | |
7253 | if (group_leader->ctx != ctx) | |
7254 | goto err_context; | |
7255 | } | |
7256 | ||
3b6f9e5c PM |
7257 | /* |
7258 | * Only a group leader can be exclusive or pinned | |
7259 | */ | |
0d48696f | 7260 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 7261 | goto err_context; |
ac9721f3 PZ |
7262 | } |
7263 | ||
7264 | if (output_event) { | |
7265 | err = perf_event_set_output(event, output_event); | |
7266 | if (err) | |
c3f00c70 | 7267 | goto err_context; |
ac9721f3 | 7268 | } |
0793a61d | 7269 | |
a21b0b35 YD |
7270 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
7271 | f_flags); | |
ea635c64 AV |
7272 | if (IS_ERR(event_file)) { |
7273 | err = PTR_ERR(event_file); | |
c3f00c70 | 7274 | goto err_context; |
ea635c64 | 7275 | } |
9b51f66d | 7276 | |
b04243ef PZ |
7277 | if (move_group) { |
7278 | struct perf_event_context *gctx = group_leader->ctx; | |
7279 | ||
7280 | mutex_lock(&gctx->mutex); | |
46ce0fe9 | 7281 | perf_remove_from_context(group_leader, false); |
0231bb53 JO |
7282 | |
7283 | /* | |
7284 | * Removing from the context ends up with disabled | |
7285 | * event. What we want here is event in the initial | |
7286 | * startup state, ready to be add into new context. | |
7287 | */ | |
7288 | perf_event__state_init(group_leader); | |
b04243ef PZ |
7289 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7290 | group_entry) { | |
46ce0fe9 | 7291 | perf_remove_from_context(sibling, false); |
0231bb53 | 7292 | perf_event__state_init(sibling); |
b04243ef PZ |
7293 | put_ctx(gctx); |
7294 | } | |
7295 | mutex_unlock(&gctx->mutex); | |
7296 | put_ctx(gctx); | |
ea635c64 | 7297 | } |
9b51f66d | 7298 | |
ad3a37de | 7299 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 7300 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
7301 | |
7302 | if (move_group) { | |
0cda4c02 | 7303 | synchronize_rcu(); |
e2d37cd2 | 7304 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
7305 | get_ctx(ctx); |
7306 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
7307 | group_entry) { | |
e2d37cd2 | 7308 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
7309 | get_ctx(ctx); |
7310 | } | |
7311 | } | |
7312 | ||
e2d37cd2 | 7313 | perf_install_in_context(ctx, event, event->cpu); |
fe4b04fa | 7314 | perf_unpin_context(ctx); |
d859e29f | 7315 | mutex_unlock(&ctx->mutex); |
9b51f66d | 7316 | |
fbfc623f YZ |
7317 | put_online_cpus(); |
7318 | ||
cdd6c482 | 7319 | event->owner = current; |
8882135b | 7320 | |
cdd6c482 IM |
7321 | mutex_lock(¤t->perf_event_mutex); |
7322 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
7323 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 7324 | |
c320c7b7 ACM |
7325 | /* |
7326 | * Precalculate sample_data sizes | |
7327 | */ | |
7328 | perf_event__header_size(event); | |
6844c09d | 7329 | perf_event__id_header_size(event); |
c320c7b7 | 7330 | |
8a49542c PZ |
7331 | /* |
7332 | * Drop the reference on the group_event after placing the | |
7333 | * new event on the sibling_list. This ensures destruction | |
7334 | * of the group leader will find the pointer to itself in | |
7335 | * perf_group_detach(). | |
7336 | */ | |
2903ff01 | 7337 | fdput(group); |
ea635c64 AV |
7338 | fd_install(event_fd, event_file); |
7339 | return event_fd; | |
0793a61d | 7340 | |
c3f00c70 | 7341 | err_context: |
fe4b04fa | 7342 | perf_unpin_context(ctx); |
ea635c64 | 7343 | put_ctx(ctx); |
c6be5a5c | 7344 | err_alloc: |
ea635c64 | 7345 | free_event(event); |
1f4ee503 | 7346 | err_cpus: |
fbfc623f | 7347 | put_online_cpus(); |
1f4ee503 | 7348 | err_task: |
e7d0bc04 PZ |
7349 | if (task) |
7350 | put_task_struct(task); | |
89a1e187 | 7351 | err_group_fd: |
2903ff01 | 7352 | fdput(group); |
ea635c64 AV |
7353 | err_fd: |
7354 | put_unused_fd(event_fd); | |
dc86cabe | 7355 | return err; |
0793a61d TG |
7356 | } |
7357 | ||
fb0459d7 AV |
7358 | /** |
7359 | * perf_event_create_kernel_counter | |
7360 | * | |
7361 | * @attr: attributes of the counter to create | |
7362 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 7363 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
7364 | */ |
7365 | struct perf_event * | |
7366 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 7367 | struct task_struct *task, |
4dc0da86 AK |
7368 | perf_overflow_handler_t overflow_handler, |
7369 | void *context) | |
fb0459d7 | 7370 | { |
fb0459d7 | 7371 | struct perf_event_context *ctx; |
c3f00c70 | 7372 | struct perf_event *event; |
fb0459d7 | 7373 | int err; |
d859e29f | 7374 | |
fb0459d7 AV |
7375 | /* |
7376 | * Get the target context (task or percpu): | |
7377 | */ | |
d859e29f | 7378 | |
4dc0da86 AK |
7379 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
7380 | overflow_handler, context); | |
c3f00c70 PZ |
7381 | if (IS_ERR(event)) { |
7382 | err = PTR_ERR(event); | |
7383 | goto err; | |
7384 | } | |
d859e29f | 7385 | |
f8697762 JO |
7386 | /* Mark owner so we could distinguish it from user events. */ |
7387 | event->owner = EVENT_OWNER_KERNEL; | |
7388 | ||
766d6c07 FW |
7389 | account_event(event); |
7390 | ||
38a81da2 | 7391 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
7392 | if (IS_ERR(ctx)) { |
7393 | err = PTR_ERR(ctx); | |
c3f00c70 | 7394 | goto err_free; |
d859e29f | 7395 | } |
fb0459d7 | 7396 | |
fb0459d7 AV |
7397 | WARN_ON_ONCE(ctx->parent_ctx); |
7398 | mutex_lock(&ctx->mutex); | |
7399 | perf_install_in_context(ctx, event, cpu); | |
fe4b04fa | 7400 | perf_unpin_context(ctx); |
fb0459d7 AV |
7401 | mutex_unlock(&ctx->mutex); |
7402 | ||
fb0459d7 AV |
7403 | return event; |
7404 | ||
c3f00c70 PZ |
7405 | err_free: |
7406 | free_event(event); | |
7407 | err: | |
c6567f64 | 7408 | return ERR_PTR(err); |
9b51f66d | 7409 | } |
fb0459d7 | 7410 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7411 | |
0cda4c02 YZ |
7412 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7413 | { | |
7414 | struct perf_event_context *src_ctx; | |
7415 | struct perf_event_context *dst_ctx; | |
7416 | struct perf_event *event, *tmp; | |
7417 | LIST_HEAD(events); | |
7418 | ||
7419 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7420 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7421 | ||
7422 | mutex_lock(&src_ctx->mutex); | |
7423 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
7424 | event_entry) { | |
46ce0fe9 | 7425 | perf_remove_from_context(event, false); |
9a545de0 | 7426 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 7427 | put_ctx(src_ctx); |
9886167d | 7428 | list_add(&event->migrate_entry, &events); |
0cda4c02 YZ |
7429 | } |
7430 | mutex_unlock(&src_ctx->mutex); | |
7431 | ||
7432 | synchronize_rcu(); | |
7433 | ||
7434 | mutex_lock(&dst_ctx->mutex); | |
9886167d PZ |
7435 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
7436 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
7437 | if (event->state >= PERF_EVENT_STATE_OFF) |
7438 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 7439 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
7440 | perf_install_in_context(dst_ctx, event, dst_cpu); |
7441 | get_ctx(dst_ctx); | |
7442 | } | |
7443 | mutex_unlock(&dst_ctx->mutex); | |
7444 | } | |
7445 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7446 | ||
cdd6c482 | 7447 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7448 | struct task_struct *child) |
d859e29f | 7449 | { |
cdd6c482 | 7450 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7451 | u64 child_val; |
d859e29f | 7452 | |
cdd6c482 IM |
7453 | if (child_event->attr.inherit_stat) |
7454 | perf_event_read_event(child_event, child); | |
38b200d6 | 7455 | |
b5e58793 | 7456 | child_val = perf_event_count(child_event); |
d859e29f PM |
7457 | |
7458 | /* | |
7459 | * Add back the child's count to the parent's count: | |
7460 | */ | |
a6e6dea6 | 7461 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7462 | atomic64_add(child_event->total_time_enabled, |
7463 | &parent_event->child_total_time_enabled); | |
7464 | atomic64_add(child_event->total_time_running, | |
7465 | &parent_event->child_total_time_running); | |
d859e29f PM |
7466 | |
7467 | /* | |
cdd6c482 | 7468 | * Remove this event from the parent's list |
d859e29f | 7469 | */ |
cdd6c482 IM |
7470 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7471 | mutex_lock(&parent_event->child_mutex); | |
7472 | list_del_init(&child_event->child_list); | |
7473 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
7474 | |
7475 | /* | |
cdd6c482 | 7476 | * Release the parent event, if this was the last |
d859e29f PM |
7477 | * reference to it. |
7478 | */ | |
a6fa941d | 7479 | put_event(parent_event); |
d859e29f PM |
7480 | } |
7481 | ||
9b51f66d | 7482 | static void |
cdd6c482 IM |
7483 | __perf_event_exit_task(struct perf_event *child_event, |
7484 | struct perf_event_context *child_ctx, | |
38b200d6 | 7485 | struct task_struct *child) |
9b51f66d | 7486 | { |
1903d50c PZ |
7487 | /* |
7488 | * Do not destroy the 'original' grouping; because of the context | |
7489 | * switch optimization the original events could've ended up in a | |
7490 | * random child task. | |
7491 | * | |
7492 | * If we were to destroy the original group, all group related | |
7493 | * operations would cease to function properly after this random | |
7494 | * child dies. | |
7495 | * | |
7496 | * Do destroy all inherited groups, we don't care about those | |
7497 | * and being thorough is better. | |
7498 | */ | |
7499 | perf_remove_from_context(child_event, !!child_event->parent); | |
0cc0c027 | 7500 | |
9b51f66d | 7501 | /* |
38b435b1 | 7502 | * It can happen that the parent exits first, and has events |
9b51f66d | 7503 | * that are still around due to the child reference. These |
38b435b1 | 7504 | * events need to be zapped. |
9b51f66d | 7505 | */ |
38b435b1 | 7506 | if (child_event->parent) { |
cdd6c482 IM |
7507 | sync_child_event(child_event, child); |
7508 | free_event(child_event); | |
4bcf349a | 7509 | } |
9b51f66d IM |
7510 | } |
7511 | ||
8dc85d54 | 7512 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7513 | { |
ebf905fc | 7514 | struct perf_event *child_event, *next; |
4a1c0f26 | 7515 | struct perf_event_context *child_ctx, *parent_ctx; |
a63eaf34 | 7516 | unsigned long flags; |
9b51f66d | 7517 | |
8dc85d54 | 7518 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7519 | perf_event_task(child, NULL, 0); |
9b51f66d | 7520 | return; |
9f498cc5 | 7521 | } |
9b51f66d | 7522 | |
a63eaf34 | 7523 | local_irq_save(flags); |
ad3a37de PM |
7524 | /* |
7525 | * We can't reschedule here because interrupts are disabled, | |
7526 | * and either child is current or it is a task that can't be | |
7527 | * scheduled, so we are now safe from rescheduling changing | |
7528 | * our context. | |
7529 | */ | |
806839b2 | 7530 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7531 | |
7532 | /* | |
7533 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7534 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7535 | * incremented the context's refcount before we do put_ctx below. |
7536 | */ | |
e625cce1 | 7537 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7538 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7539 | child->perf_event_ctxp[ctxn] = NULL; |
4a1c0f26 PZ |
7540 | |
7541 | /* | |
7542 | * In order to avoid freeing: child_ctx->parent_ctx->task | |
7543 | * under perf_event_context::lock, grab another reference. | |
7544 | */ | |
7545 | parent_ctx = child_ctx->parent_ctx; | |
7546 | if (parent_ctx) | |
7547 | get_ctx(parent_ctx); | |
7548 | ||
71a851b4 PZ |
7549 | /* |
7550 | * If this context is a clone; unclone it so it can't get | |
7551 | * swapped to another process while we're removing all | |
cdd6c482 | 7552 | * the events from it. |
71a851b4 PZ |
7553 | */ |
7554 | unclone_ctx(child_ctx); | |
5e942bb3 | 7555 | update_context_time(child_ctx); |
e625cce1 | 7556 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 | 7557 | |
4a1c0f26 PZ |
7558 | /* |
7559 | * Now that we no longer hold perf_event_context::lock, drop | |
7560 | * our extra child_ctx->parent_ctx reference. | |
7561 | */ | |
7562 | if (parent_ctx) | |
7563 | put_ctx(parent_ctx); | |
7564 | ||
9f498cc5 | 7565 | /* |
cdd6c482 IM |
7566 | * Report the task dead after unscheduling the events so that we |
7567 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7568 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7569 | */ |
cdd6c482 | 7570 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 7571 | |
66fff224 PZ |
7572 | /* |
7573 | * We can recurse on the same lock type through: | |
7574 | * | |
cdd6c482 IM |
7575 | * __perf_event_exit_task() |
7576 | * sync_child_event() | |
a6fa941d AV |
7577 | * put_event() |
7578 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
7579 | * |
7580 | * But since its the parent context it won't be the same instance. | |
7581 | */ | |
a0507c84 | 7582 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 7583 | |
ebf905fc | 7584 | list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) |
cdd6c482 | 7585 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 | 7586 | |
a63eaf34 PM |
7587 | mutex_unlock(&child_ctx->mutex); |
7588 | ||
7589 | put_ctx(child_ctx); | |
9b51f66d IM |
7590 | } |
7591 | ||
8dc85d54 PZ |
7592 | /* |
7593 | * When a child task exits, feed back event values to parent events. | |
7594 | */ | |
7595 | void perf_event_exit_task(struct task_struct *child) | |
7596 | { | |
8882135b | 7597 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
7598 | int ctxn; |
7599 | ||
8882135b PZ |
7600 | mutex_lock(&child->perf_event_mutex); |
7601 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
7602 | owner_entry) { | |
7603 | list_del_init(&event->owner_entry); | |
7604 | ||
7605 | /* | |
7606 | * Ensure the list deletion is visible before we clear | |
7607 | * the owner, closes a race against perf_release() where | |
7608 | * we need to serialize on the owner->perf_event_mutex. | |
7609 | */ | |
7610 | smp_wmb(); | |
7611 | event->owner = NULL; | |
7612 | } | |
7613 | mutex_unlock(&child->perf_event_mutex); | |
7614 | ||
8dc85d54 PZ |
7615 | for_each_task_context_nr(ctxn) |
7616 | perf_event_exit_task_context(child, ctxn); | |
7617 | } | |
7618 | ||
889ff015 FW |
7619 | static void perf_free_event(struct perf_event *event, |
7620 | struct perf_event_context *ctx) | |
7621 | { | |
7622 | struct perf_event *parent = event->parent; | |
7623 | ||
7624 | if (WARN_ON_ONCE(!parent)) | |
7625 | return; | |
7626 | ||
7627 | mutex_lock(&parent->child_mutex); | |
7628 | list_del_init(&event->child_list); | |
7629 | mutex_unlock(&parent->child_mutex); | |
7630 | ||
a6fa941d | 7631 | put_event(parent); |
889ff015 | 7632 | |
8a49542c | 7633 | perf_group_detach(event); |
889ff015 FW |
7634 | list_del_event(event, ctx); |
7635 | free_event(event); | |
7636 | } | |
7637 | ||
bbbee908 PZ |
7638 | /* |
7639 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 7640 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 7641 | */ |
cdd6c482 | 7642 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 7643 | { |
8dc85d54 | 7644 | struct perf_event_context *ctx; |
cdd6c482 | 7645 | struct perf_event *event, *tmp; |
8dc85d54 | 7646 | int ctxn; |
bbbee908 | 7647 | |
8dc85d54 PZ |
7648 | for_each_task_context_nr(ctxn) { |
7649 | ctx = task->perf_event_ctxp[ctxn]; | |
7650 | if (!ctx) | |
7651 | continue; | |
bbbee908 | 7652 | |
8dc85d54 | 7653 | mutex_lock(&ctx->mutex); |
bbbee908 | 7654 | again: |
8dc85d54 PZ |
7655 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
7656 | group_entry) | |
7657 | perf_free_event(event, ctx); | |
bbbee908 | 7658 | |
8dc85d54 PZ |
7659 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
7660 | group_entry) | |
7661 | perf_free_event(event, ctx); | |
bbbee908 | 7662 | |
8dc85d54 PZ |
7663 | if (!list_empty(&ctx->pinned_groups) || |
7664 | !list_empty(&ctx->flexible_groups)) | |
7665 | goto again; | |
bbbee908 | 7666 | |
8dc85d54 | 7667 | mutex_unlock(&ctx->mutex); |
bbbee908 | 7668 | |
8dc85d54 PZ |
7669 | put_ctx(ctx); |
7670 | } | |
889ff015 FW |
7671 | } |
7672 | ||
4e231c79 PZ |
7673 | void perf_event_delayed_put(struct task_struct *task) |
7674 | { | |
7675 | int ctxn; | |
7676 | ||
7677 | for_each_task_context_nr(ctxn) | |
7678 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7679 | } | |
7680 | ||
97dee4f3 PZ |
7681 | /* |
7682 | * inherit a event from parent task to child task: | |
7683 | */ | |
7684 | static struct perf_event * | |
7685 | inherit_event(struct perf_event *parent_event, | |
7686 | struct task_struct *parent, | |
7687 | struct perf_event_context *parent_ctx, | |
7688 | struct task_struct *child, | |
7689 | struct perf_event *group_leader, | |
7690 | struct perf_event_context *child_ctx) | |
7691 | { | |
7692 | struct perf_event *child_event; | |
cee010ec | 7693 | unsigned long flags; |
97dee4f3 PZ |
7694 | |
7695 | /* | |
7696 | * Instead of creating recursive hierarchies of events, | |
7697 | * we link inherited events back to the original parent, | |
7698 | * which has a filp for sure, which we use as the reference | |
7699 | * count: | |
7700 | */ | |
7701 | if (parent_event->parent) | |
7702 | parent_event = parent_event->parent; | |
7703 | ||
7704 | child_event = perf_event_alloc(&parent_event->attr, | |
7705 | parent_event->cpu, | |
d580ff86 | 7706 | child, |
97dee4f3 | 7707 | group_leader, parent_event, |
4dc0da86 | 7708 | NULL, NULL); |
97dee4f3 PZ |
7709 | if (IS_ERR(child_event)) |
7710 | return child_event; | |
a6fa941d AV |
7711 | |
7712 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7713 | free_event(child_event); | |
7714 | return NULL; | |
7715 | } | |
7716 | ||
97dee4f3 PZ |
7717 | get_ctx(child_ctx); |
7718 | ||
7719 | /* | |
7720 | * Make the child state follow the state of the parent event, | |
7721 | * not its attr.disabled bit. We hold the parent's mutex, | |
7722 | * so we won't race with perf_event_{en, dis}able_family. | |
7723 | */ | |
7724 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7725 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7726 | else | |
7727 | child_event->state = PERF_EVENT_STATE_OFF; | |
7728 | ||
7729 | if (parent_event->attr.freq) { | |
7730 | u64 sample_period = parent_event->hw.sample_period; | |
7731 | struct hw_perf_event *hwc = &child_event->hw; | |
7732 | ||
7733 | hwc->sample_period = sample_period; | |
7734 | hwc->last_period = sample_period; | |
7735 | ||
7736 | local64_set(&hwc->period_left, sample_period); | |
7737 | } | |
7738 | ||
7739 | child_event->ctx = child_ctx; | |
7740 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7741 | child_event->overflow_handler_context |
7742 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7743 | |
614b6780 TG |
7744 | /* |
7745 | * Precalculate sample_data sizes | |
7746 | */ | |
7747 | perf_event__header_size(child_event); | |
6844c09d | 7748 | perf_event__id_header_size(child_event); |
614b6780 | 7749 | |
97dee4f3 PZ |
7750 | /* |
7751 | * Link it up in the child's context: | |
7752 | */ | |
cee010ec | 7753 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7754 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7755 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7756 | |
97dee4f3 PZ |
7757 | /* |
7758 | * Link this into the parent event's child list | |
7759 | */ | |
7760 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7761 | mutex_lock(&parent_event->child_mutex); | |
7762 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7763 | mutex_unlock(&parent_event->child_mutex); | |
7764 | ||
7765 | return child_event; | |
7766 | } | |
7767 | ||
7768 | static int inherit_group(struct perf_event *parent_event, | |
7769 | struct task_struct *parent, | |
7770 | struct perf_event_context *parent_ctx, | |
7771 | struct task_struct *child, | |
7772 | struct perf_event_context *child_ctx) | |
7773 | { | |
7774 | struct perf_event *leader; | |
7775 | struct perf_event *sub; | |
7776 | struct perf_event *child_ctr; | |
7777 | ||
7778 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7779 | child, NULL, child_ctx); | |
7780 | if (IS_ERR(leader)) | |
7781 | return PTR_ERR(leader); | |
7782 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7783 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7784 | child, leader, child_ctx); | |
7785 | if (IS_ERR(child_ctr)) | |
7786 | return PTR_ERR(child_ctr); | |
7787 | } | |
7788 | return 0; | |
889ff015 FW |
7789 | } |
7790 | ||
7791 | static int | |
7792 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7793 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7794 | struct task_struct *child, int ctxn, |
889ff015 FW |
7795 | int *inherited_all) |
7796 | { | |
7797 | int ret; | |
8dc85d54 | 7798 | struct perf_event_context *child_ctx; |
889ff015 FW |
7799 | |
7800 | if (!event->attr.inherit) { | |
7801 | *inherited_all = 0; | |
7802 | return 0; | |
bbbee908 PZ |
7803 | } |
7804 | ||
fe4b04fa | 7805 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7806 | if (!child_ctx) { |
7807 | /* | |
7808 | * This is executed from the parent task context, so | |
7809 | * inherit events that have been marked for cloning. | |
7810 | * First allocate and initialize a context for the | |
7811 | * child. | |
7812 | */ | |
bbbee908 | 7813 | |
734df5ab | 7814 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
7815 | if (!child_ctx) |
7816 | return -ENOMEM; | |
bbbee908 | 7817 | |
8dc85d54 | 7818 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7819 | } |
7820 | ||
7821 | ret = inherit_group(event, parent, parent_ctx, | |
7822 | child, child_ctx); | |
7823 | ||
7824 | if (ret) | |
7825 | *inherited_all = 0; | |
7826 | ||
7827 | return ret; | |
bbbee908 PZ |
7828 | } |
7829 | ||
9b51f66d | 7830 | /* |
cdd6c482 | 7831 | * Initialize the perf_event context in task_struct |
9b51f66d | 7832 | */ |
985c8dcb | 7833 | static int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7834 | { |
889ff015 | 7835 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7836 | struct perf_event_context *cloned_ctx; |
7837 | struct perf_event *event; | |
9b51f66d | 7838 | struct task_struct *parent = current; |
564c2b21 | 7839 | int inherited_all = 1; |
dddd3379 | 7840 | unsigned long flags; |
6ab423e0 | 7841 | int ret = 0; |
9b51f66d | 7842 | |
8dc85d54 | 7843 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7844 | return 0; |
7845 | ||
ad3a37de | 7846 | /* |
25346b93 PM |
7847 | * If the parent's context is a clone, pin it so it won't get |
7848 | * swapped under us. | |
ad3a37de | 7849 | */ |
8dc85d54 | 7850 | parent_ctx = perf_pin_task_context(parent, ctxn); |
ffb4ef21 PZ |
7851 | if (!parent_ctx) |
7852 | return 0; | |
25346b93 | 7853 | |
ad3a37de PM |
7854 | /* |
7855 | * No need to check if parent_ctx != NULL here; since we saw | |
7856 | * it non-NULL earlier, the only reason for it to become NULL | |
7857 | * is if we exit, and since we're currently in the middle of | |
7858 | * a fork we can't be exiting at the same time. | |
7859 | */ | |
ad3a37de | 7860 | |
9b51f66d IM |
7861 | /* |
7862 | * Lock the parent list. No need to lock the child - not PID | |
7863 | * hashed yet and not running, so nobody can access it. | |
7864 | */ | |
d859e29f | 7865 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7866 | |
7867 | /* | |
7868 | * We dont have to disable NMIs - we are only looking at | |
7869 | * the list, not manipulating it: | |
7870 | */ | |
889ff015 | 7871 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7872 | ret = inherit_task_group(event, parent, parent_ctx, |
7873 | child, ctxn, &inherited_all); | |
889ff015 FW |
7874 | if (ret) |
7875 | break; | |
7876 | } | |
b93f7978 | 7877 | |
dddd3379 TG |
7878 | /* |
7879 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7880 | * to allocations, but we need to prevent rotation because | |
7881 | * rotate_ctx() will change the list from interrupt context. | |
7882 | */ | |
7883 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7884 | parent_ctx->rotate_disable = 1; | |
7885 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7886 | ||
889ff015 | 7887 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7888 | ret = inherit_task_group(event, parent, parent_ctx, |
7889 | child, ctxn, &inherited_all); | |
889ff015 | 7890 | if (ret) |
9b51f66d | 7891 | break; |
564c2b21 PM |
7892 | } |
7893 | ||
dddd3379 TG |
7894 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7895 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7896 | |
8dc85d54 | 7897 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7898 | |
05cbaa28 | 7899 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7900 | /* |
7901 | * Mark the child context as a clone of the parent | |
7902 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7903 | * |
7904 | * Note that if the parent is a clone, the holding of | |
7905 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7906 | */ |
c5ed5145 | 7907 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7908 | if (cloned_ctx) { |
7909 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7910 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7911 | } else { |
7912 | child_ctx->parent_ctx = parent_ctx; | |
7913 | child_ctx->parent_gen = parent_ctx->generation; | |
7914 | } | |
7915 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7916 | } |
7917 | ||
c5ed5145 | 7918 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7919 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7920 | |
25346b93 | 7921 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7922 | put_ctx(parent_ctx); |
ad3a37de | 7923 | |
6ab423e0 | 7924 | return ret; |
9b51f66d IM |
7925 | } |
7926 | ||
8dc85d54 PZ |
7927 | /* |
7928 | * Initialize the perf_event context in task_struct | |
7929 | */ | |
7930 | int perf_event_init_task(struct task_struct *child) | |
7931 | { | |
7932 | int ctxn, ret; | |
7933 | ||
8550d7cb ON |
7934 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7935 | mutex_init(&child->perf_event_mutex); | |
7936 | INIT_LIST_HEAD(&child->perf_event_list); | |
7937 | ||
8dc85d54 PZ |
7938 | for_each_task_context_nr(ctxn) { |
7939 | ret = perf_event_init_context(child, ctxn); | |
7940 | if (ret) | |
7941 | return ret; | |
7942 | } | |
7943 | ||
7944 | return 0; | |
7945 | } | |
7946 | ||
220b140b PM |
7947 | static void __init perf_event_init_all_cpus(void) |
7948 | { | |
b28ab83c | 7949 | struct swevent_htable *swhash; |
220b140b | 7950 | int cpu; |
220b140b PM |
7951 | |
7952 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7953 | swhash = &per_cpu(swevent_htable, cpu); |
7954 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7955 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7956 | } |
7957 | } | |
7958 | ||
0db0628d | 7959 | static void perf_event_init_cpu(int cpu) |
0793a61d | 7960 | { |
108b02cf | 7961 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7962 | |
b28ab83c | 7963 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 7964 | swhash->online = true; |
4536e4d1 | 7965 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7966 | struct swevent_hlist *hlist; |
7967 | ||
b28ab83c PZ |
7968 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7969 | WARN_ON(!hlist); | |
7970 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7971 | } |
b28ab83c | 7972 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7973 | } |
7974 | ||
c277443c | 7975 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7976 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7977 | { |
e9d2b064 PZ |
7978 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7979 | ||
7980 | WARN_ON(!irqs_disabled()); | |
7981 | ||
7982 | list_del_init(&cpuctx->rotation_list); | |
7983 | } | |
7984 | ||
108b02cf | 7985 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7986 | { |
46ce0fe9 | 7987 | struct remove_event re = { .detach_group = false }; |
108b02cf | 7988 | struct perf_event_context *ctx = __info; |
0793a61d | 7989 | |
108b02cf | 7990 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7991 | |
e3703f8c | 7992 | rcu_read_lock(); |
46ce0fe9 PZ |
7993 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
7994 | __perf_remove_from_context(&re); | |
e3703f8c | 7995 | rcu_read_unlock(); |
0793a61d | 7996 | } |
108b02cf PZ |
7997 | |
7998 | static void perf_event_exit_cpu_context(int cpu) | |
7999 | { | |
8000 | struct perf_event_context *ctx; | |
8001 | struct pmu *pmu; | |
8002 | int idx; | |
8003 | ||
8004 | idx = srcu_read_lock(&pmus_srcu); | |
8005 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 8006 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
8007 | |
8008 | mutex_lock(&ctx->mutex); | |
8009 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
8010 | mutex_unlock(&ctx->mutex); | |
8011 | } | |
8012 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
8013 | } |
8014 | ||
cdd6c482 | 8015 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 8016 | { |
b28ab83c | 8017 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 8018 | |
e3703f8c PZ |
8019 | perf_event_exit_cpu_context(cpu); |
8020 | ||
b28ab83c | 8021 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8022 | swhash->online = false; |
b28ab83c PZ |
8023 | swevent_hlist_release(swhash); |
8024 | mutex_unlock(&swhash->hlist_mutex); | |
0793a61d TG |
8025 | } |
8026 | #else | |
cdd6c482 | 8027 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
8028 | #endif |
8029 | ||
c277443c PZ |
8030 | static int |
8031 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
8032 | { | |
8033 | int cpu; | |
8034 | ||
8035 | for_each_online_cpu(cpu) | |
8036 | perf_event_exit_cpu(cpu); | |
8037 | ||
8038 | return NOTIFY_OK; | |
8039 | } | |
8040 | ||
8041 | /* | |
8042 | * Run the perf reboot notifier at the very last possible moment so that | |
8043 | * the generic watchdog code runs as long as possible. | |
8044 | */ | |
8045 | static struct notifier_block perf_reboot_notifier = { | |
8046 | .notifier_call = perf_reboot, | |
8047 | .priority = INT_MIN, | |
8048 | }; | |
8049 | ||
0db0628d | 8050 | static int |
0793a61d TG |
8051 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
8052 | { | |
8053 | unsigned int cpu = (long)hcpu; | |
8054 | ||
4536e4d1 | 8055 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
8056 | |
8057 | case CPU_UP_PREPARE: | |
5e11637e | 8058 | case CPU_DOWN_FAILED: |
cdd6c482 | 8059 | perf_event_init_cpu(cpu); |
0793a61d TG |
8060 | break; |
8061 | ||
5e11637e | 8062 | case CPU_UP_CANCELED: |
0793a61d | 8063 | case CPU_DOWN_PREPARE: |
cdd6c482 | 8064 | perf_event_exit_cpu(cpu); |
0793a61d | 8065 | break; |
0793a61d TG |
8066 | default: |
8067 | break; | |
8068 | } | |
8069 | ||
8070 | return NOTIFY_OK; | |
8071 | } | |
8072 | ||
cdd6c482 | 8073 | void __init perf_event_init(void) |
0793a61d | 8074 | { |
3c502e7a JW |
8075 | int ret; |
8076 | ||
2e80a82a PZ |
8077 | idr_init(&pmu_idr); |
8078 | ||
220b140b | 8079 | perf_event_init_all_cpus(); |
b0a873eb | 8080 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
8081 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
8082 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
8083 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
8084 | perf_tp_register(); |
8085 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 8086 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
8087 | |
8088 | ret = init_hw_breakpoint(); | |
8089 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
8090 | |
8091 | /* do not patch jump label more than once per second */ | |
8092 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
8093 | |
8094 | /* | |
8095 | * Build time assertion that we keep the data_head at the intended | |
8096 | * location. IOW, validation we got the __reserved[] size right. | |
8097 | */ | |
8098 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
8099 | != 1024); | |
0793a61d | 8100 | } |
abe43400 PZ |
8101 | |
8102 | static int __init perf_event_sysfs_init(void) | |
8103 | { | |
8104 | struct pmu *pmu; | |
8105 | int ret; | |
8106 | ||
8107 | mutex_lock(&pmus_lock); | |
8108 | ||
8109 | ret = bus_register(&pmu_bus); | |
8110 | if (ret) | |
8111 | goto unlock; | |
8112 | ||
8113 | list_for_each_entry(pmu, &pmus, entry) { | |
8114 | if (!pmu->name || pmu->type < 0) | |
8115 | continue; | |
8116 | ||
8117 | ret = pmu_dev_alloc(pmu); | |
8118 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
8119 | } | |
8120 | pmu_bus_running = 1; | |
8121 | ret = 0; | |
8122 | ||
8123 | unlock: | |
8124 | mutex_unlock(&pmus_lock); | |
8125 | ||
8126 | return ret; | |
8127 | } | |
8128 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
8129 | |
8130 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
8131 | static struct cgroup_subsys_state * |
8132 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
8133 | { |
8134 | struct perf_cgroup *jc; | |
e5d1367f | 8135 | |
1b15d055 | 8136 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
8137 | if (!jc) |
8138 | return ERR_PTR(-ENOMEM); | |
8139 | ||
e5d1367f SE |
8140 | jc->info = alloc_percpu(struct perf_cgroup_info); |
8141 | if (!jc->info) { | |
8142 | kfree(jc); | |
8143 | return ERR_PTR(-ENOMEM); | |
8144 | } | |
8145 | ||
e5d1367f SE |
8146 | return &jc->css; |
8147 | } | |
8148 | ||
eb95419b | 8149 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 8150 | { |
eb95419b TH |
8151 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
8152 | ||
e5d1367f SE |
8153 | free_percpu(jc->info); |
8154 | kfree(jc); | |
8155 | } | |
8156 | ||
8157 | static int __perf_cgroup_move(void *info) | |
8158 | { | |
8159 | struct task_struct *task = info; | |
8160 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
8161 | return 0; | |
8162 | } | |
8163 | ||
eb95419b TH |
8164 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
8165 | struct cgroup_taskset *tset) | |
e5d1367f | 8166 | { |
bb9d97b6 TH |
8167 | struct task_struct *task; |
8168 | ||
924f0d9a | 8169 | cgroup_taskset_for_each(task, tset) |
bb9d97b6 | 8170 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8171 | } |
8172 | ||
eb95419b TH |
8173 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
8174 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 8175 | struct task_struct *task) |
e5d1367f SE |
8176 | { |
8177 | /* | |
8178 | * cgroup_exit() is called in the copy_process() failure path. | |
8179 | * Ignore this case since the task hasn't ran yet, this avoids | |
8180 | * trying to poke a half freed task state from generic code. | |
8181 | */ | |
8182 | if (!(task->flags & PF_EXITING)) | |
8183 | return; | |
8184 | ||
bb9d97b6 | 8185 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8186 | } |
8187 | ||
073219e9 | 8188 | struct cgroup_subsys perf_event_cgrp_subsys = { |
92fb9748 TH |
8189 | .css_alloc = perf_cgroup_css_alloc, |
8190 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 8191 | .exit = perf_cgroup_exit, |
bb9d97b6 | 8192 | .attach = perf_cgroup_attach, |
e5d1367f SE |
8193 | }; |
8194 | #endif /* CONFIG_CGROUP_PERF */ |